diff --git a/optimized/ami/100ms/ls_eend_ami_100ms.mlmodelc/analytics/coremldata.bin b/optimized/ami/100ms/ls_eend_ami_100ms.mlmodelc/analytics/coremldata.bin index 823d91cf23a2bb454a5b6dd99a9ae05da88006bc..e336d5a767baed9689302dba203620df3bd2beb7 100644 --- a/optimized/ami/100ms/ls_eend_ami_100ms.mlmodelc/analytics/coremldata.bin +++ b/optimized/ami/100ms/ls_eend_ami_100ms.mlmodelc/analytics/coremldata.bin @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:9880c37e22316efbaf285d84d1a4eecab31657806256346bd1939fcf4a775924 +oid sha256:acd29bd7bc0274452e77f5735e705d136916e5bd978ca14d6a383c7a2018618a size 243 diff --git a/optimized/ami/100ms/ls_eend_ami_100ms.mlmodelc/coremldata.bin b/optimized/ami/100ms/ls_eend_ami_100ms.mlmodelc/coremldata.bin index 6eed47e0dd8f2c62579405217a9755f463661f78..56d90a939a4077bbce418ac68249e3597338a6c8 100644 --- a/optimized/ami/100ms/ls_eend_ami_100ms.mlmodelc/coremldata.bin +++ b/optimized/ami/100ms/ls_eend_ami_100ms.mlmodelc/coremldata.bin @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:914dd099519cff6e07d71c8a29081c50c800030c21d635e1a226ec7286819c8d -size 1292 +oid sha256:2c45f751d1fdb81520abae505c4a969a6b817d816487ce6e901e0e02e60a5f0b +size 1395 diff --git a/optimized/ami/100ms/ls_eend_ami_100ms.mlmodelc/metadata.json b/optimized/ami/100ms/ls_eend_ami_100ms.mlmodelc/metadata.json index 3dc3df54de7517ae88aad148997bb18d28dc5797..c4e7d5f6666d406a128108a844274d48bb01a2ae 100644 --- a/optimized/ami/100ms/ls_eend_ami_100ms.mlmodelc/metadata.json +++ b/optimized/ami/100ms/ls_eend_ami_100ms.mlmodelc/metadata.json @@ -1,7 +1,7 @@ [ { "metadataOutputVersion" : "3.0", - "shortDescription" : "LS-EEND AMI streaming diarizer (pipeline, T=1, max_speakers=4)", + "shortDescription" : "LS-EEND AMI streaming diarizer (pipeline, T=1, max_speakers=4, layout=raw_mel, cu=all)", "outputSchema" : [ { "hasShapeFlexibility" : "0", @@ -81,12 +81,12 @@ "specificationVersion" : 8, "mlProgramOperationTypeHistogram" : { "Ios16.reduceL2Norm" : 2, - "Ios17.reshape" : 66, + "Ios17.reshape" : 67, "Ios16.softmax" : 2, "Ios17.matmul" : 29, "Ios17.transpose" : 57, "Split" : 4, - "Ios17.expandDims" : 3, + "Ios17.expandDims" : 4, "Ios17.add" : 46, "Ios16.sigmoid" : 5, "Ios17.sliceByIndex" : 36, @@ -128,9 +128,9 @@ "hasShapeFlexibility" : "0", "isOptional" : "0", "dataType" : "Float32", - "formattedType" : "MultiArray (Float32 1 × 1 × 345)", + "formattedType" : "MultiArray (Float32 1 × 15 × 23)", "shortDescription" : "", - "shape" : "[1, 1, 345]", + "shape" : "[1, 15, 23]", "name" : "features", "type" : "MultiArray" }, @@ -206,8 +206,8 @@ } ], "userDefinedMetadata" : { - "com.github.apple.coremltools.conversion_date" : "2026-04-16", - "config" : "{\"model_name\": \"ami\", \"model_label\": \"AMI\", \"variant\": \"pipeline\", \"chunk_size\": 1, \"step_duration_ms\": 100, \"frame_duration_ms\": 100, \"frame_duration_seconds\": 0.1, \"max_speakers\": 4, \"max_nspks\": 6, \"n_units\": 256, \"n_heads\": 4, \"enc_n_layers\": 4, \"dec_n_layers\": 2, \"conv_kernel_size\": 16, \"conv_delay\": 9, \"sample_rate\": 8000, \"win_length\": 200, \"hop_length\": 80, \"n_mels\": 23, \"context_size\": 7, \"subsampling\": 10, \"feat_type\": \"logmel23_cummn\", \"pure_roll\": true}", + "com.github.apple.coremltools.conversion_date" : "2026-04-18", + "config" : "{\"model_name\": \"ami\", \"model_label\": \"AMI\", \"variant\": \"pipeline\", \"chunk_size\": 1, \"step_duration_ms\": 100, \"frame_duration_ms\": 100, \"frame_duration_seconds\": 0.1, \"max_speakers\": 4, \"max_nspks\": 6, \"n_units\": 256, \"n_heads\": 4, \"enc_n_layers\": 4, \"dec_n_layers\": 2, \"conv_kernel_size\": 16, \"conv_delay\": 9, \"sample_rate\": 8000, \"win_length\": 200, \"hop_length\": 80, \"n_mels\": 23, \"context_size\": 7, \"subsampling\": 10, \"feat_type\": \"logmel23_cummn\", \"pure_roll\": true, \"input_layout\": \"raw_mel\", \"compute_units_export\": \"all\", \"raw_mel_length\": 15}", "com.github.apple.coremltools.source" : "torch==2.6.0", "com.github.apple.coremltools.version" : "9.0", "com.github.apple.coremltools.source_dialect" : "TorchScript" diff --git a/optimized/ami/100ms/ls_eend_ami_100ms.mlmodelc/model.mil b/optimized/ami/100ms/ls_eend_ami_100ms.mlmodelc/model.mil index 69b8fb57d7fa52c2c479b44e6c51ee2bf32cccf4..2ac962d651b9b73a7bea06a0e9008c399748989e 100644 --- a/optimized/ami/100ms/ls_eend_ami_100ms.mlmodelc/model.mil +++ b/optimized/ami/100ms/ls_eend_ami_100ms.mlmodelc/model.mil @@ -1,233 +1,239 @@ program(1.0) [buildInfo = dict, tensor>({{"coremlc-component-MIL", "3510.2.1"}, {"coremlc-version", "3500.32.1"}, {"coremltools-component-torch", "2.6.0"}, {"coremltools-source-dialect", "TorchScript"}, {"coremltools-version", "9.0"}})] { - func main(tensor cnn_window, tensor dec_kv, tensor dec_scale, tensor enc_conv_cache, tensor enc_kv, tensor enc_scale, tensor features, tensor valid_mask) { - tensor encoder_cnn_bias = const()[name = tensor("encoder_cnn_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(64)))]; - tensor encoder_cnn_weight = const()[name = tensor("encoder_cnn_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1152)))]; - tensor encoder__causal_mask = const()[name = tensor("encoder__causal_mask"), val = tensor([[0x1p+0]])]; - tensor encoder__t_index = const()[name = tensor("encoder__t_index"), val = tensor([0x1p+0])]; - tensor encoder_input_projection_linear_bias = const()[name = tensor("encoder_input_projection_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4981952)))]; - tensor encoder_input_projection_linear_weight = const()[name = tensor("encoder_input_projection_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4983040)))]; - tensor encoder_pre_layer_norm_bias = const()[name = tensor("encoder_pre_layer_norm_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(5336384)))]; - tensor encoder_pre_layer_norm_weight = const()[name = tensor("encoder_pre_layer_norm_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(5337472)))]; - tensor encoder_ffn1_0_module_sequential_0_bias = const()[name = tensor("encoder_ffn1_0_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(5338560)))]; - tensor encoder_ffn1_0_module_sequential_0_weight = const()[name = tensor("encoder_ffn1_0_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(5339648)))]; - tensor encoder_ffn1_0_module_sequential_1_linear_bias = const()[name = tensor("encoder_ffn1_0_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(5340736)))]; - tensor encoder_ffn1_0_module_sequential_1_linear_weight = const()[name = tensor("encoder_ffn1_0_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(5344896)))]; - tensor encoder_ffn1_0_module_sequential_4_linear_bias = const()[name = tensor("encoder_ffn1_0_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(6393536)))]; - tensor encoder_ffn1_0_module_sequential_4_linear_weight = const()[name = tensor("encoder_ffn1_0_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(6394624)))]; - tensor encoder_ret_lns_0_bias = const()[name = tensor("encoder_ret_lns_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7443264)))]; - tensor encoder_ret_lns_0_weight = const()[name = tensor("encoder_ret_lns_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7444352)))]; - tensor encoder_q_proj_0_bias = const()[name = tensor("encoder_q_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7445440)))]; - tensor encoder_q_proj_0_weight = const()[name = tensor("encoder_q_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7446528)))]; - tensor encoder_k_proj_0_bias = const()[name = tensor("encoder_k_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7708736)))]; - tensor encoder_k_proj_0_weight = const()[name = tensor("encoder_k_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7709824)))]; - tensor encoder_v_proj_0_bias = const()[name = tensor("encoder_v_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7972032)))]; - tensor encoder_v_proj_0_weight = const()[name = tensor("encoder_v_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7973120)))]; - tensor encoder_g_proj_0_bias = const()[name = tensor("encoder_g_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8235328)))]; - tensor encoder_g_proj_0_weight = const()[name = tensor("encoder_g_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8236416)))]; - tensor encoder_out_proj_0_bias = const()[name = tensor("encoder_out_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8498624)))]; - tensor encoder_out_proj_0_weight = const()[name = tensor("encoder_out_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8499712)))]; - tensor encoder_conv_module_0_sequential_0_bias = const()[name = tensor("encoder_conv_module_0_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8761920)))]; - tensor encoder_conv_module_0_sequential_0_weight = const()[name = tensor("encoder_conv_module_0_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8763008)))]; - tensor encoder_conv_module_0_sequential_2_conv_bias = const()[name = tensor("encoder_conv_module_0_sequential_2_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8764096)))]; - tensor encoder_conv_module_0_sequential_2_conv_weight = const()[name = tensor("encoder_conv_module_0_sequential_2_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8766208)))]; - tensor encoder_conv_module_0_sequential_4_conv_weight = const()[name = tensor("encoder_conv_module_0_sequential_4_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9290560)))]; - tensor encoder_conv_module_0_sequential_5_running_var = const()[name = tensor("encoder_conv_module_0_sequential_5_running_var"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9307008)))]; - tensor encoder_conv_module_0_sequential_5_running_mean = const()[name = tensor("encoder_conv_module_0_sequential_5_running_mean"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9308096)))]; - tensor encoder_conv_module_0_sequential_5_bias = const()[name = tensor("encoder_conv_module_0_sequential_5_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9309184)))]; - tensor encoder_conv_module_0_sequential_5_weight = const()[name = tensor("encoder_conv_module_0_sequential_5_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9310272)))]; - tensor encoder_conv_module_0_sequential_7_conv_bias = const()[name = tensor("encoder_conv_module_0_sequential_7_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9311360)))]; - tensor encoder_conv_module_0_sequential_7_conv_weight = const()[name = tensor("encoder_conv_module_0_sequential_7_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9312448)))]; - tensor encoder_ffn2_0_module_sequential_0_bias = const()[name = tensor("encoder_ffn2_0_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9574656)))]; - tensor encoder_ffn2_0_module_sequential_0_weight = const()[name = tensor("encoder_ffn2_0_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9575744)))]; - tensor encoder_ffn2_0_module_sequential_1_linear_bias = const()[name = tensor("encoder_ffn2_0_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9576832)))]; - tensor encoder_ffn2_0_module_sequential_1_linear_weight = const()[name = tensor("encoder_ffn2_0_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9580992)))]; - tensor encoder_ffn2_0_module_sequential_4_linear_bias = const()[name = tensor("encoder_ffn2_0_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(10629632)))]; - tensor encoder_ffn2_0_module_sequential_4_linear_weight = const()[name = tensor("encoder_ffn2_0_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(10630720)))]; - tensor encoder_layer_norm_0_bias = const()[name = tensor("encoder_layer_norm_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11679360)))]; - tensor encoder_layer_norm_0_weight = const()[name = tensor("encoder_layer_norm_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11680448)))]; - tensor encoder_ffn1_1_module_sequential_0_bias = const()[name = tensor("encoder_ffn1_1_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11681536)))]; - tensor encoder_ffn1_1_module_sequential_0_weight = const()[name = tensor("encoder_ffn1_1_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11682624)))]; - tensor encoder_ffn1_1_module_sequential_1_linear_bias = const()[name = tensor("encoder_ffn1_1_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11683712)))]; - tensor encoder_ffn1_1_module_sequential_1_linear_weight = const()[name = tensor("encoder_ffn1_1_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11687872)))]; - tensor encoder_ffn1_1_module_sequential_4_linear_bias = const()[name = tensor("encoder_ffn1_1_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(12736512)))]; - tensor encoder_ffn1_1_module_sequential_4_linear_weight = const()[name = tensor("encoder_ffn1_1_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(12737600)))]; - tensor encoder_ret_lns_1_bias = const()[name = tensor("encoder_ret_lns_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(13786240)))]; - tensor encoder_ret_lns_1_weight = const()[name = tensor("encoder_ret_lns_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(13787328)))]; - tensor encoder_q_proj_1_bias = const()[name = tensor("encoder_q_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(13788416)))]; - tensor encoder_q_proj_1_weight = const()[name = tensor("encoder_q_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(13789504)))]; - tensor encoder_k_proj_1_bias = const()[name = tensor("encoder_k_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14051712)))]; - tensor encoder_k_proj_1_weight = const()[name = tensor("encoder_k_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14052800)))]; - tensor encoder_v_proj_1_bias = const()[name = tensor("encoder_v_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14315008)))]; - tensor encoder_v_proj_1_weight = const()[name = tensor("encoder_v_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14316096)))]; - tensor encoder_g_proj_1_bias = const()[name = tensor("encoder_g_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14578304)))]; - tensor encoder_g_proj_1_weight = const()[name = tensor("encoder_g_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14579392)))]; - tensor encoder_out_proj_1_bias = const()[name = tensor("encoder_out_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14841600)))]; - tensor encoder_out_proj_1_weight = const()[name = tensor("encoder_out_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14842688)))]; - tensor encoder_conv_module_1_sequential_0_bias = const()[name = tensor("encoder_conv_module_1_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15104896)))]; - tensor encoder_conv_module_1_sequential_0_weight = const()[name = tensor("encoder_conv_module_1_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15105984)))]; - tensor encoder_conv_module_1_sequential_2_conv_bias = const()[name = tensor("encoder_conv_module_1_sequential_2_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15107072)))]; - tensor encoder_conv_module_1_sequential_2_conv_weight = const()[name = tensor("encoder_conv_module_1_sequential_2_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15109184)))]; - tensor encoder_conv_module_1_sequential_4_conv_weight = const()[name = tensor("encoder_conv_module_1_sequential_4_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15633536)))]; - tensor encoder_conv_module_1_sequential_5_running_var = const()[name = tensor("encoder_conv_module_1_sequential_5_running_var"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15649984)))]; - tensor encoder_conv_module_1_sequential_5_running_mean = const()[name = tensor("encoder_conv_module_1_sequential_5_running_mean"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15651072)))]; - tensor encoder_conv_module_1_sequential_5_bias = const()[name = tensor("encoder_conv_module_1_sequential_5_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15652160)))]; - tensor encoder_conv_module_1_sequential_5_weight = const()[name = tensor("encoder_conv_module_1_sequential_5_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15653248)))]; - tensor encoder_conv_module_1_sequential_7_conv_bias = const()[name = tensor("encoder_conv_module_1_sequential_7_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15654336)))]; - tensor encoder_conv_module_1_sequential_7_conv_weight = const()[name = tensor("encoder_conv_module_1_sequential_7_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15655424)))]; - tensor encoder_ffn2_1_module_sequential_0_bias = const()[name = tensor("encoder_ffn2_1_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15917632)))]; - tensor encoder_ffn2_1_module_sequential_0_weight = const()[name = tensor("encoder_ffn2_1_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15918720)))]; - tensor encoder_ffn2_1_module_sequential_1_linear_bias = const()[name = tensor("encoder_ffn2_1_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15919808)))]; - tensor encoder_ffn2_1_module_sequential_1_linear_weight = const()[name = tensor("encoder_ffn2_1_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15923968)))]; - tensor encoder_ffn2_1_module_sequential_4_linear_bias = const()[name = tensor("encoder_ffn2_1_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(16972608)))]; - tensor encoder_ffn2_1_module_sequential_4_linear_weight = const()[name = tensor("encoder_ffn2_1_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(16973696)))]; - tensor encoder_layer_norm_1_bias = const()[name = tensor("encoder_layer_norm_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18022336)))]; - tensor encoder_layer_norm_1_weight = const()[name = tensor("encoder_layer_norm_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18023424)))]; - tensor encoder_ffn1_2_module_sequential_0_bias = const()[name = tensor("encoder_ffn1_2_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18024512)))]; - tensor encoder_ffn1_2_module_sequential_0_weight = const()[name = tensor("encoder_ffn1_2_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18025600)))]; - tensor encoder_ffn1_2_module_sequential_1_linear_bias = const()[name = tensor("encoder_ffn1_2_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18026688)))]; - tensor encoder_ffn1_2_module_sequential_1_linear_weight = const()[name = tensor("encoder_ffn1_2_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18030848)))]; - tensor encoder_ffn1_2_module_sequential_4_linear_bias = const()[name = tensor("encoder_ffn1_2_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(19079488)))]; - tensor encoder_ffn1_2_module_sequential_4_linear_weight = const()[name = tensor("encoder_ffn1_2_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(19080576)))]; - tensor encoder_ret_lns_2_bias = const()[name = tensor("encoder_ret_lns_2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20129216)))]; - tensor encoder_ret_lns_2_weight = const()[name = tensor("encoder_ret_lns_2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20130304)))]; - tensor encoder_q_proj_2_bias = const()[name = tensor("encoder_q_proj_2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20131392)))]; - tensor encoder_q_proj_2_weight = const()[name = tensor("encoder_q_proj_2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20132480)))]; - tensor encoder_k_proj_2_bias = const()[name = tensor("encoder_k_proj_2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20394688)))]; - tensor encoder_k_proj_2_weight = const()[name = tensor("encoder_k_proj_2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20395776)))]; - tensor encoder_v_proj_2_bias = const()[name = tensor("encoder_v_proj_2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20657984)))]; - tensor encoder_v_proj_2_weight = const()[name = tensor("encoder_v_proj_2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20659072)))]; - tensor encoder_g_proj_2_bias = const()[name = tensor("encoder_g_proj_2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20921280)))]; - tensor encoder_g_proj_2_weight = const()[name = tensor("encoder_g_proj_2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20922368)))]; - tensor encoder_out_proj_2_bias = const()[name = tensor("encoder_out_proj_2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21184576)))]; - tensor encoder_out_proj_2_weight = const()[name = tensor("encoder_out_proj_2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21185664)))]; - tensor encoder_conv_module_2_sequential_0_bias = const()[name = tensor("encoder_conv_module_2_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21447872)))]; - tensor encoder_conv_module_2_sequential_0_weight = const()[name = tensor("encoder_conv_module_2_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21448960)))]; - tensor encoder_conv_module_2_sequential_2_conv_bias = const()[name = tensor("encoder_conv_module_2_sequential_2_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21450048)))]; - tensor encoder_conv_module_2_sequential_2_conv_weight = const()[name = tensor("encoder_conv_module_2_sequential_2_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21452160)))]; - tensor encoder_conv_module_2_sequential_4_conv_weight = const()[name = tensor("encoder_conv_module_2_sequential_4_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21976512)))]; - tensor encoder_conv_module_2_sequential_5_running_var = const()[name = tensor("encoder_conv_module_2_sequential_5_running_var"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21992960)))]; - tensor encoder_conv_module_2_sequential_5_running_mean = const()[name = tensor("encoder_conv_module_2_sequential_5_running_mean"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21994048)))]; - tensor encoder_conv_module_2_sequential_5_bias = const()[name = tensor("encoder_conv_module_2_sequential_5_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21995136)))]; - tensor encoder_conv_module_2_sequential_5_weight = const()[name = tensor("encoder_conv_module_2_sequential_5_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21996224)))]; - tensor encoder_conv_module_2_sequential_7_conv_bias = const()[name = tensor("encoder_conv_module_2_sequential_7_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21997312)))]; - tensor encoder_conv_module_2_sequential_7_conv_weight = const()[name = tensor("encoder_conv_module_2_sequential_7_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21998400)))]; - tensor encoder_ffn2_2_module_sequential_0_bias = const()[name = tensor("encoder_ffn2_2_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(22260608)))]; - tensor encoder_ffn2_2_module_sequential_0_weight = const()[name = tensor("encoder_ffn2_2_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(22261696)))]; - tensor encoder_ffn2_2_module_sequential_1_linear_bias = const()[name = tensor("encoder_ffn2_2_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(22262784)))]; - tensor encoder_ffn2_2_module_sequential_1_linear_weight = const()[name = tensor("encoder_ffn2_2_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(22266944)))]; - tensor encoder_ffn2_2_module_sequential_4_linear_bias = const()[name = tensor("encoder_ffn2_2_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(23315584)))]; - tensor encoder_ffn2_2_module_sequential_4_linear_weight = const()[name = tensor("encoder_ffn2_2_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(23316672)))]; - tensor encoder_layer_norm_2_bias = const()[name = tensor("encoder_layer_norm_2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(24365312)))]; - tensor encoder_layer_norm_2_weight = const()[name = tensor("encoder_layer_norm_2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(24366400)))]; - tensor encoder_ffn1_3_module_sequential_0_bias = const()[name = tensor("encoder_ffn1_3_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(24367488)))]; - tensor encoder_ffn1_3_module_sequential_0_weight = const()[name = tensor("encoder_ffn1_3_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(24368576)))]; - tensor encoder_ffn1_3_module_sequential_1_linear_bias = const()[name = tensor("encoder_ffn1_3_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(24369664)))]; - tensor encoder_ffn1_3_module_sequential_1_linear_weight = const()[name = tensor("encoder_ffn1_3_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(24373824)))]; - tensor encoder_ffn1_3_module_sequential_4_linear_bias = const()[name = tensor("encoder_ffn1_3_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(25422464)))]; - tensor encoder_ffn1_3_module_sequential_4_linear_weight = const()[name = tensor("encoder_ffn1_3_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(25423552)))]; - tensor encoder_ret_lns_3_bias = const()[name = tensor("encoder_ret_lns_3_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(26472192)))]; - tensor encoder_ret_lns_3_weight = const()[name = tensor("encoder_ret_lns_3_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(26473280)))]; - tensor encoder_q_proj_3_bias = const()[name = tensor("encoder_q_proj_3_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(26474368)))]; - tensor encoder_q_proj_3_weight = const()[name = tensor("encoder_q_proj_3_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(26475456)))]; - tensor encoder_k_proj_3_bias = const()[name = tensor("encoder_k_proj_3_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(26737664)))]; - tensor encoder_k_proj_3_weight = const()[name = tensor("encoder_k_proj_3_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(26738752)))]; - tensor encoder_v_proj_3_bias = const()[name = tensor("encoder_v_proj_3_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27000960)))]; - tensor encoder_v_proj_3_weight = const()[name = tensor("encoder_v_proj_3_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27002048)))]; - tensor encoder_g_proj_3_bias = const()[name = tensor("encoder_g_proj_3_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27264256)))]; - tensor encoder_g_proj_3_weight = const()[name = tensor("encoder_g_proj_3_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27265344)))]; - tensor encoder_out_proj_3_bias = const()[name = tensor("encoder_out_proj_3_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27527552)))]; - tensor encoder_out_proj_3_weight = const()[name = tensor("encoder_out_proj_3_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27528640)))]; - tensor encoder_conv_module_3_sequential_0_bias = const()[name = tensor("encoder_conv_module_3_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27790848)))]; - tensor encoder_conv_module_3_sequential_0_weight = const()[name = tensor("encoder_conv_module_3_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27791936)))]; - tensor encoder_conv_module_3_sequential_2_conv_bias = const()[name = tensor("encoder_conv_module_3_sequential_2_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27793024)))]; - tensor encoder_conv_module_3_sequential_2_conv_weight = const()[name = tensor("encoder_conv_module_3_sequential_2_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27795136)))]; - tensor encoder_conv_module_3_sequential_4_conv_weight = const()[name = tensor("encoder_conv_module_3_sequential_4_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28319488)))]; - tensor encoder_conv_module_3_sequential_5_running_var = const()[name = tensor("encoder_conv_module_3_sequential_5_running_var"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28335936)))]; - tensor encoder_conv_module_3_sequential_5_running_mean = const()[name = tensor("encoder_conv_module_3_sequential_5_running_mean"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28337024)))]; - tensor encoder_conv_module_3_sequential_5_bias = const()[name = tensor("encoder_conv_module_3_sequential_5_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28338112)))]; - tensor encoder_conv_module_3_sequential_5_weight = const()[name = tensor("encoder_conv_module_3_sequential_5_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28339200)))]; - tensor encoder_conv_module_3_sequential_7_conv_bias = const()[name = tensor("encoder_conv_module_3_sequential_7_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28340288)))]; - tensor encoder_conv_module_3_sequential_7_conv_weight = const()[name = tensor("encoder_conv_module_3_sequential_7_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28341376)))]; - tensor encoder_ffn2_3_module_sequential_0_bias = const()[name = tensor("encoder_ffn2_3_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28603584)))]; - tensor encoder_ffn2_3_module_sequential_0_weight = const()[name = tensor("encoder_ffn2_3_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28604672)))]; - tensor encoder_ffn2_3_module_sequential_1_linear_bias = const()[name = tensor("encoder_ffn2_3_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28605760)))]; - tensor encoder_ffn2_3_module_sequential_1_linear_weight = const()[name = tensor("encoder_ffn2_3_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28609920)))]; - tensor encoder_ffn2_3_module_sequential_4_linear_bias = const()[name = tensor("encoder_ffn2_3_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(29658560)))]; - tensor encoder_ffn2_3_module_sequential_4_linear_weight = const()[name = tensor("encoder_ffn2_3_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(29659648)))]; - tensor encoder_layer_norm_3_bias = const()[name = tensor("encoder_layer_norm_3_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(30708288)))]; - tensor encoder_layer_norm_3_weight = const()[name = tensor("encoder_layer_norm_3_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(30709376)))]; - tensor decoder_convert_bias = const()[name = tensor("decoder_convert_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(30710464)))]; - tensor decoder_convert_weight = const()[name = tensor("decoder_convert_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(30711552)))]; - tensor decoder_q_proj_0_bias = const()[name = tensor("decoder_q_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(31235904)))]; - tensor decoder_q_proj_0_weight = const()[name = tensor("decoder_q_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(31236992)))]; - tensor decoder_k_proj_0_bias = const()[name = tensor("decoder_k_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(31499200)))]; - tensor decoder_k_proj_0_weight = const()[name = tensor("decoder_k_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(31500288)))]; - tensor decoder_v_proj_0_bias = const()[name = tensor("decoder_v_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(31762496)))]; - tensor decoder_v_proj_0_weight = const()[name = tensor("decoder_v_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(31763584)))]; - tensor decoder_g_proj_0_bias = const()[name = tensor("decoder_g_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32025792)))]; - tensor decoder_g_proj_0_weight = const()[name = tensor("decoder_g_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32026880)))]; - tensor decoder_out_proj_0_bias = const()[name = tensor("decoder_out_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32289088)))]; - tensor decoder_out_proj_0_weight = const()[name = tensor("decoder_out_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32290176)))]; - tensor decoder_norm11_0_bias = const()[name = tensor("decoder_norm11_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32552384)))]; - tensor decoder_norm11_0_weight = const()[name = tensor("decoder_norm11_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32553472)))]; - tensor decoder_self_attn2_0_out_proj_bias = const()[name = tensor("decoder_self_attn2_0_out_proj_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32554560)))]; - tensor decoder_self_attn2_0_out_proj_weight = const()[name = tensor("decoder_self_attn2_0_out_proj_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32555648)))]; - tensor decoder_self_attn2_0_in_proj_bias = const()[name = tensor("decoder_self_attn2_0_in_proj_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32817856)))]; - tensor decoder_self_attn2_0_in_proj_weight = const()[name = tensor("decoder_self_attn2_0_in_proj_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32820992)))]; - tensor decoder_norm21_0_bias = const()[name = tensor("decoder_norm21_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(33607488)))]; - tensor decoder_norm21_0_weight = const()[name = tensor("decoder_norm21_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(33608576)))]; - tensor decoder_linear1_0_bias = const()[name = tensor("decoder_linear1_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(33609664)))]; - tensor decoder_linear1_0_weight = const()[name = tensor("decoder_linear1_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(33617920)))]; - tensor decoder_linear2_0_bias = const()[name = tensor("decoder_linear2_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(35715136)))]; - tensor decoder_linear2_0_weight = const()[name = tensor("decoder_linear2_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(35716224)))]; - tensor decoder_norm22_0_bias = const()[name = tensor("decoder_norm22_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(37813440)))]; - tensor decoder_norm22_0_weight = const()[name = tensor("decoder_norm22_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(37814528)))]; - tensor decoder_q_proj_1_bias = const()[name = tensor("decoder_q_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(37815616)))]; - tensor decoder_q_proj_1_weight = const()[name = tensor("decoder_q_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(37816704)))]; - tensor decoder_k_proj_1_bias = const()[name = tensor("decoder_k_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38078912)))]; - tensor decoder_k_proj_1_weight = const()[name = tensor("decoder_k_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38080000)))]; - tensor decoder_v_proj_1_bias = const()[name = tensor("decoder_v_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38342208)))]; - tensor decoder_v_proj_1_weight = const()[name = tensor("decoder_v_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38343296)))]; - tensor decoder_g_proj_1_bias = const()[name = tensor("decoder_g_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38605504)))]; - tensor decoder_g_proj_1_weight = const()[name = tensor("decoder_g_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38606592)))]; - tensor decoder_out_proj_1_bias = const()[name = tensor("decoder_out_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38868800)))]; - tensor decoder_out_proj_1_weight = const()[name = tensor("decoder_out_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38869888)))]; - tensor decoder_norm11_1_bias = const()[name = tensor("decoder_norm11_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(39132096)))]; - tensor decoder_norm11_1_weight = const()[name = tensor("decoder_norm11_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(39133184)))]; - tensor decoder_self_attn2_1_out_proj_bias = const()[name = tensor("decoder_self_attn2_1_out_proj_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(39134272)))]; - tensor decoder_self_attn2_1_out_proj_weight = const()[name = tensor("decoder_self_attn2_1_out_proj_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(39135360)))]; - tensor decoder_self_attn2_1_in_proj_bias = const()[name = tensor("decoder_self_attn2_1_in_proj_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(39397568)))]; - tensor decoder_self_attn2_1_in_proj_weight = const()[name = tensor("decoder_self_attn2_1_in_proj_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(39400704)))]; - tensor decoder_norm21_1_bias = const()[name = tensor("decoder_norm21_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(40187200)))]; - tensor decoder_norm21_1_weight = const()[name = tensor("decoder_norm21_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(40188288)))]; - tensor decoder_linear1_1_bias = const()[name = tensor("decoder_linear1_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(40189376)))]; - tensor decoder_linear1_1_weight = const()[name = tensor("decoder_linear1_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(40197632)))]; - tensor decoder_linear2_1_bias = const()[name = tensor("decoder_linear2_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(42294848)))]; - tensor decoder_linear2_1_weight = const()[name = tensor("decoder_linear2_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(42295936)))]; - tensor decoder_norm22_1_bias = const()[name = tensor("decoder_norm22_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(44393152)))]; - tensor decoder_norm22_1_weight = const()[name = tensor("decoder_norm22_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(44394240)))]; - tensor var_11 = const()[name = tensor("op_11"), val = tensor(0x1.5798eep-27)]; - tensor var_17 = const()[name = tensor("op_17"), val = tensor(0x1.0c6f7ap-20)]; - tensor var_20 = const()[name = tensor("op_20"), val = tensor(2)]; - tensor var_23 = const()[name = tensor("op_23"), val = tensor(0)]; - tensor var_26 = const()[name = tensor("op_26"), val = tensor(1)]; - tensor var_28 = const()[name = tensor("op_28"), val = tensor(0x1.4f8b58p-17)]; - tensor var_29 = const()[name = tensor("op_29"), val = tensor(true)]; - tensor input_1 = linear(bias = encoder_input_projection_linear_bias, weight = encoder_input_projection_linear_weight, x = features)[name = tensor("linear_0")]; - tensor input_3_axes_0 = const()[name = tensor("input_3_axes_0"), val = tensor([-1])]; - tensor input_3 = layer_norm(axes = input_3_axes_0, beta = encoder_pre_layer_norm_bias, epsilon = var_28, gamma = encoder_pre_layer_norm_weight, x = input_1)[name = tensor("input_3")]; + func main(tensor cnn_window, tensor dec_kv, tensor dec_scale, tensor enc_conv_cache, tensor enc_kv, tensor enc_scale, tensor features, tensor valid_mask) { + tensor inner_encoder_cnn_bias = const()[name = tensor("inner_encoder_cnn_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(64)))]; + tensor inner_encoder_cnn_weight = const()[name = tensor("inner_encoder_cnn_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1152)))]; + tensor inner_encoder__causal_mask = const()[name = tensor("inner_encoder__causal_mask"), val = tensor([[0x1p+0]])]; + tensor inner_encoder__t_index = const()[name = tensor("inner_encoder__t_index"), val = tensor([0x1p+0])]; + tensor inner_encoder_input_projection_linear_bias = const()[name = tensor("inner_encoder_input_projection_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4981952)))]; + tensor inner_encoder_input_projection_linear_weight = const()[name = tensor("inner_encoder_input_projection_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4983040)))]; + tensor inner_encoder_pre_layer_norm_bias = const()[name = tensor("inner_encoder_pre_layer_norm_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(5336384)))]; + tensor inner_encoder_pre_layer_norm_weight = const()[name = tensor("inner_encoder_pre_layer_norm_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(5337472)))]; + tensor inner_encoder_ffn1_0_module_sequential_0_bias = const()[name = tensor("inner_encoder_ffn1_0_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(5338560)))]; + tensor inner_encoder_ffn1_0_module_sequential_0_weight = const()[name = tensor("inner_encoder_ffn1_0_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(5339648)))]; + tensor inner_encoder_ffn1_0_module_sequential_1_linear_bias = const()[name = tensor("inner_encoder_ffn1_0_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(5340736)))]; + tensor inner_encoder_ffn1_0_module_sequential_1_linear_weight = const()[name = tensor("inner_encoder_ffn1_0_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(5344896)))]; + tensor inner_encoder_ffn1_0_module_sequential_4_linear_bias = const()[name = tensor("inner_encoder_ffn1_0_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(6393536)))]; + tensor inner_encoder_ffn1_0_module_sequential_4_linear_weight = const()[name = tensor("inner_encoder_ffn1_0_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(6394624)))]; + tensor inner_encoder_ret_lns_0_bias = const()[name = tensor("inner_encoder_ret_lns_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7443264)))]; + tensor inner_encoder_ret_lns_0_weight = const()[name = tensor("inner_encoder_ret_lns_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7444352)))]; + tensor inner_encoder_q_proj_0_bias = const()[name = tensor("inner_encoder_q_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7445440)))]; + tensor inner_encoder_q_proj_0_weight = const()[name = tensor("inner_encoder_q_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7446528)))]; + tensor inner_encoder_k_proj_0_bias = const()[name = tensor("inner_encoder_k_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7708736)))]; + tensor inner_encoder_k_proj_0_weight = const()[name = tensor("inner_encoder_k_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7709824)))]; + tensor inner_encoder_v_proj_0_bias = const()[name = tensor("inner_encoder_v_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7972032)))]; + tensor inner_encoder_v_proj_0_weight = const()[name = tensor("inner_encoder_v_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7973120)))]; + tensor inner_encoder_g_proj_0_bias = const()[name = tensor("inner_encoder_g_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8235328)))]; + tensor inner_encoder_g_proj_0_weight = const()[name = tensor("inner_encoder_g_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8236416)))]; + tensor inner_encoder_out_proj_0_bias = const()[name = tensor("inner_encoder_out_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8498624)))]; + tensor inner_encoder_out_proj_0_weight = const()[name = tensor("inner_encoder_out_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8499712)))]; + tensor inner_encoder_conv_module_0_sequential_0_bias = const()[name = tensor("inner_encoder_conv_module_0_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8761920)))]; + tensor inner_encoder_conv_module_0_sequential_0_weight = const()[name = tensor("inner_encoder_conv_module_0_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8763008)))]; + tensor inner_encoder_conv_module_0_sequential_2_conv_bias = const()[name = tensor("inner_encoder_conv_module_0_sequential_2_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8764096)))]; + tensor inner_encoder_conv_module_0_sequential_2_conv_weight = const()[name = tensor("inner_encoder_conv_module_0_sequential_2_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8766208)))]; + tensor inner_encoder_conv_module_0_sequential_4_conv_weight = const()[name = tensor("inner_encoder_conv_module_0_sequential_4_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9290560)))]; + tensor inner_encoder_conv_module_0_sequential_5_running_var = const()[name = tensor("inner_encoder_conv_module_0_sequential_5_running_var"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9307008)))]; + tensor inner_encoder_conv_module_0_sequential_5_running_mean = const()[name = tensor("inner_encoder_conv_module_0_sequential_5_running_mean"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9308096)))]; + tensor inner_encoder_conv_module_0_sequential_5_bias = const()[name = tensor("inner_encoder_conv_module_0_sequential_5_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9309184)))]; + tensor inner_encoder_conv_module_0_sequential_5_weight = const()[name = tensor("inner_encoder_conv_module_0_sequential_5_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9310272)))]; + tensor inner_encoder_conv_module_0_sequential_7_conv_bias = const()[name = tensor("inner_encoder_conv_module_0_sequential_7_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9311360)))]; + tensor inner_encoder_conv_module_0_sequential_7_conv_weight = const()[name = tensor("inner_encoder_conv_module_0_sequential_7_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9312448)))]; + tensor inner_encoder_ffn2_0_module_sequential_0_bias = const()[name = tensor("inner_encoder_ffn2_0_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9574656)))]; + tensor inner_encoder_ffn2_0_module_sequential_0_weight = const()[name = tensor("inner_encoder_ffn2_0_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9575744)))]; + tensor inner_encoder_ffn2_0_module_sequential_1_linear_bias = const()[name = tensor("inner_encoder_ffn2_0_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9576832)))]; + tensor inner_encoder_ffn2_0_module_sequential_1_linear_weight = const()[name = tensor("inner_encoder_ffn2_0_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9580992)))]; + tensor inner_encoder_ffn2_0_module_sequential_4_linear_bias = const()[name = tensor("inner_encoder_ffn2_0_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(10629632)))]; + tensor inner_encoder_ffn2_0_module_sequential_4_linear_weight = const()[name = tensor("inner_encoder_ffn2_0_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(10630720)))]; + tensor inner_encoder_layer_norm_0_bias = const()[name = tensor("inner_encoder_layer_norm_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11679360)))]; + tensor inner_encoder_layer_norm_0_weight = const()[name = tensor("inner_encoder_layer_norm_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11680448)))]; + tensor inner_encoder_ffn1_1_module_sequential_0_bias = const()[name = tensor("inner_encoder_ffn1_1_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11681536)))]; + tensor inner_encoder_ffn1_1_module_sequential_0_weight = const()[name = tensor("inner_encoder_ffn1_1_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11682624)))]; + tensor inner_encoder_ffn1_1_module_sequential_1_linear_bias = const()[name = tensor("inner_encoder_ffn1_1_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11683712)))]; + tensor inner_encoder_ffn1_1_module_sequential_1_linear_weight = const()[name = tensor("inner_encoder_ffn1_1_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11687872)))]; + tensor inner_encoder_ffn1_1_module_sequential_4_linear_bias = const()[name = tensor("inner_encoder_ffn1_1_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(12736512)))]; + tensor inner_encoder_ffn1_1_module_sequential_4_linear_weight = const()[name = tensor("inner_encoder_ffn1_1_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(12737600)))]; + tensor inner_encoder_ret_lns_1_bias = const()[name = tensor("inner_encoder_ret_lns_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(13786240)))]; + tensor inner_encoder_ret_lns_1_weight = const()[name = tensor("inner_encoder_ret_lns_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(13787328)))]; + tensor inner_encoder_q_proj_1_bias = const()[name = tensor("inner_encoder_q_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(13788416)))]; + tensor inner_encoder_q_proj_1_weight = const()[name = tensor("inner_encoder_q_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(13789504)))]; + tensor inner_encoder_k_proj_1_bias = const()[name = tensor("inner_encoder_k_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14051712)))]; + tensor inner_encoder_k_proj_1_weight = const()[name = tensor("inner_encoder_k_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14052800)))]; + tensor inner_encoder_v_proj_1_bias = const()[name = tensor("inner_encoder_v_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14315008)))]; + tensor inner_encoder_v_proj_1_weight = const()[name = tensor("inner_encoder_v_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14316096)))]; + tensor inner_encoder_g_proj_1_bias = const()[name = tensor("inner_encoder_g_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14578304)))]; + tensor inner_encoder_g_proj_1_weight = const()[name = tensor("inner_encoder_g_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14579392)))]; + tensor inner_encoder_out_proj_1_bias = const()[name = tensor("inner_encoder_out_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14841600)))]; + tensor inner_encoder_out_proj_1_weight = const()[name = tensor("inner_encoder_out_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14842688)))]; + tensor inner_encoder_conv_module_1_sequential_0_bias = const()[name = tensor("inner_encoder_conv_module_1_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15104896)))]; + tensor inner_encoder_conv_module_1_sequential_0_weight = const()[name = tensor("inner_encoder_conv_module_1_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15105984)))]; + tensor inner_encoder_conv_module_1_sequential_2_conv_bias = const()[name = tensor("inner_encoder_conv_module_1_sequential_2_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15107072)))]; + tensor inner_encoder_conv_module_1_sequential_2_conv_weight = const()[name = tensor("inner_encoder_conv_module_1_sequential_2_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15109184)))]; + tensor inner_encoder_conv_module_1_sequential_4_conv_weight = const()[name = tensor("inner_encoder_conv_module_1_sequential_4_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15633536)))]; + tensor inner_encoder_conv_module_1_sequential_5_running_var = const()[name = tensor("inner_encoder_conv_module_1_sequential_5_running_var"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15649984)))]; + tensor inner_encoder_conv_module_1_sequential_5_running_mean = const()[name = tensor("inner_encoder_conv_module_1_sequential_5_running_mean"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15651072)))]; + tensor inner_encoder_conv_module_1_sequential_5_bias = const()[name = tensor("inner_encoder_conv_module_1_sequential_5_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15652160)))]; + tensor inner_encoder_conv_module_1_sequential_5_weight = const()[name = tensor("inner_encoder_conv_module_1_sequential_5_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15653248)))]; + tensor inner_encoder_conv_module_1_sequential_7_conv_bias = const()[name = tensor("inner_encoder_conv_module_1_sequential_7_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15654336)))]; + tensor inner_encoder_conv_module_1_sequential_7_conv_weight = const()[name = tensor("inner_encoder_conv_module_1_sequential_7_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15655424)))]; + tensor inner_encoder_ffn2_1_module_sequential_0_bias = const()[name = tensor("inner_encoder_ffn2_1_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15917632)))]; + tensor inner_encoder_ffn2_1_module_sequential_0_weight = const()[name = tensor("inner_encoder_ffn2_1_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15918720)))]; + tensor inner_encoder_ffn2_1_module_sequential_1_linear_bias = const()[name = tensor("inner_encoder_ffn2_1_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15919808)))]; + tensor inner_encoder_ffn2_1_module_sequential_1_linear_weight = const()[name = tensor("inner_encoder_ffn2_1_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15923968)))]; + tensor inner_encoder_ffn2_1_module_sequential_4_linear_bias = const()[name = tensor("inner_encoder_ffn2_1_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(16972608)))]; + tensor inner_encoder_ffn2_1_module_sequential_4_linear_weight = const()[name = tensor("inner_encoder_ffn2_1_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(16973696)))]; + tensor inner_encoder_layer_norm_1_bias = const()[name = tensor("inner_encoder_layer_norm_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18022336)))]; + tensor inner_encoder_layer_norm_1_weight = const()[name = tensor("inner_encoder_layer_norm_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18023424)))]; + tensor inner_encoder_ffn1_2_module_sequential_0_bias = const()[name = tensor("inner_encoder_ffn1_2_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18024512)))]; + tensor inner_encoder_ffn1_2_module_sequential_0_weight = const()[name = tensor("inner_encoder_ffn1_2_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18025600)))]; + tensor inner_encoder_ffn1_2_module_sequential_1_linear_bias = const()[name = tensor("inner_encoder_ffn1_2_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18026688)))]; + tensor inner_encoder_ffn1_2_module_sequential_1_linear_weight = const()[name = tensor("inner_encoder_ffn1_2_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18030848)))]; + tensor inner_encoder_ffn1_2_module_sequential_4_linear_bias = const()[name = tensor("inner_encoder_ffn1_2_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(19079488)))]; + tensor inner_encoder_ffn1_2_module_sequential_4_linear_weight = const()[name = tensor("inner_encoder_ffn1_2_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(19080576)))]; + tensor inner_encoder_ret_lns_2_bias = const()[name = tensor("inner_encoder_ret_lns_2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20129216)))]; + tensor inner_encoder_ret_lns_2_weight = const()[name = tensor("inner_encoder_ret_lns_2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20130304)))]; + tensor inner_encoder_q_proj_2_bias = const()[name = tensor("inner_encoder_q_proj_2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20131392)))]; + tensor inner_encoder_q_proj_2_weight = const()[name = tensor("inner_encoder_q_proj_2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20132480)))]; + tensor inner_encoder_k_proj_2_bias = const()[name = tensor("inner_encoder_k_proj_2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20394688)))]; + tensor inner_encoder_k_proj_2_weight = const()[name = tensor("inner_encoder_k_proj_2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20395776)))]; + tensor inner_encoder_v_proj_2_bias = const()[name = tensor("inner_encoder_v_proj_2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20657984)))]; + tensor inner_encoder_v_proj_2_weight = const()[name = tensor("inner_encoder_v_proj_2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20659072)))]; + tensor inner_encoder_g_proj_2_bias = const()[name = tensor("inner_encoder_g_proj_2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20921280)))]; + tensor inner_encoder_g_proj_2_weight = const()[name = tensor("inner_encoder_g_proj_2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20922368)))]; + tensor inner_encoder_out_proj_2_bias = const()[name = tensor("inner_encoder_out_proj_2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21184576)))]; + tensor inner_encoder_out_proj_2_weight = const()[name = tensor("inner_encoder_out_proj_2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21185664)))]; + tensor inner_encoder_conv_module_2_sequential_0_bias = const()[name = tensor("inner_encoder_conv_module_2_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21447872)))]; + tensor inner_encoder_conv_module_2_sequential_0_weight = const()[name = tensor("inner_encoder_conv_module_2_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21448960)))]; + tensor inner_encoder_conv_module_2_sequential_2_conv_bias = const()[name = tensor("inner_encoder_conv_module_2_sequential_2_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21450048)))]; + tensor inner_encoder_conv_module_2_sequential_2_conv_weight = const()[name = tensor("inner_encoder_conv_module_2_sequential_2_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21452160)))]; + tensor inner_encoder_conv_module_2_sequential_4_conv_weight = const()[name = tensor("inner_encoder_conv_module_2_sequential_4_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21976512)))]; + tensor inner_encoder_conv_module_2_sequential_5_running_var = const()[name = tensor("inner_encoder_conv_module_2_sequential_5_running_var"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21992960)))]; + tensor inner_encoder_conv_module_2_sequential_5_running_mean = const()[name = tensor("inner_encoder_conv_module_2_sequential_5_running_mean"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21994048)))]; + tensor inner_encoder_conv_module_2_sequential_5_bias = const()[name = tensor("inner_encoder_conv_module_2_sequential_5_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21995136)))]; + tensor inner_encoder_conv_module_2_sequential_5_weight = const()[name = tensor("inner_encoder_conv_module_2_sequential_5_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21996224)))]; + tensor inner_encoder_conv_module_2_sequential_7_conv_bias = const()[name = tensor("inner_encoder_conv_module_2_sequential_7_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21997312)))]; + tensor inner_encoder_conv_module_2_sequential_7_conv_weight = const()[name = tensor("inner_encoder_conv_module_2_sequential_7_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21998400)))]; + tensor inner_encoder_ffn2_2_module_sequential_0_bias = const()[name = tensor("inner_encoder_ffn2_2_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(22260608)))]; + tensor inner_encoder_ffn2_2_module_sequential_0_weight = const()[name = tensor("inner_encoder_ffn2_2_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(22261696)))]; + tensor inner_encoder_ffn2_2_module_sequential_1_linear_bias = const()[name = tensor("inner_encoder_ffn2_2_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(22262784)))]; + tensor inner_encoder_ffn2_2_module_sequential_1_linear_weight = const()[name = tensor("inner_encoder_ffn2_2_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(22266944)))]; + tensor inner_encoder_ffn2_2_module_sequential_4_linear_bias = const()[name = tensor("inner_encoder_ffn2_2_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(23315584)))]; + tensor inner_encoder_ffn2_2_module_sequential_4_linear_weight = const()[name = tensor("inner_encoder_ffn2_2_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(23316672)))]; + tensor inner_encoder_layer_norm_2_bias = const()[name = tensor("inner_encoder_layer_norm_2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(24365312)))]; + tensor inner_encoder_layer_norm_2_weight = const()[name = tensor("inner_encoder_layer_norm_2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(24366400)))]; + tensor inner_encoder_ffn1_3_module_sequential_0_bias = const()[name = tensor("inner_encoder_ffn1_3_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(24367488)))]; + tensor inner_encoder_ffn1_3_module_sequential_0_weight = const()[name = tensor("inner_encoder_ffn1_3_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(24368576)))]; + tensor inner_encoder_ffn1_3_module_sequential_1_linear_bias = const()[name = tensor("inner_encoder_ffn1_3_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(24369664)))]; + tensor inner_encoder_ffn1_3_module_sequential_1_linear_weight = const()[name = tensor("inner_encoder_ffn1_3_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(24373824)))]; + tensor inner_encoder_ffn1_3_module_sequential_4_linear_bias = const()[name = tensor("inner_encoder_ffn1_3_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(25422464)))]; + tensor inner_encoder_ffn1_3_module_sequential_4_linear_weight = const()[name = tensor("inner_encoder_ffn1_3_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(25423552)))]; + tensor inner_encoder_ret_lns_3_bias = const()[name = tensor("inner_encoder_ret_lns_3_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(26472192)))]; + tensor inner_encoder_ret_lns_3_weight = const()[name = tensor("inner_encoder_ret_lns_3_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(26473280)))]; + tensor inner_encoder_q_proj_3_bias = const()[name = tensor("inner_encoder_q_proj_3_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(26474368)))]; + tensor inner_encoder_q_proj_3_weight = const()[name = tensor("inner_encoder_q_proj_3_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(26475456)))]; + tensor inner_encoder_k_proj_3_bias = const()[name = tensor("inner_encoder_k_proj_3_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(26737664)))]; + tensor inner_encoder_k_proj_3_weight = const()[name = tensor("inner_encoder_k_proj_3_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(26738752)))]; + tensor inner_encoder_v_proj_3_bias = const()[name = tensor("inner_encoder_v_proj_3_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27000960)))]; + tensor inner_encoder_v_proj_3_weight = const()[name = tensor("inner_encoder_v_proj_3_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27002048)))]; + tensor inner_encoder_g_proj_3_bias = const()[name = tensor("inner_encoder_g_proj_3_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27264256)))]; + tensor inner_encoder_g_proj_3_weight = const()[name = tensor("inner_encoder_g_proj_3_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27265344)))]; + tensor inner_encoder_out_proj_3_bias = const()[name = tensor("inner_encoder_out_proj_3_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27527552)))]; + tensor inner_encoder_out_proj_3_weight = const()[name = tensor("inner_encoder_out_proj_3_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27528640)))]; + tensor inner_encoder_conv_module_3_sequential_0_bias = const()[name = tensor("inner_encoder_conv_module_3_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27790848)))]; + tensor inner_encoder_conv_module_3_sequential_0_weight = const()[name = tensor("inner_encoder_conv_module_3_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27791936)))]; + tensor inner_encoder_conv_module_3_sequential_2_conv_bias = const()[name = tensor("inner_encoder_conv_module_3_sequential_2_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27793024)))]; + tensor inner_encoder_conv_module_3_sequential_2_conv_weight = const()[name = tensor("inner_encoder_conv_module_3_sequential_2_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27795136)))]; + tensor inner_encoder_conv_module_3_sequential_4_conv_weight = const()[name = tensor("inner_encoder_conv_module_3_sequential_4_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28319488)))]; + tensor inner_encoder_conv_module_3_sequential_5_running_var = const()[name = tensor("inner_encoder_conv_module_3_sequential_5_running_var"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28335936)))]; + tensor inner_encoder_conv_module_3_sequential_5_running_mean = const()[name = tensor("inner_encoder_conv_module_3_sequential_5_running_mean"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28337024)))]; + tensor inner_encoder_conv_module_3_sequential_5_bias = const()[name = tensor("inner_encoder_conv_module_3_sequential_5_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28338112)))]; + tensor inner_encoder_conv_module_3_sequential_5_weight = const()[name = tensor("inner_encoder_conv_module_3_sequential_5_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28339200)))]; + tensor inner_encoder_conv_module_3_sequential_7_conv_bias = const()[name = tensor("inner_encoder_conv_module_3_sequential_7_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28340288)))]; + tensor inner_encoder_conv_module_3_sequential_7_conv_weight = const()[name = tensor("inner_encoder_conv_module_3_sequential_7_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28341376)))]; + tensor inner_encoder_ffn2_3_module_sequential_0_bias = const()[name = tensor("inner_encoder_ffn2_3_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28603584)))]; + tensor inner_encoder_ffn2_3_module_sequential_0_weight = const()[name = tensor("inner_encoder_ffn2_3_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28604672)))]; + tensor inner_encoder_ffn2_3_module_sequential_1_linear_bias = const()[name = tensor("inner_encoder_ffn2_3_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28605760)))]; + tensor inner_encoder_ffn2_3_module_sequential_1_linear_weight = const()[name = tensor("inner_encoder_ffn2_3_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28609920)))]; + tensor inner_encoder_ffn2_3_module_sequential_4_linear_bias = const()[name = tensor("inner_encoder_ffn2_3_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(29658560)))]; + tensor inner_encoder_ffn2_3_module_sequential_4_linear_weight = const()[name = tensor("inner_encoder_ffn2_3_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(29659648)))]; + tensor inner_encoder_layer_norm_3_bias = const()[name = tensor("inner_encoder_layer_norm_3_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(30708288)))]; + tensor inner_encoder_layer_norm_3_weight = const()[name = tensor("inner_encoder_layer_norm_3_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(30709376)))]; + tensor inner_decoder_convert_bias = const()[name = tensor("inner_decoder_convert_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(30710464)))]; + tensor inner_decoder_convert_weight = const()[name = tensor("inner_decoder_convert_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(30711552)))]; + tensor inner_decoder_q_proj_0_bias = const()[name = tensor("inner_decoder_q_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(31235904)))]; + tensor inner_decoder_q_proj_0_weight = const()[name = tensor("inner_decoder_q_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(31236992)))]; + tensor inner_decoder_k_proj_0_bias = const()[name = tensor("inner_decoder_k_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(31499200)))]; + tensor inner_decoder_k_proj_0_weight = const()[name = tensor("inner_decoder_k_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(31500288)))]; + tensor inner_decoder_v_proj_0_bias = const()[name = tensor("inner_decoder_v_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(31762496)))]; + tensor inner_decoder_v_proj_0_weight = const()[name = tensor("inner_decoder_v_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(31763584)))]; + tensor inner_decoder_g_proj_0_bias = const()[name = tensor("inner_decoder_g_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32025792)))]; + tensor inner_decoder_g_proj_0_weight = const()[name = tensor("inner_decoder_g_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32026880)))]; + tensor inner_decoder_out_proj_0_bias = const()[name = tensor("inner_decoder_out_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32289088)))]; + tensor inner_decoder_out_proj_0_weight = const()[name = tensor("inner_decoder_out_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32290176)))]; + tensor inner_decoder_norm11_0_bias = const()[name = tensor("inner_decoder_norm11_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32552384)))]; + tensor inner_decoder_norm11_0_weight = const()[name = tensor("inner_decoder_norm11_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32553472)))]; + tensor inner_decoder_self_attn2_0_out_proj_bias = const()[name = tensor("inner_decoder_self_attn2_0_out_proj_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32554560)))]; + tensor inner_decoder_self_attn2_0_out_proj_weight = const()[name = tensor("inner_decoder_self_attn2_0_out_proj_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32555648)))]; + tensor inner_decoder_self_attn2_0_in_proj_bias = const()[name = tensor("inner_decoder_self_attn2_0_in_proj_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32817856)))]; + tensor inner_decoder_self_attn2_0_in_proj_weight = const()[name = tensor("inner_decoder_self_attn2_0_in_proj_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32820992)))]; + tensor inner_decoder_norm21_0_bias = const()[name = tensor("inner_decoder_norm21_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(33607488)))]; + tensor inner_decoder_norm21_0_weight = const()[name = tensor("inner_decoder_norm21_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(33608576)))]; + tensor inner_decoder_linear1_0_bias = const()[name = tensor("inner_decoder_linear1_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(33609664)))]; + tensor inner_decoder_linear1_0_weight = const()[name = tensor("inner_decoder_linear1_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(33617920)))]; + tensor inner_decoder_linear2_0_bias = const()[name = tensor("inner_decoder_linear2_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(35715136)))]; + tensor inner_decoder_linear2_0_weight = const()[name = tensor("inner_decoder_linear2_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(35716224)))]; + tensor inner_decoder_norm22_0_bias = const()[name = tensor("inner_decoder_norm22_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(37813440)))]; + tensor inner_decoder_norm22_0_weight = const()[name = tensor("inner_decoder_norm22_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(37814528)))]; + tensor inner_decoder_q_proj_1_bias = const()[name = tensor("inner_decoder_q_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(37815616)))]; + tensor inner_decoder_q_proj_1_weight = const()[name = tensor("inner_decoder_q_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(37816704)))]; + tensor inner_decoder_k_proj_1_bias = const()[name = tensor("inner_decoder_k_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38078912)))]; + tensor inner_decoder_k_proj_1_weight = const()[name = tensor("inner_decoder_k_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38080000)))]; + tensor inner_decoder_v_proj_1_bias = const()[name = tensor("inner_decoder_v_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38342208)))]; + tensor inner_decoder_v_proj_1_weight = const()[name = tensor("inner_decoder_v_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38343296)))]; + tensor inner_decoder_g_proj_1_bias = const()[name = tensor("inner_decoder_g_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38605504)))]; + tensor inner_decoder_g_proj_1_weight = const()[name = tensor("inner_decoder_g_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38606592)))]; + tensor inner_decoder_out_proj_1_bias = const()[name = tensor("inner_decoder_out_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38868800)))]; + tensor inner_decoder_out_proj_1_weight = const()[name = tensor("inner_decoder_out_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38869888)))]; + tensor inner_decoder_norm11_1_bias = const()[name = tensor("inner_decoder_norm11_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(39132096)))]; + tensor inner_decoder_norm11_1_weight = const()[name = tensor("inner_decoder_norm11_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(39133184)))]; + tensor inner_decoder_self_attn2_1_out_proj_bias = const()[name = tensor("inner_decoder_self_attn2_1_out_proj_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(39134272)))]; + tensor inner_decoder_self_attn2_1_out_proj_weight = const()[name = tensor("inner_decoder_self_attn2_1_out_proj_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(39135360)))]; + tensor inner_decoder_self_attn2_1_in_proj_bias = const()[name = tensor("inner_decoder_self_attn2_1_in_proj_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(39397568)))]; + tensor inner_decoder_self_attn2_1_in_proj_weight = const()[name = tensor("inner_decoder_self_attn2_1_in_proj_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(39400704)))]; + tensor inner_decoder_norm21_1_bias = const()[name = tensor("inner_decoder_norm21_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(40187200)))]; + tensor inner_decoder_norm21_1_weight = const()[name = tensor("inner_decoder_norm21_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(40188288)))]; + tensor inner_decoder_linear1_1_bias = const()[name = tensor("inner_decoder_linear1_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(40189376)))]; + tensor inner_decoder_linear1_1_weight = const()[name = tensor("inner_decoder_linear1_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(40197632)))]; + tensor inner_decoder_linear2_1_bias = const()[name = tensor("inner_decoder_linear2_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(42294848)))]; + tensor inner_decoder_linear2_1_weight = const()[name = tensor("inner_decoder_linear2_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(42295936)))]; + tensor inner_decoder_norm22_1_bias = const()[name = tensor("inner_decoder_norm22_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(44393152)))]; + tensor inner_decoder_norm22_1_weight = const()[name = tensor("inner_decoder_norm22_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(44394240)))]; + tensor stacked_axes_0 = const()[name = tensor("stacked_axes_0"), val = tensor([1])]; + tensor stacked = expand_dims(axes = stacked_axes_0, x = features)[name = tensor("stacked")]; + tensor var_26 = const()[name = tensor("op_26"), val = tensor([1, 1, 345])]; + tensor input_1 = reshape(shape = var_26, x = stacked)[name = tensor("input_1")]; + tensor var_28 = const()[name = tensor("op_28"), val = tensor(0x1p+0)]; + tensor var_34 = const()[name = tensor("op_34"), val = tensor(true)]; + tensor var_35 = const()[name = tensor("op_35"), val = tensor(0x1.4f8b58p-17)]; + tensor var_38 = const()[name = tensor("op_38"), val = tensor(0)]; + tensor var_40 = const()[name = tensor("op_40"), val = tensor(2)]; + tensor var_41 = const()[name = tensor("op_41"), val = tensor(-1)]; + tensor var_43 = const()[name = tensor("op_43"), val = tensor(0x1.0c6f7ap-20)]; + tensor var_48 = const()[name = tensor("op_48"), val = tensor(0x1.5798eep-27)]; + tensor var_52 = const()[name = tensor("op_52"), val = tensor(1)]; + tensor input_3 = linear(bias = inner_encoder_input_projection_linear_bias, weight = inner_encoder_input_projection_linear_weight, x = input_1)[name = tensor("linear_0")]; tensor input_5_axes_0 = const()[name = tensor("input_5_axes_0"), val = tensor([-1])]; - tensor input_5 = layer_norm(axes = input_5_axes_0, beta = encoder_ffn1_0_module_sequential_0_bias, epsilon = var_28, gamma = encoder_ffn1_0_module_sequential_0_weight, x = input_3)[name = tensor("input_5")]; - tensor inputs_1 = linear(bias = encoder_ffn1_0_module_sequential_1_linear_bias, weight = encoder_ffn1_0_module_sequential_1_linear_weight, x = input_5)[name = tensor("linear_1")]; - tensor input_7 = silu(x = inputs_1)[name = tensor("input_7")]; - tensor input_11 = linear(bias = encoder_ffn1_0_module_sequential_4_linear_bias, weight = encoder_ffn1_0_module_sequential_4_linear_weight, x = input_7)[name = tensor("linear_2")]; - tensor var_147 = const()[name = tensor("op_147"), val = tensor(0x1p-1)]; - tensor var_148 = mul(x = input_11, y = var_147)[name = tensor("op_148")]; - tensor input_13 = add(x = var_148, y = input_3)[name = tensor("input_13")]; + tensor input_5 = layer_norm(axes = input_5_axes_0, beta = inner_encoder_pre_layer_norm_bias, epsilon = var_35, gamma = inner_encoder_pre_layer_norm_weight, x = input_3)[name = tensor("input_5")]; + tensor input_7_axes_0 = const()[name = tensor("input_7_axes_0"), val = tensor([-1])]; + tensor input_7 = layer_norm(axes = input_7_axes_0, beta = inner_encoder_ffn1_0_module_sequential_0_bias, epsilon = var_35, gamma = inner_encoder_ffn1_0_module_sequential_0_weight, x = input_5)[name = tensor("input_7")]; + tensor inputs_1 = linear(bias = inner_encoder_ffn1_0_module_sequential_1_linear_bias, weight = inner_encoder_ffn1_0_module_sequential_1_linear_weight, x = input_7)[name = tensor("linear_1")]; + tensor input_9 = silu(x = inputs_1)[name = tensor("input_9")]; + tensor input_13 = linear(bias = inner_encoder_ffn1_0_module_sequential_4_linear_bias, weight = inner_encoder_ffn1_0_module_sequential_4_linear_weight, x = input_9)[name = tensor("linear_2")]; + tensor var_173 = const()[name = tensor("op_173"), val = tensor(0x1p-1)]; + tensor var_174 = mul(x = input_13, y = var_173)[name = tensor("op_174")]; + tensor input_15 = add(x = var_174, y = input_5)[name = tensor("input_15")]; tensor x_1_axes_0 = const()[name = tensor("x_1_axes_0"), val = tensor([-1])]; - tensor x_1 = layer_norm(axes = x_1_axes_0, beta = encoder_ret_lns_0_bias, epsilon = var_28, gamma = encoder_ret_lns_0_weight, x = input_13)[name = tensor("x_1")]; + tensor x_1 = layer_norm(axes = x_1_axes_0, beta = inner_encoder_ret_lns_0_bias, epsilon = var_35, gamma = inner_encoder_ret_lns_0_weight, x = input_15)[name = tensor("x_1")]; tensor prev_kv_1_begin_0 = const()[name = tensor("prev_kv_1_begin_0"), val = tensor([0, 0, 0, 0, 0])]; tensor prev_kv_1_end_0 = const()[name = tensor("prev_kv_1_end_0"), val = tensor([1, 1, 4, 64, 64])]; tensor prev_kv_1_end_mask_0 = const()[name = tensor("prev_kv_1_end_mask_0"), val = tensor([false, true, true, true, true])]; @@ -238,139 +244,139 @@ program(1.0) tensor prev_scale_1_end_mask_0 = const()[name = tensor("prev_scale_1_end_mask_0"), val = tensor([false, true])]; tensor prev_scale_1_squeeze_mask_0 = const()[name = tensor("prev_scale_1_squeeze_mask_0"), val = tensor([true, false])]; tensor prev_scale_1 = slice_by_index(begin = prev_scale_1_begin_0, end = prev_scale_1_end_0, end_mask = prev_scale_1_end_mask_0, squeeze_mask = prev_scale_1_squeeze_mask_0, x = enc_scale)[name = tensor("prev_scale_1")]; - tensor var_162 = linear(bias = encoder_q_proj_0_bias, weight = encoder_q_proj_0_weight, x = x_1)[name = tensor("linear_3")]; - tensor var_163 = const()[name = tensor("op_163"), val = tensor([1, 1, 4, 64])]; - tensor var_164 = reshape(shape = var_163, x = var_162)[name = tensor("op_164")]; + tensor var_188 = linear(bias = inner_encoder_q_proj_0_bias, weight = inner_encoder_q_proj_0_weight, x = x_1)[name = tensor("linear_3")]; + tensor var_189 = const()[name = tensor("op_189"), val = tensor([1, 1, 4, 64])]; + tensor var_190 = reshape(shape = var_189, x = var_188)[name = tensor("op_190")]; tensor q_1_perm_0 = const()[name = tensor("q_1_perm_0"), val = tensor([0, 2, 1, 3])]; - tensor var_168 = linear(bias = encoder_k_proj_0_bias, weight = encoder_k_proj_0_weight, x = x_1)[name = tensor("linear_4")]; - tensor var_169 = const()[name = tensor("op_169"), val = tensor(0x1p-3)]; - tensor var_170 = mul(x = var_168, y = var_169)[name = tensor("op_170")]; - tensor var_171 = const()[name = tensor("op_171"), val = tensor([1, 1, 4, 64])]; - tensor var_172 = reshape(shape = var_171, x = var_170)[name = tensor("op_172")]; + tensor var_194 = linear(bias = inner_encoder_k_proj_0_bias, weight = inner_encoder_k_proj_0_weight, x = x_1)[name = tensor("linear_4")]; + tensor var_195 = const()[name = tensor("op_195"), val = tensor(0x1p-3)]; + tensor var_196 = mul(x = var_194, y = var_195)[name = tensor("op_196")]; + tensor var_197 = const()[name = tensor("op_197"), val = tensor([1, 1, 4, 64])]; + tensor var_198 = reshape(shape = var_197, x = var_196)[name = tensor("op_198")]; tensor k_1_perm_0 = const()[name = tensor("k_1_perm_0"), val = tensor([0, 2, 1, 3])]; - tensor var_176 = linear(bias = encoder_v_proj_0_bias, weight = encoder_v_proj_0_weight, x = x_1)[name = tensor("linear_5")]; - tensor var_177 = const()[name = tensor("op_177"), val = tensor([1, 1, 4, 64])]; - tensor var_178 = reshape(shape = var_177, x = var_176)[name = tensor("op_178")]; + tensor var_202 = linear(bias = inner_encoder_v_proj_0_bias, weight = inner_encoder_v_proj_0_weight, x = x_1)[name = tensor("linear_5")]; + tensor var_203 = const()[name = tensor("op_203"), val = tensor([1, 1, 4, 64])]; + tensor var_204 = reshape(shape = var_203, x = var_202)[name = tensor("op_204")]; tensor v_1_perm_0 = const()[name = tensor("v_1_perm_0"), val = tensor([0, 2, 1, 3])]; - tensor input_17 = linear(bias = encoder_g_proj_0_bias, weight = encoder_g_proj_0_weight, x = x_1)[name = tensor("linear_6")]; + tensor input_19 = linear(bias = inner_encoder_g_proj_0_bias, weight = inner_encoder_g_proj_0_weight, x = x_1)[name = tensor("linear_6")]; tensor sqrt_s0_1 = sqrt(x = prev_scale_1)[name = tensor("sqrt_s0_1")]; - tensor s_t_1 = add(x = prev_scale_1, y = encoder__t_index)[name = tensor("s_t_1")]; + tensor s_t_1 = add(x = prev_scale_1, y = inner_encoder__t_index)[name = tensor("s_t_1")]; tensor sqrt_s_t_1 = sqrt(x = s_t_1)[name = tensor("sqrt_s_t_1")]; tensor qk_1_transpose_x_1 = const()[name = tensor("qk_1_transpose_x_1"), val = tensor(false)]; tensor qk_1_transpose_y_1 = const()[name = tensor("qk_1_transpose_y_1"), val = tensor(true)]; - tensor k_1 = transpose(perm = k_1_perm_0, x = var_172)[name = tensor("transpose_57")]; - tensor q_1 = transpose(perm = q_1_perm_0, x = var_164)[name = tensor("transpose_58")]; + tensor k_1 = transpose(perm = k_1_perm_0, x = var_198)[name = tensor("transpose_57")]; + tensor q_1 = transpose(perm = q_1_perm_0, x = var_190)[name = tensor("transpose_58")]; tensor qk_1 = matmul(transpose_x = qk_1_transpose_x_1, transpose_y = qk_1_transpose_y_1, x = q_1, y = k_1)[name = tensor("qk_1")]; - tensor var_188 = const()[name = tensor("op_188"), val = tensor([1, 1])]; - tensor var_189 = reshape(shape = var_188, x = sqrt_s_t_1)[name = tensor("op_189")]; - tensor M_1 = real_div(x = encoder__causal_mask, y = var_189)[name = tensor("M_1")]; - tensor var_191 = mul(x = qk_1, y = M_1)[name = tensor("op_191")]; + tensor var_214 = const()[name = tensor("op_214"), val = tensor([1, 1])]; + tensor var_215 = reshape(shape = var_214, x = sqrt_s_t_1)[name = tensor("op_215")]; + tensor M_1 = real_div(x = inner_encoder__causal_mask, y = var_215)[name = tensor("M_1")]; + tensor var_217 = mul(x = qk_1, y = M_1)[name = tensor("op_217")]; tensor inner_1_transpose_x_0 = const()[name = tensor("inner_1_transpose_x_0"), val = tensor(false)]; tensor inner_1_transpose_y_0 = const()[name = tensor("inner_1_transpose_y_0"), val = tensor(false)]; - tensor v_1 = transpose(perm = v_1_perm_0, x = var_178)[name = tensor("transpose_56")]; - tensor inner_1 = matmul(transpose_x = inner_1_transpose_x_0, transpose_y = inner_1_transpose_y_0, x = var_191, y = v_1)[name = tensor("inner_1")]; - tensor var_193_transpose_x_0 = const()[name = tensor("op_193_transpose_x_0"), val = tensor(false)]; - tensor var_193_transpose_y_0 = const()[name = tensor("op_193_transpose_y_0"), val = tensor(false)]; - tensor var_193 = matmul(transpose_x = var_193_transpose_x_0, transpose_y = var_193_transpose_y_0, x = q_1, y = prev_kv_1)[name = tensor("op_193")]; - tensor var_194 = real_div(x = sqrt_s0_1, y = sqrt_s_t_1)[name = tensor("op_194")]; - tensor var_195 = const()[name = tensor("op_195"), val = tensor([1, 1, 1, 1])]; - tensor var_196 = reshape(shape = var_195, x = var_194)[name = tensor("op_196")]; - tensor cross_1 = mul(x = var_193, y = var_196)[name = tensor("cross_1")]; + tensor v_1 = transpose(perm = v_1_perm_0, x = var_204)[name = tensor("transpose_56")]; + tensor inner_1 = matmul(transpose_x = inner_1_transpose_x_0, transpose_y = inner_1_transpose_y_0, x = var_217, y = v_1)[name = tensor("inner_1")]; + tensor var_219_transpose_x_0 = const()[name = tensor("op_219_transpose_x_0"), val = tensor(false)]; + tensor var_219_transpose_y_0 = const()[name = tensor("op_219_transpose_y_0"), val = tensor(false)]; + tensor var_219 = matmul(transpose_x = var_219_transpose_x_0, transpose_y = var_219_transpose_y_0, x = q_1, y = prev_kv_1)[name = tensor("op_219")]; + tensor var_220 = real_div(x = sqrt_s0_1, y = sqrt_s_t_1)[name = tensor("op_220")]; + tensor var_221 = const()[name = tensor("op_221"), val = tensor([1, 1, 1, 1])]; + tensor var_222 = reshape(shape = var_221, x = var_220)[name = tensor("op_222")]; + tensor cross_1 = mul(x = var_219, y = var_222)[name = tensor("cross_1")]; tensor out_1 = add(x = inner_1, y = cross_1)[name = tensor("out_1")]; - tensor var_199 = mul(x = prev_kv_1, y = sqrt_s0_1)[name = tensor("op_199")]; - tensor var_201_transpose_x_1 = const()[name = tensor("op_201_transpose_x_1"), val = tensor(true)]; - tensor var_201_transpose_y_1 = const()[name = tensor("op_201_transpose_y_1"), val = tensor(false)]; - tensor var_201 = matmul(transpose_x = var_201_transpose_x_1, transpose_y = var_201_transpose_y_1, x = k_1, y = v_1)[name = tensor("op_201")]; - tensor new_kv_unnorm_1 = add(x = var_199, y = var_201)[name = tensor("new_kv_unnorm_1")]; - tensor var_203 = const()[name = tensor("op_203"), val = tensor(0x1p+0)]; - tensor new_scale_1 = add(x = prev_scale_1, y = var_203)[name = tensor("new_scale_1")]; - tensor var_205 = sqrt(x = new_scale_1)[name = tensor("op_205")]; - tensor var_206 = real_div(x = new_kv_unnorm_1, y = var_205)[name = tensor("op_206")]; - tensor var_207_perm_0 = const()[name = tensor("op_207_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_225 = mul(x = prev_kv_1, y = sqrt_s0_1)[name = tensor("op_225")]; + tensor var_227_transpose_x_1 = const()[name = tensor("op_227_transpose_x_1"), val = tensor(true)]; + tensor var_227_transpose_y_1 = const()[name = tensor("op_227_transpose_y_1"), val = tensor(false)]; + tensor var_227 = matmul(transpose_x = var_227_transpose_x_1, transpose_y = var_227_transpose_y_1, x = k_1, y = v_1)[name = tensor("op_227")]; + tensor new_kv_unnorm_1 = add(x = var_225, y = var_227)[name = tensor("new_kv_unnorm_1")]; + tensor var_229 = const()[name = tensor("op_229"), val = tensor(0x1p+0)]; + tensor new_scale_1 = add(x = prev_scale_1, y = var_229)[name = tensor("new_scale_1")]; + tensor var_231 = sqrt(x = new_scale_1)[name = tensor("op_231")]; + tensor var_232 = real_div(x = new_kv_unnorm_1, y = var_231)[name = tensor("op_232")]; + tensor var_233_perm_0 = const()[name = tensor("op_233_perm_0"), val = tensor([0, 2, 1, 3])]; tensor out_3_axes_0 = const()[name = tensor("out_3_axes_0"), val = tensor([-1])]; - tensor var_207 = transpose(perm = var_207_perm_0, x = out_1)[name = tensor("transpose_55")]; - tensor out_3 = layer_norm(axes = out_3_axes_0, epsilon = var_17, x = var_207)[name = tensor("out_3")]; - tensor var_211 = const()[name = tensor("op_211"), val = tensor([1, 1, 256])]; - tensor out_5 = reshape(shape = var_211, x = out_3)[name = tensor("out_5")]; - tensor var_213 = silu(x = input_17)[name = tensor("op_213")]; - tensor input_19 = mul(x = var_213, y = out_5)[name = tensor("input_19")]; - tensor ret_out_1 = linear(bias = encoder_out_proj_0_bias, weight = encoder_out_proj_0_weight, x = input_19)[name = tensor("linear_7")]; - tensor x_3 = add(x = input_13, y = ret_out_1)[name = tensor("x_3")]; + tensor var_233 = transpose(perm = var_233_perm_0, x = out_1)[name = tensor("transpose_55")]; + tensor out_3 = layer_norm(axes = out_3_axes_0, epsilon = var_43, x = var_233)[name = tensor("out_3")]; + tensor var_237 = const()[name = tensor("op_237"), val = tensor([1, 1, 256])]; + tensor out_5 = reshape(shape = var_237, x = out_3)[name = tensor("out_5")]; + tensor var_239 = silu(x = input_19)[name = tensor("op_239")]; + tensor input_21 = mul(x = var_239, y = out_5)[name = tensor("input_21")]; + tensor ret_out_1 = linear(bias = inner_encoder_out_proj_0_bias, weight = inner_encoder_out_proj_0_weight, x = input_21)[name = tensor("linear_7")]; + tensor x_3 = add(x = input_15, y = ret_out_1)[name = tensor("x_3")]; tensor window_1_begin_0 = const()[name = tensor("window_1_begin_0"), val = tensor([0, 0, 0, 0])]; tensor window_1_end_0 = const()[name = tensor("window_1_end_0"), val = tensor([1, 1, 16, 256])]; tensor window_1_end_mask_0 = const()[name = tensor("window_1_end_mask_0"), val = tensor([false, true, true, true])]; tensor window_1_squeeze_mask_0 = const()[name = tensor("window_1_squeeze_mask_0"), val = tensor([true, false, false, false])]; tensor window_1 = slice_by_index(begin = window_1_begin_0, end = window_1_end_0, end_mask = window_1_end_mask_0, squeeze_mask = window_1_squeeze_mask_0, x = enc_conv_cache)[name = tensor("window_1")]; - tensor var_224_begin_0 = const()[name = tensor("op_224_begin_0"), val = tensor([0, 1, 0])]; - tensor var_224_end_0 = const()[name = tensor("op_224_end_0"), val = tensor([1, 16, 256])]; - tensor var_224_end_mask_0 = const()[name = tensor("op_224_end_mask_0"), val = tensor([true, true, true])]; - tensor var_224 = slice_by_index(begin = var_224_begin_0, end = var_224_end_0, end_mask = var_224_end_mask_0, x = window_1)[name = tensor("op_224")]; + tensor var_250_begin_0 = const()[name = tensor("op_250_begin_0"), val = tensor([0, 1, 0])]; + tensor var_250_end_0 = const()[name = tensor("op_250_end_0"), val = tensor([1, 16, 256])]; + tensor var_250_end_mask_0 = const()[name = tensor("op_250_end_mask_0"), val = tensor([true, true, true])]; + tensor var_250 = slice_by_index(begin = var_250_begin_0, end = var_250_end_0, end_mask = var_250_end_mask_0, x = window_1)[name = tensor("op_250")]; tensor window_3_interleave_0 = const()[name = tensor("window_3_interleave_0"), val = tensor(false)]; - tensor window_3 = concat(axis = var_26, interleave = window_3_interleave_0, values = (var_224, x_3))[name = tensor("window_3")]; - tensor input_21_interleave_0 = const()[name = tensor("input_21_interleave_0"), val = tensor(false)]; - tensor input_21 = concat(axis = var_23, interleave = input_21_interleave_0, values = window_3)[name = tensor("input_21")]; + tensor window_3 = concat(axis = var_52, interleave = window_3_interleave_0, values = (var_250, x_3))[name = tensor("window_3")]; + tensor input_23_interleave_0 = const()[name = tensor("input_23_interleave_0"), val = tensor(false)]; + tensor input_23 = concat(axis = var_38, interleave = input_23_interleave_0, values = window_3)[name = tensor("input_23")]; tensor x_5_axes_0 = const()[name = tensor("x_5_axes_0"), val = tensor([-1])]; - tensor x_5 = layer_norm(axes = x_5_axes_0, beta = encoder_conv_module_0_sequential_0_bias, epsilon = var_28, gamma = encoder_conv_module_0_sequential_0_weight, x = input_21)[name = tensor("x_5")]; - tensor input_23_perm_0 = const()[name = tensor("input_23_perm_0"), val = tensor([0, 2, 1])]; + tensor x_5 = layer_norm(axes = x_5_axes_0, beta = inner_encoder_conv_module_0_sequential_0_bias, epsilon = var_35, gamma = inner_encoder_conv_module_0_sequential_0_weight, x = input_23)[name = tensor("x_5")]; + tensor input_25_perm_0 = const()[name = tensor("input_25_perm_0"), val = tensor([0, 2, 1])]; tensor inputs_3_pad_type_0 = const()[name = tensor("inputs_3_pad_type_0"), val = tensor("valid")]; tensor inputs_3_strides_0 = const()[name = tensor("inputs_3_strides_0"), val = tensor([1])]; tensor inputs_3_pad_0 = const()[name = tensor("inputs_3_pad_0"), val = tensor([0, 0])]; tensor inputs_3_dilations_0 = const()[name = tensor("inputs_3_dilations_0"), val = tensor([1])]; tensor inputs_3_groups_0 = const()[name = tensor("inputs_3_groups_0"), val = tensor(1)]; - tensor input_23 = transpose(perm = input_23_perm_0, x = x_5)[name = tensor("transpose_54")]; - tensor inputs_3 = conv(bias = encoder_conv_module_0_sequential_2_conv_bias, dilations = inputs_3_dilations_0, groups = inputs_3_groups_0, pad = inputs_3_pad_0, pad_type = inputs_3_pad_type_0, strides = inputs_3_strides_0, weight = encoder_conv_module_0_sequential_2_conv_weight, x = input_23)[name = tensor("inputs_3")]; - tensor var_249_split_sizes_0 = const()[name = tensor("op_249_split_sizes_0"), val = tensor([256, 256])]; - tensor var_249_axis_0 = const()[name = tensor("op_249_axis_0"), val = tensor(1)]; - tensor var_249_0, tensor var_249_1 = split(axis = var_249_axis_0, split_sizes = var_249_split_sizes_0, x = inputs_3)[name = tensor("op_249")]; - tensor var_251 = sigmoid(x = var_249_1)[name = tensor("op_251")]; - tensor inputs_5 = mul(x = var_249_0, y = var_251)[name = tensor("inputs_5")]; + tensor input_25 = transpose(perm = input_25_perm_0, x = x_5)[name = tensor("transpose_54")]; + tensor inputs_3 = conv(bias = inner_encoder_conv_module_0_sequential_2_conv_bias, dilations = inputs_3_dilations_0, groups = inputs_3_groups_0, pad = inputs_3_pad_0, pad_type = inputs_3_pad_type_0, strides = inputs_3_strides_0, weight = inner_encoder_conv_module_0_sequential_2_conv_weight, x = input_25)[name = tensor("inputs_3")]; + tensor var_275_split_sizes_0 = const()[name = tensor("op_275_split_sizes_0"), val = tensor([256, 256])]; + tensor var_275_axis_0 = const()[name = tensor("op_275_axis_0"), val = tensor(1)]; + tensor var_275_0, tensor var_275_1 = split(axis = var_275_axis_0, split_sizes = var_275_split_sizes_0, x = inputs_3)[name = tensor("op_275")]; + tensor var_277 = sigmoid(x = var_275_1)[name = tensor("op_277")]; + tensor inputs_5 = mul(x = var_275_0, y = var_277)[name = tensor("inputs_5")]; tensor outputs_aug_1_pad_type_0 = const()[name = tensor("outputs_aug_1_pad_type_0"), val = tensor("custom")]; tensor outputs_aug_1_pad_0 = const()[name = tensor("outputs_aug_1_pad_0"), val = tensor([15, 15])]; tensor outputs_aug_1_groups_0 = const()[name = tensor("outputs_aug_1_groups_0"), val = tensor(256)]; tensor outputs_aug_1_strides_0 = const()[name = tensor("outputs_aug_1_strides_0"), val = tensor([1])]; tensor outputs_aug_1_dilations_0 = const()[name = tensor("outputs_aug_1_dilations_0"), val = tensor([1])]; - tensor outputs_aug_1 = conv(dilations = outputs_aug_1_dilations_0, groups = outputs_aug_1_groups_0, pad = outputs_aug_1_pad_0, pad_type = outputs_aug_1_pad_type_0, strides = outputs_aug_1_strides_0, weight = encoder_conv_module_0_sequential_4_conv_weight, x = inputs_5)[name = tensor("outputs_aug_1")]; - tensor input_25_begin_0 = const()[name = tensor("input_25_begin_0"), val = tensor([0, 0, 0])]; - tensor input_25_end_0 = const()[name = tensor("input_25_end_0"), val = tensor([1, 256, 16])]; - tensor input_25_end_mask_0 = const()[name = tensor("input_25_end_mask_0"), val = tensor([true, true, false])]; - tensor input_25 = slice_by_index(begin = input_25_begin_0, end = input_25_end_0, end_mask = input_25_end_mask_0, x = outputs_aug_1)[name = tensor("input_25")]; - tensor inputs_7 = batch_norm(beta = encoder_conv_module_0_sequential_5_bias, epsilon = var_28, gamma = encoder_conv_module_0_sequential_5_weight, mean = encoder_conv_module_0_sequential_5_running_mean, variance = encoder_conv_module_0_sequential_5_running_var, x = input_25)[name = tensor("inputs_7")]; - tensor input_27 = silu(x = inputs_7)[name = tensor("input_27")]; - tensor input_29_pad_type_0 = const()[name = tensor("input_29_pad_type_0"), val = tensor("valid")]; - tensor input_29_strides_0 = const()[name = tensor("input_29_strides_0"), val = tensor([1])]; - tensor input_29_pad_0 = const()[name = tensor("input_29_pad_0"), val = tensor([0, 0])]; - tensor input_29_dilations_0 = const()[name = tensor("input_29_dilations_0"), val = tensor([1])]; - tensor input_29_groups_0 = const()[name = tensor("input_29_groups_0"), val = tensor(1)]; - tensor input_29 = conv(bias = encoder_conv_module_0_sequential_7_conv_bias, dilations = input_29_dilations_0, groups = input_29_groups_0, pad = input_29_pad_0, pad_type = input_29_pad_type_0, strides = input_29_strides_0, weight = encoder_conv_module_0_sequential_7_conv_weight, x = input_27)[name = tensor("input_29")]; + tensor outputs_aug_1 = conv(dilations = outputs_aug_1_dilations_0, groups = outputs_aug_1_groups_0, pad = outputs_aug_1_pad_0, pad_type = outputs_aug_1_pad_type_0, strides = outputs_aug_1_strides_0, weight = inner_encoder_conv_module_0_sequential_4_conv_weight, x = inputs_5)[name = tensor("outputs_aug_1")]; + tensor input_27_begin_0 = const()[name = tensor("input_27_begin_0"), val = tensor([0, 0, 0])]; + tensor input_27_end_0 = const()[name = tensor("input_27_end_0"), val = tensor([1, 256, 16])]; + tensor input_27_end_mask_0 = const()[name = tensor("input_27_end_mask_0"), val = tensor([true, true, false])]; + tensor input_27 = slice_by_index(begin = input_27_begin_0, end = input_27_end_0, end_mask = input_27_end_mask_0, x = outputs_aug_1)[name = tensor("input_27")]; + tensor inputs_7 = batch_norm(beta = inner_encoder_conv_module_0_sequential_5_bias, epsilon = var_35, gamma = inner_encoder_conv_module_0_sequential_5_weight, mean = inner_encoder_conv_module_0_sequential_5_running_mean, variance = inner_encoder_conv_module_0_sequential_5_running_var, x = input_27)[name = tensor("inputs_7")]; + tensor input_29 = silu(x = inputs_7)[name = tensor("input_29")]; + tensor input_31_pad_type_0 = const()[name = tensor("input_31_pad_type_0"), val = tensor("valid")]; + tensor input_31_strides_0 = const()[name = tensor("input_31_strides_0"), val = tensor([1])]; + tensor input_31_pad_0 = const()[name = tensor("input_31_pad_0"), val = tensor([0, 0])]; + tensor input_31_dilations_0 = const()[name = tensor("input_31_dilations_0"), val = tensor([1])]; + tensor input_31_groups_0 = const()[name = tensor("input_31_groups_0"), val = tensor(1)]; + tensor input_31 = conv(bias = inner_encoder_conv_module_0_sequential_7_conv_bias, dilations = input_31_dilations_0, groups = input_31_groups_0, pad = input_31_pad_0, pad_type = input_31_pad_type_0, strides = input_31_strides_0, weight = inner_encoder_conv_module_0_sequential_7_conv_weight, x = input_29)[name = tensor("input_31")]; tensor conv_out_1_perm_0 = const()[name = tensor("conv_out_1_perm_0"), val = tensor([0, 2, 1])]; - tensor var_282_begin_0 = const()[name = tensor("op_282_begin_0"), val = tensor([0, -1, 0])]; - tensor var_282_end_0 = const()[name = tensor("op_282_end_0"), val = tensor([1, 16, 256])]; - tensor var_282_end_mask_0 = const()[name = tensor("op_282_end_mask_0"), val = tensor([true, true, true])]; - tensor conv_out_1 = transpose(perm = conv_out_1_perm_0, x = input_29)[name = tensor("transpose_53")]; - tensor var_282 = slice_by_index(begin = var_282_begin_0, end = var_282_end_0, end_mask = var_282_end_mask_0, x = conv_out_1)[name = tensor("op_282")]; - tensor var_284_perm_0 = const()[name = tensor("op_284_perm_0"), val = tensor([1, 0, 2])]; - tensor var_284 = transpose(perm = var_284_perm_0, x = var_282)[name = tensor("transpose_52")]; - tensor input_31 = add(x = x_3, y = var_284)[name = tensor("input_31")]; - tensor input_33_axes_0 = const()[name = tensor("input_33_axes_0"), val = tensor([-1])]; - tensor input_33 = layer_norm(axes = input_33_axes_0, beta = encoder_ffn2_0_module_sequential_0_bias, epsilon = var_28, gamma = encoder_ffn2_0_module_sequential_0_weight, x = input_31)[name = tensor("input_33")]; - tensor inputs_9 = linear(bias = encoder_ffn2_0_module_sequential_1_linear_bias, weight = encoder_ffn2_0_module_sequential_1_linear_weight, x = input_33)[name = tensor("linear_8")]; - tensor input_35 = silu(x = inputs_9)[name = tensor("input_35")]; - tensor input_39 = linear(bias = encoder_ffn2_0_module_sequential_4_linear_bias, weight = encoder_ffn2_0_module_sequential_4_linear_weight, x = input_35)[name = tensor("linear_9")]; - tensor var_307 = const()[name = tensor("op_307"), val = tensor(0x1p-1)]; - tensor var_308 = mul(x = input_39, y = var_307)[name = tensor("op_308")]; - tensor input_41 = add(x = var_308, y = input_31)[name = tensor("input_41")]; - tensor input_43_axes_0 = const()[name = tensor("input_43_axes_0"), val = tensor([-1])]; - tensor input_43 = layer_norm(axes = input_43_axes_0, beta = encoder_layer_norm_0_bias, epsilon = var_28, gamma = encoder_layer_norm_0_weight, x = input_41)[name = tensor("input_43")]; + tensor var_308_begin_0 = const()[name = tensor("op_308_begin_0"), val = tensor([0, -1, 0])]; + tensor var_308_end_0 = const()[name = tensor("op_308_end_0"), val = tensor([1, 16, 256])]; + tensor var_308_end_mask_0 = const()[name = tensor("op_308_end_mask_0"), val = tensor([true, true, true])]; + tensor conv_out_1 = transpose(perm = conv_out_1_perm_0, x = input_31)[name = tensor("transpose_53")]; + tensor var_308 = slice_by_index(begin = var_308_begin_0, end = var_308_end_0, end_mask = var_308_end_mask_0, x = conv_out_1)[name = tensor("op_308")]; + tensor var_310_perm_0 = const()[name = tensor("op_310_perm_0"), val = tensor([1, 0, 2])]; + tensor var_310 = transpose(perm = var_310_perm_0, x = var_308)[name = tensor("transpose_52")]; + tensor input_33 = add(x = x_3, y = var_310)[name = tensor("input_33")]; + tensor input_35_axes_0 = const()[name = tensor("input_35_axes_0"), val = tensor([-1])]; + tensor input_35 = layer_norm(axes = input_35_axes_0, beta = inner_encoder_ffn2_0_module_sequential_0_bias, epsilon = var_35, gamma = inner_encoder_ffn2_0_module_sequential_0_weight, x = input_33)[name = tensor("input_35")]; + tensor inputs_9 = linear(bias = inner_encoder_ffn2_0_module_sequential_1_linear_bias, weight = inner_encoder_ffn2_0_module_sequential_1_linear_weight, x = input_35)[name = tensor("linear_8")]; + tensor input_37 = silu(x = inputs_9)[name = tensor("input_37")]; + tensor input_41 = linear(bias = inner_encoder_ffn2_0_module_sequential_4_linear_bias, weight = inner_encoder_ffn2_0_module_sequential_4_linear_weight, x = input_37)[name = tensor("linear_9")]; + tensor var_333 = const()[name = tensor("op_333"), val = tensor(0x1p-1)]; + tensor var_334 = mul(x = input_41, y = var_333)[name = tensor("op_334")]; + tensor input_43 = add(x = var_334, y = input_33)[name = tensor("input_43")]; tensor input_45_axes_0 = const()[name = tensor("input_45_axes_0"), val = tensor([-1])]; - tensor input_45 = layer_norm(axes = input_45_axes_0, beta = encoder_ffn1_1_module_sequential_0_bias, epsilon = var_28, gamma = encoder_ffn1_1_module_sequential_0_weight, x = input_43)[name = tensor("input_45")]; - tensor inputs_11 = linear(bias = encoder_ffn1_1_module_sequential_1_linear_bias, weight = encoder_ffn1_1_module_sequential_1_linear_weight, x = input_45)[name = tensor("linear_10")]; - tensor input_47 = silu(x = inputs_11)[name = tensor("input_47")]; - tensor input_51 = linear(bias = encoder_ffn1_1_module_sequential_4_linear_bias, weight = encoder_ffn1_1_module_sequential_4_linear_weight, x = input_47)[name = tensor("linear_11")]; - tensor var_337 = const()[name = tensor("op_337"), val = tensor(0x1p-1)]; - tensor var_338 = mul(x = input_51, y = var_337)[name = tensor("op_338")]; - tensor input_53 = add(x = var_338, y = input_43)[name = tensor("input_53")]; + tensor input_45 = layer_norm(axes = input_45_axes_0, beta = inner_encoder_layer_norm_0_bias, epsilon = var_35, gamma = inner_encoder_layer_norm_0_weight, x = input_43)[name = tensor("input_45")]; + tensor input_47_axes_0 = const()[name = tensor("input_47_axes_0"), val = tensor([-1])]; + tensor input_47 = layer_norm(axes = input_47_axes_0, beta = inner_encoder_ffn1_1_module_sequential_0_bias, epsilon = var_35, gamma = inner_encoder_ffn1_1_module_sequential_0_weight, x = input_45)[name = tensor("input_47")]; + tensor inputs_11 = linear(bias = inner_encoder_ffn1_1_module_sequential_1_linear_bias, weight = inner_encoder_ffn1_1_module_sequential_1_linear_weight, x = input_47)[name = tensor("linear_10")]; + tensor input_49 = silu(x = inputs_11)[name = tensor("input_49")]; + tensor input_53 = linear(bias = inner_encoder_ffn1_1_module_sequential_4_linear_bias, weight = inner_encoder_ffn1_1_module_sequential_4_linear_weight, x = input_49)[name = tensor("linear_11")]; + tensor var_363 = const()[name = tensor("op_363"), val = tensor(0x1p-1)]; + tensor var_364 = mul(x = input_53, y = var_363)[name = tensor("op_364")]; + tensor input_55 = add(x = var_364, y = input_45)[name = tensor("input_55")]; tensor x_7_axes_0 = const()[name = tensor("x_7_axes_0"), val = tensor([-1])]; - tensor x_7 = layer_norm(axes = x_7_axes_0, beta = encoder_ret_lns_1_bias, epsilon = var_28, gamma = encoder_ret_lns_1_weight, x = input_53)[name = tensor("x_7")]; + tensor x_7 = layer_norm(axes = x_7_axes_0, beta = inner_encoder_ret_lns_1_bias, epsilon = var_35, gamma = inner_encoder_ret_lns_1_weight, x = input_55)[name = tensor("x_7")]; tensor prev_kv_3_begin_0 = const()[name = tensor("prev_kv_3_begin_0"), val = tensor([1, 0, 0, 0, 0])]; tensor prev_kv_3_end_0 = const()[name = tensor("prev_kv_3_end_0"), val = tensor([2, 1, 4, 64, 64])]; tensor prev_kv_3_end_mask_0 = const()[name = tensor("prev_kv_3_end_mask_0"), val = tensor([false, true, true, true, true])]; @@ -381,139 +387,139 @@ program(1.0) tensor prev_scale_3_end_mask_0 = const()[name = tensor("prev_scale_3_end_mask_0"), val = tensor([false, true])]; tensor prev_scale_3_squeeze_mask_0 = const()[name = tensor("prev_scale_3_squeeze_mask_0"), val = tensor([true, false])]; tensor prev_scale_3 = slice_by_index(begin = prev_scale_3_begin_0, end = prev_scale_3_end_0, end_mask = prev_scale_3_end_mask_0, squeeze_mask = prev_scale_3_squeeze_mask_0, x = enc_scale)[name = tensor("prev_scale_3")]; - tensor var_352 = linear(bias = encoder_q_proj_1_bias, weight = encoder_q_proj_1_weight, x = x_7)[name = tensor("linear_12")]; - tensor var_353 = const()[name = tensor("op_353"), val = tensor([1, 1, 4, 64])]; - tensor var_354 = reshape(shape = var_353, x = var_352)[name = tensor("op_354")]; + tensor var_378 = linear(bias = inner_encoder_q_proj_1_bias, weight = inner_encoder_q_proj_1_weight, x = x_7)[name = tensor("linear_12")]; + tensor var_379 = const()[name = tensor("op_379"), val = tensor([1, 1, 4, 64])]; + tensor var_380 = reshape(shape = var_379, x = var_378)[name = tensor("op_380")]; tensor q_3_perm_0 = const()[name = tensor("q_3_perm_0"), val = tensor([0, 2, 1, 3])]; - tensor var_358 = linear(bias = encoder_k_proj_1_bias, weight = encoder_k_proj_1_weight, x = x_7)[name = tensor("linear_13")]; - tensor var_359 = const()[name = tensor("op_359"), val = tensor(0x1p-3)]; - tensor var_360 = mul(x = var_358, y = var_359)[name = tensor("op_360")]; - tensor var_361 = const()[name = tensor("op_361"), val = tensor([1, 1, 4, 64])]; - tensor var_362 = reshape(shape = var_361, x = var_360)[name = tensor("op_362")]; + tensor var_384 = linear(bias = inner_encoder_k_proj_1_bias, weight = inner_encoder_k_proj_1_weight, x = x_7)[name = tensor("linear_13")]; + tensor var_385 = const()[name = tensor("op_385"), val = tensor(0x1p-3)]; + tensor var_386 = mul(x = var_384, y = var_385)[name = tensor("op_386")]; + tensor var_387 = const()[name = tensor("op_387"), val = tensor([1, 1, 4, 64])]; + tensor var_388 = reshape(shape = var_387, x = var_386)[name = tensor("op_388")]; tensor k_3_perm_0 = const()[name = tensor("k_3_perm_0"), val = tensor([0, 2, 1, 3])]; - tensor var_366 = linear(bias = encoder_v_proj_1_bias, weight = encoder_v_proj_1_weight, x = x_7)[name = tensor("linear_14")]; - tensor var_367 = const()[name = tensor("op_367"), val = tensor([1, 1, 4, 64])]; - tensor var_368 = reshape(shape = var_367, x = var_366)[name = tensor("op_368")]; + tensor var_392 = linear(bias = inner_encoder_v_proj_1_bias, weight = inner_encoder_v_proj_1_weight, x = x_7)[name = tensor("linear_14")]; + tensor var_393 = const()[name = tensor("op_393"), val = tensor([1, 1, 4, 64])]; + tensor var_394 = reshape(shape = var_393, x = var_392)[name = tensor("op_394")]; tensor v_3_perm_0 = const()[name = tensor("v_3_perm_0"), val = tensor([0, 2, 1, 3])]; - tensor input_57 = linear(bias = encoder_g_proj_1_bias, weight = encoder_g_proj_1_weight, x = x_7)[name = tensor("linear_15")]; + tensor input_59 = linear(bias = inner_encoder_g_proj_1_bias, weight = inner_encoder_g_proj_1_weight, x = x_7)[name = tensor("linear_15")]; tensor sqrt_s0_3 = sqrt(x = prev_scale_3)[name = tensor("sqrt_s0_3")]; - tensor s_t_3 = add(x = prev_scale_3, y = encoder__t_index)[name = tensor("s_t_3")]; + tensor s_t_3 = add(x = prev_scale_3, y = inner_encoder__t_index)[name = tensor("s_t_3")]; tensor sqrt_s_t_3 = sqrt(x = s_t_3)[name = tensor("sqrt_s_t_3")]; tensor qk_3_transpose_x_1 = const()[name = tensor("qk_3_transpose_x_1"), val = tensor(false)]; tensor qk_3_transpose_y_1 = const()[name = tensor("qk_3_transpose_y_1"), val = tensor(true)]; - tensor k_3 = transpose(perm = k_3_perm_0, x = var_362)[name = tensor("transpose_50")]; - tensor q_3 = transpose(perm = q_3_perm_0, x = var_354)[name = tensor("transpose_51")]; + tensor k_3 = transpose(perm = k_3_perm_0, x = var_388)[name = tensor("transpose_50")]; + tensor q_3 = transpose(perm = q_3_perm_0, x = var_380)[name = tensor("transpose_51")]; tensor qk_3 = matmul(transpose_x = qk_3_transpose_x_1, transpose_y = qk_3_transpose_y_1, x = q_3, y = k_3)[name = tensor("qk_3")]; - tensor var_378 = const()[name = tensor("op_378"), val = tensor([1, 1])]; - tensor var_379 = reshape(shape = var_378, x = sqrt_s_t_3)[name = tensor("op_379")]; - tensor M_3 = real_div(x = encoder__causal_mask, y = var_379)[name = tensor("M_3")]; - tensor var_381 = mul(x = qk_3, y = M_3)[name = tensor("op_381")]; + tensor var_404 = const()[name = tensor("op_404"), val = tensor([1, 1])]; + tensor var_405 = reshape(shape = var_404, x = sqrt_s_t_3)[name = tensor("op_405")]; + tensor M_3 = real_div(x = inner_encoder__causal_mask, y = var_405)[name = tensor("M_3")]; + tensor var_407 = mul(x = qk_3, y = M_3)[name = tensor("op_407")]; tensor inner_3_transpose_x_0 = const()[name = tensor("inner_3_transpose_x_0"), val = tensor(false)]; tensor inner_3_transpose_y_0 = const()[name = tensor("inner_3_transpose_y_0"), val = tensor(false)]; - tensor v_3 = transpose(perm = v_3_perm_0, x = var_368)[name = tensor("transpose_49")]; - tensor inner_3 = matmul(transpose_x = inner_3_transpose_x_0, transpose_y = inner_3_transpose_y_0, x = var_381, y = v_3)[name = tensor("inner_3")]; - tensor var_383_transpose_x_0 = const()[name = tensor("op_383_transpose_x_0"), val = tensor(false)]; - tensor var_383_transpose_y_0 = const()[name = tensor("op_383_transpose_y_0"), val = tensor(false)]; - tensor var_383 = matmul(transpose_x = var_383_transpose_x_0, transpose_y = var_383_transpose_y_0, x = q_3, y = prev_kv_3)[name = tensor("op_383")]; - tensor var_384 = real_div(x = sqrt_s0_3, y = sqrt_s_t_3)[name = tensor("op_384")]; - tensor var_385 = const()[name = tensor("op_385"), val = tensor([1, 1, 1, 1])]; - tensor var_386 = reshape(shape = var_385, x = var_384)[name = tensor("op_386")]; - tensor cross_3 = mul(x = var_383, y = var_386)[name = tensor("cross_3")]; + tensor v_3 = transpose(perm = v_3_perm_0, x = var_394)[name = tensor("transpose_49")]; + tensor inner_3 = matmul(transpose_x = inner_3_transpose_x_0, transpose_y = inner_3_transpose_y_0, x = var_407, y = v_3)[name = tensor("inner_3")]; + tensor var_409_transpose_x_0 = const()[name = tensor("op_409_transpose_x_0"), val = tensor(false)]; + tensor var_409_transpose_y_0 = const()[name = tensor("op_409_transpose_y_0"), val = tensor(false)]; + tensor var_409 = matmul(transpose_x = var_409_transpose_x_0, transpose_y = var_409_transpose_y_0, x = q_3, y = prev_kv_3)[name = tensor("op_409")]; + tensor var_410 = real_div(x = sqrt_s0_3, y = sqrt_s_t_3)[name = tensor("op_410")]; + tensor var_411 = const()[name = tensor("op_411"), val = tensor([1, 1, 1, 1])]; + tensor var_412 = reshape(shape = var_411, x = var_410)[name = tensor("op_412")]; + tensor cross_3 = mul(x = var_409, y = var_412)[name = tensor("cross_3")]; tensor out_7 = add(x = inner_3, y = cross_3)[name = tensor("out_7")]; - tensor var_389 = mul(x = prev_kv_3, y = sqrt_s0_3)[name = tensor("op_389")]; - tensor var_391_transpose_x_1 = const()[name = tensor("op_391_transpose_x_1"), val = tensor(true)]; - tensor var_391_transpose_y_1 = const()[name = tensor("op_391_transpose_y_1"), val = tensor(false)]; - tensor var_391 = matmul(transpose_x = var_391_transpose_x_1, transpose_y = var_391_transpose_y_1, x = k_3, y = v_3)[name = tensor("op_391")]; - tensor new_kv_unnorm_3 = add(x = var_389, y = var_391)[name = tensor("new_kv_unnorm_3")]; - tensor var_393 = const()[name = tensor("op_393"), val = tensor(0x1p+0)]; - tensor new_scale_3 = add(x = prev_scale_3, y = var_393)[name = tensor("new_scale_3")]; - tensor var_395 = sqrt(x = new_scale_3)[name = tensor("op_395")]; - tensor var_396 = real_div(x = new_kv_unnorm_3, y = var_395)[name = tensor("op_396")]; - tensor var_397_perm_0 = const()[name = tensor("op_397_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_415 = mul(x = prev_kv_3, y = sqrt_s0_3)[name = tensor("op_415")]; + tensor var_417_transpose_x_1 = const()[name = tensor("op_417_transpose_x_1"), val = tensor(true)]; + tensor var_417_transpose_y_1 = const()[name = tensor("op_417_transpose_y_1"), val = tensor(false)]; + tensor var_417 = matmul(transpose_x = var_417_transpose_x_1, transpose_y = var_417_transpose_y_1, x = k_3, y = v_3)[name = tensor("op_417")]; + tensor new_kv_unnorm_3 = add(x = var_415, y = var_417)[name = tensor("new_kv_unnorm_3")]; + tensor var_419 = const()[name = tensor("op_419"), val = tensor(0x1p+0)]; + tensor new_scale_3 = add(x = prev_scale_3, y = var_419)[name = tensor("new_scale_3")]; + tensor var_421 = sqrt(x = new_scale_3)[name = tensor("op_421")]; + tensor var_422 = real_div(x = new_kv_unnorm_3, y = var_421)[name = tensor("op_422")]; + tensor var_423_perm_0 = const()[name = tensor("op_423_perm_0"), val = tensor([0, 2, 1, 3])]; tensor out_9_axes_0 = const()[name = tensor("out_9_axes_0"), val = tensor([-1])]; - tensor var_397 = transpose(perm = var_397_perm_0, x = out_7)[name = tensor("transpose_48")]; - tensor out_9 = layer_norm(axes = out_9_axes_0, epsilon = var_17, x = var_397)[name = tensor("out_9")]; - tensor var_401 = const()[name = tensor("op_401"), val = tensor([1, 1, 256])]; - tensor out_11 = reshape(shape = var_401, x = out_9)[name = tensor("out_11")]; - tensor var_403 = silu(x = input_57)[name = tensor("op_403")]; - tensor input_59 = mul(x = var_403, y = out_11)[name = tensor("input_59")]; - tensor ret_out_3 = linear(bias = encoder_out_proj_1_bias, weight = encoder_out_proj_1_weight, x = input_59)[name = tensor("linear_16")]; - tensor x_9 = add(x = input_53, y = ret_out_3)[name = tensor("x_9")]; + tensor var_423 = transpose(perm = var_423_perm_0, x = out_7)[name = tensor("transpose_48")]; + tensor out_9 = layer_norm(axes = out_9_axes_0, epsilon = var_43, x = var_423)[name = tensor("out_9")]; + tensor var_427 = const()[name = tensor("op_427"), val = tensor([1, 1, 256])]; + tensor out_11 = reshape(shape = var_427, x = out_9)[name = tensor("out_11")]; + tensor var_429 = silu(x = input_59)[name = tensor("op_429")]; + tensor input_61 = mul(x = var_429, y = out_11)[name = tensor("input_61")]; + tensor ret_out_3 = linear(bias = inner_encoder_out_proj_1_bias, weight = inner_encoder_out_proj_1_weight, x = input_61)[name = tensor("linear_16")]; + tensor x_9 = add(x = input_55, y = ret_out_3)[name = tensor("x_9")]; tensor window_5_begin_0 = const()[name = tensor("window_5_begin_0"), val = tensor([1, 0, 0, 0])]; tensor window_5_end_0 = const()[name = tensor("window_5_end_0"), val = tensor([2, 1, 16, 256])]; tensor window_5_end_mask_0 = const()[name = tensor("window_5_end_mask_0"), val = tensor([false, true, true, true])]; tensor window_5_squeeze_mask_0 = const()[name = tensor("window_5_squeeze_mask_0"), val = tensor([true, false, false, false])]; tensor window_5 = slice_by_index(begin = window_5_begin_0, end = window_5_end_0, end_mask = window_5_end_mask_0, squeeze_mask = window_5_squeeze_mask_0, x = enc_conv_cache)[name = tensor("window_5")]; - tensor var_414_begin_0 = const()[name = tensor("op_414_begin_0"), val = tensor([0, 1, 0])]; - tensor var_414_end_0 = const()[name = tensor("op_414_end_0"), val = tensor([1, 16, 256])]; - tensor var_414_end_mask_0 = const()[name = tensor("op_414_end_mask_0"), val = tensor([true, true, true])]; - tensor var_414 = slice_by_index(begin = var_414_begin_0, end = var_414_end_0, end_mask = var_414_end_mask_0, x = window_5)[name = tensor("op_414")]; + tensor var_440_begin_0 = const()[name = tensor("op_440_begin_0"), val = tensor([0, 1, 0])]; + tensor var_440_end_0 = const()[name = tensor("op_440_end_0"), val = tensor([1, 16, 256])]; + tensor var_440_end_mask_0 = const()[name = tensor("op_440_end_mask_0"), val = tensor([true, true, true])]; + tensor var_440 = slice_by_index(begin = var_440_begin_0, end = var_440_end_0, end_mask = var_440_end_mask_0, x = window_5)[name = tensor("op_440")]; tensor window_7_interleave_0 = const()[name = tensor("window_7_interleave_0"), val = tensor(false)]; - tensor window_7 = concat(axis = var_26, interleave = window_7_interleave_0, values = (var_414, x_9))[name = tensor("window_7")]; - tensor input_61_interleave_0 = const()[name = tensor("input_61_interleave_0"), val = tensor(false)]; - tensor input_61 = concat(axis = var_23, interleave = input_61_interleave_0, values = window_7)[name = tensor("input_61")]; + tensor window_7 = concat(axis = var_52, interleave = window_7_interleave_0, values = (var_440, x_9))[name = tensor("window_7")]; + tensor input_63_interleave_0 = const()[name = tensor("input_63_interleave_0"), val = tensor(false)]; + tensor input_63 = concat(axis = var_38, interleave = input_63_interleave_0, values = window_7)[name = tensor("input_63")]; tensor x_11_axes_0 = const()[name = tensor("x_11_axes_0"), val = tensor([-1])]; - tensor x_11 = layer_norm(axes = x_11_axes_0, beta = encoder_conv_module_1_sequential_0_bias, epsilon = var_28, gamma = encoder_conv_module_1_sequential_0_weight, x = input_61)[name = tensor("x_11")]; - tensor input_63_perm_0 = const()[name = tensor("input_63_perm_0"), val = tensor([0, 2, 1])]; + tensor x_11 = layer_norm(axes = x_11_axes_0, beta = inner_encoder_conv_module_1_sequential_0_bias, epsilon = var_35, gamma = inner_encoder_conv_module_1_sequential_0_weight, x = input_63)[name = tensor("x_11")]; + tensor input_65_perm_0 = const()[name = tensor("input_65_perm_0"), val = tensor([0, 2, 1])]; tensor inputs_13_pad_type_0 = const()[name = tensor("inputs_13_pad_type_0"), val = tensor("valid")]; tensor inputs_13_strides_0 = const()[name = tensor("inputs_13_strides_0"), val = tensor([1])]; tensor inputs_13_pad_0 = const()[name = tensor("inputs_13_pad_0"), val = tensor([0, 0])]; tensor inputs_13_dilations_0 = const()[name = tensor("inputs_13_dilations_0"), val = tensor([1])]; tensor inputs_13_groups_0 = const()[name = tensor("inputs_13_groups_0"), val = tensor(1)]; - tensor input_63 = transpose(perm = input_63_perm_0, x = x_11)[name = tensor("transpose_47")]; - tensor inputs_13 = conv(bias = encoder_conv_module_1_sequential_2_conv_bias, dilations = inputs_13_dilations_0, groups = inputs_13_groups_0, pad = inputs_13_pad_0, pad_type = inputs_13_pad_type_0, strides = inputs_13_strides_0, weight = encoder_conv_module_1_sequential_2_conv_weight, x = input_63)[name = tensor("inputs_13")]; - tensor var_439_split_sizes_0 = const()[name = tensor("op_439_split_sizes_0"), val = tensor([256, 256])]; - tensor var_439_axis_0 = const()[name = tensor("op_439_axis_0"), val = tensor(1)]; - tensor var_439_0, tensor var_439_1 = split(axis = var_439_axis_0, split_sizes = var_439_split_sizes_0, x = inputs_13)[name = tensor("op_439")]; - tensor var_441 = sigmoid(x = var_439_1)[name = tensor("op_441")]; - tensor inputs_15 = mul(x = var_439_0, y = var_441)[name = tensor("inputs_15")]; + tensor input_65 = transpose(perm = input_65_perm_0, x = x_11)[name = tensor("transpose_47")]; + tensor inputs_13 = conv(bias = inner_encoder_conv_module_1_sequential_2_conv_bias, dilations = inputs_13_dilations_0, groups = inputs_13_groups_0, pad = inputs_13_pad_0, pad_type = inputs_13_pad_type_0, strides = inputs_13_strides_0, weight = inner_encoder_conv_module_1_sequential_2_conv_weight, x = input_65)[name = tensor("inputs_13")]; + tensor var_465_split_sizes_0 = const()[name = tensor("op_465_split_sizes_0"), val = tensor([256, 256])]; + tensor var_465_axis_0 = const()[name = tensor("op_465_axis_0"), val = tensor(1)]; + tensor var_465_0, tensor var_465_1 = split(axis = var_465_axis_0, split_sizes = var_465_split_sizes_0, x = inputs_13)[name = tensor("op_465")]; + tensor var_467 = sigmoid(x = var_465_1)[name = tensor("op_467")]; + tensor inputs_15 = mul(x = var_465_0, y = var_467)[name = tensor("inputs_15")]; tensor outputs_aug_3_pad_type_0 = const()[name = tensor("outputs_aug_3_pad_type_0"), val = tensor("custom")]; tensor outputs_aug_3_pad_0 = const()[name = tensor("outputs_aug_3_pad_0"), val = tensor([15, 15])]; tensor outputs_aug_3_groups_0 = const()[name = tensor("outputs_aug_3_groups_0"), val = tensor(256)]; tensor outputs_aug_3_strides_0 = const()[name = tensor("outputs_aug_3_strides_0"), val = tensor([1])]; tensor outputs_aug_3_dilations_0 = const()[name = tensor("outputs_aug_3_dilations_0"), val = tensor([1])]; - tensor outputs_aug_3 = conv(dilations = outputs_aug_3_dilations_0, groups = outputs_aug_3_groups_0, pad = outputs_aug_3_pad_0, pad_type = outputs_aug_3_pad_type_0, strides = outputs_aug_3_strides_0, weight = encoder_conv_module_1_sequential_4_conv_weight, x = inputs_15)[name = tensor("outputs_aug_3")]; - tensor input_65_begin_0 = const()[name = tensor("input_65_begin_0"), val = tensor([0, 0, 0])]; - tensor input_65_end_0 = const()[name = tensor("input_65_end_0"), val = tensor([1, 256, 16])]; - tensor input_65_end_mask_0 = const()[name = tensor("input_65_end_mask_0"), val = tensor([true, true, false])]; - tensor input_65 = slice_by_index(begin = input_65_begin_0, end = input_65_end_0, end_mask = input_65_end_mask_0, x = outputs_aug_3)[name = tensor("input_65")]; - tensor inputs_17 = batch_norm(beta = encoder_conv_module_1_sequential_5_bias, epsilon = var_28, gamma = encoder_conv_module_1_sequential_5_weight, mean = encoder_conv_module_1_sequential_5_running_mean, variance = encoder_conv_module_1_sequential_5_running_var, x = input_65)[name = tensor("inputs_17")]; - tensor input_67 = silu(x = inputs_17)[name = tensor("input_67")]; - tensor input_69_pad_type_0 = const()[name = tensor("input_69_pad_type_0"), val = tensor("valid")]; - tensor input_69_strides_0 = const()[name = tensor("input_69_strides_0"), val = tensor([1])]; - tensor input_69_pad_0 = const()[name = tensor("input_69_pad_0"), val = tensor([0, 0])]; - tensor input_69_dilations_0 = const()[name = tensor("input_69_dilations_0"), val = tensor([1])]; - tensor input_69_groups_0 = const()[name = tensor("input_69_groups_0"), val = tensor(1)]; - tensor input_69 = conv(bias = encoder_conv_module_1_sequential_7_conv_bias, dilations = input_69_dilations_0, groups = input_69_groups_0, pad = input_69_pad_0, pad_type = input_69_pad_type_0, strides = input_69_strides_0, weight = encoder_conv_module_1_sequential_7_conv_weight, x = input_67)[name = tensor("input_69")]; + tensor outputs_aug_3 = conv(dilations = outputs_aug_3_dilations_0, groups = outputs_aug_3_groups_0, pad = outputs_aug_3_pad_0, pad_type = outputs_aug_3_pad_type_0, strides = outputs_aug_3_strides_0, weight = inner_encoder_conv_module_1_sequential_4_conv_weight, x = inputs_15)[name = tensor("outputs_aug_3")]; + tensor input_67_begin_0 = const()[name = tensor("input_67_begin_0"), val = tensor([0, 0, 0])]; + tensor input_67_end_0 = const()[name = tensor("input_67_end_0"), val = tensor([1, 256, 16])]; + tensor input_67_end_mask_0 = const()[name = tensor("input_67_end_mask_0"), val = tensor([true, true, false])]; + tensor input_67 = slice_by_index(begin = input_67_begin_0, end = input_67_end_0, end_mask = input_67_end_mask_0, x = outputs_aug_3)[name = tensor("input_67")]; + tensor inputs_17 = batch_norm(beta = inner_encoder_conv_module_1_sequential_5_bias, epsilon = var_35, gamma = inner_encoder_conv_module_1_sequential_5_weight, mean = inner_encoder_conv_module_1_sequential_5_running_mean, variance = inner_encoder_conv_module_1_sequential_5_running_var, x = input_67)[name = tensor("inputs_17")]; + tensor input_69 = silu(x = inputs_17)[name = tensor("input_69")]; + tensor input_71_pad_type_0 = const()[name = tensor("input_71_pad_type_0"), val = tensor("valid")]; + tensor input_71_strides_0 = const()[name = tensor("input_71_strides_0"), val = tensor([1])]; + tensor input_71_pad_0 = const()[name = tensor("input_71_pad_0"), val = tensor([0, 0])]; + tensor input_71_dilations_0 = const()[name = tensor("input_71_dilations_0"), val = tensor([1])]; + tensor input_71_groups_0 = const()[name = tensor("input_71_groups_0"), val = tensor(1)]; + tensor input_71 = conv(bias = inner_encoder_conv_module_1_sequential_7_conv_bias, dilations = input_71_dilations_0, groups = input_71_groups_0, pad = input_71_pad_0, pad_type = input_71_pad_type_0, strides = input_71_strides_0, weight = inner_encoder_conv_module_1_sequential_7_conv_weight, x = input_69)[name = tensor("input_71")]; tensor conv_out_3_perm_0 = const()[name = tensor("conv_out_3_perm_0"), val = tensor([0, 2, 1])]; - tensor var_472_begin_0 = const()[name = tensor("op_472_begin_0"), val = tensor([0, -1, 0])]; - tensor var_472_end_0 = const()[name = tensor("op_472_end_0"), val = tensor([1, 16, 256])]; - tensor var_472_end_mask_0 = const()[name = tensor("op_472_end_mask_0"), val = tensor([true, true, true])]; - tensor conv_out_3 = transpose(perm = conv_out_3_perm_0, x = input_69)[name = tensor("transpose_46")]; - tensor var_472 = slice_by_index(begin = var_472_begin_0, end = var_472_end_0, end_mask = var_472_end_mask_0, x = conv_out_3)[name = tensor("op_472")]; - tensor var_474_perm_0 = const()[name = tensor("op_474_perm_0"), val = tensor([1, 0, 2])]; - tensor var_474 = transpose(perm = var_474_perm_0, x = var_472)[name = tensor("transpose_45")]; - tensor input_71 = add(x = x_9, y = var_474)[name = tensor("input_71")]; - tensor input_73_axes_0 = const()[name = tensor("input_73_axes_0"), val = tensor([-1])]; - tensor input_73 = layer_norm(axes = input_73_axes_0, beta = encoder_ffn2_1_module_sequential_0_bias, epsilon = var_28, gamma = encoder_ffn2_1_module_sequential_0_weight, x = input_71)[name = tensor("input_73")]; - tensor inputs_19 = linear(bias = encoder_ffn2_1_module_sequential_1_linear_bias, weight = encoder_ffn2_1_module_sequential_1_linear_weight, x = input_73)[name = tensor("linear_17")]; - tensor input_75 = silu(x = inputs_19)[name = tensor("input_75")]; - tensor input_79 = linear(bias = encoder_ffn2_1_module_sequential_4_linear_bias, weight = encoder_ffn2_1_module_sequential_4_linear_weight, x = input_75)[name = tensor("linear_18")]; - tensor var_497 = const()[name = tensor("op_497"), val = tensor(0x1p-1)]; - tensor var_498 = mul(x = input_79, y = var_497)[name = tensor("op_498")]; - tensor input_81 = add(x = var_498, y = input_71)[name = tensor("input_81")]; - tensor input_83_axes_0 = const()[name = tensor("input_83_axes_0"), val = tensor([-1])]; - tensor input_83 = layer_norm(axes = input_83_axes_0, beta = encoder_layer_norm_1_bias, epsilon = var_28, gamma = encoder_layer_norm_1_weight, x = input_81)[name = tensor("input_83")]; + tensor var_498_begin_0 = const()[name = tensor("op_498_begin_0"), val = tensor([0, -1, 0])]; + tensor var_498_end_0 = const()[name = tensor("op_498_end_0"), val = tensor([1, 16, 256])]; + tensor var_498_end_mask_0 = const()[name = tensor("op_498_end_mask_0"), val = tensor([true, true, true])]; + tensor conv_out_3 = transpose(perm = conv_out_3_perm_0, x = input_71)[name = tensor("transpose_46")]; + tensor var_498 = slice_by_index(begin = var_498_begin_0, end = var_498_end_0, end_mask = var_498_end_mask_0, x = conv_out_3)[name = tensor("op_498")]; + tensor var_500_perm_0 = const()[name = tensor("op_500_perm_0"), val = tensor([1, 0, 2])]; + tensor var_500 = transpose(perm = var_500_perm_0, x = var_498)[name = tensor("transpose_45")]; + tensor input_73 = add(x = x_9, y = var_500)[name = tensor("input_73")]; + tensor input_75_axes_0 = const()[name = tensor("input_75_axes_0"), val = tensor([-1])]; + tensor input_75 = layer_norm(axes = input_75_axes_0, beta = inner_encoder_ffn2_1_module_sequential_0_bias, epsilon = var_35, gamma = inner_encoder_ffn2_1_module_sequential_0_weight, x = input_73)[name = tensor("input_75")]; + tensor inputs_19 = linear(bias = inner_encoder_ffn2_1_module_sequential_1_linear_bias, weight = inner_encoder_ffn2_1_module_sequential_1_linear_weight, x = input_75)[name = tensor("linear_17")]; + tensor input_77 = silu(x = inputs_19)[name = tensor("input_77")]; + tensor input_81 = linear(bias = inner_encoder_ffn2_1_module_sequential_4_linear_bias, weight = inner_encoder_ffn2_1_module_sequential_4_linear_weight, x = input_77)[name = tensor("linear_18")]; + tensor var_523 = const()[name = tensor("op_523"), val = tensor(0x1p-1)]; + tensor var_524 = mul(x = input_81, y = var_523)[name = tensor("op_524")]; + tensor input_83 = add(x = var_524, y = input_73)[name = tensor("input_83")]; tensor input_85_axes_0 = const()[name = tensor("input_85_axes_0"), val = tensor([-1])]; - tensor input_85 = layer_norm(axes = input_85_axes_0, beta = encoder_ffn1_2_module_sequential_0_bias, epsilon = var_28, gamma = encoder_ffn1_2_module_sequential_0_weight, x = input_83)[name = tensor("input_85")]; - tensor inputs_21 = linear(bias = encoder_ffn1_2_module_sequential_1_linear_bias, weight = encoder_ffn1_2_module_sequential_1_linear_weight, x = input_85)[name = tensor("linear_19")]; - tensor input_87 = silu(x = inputs_21)[name = tensor("input_87")]; - tensor input_91 = linear(bias = encoder_ffn1_2_module_sequential_4_linear_bias, weight = encoder_ffn1_2_module_sequential_4_linear_weight, x = input_87)[name = tensor("linear_20")]; - tensor var_527 = const()[name = tensor("op_527"), val = tensor(0x1p-1)]; - tensor var_528 = mul(x = input_91, y = var_527)[name = tensor("op_528")]; - tensor input_93 = add(x = var_528, y = input_83)[name = tensor("input_93")]; + tensor input_85 = layer_norm(axes = input_85_axes_0, beta = inner_encoder_layer_norm_1_bias, epsilon = var_35, gamma = inner_encoder_layer_norm_1_weight, x = input_83)[name = tensor("input_85")]; + tensor input_87_axes_0 = const()[name = tensor("input_87_axes_0"), val = tensor([-1])]; + tensor input_87 = layer_norm(axes = input_87_axes_0, beta = inner_encoder_ffn1_2_module_sequential_0_bias, epsilon = var_35, gamma = inner_encoder_ffn1_2_module_sequential_0_weight, x = input_85)[name = tensor("input_87")]; + tensor inputs_21 = linear(bias = inner_encoder_ffn1_2_module_sequential_1_linear_bias, weight = inner_encoder_ffn1_2_module_sequential_1_linear_weight, x = input_87)[name = tensor("linear_19")]; + tensor input_89 = silu(x = inputs_21)[name = tensor("input_89")]; + tensor input_93 = linear(bias = inner_encoder_ffn1_2_module_sequential_4_linear_bias, weight = inner_encoder_ffn1_2_module_sequential_4_linear_weight, x = input_89)[name = tensor("linear_20")]; + tensor var_553 = const()[name = tensor("op_553"), val = tensor(0x1p-1)]; + tensor var_554 = mul(x = input_93, y = var_553)[name = tensor("op_554")]; + tensor input_95 = add(x = var_554, y = input_85)[name = tensor("input_95")]; tensor x_13_axes_0 = const()[name = tensor("x_13_axes_0"), val = tensor([-1])]; - tensor x_13 = layer_norm(axes = x_13_axes_0, beta = encoder_ret_lns_2_bias, epsilon = var_28, gamma = encoder_ret_lns_2_weight, x = input_93)[name = tensor("x_13")]; + tensor x_13 = layer_norm(axes = x_13_axes_0, beta = inner_encoder_ret_lns_2_bias, epsilon = var_35, gamma = inner_encoder_ret_lns_2_weight, x = input_95)[name = tensor("x_13")]; tensor prev_kv_5_begin_0 = const()[name = tensor("prev_kv_5_begin_0"), val = tensor([2, 0, 0, 0, 0])]; tensor prev_kv_5_end_0 = const()[name = tensor("prev_kv_5_end_0"), val = tensor([3, 1, 4, 64, 64])]; tensor prev_kv_5_end_mask_0 = const()[name = tensor("prev_kv_5_end_mask_0"), val = tensor([false, true, true, true, true])]; @@ -524,139 +530,139 @@ program(1.0) tensor prev_scale_5_end_mask_0 = const()[name = tensor("prev_scale_5_end_mask_0"), val = tensor([false, true])]; tensor prev_scale_5_squeeze_mask_0 = const()[name = tensor("prev_scale_5_squeeze_mask_0"), val = tensor([true, false])]; tensor prev_scale_5 = slice_by_index(begin = prev_scale_5_begin_0, end = prev_scale_5_end_0, end_mask = prev_scale_5_end_mask_0, squeeze_mask = prev_scale_5_squeeze_mask_0, x = enc_scale)[name = tensor("prev_scale_5")]; - tensor var_542 = linear(bias = encoder_q_proj_2_bias, weight = encoder_q_proj_2_weight, x = x_13)[name = tensor("linear_21")]; - tensor var_543 = const()[name = tensor("op_543"), val = tensor([1, 1, 4, 64])]; - tensor var_544 = reshape(shape = var_543, x = var_542)[name = tensor("op_544")]; + tensor var_568 = linear(bias = inner_encoder_q_proj_2_bias, weight = inner_encoder_q_proj_2_weight, x = x_13)[name = tensor("linear_21")]; + tensor var_569 = const()[name = tensor("op_569"), val = tensor([1, 1, 4, 64])]; + tensor var_570 = reshape(shape = var_569, x = var_568)[name = tensor("op_570")]; tensor q_5_perm_0 = const()[name = tensor("q_5_perm_0"), val = tensor([0, 2, 1, 3])]; - tensor var_548 = linear(bias = encoder_k_proj_2_bias, weight = encoder_k_proj_2_weight, x = x_13)[name = tensor("linear_22")]; - tensor var_549 = const()[name = tensor("op_549"), val = tensor(0x1p-3)]; - tensor var_550 = mul(x = var_548, y = var_549)[name = tensor("op_550")]; - tensor var_551 = const()[name = tensor("op_551"), val = tensor([1, 1, 4, 64])]; - tensor var_552 = reshape(shape = var_551, x = var_550)[name = tensor("op_552")]; + tensor var_574 = linear(bias = inner_encoder_k_proj_2_bias, weight = inner_encoder_k_proj_2_weight, x = x_13)[name = tensor("linear_22")]; + tensor var_575 = const()[name = tensor("op_575"), val = tensor(0x1p-3)]; + tensor var_576 = mul(x = var_574, y = var_575)[name = tensor("op_576")]; + tensor var_577 = const()[name = tensor("op_577"), val = tensor([1, 1, 4, 64])]; + tensor var_578 = reshape(shape = var_577, x = var_576)[name = tensor("op_578")]; tensor k_5_perm_0 = const()[name = tensor("k_5_perm_0"), val = tensor([0, 2, 1, 3])]; - tensor var_556 = linear(bias = encoder_v_proj_2_bias, weight = encoder_v_proj_2_weight, x = x_13)[name = tensor("linear_23")]; - tensor var_557 = const()[name = tensor("op_557"), val = tensor([1, 1, 4, 64])]; - tensor var_558 = reshape(shape = var_557, x = var_556)[name = tensor("op_558")]; + tensor var_582 = linear(bias = inner_encoder_v_proj_2_bias, weight = inner_encoder_v_proj_2_weight, x = x_13)[name = tensor("linear_23")]; + tensor var_583 = const()[name = tensor("op_583"), val = tensor([1, 1, 4, 64])]; + tensor var_584 = reshape(shape = var_583, x = var_582)[name = tensor("op_584")]; tensor v_5_perm_0 = const()[name = tensor("v_5_perm_0"), val = tensor([0, 2, 1, 3])]; - tensor input_97 = linear(bias = encoder_g_proj_2_bias, weight = encoder_g_proj_2_weight, x = x_13)[name = tensor("linear_24")]; + tensor input_99 = linear(bias = inner_encoder_g_proj_2_bias, weight = inner_encoder_g_proj_2_weight, x = x_13)[name = tensor("linear_24")]; tensor sqrt_s0_5 = sqrt(x = prev_scale_5)[name = tensor("sqrt_s0_5")]; - tensor s_t_5 = add(x = prev_scale_5, y = encoder__t_index)[name = tensor("s_t_5")]; + tensor s_t_5 = add(x = prev_scale_5, y = inner_encoder__t_index)[name = tensor("s_t_5")]; tensor sqrt_s_t_5 = sqrt(x = s_t_5)[name = tensor("sqrt_s_t_5")]; tensor qk_5_transpose_x_1 = const()[name = tensor("qk_5_transpose_x_1"), val = tensor(false)]; tensor qk_5_transpose_y_1 = const()[name = tensor("qk_5_transpose_y_1"), val = tensor(true)]; - tensor k_5 = transpose(perm = k_5_perm_0, x = var_552)[name = tensor("transpose_43")]; - tensor q_5 = transpose(perm = q_5_perm_0, x = var_544)[name = tensor("transpose_44")]; + tensor k_5 = transpose(perm = k_5_perm_0, x = var_578)[name = tensor("transpose_43")]; + tensor q_5 = transpose(perm = q_5_perm_0, x = var_570)[name = tensor("transpose_44")]; tensor qk_5 = matmul(transpose_x = qk_5_transpose_x_1, transpose_y = qk_5_transpose_y_1, x = q_5, y = k_5)[name = tensor("qk_5")]; - tensor var_568 = const()[name = tensor("op_568"), val = tensor([1, 1])]; - tensor var_569 = reshape(shape = var_568, x = sqrt_s_t_5)[name = tensor("op_569")]; - tensor M_5 = real_div(x = encoder__causal_mask, y = var_569)[name = tensor("M_5")]; - tensor var_571 = mul(x = qk_5, y = M_5)[name = tensor("op_571")]; + tensor var_594 = const()[name = tensor("op_594"), val = tensor([1, 1])]; + tensor var_595 = reshape(shape = var_594, x = sqrt_s_t_5)[name = tensor("op_595")]; + tensor M_5 = real_div(x = inner_encoder__causal_mask, y = var_595)[name = tensor("M_5")]; + tensor var_597 = mul(x = qk_5, y = M_5)[name = tensor("op_597")]; tensor inner_5_transpose_x_0 = const()[name = tensor("inner_5_transpose_x_0"), val = tensor(false)]; tensor inner_5_transpose_y_0 = const()[name = tensor("inner_5_transpose_y_0"), val = tensor(false)]; - tensor v_5 = transpose(perm = v_5_perm_0, x = var_558)[name = tensor("transpose_42")]; - tensor inner_5 = matmul(transpose_x = inner_5_transpose_x_0, transpose_y = inner_5_transpose_y_0, x = var_571, y = v_5)[name = tensor("inner_5")]; - tensor var_573_transpose_x_0 = const()[name = tensor("op_573_transpose_x_0"), val = tensor(false)]; - tensor var_573_transpose_y_0 = const()[name = tensor("op_573_transpose_y_0"), val = tensor(false)]; - tensor var_573 = matmul(transpose_x = var_573_transpose_x_0, transpose_y = var_573_transpose_y_0, x = q_5, y = prev_kv_5)[name = tensor("op_573")]; - tensor var_574 = real_div(x = sqrt_s0_5, y = sqrt_s_t_5)[name = tensor("op_574")]; - tensor var_575 = const()[name = tensor("op_575"), val = tensor([1, 1, 1, 1])]; - tensor var_576 = reshape(shape = var_575, x = var_574)[name = tensor("op_576")]; - tensor cross_5 = mul(x = var_573, y = var_576)[name = tensor("cross_5")]; + tensor v_5 = transpose(perm = v_5_perm_0, x = var_584)[name = tensor("transpose_42")]; + tensor inner_5 = matmul(transpose_x = inner_5_transpose_x_0, transpose_y = inner_5_transpose_y_0, x = var_597, y = v_5)[name = tensor("inner_5")]; + tensor var_599_transpose_x_0 = const()[name = tensor("op_599_transpose_x_0"), val = tensor(false)]; + tensor var_599_transpose_y_0 = const()[name = tensor("op_599_transpose_y_0"), val = tensor(false)]; + tensor var_599 = matmul(transpose_x = var_599_transpose_x_0, transpose_y = var_599_transpose_y_0, x = q_5, y = prev_kv_5)[name = tensor("op_599")]; + tensor var_600 = real_div(x = sqrt_s0_5, y = sqrt_s_t_5)[name = tensor("op_600")]; + tensor var_601 = const()[name = tensor("op_601"), val = tensor([1, 1, 1, 1])]; + tensor var_602 = reshape(shape = var_601, x = var_600)[name = tensor("op_602")]; + tensor cross_5 = mul(x = var_599, y = var_602)[name = tensor("cross_5")]; tensor out_13 = add(x = inner_5, y = cross_5)[name = tensor("out_13")]; - tensor var_579 = mul(x = prev_kv_5, y = sqrt_s0_5)[name = tensor("op_579")]; - tensor var_581_transpose_x_1 = const()[name = tensor("op_581_transpose_x_1"), val = tensor(true)]; - tensor var_581_transpose_y_1 = const()[name = tensor("op_581_transpose_y_1"), val = tensor(false)]; - tensor var_581 = matmul(transpose_x = var_581_transpose_x_1, transpose_y = var_581_transpose_y_1, x = k_5, y = v_5)[name = tensor("op_581")]; - tensor new_kv_unnorm_5 = add(x = var_579, y = var_581)[name = tensor("new_kv_unnorm_5")]; - tensor var_583 = const()[name = tensor("op_583"), val = tensor(0x1p+0)]; - tensor new_scale_5 = add(x = prev_scale_5, y = var_583)[name = tensor("new_scale_5")]; - tensor var_585 = sqrt(x = new_scale_5)[name = tensor("op_585")]; - tensor var_586 = real_div(x = new_kv_unnorm_5, y = var_585)[name = tensor("op_586")]; - tensor var_587_perm_0 = const()[name = tensor("op_587_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_605 = mul(x = prev_kv_5, y = sqrt_s0_5)[name = tensor("op_605")]; + tensor var_607_transpose_x_1 = const()[name = tensor("op_607_transpose_x_1"), val = tensor(true)]; + tensor var_607_transpose_y_1 = const()[name = tensor("op_607_transpose_y_1"), val = tensor(false)]; + tensor var_607 = matmul(transpose_x = var_607_transpose_x_1, transpose_y = var_607_transpose_y_1, x = k_5, y = v_5)[name = tensor("op_607")]; + tensor new_kv_unnorm_5 = add(x = var_605, y = var_607)[name = tensor("new_kv_unnorm_5")]; + tensor var_609 = const()[name = tensor("op_609"), val = tensor(0x1p+0)]; + tensor new_scale_5 = add(x = prev_scale_5, y = var_609)[name = tensor("new_scale_5")]; + tensor var_611 = sqrt(x = new_scale_5)[name = tensor("op_611")]; + tensor var_612 = real_div(x = new_kv_unnorm_5, y = var_611)[name = tensor("op_612")]; + tensor var_613_perm_0 = const()[name = tensor("op_613_perm_0"), val = tensor([0, 2, 1, 3])]; tensor out_15_axes_0 = const()[name = tensor("out_15_axes_0"), val = tensor([-1])]; - tensor var_587 = transpose(perm = var_587_perm_0, x = out_13)[name = tensor("transpose_41")]; - tensor out_15 = layer_norm(axes = out_15_axes_0, epsilon = var_17, x = var_587)[name = tensor("out_15")]; - tensor var_591 = const()[name = tensor("op_591"), val = tensor([1, 1, 256])]; - tensor out_17 = reshape(shape = var_591, x = out_15)[name = tensor("out_17")]; - tensor var_593 = silu(x = input_97)[name = tensor("op_593")]; - tensor input_99 = mul(x = var_593, y = out_17)[name = tensor("input_99")]; - tensor ret_out_5 = linear(bias = encoder_out_proj_2_bias, weight = encoder_out_proj_2_weight, x = input_99)[name = tensor("linear_25")]; - tensor x_15 = add(x = input_93, y = ret_out_5)[name = tensor("x_15")]; + tensor var_613 = transpose(perm = var_613_perm_0, x = out_13)[name = tensor("transpose_41")]; + tensor out_15 = layer_norm(axes = out_15_axes_0, epsilon = var_43, x = var_613)[name = tensor("out_15")]; + tensor var_617 = const()[name = tensor("op_617"), val = tensor([1, 1, 256])]; + tensor out_17 = reshape(shape = var_617, x = out_15)[name = tensor("out_17")]; + tensor var_619 = silu(x = input_99)[name = tensor("op_619")]; + tensor input_101 = mul(x = var_619, y = out_17)[name = tensor("input_101")]; + tensor ret_out_5 = linear(bias = inner_encoder_out_proj_2_bias, weight = inner_encoder_out_proj_2_weight, x = input_101)[name = tensor("linear_25")]; + tensor x_15 = add(x = input_95, y = ret_out_5)[name = tensor("x_15")]; tensor window_9_begin_0 = const()[name = tensor("window_9_begin_0"), val = tensor([2, 0, 0, 0])]; tensor window_9_end_0 = const()[name = tensor("window_9_end_0"), val = tensor([3, 1, 16, 256])]; tensor window_9_end_mask_0 = const()[name = tensor("window_9_end_mask_0"), val = tensor([false, true, true, true])]; tensor window_9_squeeze_mask_0 = const()[name = tensor("window_9_squeeze_mask_0"), val = tensor([true, false, false, false])]; tensor window_9 = slice_by_index(begin = window_9_begin_0, end = window_9_end_0, end_mask = window_9_end_mask_0, squeeze_mask = window_9_squeeze_mask_0, x = enc_conv_cache)[name = tensor("window_9")]; - tensor var_604_begin_0 = const()[name = tensor("op_604_begin_0"), val = tensor([0, 1, 0])]; - tensor var_604_end_0 = const()[name = tensor("op_604_end_0"), val = tensor([1, 16, 256])]; - tensor var_604_end_mask_0 = const()[name = tensor("op_604_end_mask_0"), val = tensor([true, true, true])]; - tensor var_604 = slice_by_index(begin = var_604_begin_0, end = var_604_end_0, end_mask = var_604_end_mask_0, x = window_9)[name = tensor("op_604")]; + tensor var_630_begin_0 = const()[name = tensor("op_630_begin_0"), val = tensor([0, 1, 0])]; + tensor var_630_end_0 = const()[name = tensor("op_630_end_0"), val = tensor([1, 16, 256])]; + tensor var_630_end_mask_0 = const()[name = tensor("op_630_end_mask_0"), val = tensor([true, true, true])]; + tensor var_630 = slice_by_index(begin = var_630_begin_0, end = var_630_end_0, end_mask = var_630_end_mask_0, x = window_9)[name = tensor("op_630")]; tensor window_11_interleave_0 = const()[name = tensor("window_11_interleave_0"), val = tensor(false)]; - tensor window_11 = concat(axis = var_26, interleave = window_11_interleave_0, values = (var_604, x_15))[name = tensor("window_11")]; - tensor input_101_interleave_0 = const()[name = tensor("input_101_interleave_0"), val = tensor(false)]; - tensor input_101 = concat(axis = var_23, interleave = input_101_interleave_0, values = window_11)[name = tensor("input_101")]; + tensor window_11 = concat(axis = var_52, interleave = window_11_interleave_0, values = (var_630, x_15))[name = tensor("window_11")]; + tensor input_103_interleave_0 = const()[name = tensor("input_103_interleave_0"), val = tensor(false)]; + tensor input_103 = concat(axis = var_38, interleave = input_103_interleave_0, values = window_11)[name = tensor("input_103")]; tensor x_17_axes_0 = const()[name = tensor("x_17_axes_0"), val = tensor([-1])]; - tensor x_17 = layer_norm(axes = x_17_axes_0, beta = encoder_conv_module_2_sequential_0_bias, epsilon = var_28, gamma = encoder_conv_module_2_sequential_0_weight, x = input_101)[name = tensor("x_17")]; - tensor input_103_perm_0 = const()[name = tensor("input_103_perm_0"), val = tensor([0, 2, 1])]; + tensor x_17 = layer_norm(axes = x_17_axes_0, beta = inner_encoder_conv_module_2_sequential_0_bias, epsilon = var_35, gamma = inner_encoder_conv_module_2_sequential_0_weight, x = input_103)[name = tensor("x_17")]; + tensor input_105_perm_0 = const()[name = tensor("input_105_perm_0"), val = tensor([0, 2, 1])]; tensor inputs_23_pad_type_0 = const()[name = tensor("inputs_23_pad_type_0"), val = tensor("valid")]; tensor inputs_23_strides_0 = const()[name = tensor("inputs_23_strides_0"), val = tensor([1])]; tensor inputs_23_pad_0 = const()[name = tensor("inputs_23_pad_0"), val = tensor([0, 0])]; tensor inputs_23_dilations_0 = const()[name = tensor("inputs_23_dilations_0"), val = tensor([1])]; tensor inputs_23_groups_0 = const()[name = tensor("inputs_23_groups_0"), val = tensor(1)]; - tensor input_103 = transpose(perm = input_103_perm_0, x = x_17)[name = tensor("transpose_40")]; - tensor inputs_23 = conv(bias = encoder_conv_module_2_sequential_2_conv_bias, dilations = inputs_23_dilations_0, groups = inputs_23_groups_0, pad = inputs_23_pad_0, pad_type = inputs_23_pad_type_0, strides = inputs_23_strides_0, weight = encoder_conv_module_2_sequential_2_conv_weight, x = input_103)[name = tensor("inputs_23")]; - tensor var_629_split_sizes_0 = const()[name = tensor("op_629_split_sizes_0"), val = tensor([256, 256])]; - tensor var_629_axis_0 = const()[name = tensor("op_629_axis_0"), val = tensor(1)]; - tensor var_629_0, tensor var_629_1 = split(axis = var_629_axis_0, split_sizes = var_629_split_sizes_0, x = inputs_23)[name = tensor("op_629")]; - tensor var_631 = sigmoid(x = var_629_1)[name = tensor("op_631")]; - tensor inputs_25 = mul(x = var_629_0, y = var_631)[name = tensor("inputs_25")]; + tensor input_105 = transpose(perm = input_105_perm_0, x = x_17)[name = tensor("transpose_40")]; + tensor inputs_23 = conv(bias = inner_encoder_conv_module_2_sequential_2_conv_bias, dilations = inputs_23_dilations_0, groups = inputs_23_groups_0, pad = inputs_23_pad_0, pad_type = inputs_23_pad_type_0, strides = inputs_23_strides_0, weight = inner_encoder_conv_module_2_sequential_2_conv_weight, x = input_105)[name = tensor("inputs_23")]; + tensor var_655_split_sizes_0 = const()[name = tensor("op_655_split_sizes_0"), val = tensor([256, 256])]; + tensor var_655_axis_0 = const()[name = tensor("op_655_axis_0"), val = tensor(1)]; + tensor var_655_0, tensor var_655_1 = split(axis = var_655_axis_0, split_sizes = var_655_split_sizes_0, x = inputs_23)[name = tensor("op_655")]; + tensor var_657 = sigmoid(x = var_655_1)[name = tensor("op_657")]; + tensor inputs_25 = mul(x = var_655_0, y = var_657)[name = tensor("inputs_25")]; tensor outputs_aug_5_pad_type_0 = const()[name = tensor("outputs_aug_5_pad_type_0"), val = tensor("custom")]; tensor outputs_aug_5_pad_0 = const()[name = tensor("outputs_aug_5_pad_0"), val = tensor([15, 15])]; tensor outputs_aug_5_groups_0 = const()[name = tensor("outputs_aug_5_groups_0"), val = tensor(256)]; tensor outputs_aug_5_strides_0 = const()[name = tensor("outputs_aug_5_strides_0"), val = tensor([1])]; tensor outputs_aug_5_dilations_0 = const()[name = tensor("outputs_aug_5_dilations_0"), val = tensor([1])]; - tensor outputs_aug_5 = conv(dilations = outputs_aug_5_dilations_0, groups = outputs_aug_5_groups_0, pad = outputs_aug_5_pad_0, pad_type = outputs_aug_5_pad_type_0, strides = outputs_aug_5_strides_0, weight = encoder_conv_module_2_sequential_4_conv_weight, x = inputs_25)[name = tensor("outputs_aug_5")]; - tensor input_105_begin_0 = const()[name = tensor("input_105_begin_0"), val = tensor([0, 0, 0])]; - tensor input_105_end_0 = const()[name = tensor("input_105_end_0"), val = tensor([1, 256, 16])]; - tensor input_105_end_mask_0 = const()[name = tensor("input_105_end_mask_0"), val = tensor([true, true, false])]; - tensor input_105 = slice_by_index(begin = input_105_begin_0, end = input_105_end_0, end_mask = input_105_end_mask_0, x = outputs_aug_5)[name = tensor("input_105")]; - tensor inputs_27 = batch_norm(beta = encoder_conv_module_2_sequential_5_bias, epsilon = var_28, gamma = encoder_conv_module_2_sequential_5_weight, mean = encoder_conv_module_2_sequential_5_running_mean, variance = encoder_conv_module_2_sequential_5_running_var, x = input_105)[name = tensor("inputs_27")]; - tensor input_107 = silu(x = inputs_27)[name = tensor("input_107")]; - tensor input_109_pad_type_0 = const()[name = tensor("input_109_pad_type_0"), val = tensor("valid")]; - tensor input_109_strides_0 = const()[name = tensor("input_109_strides_0"), val = tensor([1])]; - tensor input_109_pad_0 = const()[name = tensor("input_109_pad_0"), val = tensor([0, 0])]; - tensor input_109_dilations_0 = const()[name = tensor("input_109_dilations_0"), val = tensor([1])]; - tensor input_109_groups_0 = const()[name = tensor("input_109_groups_0"), val = tensor(1)]; - tensor input_109 = conv(bias = encoder_conv_module_2_sequential_7_conv_bias, dilations = input_109_dilations_0, groups = input_109_groups_0, pad = input_109_pad_0, pad_type = input_109_pad_type_0, strides = input_109_strides_0, weight = encoder_conv_module_2_sequential_7_conv_weight, x = input_107)[name = tensor("input_109")]; + tensor outputs_aug_5 = conv(dilations = outputs_aug_5_dilations_0, groups = outputs_aug_5_groups_0, pad = outputs_aug_5_pad_0, pad_type = outputs_aug_5_pad_type_0, strides = outputs_aug_5_strides_0, weight = inner_encoder_conv_module_2_sequential_4_conv_weight, x = inputs_25)[name = tensor("outputs_aug_5")]; + tensor input_107_begin_0 = const()[name = tensor("input_107_begin_0"), val = tensor([0, 0, 0])]; + tensor input_107_end_0 = const()[name = tensor("input_107_end_0"), val = tensor([1, 256, 16])]; + tensor input_107_end_mask_0 = const()[name = tensor("input_107_end_mask_0"), val = tensor([true, true, false])]; + tensor input_107 = slice_by_index(begin = input_107_begin_0, end = input_107_end_0, end_mask = input_107_end_mask_0, x = outputs_aug_5)[name = tensor("input_107")]; + tensor inputs_27 = batch_norm(beta = inner_encoder_conv_module_2_sequential_5_bias, epsilon = var_35, gamma = inner_encoder_conv_module_2_sequential_5_weight, mean = inner_encoder_conv_module_2_sequential_5_running_mean, variance = inner_encoder_conv_module_2_sequential_5_running_var, x = input_107)[name = tensor("inputs_27")]; + tensor input_109 = silu(x = inputs_27)[name = tensor("input_109")]; + tensor input_111_pad_type_0 = const()[name = tensor("input_111_pad_type_0"), val = tensor("valid")]; + tensor input_111_strides_0 = const()[name = tensor("input_111_strides_0"), val = tensor([1])]; + tensor input_111_pad_0 = const()[name = tensor("input_111_pad_0"), val = tensor([0, 0])]; + tensor input_111_dilations_0 = const()[name = tensor("input_111_dilations_0"), val = tensor([1])]; + tensor input_111_groups_0 = const()[name = tensor("input_111_groups_0"), val = tensor(1)]; + tensor input_111 = conv(bias = inner_encoder_conv_module_2_sequential_7_conv_bias, dilations = input_111_dilations_0, groups = input_111_groups_0, pad = input_111_pad_0, pad_type = input_111_pad_type_0, strides = input_111_strides_0, weight = inner_encoder_conv_module_2_sequential_7_conv_weight, x = input_109)[name = tensor("input_111")]; tensor conv_out_5_perm_0 = const()[name = tensor("conv_out_5_perm_0"), val = tensor([0, 2, 1])]; - tensor var_662_begin_0 = const()[name = tensor("op_662_begin_0"), val = tensor([0, -1, 0])]; - tensor var_662_end_0 = const()[name = tensor("op_662_end_0"), val = tensor([1, 16, 256])]; - tensor var_662_end_mask_0 = const()[name = tensor("op_662_end_mask_0"), val = tensor([true, true, true])]; - tensor conv_out_5 = transpose(perm = conv_out_5_perm_0, x = input_109)[name = tensor("transpose_39")]; - tensor var_662 = slice_by_index(begin = var_662_begin_0, end = var_662_end_0, end_mask = var_662_end_mask_0, x = conv_out_5)[name = tensor("op_662")]; - tensor var_664_perm_0 = const()[name = tensor("op_664_perm_0"), val = tensor([1, 0, 2])]; - tensor var_664 = transpose(perm = var_664_perm_0, x = var_662)[name = tensor("transpose_38")]; - tensor input_111 = add(x = x_15, y = var_664)[name = tensor("input_111")]; - tensor input_113_axes_0 = const()[name = tensor("input_113_axes_0"), val = tensor([-1])]; - tensor input_113 = layer_norm(axes = input_113_axes_0, beta = encoder_ffn2_2_module_sequential_0_bias, epsilon = var_28, gamma = encoder_ffn2_2_module_sequential_0_weight, x = input_111)[name = tensor("input_113")]; - tensor inputs_29 = linear(bias = encoder_ffn2_2_module_sequential_1_linear_bias, weight = encoder_ffn2_2_module_sequential_1_linear_weight, x = input_113)[name = tensor("linear_26")]; - tensor input_115 = silu(x = inputs_29)[name = tensor("input_115")]; - tensor input_119 = linear(bias = encoder_ffn2_2_module_sequential_4_linear_bias, weight = encoder_ffn2_2_module_sequential_4_linear_weight, x = input_115)[name = tensor("linear_27")]; - tensor var_687 = const()[name = tensor("op_687"), val = tensor(0x1p-1)]; - tensor var_688 = mul(x = input_119, y = var_687)[name = tensor("op_688")]; - tensor input_121 = add(x = var_688, y = input_111)[name = tensor("input_121")]; - tensor input_123_axes_0 = const()[name = tensor("input_123_axes_0"), val = tensor([-1])]; - tensor input_123 = layer_norm(axes = input_123_axes_0, beta = encoder_layer_norm_2_bias, epsilon = var_28, gamma = encoder_layer_norm_2_weight, x = input_121)[name = tensor("input_123")]; + tensor var_688_begin_0 = const()[name = tensor("op_688_begin_0"), val = tensor([0, -1, 0])]; + tensor var_688_end_0 = const()[name = tensor("op_688_end_0"), val = tensor([1, 16, 256])]; + tensor var_688_end_mask_0 = const()[name = tensor("op_688_end_mask_0"), val = tensor([true, true, true])]; + tensor conv_out_5 = transpose(perm = conv_out_5_perm_0, x = input_111)[name = tensor("transpose_39")]; + tensor var_688 = slice_by_index(begin = var_688_begin_0, end = var_688_end_0, end_mask = var_688_end_mask_0, x = conv_out_5)[name = tensor("op_688")]; + tensor var_690_perm_0 = const()[name = tensor("op_690_perm_0"), val = tensor([1, 0, 2])]; + tensor var_690 = transpose(perm = var_690_perm_0, x = var_688)[name = tensor("transpose_38")]; + tensor input_113 = add(x = x_15, y = var_690)[name = tensor("input_113")]; + tensor input_115_axes_0 = const()[name = tensor("input_115_axes_0"), val = tensor([-1])]; + tensor input_115 = layer_norm(axes = input_115_axes_0, beta = inner_encoder_ffn2_2_module_sequential_0_bias, epsilon = var_35, gamma = inner_encoder_ffn2_2_module_sequential_0_weight, x = input_113)[name = tensor("input_115")]; + tensor inputs_29 = linear(bias = inner_encoder_ffn2_2_module_sequential_1_linear_bias, weight = inner_encoder_ffn2_2_module_sequential_1_linear_weight, x = input_115)[name = tensor("linear_26")]; + tensor input_117 = silu(x = inputs_29)[name = tensor("input_117")]; + tensor input_121 = linear(bias = inner_encoder_ffn2_2_module_sequential_4_linear_bias, weight = inner_encoder_ffn2_2_module_sequential_4_linear_weight, x = input_117)[name = tensor("linear_27")]; + tensor var_713 = const()[name = tensor("op_713"), val = tensor(0x1p-1)]; + tensor var_714 = mul(x = input_121, y = var_713)[name = tensor("op_714")]; + tensor input_123 = add(x = var_714, y = input_113)[name = tensor("input_123")]; tensor input_125_axes_0 = const()[name = tensor("input_125_axes_0"), val = tensor([-1])]; - tensor input_125 = layer_norm(axes = input_125_axes_0, beta = encoder_ffn1_3_module_sequential_0_bias, epsilon = var_28, gamma = encoder_ffn1_3_module_sequential_0_weight, x = input_123)[name = tensor("input_125")]; - tensor inputs_31 = linear(bias = encoder_ffn1_3_module_sequential_1_linear_bias, weight = encoder_ffn1_3_module_sequential_1_linear_weight, x = input_125)[name = tensor("linear_28")]; - tensor input_127 = silu(x = inputs_31)[name = tensor("input_127")]; - tensor input_131 = linear(bias = encoder_ffn1_3_module_sequential_4_linear_bias, weight = encoder_ffn1_3_module_sequential_4_linear_weight, x = input_127)[name = tensor("linear_29")]; - tensor var_717 = const()[name = tensor("op_717"), val = tensor(0x1p-1)]; - tensor var_718 = mul(x = input_131, y = var_717)[name = tensor("op_718")]; - tensor input_133 = add(x = var_718, y = input_123)[name = tensor("input_133")]; + tensor input_125 = layer_norm(axes = input_125_axes_0, beta = inner_encoder_layer_norm_2_bias, epsilon = var_35, gamma = inner_encoder_layer_norm_2_weight, x = input_123)[name = tensor("input_125")]; + tensor input_127_axes_0 = const()[name = tensor("input_127_axes_0"), val = tensor([-1])]; + tensor input_127 = layer_norm(axes = input_127_axes_0, beta = inner_encoder_ffn1_3_module_sequential_0_bias, epsilon = var_35, gamma = inner_encoder_ffn1_3_module_sequential_0_weight, x = input_125)[name = tensor("input_127")]; + tensor inputs_31 = linear(bias = inner_encoder_ffn1_3_module_sequential_1_linear_bias, weight = inner_encoder_ffn1_3_module_sequential_1_linear_weight, x = input_127)[name = tensor("linear_28")]; + tensor input_129 = silu(x = inputs_31)[name = tensor("input_129")]; + tensor input_133 = linear(bias = inner_encoder_ffn1_3_module_sequential_4_linear_bias, weight = inner_encoder_ffn1_3_module_sequential_4_linear_weight, x = input_129)[name = tensor("linear_29")]; + tensor var_743 = const()[name = tensor("op_743"), val = tensor(0x1p-1)]; + tensor var_744 = mul(x = input_133, y = var_743)[name = tensor("op_744")]; + tensor input_135 = add(x = var_744, y = input_125)[name = tensor("input_135")]; tensor x_19_axes_0 = const()[name = tensor("x_19_axes_0"), val = tensor([-1])]; - tensor x_19 = layer_norm(axes = x_19_axes_0, beta = encoder_ret_lns_3_bias, epsilon = var_28, gamma = encoder_ret_lns_3_weight, x = input_133)[name = tensor("x_19")]; + tensor x_19 = layer_norm(axes = x_19_axes_0, beta = inner_encoder_ret_lns_3_bias, epsilon = var_35, gamma = inner_encoder_ret_lns_3_weight, x = input_135)[name = tensor("x_19")]; tensor prev_kv_7_begin_0 = const()[name = tensor("prev_kv_7_begin_0"), val = tensor([3, 0, 0, 0, 0])]; tensor prev_kv_7_end_0 = const()[name = tensor("prev_kv_7_end_0"), val = tensor([4, 1, 4, 64, 64])]; tensor prev_kv_7_end_mask_0 = const()[name = tensor("prev_kv_7_end_mask_0"), val = tensor([false, true, true, true, true])]; @@ -667,175 +673,168 @@ program(1.0) tensor prev_scale_7_end_mask_0 = const()[name = tensor("prev_scale_7_end_mask_0"), val = tensor([false, true])]; tensor prev_scale_7_squeeze_mask_0 = const()[name = tensor("prev_scale_7_squeeze_mask_0"), val = tensor([true, false])]; tensor prev_scale_7 = slice_by_index(begin = prev_scale_7_begin_0, end = prev_scale_7_end_0, end_mask = prev_scale_7_end_mask_0, squeeze_mask = prev_scale_7_squeeze_mask_0, x = enc_scale)[name = tensor("prev_scale_7")]; - tensor var_732 = linear(bias = encoder_q_proj_3_bias, weight = encoder_q_proj_3_weight, x = x_19)[name = tensor("linear_30")]; - tensor var_733 = const()[name = tensor("op_733"), val = tensor([1, 1, 4, 64])]; - tensor var_734 = reshape(shape = var_733, x = var_732)[name = tensor("op_734")]; + tensor var_758 = linear(bias = inner_encoder_q_proj_3_bias, weight = inner_encoder_q_proj_3_weight, x = x_19)[name = tensor("linear_30")]; + tensor var_759 = const()[name = tensor("op_759"), val = tensor([1, 1, 4, 64])]; + tensor var_760 = reshape(shape = var_759, x = var_758)[name = tensor("op_760")]; tensor q_7_perm_0 = const()[name = tensor("q_7_perm_0"), val = tensor([0, 2, 1, 3])]; - tensor var_738 = linear(bias = encoder_k_proj_3_bias, weight = encoder_k_proj_3_weight, x = x_19)[name = tensor("linear_31")]; - tensor var_739 = const()[name = tensor("op_739"), val = tensor(0x1p-3)]; - tensor var_740 = mul(x = var_738, y = var_739)[name = tensor("op_740")]; - tensor var_741 = const()[name = tensor("op_741"), val = tensor([1, 1, 4, 64])]; - tensor var_742 = reshape(shape = var_741, x = var_740)[name = tensor("op_742")]; + tensor var_764 = linear(bias = inner_encoder_k_proj_3_bias, weight = inner_encoder_k_proj_3_weight, x = x_19)[name = tensor("linear_31")]; + tensor var_765 = const()[name = tensor("op_765"), val = tensor(0x1p-3)]; + tensor var_766 = mul(x = var_764, y = var_765)[name = tensor("op_766")]; + tensor var_767 = const()[name = tensor("op_767"), val = tensor([1, 1, 4, 64])]; + tensor var_768 = reshape(shape = var_767, x = var_766)[name = tensor("op_768")]; tensor k_7_perm_0 = const()[name = tensor("k_7_perm_0"), val = tensor([0, 2, 1, 3])]; - tensor var_746 = linear(bias = encoder_v_proj_3_bias, weight = encoder_v_proj_3_weight, x = x_19)[name = tensor("linear_32")]; - tensor var_747 = const()[name = tensor("op_747"), val = tensor([1, 1, 4, 64])]; - tensor var_748 = reshape(shape = var_747, x = var_746)[name = tensor("op_748")]; + tensor var_772 = linear(bias = inner_encoder_v_proj_3_bias, weight = inner_encoder_v_proj_3_weight, x = x_19)[name = tensor("linear_32")]; + tensor var_773 = const()[name = tensor("op_773"), val = tensor([1, 1, 4, 64])]; + tensor var_774 = reshape(shape = var_773, x = var_772)[name = tensor("op_774")]; tensor v_7_perm_0 = const()[name = tensor("v_7_perm_0"), val = tensor([0, 2, 1, 3])]; - tensor input_137 = linear(bias = encoder_g_proj_3_bias, weight = encoder_g_proj_3_weight, x = x_19)[name = tensor("linear_33")]; + tensor input_139 = linear(bias = inner_encoder_g_proj_3_bias, weight = inner_encoder_g_proj_3_weight, x = x_19)[name = tensor("linear_33")]; tensor sqrt_s0_7 = sqrt(x = prev_scale_7)[name = tensor("sqrt_s0_7")]; - tensor s_t_7 = add(x = prev_scale_7, y = encoder__t_index)[name = tensor("s_t_7")]; + tensor s_t_7 = add(x = prev_scale_7, y = inner_encoder__t_index)[name = tensor("s_t_7")]; tensor sqrt_s_t_7 = sqrt(x = s_t_7)[name = tensor("sqrt_s_t_7")]; tensor qk_7_transpose_x_1 = const()[name = tensor("qk_7_transpose_x_1"), val = tensor(false)]; tensor qk_7_transpose_y_1 = const()[name = tensor("qk_7_transpose_y_1"), val = tensor(true)]; - tensor k_7 = transpose(perm = k_7_perm_0, x = var_742)[name = tensor("transpose_36")]; - tensor q_7 = transpose(perm = q_7_perm_0, x = var_734)[name = tensor("transpose_37")]; + tensor k_7 = transpose(perm = k_7_perm_0, x = var_768)[name = tensor("transpose_36")]; + tensor q_7 = transpose(perm = q_7_perm_0, x = var_760)[name = tensor("transpose_37")]; tensor qk_7 = matmul(transpose_x = qk_7_transpose_x_1, transpose_y = qk_7_transpose_y_1, x = q_7, y = k_7)[name = tensor("qk_7")]; - tensor var_758 = const()[name = tensor("op_758"), val = tensor([1, 1])]; - tensor var_759 = reshape(shape = var_758, x = sqrt_s_t_7)[name = tensor("op_759")]; - tensor M_7 = real_div(x = encoder__causal_mask, y = var_759)[name = tensor("M_7")]; - tensor var_761 = mul(x = qk_7, y = M_7)[name = tensor("op_761")]; + tensor var_784 = const()[name = tensor("op_784"), val = tensor([1, 1])]; + tensor var_785 = reshape(shape = var_784, x = sqrt_s_t_7)[name = tensor("op_785")]; + tensor M_7 = real_div(x = inner_encoder__causal_mask, y = var_785)[name = tensor("M_7")]; + tensor var_787 = mul(x = qk_7, y = M_7)[name = tensor("op_787")]; tensor inner_7_transpose_x_0 = const()[name = tensor("inner_7_transpose_x_0"), val = tensor(false)]; tensor inner_7_transpose_y_0 = const()[name = tensor("inner_7_transpose_y_0"), val = tensor(false)]; - tensor v_7 = transpose(perm = v_7_perm_0, x = var_748)[name = tensor("transpose_35")]; - tensor inner_7 = matmul(transpose_x = inner_7_transpose_x_0, transpose_y = inner_7_transpose_y_0, x = var_761, y = v_7)[name = tensor("inner_7")]; - tensor var_763_transpose_x_0 = const()[name = tensor("op_763_transpose_x_0"), val = tensor(false)]; - tensor var_763_transpose_y_0 = const()[name = tensor("op_763_transpose_y_0"), val = tensor(false)]; - tensor var_763 = matmul(transpose_x = var_763_transpose_x_0, transpose_y = var_763_transpose_y_0, x = q_7, y = prev_kv_7)[name = tensor("op_763")]; - tensor var_764 = real_div(x = sqrt_s0_7, y = sqrt_s_t_7)[name = tensor("op_764")]; - tensor var_765 = const()[name = tensor("op_765"), val = tensor([1, 1, 1, 1])]; - tensor var_766 = reshape(shape = var_765, x = var_764)[name = tensor("op_766")]; - tensor cross_7 = mul(x = var_763, y = var_766)[name = tensor("cross_7")]; + tensor v_7 = transpose(perm = v_7_perm_0, x = var_774)[name = tensor("transpose_35")]; + tensor inner_7 = matmul(transpose_x = inner_7_transpose_x_0, transpose_y = inner_7_transpose_y_0, x = var_787, y = v_7)[name = tensor("inner_7")]; + tensor var_789_transpose_x_0 = const()[name = tensor("op_789_transpose_x_0"), val = tensor(false)]; + tensor var_789_transpose_y_0 = const()[name = tensor("op_789_transpose_y_0"), val = tensor(false)]; + tensor var_789 = matmul(transpose_x = var_789_transpose_x_0, transpose_y = var_789_transpose_y_0, x = q_7, y = prev_kv_7)[name = tensor("op_789")]; + tensor var_790 = real_div(x = sqrt_s0_7, y = sqrt_s_t_7)[name = tensor("op_790")]; + tensor var_791 = const()[name = tensor("op_791"), val = tensor([1, 1, 1, 1])]; + tensor var_792 = reshape(shape = var_791, x = var_790)[name = tensor("op_792")]; + tensor cross_7 = mul(x = var_789, y = var_792)[name = tensor("cross_7")]; tensor out_19 = add(x = inner_7, y = cross_7)[name = tensor("out_19")]; - tensor var_769 = mul(x = prev_kv_7, y = sqrt_s0_7)[name = tensor("op_769")]; - tensor var_771_transpose_x_1 = const()[name = tensor("op_771_transpose_x_1"), val = tensor(true)]; - tensor var_771_transpose_y_1 = const()[name = tensor("op_771_transpose_y_1"), val = tensor(false)]; - tensor var_771 = matmul(transpose_x = var_771_transpose_x_1, transpose_y = var_771_transpose_y_1, x = k_7, y = v_7)[name = tensor("op_771")]; - tensor new_kv_unnorm_7 = add(x = var_769, y = var_771)[name = tensor("new_kv_unnorm_7")]; - tensor var_773 = const()[name = tensor("op_773"), val = tensor(0x1p+0)]; - tensor new_scale_7 = add(x = prev_scale_7, y = var_773)[name = tensor("new_scale_7")]; - tensor var_775 = sqrt(x = new_scale_7)[name = tensor("op_775")]; - tensor nkv_1 = real_div(x = new_kv_unnorm_7, y = var_775)[name = tensor("nkv_1")]; - tensor var_777_perm_0 = const()[name = tensor("op_777_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_795 = mul(x = prev_kv_7, y = sqrt_s0_7)[name = tensor("op_795")]; + tensor var_797_transpose_x_1 = const()[name = tensor("op_797_transpose_x_1"), val = tensor(true)]; + tensor var_797_transpose_y_1 = const()[name = tensor("op_797_transpose_y_1"), val = tensor(false)]; + tensor var_797 = matmul(transpose_x = var_797_transpose_x_1, transpose_y = var_797_transpose_y_1, x = k_7, y = v_7)[name = tensor("op_797")]; + tensor new_kv_unnorm_7 = add(x = var_795, y = var_797)[name = tensor("new_kv_unnorm_7")]; + tensor var_799 = const()[name = tensor("op_799"), val = tensor(0x1p+0)]; + tensor new_scale_7 = add(x = prev_scale_7, y = var_799)[name = tensor("new_scale_7")]; + tensor var_801 = sqrt(x = new_scale_7)[name = tensor("op_801")]; + tensor nkv_1 = real_div(x = new_kv_unnorm_7, y = var_801)[name = tensor("nkv_1")]; + tensor var_803_perm_0 = const()[name = tensor("op_803_perm_0"), val = tensor([0, 2, 1, 3])]; tensor out_21_axes_0 = const()[name = tensor("out_21_axes_0"), val = tensor([-1])]; - tensor var_777 = transpose(perm = var_777_perm_0, x = out_19)[name = tensor("transpose_34")]; - tensor out_21 = layer_norm(axes = out_21_axes_0, epsilon = var_17, x = var_777)[name = tensor("out_21")]; - tensor var_781 = const()[name = tensor("op_781"), val = tensor([1, 1, 256])]; - tensor out_23 = reshape(shape = var_781, x = out_21)[name = tensor("out_23")]; - tensor var_783 = silu(x = input_137)[name = tensor("op_783")]; - tensor input_139 = mul(x = var_783, y = out_23)[name = tensor("input_139")]; - tensor ret_out_7 = linear(bias = encoder_out_proj_3_bias, weight = encoder_out_proj_3_weight, x = input_139)[name = tensor("linear_34")]; - tensor x_21 = add(x = input_133, y = ret_out_7)[name = tensor("x_21")]; + tensor var_803 = transpose(perm = var_803_perm_0, x = out_19)[name = tensor("transpose_34")]; + tensor out_21 = layer_norm(axes = out_21_axes_0, epsilon = var_43, x = var_803)[name = tensor("out_21")]; + tensor var_807 = const()[name = tensor("op_807"), val = tensor([1, 1, 256])]; + tensor out_23 = reshape(shape = var_807, x = out_21)[name = tensor("out_23")]; + tensor var_809 = silu(x = input_139)[name = tensor("op_809")]; + tensor input_141 = mul(x = var_809, y = out_23)[name = tensor("input_141")]; + tensor ret_out_7 = linear(bias = inner_encoder_out_proj_3_bias, weight = inner_encoder_out_proj_3_weight, x = input_141)[name = tensor("linear_34")]; + tensor x_21 = add(x = input_135, y = ret_out_7)[name = tensor("x_21")]; tensor window_13_begin_0 = const()[name = tensor("window_13_begin_0"), val = tensor([3, 0, 0, 0])]; tensor window_13_end_0 = const()[name = tensor("window_13_end_0"), val = tensor([4, 1, 16, 256])]; tensor window_13_end_mask_0 = const()[name = tensor("window_13_end_mask_0"), val = tensor([false, true, true, true])]; tensor window_13_squeeze_mask_0 = const()[name = tensor("window_13_squeeze_mask_0"), val = tensor([true, false, false, false])]; tensor window_13 = slice_by_index(begin = window_13_begin_0, end = window_13_end_0, end_mask = window_13_end_mask_0, squeeze_mask = window_13_squeeze_mask_0, x = enc_conv_cache)[name = tensor("window_13")]; - tensor var_794_begin_0 = const()[name = tensor("op_794_begin_0"), val = tensor([0, 1, 0])]; - tensor var_794_end_0 = const()[name = tensor("op_794_end_0"), val = tensor([1, 16, 256])]; - tensor var_794_end_mask_0 = const()[name = tensor("op_794_end_mask_0"), val = tensor([true, true, true])]; - tensor var_794 = slice_by_index(begin = var_794_begin_0, end = var_794_end_0, end_mask = var_794_end_mask_0, x = window_13)[name = tensor("op_794")]; + tensor var_820_begin_0 = const()[name = tensor("op_820_begin_0"), val = tensor([0, 1, 0])]; + tensor var_820_end_0 = const()[name = tensor("op_820_end_0"), val = tensor([1, 16, 256])]; + tensor var_820_end_mask_0 = const()[name = tensor("op_820_end_mask_0"), val = tensor([true, true, true])]; + tensor var_820 = slice_by_index(begin = var_820_begin_0, end = var_820_end_0, end_mask = var_820_end_mask_0, x = window_13)[name = tensor("op_820")]; tensor window_interleave_0 = const()[name = tensor("window_interleave_0"), val = tensor(false)]; - tensor window = concat(axis = var_26, interleave = window_interleave_0, values = (var_794, x_21))[name = tensor("window")]; - tensor input_141_interleave_0 = const()[name = tensor("input_141_interleave_0"), val = tensor(false)]; - tensor input_141 = concat(axis = var_23, interleave = input_141_interleave_0, values = window)[name = tensor("input_141")]; + tensor window = concat(axis = var_52, interleave = window_interleave_0, values = (var_820, x_21))[name = tensor("window")]; + tensor input_143_interleave_0 = const()[name = tensor("input_143_interleave_0"), val = tensor(false)]; + tensor input_143 = concat(axis = var_38, interleave = input_143_interleave_0, values = window)[name = tensor("input_143")]; tensor x_23_axes_0 = const()[name = tensor("x_23_axes_0"), val = tensor([-1])]; - tensor x_23 = layer_norm(axes = x_23_axes_0, beta = encoder_conv_module_3_sequential_0_bias, epsilon = var_28, gamma = encoder_conv_module_3_sequential_0_weight, x = input_141)[name = tensor("x_23")]; - tensor input_143_perm_0 = const()[name = tensor("input_143_perm_0"), val = tensor([0, 2, 1])]; + tensor x_23 = layer_norm(axes = x_23_axes_0, beta = inner_encoder_conv_module_3_sequential_0_bias, epsilon = var_35, gamma = inner_encoder_conv_module_3_sequential_0_weight, x = input_143)[name = tensor("x_23")]; + tensor input_145_perm_0 = const()[name = tensor("input_145_perm_0"), val = tensor([0, 2, 1])]; tensor inputs_33_pad_type_0 = const()[name = tensor("inputs_33_pad_type_0"), val = tensor("valid")]; tensor inputs_33_strides_0 = const()[name = tensor("inputs_33_strides_0"), val = tensor([1])]; tensor inputs_33_pad_0 = const()[name = tensor("inputs_33_pad_0"), val = tensor([0, 0])]; tensor inputs_33_dilations_0 = const()[name = tensor("inputs_33_dilations_0"), val = tensor([1])]; tensor inputs_33_groups_0 = const()[name = tensor("inputs_33_groups_0"), val = tensor(1)]; - tensor input_143 = transpose(perm = input_143_perm_0, x = x_23)[name = tensor("transpose_33")]; - tensor inputs_33 = conv(bias = encoder_conv_module_3_sequential_2_conv_bias, dilations = inputs_33_dilations_0, groups = inputs_33_groups_0, pad = inputs_33_pad_0, pad_type = inputs_33_pad_type_0, strides = inputs_33_strides_0, weight = encoder_conv_module_3_sequential_2_conv_weight, x = input_143)[name = tensor("inputs_33")]; - tensor var_819_split_sizes_0 = const()[name = tensor("op_819_split_sizes_0"), val = tensor([256, 256])]; - tensor var_819_axis_0 = const()[name = tensor("op_819_axis_0"), val = tensor(1)]; - tensor var_819_0, tensor var_819_1 = split(axis = var_819_axis_0, split_sizes = var_819_split_sizes_0, x = inputs_33)[name = tensor("op_819")]; - tensor var_821 = sigmoid(x = var_819_1)[name = tensor("op_821")]; - tensor inputs_35 = mul(x = var_819_0, y = var_821)[name = tensor("inputs_35")]; + tensor input_145 = transpose(perm = input_145_perm_0, x = x_23)[name = tensor("transpose_33")]; + tensor inputs_33 = conv(bias = inner_encoder_conv_module_3_sequential_2_conv_bias, dilations = inputs_33_dilations_0, groups = inputs_33_groups_0, pad = inputs_33_pad_0, pad_type = inputs_33_pad_type_0, strides = inputs_33_strides_0, weight = inner_encoder_conv_module_3_sequential_2_conv_weight, x = input_145)[name = tensor("inputs_33")]; + tensor var_845_split_sizes_0 = const()[name = tensor("op_845_split_sizes_0"), val = tensor([256, 256])]; + tensor var_845_axis_0 = const()[name = tensor("op_845_axis_0"), val = tensor(1)]; + tensor var_845_0, tensor var_845_1 = split(axis = var_845_axis_0, split_sizes = var_845_split_sizes_0, x = inputs_33)[name = tensor("op_845")]; + tensor var_847 = sigmoid(x = var_845_1)[name = tensor("op_847")]; + tensor inputs_35 = mul(x = var_845_0, y = var_847)[name = tensor("inputs_35")]; tensor outputs_aug_pad_type_0 = const()[name = tensor("outputs_aug_pad_type_0"), val = tensor("custom")]; tensor outputs_aug_pad_0 = const()[name = tensor("outputs_aug_pad_0"), val = tensor([15, 15])]; tensor outputs_aug_groups_0 = const()[name = tensor("outputs_aug_groups_0"), val = tensor(256)]; tensor outputs_aug_strides_0 = const()[name = tensor("outputs_aug_strides_0"), val = tensor([1])]; tensor outputs_aug_dilations_0 = const()[name = tensor("outputs_aug_dilations_0"), val = tensor([1])]; - tensor outputs_aug = conv(dilations = outputs_aug_dilations_0, groups = outputs_aug_groups_0, pad = outputs_aug_pad_0, pad_type = outputs_aug_pad_type_0, strides = outputs_aug_strides_0, weight = encoder_conv_module_3_sequential_4_conv_weight, x = inputs_35)[name = tensor("outputs_aug")]; - tensor input_145_begin_0 = const()[name = tensor("input_145_begin_0"), val = tensor([0, 0, 0])]; - tensor input_145_end_0 = const()[name = tensor("input_145_end_0"), val = tensor([1, 256, 16])]; - tensor input_145_end_mask_0 = const()[name = tensor("input_145_end_mask_0"), val = tensor([true, true, false])]; - tensor input_145 = slice_by_index(begin = input_145_begin_0, end = input_145_end_0, end_mask = input_145_end_mask_0, x = outputs_aug)[name = tensor("input_145")]; - tensor inputs_37 = batch_norm(beta = encoder_conv_module_3_sequential_5_bias, epsilon = var_28, gamma = encoder_conv_module_3_sequential_5_weight, mean = encoder_conv_module_3_sequential_5_running_mean, variance = encoder_conv_module_3_sequential_5_running_var, x = input_145)[name = tensor("inputs_37")]; - tensor input_147 = silu(x = inputs_37)[name = tensor("input_147")]; - tensor input_149_pad_type_0 = const()[name = tensor("input_149_pad_type_0"), val = tensor("valid")]; - tensor input_149_strides_0 = const()[name = tensor("input_149_strides_0"), val = tensor([1])]; - tensor input_149_pad_0 = const()[name = tensor("input_149_pad_0"), val = tensor([0, 0])]; - tensor input_149_dilations_0 = const()[name = tensor("input_149_dilations_0"), val = tensor([1])]; - tensor input_149_groups_0 = const()[name = tensor("input_149_groups_0"), val = tensor(1)]; - tensor input_149 = conv(bias = encoder_conv_module_3_sequential_7_conv_bias, dilations = input_149_dilations_0, groups = input_149_groups_0, pad = input_149_pad_0, pad_type = input_149_pad_type_0, strides = input_149_strides_0, weight = encoder_conv_module_3_sequential_7_conv_weight, x = input_147)[name = tensor("input_149")]; + tensor outputs_aug = conv(dilations = outputs_aug_dilations_0, groups = outputs_aug_groups_0, pad = outputs_aug_pad_0, pad_type = outputs_aug_pad_type_0, strides = outputs_aug_strides_0, weight = inner_encoder_conv_module_3_sequential_4_conv_weight, x = inputs_35)[name = tensor("outputs_aug")]; + tensor input_147_begin_0 = const()[name = tensor("input_147_begin_0"), val = tensor([0, 0, 0])]; + tensor input_147_end_0 = const()[name = tensor("input_147_end_0"), val = tensor([1, 256, 16])]; + tensor input_147_end_mask_0 = const()[name = tensor("input_147_end_mask_0"), val = tensor([true, true, false])]; + tensor input_147 = slice_by_index(begin = input_147_begin_0, end = input_147_end_0, end_mask = input_147_end_mask_0, x = outputs_aug)[name = tensor("input_147")]; + tensor inputs_37 = batch_norm(beta = inner_encoder_conv_module_3_sequential_5_bias, epsilon = var_35, gamma = inner_encoder_conv_module_3_sequential_5_weight, mean = inner_encoder_conv_module_3_sequential_5_running_mean, variance = inner_encoder_conv_module_3_sequential_5_running_var, x = input_147)[name = tensor("inputs_37")]; + tensor input_149 = silu(x = inputs_37)[name = tensor("input_149")]; + tensor input_151_pad_type_0 = const()[name = tensor("input_151_pad_type_0"), val = tensor("valid")]; + tensor input_151_strides_0 = const()[name = tensor("input_151_strides_0"), val = tensor([1])]; + tensor input_151_pad_0 = const()[name = tensor("input_151_pad_0"), val = tensor([0, 0])]; + tensor input_151_dilations_0 = const()[name = tensor("input_151_dilations_0"), val = tensor([1])]; + tensor input_151_groups_0 = const()[name = tensor("input_151_groups_0"), val = tensor(1)]; + tensor input_151 = conv(bias = inner_encoder_conv_module_3_sequential_7_conv_bias, dilations = input_151_dilations_0, groups = input_151_groups_0, pad = input_151_pad_0, pad_type = input_151_pad_type_0, strides = input_151_strides_0, weight = inner_encoder_conv_module_3_sequential_7_conv_weight, x = input_149)[name = tensor("input_151")]; tensor conv_out_7_perm_0 = const()[name = tensor("conv_out_7_perm_0"), val = tensor([0, 2, 1])]; - tensor var_852_begin_0 = const()[name = tensor("op_852_begin_0"), val = tensor([0, -1, 0])]; - tensor var_852_end_0 = const()[name = tensor("op_852_end_0"), val = tensor([1, 16, 256])]; - tensor var_852_end_mask_0 = const()[name = tensor("op_852_end_mask_0"), val = tensor([true, true, true])]; - tensor conv_out_7 = transpose(perm = conv_out_7_perm_0, x = input_149)[name = tensor("transpose_32")]; - tensor var_852 = slice_by_index(begin = var_852_begin_0, end = var_852_end_0, end_mask = var_852_end_mask_0, x = conv_out_7)[name = tensor("op_852")]; - tensor var_854_perm_0 = const()[name = tensor("op_854_perm_0"), val = tensor([1, 0, 2])]; - tensor var_854 = transpose(perm = var_854_perm_0, x = var_852)[name = tensor("transpose_31")]; - tensor input_151 = add(x = x_21, y = var_854)[name = tensor("input_151")]; - tensor input_153_axes_0 = const()[name = tensor("input_153_axes_0"), val = tensor([-1])]; - tensor input_153 = layer_norm(axes = input_153_axes_0, beta = encoder_ffn2_3_module_sequential_0_bias, epsilon = var_28, gamma = encoder_ffn2_3_module_sequential_0_weight, x = input_151)[name = tensor("input_153")]; - tensor inputs = linear(bias = encoder_ffn2_3_module_sequential_1_linear_bias, weight = encoder_ffn2_3_module_sequential_1_linear_weight, x = input_153)[name = tensor("linear_35")]; - tensor input_155 = silu(x = inputs)[name = tensor("input_155")]; - tensor input_159 = linear(bias = encoder_ffn2_3_module_sequential_4_linear_bias, weight = encoder_ffn2_3_module_sequential_4_linear_weight, x = input_155)[name = tensor("linear_36")]; - tensor var_877 = const()[name = tensor("op_877"), val = tensor(0x1p-1)]; - tensor var_878 = mul(x = input_159, y = var_877)[name = tensor("op_878")]; - tensor input_161 = add(x = var_878, y = input_151)[name = tensor("input_161")]; + tensor var_878_begin_0 = const()[name = tensor("op_878_begin_0"), val = tensor([0, -1, 0])]; + tensor var_878_end_0 = const()[name = tensor("op_878_end_0"), val = tensor([1, 16, 256])]; + tensor var_878_end_mask_0 = const()[name = tensor("op_878_end_mask_0"), val = tensor([true, true, true])]; + tensor conv_out_7 = transpose(perm = conv_out_7_perm_0, x = input_151)[name = tensor("transpose_32")]; + tensor var_878 = slice_by_index(begin = var_878_begin_0, end = var_878_end_0, end_mask = var_878_end_mask_0, x = conv_out_7)[name = tensor("op_878")]; + tensor var_880_perm_0 = const()[name = tensor("op_880_perm_0"), val = tensor([1, 0, 2])]; + tensor var_880 = transpose(perm = var_880_perm_0, x = var_878)[name = tensor("transpose_31")]; + tensor input_153 = add(x = x_21, y = var_880)[name = tensor("input_153")]; + tensor input_155_axes_0 = const()[name = tensor("input_155_axes_0"), val = tensor([-1])]; + tensor input_155 = layer_norm(axes = input_155_axes_0, beta = inner_encoder_ffn2_3_module_sequential_0_bias, epsilon = var_35, gamma = inner_encoder_ffn2_3_module_sequential_0_weight, x = input_153)[name = tensor("input_155")]; + tensor inputs = linear(bias = inner_encoder_ffn2_3_module_sequential_1_linear_bias, weight = inner_encoder_ffn2_3_module_sequential_1_linear_weight, x = input_155)[name = tensor("linear_35")]; + tensor input_157 = silu(x = inputs)[name = tensor("input_157")]; + tensor input_161 = linear(bias = inner_encoder_ffn2_3_module_sequential_4_linear_bias, weight = inner_encoder_ffn2_3_module_sequential_4_linear_weight, x = input_157)[name = tensor("linear_36")]; + tensor var_903 = const()[name = tensor("op_903"), val = tensor(0x1p-1)]; + tensor var_904 = mul(x = input_161, y = var_903)[name = tensor("op_904")]; + tensor input_163 = add(x = var_904, y = input_153)[name = tensor("input_163")]; tensor x_25_axes_0 = const()[name = tensor("x_25_axes_0"), val = tensor([-1])]; - tensor x_25 = layer_norm(axes = x_25_axes_0, beta = encoder_layer_norm_3_bias, epsilon = var_28, gamma = encoder_layer_norm_3_weight, x = input_161)[name = tensor("x_25")]; + tensor x_25 = layer_norm(axes = x_25_axes_0, beta = inner_encoder_layer_norm_3_bias, epsilon = var_35, gamma = inner_encoder_layer_norm_3_weight, x = input_163)[name = tensor("x_25")]; tensor x_ct_perm_0 = const()[name = tensor("x_ct_perm_0"), val = tensor([0, 2, 1])]; tensor cat_interleave_0 = const()[name = tensor("cat_interleave_0"), val = tensor(false)]; tensor x_ct = transpose(perm = x_ct_perm_0, x = x_25)[name = tensor("transpose_30")]; - tensor cat = concat(axis = var_20, interleave = cat_interleave_0, values = (cnn_window, x_ct))[name = tensor("cat")]; + tensor cat = concat(axis = var_40, interleave = cat_interleave_0, values = (cnn_window, x_ct))[name = tensor("cat")]; tensor conv_out_pad_type_0 = const()[name = tensor("conv_out_pad_type_0"), val = tensor("valid")]; tensor conv_out_strides_0 = const()[name = tensor("conv_out_strides_0"), val = tensor([1])]; tensor conv_out_pad_0 = const()[name = tensor("conv_out_pad_0"), val = tensor([0, 0])]; tensor conv_out_dilations_0 = const()[name = tensor("conv_out_dilations_0"), val = tensor([1])]; tensor conv_out_groups_0 = const()[name = tensor("conv_out_groups_0"), val = tensor(1)]; - tensor conv_out = conv(bias = encoder_cnn_bias, dilations = conv_out_dilations_0, groups = conv_out_groups_0, pad = conv_out_pad_0, pad_type = conv_out_pad_type_0, strides = conv_out_strides_0, weight = encoder_cnn_weight, x = cat)[name = tensor("conv_out")]; - tensor var_896_begin_0 = const()[name = tensor("op_896_begin_0"), val = tensor([0, 0, 1])]; - tensor var_896_end_0 = const()[name = tensor("op_896_end_0"), val = tensor([1, 256, 19])]; - tensor var_896_end_mask_0 = const()[name = tensor("op_896_end_mask_0"), val = tensor([true, true, true])]; - tensor cnn_window_new = slice_by_index(begin = var_896_begin_0, end = var_896_end_0, end_mask = var_896_end_mask_0, x = cat)[name = tensor("op_896")]; - tensor input_163_perm_0 = const()[name = tensor("input_163_perm_0"), val = tensor([0, 2, 1])]; - tensor var_898 = const()[name = tensor("op_898"), val = tensor([-1])]; - tensor input_163 = transpose(perm = input_163_perm_0, x = conv_out)[name = tensor("transpose_29")]; - tensor var_899 = reduce_l2_norm(axes = var_898, keep_dims = var_29, x = input_163)[name = tensor("op_899")]; + tensor conv_out = conv(bias = inner_encoder_cnn_bias, dilations = conv_out_dilations_0, groups = conv_out_groups_0, pad = conv_out_pad_0, pad_type = conv_out_pad_type_0, strides = conv_out_strides_0, weight = inner_encoder_cnn_weight, x = cat)[name = tensor("conv_out")]; + tensor var_922_begin_0 = const()[name = tensor("op_922_begin_0"), val = tensor([0, 0, 1])]; + tensor var_922_end_0 = const()[name = tensor("op_922_end_0"), val = tensor([1, 256, 19])]; + tensor var_922_end_mask_0 = const()[name = tensor("op_922_end_mask_0"), val = tensor([true, true, true])]; + tensor cnn_window_new = slice_by_index(begin = var_922_begin_0, end = var_922_end_0, end_mask = var_922_end_mask_0, x = cat)[name = tensor("op_922")]; + tensor input_165_perm_0 = const()[name = tensor("input_165_perm_0"), val = tensor([0, 2, 1])]; + tensor var_924 = const()[name = tensor("op_924"), val = tensor([-1])]; + tensor input_165 = transpose(perm = input_165_perm_0, x = conv_out)[name = tensor("transpose_29")]; + tensor var_925 = reduce_l2_norm(axes = var_924, keep_dims = var_34, x = input_165)[name = tensor("op_925")]; tensor const_12 = const()[name = tensor("const_12"), val = tensor(0x1.fffffep+127)]; - tensor clip_0 = clip(alpha = var_11, beta = const_12, x = var_899)[name = tensor("clip_0")]; - tensor emb = real_div(x = input_163, y = clip_0)[name = tensor("emb")]; - tensor var_903_axis_0 = const()[name = tensor("op_903_axis_0"), val = tensor(0)]; - tensor enc_kv_new = stack(axis = var_903_axis_0, values = (var_206, var_396, var_586, nkv_1))[name = tensor("op_903")]; - tensor var_905_axis_0 = const()[name = tensor("op_905_axis_0"), val = tensor(0)]; - tensor enc_scale_new = stack(axis = var_905_axis_0, values = (new_scale_1, new_scale_3, new_scale_5, new_scale_7))[name = tensor("op_905")]; - tensor var_907_axis_0 = const()[name = tensor("op_907_axis_0"), val = tensor(0)]; - tensor enc_conv_cache_new = stack(axis = var_907_axis_0, values = (window_3, window_7, window_11, window))[name = tensor("op_907")]; - tensor var_916 = const()[name = tensor("op_916"), val = tensor(0x1.5798eep-27)]; - tensor var_921 = const()[name = tensor("op_921"), val = tensor(0x1.4f8b58p-17)]; - tensor var_923 = const()[name = tensor("op_923"), val = tensor(0x1.0c6f7ap-20)]; - tensor var_924 = const()[name = tensor("op_924"), val = tensor(true)]; - tensor var_926 = const()[name = tensor("op_926"), val = tensor(0x1p+0)]; - tensor var_930 = const()[name = tensor("op_930"), val = tensor(-1)]; - tensor var_936 = const()[name = tensor("op_936"), val = tensor(0)]; - tensor var_993 = const()[name = tensor("op_993"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(44395328)))]; - tensor var_998_axes_0 = const()[name = tensor("op_998_axes_0"), val = tensor([2])]; - tensor var_998 = expand_dims(axes = var_998_axes_0, x = emb)[name = tensor("op_998")]; + tensor clip_0 = clip(alpha = var_48, beta = const_12, x = var_925)[name = tensor("clip_0")]; + tensor emb = real_div(x = input_165, y = clip_0)[name = tensor("emb")]; + tensor var_929_axis_0 = const()[name = tensor("op_929_axis_0"), val = tensor(0)]; + tensor enc_kv_new = stack(axis = var_929_axis_0, values = (var_232, var_422, var_612, nkv_1))[name = tensor("op_929")]; + tensor var_931_axis_0 = const()[name = tensor("op_931_axis_0"), val = tensor(0)]; + tensor enc_scale_new = stack(axis = var_931_axis_0, values = (new_scale_1, new_scale_3, new_scale_5, new_scale_7))[name = tensor("op_931")]; + tensor var_933_axis_0 = const()[name = tensor("op_933_axis_0"), val = tensor(0)]; + tensor enc_conv_cache_new = stack(axis = var_933_axis_0, values = (window_3, window_7, window_11, window))[name = tensor("op_933")]; + tensor var_996 = const()[name = tensor("op_996"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(44395328)))]; + tensor var_1001_axes_0 = const()[name = tensor("op_1001_axes_0"), val = tensor([2])]; + tensor var_1001 = expand_dims(axes = var_1001_axes_0, x = emb)[name = tensor("op_1001")]; tensor emb_exp_reps_0 = const()[name = tensor("emb_exp_reps_0"), val = tensor([1, 1, 6, 1])]; - tensor emb_exp = tile(reps = emb_exp_reps_0, x = var_998)[name = tensor("emb_exp")]; - tensor input_165_interleave_0 = const()[name = tensor("input_165_interleave_0"), val = tensor(false)]; - tensor input_165 = concat(axis = var_930, interleave = input_165_interleave_0, values = (emb_exp, var_993))[name = tensor("input_165")]; - tensor x_27 = linear(bias = decoder_convert_bias, weight = decoder_convert_weight, x = input_165)[name = tensor("linear_37")]; - tensor var_1006_perm_0 = const()[name = tensor("op_1006_perm_0"), val = tensor([0, 2, 1, 3])]; - tensor var_1010 = const()[name = tensor("op_1010"), val = tensor([6, 1, 256])]; - tensor var_1006 = transpose(perm = var_1006_perm_0, x = x_27)[name = tensor("transpose_28")]; - tensor x_29 = reshape(shape = var_1010, x = var_1006)[name = tensor("x_29")]; + tensor emb_exp = tile(reps = emb_exp_reps_0, x = var_1001)[name = tensor("emb_exp")]; + tensor input_167_interleave_0 = const()[name = tensor("input_167_interleave_0"), val = tensor(false)]; + tensor input_167 = concat(axis = var_41, interleave = input_167_interleave_0, values = (emb_exp, var_996))[name = tensor("input_167")]; + tensor x_27 = linear(bias = inner_decoder_convert_bias, weight = inner_decoder_convert_weight, x = input_167)[name = tensor("linear_37")]; + tensor var_1009_perm_0 = const()[name = tensor("op_1009_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_1013 = const()[name = tensor("op_1013"), val = tensor([6, 1, 256])]; + tensor var_1009 = transpose(perm = var_1009_perm_0, x = x_27)[name = tensor("transpose_28")]; + tensor x_29 = reshape(shape = var_1013, x = var_1009)[name = tensor("x_29")]; tensor prev_kv_9_begin_0 = const()[name = tensor("prev_kv_9_begin_0"), val = tensor([0, 0, 0, 0, 0])]; tensor prev_kv_9_end_0 = const()[name = tensor("prev_kv_9_end_0"), val = tensor([1, 6, 4, 64, 64])]; tensor prev_kv_9_end_mask_0 = const()[name = tensor("prev_kv_9_end_mask_0"), val = tensor([false, true, true, true, true])]; @@ -846,131 +845,131 @@ program(1.0) tensor prev_scale_9_end_mask_0 = const()[name = tensor("prev_scale_9_end_mask_0"), val = tensor([false, true])]; tensor prev_scale_9_squeeze_mask_0 = const()[name = tensor("prev_scale_9_squeeze_mask_0"), val = tensor([true, false])]; tensor prev_scale_9 = slice_by_index(begin = prev_scale_9_begin_0, end = prev_scale_9_end_0, end_mask = prev_scale_9_end_mask_0, squeeze_mask = prev_scale_9_squeeze_mask_0, x = dec_scale)[name = tensor("prev_scale_9")]; - tensor var_1018 = linear(bias = decoder_q_proj_0_bias, weight = decoder_q_proj_0_weight, x = x_29)[name = tensor("linear_38")]; - tensor var_1019 = const()[name = tensor("op_1019"), val = tensor([6, 1, 4, 64])]; - tensor var_1020 = reshape(shape = var_1019, x = var_1018)[name = tensor("op_1020")]; + tensor var_1021 = linear(bias = inner_decoder_q_proj_0_bias, weight = inner_decoder_q_proj_0_weight, x = x_29)[name = tensor("linear_38")]; + tensor var_1022 = const()[name = tensor("op_1022"), val = tensor([6, 1, 4, 64])]; + tensor var_1023 = reshape(shape = var_1022, x = var_1021)[name = tensor("op_1023")]; tensor q_9_perm_0 = const()[name = tensor("q_9_perm_0"), val = tensor([0, 2, 1, 3])]; - tensor var_1024 = linear(bias = decoder_k_proj_0_bias, weight = decoder_k_proj_0_weight, x = x_29)[name = tensor("linear_39")]; - tensor var_1025 = const()[name = tensor("op_1025"), val = tensor(0x1p-3)]; - tensor var_1026 = mul(x = var_1024, y = var_1025)[name = tensor("op_1026")]; - tensor var_1027 = const()[name = tensor("op_1027"), val = tensor([6, 1, 4, 64])]; - tensor var_1028 = reshape(shape = var_1027, x = var_1026)[name = tensor("op_1028")]; + tensor var_1027 = linear(bias = inner_decoder_k_proj_0_bias, weight = inner_decoder_k_proj_0_weight, x = x_29)[name = tensor("linear_39")]; + tensor var_1028 = const()[name = tensor("op_1028"), val = tensor(0x1p-3)]; + tensor var_1029 = mul(x = var_1027, y = var_1028)[name = tensor("op_1029")]; + tensor var_1030 = const()[name = tensor("op_1030"), val = tensor([6, 1, 4, 64])]; + tensor var_1031 = reshape(shape = var_1030, x = var_1029)[name = tensor("op_1031")]; tensor k_9_perm_0 = const()[name = tensor("k_9_perm_0"), val = tensor([0, 2, 1, 3])]; - tensor var_1032 = linear(bias = decoder_v_proj_0_bias, weight = decoder_v_proj_0_weight, x = x_29)[name = tensor("linear_40")]; - tensor var_1033 = const()[name = tensor("op_1033"), val = tensor([6, 1, 4, 64])]; - tensor var_1034 = reshape(shape = var_1033, x = var_1032)[name = tensor("op_1034")]; + tensor var_1035 = linear(bias = inner_decoder_v_proj_0_bias, weight = inner_decoder_v_proj_0_weight, x = x_29)[name = tensor("linear_40")]; + tensor var_1036 = const()[name = tensor("op_1036"), val = tensor([6, 1, 4, 64])]; + tensor var_1037 = reshape(shape = var_1036, x = var_1035)[name = tensor("op_1037")]; tensor v_9_perm_0 = const()[name = tensor("v_9_perm_0"), val = tensor([0, 2, 1, 3])]; - tensor input_169 = linear(bias = decoder_g_proj_0_bias, weight = decoder_g_proj_0_weight, x = x_29)[name = tensor("linear_41")]; + tensor input_171 = linear(bias = inner_decoder_g_proj_0_bias, weight = inner_decoder_g_proj_0_weight, x = x_29)[name = tensor("linear_41")]; tensor cumsum_mask_1_exclusive_0 = const()[name = tensor("cumsum_mask_1_exclusive_0"), val = tensor(false)]; tensor cumsum_mask_1_reverse_0 = const()[name = tensor("cumsum_mask_1_reverse_0"), val = tensor(false)]; - tensor cumsum_mask_1 = cumsum(axis = var_936, exclusive = cumsum_mask_1_exclusive_0, reverse = cumsum_mask_1_reverse_0, x = valid_mask)[name = tensor("cumsum_mask_1")]; + tensor cumsum_mask_1 = cumsum(axis = var_38, exclusive = cumsum_mask_1_exclusive_0, reverse = cumsum_mask_1_reverse_0, x = valid_mask)[name = tensor("cumsum_mask_1")]; tensor sqrt_s0_9 = sqrt(x = prev_scale_9)[name = tensor("sqrt_s0_9")]; tensor s_t_9 = add(x = prev_scale_9, y = cumsum_mask_1)[name = tensor("s_t_9")]; tensor const_20 = const()[name = tensor("const_20"), val = tensor(0x1.fffffep+127)]; - tensor clip_1 = clip(alpha = var_926, beta = const_20, x = s_t_9)[name = tensor("clip_1")]; + tensor clip_1 = clip(alpha = var_28, beta = const_20, x = s_t_9)[name = tensor("clip_1")]; tensor sqrt_s_t_9 = sqrt(x = clip_1)[name = tensor("sqrt_s_t_9")]; tensor qk_9_transpose_x_1 = const()[name = tensor("qk_9_transpose_x_1"), val = tensor(false)]; tensor qk_9_transpose_y_1 = const()[name = tensor("qk_9_transpose_y_1"), val = tensor(true)]; - tensor k_9 = transpose(perm = k_9_perm_0, x = var_1028)[name = tensor("transpose_26")]; - tensor q_9 = transpose(perm = q_9_perm_0, x = var_1020)[name = tensor("transpose_27")]; + tensor k_9 = transpose(perm = k_9_perm_0, x = var_1031)[name = tensor("transpose_26")]; + tensor q_9 = transpose(perm = q_9_perm_0, x = var_1023)[name = tensor("transpose_27")]; tensor qk_9 = matmul(transpose_x = qk_9_transpose_x_1, transpose_y = qk_9_transpose_y_1, x = q_9, y = k_9)[name = tensor("qk_9")]; - tensor var_1046 = const()[name = tensor("op_1046"), val = tensor([1, 1])]; - tensor var_1047 = reshape(shape = var_1046, x = valid_mask)[name = tensor("op_1047")]; tensor var_1049 = const()[name = tensor("op_1049"), val = tensor([1, 1])]; - tensor var_1050 = reshape(shape = var_1049, x = sqrt_s_t_9)[name = tensor("op_1050")]; - tensor M_9 = real_div(x = var_1047, y = var_1050)[name = tensor("M_9")]; - tensor var_1052 = mul(x = qk_9, y = M_9)[name = tensor("op_1052")]; + tensor var_1050 = reshape(shape = var_1049, x = valid_mask)[name = tensor("op_1050")]; + tensor var_1052 = const()[name = tensor("op_1052"), val = tensor([1, 1])]; + tensor var_1053 = reshape(shape = var_1052, x = sqrt_s_t_9)[name = tensor("op_1053")]; + tensor M_9 = real_div(x = var_1050, y = var_1053)[name = tensor("M_9")]; + tensor var_1055 = mul(x = qk_9, y = M_9)[name = tensor("op_1055")]; tensor inner_9_transpose_x_0 = const()[name = tensor("inner_9_transpose_x_0"), val = tensor(false)]; tensor inner_9_transpose_y_0 = const()[name = tensor("inner_9_transpose_y_0"), val = tensor(false)]; - tensor v_9 = transpose(perm = v_9_perm_0, x = var_1034)[name = tensor("transpose_25")]; - tensor inner_9 = matmul(transpose_x = inner_9_transpose_x_0, transpose_y = inner_9_transpose_y_0, x = var_1052, y = v_9)[name = tensor("inner_9")]; - tensor var_1054_transpose_x_0 = const()[name = tensor("op_1054_transpose_x_0"), val = tensor(false)]; - tensor var_1054_transpose_y_0 = const()[name = tensor("op_1054_transpose_y_0"), val = tensor(false)]; - tensor var_1054 = matmul(transpose_x = var_1054_transpose_x_0, transpose_y = var_1054_transpose_y_0, x = q_9, y = prev_kv_9)[name = tensor("op_1054")]; - tensor var_1055 = real_div(x = sqrt_s0_9, y = sqrt_s_t_9)[name = tensor("op_1055")]; - tensor var_1056 = const()[name = tensor("op_1056"), val = tensor([1, 1, 1, 1])]; - tensor var_1057 = reshape(shape = var_1056, x = var_1055)[name = tensor("op_1057")]; - tensor cross_9 = mul(x = var_1054, y = var_1057)[name = tensor("cross_9")]; + tensor v_9 = transpose(perm = v_9_perm_0, x = var_1037)[name = tensor("transpose_25")]; + tensor inner_9 = matmul(transpose_x = inner_9_transpose_x_0, transpose_y = inner_9_transpose_y_0, x = var_1055, y = v_9)[name = tensor("inner_9")]; + tensor var_1057_transpose_x_0 = const()[name = tensor("op_1057_transpose_x_0"), val = tensor(false)]; + tensor var_1057_transpose_y_0 = const()[name = tensor("op_1057_transpose_y_0"), val = tensor(false)]; + tensor var_1057 = matmul(transpose_x = var_1057_transpose_x_0, transpose_y = var_1057_transpose_y_0, x = q_9, y = prev_kv_9)[name = tensor("op_1057")]; + tensor var_1058 = real_div(x = sqrt_s0_9, y = sqrt_s_t_9)[name = tensor("op_1058")]; + tensor var_1059 = const()[name = tensor("op_1059"), val = tensor([1, 1, 1, 1])]; + tensor var_1060 = reshape(shape = var_1059, x = var_1058)[name = tensor("op_1060")]; + tensor cross_9 = mul(x = var_1057, y = var_1060)[name = tensor("cross_9")]; tensor out_25 = add(x = inner_9, y = cross_9)[name = tensor("out_25")]; - tensor var_1060 = const()[name = tensor("op_1060"), val = tensor([1, 1, 1, 1])]; - tensor var_1061 = reshape(shape = var_1060, x = valid_mask)[name = tensor("op_1061")]; - tensor v_masked_1 = mul(x = v_9, y = var_1061)[name = tensor("v_masked_1")]; - tensor var_1063 = mul(x = prev_kv_9, y = sqrt_s0_9)[name = tensor("op_1063")]; - tensor var_1065_transpose_x_1 = const()[name = tensor("op_1065_transpose_x_1"), val = tensor(true)]; - tensor var_1065_transpose_y_1 = const()[name = tensor("op_1065_transpose_y_1"), val = tensor(false)]; - tensor var_1065 = matmul(transpose_x = var_1065_transpose_x_1, transpose_y = var_1065_transpose_y_1, x = k_9, y = v_masked_1)[name = tensor("op_1065")]; - tensor new_kv_unnorm_9 = add(x = var_1063, y = var_1065)[name = tensor("new_kv_unnorm_9")]; - tensor var_1067_keep_dims_0 = const()[name = tensor("op_1067_keep_dims_0"), val = tensor(false)]; - tensor var_1067 = reduce_sum(keep_dims = var_1067_keep_dims_0, x = valid_mask)[name = tensor("op_1067")]; - tensor var_1068 = const()[name = tensor("op_1068"), val = tensor([1])]; - tensor var_1069 = reshape(shape = var_1068, x = var_1067)[name = tensor("op_1069")]; - tensor new_scale_9 = add(x = prev_scale_9, y = var_1069)[name = tensor("new_scale_9")]; + tensor var_1063 = const()[name = tensor("op_1063"), val = tensor([1, 1, 1, 1])]; + tensor var_1064 = reshape(shape = var_1063, x = valid_mask)[name = tensor("op_1064")]; + tensor v_masked_1 = mul(x = v_9, y = var_1064)[name = tensor("v_masked_1")]; + tensor var_1066 = mul(x = prev_kv_9, y = sqrt_s0_9)[name = tensor("op_1066")]; + tensor var_1068_transpose_x_1 = const()[name = tensor("op_1068_transpose_x_1"), val = tensor(true)]; + tensor var_1068_transpose_y_1 = const()[name = tensor("op_1068_transpose_y_1"), val = tensor(false)]; + tensor var_1068 = matmul(transpose_x = var_1068_transpose_x_1, transpose_y = var_1068_transpose_y_1, x = k_9, y = v_masked_1)[name = tensor("op_1068")]; + tensor new_kv_unnorm_9 = add(x = var_1066, y = var_1068)[name = tensor("new_kv_unnorm_9")]; + tensor var_1070_keep_dims_0 = const()[name = tensor("op_1070_keep_dims_0"), val = tensor(false)]; + tensor var_1070 = reduce_sum(keep_dims = var_1070_keep_dims_0, x = valid_mask)[name = tensor("op_1070")]; + tensor var_1071 = const()[name = tensor("op_1071"), val = tensor([1])]; + tensor var_1072 = reshape(shape = var_1071, x = var_1070)[name = tensor("op_1072")]; + tensor new_scale_9 = add(x = prev_scale_9, y = var_1072)[name = tensor("new_scale_9")]; tensor const_21 = const()[name = tensor("const_21"), val = tensor(0x1.fffffep+127)]; - tensor clip_2 = clip(alpha = var_926, beta = const_21, x = new_scale_9)[name = tensor("clip_2")]; + tensor clip_2 = clip(alpha = var_28, beta = const_21, x = new_scale_9)[name = tensor("clip_2")]; tensor sqrt_new_scale_1 = sqrt(x = clip_2)[name = tensor("sqrt_new_scale_1")]; - tensor var_1073 = real_div(x = new_kv_unnorm_9, y = sqrt_new_scale_1)[name = tensor("op_1073")]; - tensor var_1074_perm_0 = const()[name = tensor("op_1074_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_1076 = real_div(x = new_kv_unnorm_9, y = sqrt_new_scale_1)[name = tensor("op_1076")]; + tensor var_1077_perm_0 = const()[name = tensor("op_1077_perm_0"), val = tensor([0, 2, 1, 3])]; tensor out_27_axes_0 = const()[name = tensor("out_27_axes_0"), val = tensor([-1])]; - tensor var_1074 = transpose(perm = var_1074_perm_0, x = out_25)[name = tensor("transpose_24")]; - tensor out_27 = layer_norm(axes = out_27_axes_0, epsilon = var_923, x = var_1074)[name = tensor("out_27")]; - tensor var_1078 = const()[name = tensor("op_1078"), val = tensor([6, 1, 256])]; - tensor out_29 = reshape(shape = var_1078, x = out_27)[name = tensor("out_29")]; - tensor var_1080 = silu(x = input_169)[name = tensor("op_1080")]; - tensor input_171 = mul(x = var_1080, y = out_29)[name = tensor("input_171")]; - tensor ret_out_9 = linear(bias = decoder_out_proj_0_bias, weight = decoder_out_proj_0_weight, x = input_171)[name = tensor("linear_42")]; - tensor input_173 = add(x = x_29, y = ret_out_9)[name = tensor("input_173")]; + tensor var_1077 = transpose(perm = var_1077_perm_0, x = out_25)[name = tensor("transpose_24")]; + tensor out_27 = layer_norm(axes = out_27_axes_0, epsilon = var_43, x = var_1077)[name = tensor("out_27")]; + tensor var_1081 = const()[name = tensor("op_1081"), val = tensor([6, 1, 256])]; + tensor out_29 = reshape(shape = var_1081, x = out_27)[name = tensor("out_29")]; + tensor var_1083 = silu(x = input_171)[name = tensor("op_1083")]; + tensor input_173 = mul(x = var_1083, y = out_29)[name = tensor("input_173")]; + tensor ret_out_9 = linear(bias = inner_decoder_out_proj_0_bias, weight = inner_decoder_out_proj_0_weight, x = input_173)[name = tensor("linear_42")]; + tensor input_175 = add(x = x_29, y = ret_out_9)[name = tensor("input_175")]; tensor xt_1_axes_0 = const()[name = tensor("xt_1_axes_0"), val = tensor([-1])]; - tensor xt_1 = layer_norm(axes = xt_1_axes_0, beta = decoder_norm11_0_bias, epsilon = var_921, gamma = decoder_norm11_0_weight, x = input_173)[name = tensor("xt_1")]; - tensor var_1090 = const()[name = tensor("op_1090"), val = tensor([1, 6, 1, 256])]; - tensor var_1091 = reshape(shape = var_1090, x = xt_1)[name = tensor("op_1091")]; - tensor var_1092_perm_0 = const()[name = tensor("op_1092_perm_0"), val = tensor([0, 2, 1, 3])]; - tensor var_1095 = const()[name = tensor("op_1095"), val = tensor([1, 6, 256])]; - tensor var_1092 = transpose(perm = var_1092_perm_0, x = var_1091)[name = tensor("transpose_23")]; - tensor query_1 = reshape(shape = var_1095, x = var_1092)[name = tensor("query_1")]; + tensor xt_1 = layer_norm(axes = xt_1_axes_0, beta = inner_decoder_norm11_0_bias, epsilon = var_35, gamma = inner_decoder_norm11_0_weight, x = input_175)[name = tensor("xt_1")]; + tensor var_1093 = const()[name = tensor("op_1093"), val = tensor([1, 6, 1, 256])]; + tensor var_1094 = reshape(shape = var_1093, x = xt_1)[name = tensor("op_1094")]; + tensor var_1095_perm_0 = const()[name = tensor("op_1095_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_1098 = const()[name = tensor("op_1098"), val = tensor([1, 6, 256])]; + tensor var_1095 = transpose(perm = var_1095_perm_0, x = var_1094)[name = tensor("transpose_23")]; + tensor query_1 = reshape(shape = var_1098, x = var_1095)[name = tensor("query_1")]; tensor query_3_perm_0 = const()[name = tensor("query_3_perm_0"), val = tensor([1, 0, 2])]; tensor query_3 = transpose(perm = query_3_perm_0, x = query_1)[name = tensor("transpose_22")]; - tensor var_1118 = linear(bias = decoder_self_attn2_0_in_proj_bias, weight = decoder_self_attn2_0_in_proj_weight, x = query_3)[name = tensor("linear_43")]; + tensor var_1121 = linear(bias = inner_decoder_self_attn2_0_in_proj_bias, weight = inner_decoder_self_attn2_0_in_proj_weight, x = query_3)[name = tensor("linear_43")]; tensor concat_1 = const()[name = tensor("concat_1"), val = tensor([6, 1, 3, 256])]; - tensor var_1120 = reshape(shape = concat_1, x = var_1118)[name = tensor("op_1120")]; - tensor var_1121_axes_0 = const()[name = tensor("op_1121_axes_0"), val = tensor([0])]; - tensor var_1121 = expand_dims(axes = var_1121_axes_0, x = var_1120)[name = tensor("op_1121")]; - tensor var_1122_perm_0 = const()[name = tensor("op_1122_perm_0"), val = tensor([-2, 1, 2, 0, 4])]; - tensor var_1123_axes_0 = const()[name = tensor("op_1123_axes_0"), val = tensor([-2])]; - tensor var_1122 = transpose(perm = var_1122_perm_0, x = var_1121)[name = tensor("transpose_21")]; - tensor var_1123 = squeeze(axes = var_1123_axes_0, x = var_1122)[name = tensor("op_1123")]; + tensor var_1123 = reshape(shape = concat_1, x = var_1121)[name = tensor("op_1123")]; + tensor var_1124_axes_0 = const()[name = tensor("op_1124_axes_0"), val = tensor([0])]; + tensor var_1124 = expand_dims(axes = var_1124_axes_0, x = var_1123)[name = tensor("op_1124")]; + tensor var_1125_perm_0 = const()[name = tensor("op_1125_perm_0"), val = tensor([-2, 1, 2, 0, 4])]; + tensor var_1126_axes_0 = const()[name = tensor("op_1126_axes_0"), val = tensor([-2])]; + tensor var_1125 = transpose(perm = var_1125_perm_0, x = var_1124)[name = tensor("transpose_21")]; + tensor var_1126 = squeeze(axes = var_1126_axes_0, x = var_1125)[name = tensor("op_1126")]; tensor q_11_begin_0 = const()[name = tensor("q_11_begin_0"), val = tensor([0, 0, 0, 0])]; tensor q_11_end_0 = const()[name = tensor("q_11_end_0"), val = tensor([1, 6, 1, 256])]; tensor q_11_end_mask_0 = const()[name = tensor("q_11_end_mask_0"), val = tensor([false, true, true, true])]; tensor q_11_squeeze_mask_0 = const()[name = tensor("q_11_squeeze_mask_0"), val = tensor([true, false, false, false])]; - tensor q_11 = slice_by_index(begin = q_11_begin_0, end = q_11_end_0, end_mask = q_11_end_mask_0, squeeze_mask = q_11_squeeze_mask_0, x = var_1123)[name = tensor("q_11")]; + tensor q_11 = slice_by_index(begin = q_11_begin_0, end = q_11_end_0, end_mask = q_11_end_mask_0, squeeze_mask = q_11_squeeze_mask_0, x = var_1126)[name = tensor("q_11")]; tensor k_11_begin_0 = const()[name = tensor("k_11_begin_0"), val = tensor([1, 0, 0, 0])]; tensor k_11_end_0 = const()[name = tensor("k_11_end_0"), val = tensor([2, 6, 1, 256])]; tensor k_11_end_mask_0 = const()[name = tensor("k_11_end_mask_0"), val = tensor([false, true, true, true])]; tensor k_11_squeeze_mask_0 = const()[name = tensor("k_11_squeeze_mask_0"), val = tensor([true, false, false, false])]; - tensor k_11 = slice_by_index(begin = k_11_begin_0, end = k_11_end_0, end_mask = k_11_end_mask_0, squeeze_mask = k_11_squeeze_mask_0, x = var_1123)[name = tensor("k_11")]; + tensor k_11 = slice_by_index(begin = k_11_begin_0, end = k_11_end_0, end_mask = k_11_end_mask_0, squeeze_mask = k_11_squeeze_mask_0, x = var_1126)[name = tensor("k_11")]; tensor v_11_begin_0 = const()[name = tensor("v_11_begin_0"), val = tensor([2, 0, 0, 0])]; tensor v_11_end_0 = const()[name = tensor("v_11_end_0"), val = tensor([3, 6, 1, 256])]; tensor v_11_end_mask_0 = const()[name = tensor("v_11_end_mask_0"), val = tensor([false, true, true, true])]; tensor v_11_squeeze_mask_0 = const()[name = tensor("v_11_squeeze_mask_0"), val = tensor([true, false, false, false])]; - tensor v_11 = slice_by_index(begin = v_11_begin_0, end = v_11_end_0, end_mask = v_11_end_mask_0, squeeze_mask = v_11_squeeze_mask_0, x = var_1123)[name = tensor("v_11")]; - tensor var_1131 = const()[name = tensor("op_1131"), val = tensor([6, 4, 64])]; - tensor var_1132 = reshape(shape = var_1131, x = q_11)[name = tensor("op_1132")]; + tensor v_11 = slice_by_index(begin = v_11_begin_0, end = v_11_end_0, end_mask = v_11_end_mask_0, squeeze_mask = v_11_squeeze_mask_0, x = var_1126)[name = tensor("v_11")]; + tensor var_1134 = const()[name = tensor("op_1134"), val = tensor([6, 4, 64])]; + tensor var_1135 = reshape(shape = var_1134, x = q_11)[name = tensor("op_1135")]; tensor q_13_perm_0 = const()[name = tensor("q_13_perm_0"), val = tensor([1, 0, 2])]; - tensor var_1138 = const()[name = tensor("op_1138"), val = tensor([6, 4, 64])]; - tensor var_1139 = reshape(shape = var_1138, x = k_11)[name = tensor("op_1139")]; + tensor var_1141 = const()[name = tensor("op_1141"), val = tensor([6, 4, 64])]; + tensor var_1142 = reshape(shape = var_1141, x = k_11)[name = tensor("op_1142")]; tensor k_13_perm_0 = const()[name = tensor("k_13_perm_0"), val = tensor([1, 0, 2])]; - tensor var_1145 = const()[name = tensor("op_1145"), val = tensor([6, 4, 64])]; - tensor var_1146 = reshape(shape = var_1145, x = v_11)[name = tensor("op_1146")]; + tensor var_1148 = const()[name = tensor("op_1148"), val = tensor([6, 4, 64])]; + tensor var_1149 = reshape(shape = var_1148, x = v_11)[name = tensor("op_1149")]; tensor v_13_perm_0 = const()[name = tensor("v_13_perm_0"), val = tensor([1, 0, 2])]; - tensor var_1149 = const()[name = tensor("op_1149"), val = tensor([1, 4, 6, 64])]; - tensor q_13 = transpose(perm = q_13_perm_0, x = var_1132)[name = tensor("transpose_20")]; - tensor q_15 = reshape(shape = var_1149, x = q_13)[name = tensor("q_15")]; - tensor var_1151 = const()[name = tensor("op_1151"), val = tensor([1, 4, 6, 64])]; - tensor k_13 = transpose(perm = k_13_perm_0, x = var_1139)[name = tensor("transpose_19")]; - tensor k_15 = reshape(shape = var_1151, x = k_13)[name = tensor("k_15")]; - tensor var_1153 = const()[name = tensor("op_1153"), val = tensor([1, 4, 6, 64])]; - tensor v_13 = transpose(perm = v_13_perm_0, x = var_1146)[name = tensor("transpose_18")]; - tensor v_15 = reshape(shape = var_1153, x = v_13)[name = tensor("v_15")]; + tensor var_1152 = const()[name = tensor("op_1152"), val = tensor([1, 4, 6, 64])]; + tensor q_13 = transpose(perm = q_13_perm_0, x = var_1135)[name = tensor("transpose_20")]; + tensor q_15 = reshape(shape = var_1152, x = q_13)[name = tensor("q_15")]; + tensor var_1154 = const()[name = tensor("op_1154"), val = tensor([1, 4, 6, 64])]; + tensor k_13 = transpose(perm = k_13_perm_0, x = var_1142)[name = tensor("transpose_19")]; + tensor k_15 = reshape(shape = var_1154, x = k_13)[name = tensor("k_15")]; + tensor var_1156 = const()[name = tensor("op_1156"), val = tensor([1, 4, 6, 64])]; + tensor v_13 = transpose(perm = v_13_perm_0, x = var_1149)[name = tensor("transpose_18")]; + tensor v_15 = reshape(shape = var_1156, x = v_13)[name = tensor("v_15")]; tensor mul_1_y_0 = const()[name = tensor("mul_1_y_0"), val = tensor(0x1p-3)]; tensor mul_1 = mul(x = q_15, y = mul_1_y_0)[name = tensor("mul_1")]; tensor matmul_0_transpose_y_0 = const()[name = tensor("matmul_0_transpose_y_0"), val = tensor(true)]; @@ -981,30 +980,30 @@ program(1.0) tensor attn_output_1_transpose_x_0 = const()[name = tensor("attn_output_1_transpose_x_0"), val = tensor(false)]; tensor attn_output_1_transpose_y_0 = const()[name = tensor("attn_output_1_transpose_y_0"), val = tensor(false)]; tensor attn_output_1 = matmul(transpose_x = attn_output_1_transpose_x_0, transpose_y = attn_output_1_transpose_y_0, x = softmax_0, y = v_15)[name = tensor("attn_output_1")]; - tensor var_1156 = const()[name = tensor("op_1156"), val = tensor([2, 0, 1, 3])]; - tensor var_1161 = const()[name = tensor("op_1161"), val = tensor([6, 256])]; - tensor var_1157 = transpose(perm = var_1156, x = attn_output_1)[name = tensor("transpose_17")]; - tensor attn_output_3 = reshape(shape = var_1161, x = var_1157)[name = tensor("attn_output_3")]; - tensor attn_output_5 = linear(bias = decoder_self_attn2_0_out_proj_bias, weight = decoder_self_attn2_0_out_proj_weight, x = attn_output_3)[name = tensor("linear_44")]; - tensor var_1165 = const()[name = tensor("op_1165"), val = tensor([6, 1, 256])]; - tensor attn_output_7 = reshape(shape = var_1165, x = attn_output_5)[name = tensor("attn_output_7")]; + tensor var_1159 = const()[name = tensor("op_1159"), val = tensor([2, 0, 1, 3])]; + tensor var_1164 = const()[name = tensor("op_1164"), val = tensor([6, 256])]; + tensor var_1160 = transpose(perm = var_1159, x = attn_output_1)[name = tensor("transpose_17")]; + tensor attn_output_3 = reshape(shape = var_1164, x = var_1160)[name = tensor("attn_output_3")]; + tensor attn_output_5 = linear(bias = inner_decoder_self_attn2_0_out_proj_bias, weight = inner_decoder_self_attn2_0_out_proj_weight, x = attn_output_3)[name = tensor("linear_44")]; + tensor var_1168 = const()[name = tensor("op_1168"), val = tensor([6, 1, 256])]; + tensor attn_output_7 = reshape(shape = var_1168, x = attn_output_5)[name = tensor("attn_output_7")]; tensor sa2_out_1_perm_0 = const()[name = tensor("sa2_out_1_perm_0"), val = tensor([1, 0, 2])]; tensor sa2_out_1 = transpose(perm = sa2_out_1_perm_0, x = attn_output_7)[name = tensor("transpose_16")]; - tensor input_175 = add(x = query_1, y = sa2_out_1)[name = tensor("input_175")]; - tensor input_177_axes_0 = const()[name = tensor("input_177_axes_0"), val = tensor([-1])]; - tensor input_177 = layer_norm(axes = input_177_axes_0, beta = decoder_norm21_0_bias, epsilon = var_921, gamma = decoder_norm21_0_weight, x = input_175)[name = tensor("input_177")]; - tensor input_179 = linear(bias = decoder_linear1_0_bias, weight = decoder_linear1_0_weight, x = input_177)[name = tensor("linear_45")]; - tensor input_181 = relu(x = input_179)[name = tensor("input_181")]; - tensor ffn_out_1 = linear(bias = decoder_linear2_0_bias, weight = decoder_linear2_0_weight, x = input_181)[name = tensor("linear_46")]; - tensor input_183 = add(x = input_177, y = ffn_out_1)[name = tensor("input_183")]; + tensor input_177 = add(x = query_1, y = sa2_out_1)[name = tensor("input_177")]; + tensor input_179_axes_0 = const()[name = tensor("input_179_axes_0"), val = tensor([-1])]; + tensor input_179 = layer_norm(axes = input_179_axes_0, beta = inner_decoder_norm21_0_bias, epsilon = var_35, gamma = inner_decoder_norm21_0_weight, x = input_177)[name = tensor("input_179")]; + tensor input_181 = linear(bias = inner_decoder_linear1_0_bias, weight = inner_decoder_linear1_0_weight, x = input_179)[name = tensor("linear_45")]; + tensor input_183 = relu(x = input_181)[name = tensor("input_183")]; + tensor ffn_out_1 = linear(bias = inner_decoder_linear2_0_bias, weight = inner_decoder_linear2_0_weight, x = input_183)[name = tensor("linear_46")]; + tensor input_185 = add(x = input_179, y = ffn_out_1)[name = tensor("input_185")]; tensor xt_3_axes_0 = const()[name = tensor("xt_3_axes_0"), val = tensor([-1])]; - tensor xt_3 = layer_norm(axes = xt_3_axes_0, beta = decoder_norm22_0_bias, epsilon = var_921, gamma = decoder_norm22_0_weight, x = input_183)[name = tensor("xt_3")]; - tensor var_1185 = const()[name = tensor("op_1185"), val = tensor([1, 1, 6, 256])]; - tensor x_31 = reshape(shape = var_1185, x = xt_3)[name = tensor("x_31")]; - tensor var_1187_perm_0 = const()[name = tensor("op_1187_perm_0"), val = tensor([0, 2, 1, 3])]; - tensor var_1191 = const()[name = tensor("op_1191"), val = tensor([6, 1, 256])]; - tensor var_1187 = transpose(perm = var_1187_perm_0, x = x_31)[name = tensor("transpose_15")]; - tensor x = reshape(shape = var_1191, x = var_1187)[name = tensor("x")]; + tensor xt_3 = layer_norm(axes = xt_3_axes_0, beta = inner_decoder_norm22_0_bias, epsilon = var_35, gamma = inner_decoder_norm22_0_weight, x = input_185)[name = tensor("xt_3")]; + tensor var_1188 = const()[name = tensor("op_1188"), val = tensor([1, 1, 6, 256])]; + tensor x_31 = reshape(shape = var_1188, x = xt_3)[name = tensor("x_31")]; + tensor var_1190_perm_0 = const()[name = tensor("op_1190_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_1194 = const()[name = tensor("op_1194"), val = tensor([6, 1, 256])]; + tensor var_1190 = transpose(perm = var_1190_perm_0, x = x_31)[name = tensor("transpose_15")]; + tensor x = reshape(shape = var_1194, x = var_1190)[name = tensor("x")]; tensor prev_kv_begin_0 = const()[name = tensor("prev_kv_begin_0"), val = tensor([1, 0, 0, 0, 0])]; tensor prev_kv_end_0 = const()[name = tensor("prev_kv_end_0"), val = tensor([2, 6, 4, 64, 64])]; tensor prev_kv_end_mask_0 = const()[name = tensor("prev_kv_end_mask_0"), val = tensor([false, true, true, true, true])]; @@ -1015,120 +1014,120 @@ program(1.0) tensor prev_scale_end_mask_0 = const()[name = tensor("prev_scale_end_mask_0"), val = tensor([false, true])]; tensor prev_scale_squeeze_mask_0 = const()[name = tensor("prev_scale_squeeze_mask_0"), val = tensor([true, false])]; tensor prev_scale = slice_by_index(begin = prev_scale_begin_0, end = prev_scale_end_0, end_mask = prev_scale_end_mask_0, squeeze_mask = prev_scale_squeeze_mask_0, x = dec_scale)[name = tensor("prev_scale")]; - tensor var_1199 = linear(bias = decoder_q_proj_1_bias, weight = decoder_q_proj_1_weight, x = x)[name = tensor("linear_47")]; - tensor var_1200 = const()[name = tensor("op_1200"), val = tensor([6, 1, 4, 64])]; - tensor var_1201 = reshape(shape = var_1200, x = var_1199)[name = tensor("op_1201")]; + tensor var_1202 = linear(bias = inner_decoder_q_proj_1_bias, weight = inner_decoder_q_proj_1_weight, x = x)[name = tensor("linear_47")]; + tensor var_1203 = const()[name = tensor("op_1203"), val = tensor([6, 1, 4, 64])]; + tensor var_1204 = reshape(shape = var_1203, x = var_1202)[name = tensor("op_1204")]; tensor q_17_perm_0 = const()[name = tensor("q_17_perm_0"), val = tensor([0, 2, 1, 3])]; - tensor var_1205 = linear(bias = decoder_k_proj_1_bias, weight = decoder_k_proj_1_weight, x = x)[name = tensor("linear_48")]; - tensor var_1206 = const()[name = tensor("op_1206"), val = tensor(0x1p-3)]; - tensor var_1207 = mul(x = var_1205, y = var_1206)[name = tensor("op_1207")]; - tensor var_1208 = const()[name = tensor("op_1208"), val = tensor([6, 1, 4, 64])]; - tensor var_1209 = reshape(shape = var_1208, x = var_1207)[name = tensor("op_1209")]; + tensor var_1208 = linear(bias = inner_decoder_k_proj_1_bias, weight = inner_decoder_k_proj_1_weight, x = x)[name = tensor("linear_48")]; + tensor var_1209 = const()[name = tensor("op_1209"), val = tensor(0x1p-3)]; + tensor var_1210 = mul(x = var_1208, y = var_1209)[name = tensor("op_1210")]; + tensor var_1211 = const()[name = tensor("op_1211"), val = tensor([6, 1, 4, 64])]; + tensor var_1212 = reshape(shape = var_1211, x = var_1210)[name = tensor("op_1212")]; tensor k_17_perm_0 = const()[name = tensor("k_17_perm_0"), val = tensor([0, 2, 1, 3])]; - tensor var_1213 = linear(bias = decoder_v_proj_1_bias, weight = decoder_v_proj_1_weight, x = x)[name = tensor("linear_49")]; - tensor var_1214 = const()[name = tensor("op_1214"), val = tensor([6, 1, 4, 64])]; - tensor var_1215 = reshape(shape = var_1214, x = var_1213)[name = tensor("op_1215")]; + tensor var_1216 = linear(bias = inner_decoder_v_proj_1_bias, weight = inner_decoder_v_proj_1_weight, x = x)[name = tensor("linear_49")]; + tensor var_1217 = const()[name = tensor("op_1217"), val = tensor([6, 1, 4, 64])]; + tensor var_1218 = reshape(shape = var_1217, x = var_1216)[name = tensor("op_1218")]; tensor v_17_perm_0 = const()[name = tensor("v_17_perm_0"), val = tensor([0, 2, 1, 3])]; - tensor input_187 = linear(bias = decoder_g_proj_1_bias, weight = decoder_g_proj_1_weight, x = x)[name = tensor("linear_50")]; + tensor input_189 = linear(bias = inner_decoder_g_proj_1_bias, weight = inner_decoder_g_proj_1_weight, x = x)[name = tensor("linear_50")]; tensor sqrt_s0 = sqrt(x = prev_scale)[name = tensor("sqrt_s0")]; tensor s_t = add(x = prev_scale, y = cumsum_mask_1)[name = tensor("s_t")]; tensor const_32 = const()[name = tensor("const_32"), val = tensor(0x1.fffffep+127)]; - tensor clip_3 = clip(alpha = var_926, beta = const_32, x = s_t)[name = tensor("clip_3")]; + tensor clip_3 = clip(alpha = var_28, beta = const_32, x = s_t)[name = tensor("clip_3")]; tensor sqrt_s_t = sqrt(x = clip_3)[name = tensor("sqrt_s_t")]; tensor qk_transpose_x_1 = const()[name = tensor("qk_transpose_x_1"), val = tensor(false)]; tensor qk_transpose_y_1 = const()[name = tensor("qk_transpose_y_1"), val = tensor(true)]; - tensor k_17 = transpose(perm = k_17_perm_0, x = var_1209)[name = tensor("transpose_13")]; - tensor q_17 = transpose(perm = q_17_perm_0, x = var_1201)[name = tensor("transpose_14")]; + tensor k_17 = transpose(perm = k_17_perm_0, x = var_1212)[name = tensor("transpose_13")]; + tensor q_17 = transpose(perm = q_17_perm_0, x = var_1204)[name = tensor("transpose_14")]; tensor qk = matmul(transpose_x = qk_transpose_x_1, transpose_y = qk_transpose_y_1, x = q_17, y = k_17)[name = tensor("qk")]; - tensor var_1230 = const()[name = tensor("op_1230"), val = tensor([1, 1])]; - tensor var_1231 = reshape(shape = var_1230, x = sqrt_s_t)[name = tensor("op_1231")]; - tensor M = real_div(x = var_1047, y = var_1231)[name = tensor("M")]; - tensor var_1233 = mul(x = qk, y = M)[name = tensor("op_1233")]; - tensor inner_transpose_x_0 = const()[name = tensor("inner_transpose_x_0"), val = tensor(false)]; - tensor inner_transpose_y_0 = const()[name = tensor("inner_transpose_y_0"), val = tensor(false)]; - tensor v_17 = transpose(perm = v_17_perm_0, x = var_1215)[name = tensor("transpose_12")]; - tensor inner = matmul(transpose_x = inner_transpose_x_0, transpose_y = inner_transpose_y_0, x = var_1233, y = v_17)[name = tensor("inner")]; - tensor var_1235_transpose_x_0 = const()[name = tensor("op_1235_transpose_x_0"), val = tensor(false)]; - tensor var_1235_transpose_y_0 = const()[name = tensor("op_1235_transpose_y_0"), val = tensor(false)]; - tensor var_1235 = matmul(transpose_x = var_1235_transpose_x_0, transpose_y = var_1235_transpose_y_0, x = q_17, y = prev_kv)[name = tensor("op_1235")]; - tensor var_1236 = real_div(x = sqrt_s0, y = sqrt_s_t)[name = tensor("op_1236")]; - tensor var_1237 = const()[name = tensor("op_1237"), val = tensor([1, 1, 1, 1])]; - tensor var_1238 = reshape(shape = var_1237, x = var_1236)[name = tensor("op_1238")]; - tensor cross = mul(x = var_1235, y = var_1238)[name = tensor("cross")]; - tensor out_31 = add(x = inner, y = cross)[name = tensor("out_31")]; - tensor v_masked = mul(x = v_17, y = var_1061)[name = tensor("v_masked")]; - tensor var_1244 = mul(x = prev_kv, y = sqrt_s0)[name = tensor("op_1244")]; - tensor var_1246_transpose_x_1 = const()[name = tensor("op_1246_transpose_x_1"), val = tensor(true)]; - tensor var_1246_transpose_y_1 = const()[name = tensor("op_1246_transpose_y_1"), val = tensor(false)]; - tensor var_1246 = matmul(transpose_x = var_1246_transpose_x_1, transpose_y = var_1246_transpose_y_1, x = k_17, y = v_masked)[name = tensor("op_1246")]; - tensor new_kv_unnorm = add(x = var_1244, y = var_1246)[name = tensor("new_kv_unnorm")]; - tensor new_scale = add(x = prev_scale, y = var_1069)[name = tensor("new_scale")]; + tensor var_1233 = const()[name = tensor("op_1233"), val = tensor([1, 1])]; + tensor var_1234 = reshape(shape = var_1233, x = sqrt_s_t)[name = tensor("op_1234")]; + tensor M = real_div(x = var_1050, y = var_1234)[name = tensor("M")]; + tensor var_1236 = mul(x = qk, y = M)[name = tensor("op_1236")]; + tensor inner_11_transpose_x_0 = const()[name = tensor("inner_11_transpose_x_0"), val = tensor(false)]; + tensor inner_11_transpose_y_0 = const()[name = tensor("inner_11_transpose_y_0"), val = tensor(false)]; + tensor v_17 = transpose(perm = v_17_perm_0, x = var_1218)[name = tensor("transpose_12")]; + tensor inner_11 = matmul(transpose_x = inner_11_transpose_x_0, transpose_y = inner_11_transpose_y_0, x = var_1236, y = v_17)[name = tensor("inner_11")]; + tensor var_1238_transpose_x_0 = const()[name = tensor("op_1238_transpose_x_0"), val = tensor(false)]; + tensor var_1238_transpose_y_0 = const()[name = tensor("op_1238_transpose_y_0"), val = tensor(false)]; + tensor var_1238 = matmul(transpose_x = var_1238_transpose_x_0, transpose_y = var_1238_transpose_y_0, x = q_17, y = prev_kv)[name = tensor("op_1238")]; + tensor var_1239 = real_div(x = sqrt_s0, y = sqrt_s_t)[name = tensor("op_1239")]; + tensor var_1240 = const()[name = tensor("op_1240"), val = tensor([1, 1, 1, 1])]; + tensor var_1241 = reshape(shape = var_1240, x = var_1239)[name = tensor("op_1241")]; + tensor cross = mul(x = var_1238, y = var_1241)[name = tensor("cross")]; + tensor out_31 = add(x = inner_11, y = cross)[name = tensor("out_31")]; + tensor v_masked = mul(x = v_17, y = var_1064)[name = tensor("v_masked")]; + tensor var_1247 = mul(x = prev_kv, y = sqrt_s0)[name = tensor("op_1247")]; + tensor var_1249_transpose_x_1 = const()[name = tensor("op_1249_transpose_x_1"), val = tensor(true)]; + tensor var_1249_transpose_y_1 = const()[name = tensor("op_1249_transpose_y_1"), val = tensor(false)]; + tensor var_1249 = matmul(transpose_x = var_1249_transpose_x_1, transpose_y = var_1249_transpose_y_1, x = k_17, y = v_masked)[name = tensor("op_1249")]; + tensor new_kv_unnorm = add(x = var_1247, y = var_1249)[name = tensor("new_kv_unnorm")]; + tensor new_scale = add(x = prev_scale, y = var_1072)[name = tensor("new_scale")]; tensor const_33 = const()[name = tensor("const_33"), val = tensor(0x1.fffffep+127)]; - tensor clip_4 = clip(alpha = var_926, beta = const_33, x = new_scale)[name = tensor("clip_4")]; + tensor clip_4 = clip(alpha = var_28, beta = const_33, x = new_scale)[name = tensor("clip_4")]; tensor sqrt_new_scale = sqrt(x = clip_4)[name = tensor("sqrt_new_scale")]; tensor nkv = real_div(x = new_kv_unnorm, y = sqrt_new_scale)[name = tensor("nkv")]; - tensor var_1255_perm_0 = const()[name = tensor("op_1255_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_1258_perm_0 = const()[name = tensor("op_1258_perm_0"), val = tensor([0, 2, 1, 3])]; tensor out_33_axes_0 = const()[name = tensor("out_33_axes_0"), val = tensor([-1])]; - tensor var_1255 = transpose(perm = var_1255_perm_0, x = out_31)[name = tensor("transpose_11")]; - tensor out_33 = layer_norm(axes = out_33_axes_0, epsilon = var_923, x = var_1255)[name = tensor("out_33")]; - tensor var_1259 = const()[name = tensor("op_1259"), val = tensor([6, 1, 256])]; - tensor out = reshape(shape = var_1259, x = out_33)[name = tensor("out")]; - tensor var_1261 = silu(x = input_187)[name = tensor("op_1261")]; - tensor input_189 = mul(x = var_1261, y = out)[name = tensor("input_189")]; - tensor ret_out = linear(bias = decoder_out_proj_1_bias, weight = decoder_out_proj_1_weight, x = input_189)[name = tensor("linear_51")]; - tensor input_191 = add(x = x, y = ret_out)[name = tensor("input_191")]; + tensor var_1258 = transpose(perm = var_1258_perm_0, x = out_31)[name = tensor("transpose_11")]; + tensor out_33 = layer_norm(axes = out_33_axes_0, epsilon = var_43, x = var_1258)[name = tensor("out_33")]; + tensor var_1262 = const()[name = tensor("op_1262"), val = tensor([6, 1, 256])]; + tensor out = reshape(shape = var_1262, x = out_33)[name = tensor("out")]; + tensor var_1264 = silu(x = input_189)[name = tensor("op_1264")]; + tensor input_191 = mul(x = var_1264, y = out)[name = tensor("input_191")]; + tensor ret_out = linear(bias = inner_decoder_out_proj_1_bias, weight = inner_decoder_out_proj_1_weight, x = input_191)[name = tensor("linear_51")]; + tensor input_193 = add(x = x, y = ret_out)[name = tensor("input_193")]; tensor xt_5_axes_0 = const()[name = tensor("xt_5_axes_0"), val = tensor([-1])]; - tensor xt_5 = layer_norm(axes = xt_5_axes_0, beta = decoder_norm11_1_bias, epsilon = var_921, gamma = decoder_norm11_1_weight, x = input_191)[name = tensor("xt_5")]; - tensor var_1271 = const()[name = tensor("op_1271"), val = tensor([1, 6, 1, 256])]; - tensor var_1272 = reshape(shape = var_1271, x = xt_5)[name = tensor("op_1272")]; - tensor var_1273_perm_0 = const()[name = tensor("op_1273_perm_0"), val = tensor([0, 2, 1, 3])]; - tensor var_1276 = const()[name = tensor("op_1276"), val = tensor([1, 6, 256])]; - tensor var_1273 = transpose(perm = var_1273_perm_0, x = var_1272)[name = tensor("transpose_10")]; - tensor query_5 = reshape(shape = var_1276, x = var_1273)[name = tensor("query_5")]; + tensor xt_5 = layer_norm(axes = xt_5_axes_0, beta = inner_decoder_norm11_1_bias, epsilon = var_35, gamma = inner_decoder_norm11_1_weight, x = input_193)[name = tensor("xt_5")]; + tensor var_1274 = const()[name = tensor("op_1274"), val = tensor([1, 6, 1, 256])]; + tensor var_1275 = reshape(shape = var_1274, x = xt_5)[name = tensor("op_1275")]; + tensor var_1276_perm_0 = const()[name = tensor("op_1276_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_1279 = const()[name = tensor("op_1279"), val = tensor([1, 6, 256])]; + tensor var_1276 = transpose(perm = var_1276_perm_0, x = var_1275)[name = tensor("transpose_10")]; + tensor query_5 = reshape(shape = var_1279, x = var_1276)[name = tensor("query_5")]; tensor query_perm_0 = const()[name = tensor("query_perm_0"), val = tensor([1, 0, 2])]; tensor query = transpose(perm = query_perm_0, x = query_5)[name = tensor("transpose_9")]; - tensor var_1299 = linear(bias = decoder_self_attn2_1_in_proj_bias, weight = decoder_self_attn2_1_in_proj_weight, x = query)[name = tensor("linear_52")]; + tensor var_1302 = linear(bias = inner_decoder_self_attn2_1_in_proj_bias, weight = inner_decoder_self_attn2_1_in_proj_weight, x = query)[name = tensor("linear_52")]; tensor concat_2 = const()[name = tensor("concat_2"), val = tensor([6, 1, 3, 256])]; - tensor var_1301 = reshape(shape = concat_2, x = var_1299)[name = tensor("op_1301")]; - tensor var_1302_axes_0 = const()[name = tensor("op_1302_axes_0"), val = tensor([0])]; - tensor var_1302 = expand_dims(axes = var_1302_axes_0, x = var_1301)[name = tensor("op_1302")]; - tensor var_1303_perm_0 = const()[name = tensor("op_1303_perm_0"), val = tensor([-2, 1, 2, 0, 4])]; - tensor var_1304_axes_0 = const()[name = tensor("op_1304_axes_0"), val = tensor([-2])]; - tensor var_1303 = transpose(perm = var_1303_perm_0, x = var_1302)[name = tensor("transpose_8")]; - tensor var_1304 = squeeze(axes = var_1304_axes_0, x = var_1303)[name = tensor("op_1304")]; + tensor var_1304 = reshape(shape = concat_2, x = var_1302)[name = tensor("op_1304")]; + tensor var_1305_axes_0 = const()[name = tensor("op_1305_axes_0"), val = tensor([0])]; + tensor var_1305 = expand_dims(axes = var_1305_axes_0, x = var_1304)[name = tensor("op_1305")]; + tensor var_1306_perm_0 = const()[name = tensor("op_1306_perm_0"), val = tensor([-2, 1, 2, 0, 4])]; + tensor var_1307_axes_0 = const()[name = tensor("op_1307_axes_0"), val = tensor([-2])]; + tensor var_1306 = transpose(perm = var_1306_perm_0, x = var_1305)[name = tensor("transpose_8")]; + tensor var_1307 = squeeze(axes = var_1307_axes_0, x = var_1306)[name = tensor("op_1307")]; tensor q_19_begin_0 = const()[name = tensor("q_19_begin_0"), val = tensor([0, 0, 0, 0])]; tensor q_19_end_0 = const()[name = tensor("q_19_end_0"), val = tensor([1, 6, 1, 256])]; tensor q_19_end_mask_0 = const()[name = tensor("q_19_end_mask_0"), val = tensor([false, true, true, true])]; tensor q_19_squeeze_mask_0 = const()[name = tensor("q_19_squeeze_mask_0"), val = tensor([true, false, false, false])]; - tensor q_19 = slice_by_index(begin = q_19_begin_0, end = q_19_end_0, end_mask = q_19_end_mask_0, squeeze_mask = q_19_squeeze_mask_0, x = var_1304)[name = tensor("q_19")]; + tensor q_19 = slice_by_index(begin = q_19_begin_0, end = q_19_end_0, end_mask = q_19_end_mask_0, squeeze_mask = q_19_squeeze_mask_0, x = var_1307)[name = tensor("q_19")]; tensor k_19_begin_0 = const()[name = tensor("k_19_begin_0"), val = tensor([1, 0, 0, 0])]; tensor k_19_end_0 = const()[name = tensor("k_19_end_0"), val = tensor([2, 6, 1, 256])]; tensor k_19_end_mask_0 = const()[name = tensor("k_19_end_mask_0"), val = tensor([false, true, true, true])]; tensor k_19_squeeze_mask_0 = const()[name = tensor("k_19_squeeze_mask_0"), val = tensor([true, false, false, false])]; - tensor k_19 = slice_by_index(begin = k_19_begin_0, end = k_19_end_0, end_mask = k_19_end_mask_0, squeeze_mask = k_19_squeeze_mask_0, x = var_1304)[name = tensor("k_19")]; + tensor k_19 = slice_by_index(begin = k_19_begin_0, end = k_19_end_0, end_mask = k_19_end_mask_0, squeeze_mask = k_19_squeeze_mask_0, x = var_1307)[name = tensor("k_19")]; tensor v_19_begin_0 = const()[name = tensor("v_19_begin_0"), val = tensor([2, 0, 0, 0])]; tensor v_19_end_0 = const()[name = tensor("v_19_end_0"), val = tensor([3, 6, 1, 256])]; tensor v_19_end_mask_0 = const()[name = tensor("v_19_end_mask_0"), val = tensor([false, true, true, true])]; tensor v_19_squeeze_mask_0 = const()[name = tensor("v_19_squeeze_mask_0"), val = tensor([true, false, false, false])]; - tensor v_19 = slice_by_index(begin = v_19_begin_0, end = v_19_end_0, end_mask = v_19_end_mask_0, squeeze_mask = v_19_squeeze_mask_0, x = var_1304)[name = tensor("v_19")]; - tensor var_1312 = const()[name = tensor("op_1312"), val = tensor([6, 4, 64])]; - tensor var_1313 = reshape(shape = var_1312, x = q_19)[name = tensor("op_1313")]; + tensor v_19 = slice_by_index(begin = v_19_begin_0, end = v_19_end_0, end_mask = v_19_end_mask_0, squeeze_mask = v_19_squeeze_mask_0, x = var_1307)[name = tensor("v_19")]; + tensor var_1315 = const()[name = tensor("op_1315"), val = tensor([6, 4, 64])]; + tensor var_1316 = reshape(shape = var_1315, x = q_19)[name = tensor("op_1316")]; tensor q_21_perm_0 = const()[name = tensor("q_21_perm_0"), val = tensor([1, 0, 2])]; - tensor var_1319 = const()[name = tensor("op_1319"), val = tensor([6, 4, 64])]; - tensor var_1320 = reshape(shape = var_1319, x = k_19)[name = tensor("op_1320")]; + tensor var_1322 = const()[name = tensor("op_1322"), val = tensor([6, 4, 64])]; + tensor var_1323 = reshape(shape = var_1322, x = k_19)[name = tensor("op_1323")]; tensor k_21_perm_0 = const()[name = tensor("k_21_perm_0"), val = tensor([1, 0, 2])]; - tensor var_1326 = const()[name = tensor("op_1326"), val = tensor([6, 4, 64])]; - tensor var_1327 = reshape(shape = var_1326, x = v_19)[name = tensor("op_1327")]; + tensor var_1329 = const()[name = tensor("op_1329"), val = tensor([6, 4, 64])]; + tensor var_1330 = reshape(shape = var_1329, x = v_19)[name = tensor("op_1330")]; tensor v_21_perm_0 = const()[name = tensor("v_21_perm_0"), val = tensor([1, 0, 2])]; - tensor var_1330 = const()[name = tensor("op_1330"), val = tensor([1, 4, 6, 64])]; - tensor q_21 = transpose(perm = q_21_perm_0, x = var_1313)[name = tensor("transpose_7")]; - tensor q = reshape(shape = var_1330, x = q_21)[name = tensor("q")]; - tensor var_1332 = const()[name = tensor("op_1332"), val = tensor([1, 4, 6, 64])]; - tensor k_21 = transpose(perm = k_21_perm_0, x = var_1320)[name = tensor("transpose_6")]; - tensor k = reshape(shape = var_1332, x = k_21)[name = tensor("k")]; - tensor var_1334 = const()[name = tensor("op_1334"), val = tensor([1, 4, 6, 64])]; - tensor v_21 = transpose(perm = v_21_perm_0, x = var_1327)[name = tensor("transpose_5")]; - tensor v = reshape(shape = var_1334, x = v_21)[name = tensor("v")]; + tensor var_1333 = const()[name = tensor("op_1333"), val = tensor([1, 4, 6, 64])]; + tensor q_21 = transpose(perm = q_21_perm_0, x = var_1316)[name = tensor("transpose_7")]; + tensor q = reshape(shape = var_1333, x = q_21)[name = tensor("q")]; + tensor var_1335 = const()[name = tensor("op_1335"), val = tensor([1, 4, 6, 64])]; + tensor k_21 = transpose(perm = k_21_perm_0, x = var_1323)[name = tensor("transpose_6")]; + tensor k = reshape(shape = var_1335, x = k_21)[name = tensor("k")]; + tensor var_1337 = const()[name = tensor("op_1337"), val = tensor([1, 4, 6, 64])]; + tensor v_21 = transpose(perm = v_21_perm_0, x = var_1330)[name = tensor("transpose_5")]; + tensor v = reshape(shape = var_1337, x = v_21)[name = tensor("v")]; tensor mul_3_y_0 = const()[name = tensor("mul_3_y_0"), val = tensor(0x1p-3)]; tensor mul_3 = mul(x = q, y = mul_3_y_0)[name = tensor("mul_3")]; tensor matmul_1_transpose_y_0 = const()[name = tensor("matmul_1_transpose_y_0"), val = tensor(true)]; @@ -1139,34 +1138,34 @@ program(1.0) tensor attn_output_9_transpose_x_0 = const()[name = tensor("attn_output_9_transpose_x_0"), val = tensor(false)]; tensor attn_output_9_transpose_y_0 = const()[name = tensor("attn_output_9_transpose_y_0"), val = tensor(false)]; tensor attn_output_9 = matmul(transpose_x = attn_output_9_transpose_x_0, transpose_y = attn_output_9_transpose_y_0, x = softmax_1, y = v)[name = tensor("attn_output_9")]; - tensor var_1337 = const()[name = tensor("op_1337"), val = tensor([2, 0, 1, 3])]; - tensor var_1342 = const()[name = tensor("op_1342"), val = tensor([6, 256])]; - tensor var_1338 = transpose(perm = var_1337, x = attn_output_9)[name = tensor("transpose_4")]; - tensor attn_output_11 = reshape(shape = var_1342, x = var_1338)[name = tensor("attn_output_11")]; - tensor attn_output_13 = linear(bias = decoder_self_attn2_1_out_proj_bias, weight = decoder_self_attn2_1_out_proj_weight, x = attn_output_11)[name = tensor("linear_53")]; - tensor var_1346 = const()[name = tensor("op_1346"), val = tensor([6, 1, 256])]; - tensor attn_output = reshape(shape = var_1346, x = attn_output_13)[name = tensor("attn_output")]; + tensor var_1340 = const()[name = tensor("op_1340"), val = tensor([2, 0, 1, 3])]; + tensor var_1345 = const()[name = tensor("op_1345"), val = tensor([6, 256])]; + tensor var_1341 = transpose(perm = var_1340, x = attn_output_9)[name = tensor("transpose_4")]; + tensor attn_output_11 = reshape(shape = var_1345, x = var_1341)[name = tensor("attn_output_11")]; + tensor attn_output_13 = linear(bias = inner_decoder_self_attn2_1_out_proj_bias, weight = inner_decoder_self_attn2_1_out_proj_weight, x = attn_output_11)[name = tensor("linear_53")]; + tensor var_1349 = const()[name = tensor("op_1349"), val = tensor([6, 1, 256])]; + tensor attn_output = reshape(shape = var_1349, x = attn_output_13)[name = tensor("attn_output")]; tensor sa2_out_perm_0 = const()[name = tensor("sa2_out_perm_0"), val = tensor([1, 0, 2])]; tensor sa2_out = transpose(perm = sa2_out_perm_0, x = attn_output)[name = tensor("transpose_3")]; - tensor input_193 = add(x = query_5, y = sa2_out)[name = tensor("input_193")]; - tensor input_195_axes_0 = const()[name = tensor("input_195_axes_0"), val = tensor([-1])]; - tensor input_195 = layer_norm(axes = input_195_axes_0, beta = decoder_norm21_1_bias, epsilon = var_921, gamma = decoder_norm21_1_weight, x = input_193)[name = tensor("input_195")]; - tensor input_197 = linear(bias = decoder_linear1_1_bias, weight = decoder_linear1_1_weight, x = input_195)[name = tensor("linear_54")]; - tensor input_199 = relu(x = input_197)[name = tensor("input_199")]; - tensor ffn_out = linear(bias = decoder_linear2_1_bias, weight = decoder_linear2_1_weight, x = input_199)[name = tensor("linear_55")]; - tensor input_201 = add(x = input_195, y = ffn_out)[name = tensor("input_201")]; + tensor input_195 = add(x = query_5, y = sa2_out)[name = tensor("input_195")]; + tensor input_197_axes_0 = const()[name = tensor("input_197_axes_0"), val = tensor([-1])]; + tensor input_197 = layer_norm(axes = input_197_axes_0, beta = inner_decoder_norm21_1_bias, epsilon = var_35, gamma = inner_decoder_norm21_1_weight, x = input_195)[name = tensor("input_197")]; + tensor input_199 = linear(bias = inner_decoder_linear1_1_bias, weight = inner_decoder_linear1_1_weight, x = input_197)[name = tensor("linear_54")]; + tensor input_201 = relu(x = input_199)[name = tensor("input_201")]; + tensor ffn_out = linear(bias = inner_decoder_linear2_1_bias, weight = inner_decoder_linear2_1_weight, x = input_201)[name = tensor("linear_55")]; + tensor input_203 = add(x = input_197, y = ffn_out)[name = tensor("input_203")]; tensor xt_axes_0 = const()[name = tensor("xt_axes_0"), val = tensor([-1])]; - tensor xt = layer_norm(axes = xt_axes_0, beta = decoder_norm22_1_bias, epsilon = var_921, gamma = decoder_norm22_1_weight, x = input_201)[name = tensor("xt")]; - tensor var_1366 = const()[name = tensor("op_1366"), val = tensor([1, 1, 6, 256])]; - tensor input = reshape(shape = var_1366, x = xt)[name = tensor("input")]; - tensor var_1368 = const()[name = tensor("op_1368"), val = tensor([-1])]; - tensor var_1369 = reduce_l2_norm(axes = var_1368, keep_dims = var_924, x = input)[name = tensor("op_1369")]; + tensor xt = layer_norm(axes = xt_axes_0, beta = inner_decoder_norm22_1_bias, epsilon = var_35, gamma = inner_decoder_norm22_1_weight, x = input_203)[name = tensor("xt")]; + tensor var_1369 = const()[name = tensor("op_1369"), val = tensor([1, 1, 6, 256])]; + tensor input = reshape(shape = var_1369, x = xt)[name = tensor("input")]; + tensor var_1371 = const()[name = tensor("op_1371"), val = tensor([-1])]; + tensor var_1372 = reduce_l2_norm(axes = var_1371, keep_dims = var_34, x = input)[name = tensor("op_1372")]; tensor const_42 = const()[name = tensor("const_42"), val = tensor(0x1.fffffep+127)]; - tensor clip_5 = clip(alpha = var_916, beta = const_42, x = var_1369)[name = tensor("clip_5")]; - tensor var_1371 = real_div(x = input, y = clip_5)[name = tensor("op_1371")]; + tensor clip_5 = clip(alpha = var_48, beta = const_42, x = var_1372)[name = tensor("clip_5")]; + tensor var_1374 = real_div(x = input, y = clip_5)[name = tensor("op_1374")]; tensor transpose_1_perm_0 = const()[name = tensor("transpose_1_perm_0"), val = tensor([0, 1, 3, 2])]; tensor concat_7 = const()[name = tensor("concat_7"), val = tensor([1, 256, 6])]; - tensor transpose_1 = transpose(perm = transpose_1_perm_0, x = var_1371)[name = tensor("transpose_2")]; + tensor transpose_1 = transpose(perm = transpose_1_perm_0, x = var_1374)[name = tensor("transpose_2")]; tensor reshape_1 = reshape(shape = concat_7, x = transpose_1)[name = tensor("reshape_1")]; tensor matmul_2_transpose_x_0 = const()[name = tensor("matmul_2_transpose_x_0"), val = tensor(false)]; tensor matmul_2_transpose_y_0 = const()[name = tensor("matmul_2_transpose_y_0"), val = tensor(false)]; @@ -1175,10 +1174,10 @@ program(1.0) tensor output_end_0 = const()[name = tensor("output_end_0"), val = tensor([1, 1, 5])]; tensor output_end_mask_0 = const()[name = tensor("output_end_mask_0"), val = tensor([true, true, false])]; tensor output = slice_by_index(begin = output_begin_0, end = output_end_0, end_mask = output_end_mask_0, x = matmul_2)[name = tensor("output")]; - tensor probs = sigmoid(x = output)[name = tensor("op_1375")]; - tensor var_1377_axis_0 = const()[name = tensor("op_1377_axis_0"), val = tensor(0)]; - tensor dec_kv_new = stack(axis = var_1377_axis_0, values = (var_1073, nkv))[name = tensor("op_1377")]; - tensor var_1379_axis_0 = const()[name = tensor("op_1379_axis_0"), val = tensor(0)]; - tensor dec_scale_new = stack(axis = var_1379_axis_0, values = (new_scale_9, new_scale))[name = tensor("op_1379")]; + tensor probs = sigmoid(x = output)[name = tensor("op_1378")]; + tensor var_1380_axis_0 = const()[name = tensor("op_1380_axis_0"), val = tensor(0)]; + tensor dec_kv_new = stack(axis = var_1380_axis_0, values = (var_1076, nkv))[name = tensor("op_1380")]; + tensor var_1382_axis_0 = const()[name = tensor("op_1382_axis_0"), val = tensor(0)]; + tensor dec_scale_new = stack(axis = var_1382_axis_0, values = (new_scale_9, new_scale))[name = tensor("op_1382")]; } -> (probs, enc_kv_new, enc_scale_new, enc_conv_cache_new, cnn_window_new, dec_kv_new, dec_scale_new); } \ No newline at end of file diff --git a/optimized/ami/100ms/ls_eend_ami_100ms.mlpackage/Data/com.apple.CoreML/model.mlmodel b/optimized/ami/100ms/ls_eend_ami_100ms.mlpackage/Data/com.apple.CoreML/model.mlmodel index 23b9a3151223c06a795711a72944fec0a2d73987..dc361a040c87fc3818df1830df2c72f647d6c270 100644 --- a/optimized/ami/100ms/ls_eend_ami_100ms.mlpackage/Data/com.apple.CoreML/model.mlmodel +++ b/optimized/ami/100ms/ls_eend_ami_100ms.mlpackage/Data/com.apple.CoreML/model.mlmodel @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:2d99472f0bd00e091613210906f7fb7be42e92f6019968246af665c07dab02a7 -size 171348 +oid sha256:fc177110db14e37e532be41bb47accb450f1ce89bb7876b675b2d4a25d797e1e +size 175266 diff --git a/optimized/ami/100ms/ls_eend_ami_100ms.mlpackage/Manifest.json b/optimized/ami/100ms/ls_eend_ami_100ms.mlpackage/Manifest.json index 6a6df0d6218423e5cf5828262039e34f982ae069..600ad3c5b36c6613b59be15141c9242bb273eb53 100644 --- a/optimized/ami/100ms/ls_eend_ami_100ms.mlpackage/Manifest.json +++ b/optimized/ami/100ms/ls_eend_ami_100ms.mlpackage/Manifest.json @@ -1,18 +1,18 @@ { "fileFormatVersion": "1.0.0", "itemInfoEntries": { - "25D17A51-8136-403A-8E7C-65E586D88F62": { + "3ADB4906-F6AE-45F6-8FBA-E66FCC7A8291": { "author": "com.apple.CoreML", "description": "CoreML Model Weights", "name": "weights", "path": "com.apple.CoreML/weights" }, - "966E268D-F501-43F0-BD33-968439381A4D": { + "403AED6F-5C8E-400B-80AA-18BE809ED8A3": { "author": "com.apple.CoreML", "description": "CoreML Model Specification", "name": "model.mlmodel", "path": "com.apple.CoreML/model.mlmodel" } }, - "rootModelIdentifier": "966E268D-F501-43F0-BD33-968439381A4D" + "rootModelIdentifier": "403AED6F-5C8E-400B-80AA-18BE809ED8A3" } diff --git a/optimized/ami/200ms/ls_eend_ami_200ms.mlmodelc/analytics/coremldata.bin b/optimized/ami/200ms/ls_eend_ami_200ms.mlmodelc/analytics/coremldata.bin index dbf2843795d71fc2c8081b28675d1065af2fc06c..38434f198256b9bb1ff20657e279c5ad2ef9ed87 100644 --- a/optimized/ami/200ms/ls_eend_ami_200ms.mlmodelc/analytics/coremldata.bin +++ b/optimized/ami/200ms/ls_eend_ami_200ms.mlmodelc/analytics/coremldata.bin @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:c6a81fd874a82f12a3d9273121dfc5cf2c7b7fc3ab75940416ce700d6c464715 +oid sha256:879a50af51ba6bf2014083aac4b8757dd61926e4ea7e84b03b88c8bd34bf19a7 size 243 diff --git a/optimized/ami/200ms/ls_eend_ami_200ms.mlmodelc/coremldata.bin b/optimized/ami/200ms/ls_eend_ami_200ms.mlmodelc/coremldata.bin index c2e0a156964f1117ec1437b8f2af288e3cae4b6a..33a973adb4aaa20c921d030786fdd9741671d4e6 100644 --- a/optimized/ami/200ms/ls_eend_ami_200ms.mlmodelc/coremldata.bin +++ b/optimized/ami/200ms/ls_eend_ami_200ms.mlmodelc/coremldata.bin @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:1ff7a04daff876566a5951933f1afdcf6a2bc73cf8214ddcb95845fb83a7ad66 -size 1292 +oid sha256:db2f6d6a11db778787a47cac87d5933e6c831d13f92e280b3ba1d11f75f4ba6c +size 1395 diff --git a/optimized/ami/200ms/ls_eend_ami_200ms.mlmodelc/metadata.json b/optimized/ami/200ms/ls_eend_ami_200ms.mlmodelc/metadata.json index fceb4211edaa544fa740da0dbf819fbec3b8ddf1..b44cfc54249ccb12efacb97f32bc8fa49ad4ee48 100644 --- a/optimized/ami/200ms/ls_eend_ami_200ms.mlmodelc/metadata.json +++ b/optimized/ami/200ms/ls_eend_ami_200ms.mlmodelc/metadata.json @@ -1,7 +1,7 @@ [ { "metadataOutputVersion" : "3.0", - "shortDescription" : "LS-EEND AMI streaming diarizer (pipeline, T=2, max_speakers=4)", + "shortDescription" : "LS-EEND AMI streaming diarizer (pipeline, T=2, max_speakers=4, layout=raw_mel, cu=all)", "outputSchema" : [ { "hasShapeFlexibility" : "0", @@ -81,7 +81,7 @@ "specificationVersion" : 8, "mlProgramOperationTypeHistogram" : { "Ios16.reduceL2Norm" : 2, - "Ios17.reshape" : 68, + "Ios17.reshape" : 69, "Ios16.softmax" : 2, "Ios17.matmul" : 29, "Ios17.transpose" : 57, @@ -89,7 +89,7 @@ "Ios17.expandDims" : 3, "Ios17.add" : 46, "Ios16.sigmoid" : 5, - "Ios17.sliceByIndex" : 48, + "Ios17.sliceByIndex" : 50, "Tile" : 1, "Ios16.reduceSum" : 1, "Ios17.squeeze" : 2, @@ -101,7 +101,7 @@ "Ios16.silu" : 18, "Ios17.realDiv" : 20, "Ios17.linear" : 56, - "Stack" : 5, + "Stack" : 6, "Ios17.concat" : 14, "Ios16.relu" : 2, "Ios16.cumsum" : 1, @@ -128,9 +128,9 @@ "hasShapeFlexibility" : "0", "isOptional" : "0", "dataType" : "Float32", - "formattedType" : "MultiArray (Float32 1 × 2 × 345)", + "formattedType" : "MultiArray (Float32 1 × 25 × 23)", "shortDescription" : "", - "shape" : "[1, 2, 345]", + "shape" : "[1, 25, 23]", "name" : "features", "type" : "MultiArray" }, @@ -206,8 +206,8 @@ } ], "userDefinedMetadata" : { - "com.github.apple.coremltools.conversion_date" : "2026-04-16", - "config" : "{\"model_name\": \"ami\", \"model_label\": \"AMI\", \"variant\": \"pipeline\", \"chunk_size\": 2, \"step_duration_ms\": 200, \"frame_duration_ms\": 100, \"frame_duration_seconds\": 0.1, \"max_speakers\": 4, \"max_nspks\": 6, \"n_units\": 256, \"n_heads\": 4, \"enc_n_layers\": 4, \"dec_n_layers\": 2, \"conv_kernel_size\": 16, \"conv_delay\": 9, \"sample_rate\": 8000, \"win_length\": 200, \"hop_length\": 80, \"n_mels\": 23, \"context_size\": 7, \"subsampling\": 10, \"feat_type\": \"logmel23_cummn\", \"pure_roll\": true}", + "com.github.apple.coremltools.conversion_date" : "2026-04-18", + "config" : "{\"model_name\": \"ami\", \"model_label\": \"AMI\", \"variant\": \"pipeline\", \"chunk_size\": 2, \"step_duration_ms\": 200, \"frame_duration_ms\": 100, \"frame_duration_seconds\": 0.1, \"max_speakers\": 4, \"max_nspks\": 6, \"n_units\": 256, \"n_heads\": 4, \"enc_n_layers\": 4, \"dec_n_layers\": 2, \"conv_kernel_size\": 16, \"conv_delay\": 9, \"sample_rate\": 8000, \"win_length\": 200, \"hop_length\": 80, \"n_mels\": 23, \"context_size\": 7, \"subsampling\": 10, \"feat_type\": \"logmel23_cummn\", \"pure_roll\": true, \"input_layout\": \"raw_mel\", \"compute_units_export\": \"all\", \"raw_mel_length\": 25}", "com.github.apple.coremltools.source" : "torch==2.6.0", "com.github.apple.coremltools.version" : "9.0", "com.github.apple.coremltools.source_dialect" : "TorchScript" diff --git a/optimized/ami/200ms/ls_eend_ami_200ms.mlmodelc/model.mil b/optimized/ami/200ms/ls_eend_ami_200ms.mlmodelc/model.mil index 06fd0af3c049963c87924d0249246fe8af7890b0..4e4bf7dc4c9439ee3979e99bab1dbb11b541124b 100644 --- a/optimized/ami/200ms/ls_eend_ami_200ms.mlmodelc/model.mil +++ b/optimized/ami/200ms/ls_eend_ami_200ms.mlmodelc/model.mil @@ -1,234 +1,248 @@ program(1.0) [buildInfo = dict, tensor>({{"coremlc-component-MIL", "3510.2.1"}, {"coremlc-version", "3500.32.1"}, {"coremltools-component-torch", "2.6.0"}, {"coremltools-source-dialect", "TorchScript"}, {"coremltools-version", "9.0"}})] { - func main(tensor cnn_window, tensor dec_kv, tensor dec_scale, tensor enc_conv_cache, tensor enc_kv, tensor enc_scale, tensor features, tensor valid_mask) { - tensor encoder_cnn_bias = const()[name = tensor("encoder_cnn_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(64)))]; - tensor encoder_cnn_weight = const()[name = tensor("encoder_cnn_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1152)))]; - tensor encoder__causal_mask = const()[name = tensor("encoder__causal_mask"), val = tensor([[0x1p+0, 0x0p+0], [0x1p+0, 0x1p+0]])]; - tensor encoder__t_index = const()[name = tensor("encoder__t_index"), val = tensor([0x1p+0, 0x1p+1])]; - tensor encoder_input_projection_linear_bias = const()[name = tensor("encoder_input_projection_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4981952)))]; - tensor encoder_input_projection_linear_weight = const()[name = tensor("encoder_input_projection_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4983040)))]; - tensor encoder_pre_layer_norm_bias = const()[name = tensor("encoder_pre_layer_norm_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(5336384)))]; - tensor encoder_pre_layer_norm_weight = const()[name = tensor("encoder_pre_layer_norm_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(5337472)))]; - tensor encoder_ffn1_0_module_sequential_0_bias = const()[name = tensor("encoder_ffn1_0_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(5338560)))]; - tensor encoder_ffn1_0_module_sequential_0_weight = const()[name = tensor("encoder_ffn1_0_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(5339648)))]; - tensor encoder_ffn1_0_module_sequential_1_linear_bias = const()[name = tensor("encoder_ffn1_0_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(5340736)))]; - tensor encoder_ffn1_0_module_sequential_1_linear_weight = const()[name = tensor("encoder_ffn1_0_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(5344896)))]; - tensor encoder_ffn1_0_module_sequential_4_linear_bias = const()[name = tensor("encoder_ffn1_0_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(6393536)))]; - tensor encoder_ffn1_0_module_sequential_4_linear_weight = const()[name = tensor("encoder_ffn1_0_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(6394624)))]; - tensor encoder_ret_lns_0_bias = const()[name = tensor("encoder_ret_lns_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7443264)))]; - tensor encoder_ret_lns_0_weight = const()[name = tensor("encoder_ret_lns_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7444352)))]; - tensor encoder_q_proj_0_bias = const()[name = tensor("encoder_q_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7445440)))]; - tensor encoder_q_proj_0_weight = const()[name = tensor("encoder_q_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7446528)))]; - tensor encoder_k_proj_0_bias = const()[name = tensor("encoder_k_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7708736)))]; - tensor encoder_k_proj_0_weight = const()[name = tensor("encoder_k_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7709824)))]; - tensor encoder_v_proj_0_bias = const()[name = tensor("encoder_v_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7972032)))]; - tensor encoder_v_proj_0_weight = const()[name = tensor("encoder_v_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7973120)))]; - tensor encoder_g_proj_0_bias = const()[name = tensor("encoder_g_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8235328)))]; - tensor encoder_g_proj_0_weight = const()[name = tensor("encoder_g_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8236416)))]; - tensor encoder_out_proj_0_bias = const()[name = tensor("encoder_out_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8498624)))]; - tensor encoder_out_proj_0_weight = const()[name = tensor("encoder_out_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8499712)))]; - tensor encoder_conv_module_0_sequential_0_bias = const()[name = tensor("encoder_conv_module_0_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8761920)))]; - tensor encoder_conv_module_0_sequential_0_weight = const()[name = tensor("encoder_conv_module_0_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8763008)))]; - tensor encoder_conv_module_0_sequential_2_conv_bias = const()[name = tensor("encoder_conv_module_0_sequential_2_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8764096)))]; - tensor encoder_conv_module_0_sequential_2_conv_weight = const()[name = tensor("encoder_conv_module_0_sequential_2_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8766208)))]; - tensor encoder_conv_module_0_sequential_4_conv_weight = const()[name = tensor("encoder_conv_module_0_sequential_4_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9290560)))]; - tensor encoder_conv_module_0_sequential_5_running_var = const()[name = tensor("encoder_conv_module_0_sequential_5_running_var"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9307008)))]; - tensor encoder_conv_module_0_sequential_5_running_mean = const()[name = tensor("encoder_conv_module_0_sequential_5_running_mean"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9308096)))]; - tensor encoder_conv_module_0_sequential_5_bias = const()[name = tensor("encoder_conv_module_0_sequential_5_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9309184)))]; - tensor encoder_conv_module_0_sequential_5_weight = const()[name = tensor("encoder_conv_module_0_sequential_5_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9310272)))]; - tensor encoder_conv_module_0_sequential_7_conv_bias = const()[name = tensor("encoder_conv_module_0_sequential_7_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9311360)))]; - tensor encoder_conv_module_0_sequential_7_conv_weight = const()[name = tensor("encoder_conv_module_0_sequential_7_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9312448)))]; - tensor encoder_ffn2_0_module_sequential_0_bias = const()[name = tensor("encoder_ffn2_0_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9574656)))]; - tensor encoder_ffn2_0_module_sequential_0_weight = const()[name = tensor("encoder_ffn2_0_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9575744)))]; - tensor encoder_ffn2_0_module_sequential_1_linear_bias = const()[name = tensor("encoder_ffn2_0_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9576832)))]; - tensor encoder_ffn2_0_module_sequential_1_linear_weight = const()[name = tensor("encoder_ffn2_0_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9580992)))]; - tensor encoder_ffn2_0_module_sequential_4_linear_bias = const()[name = tensor("encoder_ffn2_0_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(10629632)))]; - tensor encoder_ffn2_0_module_sequential_4_linear_weight = const()[name = tensor("encoder_ffn2_0_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(10630720)))]; - tensor encoder_layer_norm_0_bias = const()[name = tensor("encoder_layer_norm_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11679360)))]; - tensor encoder_layer_norm_0_weight = const()[name = tensor("encoder_layer_norm_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11680448)))]; - tensor encoder_ffn1_1_module_sequential_0_bias = const()[name = tensor("encoder_ffn1_1_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11681536)))]; - tensor encoder_ffn1_1_module_sequential_0_weight = const()[name = tensor("encoder_ffn1_1_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11682624)))]; - tensor encoder_ffn1_1_module_sequential_1_linear_bias = const()[name = tensor("encoder_ffn1_1_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11683712)))]; - tensor encoder_ffn1_1_module_sequential_1_linear_weight = const()[name = tensor("encoder_ffn1_1_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11687872)))]; - tensor encoder_ffn1_1_module_sequential_4_linear_bias = const()[name = tensor("encoder_ffn1_1_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(12736512)))]; - tensor encoder_ffn1_1_module_sequential_4_linear_weight = const()[name = tensor("encoder_ffn1_1_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(12737600)))]; - tensor encoder_ret_lns_1_bias = const()[name = tensor("encoder_ret_lns_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(13786240)))]; - tensor encoder_ret_lns_1_weight = const()[name = tensor("encoder_ret_lns_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(13787328)))]; - tensor encoder_q_proj_1_bias = const()[name = tensor("encoder_q_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(13788416)))]; - tensor encoder_q_proj_1_weight = const()[name = tensor("encoder_q_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(13789504)))]; - tensor encoder_k_proj_1_bias = const()[name = tensor("encoder_k_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14051712)))]; - tensor encoder_k_proj_1_weight = const()[name = tensor("encoder_k_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14052800)))]; - tensor encoder_v_proj_1_bias = const()[name = tensor("encoder_v_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14315008)))]; - tensor encoder_v_proj_1_weight = const()[name = tensor("encoder_v_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14316096)))]; - tensor encoder_g_proj_1_bias = const()[name = tensor("encoder_g_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14578304)))]; - tensor encoder_g_proj_1_weight = const()[name = tensor("encoder_g_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14579392)))]; - tensor encoder_out_proj_1_bias = const()[name = tensor("encoder_out_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14841600)))]; - tensor encoder_out_proj_1_weight = const()[name = tensor("encoder_out_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14842688)))]; - tensor encoder_conv_module_1_sequential_0_bias = const()[name = tensor("encoder_conv_module_1_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15104896)))]; - tensor encoder_conv_module_1_sequential_0_weight = const()[name = tensor("encoder_conv_module_1_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15105984)))]; - tensor encoder_conv_module_1_sequential_2_conv_bias = const()[name = tensor("encoder_conv_module_1_sequential_2_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15107072)))]; - tensor encoder_conv_module_1_sequential_2_conv_weight = const()[name = tensor("encoder_conv_module_1_sequential_2_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15109184)))]; - tensor encoder_conv_module_1_sequential_4_conv_weight = const()[name = tensor("encoder_conv_module_1_sequential_4_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15633536)))]; - tensor encoder_conv_module_1_sequential_5_running_var = const()[name = tensor("encoder_conv_module_1_sequential_5_running_var"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15649984)))]; - tensor encoder_conv_module_1_sequential_5_running_mean = const()[name = tensor("encoder_conv_module_1_sequential_5_running_mean"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15651072)))]; - tensor encoder_conv_module_1_sequential_5_bias = const()[name = tensor("encoder_conv_module_1_sequential_5_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15652160)))]; - tensor encoder_conv_module_1_sequential_5_weight = const()[name = tensor("encoder_conv_module_1_sequential_5_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15653248)))]; - tensor encoder_conv_module_1_sequential_7_conv_bias = const()[name = tensor("encoder_conv_module_1_sequential_7_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15654336)))]; - tensor encoder_conv_module_1_sequential_7_conv_weight = const()[name = tensor("encoder_conv_module_1_sequential_7_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15655424)))]; - tensor encoder_ffn2_1_module_sequential_0_bias = const()[name = tensor("encoder_ffn2_1_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15917632)))]; - tensor encoder_ffn2_1_module_sequential_0_weight = const()[name = tensor("encoder_ffn2_1_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15918720)))]; - tensor encoder_ffn2_1_module_sequential_1_linear_bias = const()[name = tensor("encoder_ffn2_1_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15919808)))]; - tensor encoder_ffn2_1_module_sequential_1_linear_weight = const()[name = tensor("encoder_ffn2_1_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15923968)))]; - tensor encoder_ffn2_1_module_sequential_4_linear_bias = const()[name = tensor("encoder_ffn2_1_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(16972608)))]; - tensor encoder_ffn2_1_module_sequential_4_linear_weight = const()[name = tensor("encoder_ffn2_1_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(16973696)))]; - tensor encoder_layer_norm_1_bias = const()[name = tensor("encoder_layer_norm_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18022336)))]; - tensor encoder_layer_norm_1_weight = const()[name = tensor("encoder_layer_norm_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18023424)))]; - tensor encoder_ffn1_2_module_sequential_0_bias = const()[name = tensor("encoder_ffn1_2_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18024512)))]; - tensor encoder_ffn1_2_module_sequential_0_weight = const()[name = tensor("encoder_ffn1_2_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18025600)))]; - tensor encoder_ffn1_2_module_sequential_1_linear_bias = const()[name = tensor("encoder_ffn1_2_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18026688)))]; - tensor encoder_ffn1_2_module_sequential_1_linear_weight = const()[name = tensor("encoder_ffn1_2_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18030848)))]; - tensor encoder_ffn1_2_module_sequential_4_linear_bias = const()[name = tensor("encoder_ffn1_2_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(19079488)))]; - tensor encoder_ffn1_2_module_sequential_4_linear_weight = const()[name = tensor("encoder_ffn1_2_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(19080576)))]; - tensor encoder_ret_lns_2_bias = const()[name = tensor("encoder_ret_lns_2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20129216)))]; - tensor encoder_ret_lns_2_weight = const()[name = tensor("encoder_ret_lns_2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20130304)))]; - tensor encoder_q_proj_2_bias = const()[name = tensor("encoder_q_proj_2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20131392)))]; - tensor encoder_q_proj_2_weight = const()[name = tensor("encoder_q_proj_2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20132480)))]; - tensor encoder_k_proj_2_bias = const()[name = tensor("encoder_k_proj_2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20394688)))]; - tensor encoder_k_proj_2_weight = const()[name = tensor("encoder_k_proj_2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20395776)))]; - tensor encoder_v_proj_2_bias = const()[name = tensor("encoder_v_proj_2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20657984)))]; - tensor encoder_v_proj_2_weight = const()[name = tensor("encoder_v_proj_2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20659072)))]; - tensor encoder_g_proj_2_bias = const()[name = tensor("encoder_g_proj_2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20921280)))]; - tensor encoder_g_proj_2_weight = const()[name = tensor("encoder_g_proj_2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20922368)))]; - tensor encoder_out_proj_2_bias = const()[name = tensor("encoder_out_proj_2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21184576)))]; - tensor encoder_out_proj_2_weight = const()[name = tensor("encoder_out_proj_2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21185664)))]; - tensor encoder_conv_module_2_sequential_0_bias = const()[name = tensor("encoder_conv_module_2_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21447872)))]; - tensor encoder_conv_module_2_sequential_0_weight = const()[name = tensor("encoder_conv_module_2_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21448960)))]; - tensor encoder_conv_module_2_sequential_2_conv_bias = const()[name = tensor("encoder_conv_module_2_sequential_2_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21450048)))]; - tensor encoder_conv_module_2_sequential_2_conv_weight = const()[name = tensor("encoder_conv_module_2_sequential_2_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21452160)))]; - tensor encoder_conv_module_2_sequential_4_conv_weight = const()[name = tensor("encoder_conv_module_2_sequential_4_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21976512)))]; - tensor encoder_conv_module_2_sequential_5_running_var = const()[name = tensor("encoder_conv_module_2_sequential_5_running_var"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21992960)))]; - tensor encoder_conv_module_2_sequential_5_running_mean = const()[name = tensor("encoder_conv_module_2_sequential_5_running_mean"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21994048)))]; - tensor encoder_conv_module_2_sequential_5_bias = const()[name = tensor("encoder_conv_module_2_sequential_5_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21995136)))]; - tensor encoder_conv_module_2_sequential_5_weight = const()[name = tensor("encoder_conv_module_2_sequential_5_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21996224)))]; - tensor encoder_conv_module_2_sequential_7_conv_bias = const()[name = tensor("encoder_conv_module_2_sequential_7_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21997312)))]; - tensor encoder_conv_module_2_sequential_7_conv_weight = const()[name = tensor("encoder_conv_module_2_sequential_7_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21998400)))]; - tensor encoder_ffn2_2_module_sequential_0_bias = const()[name = tensor("encoder_ffn2_2_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(22260608)))]; - tensor encoder_ffn2_2_module_sequential_0_weight = const()[name = tensor("encoder_ffn2_2_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(22261696)))]; - tensor encoder_ffn2_2_module_sequential_1_linear_bias = const()[name = tensor("encoder_ffn2_2_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(22262784)))]; - tensor encoder_ffn2_2_module_sequential_1_linear_weight = const()[name = tensor("encoder_ffn2_2_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(22266944)))]; - tensor encoder_ffn2_2_module_sequential_4_linear_bias = const()[name = tensor("encoder_ffn2_2_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(23315584)))]; - tensor encoder_ffn2_2_module_sequential_4_linear_weight = const()[name = tensor("encoder_ffn2_2_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(23316672)))]; - tensor encoder_layer_norm_2_bias = const()[name = tensor("encoder_layer_norm_2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(24365312)))]; - tensor encoder_layer_norm_2_weight = const()[name = tensor("encoder_layer_norm_2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(24366400)))]; - tensor encoder_ffn1_3_module_sequential_0_bias = const()[name = tensor("encoder_ffn1_3_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(24367488)))]; - tensor encoder_ffn1_3_module_sequential_0_weight = const()[name = tensor("encoder_ffn1_3_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(24368576)))]; - tensor encoder_ffn1_3_module_sequential_1_linear_bias = const()[name = tensor("encoder_ffn1_3_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(24369664)))]; - tensor encoder_ffn1_3_module_sequential_1_linear_weight = const()[name = tensor("encoder_ffn1_3_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(24373824)))]; - tensor encoder_ffn1_3_module_sequential_4_linear_bias = const()[name = tensor("encoder_ffn1_3_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(25422464)))]; - tensor encoder_ffn1_3_module_sequential_4_linear_weight = const()[name = tensor("encoder_ffn1_3_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(25423552)))]; - tensor encoder_ret_lns_3_bias = const()[name = tensor("encoder_ret_lns_3_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(26472192)))]; - tensor encoder_ret_lns_3_weight = const()[name = tensor("encoder_ret_lns_3_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(26473280)))]; - tensor encoder_q_proj_3_bias = const()[name = tensor("encoder_q_proj_3_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(26474368)))]; - tensor encoder_q_proj_3_weight = const()[name = tensor("encoder_q_proj_3_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(26475456)))]; - tensor encoder_k_proj_3_bias = const()[name = tensor("encoder_k_proj_3_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(26737664)))]; - tensor encoder_k_proj_3_weight = const()[name = tensor("encoder_k_proj_3_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(26738752)))]; - tensor encoder_v_proj_3_bias = const()[name = tensor("encoder_v_proj_3_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27000960)))]; - tensor encoder_v_proj_3_weight = const()[name = tensor("encoder_v_proj_3_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27002048)))]; - tensor encoder_g_proj_3_bias = const()[name = tensor("encoder_g_proj_3_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27264256)))]; - tensor encoder_g_proj_3_weight = const()[name = tensor("encoder_g_proj_3_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27265344)))]; - tensor encoder_out_proj_3_bias = const()[name = tensor("encoder_out_proj_3_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27527552)))]; - tensor encoder_out_proj_3_weight = const()[name = tensor("encoder_out_proj_3_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27528640)))]; - tensor encoder_conv_module_3_sequential_0_bias = const()[name = tensor("encoder_conv_module_3_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27790848)))]; - tensor encoder_conv_module_3_sequential_0_weight = const()[name = tensor("encoder_conv_module_3_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27791936)))]; - tensor encoder_conv_module_3_sequential_2_conv_bias = const()[name = tensor("encoder_conv_module_3_sequential_2_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27793024)))]; - tensor encoder_conv_module_3_sequential_2_conv_weight = const()[name = tensor("encoder_conv_module_3_sequential_2_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27795136)))]; - tensor encoder_conv_module_3_sequential_4_conv_weight = const()[name = tensor("encoder_conv_module_3_sequential_4_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28319488)))]; - tensor encoder_conv_module_3_sequential_5_running_var = const()[name = tensor("encoder_conv_module_3_sequential_5_running_var"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28335936)))]; - tensor encoder_conv_module_3_sequential_5_running_mean = const()[name = tensor("encoder_conv_module_3_sequential_5_running_mean"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28337024)))]; - tensor encoder_conv_module_3_sequential_5_bias = const()[name = tensor("encoder_conv_module_3_sequential_5_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28338112)))]; - tensor encoder_conv_module_3_sequential_5_weight = const()[name = tensor("encoder_conv_module_3_sequential_5_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28339200)))]; - tensor encoder_conv_module_3_sequential_7_conv_bias = const()[name = tensor("encoder_conv_module_3_sequential_7_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28340288)))]; - tensor encoder_conv_module_3_sequential_7_conv_weight = const()[name = tensor("encoder_conv_module_3_sequential_7_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28341376)))]; - tensor encoder_ffn2_3_module_sequential_0_bias = const()[name = tensor("encoder_ffn2_3_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28603584)))]; - tensor encoder_ffn2_3_module_sequential_0_weight = const()[name = tensor("encoder_ffn2_3_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28604672)))]; - tensor encoder_ffn2_3_module_sequential_1_linear_bias = const()[name = tensor("encoder_ffn2_3_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28605760)))]; - tensor encoder_ffn2_3_module_sequential_1_linear_weight = const()[name = tensor("encoder_ffn2_3_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28609920)))]; - tensor encoder_ffn2_3_module_sequential_4_linear_bias = const()[name = tensor("encoder_ffn2_3_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(29658560)))]; - tensor encoder_ffn2_3_module_sequential_4_linear_weight = const()[name = tensor("encoder_ffn2_3_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(29659648)))]; - tensor encoder_layer_norm_3_bias = const()[name = tensor("encoder_layer_norm_3_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(30708288)))]; - tensor encoder_layer_norm_3_weight = const()[name = tensor("encoder_layer_norm_3_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(30709376)))]; - tensor decoder__causal_mask = const()[name = tensor("decoder__causal_mask"), val = tensor([[0x1p+0, 0x0p+0], [0x1p+0, 0x1p+0]])]; - tensor decoder_convert_bias = const()[name = tensor("decoder_convert_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(30710464)))]; - tensor decoder_convert_weight = const()[name = tensor("decoder_convert_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(30711552)))]; - tensor decoder_q_proj_0_bias = const()[name = tensor("decoder_q_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(31235904)))]; - tensor decoder_q_proj_0_weight = const()[name = tensor("decoder_q_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(31236992)))]; - tensor decoder_k_proj_0_bias = const()[name = tensor("decoder_k_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(31499200)))]; - tensor decoder_k_proj_0_weight = const()[name = tensor("decoder_k_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(31500288)))]; - tensor decoder_v_proj_0_bias = const()[name = tensor("decoder_v_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(31762496)))]; - tensor decoder_v_proj_0_weight = const()[name = tensor("decoder_v_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(31763584)))]; - tensor decoder_g_proj_0_bias = const()[name = tensor("decoder_g_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32025792)))]; - tensor decoder_g_proj_0_weight = const()[name = tensor("decoder_g_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32026880)))]; - tensor decoder_out_proj_0_bias = const()[name = tensor("decoder_out_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32289088)))]; - tensor decoder_out_proj_0_weight = const()[name = tensor("decoder_out_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32290176)))]; - tensor decoder_norm11_0_bias = const()[name = tensor("decoder_norm11_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32552384)))]; - tensor decoder_norm11_0_weight = const()[name = tensor("decoder_norm11_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32553472)))]; - tensor decoder_self_attn2_0_out_proj_bias = const()[name = tensor("decoder_self_attn2_0_out_proj_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32554560)))]; - tensor decoder_self_attn2_0_out_proj_weight = const()[name = tensor("decoder_self_attn2_0_out_proj_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32555648)))]; - tensor decoder_self_attn2_0_in_proj_bias = const()[name = tensor("decoder_self_attn2_0_in_proj_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32817856)))]; - tensor decoder_self_attn2_0_in_proj_weight = const()[name = tensor("decoder_self_attn2_0_in_proj_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32820992)))]; - tensor decoder_norm21_0_bias = const()[name = tensor("decoder_norm21_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(33607488)))]; - tensor decoder_norm21_0_weight = const()[name = tensor("decoder_norm21_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(33608576)))]; - tensor decoder_linear1_0_bias = const()[name = tensor("decoder_linear1_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(33609664)))]; - tensor decoder_linear1_0_weight = const()[name = tensor("decoder_linear1_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(33617920)))]; - tensor decoder_linear2_0_bias = const()[name = tensor("decoder_linear2_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(35715136)))]; - tensor decoder_linear2_0_weight = const()[name = tensor("decoder_linear2_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(35716224)))]; - tensor decoder_norm22_0_bias = const()[name = tensor("decoder_norm22_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(37813440)))]; - tensor decoder_norm22_0_weight = const()[name = tensor("decoder_norm22_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(37814528)))]; - tensor decoder_q_proj_1_bias = const()[name = tensor("decoder_q_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(37815616)))]; - tensor decoder_q_proj_1_weight = const()[name = tensor("decoder_q_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(37816704)))]; - tensor decoder_k_proj_1_bias = const()[name = tensor("decoder_k_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38078912)))]; - tensor decoder_k_proj_1_weight = const()[name = tensor("decoder_k_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38080000)))]; - tensor decoder_v_proj_1_bias = const()[name = tensor("decoder_v_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38342208)))]; - tensor decoder_v_proj_1_weight = const()[name = tensor("decoder_v_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38343296)))]; - tensor decoder_g_proj_1_bias = const()[name = tensor("decoder_g_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38605504)))]; - tensor decoder_g_proj_1_weight = const()[name = tensor("decoder_g_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38606592)))]; - tensor decoder_out_proj_1_bias = const()[name = tensor("decoder_out_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38868800)))]; - tensor decoder_out_proj_1_weight = const()[name = tensor("decoder_out_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38869888)))]; - tensor decoder_norm11_1_bias = const()[name = tensor("decoder_norm11_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(39132096)))]; - tensor decoder_norm11_1_weight = const()[name = tensor("decoder_norm11_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(39133184)))]; - tensor decoder_self_attn2_1_out_proj_bias = const()[name = tensor("decoder_self_attn2_1_out_proj_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(39134272)))]; - tensor decoder_self_attn2_1_out_proj_weight = const()[name = tensor("decoder_self_attn2_1_out_proj_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(39135360)))]; - tensor decoder_self_attn2_1_in_proj_bias = const()[name = tensor("decoder_self_attn2_1_in_proj_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(39397568)))]; - tensor decoder_self_attn2_1_in_proj_weight = const()[name = tensor("decoder_self_attn2_1_in_proj_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(39400704)))]; - tensor decoder_norm21_1_bias = const()[name = tensor("decoder_norm21_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(40187200)))]; - tensor decoder_norm21_1_weight = const()[name = tensor("decoder_norm21_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(40188288)))]; - tensor decoder_linear1_1_bias = const()[name = tensor("decoder_linear1_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(40189376)))]; - tensor decoder_linear1_1_weight = const()[name = tensor("decoder_linear1_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(40197632)))]; - tensor decoder_linear2_1_bias = const()[name = tensor("decoder_linear2_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(42294848)))]; - tensor decoder_linear2_1_weight = const()[name = tensor("decoder_linear2_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(42295936)))]; - tensor decoder_norm22_1_bias = const()[name = tensor("decoder_norm22_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(44393152)))]; - tensor decoder_norm22_1_weight = const()[name = tensor("decoder_norm22_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(44394240)))]; - tensor var_11 = const()[name = tensor("op_11"), val = tensor(0x1.5798eep-27)]; - tensor var_17 = const()[name = tensor("op_17"), val = tensor(0x1.0c6f7ap-20)]; - tensor var_20 = const()[name = tensor("op_20"), val = tensor(2)]; - tensor var_23 = const()[name = tensor("op_23"), val = tensor(0)]; - tensor var_26 = const()[name = tensor("op_26"), val = tensor(1)]; - tensor var_28 = const()[name = tensor("op_28"), val = tensor(0x1.4f8b58p-17)]; - tensor var_29 = const()[name = tensor("op_29"), val = tensor(true)]; - tensor input_1 = linear(bias = encoder_input_projection_linear_bias, weight = encoder_input_projection_linear_weight, x = features)[name = tensor("linear_0")]; - tensor input_3_axes_0 = const()[name = tensor("input_3_axes_0"), val = tensor([-1])]; - tensor input_3 = layer_norm(axes = input_3_axes_0, beta = encoder_pre_layer_norm_bias, epsilon = var_28, gamma = encoder_pre_layer_norm_weight, x = input_1)[name = tensor("input_3")]; + func main(tensor cnn_window, tensor dec_kv, tensor dec_scale, tensor enc_conv_cache, tensor enc_kv, tensor enc_scale, tensor features, tensor valid_mask) { + tensor inner_encoder_cnn_bias = const()[name = tensor("inner_encoder_cnn_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(64)))]; + tensor inner_encoder_cnn_weight = const()[name = tensor("inner_encoder_cnn_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1152)))]; + tensor inner_encoder__causal_mask = const()[name = tensor("inner_encoder__causal_mask"), val = tensor([[0x1p+0, 0x0p+0], [0x1p+0, 0x1p+0]])]; + tensor inner_encoder__t_index = const()[name = tensor("inner_encoder__t_index"), val = tensor([0x1p+0, 0x1p+1])]; + tensor inner_encoder_input_projection_linear_bias = const()[name = tensor("inner_encoder_input_projection_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4981952)))]; + tensor inner_encoder_input_projection_linear_weight = const()[name = tensor("inner_encoder_input_projection_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4983040)))]; + tensor inner_encoder_pre_layer_norm_bias = const()[name = tensor("inner_encoder_pre_layer_norm_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(5336384)))]; + tensor inner_encoder_pre_layer_norm_weight = const()[name = tensor("inner_encoder_pre_layer_norm_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(5337472)))]; + tensor inner_encoder_ffn1_0_module_sequential_0_bias = const()[name = tensor("inner_encoder_ffn1_0_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(5338560)))]; + tensor inner_encoder_ffn1_0_module_sequential_0_weight = const()[name = tensor("inner_encoder_ffn1_0_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(5339648)))]; + tensor inner_encoder_ffn1_0_module_sequential_1_linear_bias = const()[name = tensor("inner_encoder_ffn1_0_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(5340736)))]; + tensor inner_encoder_ffn1_0_module_sequential_1_linear_weight = const()[name = tensor("inner_encoder_ffn1_0_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(5344896)))]; + tensor inner_encoder_ffn1_0_module_sequential_4_linear_bias = const()[name = tensor("inner_encoder_ffn1_0_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(6393536)))]; + tensor inner_encoder_ffn1_0_module_sequential_4_linear_weight = const()[name = tensor("inner_encoder_ffn1_0_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(6394624)))]; + tensor inner_encoder_ret_lns_0_bias = const()[name = tensor("inner_encoder_ret_lns_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7443264)))]; + tensor inner_encoder_ret_lns_0_weight = const()[name = tensor("inner_encoder_ret_lns_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7444352)))]; + tensor inner_encoder_q_proj_0_bias = const()[name = tensor("inner_encoder_q_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7445440)))]; + tensor inner_encoder_q_proj_0_weight = const()[name = tensor("inner_encoder_q_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7446528)))]; + tensor inner_encoder_k_proj_0_bias = const()[name = tensor("inner_encoder_k_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7708736)))]; + tensor inner_encoder_k_proj_0_weight = const()[name = tensor("inner_encoder_k_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7709824)))]; + tensor inner_encoder_v_proj_0_bias = const()[name = tensor("inner_encoder_v_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7972032)))]; + tensor inner_encoder_v_proj_0_weight = const()[name = tensor("inner_encoder_v_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7973120)))]; + tensor inner_encoder_g_proj_0_bias = const()[name = tensor("inner_encoder_g_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8235328)))]; + tensor inner_encoder_g_proj_0_weight = const()[name = tensor("inner_encoder_g_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8236416)))]; + tensor inner_encoder_out_proj_0_bias = const()[name = tensor("inner_encoder_out_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8498624)))]; + tensor inner_encoder_out_proj_0_weight = const()[name = tensor("inner_encoder_out_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8499712)))]; + tensor inner_encoder_conv_module_0_sequential_0_bias = const()[name = tensor("inner_encoder_conv_module_0_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8761920)))]; + tensor inner_encoder_conv_module_0_sequential_0_weight = const()[name = tensor("inner_encoder_conv_module_0_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8763008)))]; + tensor inner_encoder_conv_module_0_sequential_2_conv_bias = const()[name = tensor("inner_encoder_conv_module_0_sequential_2_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8764096)))]; + tensor inner_encoder_conv_module_0_sequential_2_conv_weight = const()[name = tensor("inner_encoder_conv_module_0_sequential_2_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8766208)))]; + tensor inner_encoder_conv_module_0_sequential_4_conv_weight = const()[name = tensor("inner_encoder_conv_module_0_sequential_4_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9290560)))]; + tensor inner_encoder_conv_module_0_sequential_5_running_var = const()[name = tensor("inner_encoder_conv_module_0_sequential_5_running_var"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9307008)))]; + tensor inner_encoder_conv_module_0_sequential_5_running_mean = const()[name = tensor("inner_encoder_conv_module_0_sequential_5_running_mean"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9308096)))]; + tensor inner_encoder_conv_module_0_sequential_5_bias = const()[name = tensor("inner_encoder_conv_module_0_sequential_5_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9309184)))]; + tensor inner_encoder_conv_module_0_sequential_5_weight = const()[name = tensor("inner_encoder_conv_module_0_sequential_5_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9310272)))]; + tensor inner_encoder_conv_module_0_sequential_7_conv_bias = const()[name = tensor("inner_encoder_conv_module_0_sequential_7_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9311360)))]; + tensor inner_encoder_conv_module_0_sequential_7_conv_weight = const()[name = tensor("inner_encoder_conv_module_0_sequential_7_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9312448)))]; + tensor inner_encoder_ffn2_0_module_sequential_0_bias = const()[name = tensor("inner_encoder_ffn2_0_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9574656)))]; + tensor inner_encoder_ffn2_0_module_sequential_0_weight = const()[name = tensor("inner_encoder_ffn2_0_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9575744)))]; + tensor inner_encoder_ffn2_0_module_sequential_1_linear_bias = const()[name = tensor("inner_encoder_ffn2_0_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9576832)))]; + tensor inner_encoder_ffn2_0_module_sequential_1_linear_weight = const()[name = tensor("inner_encoder_ffn2_0_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9580992)))]; + tensor inner_encoder_ffn2_0_module_sequential_4_linear_bias = const()[name = tensor("inner_encoder_ffn2_0_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(10629632)))]; + tensor inner_encoder_ffn2_0_module_sequential_4_linear_weight = const()[name = tensor("inner_encoder_ffn2_0_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(10630720)))]; + tensor inner_encoder_layer_norm_0_bias = const()[name = tensor("inner_encoder_layer_norm_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11679360)))]; + tensor inner_encoder_layer_norm_0_weight = const()[name = tensor("inner_encoder_layer_norm_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11680448)))]; + tensor inner_encoder_ffn1_1_module_sequential_0_bias = const()[name = tensor("inner_encoder_ffn1_1_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11681536)))]; + tensor inner_encoder_ffn1_1_module_sequential_0_weight = const()[name = tensor("inner_encoder_ffn1_1_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11682624)))]; + tensor inner_encoder_ffn1_1_module_sequential_1_linear_bias = const()[name = tensor("inner_encoder_ffn1_1_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11683712)))]; + tensor inner_encoder_ffn1_1_module_sequential_1_linear_weight = const()[name = tensor("inner_encoder_ffn1_1_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11687872)))]; + tensor inner_encoder_ffn1_1_module_sequential_4_linear_bias = const()[name = tensor("inner_encoder_ffn1_1_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(12736512)))]; + tensor inner_encoder_ffn1_1_module_sequential_4_linear_weight = const()[name = tensor("inner_encoder_ffn1_1_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(12737600)))]; + tensor inner_encoder_ret_lns_1_bias = const()[name = tensor("inner_encoder_ret_lns_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(13786240)))]; + tensor inner_encoder_ret_lns_1_weight = const()[name = tensor("inner_encoder_ret_lns_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(13787328)))]; + tensor inner_encoder_q_proj_1_bias = const()[name = tensor("inner_encoder_q_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(13788416)))]; + tensor inner_encoder_q_proj_1_weight = const()[name = tensor("inner_encoder_q_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(13789504)))]; + tensor inner_encoder_k_proj_1_bias = const()[name = tensor("inner_encoder_k_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14051712)))]; + tensor inner_encoder_k_proj_1_weight = const()[name = tensor("inner_encoder_k_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14052800)))]; + tensor inner_encoder_v_proj_1_bias = const()[name = tensor("inner_encoder_v_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14315008)))]; + tensor inner_encoder_v_proj_1_weight = const()[name = tensor("inner_encoder_v_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14316096)))]; + tensor inner_encoder_g_proj_1_bias = const()[name = tensor("inner_encoder_g_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14578304)))]; + tensor inner_encoder_g_proj_1_weight = const()[name = tensor("inner_encoder_g_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14579392)))]; + tensor inner_encoder_out_proj_1_bias = const()[name = tensor("inner_encoder_out_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14841600)))]; + tensor inner_encoder_out_proj_1_weight = const()[name = tensor("inner_encoder_out_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14842688)))]; + tensor inner_encoder_conv_module_1_sequential_0_bias = const()[name = tensor("inner_encoder_conv_module_1_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15104896)))]; + tensor inner_encoder_conv_module_1_sequential_0_weight = const()[name = tensor("inner_encoder_conv_module_1_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15105984)))]; + tensor inner_encoder_conv_module_1_sequential_2_conv_bias = const()[name = tensor("inner_encoder_conv_module_1_sequential_2_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15107072)))]; + tensor inner_encoder_conv_module_1_sequential_2_conv_weight = const()[name = tensor("inner_encoder_conv_module_1_sequential_2_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15109184)))]; + tensor inner_encoder_conv_module_1_sequential_4_conv_weight = const()[name = tensor("inner_encoder_conv_module_1_sequential_4_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15633536)))]; + tensor inner_encoder_conv_module_1_sequential_5_running_var = const()[name = tensor("inner_encoder_conv_module_1_sequential_5_running_var"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15649984)))]; + tensor inner_encoder_conv_module_1_sequential_5_running_mean = const()[name = tensor("inner_encoder_conv_module_1_sequential_5_running_mean"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15651072)))]; + tensor inner_encoder_conv_module_1_sequential_5_bias = const()[name = tensor("inner_encoder_conv_module_1_sequential_5_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15652160)))]; + tensor inner_encoder_conv_module_1_sequential_5_weight = const()[name = tensor("inner_encoder_conv_module_1_sequential_5_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15653248)))]; + tensor inner_encoder_conv_module_1_sequential_7_conv_bias = const()[name = tensor("inner_encoder_conv_module_1_sequential_7_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15654336)))]; + tensor inner_encoder_conv_module_1_sequential_7_conv_weight = const()[name = tensor("inner_encoder_conv_module_1_sequential_7_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15655424)))]; + tensor inner_encoder_ffn2_1_module_sequential_0_bias = const()[name = tensor("inner_encoder_ffn2_1_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15917632)))]; + tensor inner_encoder_ffn2_1_module_sequential_0_weight = const()[name = tensor("inner_encoder_ffn2_1_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15918720)))]; + tensor inner_encoder_ffn2_1_module_sequential_1_linear_bias = const()[name = tensor("inner_encoder_ffn2_1_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15919808)))]; + tensor inner_encoder_ffn2_1_module_sequential_1_linear_weight = const()[name = tensor("inner_encoder_ffn2_1_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15923968)))]; + tensor inner_encoder_ffn2_1_module_sequential_4_linear_bias = const()[name = tensor("inner_encoder_ffn2_1_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(16972608)))]; + tensor inner_encoder_ffn2_1_module_sequential_4_linear_weight = const()[name = tensor("inner_encoder_ffn2_1_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(16973696)))]; + tensor inner_encoder_layer_norm_1_bias = const()[name = tensor("inner_encoder_layer_norm_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18022336)))]; + tensor inner_encoder_layer_norm_1_weight = const()[name = tensor("inner_encoder_layer_norm_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18023424)))]; + tensor inner_encoder_ffn1_2_module_sequential_0_bias = const()[name = tensor("inner_encoder_ffn1_2_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18024512)))]; + tensor inner_encoder_ffn1_2_module_sequential_0_weight = const()[name = tensor("inner_encoder_ffn1_2_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18025600)))]; + tensor inner_encoder_ffn1_2_module_sequential_1_linear_bias = const()[name = tensor("inner_encoder_ffn1_2_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18026688)))]; + tensor inner_encoder_ffn1_2_module_sequential_1_linear_weight = const()[name = tensor("inner_encoder_ffn1_2_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18030848)))]; + tensor inner_encoder_ffn1_2_module_sequential_4_linear_bias = const()[name = tensor("inner_encoder_ffn1_2_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(19079488)))]; + tensor inner_encoder_ffn1_2_module_sequential_4_linear_weight = const()[name = tensor("inner_encoder_ffn1_2_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(19080576)))]; + tensor inner_encoder_ret_lns_2_bias = const()[name = tensor("inner_encoder_ret_lns_2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20129216)))]; + tensor inner_encoder_ret_lns_2_weight = const()[name = tensor("inner_encoder_ret_lns_2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20130304)))]; + tensor inner_encoder_q_proj_2_bias = const()[name = tensor("inner_encoder_q_proj_2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20131392)))]; + tensor inner_encoder_q_proj_2_weight = const()[name = tensor("inner_encoder_q_proj_2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20132480)))]; + tensor inner_encoder_k_proj_2_bias = const()[name = tensor("inner_encoder_k_proj_2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20394688)))]; + tensor inner_encoder_k_proj_2_weight = const()[name = tensor("inner_encoder_k_proj_2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20395776)))]; + tensor inner_encoder_v_proj_2_bias = const()[name = tensor("inner_encoder_v_proj_2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20657984)))]; + tensor inner_encoder_v_proj_2_weight = const()[name = tensor("inner_encoder_v_proj_2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20659072)))]; + tensor inner_encoder_g_proj_2_bias = const()[name = tensor("inner_encoder_g_proj_2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20921280)))]; + tensor inner_encoder_g_proj_2_weight = const()[name = tensor("inner_encoder_g_proj_2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20922368)))]; + tensor inner_encoder_out_proj_2_bias = const()[name = tensor("inner_encoder_out_proj_2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21184576)))]; + tensor inner_encoder_out_proj_2_weight = const()[name = tensor("inner_encoder_out_proj_2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21185664)))]; + tensor inner_encoder_conv_module_2_sequential_0_bias = const()[name = tensor("inner_encoder_conv_module_2_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21447872)))]; + tensor inner_encoder_conv_module_2_sequential_0_weight = const()[name = tensor("inner_encoder_conv_module_2_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21448960)))]; + tensor inner_encoder_conv_module_2_sequential_2_conv_bias = const()[name = tensor("inner_encoder_conv_module_2_sequential_2_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21450048)))]; + tensor inner_encoder_conv_module_2_sequential_2_conv_weight = const()[name = tensor("inner_encoder_conv_module_2_sequential_2_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21452160)))]; + tensor inner_encoder_conv_module_2_sequential_4_conv_weight = const()[name = tensor("inner_encoder_conv_module_2_sequential_4_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21976512)))]; + tensor inner_encoder_conv_module_2_sequential_5_running_var = const()[name = tensor("inner_encoder_conv_module_2_sequential_5_running_var"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21992960)))]; + tensor inner_encoder_conv_module_2_sequential_5_running_mean = const()[name = tensor("inner_encoder_conv_module_2_sequential_5_running_mean"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21994048)))]; + tensor inner_encoder_conv_module_2_sequential_5_bias = const()[name = tensor("inner_encoder_conv_module_2_sequential_5_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21995136)))]; + tensor inner_encoder_conv_module_2_sequential_5_weight = const()[name = tensor("inner_encoder_conv_module_2_sequential_5_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21996224)))]; + tensor inner_encoder_conv_module_2_sequential_7_conv_bias = const()[name = tensor("inner_encoder_conv_module_2_sequential_7_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21997312)))]; + tensor inner_encoder_conv_module_2_sequential_7_conv_weight = const()[name = tensor("inner_encoder_conv_module_2_sequential_7_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21998400)))]; + tensor inner_encoder_ffn2_2_module_sequential_0_bias = const()[name = tensor("inner_encoder_ffn2_2_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(22260608)))]; + tensor inner_encoder_ffn2_2_module_sequential_0_weight = const()[name = tensor("inner_encoder_ffn2_2_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(22261696)))]; + tensor inner_encoder_ffn2_2_module_sequential_1_linear_bias = const()[name = tensor("inner_encoder_ffn2_2_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(22262784)))]; + tensor inner_encoder_ffn2_2_module_sequential_1_linear_weight = const()[name = tensor("inner_encoder_ffn2_2_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(22266944)))]; + tensor inner_encoder_ffn2_2_module_sequential_4_linear_bias = const()[name = tensor("inner_encoder_ffn2_2_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(23315584)))]; + tensor inner_encoder_ffn2_2_module_sequential_4_linear_weight = const()[name = tensor("inner_encoder_ffn2_2_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(23316672)))]; + tensor inner_encoder_layer_norm_2_bias = const()[name = tensor("inner_encoder_layer_norm_2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(24365312)))]; + tensor inner_encoder_layer_norm_2_weight = const()[name = tensor("inner_encoder_layer_norm_2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(24366400)))]; + tensor inner_encoder_ffn1_3_module_sequential_0_bias = const()[name = tensor("inner_encoder_ffn1_3_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(24367488)))]; + tensor inner_encoder_ffn1_3_module_sequential_0_weight = const()[name = tensor("inner_encoder_ffn1_3_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(24368576)))]; + tensor inner_encoder_ffn1_3_module_sequential_1_linear_bias = const()[name = tensor("inner_encoder_ffn1_3_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(24369664)))]; + tensor inner_encoder_ffn1_3_module_sequential_1_linear_weight = const()[name = tensor("inner_encoder_ffn1_3_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(24373824)))]; + tensor inner_encoder_ffn1_3_module_sequential_4_linear_bias = const()[name = tensor("inner_encoder_ffn1_3_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(25422464)))]; + tensor inner_encoder_ffn1_3_module_sequential_4_linear_weight = const()[name = tensor("inner_encoder_ffn1_3_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(25423552)))]; + tensor inner_encoder_ret_lns_3_bias = const()[name = tensor("inner_encoder_ret_lns_3_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(26472192)))]; + tensor inner_encoder_ret_lns_3_weight = const()[name = tensor("inner_encoder_ret_lns_3_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(26473280)))]; + tensor inner_encoder_q_proj_3_bias = const()[name = tensor("inner_encoder_q_proj_3_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(26474368)))]; + tensor inner_encoder_q_proj_3_weight = const()[name = tensor("inner_encoder_q_proj_3_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(26475456)))]; + tensor inner_encoder_k_proj_3_bias = const()[name = tensor("inner_encoder_k_proj_3_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(26737664)))]; + tensor inner_encoder_k_proj_3_weight = const()[name = tensor("inner_encoder_k_proj_3_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(26738752)))]; + tensor inner_encoder_v_proj_3_bias = const()[name = tensor("inner_encoder_v_proj_3_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27000960)))]; + tensor inner_encoder_v_proj_3_weight = const()[name = tensor("inner_encoder_v_proj_3_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27002048)))]; + tensor inner_encoder_g_proj_3_bias = const()[name = tensor("inner_encoder_g_proj_3_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27264256)))]; + tensor inner_encoder_g_proj_3_weight = const()[name = tensor("inner_encoder_g_proj_3_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27265344)))]; + tensor inner_encoder_out_proj_3_bias = const()[name = tensor("inner_encoder_out_proj_3_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27527552)))]; + tensor inner_encoder_out_proj_3_weight = const()[name = tensor("inner_encoder_out_proj_3_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27528640)))]; + tensor inner_encoder_conv_module_3_sequential_0_bias = const()[name = tensor("inner_encoder_conv_module_3_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27790848)))]; + tensor inner_encoder_conv_module_3_sequential_0_weight = const()[name = tensor("inner_encoder_conv_module_3_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27791936)))]; + tensor inner_encoder_conv_module_3_sequential_2_conv_bias = const()[name = tensor("inner_encoder_conv_module_3_sequential_2_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27793024)))]; + tensor inner_encoder_conv_module_3_sequential_2_conv_weight = const()[name = tensor("inner_encoder_conv_module_3_sequential_2_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27795136)))]; + tensor inner_encoder_conv_module_3_sequential_4_conv_weight = const()[name = tensor("inner_encoder_conv_module_3_sequential_4_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28319488)))]; + tensor inner_encoder_conv_module_3_sequential_5_running_var = const()[name = tensor("inner_encoder_conv_module_3_sequential_5_running_var"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28335936)))]; + tensor inner_encoder_conv_module_3_sequential_5_running_mean = const()[name = tensor("inner_encoder_conv_module_3_sequential_5_running_mean"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28337024)))]; + tensor inner_encoder_conv_module_3_sequential_5_bias = const()[name = tensor("inner_encoder_conv_module_3_sequential_5_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28338112)))]; + tensor inner_encoder_conv_module_3_sequential_5_weight = const()[name = tensor("inner_encoder_conv_module_3_sequential_5_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28339200)))]; + tensor inner_encoder_conv_module_3_sequential_7_conv_bias = const()[name = tensor("inner_encoder_conv_module_3_sequential_7_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28340288)))]; + tensor inner_encoder_conv_module_3_sequential_7_conv_weight = const()[name = tensor("inner_encoder_conv_module_3_sequential_7_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28341376)))]; + tensor inner_encoder_ffn2_3_module_sequential_0_bias = const()[name = tensor("inner_encoder_ffn2_3_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28603584)))]; + tensor inner_encoder_ffn2_3_module_sequential_0_weight = const()[name = tensor("inner_encoder_ffn2_3_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28604672)))]; + tensor inner_encoder_ffn2_3_module_sequential_1_linear_bias = const()[name = tensor("inner_encoder_ffn2_3_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28605760)))]; + tensor inner_encoder_ffn2_3_module_sequential_1_linear_weight = const()[name = tensor("inner_encoder_ffn2_3_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28609920)))]; + tensor inner_encoder_ffn2_3_module_sequential_4_linear_bias = const()[name = tensor("inner_encoder_ffn2_3_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(29658560)))]; + tensor inner_encoder_ffn2_3_module_sequential_4_linear_weight = const()[name = tensor("inner_encoder_ffn2_3_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(29659648)))]; + tensor inner_encoder_layer_norm_3_bias = const()[name = tensor("inner_encoder_layer_norm_3_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(30708288)))]; + tensor inner_encoder_layer_norm_3_weight = const()[name = tensor("inner_encoder_layer_norm_3_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(30709376)))]; + tensor inner_decoder__causal_mask = const()[name = tensor("inner_decoder__causal_mask"), val = tensor([[0x1p+0, 0x0p+0], [0x1p+0, 0x1p+0]])]; + tensor inner_decoder_convert_bias = const()[name = tensor("inner_decoder_convert_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(30710464)))]; + tensor inner_decoder_convert_weight = const()[name = tensor("inner_decoder_convert_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(30711552)))]; + tensor inner_decoder_q_proj_0_bias = const()[name = tensor("inner_decoder_q_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(31235904)))]; + tensor inner_decoder_q_proj_0_weight = const()[name = tensor("inner_decoder_q_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(31236992)))]; + tensor inner_decoder_k_proj_0_bias = const()[name = tensor("inner_decoder_k_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(31499200)))]; + tensor inner_decoder_k_proj_0_weight = const()[name = tensor("inner_decoder_k_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(31500288)))]; + tensor inner_decoder_v_proj_0_bias = const()[name = tensor("inner_decoder_v_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(31762496)))]; + tensor inner_decoder_v_proj_0_weight = const()[name = tensor("inner_decoder_v_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(31763584)))]; + tensor inner_decoder_g_proj_0_bias = const()[name = tensor("inner_decoder_g_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32025792)))]; + tensor inner_decoder_g_proj_0_weight = const()[name = tensor("inner_decoder_g_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32026880)))]; + tensor inner_decoder_out_proj_0_bias = const()[name = tensor("inner_decoder_out_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32289088)))]; + tensor inner_decoder_out_proj_0_weight = const()[name = tensor("inner_decoder_out_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32290176)))]; + tensor inner_decoder_norm11_0_bias = const()[name = tensor("inner_decoder_norm11_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32552384)))]; + tensor inner_decoder_norm11_0_weight = const()[name = tensor("inner_decoder_norm11_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32553472)))]; + tensor inner_decoder_self_attn2_0_out_proj_bias = const()[name = tensor("inner_decoder_self_attn2_0_out_proj_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32554560)))]; + tensor inner_decoder_self_attn2_0_out_proj_weight = const()[name = tensor("inner_decoder_self_attn2_0_out_proj_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32555648)))]; + tensor inner_decoder_self_attn2_0_in_proj_bias = const()[name = tensor("inner_decoder_self_attn2_0_in_proj_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32817856)))]; + tensor inner_decoder_self_attn2_0_in_proj_weight = const()[name = tensor("inner_decoder_self_attn2_0_in_proj_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32820992)))]; + tensor inner_decoder_norm21_0_bias = const()[name = tensor("inner_decoder_norm21_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(33607488)))]; + tensor inner_decoder_norm21_0_weight = const()[name = tensor("inner_decoder_norm21_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(33608576)))]; + tensor inner_decoder_linear1_0_bias = const()[name = tensor("inner_decoder_linear1_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(33609664)))]; + tensor inner_decoder_linear1_0_weight = const()[name = tensor("inner_decoder_linear1_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(33617920)))]; + tensor inner_decoder_linear2_0_bias = const()[name = tensor("inner_decoder_linear2_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(35715136)))]; + tensor inner_decoder_linear2_0_weight = const()[name = tensor("inner_decoder_linear2_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(35716224)))]; + tensor inner_decoder_norm22_0_bias = const()[name = tensor("inner_decoder_norm22_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(37813440)))]; + tensor inner_decoder_norm22_0_weight = const()[name = tensor("inner_decoder_norm22_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(37814528)))]; + tensor inner_decoder_q_proj_1_bias = const()[name = tensor("inner_decoder_q_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(37815616)))]; + tensor inner_decoder_q_proj_1_weight = const()[name = tensor("inner_decoder_q_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(37816704)))]; + tensor inner_decoder_k_proj_1_bias = const()[name = tensor("inner_decoder_k_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38078912)))]; + tensor inner_decoder_k_proj_1_weight = const()[name = tensor("inner_decoder_k_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38080000)))]; + tensor inner_decoder_v_proj_1_bias = const()[name = tensor("inner_decoder_v_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38342208)))]; + tensor inner_decoder_v_proj_1_weight = const()[name = tensor("inner_decoder_v_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38343296)))]; + tensor inner_decoder_g_proj_1_bias = const()[name = tensor("inner_decoder_g_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38605504)))]; + tensor inner_decoder_g_proj_1_weight = const()[name = tensor("inner_decoder_g_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38606592)))]; + tensor inner_decoder_out_proj_1_bias = const()[name = tensor("inner_decoder_out_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38868800)))]; + tensor inner_decoder_out_proj_1_weight = const()[name = tensor("inner_decoder_out_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38869888)))]; + tensor inner_decoder_norm11_1_bias = const()[name = tensor("inner_decoder_norm11_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(39132096)))]; + tensor inner_decoder_norm11_1_weight = const()[name = tensor("inner_decoder_norm11_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(39133184)))]; + tensor inner_decoder_self_attn2_1_out_proj_bias = const()[name = tensor("inner_decoder_self_attn2_1_out_proj_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(39134272)))]; + tensor inner_decoder_self_attn2_1_out_proj_weight = const()[name = tensor("inner_decoder_self_attn2_1_out_proj_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(39135360)))]; + tensor inner_decoder_self_attn2_1_in_proj_bias = const()[name = tensor("inner_decoder_self_attn2_1_in_proj_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(39397568)))]; + tensor inner_decoder_self_attn2_1_in_proj_weight = const()[name = tensor("inner_decoder_self_attn2_1_in_proj_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(39400704)))]; + tensor inner_decoder_norm21_1_bias = const()[name = tensor("inner_decoder_norm21_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(40187200)))]; + tensor inner_decoder_norm21_1_weight = const()[name = tensor("inner_decoder_norm21_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(40188288)))]; + tensor inner_decoder_linear1_1_bias = const()[name = tensor("inner_decoder_linear1_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(40189376)))]; + tensor inner_decoder_linear1_1_weight = const()[name = tensor("inner_decoder_linear1_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(40197632)))]; + tensor inner_decoder_linear2_1_bias = const()[name = tensor("inner_decoder_linear2_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(42294848)))]; + tensor inner_decoder_linear2_1_weight = const()[name = tensor("inner_decoder_linear2_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(42295936)))]; + tensor inner_decoder_norm22_1_bias = const()[name = tensor("inner_decoder_norm22_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(44393152)))]; + tensor inner_decoder_norm22_1_weight = const()[name = tensor("inner_decoder_norm22_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(44394240)))]; + tensor var_19_begin_0 = const()[name = tensor("op_19_begin_0"), val = tensor([0, 0, 0])]; + tensor var_19_end_0 = const()[name = tensor("op_19_end_0"), val = tensor([1, 15, 23])]; + tensor var_19_end_mask_0 = const()[name = tensor("op_19_end_mask_0"), val = tensor([true, false, true])]; + tensor var_19 = slice_by_index(begin = var_19_begin_0, end = var_19_end_0, end_mask = var_19_end_mask_0, x = features)[name = tensor("op_19")]; + tensor var_29_begin_0 = const()[name = tensor("op_29_begin_0"), val = tensor([0, 10, 0])]; + tensor var_29_end_0 = const()[name = tensor("op_29_end_0"), val = tensor([1, 1, 23])]; + tensor var_29_end_mask_0 = const()[name = tensor("op_29_end_mask_0"), val = tensor([true, true, true])]; + tensor var_29 = slice_by_index(begin = var_29_begin_0, end = var_29_end_0, end_mask = var_29_end_mask_0, x = features)[name = tensor("op_29")]; + tensor stacked_axis_0 = const()[name = tensor("stacked_axis_0"), val = tensor(1)]; + tensor stacked = stack(axis = stacked_axis_0, values = (var_19, var_29))[name = tensor("stacked")]; + tensor var_36 = const()[name = tensor("op_36"), val = tensor([1, 2, 345])]; + tensor input_1 = reshape(shape = var_36, x = stacked)[name = tensor("input_1")]; + tensor var_38 = const()[name = tensor("op_38"), val = tensor(0x1p+0)]; + tensor var_44 = const()[name = tensor("op_44"), val = tensor(true)]; + tensor var_45 = const()[name = tensor("op_45"), val = tensor(0x1.4f8b58p-17)]; + tensor var_48 = const()[name = tensor("op_48"), val = tensor(0)]; + tensor var_50 = const()[name = tensor("op_50"), val = tensor(2)]; + tensor var_51 = const()[name = tensor("op_51"), val = tensor(-1)]; + tensor var_53 = const()[name = tensor("op_53"), val = tensor(0x1.0c6f7ap-20)]; + tensor var_58 = const()[name = tensor("op_58"), val = tensor(0x1.5798eep-27)]; + tensor var_62 = const()[name = tensor("op_62"), val = tensor(1)]; + tensor input_3 = linear(bias = inner_encoder_input_projection_linear_bias, weight = inner_encoder_input_projection_linear_weight, x = input_1)[name = tensor("linear_0")]; tensor input_5_axes_0 = const()[name = tensor("input_5_axes_0"), val = tensor([-1])]; - tensor input_5 = layer_norm(axes = input_5_axes_0, beta = encoder_ffn1_0_module_sequential_0_bias, epsilon = var_28, gamma = encoder_ffn1_0_module_sequential_0_weight, x = input_3)[name = tensor("input_5")]; - tensor inputs_1 = linear(bias = encoder_ffn1_0_module_sequential_1_linear_bias, weight = encoder_ffn1_0_module_sequential_1_linear_weight, x = input_5)[name = tensor("linear_1")]; - tensor input_7 = silu(x = inputs_1)[name = tensor("input_7")]; - tensor input_11 = linear(bias = encoder_ffn1_0_module_sequential_4_linear_bias, weight = encoder_ffn1_0_module_sequential_4_linear_weight, x = input_7)[name = tensor("linear_2")]; - tensor var_147 = const()[name = tensor("op_147"), val = tensor(0x1p-1)]; - tensor var_148 = mul(x = input_11, y = var_147)[name = tensor("op_148")]; - tensor input_13 = add(x = var_148, y = input_3)[name = tensor("input_13")]; + tensor input_5 = layer_norm(axes = input_5_axes_0, beta = inner_encoder_pre_layer_norm_bias, epsilon = var_45, gamma = inner_encoder_pre_layer_norm_weight, x = input_3)[name = tensor("input_5")]; + tensor input_7_axes_0 = const()[name = tensor("input_7_axes_0"), val = tensor([-1])]; + tensor input_7 = layer_norm(axes = input_7_axes_0, beta = inner_encoder_ffn1_0_module_sequential_0_bias, epsilon = var_45, gamma = inner_encoder_ffn1_0_module_sequential_0_weight, x = input_5)[name = tensor("input_7")]; + tensor inputs_1 = linear(bias = inner_encoder_ffn1_0_module_sequential_1_linear_bias, weight = inner_encoder_ffn1_0_module_sequential_1_linear_weight, x = input_7)[name = tensor("linear_1")]; + tensor input_9 = silu(x = inputs_1)[name = tensor("input_9")]; + tensor input_13 = linear(bias = inner_encoder_ffn1_0_module_sequential_4_linear_bias, weight = inner_encoder_ffn1_0_module_sequential_4_linear_weight, x = input_9)[name = tensor("linear_2")]; + tensor var_183 = const()[name = tensor("op_183"), val = tensor(0x1p-1)]; + tensor var_184 = mul(x = input_13, y = var_183)[name = tensor("op_184")]; + tensor input_15 = add(x = var_184, y = input_5)[name = tensor("input_15")]; tensor x_1_axes_0 = const()[name = tensor("x_1_axes_0"), val = tensor([-1])]; - tensor x_1 = layer_norm(axes = x_1_axes_0, beta = encoder_ret_lns_0_bias, epsilon = var_28, gamma = encoder_ret_lns_0_weight, x = input_13)[name = tensor("x_1")]; + tensor x_1 = layer_norm(axes = x_1_axes_0, beta = inner_encoder_ret_lns_0_bias, epsilon = var_45, gamma = inner_encoder_ret_lns_0_weight, x = input_15)[name = tensor("x_1")]; tensor prev_kv_1_begin_0 = const()[name = tensor("prev_kv_1_begin_0"), val = tensor([0, 0, 0, 0, 0])]; tensor prev_kv_1_end_0 = const()[name = tensor("prev_kv_1_end_0"), val = tensor([1, 1, 4, 64, 64])]; tensor prev_kv_1_end_mask_0 = const()[name = tensor("prev_kv_1_end_mask_0"), val = tensor([false, true, true, true, true])]; @@ -239,153 +253,153 @@ program(1.0) tensor prev_scale_1_end_mask_0 = const()[name = tensor("prev_scale_1_end_mask_0"), val = tensor([false, true])]; tensor prev_scale_1_squeeze_mask_0 = const()[name = tensor("prev_scale_1_squeeze_mask_0"), val = tensor([true, false])]; tensor prev_scale_1 = slice_by_index(begin = prev_scale_1_begin_0, end = prev_scale_1_end_0, end_mask = prev_scale_1_end_mask_0, squeeze_mask = prev_scale_1_squeeze_mask_0, x = enc_scale)[name = tensor("prev_scale_1")]; - tensor var_162 = linear(bias = encoder_q_proj_0_bias, weight = encoder_q_proj_0_weight, x = x_1)[name = tensor("linear_3")]; - tensor var_163 = const()[name = tensor("op_163"), val = tensor([1, 2, 4, 64])]; - tensor var_164 = reshape(shape = var_163, x = var_162)[name = tensor("op_164")]; + tensor var_198 = linear(bias = inner_encoder_q_proj_0_bias, weight = inner_encoder_q_proj_0_weight, x = x_1)[name = tensor("linear_3")]; + tensor var_199 = const()[name = tensor("op_199"), val = tensor([1, 2, 4, 64])]; + tensor var_200 = reshape(shape = var_199, x = var_198)[name = tensor("op_200")]; tensor q_1_perm_0 = const()[name = tensor("q_1_perm_0"), val = tensor([0, 2, 1, 3])]; - tensor var_168 = linear(bias = encoder_k_proj_0_bias, weight = encoder_k_proj_0_weight, x = x_1)[name = tensor("linear_4")]; - tensor var_169 = const()[name = tensor("op_169"), val = tensor(0x1p-3)]; - tensor var_170 = mul(x = var_168, y = var_169)[name = tensor("op_170")]; - tensor var_171 = const()[name = tensor("op_171"), val = tensor([1, 2, 4, 64])]; - tensor var_172 = reshape(shape = var_171, x = var_170)[name = tensor("op_172")]; + tensor var_204 = linear(bias = inner_encoder_k_proj_0_bias, weight = inner_encoder_k_proj_0_weight, x = x_1)[name = tensor("linear_4")]; + tensor var_205 = const()[name = tensor("op_205"), val = tensor(0x1p-3)]; + tensor var_206 = mul(x = var_204, y = var_205)[name = tensor("op_206")]; + tensor var_207 = const()[name = tensor("op_207"), val = tensor([1, 2, 4, 64])]; + tensor var_208 = reshape(shape = var_207, x = var_206)[name = tensor("op_208")]; tensor k_1_perm_0 = const()[name = tensor("k_1_perm_0"), val = tensor([0, 2, 1, 3])]; - tensor var_176 = linear(bias = encoder_v_proj_0_bias, weight = encoder_v_proj_0_weight, x = x_1)[name = tensor("linear_5")]; - tensor var_177 = const()[name = tensor("op_177"), val = tensor([1, 2, 4, 64])]; - tensor var_178 = reshape(shape = var_177, x = var_176)[name = tensor("op_178")]; + tensor var_212 = linear(bias = inner_encoder_v_proj_0_bias, weight = inner_encoder_v_proj_0_weight, x = x_1)[name = tensor("linear_5")]; + tensor var_213 = const()[name = tensor("op_213"), val = tensor([1, 2, 4, 64])]; + tensor var_214 = reshape(shape = var_213, x = var_212)[name = tensor("op_214")]; tensor v_1_perm_0 = const()[name = tensor("v_1_perm_0"), val = tensor([0, 2, 1, 3])]; - tensor input_17 = linear(bias = encoder_g_proj_0_bias, weight = encoder_g_proj_0_weight, x = x_1)[name = tensor("linear_6")]; + tensor input_19 = linear(bias = inner_encoder_g_proj_0_bias, weight = inner_encoder_g_proj_0_weight, x = x_1)[name = tensor("linear_6")]; tensor sqrt_s0_1 = sqrt(x = prev_scale_1)[name = tensor("sqrt_s0_1")]; - tensor s_t_1 = add(x = prev_scale_1, y = encoder__t_index)[name = tensor("s_t_1")]; + tensor s_t_1 = add(x = prev_scale_1, y = inner_encoder__t_index)[name = tensor("s_t_1")]; tensor sqrt_s_t_1 = sqrt(x = s_t_1)[name = tensor("sqrt_s_t_1")]; tensor qk_1_transpose_x_1 = const()[name = tensor("qk_1_transpose_x_1"), val = tensor(false)]; tensor qk_1_transpose_y_1 = const()[name = tensor("qk_1_transpose_y_1"), val = tensor(true)]; - tensor k_1 = transpose(perm = k_1_perm_0, x = var_172)[name = tensor("transpose_57")]; - tensor q_1 = transpose(perm = q_1_perm_0, x = var_164)[name = tensor("transpose_58")]; + tensor k_1 = transpose(perm = k_1_perm_0, x = var_208)[name = tensor("transpose_57")]; + tensor q_1 = transpose(perm = q_1_perm_0, x = var_200)[name = tensor("transpose_58")]; tensor qk_1 = matmul(transpose_x = qk_1_transpose_x_1, transpose_y = qk_1_transpose_y_1, x = q_1, y = k_1)[name = tensor("qk_1")]; - tensor var_188 = const()[name = tensor("op_188"), val = tensor([2, 1])]; - tensor var_189 = reshape(shape = var_188, x = sqrt_s_t_1)[name = tensor("op_189")]; - tensor M_1 = real_div(x = encoder__causal_mask, y = var_189)[name = tensor("M_1")]; - tensor var_191 = mul(x = qk_1, y = M_1)[name = tensor("op_191")]; + tensor var_224 = const()[name = tensor("op_224"), val = tensor([2, 1])]; + tensor var_225 = reshape(shape = var_224, x = sqrt_s_t_1)[name = tensor("op_225")]; + tensor M_1 = real_div(x = inner_encoder__causal_mask, y = var_225)[name = tensor("M_1")]; + tensor var_227 = mul(x = qk_1, y = M_1)[name = tensor("op_227")]; tensor inner_1_transpose_x_0 = const()[name = tensor("inner_1_transpose_x_0"), val = tensor(false)]; tensor inner_1_transpose_y_0 = const()[name = tensor("inner_1_transpose_y_0"), val = tensor(false)]; - tensor v_1 = transpose(perm = v_1_perm_0, x = var_178)[name = tensor("transpose_56")]; - tensor inner_1 = matmul(transpose_x = inner_1_transpose_x_0, transpose_y = inner_1_transpose_y_0, x = var_191, y = v_1)[name = tensor("inner_1")]; - tensor var_193_transpose_x_0 = const()[name = tensor("op_193_transpose_x_0"), val = tensor(false)]; - tensor var_193_transpose_y_0 = const()[name = tensor("op_193_transpose_y_0"), val = tensor(false)]; - tensor var_193 = matmul(transpose_x = var_193_transpose_x_0, transpose_y = var_193_transpose_y_0, x = q_1, y = prev_kv_1)[name = tensor("op_193")]; - tensor var_194 = real_div(x = sqrt_s0_1, y = sqrt_s_t_1)[name = tensor("op_194")]; - tensor var_195 = const()[name = tensor("op_195"), val = tensor([1, 1, 2, 1])]; - tensor var_196 = reshape(shape = var_195, x = var_194)[name = tensor("op_196")]; - tensor cross_1 = mul(x = var_193, y = var_196)[name = tensor("cross_1")]; + tensor v_1 = transpose(perm = v_1_perm_0, x = var_214)[name = tensor("transpose_56")]; + tensor inner_1 = matmul(transpose_x = inner_1_transpose_x_0, transpose_y = inner_1_transpose_y_0, x = var_227, y = v_1)[name = tensor("inner_1")]; + tensor var_229_transpose_x_0 = const()[name = tensor("op_229_transpose_x_0"), val = tensor(false)]; + tensor var_229_transpose_y_0 = const()[name = tensor("op_229_transpose_y_0"), val = tensor(false)]; + tensor var_229 = matmul(transpose_x = var_229_transpose_x_0, transpose_y = var_229_transpose_y_0, x = q_1, y = prev_kv_1)[name = tensor("op_229")]; + tensor var_230 = real_div(x = sqrt_s0_1, y = sqrt_s_t_1)[name = tensor("op_230")]; + tensor var_231 = const()[name = tensor("op_231"), val = tensor([1, 1, 2, 1])]; + tensor var_232 = reshape(shape = var_231, x = var_230)[name = tensor("op_232")]; + tensor cross_1 = mul(x = var_229, y = var_232)[name = tensor("cross_1")]; tensor out_1 = add(x = inner_1, y = cross_1)[name = tensor("out_1")]; - tensor var_199 = mul(x = prev_kv_1, y = sqrt_s0_1)[name = tensor("op_199")]; - tensor var_201_transpose_x_1 = const()[name = tensor("op_201_transpose_x_1"), val = tensor(true)]; - tensor var_201_transpose_y_1 = const()[name = tensor("op_201_transpose_y_1"), val = tensor(false)]; - tensor var_201 = matmul(transpose_x = var_201_transpose_x_1, transpose_y = var_201_transpose_y_1, x = k_1, y = v_1)[name = tensor("op_201")]; - tensor new_kv_unnorm_1 = add(x = var_199, y = var_201)[name = tensor("new_kv_unnorm_1")]; - tensor var_203 = const()[name = tensor("op_203"), val = tensor(0x1p+1)]; - tensor new_scale_1 = add(x = prev_scale_1, y = var_203)[name = tensor("new_scale_1")]; - tensor var_205 = sqrt(x = new_scale_1)[name = tensor("op_205")]; - tensor var_206 = real_div(x = new_kv_unnorm_1, y = var_205)[name = tensor("op_206")]; - tensor var_207_perm_0 = const()[name = tensor("op_207_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_235 = mul(x = prev_kv_1, y = sqrt_s0_1)[name = tensor("op_235")]; + tensor var_237_transpose_x_1 = const()[name = tensor("op_237_transpose_x_1"), val = tensor(true)]; + tensor var_237_transpose_y_1 = const()[name = tensor("op_237_transpose_y_1"), val = tensor(false)]; + tensor var_237 = matmul(transpose_x = var_237_transpose_x_1, transpose_y = var_237_transpose_y_1, x = k_1, y = v_1)[name = tensor("op_237")]; + tensor new_kv_unnorm_1 = add(x = var_235, y = var_237)[name = tensor("new_kv_unnorm_1")]; + tensor var_239 = const()[name = tensor("op_239"), val = tensor(0x1p+1)]; + tensor new_scale_1 = add(x = prev_scale_1, y = var_239)[name = tensor("new_scale_1")]; + tensor var_241 = sqrt(x = new_scale_1)[name = tensor("op_241")]; + tensor var_242 = real_div(x = new_kv_unnorm_1, y = var_241)[name = tensor("op_242")]; + tensor var_243_perm_0 = const()[name = tensor("op_243_perm_0"), val = tensor([0, 2, 1, 3])]; tensor out_3_axes_0 = const()[name = tensor("out_3_axes_0"), val = tensor([-1])]; - tensor var_207 = transpose(perm = var_207_perm_0, x = out_1)[name = tensor("transpose_55")]; - tensor out_3 = layer_norm(axes = out_3_axes_0, epsilon = var_17, x = var_207)[name = tensor("out_3")]; - tensor var_211 = const()[name = tensor("op_211"), val = tensor([1, 2, 256])]; - tensor out_5 = reshape(shape = var_211, x = out_3)[name = tensor("out_5")]; - tensor var_213 = silu(x = input_17)[name = tensor("op_213")]; - tensor input_19 = mul(x = var_213, y = out_5)[name = tensor("input_19")]; - tensor ret_out_1 = linear(bias = encoder_out_proj_0_bias, weight = encoder_out_proj_0_weight, x = input_19)[name = tensor("linear_7")]; - tensor x_3 = add(x = input_13, y = ret_out_1)[name = tensor("x_3")]; + tensor var_243 = transpose(perm = var_243_perm_0, x = out_1)[name = tensor("transpose_55")]; + tensor out_3 = layer_norm(axes = out_3_axes_0, epsilon = var_53, x = var_243)[name = tensor("out_3")]; + tensor var_247 = const()[name = tensor("op_247"), val = tensor([1, 2, 256])]; + tensor out_5 = reshape(shape = var_247, x = out_3)[name = tensor("out_5")]; + tensor var_249 = silu(x = input_19)[name = tensor("op_249")]; + tensor input_21 = mul(x = var_249, y = out_5)[name = tensor("input_21")]; + tensor ret_out_1 = linear(bias = inner_encoder_out_proj_0_bias, weight = inner_encoder_out_proj_0_weight, x = input_21)[name = tensor("linear_7")]; + tensor x_3 = add(x = input_15, y = ret_out_1)[name = tensor("x_3")]; tensor window_1_begin_0 = const()[name = tensor("window_1_begin_0"), val = tensor([0, 0, 0, 0])]; tensor window_1_end_0 = const()[name = tensor("window_1_end_0"), val = tensor([1, 1, 16, 256])]; tensor window_1_end_mask_0 = const()[name = tensor("window_1_end_mask_0"), val = tensor([false, true, true, true])]; tensor window_1_squeeze_mask_0 = const()[name = tensor("window_1_squeeze_mask_0"), val = tensor([true, false, false, false])]; tensor window_1 = slice_by_index(begin = window_1_begin_0, end = window_1_end_0, end_mask = window_1_end_mask_0, squeeze_mask = window_1_squeeze_mask_0, x = enc_conv_cache)[name = tensor("window_1")]; - tensor var_221_begin_0 = const()[name = tensor("op_221_begin_0"), val = tensor([0, 0, 0])]; - tensor var_221_end_0 = const()[name = tensor("op_221_end_0"), val = tensor([1, 1, 256])]; - tensor var_221_end_mask_0 = const()[name = tensor("op_221_end_mask_0"), val = tensor([true, false, true])]; - tensor var_221 = slice_by_index(begin = var_221_begin_0, end = var_221_end_0, end_mask = var_221_end_mask_0, x = x_3)[name = tensor("op_221")]; - tensor var_224_begin_0 = const()[name = tensor("op_224_begin_0"), val = tensor([0, 1, 0])]; - tensor var_224_end_0 = const()[name = tensor("op_224_end_0"), val = tensor([1, 16, 256])]; - tensor var_224_end_mask_0 = const()[name = tensor("op_224_end_mask_0"), val = tensor([true, true, true])]; - tensor var_224 = slice_by_index(begin = var_224_begin_0, end = var_224_end_0, end_mask = var_224_end_mask_0, x = window_1)[name = tensor("op_224")]; + tensor var_257_begin_0 = const()[name = tensor("op_257_begin_0"), val = tensor([0, 0, 0])]; + tensor var_257_end_0 = const()[name = tensor("op_257_end_0"), val = tensor([1, 1, 256])]; + tensor var_257_end_mask_0 = const()[name = tensor("op_257_end_mask_0"), val = tensor([true, false, true])]; + tensor var_257 = slice_by_index(begin = var_257_begin_0, end = var_257_end_0, end_mask = var_257_end_mask_0, x = x_3)[name = tensor("op_257")]; + tensor var_260_begin_0 = const()[name = tensor("op_260_begin_0"), val = tensor([0, 1, 0])]; + tensor var_260_end_0 = const()[name = tensor("op_260_end_0"), val = tensor([1, 16, 256])]; + tensor var_260_end_mask_0 = const()[name = tensor("op_260_end_mask_0"), val = tensor([true, true, true])]; + tensor var_260 = slice_by_index(begin = var_260_begin_0, end = var_260_end_0, end_mask = var_260_end_mask_0, x = window_1)[name = tensor("op_260")]; tensor window_3_interleave_0 = const()[name = tensor("window_3_interleave_0"), val = tensor(false)]; - tensor window_3 = concat(axis = var_26, interleave = window_3_interleave_0, values = (var_224, var_221))[name = tensor("window_3")]; - tensor var_229_begin_0 = const()[name = tensor("op_229_begin_0"), val = tensor([0, 1, 0])]; - tensor var_229_end_0 = const()[name = tensor("op_229_end_0"), val = tensor([1, 1, 256])]; - tensor var_229_end_mask_0 = const()[name = tensor("op_229_end_mask_0"), val = tensor([true, true, true])]; - tensor var_229 = slice_by_index(begin = var_229_begin_0, end = var_229_end_0, end_mask = var_229_end_mask_0, x = x_3)[name = tensor("op_229")]; - tensor var_232_begin_0 = const()[name = tensor("op_232_begin_0"), val = tensor([0, 1, 0])]; - tensor var_232_end_0 = const()[name = tensor("op_232_end_0"), val = tensor([1, 16, 256])]; - tensor var_232_end_mask_0 = const()[name = tensor("op_232_end_mask_0"), val = tensor([true, true, true])]; - tensor var_232 = slice_by_index(begin = var_232_begin_0, end = var_232_end_0, end_mask = var_232_end_mask_0, x = window_3)[name = tensor("op_232")]; + tensor window_3 = concat(axis = var_62, interleave = window_3_interleave_0, values = (var_260, var_257))[name = tensor("window_3")]; + tensor var_265_begin_0 = const()[name = tensor("op_265_begin_0"), val = tensor([0, 1, 0])]; + tensor var_265_end_0 = const()[name = tensor("op_265_end_0"), val = tensor([1, 1, 256])]; + tensor var_265_end_mask_0 = const()[name = tensor("op_265_end_mask_0"), val = tensor([true, true, true])]; + tensor var_265 = slice_by_index(begin = var_265_begin_0, end = var_265_end_0, end_mask = var_265_end_mask_0, x = x_3)[name = tensor("op_265")]; + tensor var_268_begin_0 = const()[name = tensor("op_268_begin_0"), val = tensor([0, 1, 0])]; + tensor var_268_end_0 = const()[name = tensor("op_268_end_0"), val = tensor([1, 16, 256])]; + tensor var_268_end_mask_0 = const()[name = tensor("op_268_end_mask_0"), val = tensor([true, true, true])]; + tensor var_268 = slice_by_index(begin = var_268_begin_0, end = var_268_end_0, end_mask = var_268_end_mask_0, x = window_3)[name = tensor("op_268")]; tensor window_5_interleave_0 = const()[name = tensor("window_5_interleave_0"), val = tensor(false)]; - tensor window_5 = concat(axis = var_26, interleave = window_5_interleave_0, values = (var_232, var_229))[name = tensor("window_5")]; - tensor input_21_interleave_0 = const()[name = tensor("input_21_interleave_0"), val = tensor(false)]; - tensor input_21 = concat(axis = var_23, interleave = input_21_interleave_0, values = (window_3, window_5))[name = tensor("input_21")]; + tensor window_5 = concat(axis = var_62, interleave = window_5_interleave_0, values = (var_268, var_265))[name = tensor("window_5")]; + tensor input_23_interleave_0 = const()[name = tensor("input_23_interleave_0"), val = tensor(false)]; + tensor input_23 = concat(axis = var_48, interleave = input_23_interleave_0, values = (window_3, window_5))[name = tensor("input_23")]; tensor x_5_axes_0 = const()[name = tensor("x_5_axes_0"), val = tensor([-1])]; - tensor x_5 = layer_norm(axes = x_5_axes_0, beta = encoder_conv_module_0_sequential_0_bias, epsilon = var_28, gamma = encoder_conv_module_0_sequential_0_weight, x = input_21)[name = tensor("x_5")]; - tensor input_23_perm_0 = const()[name = tensor("input_23_perm_0"), val = tensor([0, 2, 1])]; + tensor x_5 = layer_norm(axes = x_5_axes_0, beta = inner_encoder_conv_module_0_sequential_0_bias, epsilon = var_45, gamma = inner_encoder_conv_module_0_sequential_0_weight, x = input_23)[name = tensor("x_5")]; + tensor input_25_perm_0 = const()[name = tensor("input_25_perm_0"), val = tensor([0, 2, 1])]; tensor inputs_3_pad_type_0 = const()[name = tensor("inputs_3_pad_type_0"), val = tensor("valid")]; tensor inputs_3_strides_0 = const()[name = tensor("inputs_3_strides_0"), val = tensor([1])]; tensor inputs_3_pad_0 = const()[name = tensor("inputs_3_pad_0"), val = tensor([0, 0])]; tensor inputs_3_dilations_0 = const()[name = tensor("inputs_3_dilations_0"), val = tensor([1])]; tensor inputs_3_groups_0 = const()[name = tensor("inputs_3_groups_0"), val = tensor(1)]; - tensor input_23 = transpose(perm = input_23_perm_0, x = x_5)[name = tensor("transpose_54")]; - tensor inputs_3 = conv(bias = encoder_conv_module_0_sequential_2_conv_bias, dilations = inputs_3_dilations_0, groups = inputs_3_groups_0, pad = inputs_3_pad_0, pad_type = inputs_3_pad_type_0, strides = inputs_3_strides_0, weight = encoder_conv_module_0_sequential_2_conv_weight, x = input_23)[name = tensor("inputs_3")]; - tensor var_257_split_sizes_0 = const()[name = tensor("op_257_split_sizes_0"), val = tensor([256, 256])]; - tensor var_257_axis_0 = const()[name = tensor("op_257_axis_0"), val = tensor(1)]; - tensor var_257_0, tensor var_257_1 = split(axis = var_257_axis_0, split_sizes = var_257_split_sizes_0, x = inputs_3)[name = tensor("op_257")]; - tensor var_259 = sigmoid(x = var_257_1)[name = tensor("op_259")]; - tensor inputs_5 = mul(x = var_257_0, y = var_259)[name = tensor("inputs_5")]; + tensor input_25 = transpose(perm = input_25_perm_0, x = x_5)[name = tensor("transpose_54")]; + tensor inputs_3 = conv(bias = inner_encoder_conv_module_0_sequential_2_conv_bias, dilations = inputs_3_dilations_0, groups = inputs_3_groups_0, pad = inputs_3_pad_0, pad_type = inputs_3_pad_type_0, strides = inputs_3_strides_0, weight = inner_encoder_conv_module_0_sequential_2_conv_weight, x = input_25)[name = tensor("inputs_3")]; + tensor var_293_split_sizes_0 = const()[name = tensor("op_293_split_sizes_0"), val = tensor([256, 256])]; + tensor var_293_axis_0 = const()[name = tensor("op_293_axis_0"), val = tensor(1)]; + tensor var_293_0, tensor var_293_1 = split(axis = var_293_axis_0, split_sizes = var_293_split_sizes_0, x = inputs_3)[name = tensor("op_293")]; + tensor var_295 = sigmoid(x = var_293_1)[name = tensor("op_295")]; + tensor inputs_5 = mul(x = var_293_0, y = var_295)[name = tensor("inputs_5")]; tensor outputs_aug_1_pad_type_0 = const()[name = tensor("outputs_aug_1_pad_type_0"), val = tensor("custom")]; tensor outputs_aug_1_pad_0 = const()[name = tensor("outputs_aug_1_pad_0"), val = tensor([15, 15])]; tensor outputs_aug_1_groups_0 = const()[name = tensor("outputs_aug_1_groups_0"), val = tensor(256)]; tensor outputs_aug_1_strides_0 = const()[name = tensor("outputs_aug_1_strides_0"), val = tensor([1])]; tensor outputs_aug_1_dilations_0 = const()[name = tensor("outputs_aug_1_dilations_0"), val = tensor([1])]; - tensor outputs_aug_1 = conv(dilations = outputs_aug_1_dilations_0, groups = outputs_aug_1_groups_0, pad = outputs_aug_1_pad_0, pad_type = outputs_aug_1_pad_type_0, strides = outputs_aug_1_strides_0, weight = encoder_conv_module_0_sequential_4_conv_weight, x = inputs_5)[name = tensor("outputs_aug_1")]; - tensor input_25_begin_0 = const()[name = tensor("input_25_begin_0"), val = tensor([0, 0, 0])]; - tensor input_25_end_0 = const()[name = tensor("input_25_end_0"), val = tensor([2, 256, 16])]; - tensor input_25_end_mask_0 = const()[name = tensor("input_25_end_mask_0"), val = tensor([true, true, false])]; - tensor input_25 = slice_by_index(begin = input_25_begin_0, end = input_25_end_0, end_mask = input_25_end_mask_0, x = outputs_aug_1)[name = tensor("input_25")]; - tensor inputs_7 = batch_norm(beta = encoder_conv_module_0_sequential_5_bias, epsilon = var_28, gamma = encoder_conv_module_0_sequential_5_weight, mean = encoder_conv_module_0_sequential_5_running_mean, variance = encoder_conv_module_0_sequential_5_running_var, x = input_25)[name = tensor("inputs_7")]; - tensor input_27 = silu(x = inputs_7)[name = tensor("input_27")]; - tensor input_29_pad_type_0 = const()[name = tensor("input_29_pad_type_0"), val = tensor("valid")]; - tensor input_29_strides_0 = const()[name = tensor("input_29_strides_0"), val = tensor([1])]; - tensor input_29_pad_0 = const()[name = tensor("input_29_pad_0"), val = tensor([0, 0])]; - tensor input_29_dilations_0 = const()[name = tensor("input_29_dilations_0"), val = tensor([1])]; - tensor input_29_groups_0 = const()[name = tensor("input_29_groups_0"), val = tensor(1)]; - tensor input_29 = conv(bias = encoder_conv_module_0_sequential_7_conv_bias, dilations = input_29_dilations_0, groups = input_29_groups_0, pad = input_29_pad_0, pad_type = input_29_pad_type_0, strides = input_29_strides_0, weight = encoder_conv_module_0_sequential_7_conv_weight, x = input_27)[name = tensor("input_29")]; + tensor outputs_aug_1 = conv(dilations = outputs_aug_1_dilations_0, groups = outputs_aug_1_groups_0, pad = outputs_aug_1_pad_0, pad_type = outputs_aug_1_pad_type_0, strides = outputs_aug_1_strides_0, weight = inner_encoder_conv_module_0_sequential_4_conv_weight, x = inputs_5)[name = tensor("outputs_aug_1")]; + tensor input_27_begin_0 = const()[name = tensor("input_27_begin_0"), val = tensor([0, 0, 0])]; + tensor input_27_end_0 = const()[name = tensor("input_27_end_0"), val = tensor([2, 256, 16])]; + tensor input_27_end_mask_0 = const()[name = tensor("input_27_end_mask_0"), val = tensor([true, true, false])]; + tensor input_27 = slice_by_index(begin = input_27_begin_0, end = input_27_end_0, end_mask = input_27_end_mask_0, x = outputs_aug_1)[name = tensor("input_27")]; + tensor inputs_7 = batch_norm(beta = inner_encoder_conv_module_0_sequential_5_bias, epsilon = var_45, gamma = inner_encoder_conv_module_0_sequential_5_weight, mean = inner_encoder_conv_module_0_sequential_5_running_mean, variance = inner_encoder_conv_module_0_sequential_5_running_var, x = input_27)[name = tensor("inputs_7")]; + tensor input_29 = silu(x = inputs_7)[name = tensor("input_29")]; + tensor input_31_pad_type_0 = const()[name = tensor("input_31_pad_type_0"), val = tensor("valid")]; + tensor input_31_strides_0 = const()[name = tensor("input_31_strides_0"), val = tensor([1])]; + tensor input_31_pad_0 = const()[name = tensor("input_31_pad_0"), val = tensor([0, 0])]; + tensor input_31_dilations_0 = const()[name = tensor("input_31_dilations_0"), val = tensor([1])]; + tensor input_31_groups_0 = const()[name = tensor("input_31_groups_0"), val = tensor(1)]; + tensor input_31 = conv(bias = inner_encoder_conv_module_0_sequential_7_conv_bias, dilations = input_31_dilations_0, groups = input_31_groups_0, pad = input_31_pad_0, pad_type = input_31_pad_type_0, strides = input_31_strides_0, weight = inner_encoder_conv_module_0_sequential_7_conv_weight, x = input_29)[name = tensor("input_31")]; tensor conv_out_1_perm_0 = const()[name = tensor("conv_out_1_perm_0"), val = tensor([0, 2, 1])]; - tensor var_290_begin_0 = const()[name = tensor("op_290_begin_0"), val = tensor([0, -1, 0])]; - tensor var_290_end_0 = const()[name = tensor("op_290_end_0"), val = tensor([2, 16, 256])]; - tensor var_290_end_mask_0 = const()[name = tensor("op_290_end_mask_0"), val = tensor([true, true, true])]; - tensor conv_out_1 = transpose(perm = conv_out_1_perm_0, x = input_29)[name = tensor("transpose_53")]; - tensor var_290 = slice_by_index(begin = var_290_begin_0, end = var_290_end_0, end_mask = var_290_end_mask_0, x = conv_out_1)[name = tensor("op_290")]; - tensor var_292_perm_0 = const()[name = tensor("op_292_perm_0"), val = tensor([1, 0, 2])]; - tensor var_292 = transpose(perm = var_292_perm_0, x = var_290)[name = tensor("transpose_52")]; - tensor input_31 = add(x = x_3, y = var_292)[name = tensor("input_31")]; - tensor input_33_axes_0 = const()[name = tensor("input_33_axes_0"), val = tensor([-1])]; - tensor input_33 = layer_norm(axes = input_33_axes_0, beta = encoder_ffn2_0_module_sequential_0_bias, epsilon = var_28, gamma = encoder_ffn2_0_module_sequential_0_weight, x = input_31)[name = tensor("input_33")]; - tensor inputs_9 = linear(bias = encoder_ffn2_0_module_sequential_1_linear_bias, weight = encoder_ffn2_0_module_sequential_1_linear_weight, x = input_33)[name = tensor("linear_8")]; - tensor input_35 = silu(x = inputs_9)[name = tensor("input_35")]; - tensor input_39 = linear(bias = encoder_ffn2_0_module_sequential_4_linear_bias, weight = encoder_ffn2_0_module_sequential_4_linear_weight, x = input_35)[name = tensor("linear_9")]; - tensor var_315 = const()[name = tensor("op_315"), val = tensor(0x1p-1)]; - tensor var_316 = mul(x = input_39, y = var_315)[name = tensor("op_316")]; - tensor input_41 = add(x = var_316, y = input_31)[name = tensor("input_41")]; - tensor input_43_axes_0 = const()[name = tensor("input_43_axes_0"), val = tensor([-1])]; - tensor input_43 = layer_norm(axes = input_43_axes_0, beta = encoder_layer_norm_0_bias, epsilon = var_28, gamma = encoder_layer_norm_0_weight, x = input_41)[name = tensor("input_43")]; + tensor var_326_begin_0 = const()[name = tensor("op_326_begin_0"), val = tensor([0, -1, 0])]; + tensor var_326_end_0 = const()[name = tensor("op_326_end_0"), val = tensor([2, 16, 256])]; + tensor var_326_end_mask_0 = const()[name = tensor("op_326_end_mask_0"), val = tensor([true, true, true])]; + tensor conv_out_1 = transpose(perm = conv_out_1_perm_0, x = input_31)[name = tensor("transpose_53")]; + tensor var_326 = slice_by_index(begin = var_326_begin_0, end = var_326_end_0, end_mask = var_326_end_mask_0, x = conv_out_1)[name = tensor("op_326")]; + tensor var_328_perm_0 = const()[name = tensor("op_328_perm_0"), val = tensor([1, 0, 2])]; + tensor var_328 = transpose(perm = var_328_perm_0, x = var_326)[name = tensor("transpose_52")]; + tensor input_33 = add(x = x_3, y = var_328)[name = tensor("input_33")]; + tensor input_35_axes_0 = const()[name = tensor("input_35_axes_0"), val = tensor([-1])]; + tensor input_35 = layer_norm(axes = input_35_axes_0, beta = inner_encoder_ffn2_0_module_sequential_0_bias, epsilon = var_45, gamma = inner_encoder_ffn2_0_module_sequential_0_weight, x = input_33)[name = tensor("input_35")]; + tensor inputs_9 = linear(bias = inner_encoder_ffn2_0_module_sequential_1_linear_bias, weight = inner_encoder_ffn2_0_module_sequential_1_linear_weight, x = input_35)[name = tensor("linear_8")]; + tensor input_37 = silu(x = inputs_9)[name = tensor("input_37")]; + tensor input_41 = linear(bias = inner_encoder_ffn2_0_module_sequential_4_linear_bias, weight = inner_encoder_ffn2_0_module_sequential_4_linear_weight, x = input_37)[name = tensor("linear_9")]; + tensor var_351 = const()[name = tensor("op_351"), val = tensor(0x1p-1)]; + tensor var_352 = mul(x = input_41, y = var_351)[name = tensor("op_352")]; + tensor input_43 = add(x = var_352, y = input_33)[name = tensor("input_43")]; tensor input_45_axes_0 = const()[name = tensor("input_45_axes_0"), val = tensor([-1])]; - tensor input_45 = layer_norm(axes = input_45_axes_0, beta = encoder_ffn1_1_module_sequential_0_bias, epsilon = var_28, gamma = encoder_ffn1_1_module_sequential_0_weight, x = input_43)[name = tensor("input_45")]; - tensor inputs_11 = linear(bias = encoder_ffn1_1_module_sequential_1_linear_bias, weight = encoder_ffn1_1_module_sequential_1_linear_weight, x = input_45)[name = tensor("linear_10")]; - tensor input_47 = silu(x = inputs_11)[name = tensor("input_47")]; - tensor input_51 = linear(bias = encoder_ffn1_1_module_sequential_4_linear_bias, weight = encoder_ffn1_1_module_sequential_4_linear_weight, x = input_47)[name = tensor("linear_11")]; - tensor var_345 = const()[name = tensor("op_345"), val = tensor(0x1p-1)]; - tensor var_346 = mul(x = input_51, y = var_345)[name = tensor("op_346")]; - tensor input_53 = add(x = var_346, y = input_43)[name = tensor("input_53")]; + tensor input_45 = layer_norm(axes = input_45_axes_0, beta = inner_encoder_layer_norm_0_bias, epsilon = var_45, gamma = inner_encoder_layer_norm_0_weight, x = input_43)[name = tensor("input_45")]; + tensor input_47_axes_0 = const()[name = tensor("input_47_axes_0"), val = tensor([-1])]; + tensor input_47 = layer_norm(axes = input_47_axes_0, beta = inner_encoder_ffn1_1_module_sequential_0_bias, epsilon = var_45, gamma = inner_encoder_ffn1_1_module_sequential_0_weight, x = input_45)[name = tensor("input_47")]; + tensor inputs_11 = linear(bias = inner_encoder_ffn1_1_module_sequential_1_linear_bias, weight = inner_encoder_ffn1_1_module_sequential_1_linear_weight, x = input_47)[name = tensor("linear_10")]; + tensor input_49 = silu(x = inputs_11)[name = tensor("input_49")]; + tensor input_53 = linear(bias = inner_encoder_ffn1_1_module_sequential_4_linear_bias, weight = inner_encoder_ffn1_1_module_sequential_4_linear_weight, x = input_49)[name = tensor("linear_11")]; + tensor var_381 = const()[name = tensor("op_381"), val = tensor(0x1p-1)]; + tensor var_382 = mul(x = input_53, y = var_381)[name = tensor("op_382")]; + tensor input_55 = add(x = var_382, y = input_45)[name = tensor("input_55")]; tensor x_7_axes_0 = const()[name = tensor("x_7_axes_0"), val = tensor([-1])]; - tensor x_7 = layer_norm(axes = x_7_axes_0, beta = encoder_ret_lns_1_bias, epsilon = var_28, gamma = encoder_ret_lns_1_weight, x = input_53)[name = tensor("x_7")]; + tensor x_7 = layer_norm(axes = x_7_axes_0, beta = inner_encoder_ret_lns_1_bias, epsilon = var_45, gamma = inner_encoder_ret_lns_1_weight, x = input_55)[name = tensor("x_7")]; tensor prev_kv_3_begin_0 = const()[name = tensor("prev_kv_3_begin_0"), val = tensor([1, 0, 0, 0, 0])]; tensor prev_kv_3_end_0 = const()[name = tensor("prev_kv_3_end_0"), val = tensor([2, 1, 4, 64, 64])]; tensor prev_kv_3_end_mask_0 = const()[name = tensor("prev_kv_3_end_mask_0"), val = tensor([false, true, true, true, true])]; @@ -396,153 +410,153 @@ program(1.0) tensor prev_scale_3_end_mask_0 = const()[name = tensor("prev_scale_3_end_mask_0"), val = tensor([false, true])]; tensor prev_scale_3_squeeze_mask_0 = const()[name = tensor("prev_scale_3_squeeze_mask_0"), val = tensor([true, false])]; tensor prev_scale_3 = slice_by_index(begin = prev_scale_3_begin_0, end = prev_scale_3_end_0, end_mask = prev_scale_3_end_mask_0, squeeze_mask = prev_scale_3_squeeze_mask_0, x = enc_scale)[name = tensor("prev_scale_3")]; - tensor var_360 = linear(bias = encoder_q_proj_1_bias, weight = encoder_q_proj_1_weight, x = x_7)[name = tensor("linear_12")]; - tensor var_361 = const()[name = tensor("op_361"), val = tensor([1, 2, 4, 64])]; - tensor var_362 = reshape(shape = var_361, x = var_360)[name = tensor("op_362")]; + tensor var_396 = linear(bias = inner_encoder_q_proj_1_bias, weight = inner_encoder_q_proj_1_weight, x = x_7)[name = tensor("linear_12")]; + tensor var_397 = const()[name = tensor("op_397"), val = tensor([1, 2, 4, 64])]; + tensor var_398 = reshape(shape = var_397, x = var_396)[name = tensor("op_398")]; tensor q_3_perm_0 = const()[name = tensor("q_3_perm_0"), val = tensor([0, 2, 1, 3])]; - tensor var_366 = linear(bias = encoder_k_proj_1_bias, weight = encoder_k_proj_1_weight, x = x_7)[name = tensor("linear_13")]; - tensor var_367 = const()[name = tensor("op_367"), val = tensor(0x1p-3)]; - tensor var_368 = mul(x = var_366, y = var_367)[name = tensor("op_368")]; - tensor var_369 = const()[name = tensor("op_369"), val = tensor([1, 2, 4, 64])]; - tensor var_370 = reshape(shape = var_369, x = var_368)[name = tensor("op_370")]; + tensor var_402 = linear(bias = inner_encoder_k_proj_1_bias, weight = inner_encoder_k_proj_1_weight, x = x_7)[name = tensor("linear_13")]; + tensor var_403 = const()[name = tensor("op_403"), val = tensor(0x1p-3)]; + tensor var_404 = mul(x = var_402, y = var_403)[name = tensor("op_404")]; + tensor var_405 = const()[name = tensor("op_405"), val = tensor([1, 2, 4, 64])]; + tensor var_406 = reshape(shape = var_405, x = var_404)[name = tensor("op_406")]; tensor k_3_perm_0 = const()[name = tensor("k_3_perm_0"), val = tensor([0, 2, 1, 3])]; - tensor var_374 = linear(bias = encoder_v_proj_1_bias, weight = encoder_v_proj_1_weight, x = x_7)[name = tensor("linear_14")]; - tensor var_375 = const()[name = tensor("op_375"), val = tensor([1, 2, 4, 64])]; - tensor var_376 = reshape(shape = var_375, x = var_374)[name = tensor("op_376")]; + tensor var_410 = linear(bias = inner_encoder_v_proj_1_bias, weight = inner_encoder_v_proj_1_weight, x = x_7)[name = tensor("linear_14")]; + tensor var_411 = const()[name = tensor("op_411"), val = tensor([1, 2, 4, 64])]; + tensor var_412 = reshape(shape = var_411, x = var_410)[name = tensor("op_412")]; tensor v_3_perm_0 = const()[name = tensor("v_3_perm_0"), val = tensor([0, 2, 1, 3])]; - tensor input_57 = linear(bias = encoder_g_proj_1_bias, weight = encoder_g_proj_1_weight, x = x_7)[name = tensor("linear_15")]; + tensor input_59 = linear(bias = inner_encoder_g_proj_1_bias, weight = inner_encoder_g_proj_1_weight, x = x_7)[name = tensor("linear_15")]; tensor sqrt_s0_3 = sqrt(x = prev_scale_3)[name = tensor("sqrt_s0_3")]; - tensor s_t_3 = add(x = prev_scale_3, y = encoder__t_index)[name = tensor("s_t_3")]; + tensor s_t_3 = add(x = prev_scale_3, y = inner_encoder__t_index)[name = tensor("s_t_3")]; tensor sqrt_s_t_3 = sqrt(x = s_t_3)[name = tensor("sqrt_s_t_3")]; tensor qk_3_transpose_x_1 = const()[name = tensor("qk_3_transpose_x_1"), val = tensor(false)]; tensor qk_3_transpose_y_1 = const()[name = tensor("qk_3_transpose_y_1"), val = tensor(true)]; - tensor k_3 = transpose(perm = k_3_perm_0, x = var_370)[name = tensor("transpose_50")]; - tensor q_3 = transpose(perm = q_3_perm_0, x = var_362)[name = tensor("transpose_51")]; + tensor k_3 = transpose(perm = k_3_perm_0, x = var_406)[name = tensor("transpose_50")]; + tensor q_3 = transpose(perm = q_3_perm_0, x = var_398)[name = tensor("transpose_51")]; tensor qk_3 = matmul(transpose_x = qk_3_transpose_x_1, transpose_y = qk_3_transpose_y_1, x = q_3, y = k_3)[name = tensor("qk_3")]; - tensor var_386 = const()[name = tensor("op_386"), val = tensor([2, 1])]; - tensor var_387 = reshape(shape = var_386, x = sqrt_s_t_3)[name = tensor("op_387")]; - tensor M_3 = real_div(x = encoder__causal_mask, y = var_387)[name = tensor("M_3")]; - tensor var_389 = mul(x = qk_3, y = M_3)[name = tensor("op_389")]; + tensor var_422 = const()[name = tensor("op_422"), val = tensor([2, 1])]; + tensor var_423 = reshape(shape = var_422, x = sqrt_s_t_3)[name = tensor("op_423")]; + tensor M_3 = real_div(x = inner_encoder__causal_mask, y = var_423)[name = tensor("M_3")]; + tensor var_425 = mul(x = qk_3, y = M_3)[name = tensor("op_425")]; tensor inner_3_transpose_x_0 = const()[name = tensor("inner_3_transpose_x_0"), val = tensor(false)]; tensor inner_3_transpose_y_0 = const()[name = tensor("inner_3_transpose_y_0"), val = tensor(false)]; - tensor v_3 = transpose(perm = v_3_perm_0, x = var_376)[name = tensor("transpose_49")]; - tensor inner_3 = matmul(transpose_x = inner_3_transpose_x_0, transpose_y = inner_3_transpose_y_0, x = var_389, y = v_3)[name = tensor("inner_3")]; - tensor var_391_transpose_x_0 = const()[name = tensor("op_391_transpose_x_0"), val = tensor(false)]; - tensor var_391_transpose_y_0 = const()[name = tensor("op_391_transpose_y_0"), val = tensor(false)]; - tensor var_391 = matmul(transpose_x = var_391_transpose_x_0, transpose_y = var_391_transpose_y_0, x = q_3, y = prev_kv_3)[name = tensor("op_391")]; - tensor var_392 = real_div(x = sqrt_s0_3, y = sqrt_s_t_3)[name = tensor("op_392")]; - tensor var_393 = const()[name = tensor("op_393"), val = tensor([1, 1, 2, 1])]; - tensor var_394 = reshape(shape = var_393, x = var_392)[name = tensor("op_394")]; - tensor cross_3 = mul(x = var_391, y = var_394)[name = tensor("cross_3")]; + tensor v_3 = transpose(perm = v_3_perm_0, x = var_412)[name = tensor("transpose_49")]; + tensor inner_3 = matmul(transpose_x = inner_3_transpose_x_0, transpose_y = inner_3_transpose_y_0, x = var_425, y = v_3)[name = tensor("inner_3")]; + tensor var_427_transpose_x_0 = const()[name = tensor("op_427_transpose_x_0"), val = tensor(false)]; + tensor var_427_transpose_y_0 = const()[name = tensor("op_427_transpose_y_0"), val = tensor(false)]; + tensor var_427 = matmul(transpose_x = var_427_transpose_x_0, transpose_y = var_427_transpose_y_0, x = q_3, y = prev_kv_3)[name = tensor("op_427")]; + tensor var_428 = real_div(x = sqrt_s0_3, y = sqrt_s_t_3)[name = tensor("op_428")]; + tensor var_429 = const()[name = tensor("op_429"), val = tensor([1, 1, 2, 1])]; + tensor var_430 = reshape(shape = var_429, x = var_428)[name = tensor("op_430")]; + tensor cross_3 = mul(x = var_427, y = var_430)[name = tensor("cross_3")]; tensor out_7 = add(x = inner_3, y = cross_3)[name = tensor("out_7")]; - tensor var_397 = mul(x = prev_kv_3, y = sqrt_s0_3)[name = tensor("op_397")]; - tensor var_399_transpose_x_1 = const()[name = tensor("op_399_transpose_x_1"), val = tensor(true)]; - tensor var_399_transpose_y_1 = const()[name = tensor("op_399_transpose_y_1"), val = tensor(false)]; - tensor var_399 = matmul(transpose_x = var_399_transpose_x_1, transpose_y = var_399_transpose_y_1, x = k_3, y = v_3)[name = tensor("op_399")]; - tensor new_kv_unnorm_3 = add(x = var_397, y = var_399)[name = tensor("new_kv_unnorm_3")]; - tensor var_401 = const()[name = tensor("op_401"), val = tensor(0x1p+1)]; - tensor new_scale_3 = add(x = prev_scale_3, y = var_401)[name = tensor("new_scale_3")]; - tensor var_403 = sqrt(x = new_scale_3)[name = tensor("op_403")]; - tensor var_404 = real_div(x = new_kv_unnorm_3, y = var_403)[name = tensor("op_404")]; - tensor var_405_perm_0 = const()[name = tensor("op_405_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_433 = mul(x = prev_kv_3, y = sqrt_s0_3)[name = tensor("op_433")]; + tensor var_435_transpose_x_1 = const()[name = tensor("op_435_transpose_x_1"), val = tensor(true)]; + tensor var_435_transpose_y_1 = const()[name = tensor("op_435_transpose_y_1"), val = tensor(false)]; + tensor var_435 = matmul(transpose_x = var_435_transpose_x_1, transpose_y = var_435_transpose_y_1, x = k_3, y = v_3)[name = tensor("op_435")]; + tensor new_kv_unnorm_3 = add(x = var_433, y = var_435)[name = tensor("new_kv_unnorm_3")]; + tensor var_437 = const()[name = tensor("op_437"), val = tensor(0x1p+1)]; + tensor new_scale_3 = add(x = prev_scale_3, y = var_437)[name = tensor("new_scale_3")]; + tensor var_439 = sqrt(x = new_scale_3)[name = tensor("op_439")]; + tensor var_440 = real_div(x = new_kv_unnorm_3, y = var_439)[name = tensor("op_440")]; + tensor var_441_perm_0 = const()[name = tensor("op_441_perm_0"), val = tensor([0, 2, 1, 3])]; tensor out_9_axes_0 = const()[name = tensor("out_9_axes_0"), val = tensor([-1])]; - tensor var_405 = transpose(perm = var_405_perm_0, x = out_7)[name = tensor("transpose_48")]; - tensor out_9 = layer_norm(axes = out_9_axes_0, epsilon = var_17, x = var_405)[name = tensor("out_9")]; - tensor var_409 = const()[name = tensor("op_409"), val = tensor([1, 2, 256])]; - tensor out_11 = reshape(shape = var_409, x = out_9)[name = tensor("out_11")]; - tensor var_411 = silu(x = input_57)[name = tensor("op_411")]; - tensor input_59 = mul(x = var_411, y = out_11)[name = tensor("input_59")]; - tensor ret_out_3 = linear(bias = encoder_out_proj_1_bias, weight = encoder_out_proj_1_weight, x = input_59)[name = tensor("linear_16")]; - tensor x_9 = add(x = input_53, y = ret_out_3)[name = tensor("x_9")]; + tensor var_441 = transpose(perm = var_441_perm_0, x = out_7)[name = tensor("transpose_48")]; + tensor out_9 = layer_norm(axes = out_9_axes_0, epsilon = var_53, x = var_441)[name = tensor("out_9")]; + tensor var_445 = const()[name = tensor("op_445"), val = tensor([1, 2, 256])]; + tensor out_11 = reshape(shape = var_445, x = out_9)[name = tensor("out_11")]; + tensor var_447 = silu(x = input_59)[name = tensor("op_447")]; + tensor input_61 = mul(x = var_447, y = out_11)[name = tensor("input_61")]; + tensor ret_out_3 = linear(bias = inner_encoder_out_proj_1_bias, weight = inner_encoder_out_proj_1_weight, x = input_61)[name = tensor("linear_16")]; + tensor x_9 = add(x = input_55, y = ret_out_3)[name = tensor("x_9")]; tensor window_7_begin_0 = const()[name = tensor("window_7_begin_0"), val = tensor([1, 0, 0, 0])]; tensor window_7_end_0 = const()[name = tensor("window_7_end_0"), val = tensor([2, 1, 16, 256])]; tensor window_7_end_mask_0 = const()[name = tensor("window_7_end_mask_0"), val = tensor([false, true, true, true])]; tensor window_7_squeeze_mask_0 = const()[name = tensor("window_7_squeeze_mask_0"), val = tensor([true, false, false, false])]; tensor window_7 = slice_by_index(begin = window_7_begin_0, end = window_7_end_0, end_mask = window_7_end_mask_0, squeeze_mask = window_7_squeeze_mask_0, x = enc_conv_cache)[name = tensor("window_7")]; - tensor var_419_begin_0 = const()[name = tensor("op_419_begin_0"), val = tensor([0, 0, 0])]; - tensor var_419_end_0 = const()[name = tensor("op_419_end_0"), val = tensor([1, 1, 256])]; - tensor var_419_end_mask_0 = const()[name = tensor("op_419_end_mask_0"), val = tensor([true, false, true])]; - tensor var_419 = slice_by_index(begin = var_419_begin_0, end = var_419_end_0, end_mask = var_419_end_mask_0, x = x_9)[name = tensor("op_419")]; - tensor var_422_begin_0 = const()[name = tensor("op_422_begin_0"), val = tensor([0, 1, 0])]; - tensor var_422_end_0 = const()[name = tensor("op_422_end_0"), val = tensor([1, 16, 256])]; - tensor var_422_end_mask_0 = const()[name = tensor("op_422_end_mask_0"), val = tensor([true, true, true])]; - tensor var_422 = slice_by_index(begin = var_422_begin_0, end = var_422_end_0, end_mask = var_422_end_mask_0, x = window_7)[name = tensor("op_422")]; + tensor var_455_begin_0 = const()[name = tensor("op_455_begin_0"), val = tensor([0, 0, 0])]; + tensor var_455_end_0 = const()[name = tensor("op_455_end_0"), val = tensor([1, 1, 256])]; + tensor var_455_end_mask_0 = const()[name = tensor("op_455_end_mask_0"), val = tensor([true, false, true])]; + tensor var_455 = slice_by_index(begin = var_455_begin_0, end = var_455_end_0, end_mask = var_455_end_mask_0, x = x_9)[name = tensor("op_455")]; + tensor var_458_begin_0 = const()[name = tensor("op_458_begin_0"), val = tensor([0, 1, 0])]; + tensor var_458_end_0 = const()[name = tensor("op_458_end_0"), val = tensor([1, 16, 256])]; + tensor var_458_end_mask_0 = const()[name = tensor("op_458_end_mask_0"), val = tensor([true, true, true])]; + tensor var_458 = slice_by_index(begin = var_458_begin_0, end = var_458_end_0, end_mask = var_458_end_mask_0, x = window_7)[name = tensor("op_458")]; tensor window_9_interleave_0 = const()[name = tensor("window_9_interleave_0"), val = tensor(false)]; - tensor window_9 = concat(axis = var_26, interleave = window_9_interleave_0, values = (var_422, var_419))[name = tensor("window_9")]; - tensor var_427_begin_0 = const()[name = tensor("op_427_begin_0"), val = tensor([0, 1, 0])]; - tensor var_427_end_0 = const()[name = tensor("op_427_end_0"), val = tensor([1, 1, 256])]; - tensor var_427_end_mask_0 = const()[name = tensor("op_427_end_mask_0"), val = tensor([true, true, true])]; - tensor var_427 = slice_by_index(begin = var_427_begin_0, end = var_427_end_0, end_mask = var_427_end_mask_0, x = x_9)[name = tensor("op_427")]; - tensor var_430_begin_0 = const()[name = tensor("op_430_begin_0"), val = tensor([0, 1, 0])]; - tensor var_430_end_0 = const()[name = tensor("op_430_end_0"), val = tensor([1, 16, 256])]; - tensor var_430_end_mask_0 = const()[name = tensor("op_430_end_mask_0"), val = tensor([true, true, true])]; - tensor var_430 = slice_by_index(begin = var_430_begin_0, end = var_430_end_0, end_mask = var_430_end_mask_0, x = window_9)[name = tensor("op_430")]; + tensor window_9 = concat(axis = var_62, interleave = window_9_interleave_0, values = (var_458, var_455))[name = tensor("window_9")]; + tensor var_463_begin_0 = const()[name = tensor("op_463_begin_0"), val = tensor([0, 1, 0])]; + tensor var_463_end_0 = const()[name = tensor("op_463_end_0"), val = tensor([1, 1, 256])]; + tensor var_463_end_mask_0 = const()[name = tensor("op_463_end_mask_0"), val = tensor([true, true, true])]; + tensor var_463 = slice_by_index(begin = var_463_begin_0, end = var_463_end_0, end_mask = var_463_end_mask_0, x = x_9)[name = tensor("op_463")]; + tensor var_466_begin_0 = const()[name = tensor("op_466_begin_0"), val = tensor([0, 1, 0])]; + tensor var_466_end_0 = const()[name = tensor("op_466_end_0"), val = tensor([1, 16, 256])]; + tensor var_466_end_mask_0 = const()[name = tensor("op_466_end_mask_0"), val = tensor([true, true, true])]; + tensor var_466 = slice_by_index(begin = var_466_begin_0, end = var_466_end_0, end_mask = var_466_end_mask_0, x = window_9)[name = tensor("op_466")]; tensor window_11_interleave_0 = const()[name = tensor("window_11_interleave_0"), val = tensor(false)]; - tensor window_11 = concat(axis = var_26, interleave = window_11_interleave_0, values = (var_430, var_427))[name = tensor("window_11")]; - tensor input_61_interleave_0 = const()[name = tensor("input_61_interleave_0"), val = tensor(false)]; - tensor input_61 = concat(axis = var_23, interleave = input_61_interleave_0, values = (window_9, window_11))[name = tensor("input_61")]; + tensor window_11 = concat(axis = var_62, interleave = window_11_interleave_0, values = (var_466, var_463))[name = tensor("window_11")]; + tensor input_63_interleave_0 = const()[name = tensor("input_63_interleave_0"), val = tensor(false)]; + tensor input_63 = concat(axis = var_48, interleave = input_63_interleave_0, values = (window_9, window_11))[name = tensor("input_63")]; tensor x_11_axes_0 = const()[name = tensor("x_11_axes_0"), val = tensor([-1])]; - tensor x_11 = layer_norm(axes = x_11_axes_0, beta = encoder_conv_module_1_sequential_0_bias, epsilon = var_28, gamma = encoder_conv_module_1_sequential_0_weight, x = input_61)[name = tensor("x_11")]; - tensor input_63_perm_0 = const()[name = tensor("input_63_perm_0"), val = tensor([0, 2, 1])]; + tensor x_11 = layer_norm(axes = x_11_axes_0, beta = inner_encoder_conv_module_1_sequential_0_bias, epsilon = var_45, gamma = inner_encoder_conv_module_1_sequential_0_weight, x = input_63)[name = tensor("x_11")]; + tensor input_65_perm_0 = const()[name = tensor("input_65_perm_0"), val = tensor([0, 2, 1])]; tensor inputs_13_pad_type_0 = const()[name = tensor("inputs_13_pad_type_0"), val = tensor("valid")]; tensor inputs_13_strides_0 = const()[name = tensor("inputs_13_strides_0"), val = tensor([1])]; tensor inputs_13_pad_0 = const()[name = tensor("inputs_13_pad_0"), val = tensor([0, 0])]; tensor inputs_13_dilations_0 = const()[name = tensor("inputs_13_dilations_0"), val = tensor([1])]; tensor inputs_13_groups_0 = const()[name = tensor("inputs_13_groups_0"), val = tensor(1)]; - tensor input_63 = transpose(perm = input_63_perm_0, x = x_11)[name = tensor("transpose_47")]; - tensor inputs_13 = conv(bias = encoder_conv_module_1_sequential_2_conv_bias, dilations = inputs_13_dilations_0, groups = inputs_13_groups_0, pad = inputs_13_pad_0, pad_type = inputs_13_pad_type_0, strides = inputs_13_strides_0, weight = encoder_conv_module_1_sequential_2_conv_weight, x = input_63)[name = tensor("inputs_13")]; - tensor var_455_split_sizes_0 = const()[name = tensor("op_455_split_sizes_0"), val = tensor([256, 256])]; - tensor var_455_axis_0 = const()[name = tensor("op_455_axis_0"), val = tensor(1)]; - tensor var_455_0, tensor var_455_1 = split(axis = var_455_axis_0, split_sizes = var_455_split_sizes_0, x = inputs_13)[name = tensor("op_455")]; - tensor var_457 = sigmoid(x = var_455_1)[name = tensor("op_457")]; - tensor inputs_15 = mul(x = var_455_0, y = var_457)[name = tensor("inputs_15")]; + tensor input_65 = transpose(perm = input_65_perm_0, x = x_11)[name = tensor("transpose_47")]; + tensor inputs_13 = conv(bias = inner_encoder_conv_module_1_sequential_2_conv_bias, dilations = inputs_13_dilations_0, groups = inputs_13_groups_0, pad = inputs_13_pad_0, pad_type = inputs_13_pad_type_0, strides = inputs_13_strides_0, weight = inner_encoder_conv_module_1_sequential_2_conv_weight, x = input_65)[name = tensor("inputs_13")]; + tensor var_491_split_sizes_0 = const()[name = tensor("op_491_split_sizes_0"), val = tensor([256, 256])]; + tensor var_491_axis_0 = const()[name = tensor("op_491_axis_0"), val = tensor(1)]; + tensor var_491_0, tensor var_491_1 = split(axis = var_491_axis_0, split_sizes = var_491_split_sizes_0, x = inputs_13)[name = tensor("op_491")]; + tensor var_493 = sigmoid(x = var_491_1)[name = tensor("op_493")]; + tensor inputs_15 = mul(x = var_491_0, y = var_493)[name = tensor("inputs_15")]; tensor outputs_aug_3_pad_type_0 = const()[name = tensor("outputs_aug_3_pad_type_0"), val = tensor("custom")]; tensor outputs_aug_3_pad_0 = const()[name = tensor("outputs_aug_3_pad_0"), val = tensor([15, 15])]; tensor outputs_aug_3_groups_0 = const()[name = tensor("outputs_aug_3_groups_0"), val = tensor(256)]; tensor outputs_aug_3_strides_0 = const()[name = tensor("outputs_aug_3_strides_0"), val = tensor([1])]; tensor outputs_aug_3_dilations_0 = const()[name = tensor("outputs_aug_3_dilations_0"), val = tensor([1])]; - tensor outputs_aug_3 = conv(dilations = outputs_aug_3_dilations_0, groups = outputs_aug_3_groups_0, pad = outputs_aug_3_pad_0, pad_type = outputs_aug_3_pad_type_0, strides = outputs_aug_3_strides_0, weight = encoder_conv_module_1_sequential_4_conv_weight, x = inputs_15)[name = tensor("outputs_aug_3")]; - tensor input_65_begin_0 = const()[name = tensor("input_65_begin_0"), val = tensor([0, 0, 0])]; - tensor input_65_end_0 = const()[name = tensor("input_65_end_0"), val = tensor([2, 256, 16])]; - tensor input_65_end_mask_0 = const()[name = tensor("input_65_end_mask_0"), val = tensor([true, true, false])]; - tensor input_65 = slice_by_index(begin = input_65_begin_0, end = input_65_end_0, end_mask = input_65_end_mask_0, x = outputs_aug_3)[name = tensor("input_65")]; - tensor inputs_17 = batch_norm(beta = encoder_conv_module_1_sequential_5_bias, epsilon = var_28, gamma = encoder_conv_module_1_sequential_5_weight, mean = encoder_conv_module_1_sequential_5_running_mean, variance = encoder_conv_module_1_sequential_5_running_var, x = input_65)[name = tensor("inputs_17")]; - tensor input_67 = silu(x = inputs_17)[name = tensor("input_67")]; - tensor input_69_pad_type_0 = const()[name = tensor("input_69_pad_type_0"), val = tensor("valid")]; - tensor input_69_strides_0 = const()[name = tensor("input_69_strides_0"), val = tensor([1])]; - tensor input_69_pad_0 = const()[name = tensor("input_69_pad_0"), val = tensor([0, 0])]; - tensor input_69_dilations_0 = const()[name = tensor("input_69_dilations_0"), val = tensor([1])]; - tensor input_69_groups_0 = const()[name = tensor("input_69_groups_0"), val = tensor(1)]; - tensor input_69 = conv(bias = encoder_conv_module_1_sequential_7_conv_bias, dilations = input_69_dilations_0, groups = input_69_groups_0, pad = input_69_pad_0, pad_type = input_69_pad_type_0, strides = input_69_strides_0, weight = encoder_conv_module_1_sequential_7_conv_weight, x = input_67)[name = tensor("input_69")]; + tensor outputs_aug_3 = conv(dilations = outputs_aug_3_dilations_0, groups = outputs_aug_3_groups_0, pad = outputs_aug_3_pad_0, pad_type = outputs_aug_3_pad_type_0, strides = outputs_aug_3_strides_0, weight = inner_encoder_conv_module_1_sequential_4_conv_weight, x = inputs_15)[name = tensor("outputs_aug_3")]; + tensor input_67_begin_0 = const()[name = tensor("input_67_begin_0"), val = tensor([0, 0, 0])]; + tensor input_67_end_0 = const()[name = tensor("input_67_end_0"), val = tensor([2, 256, 16])]; + tensor input_67_end_mask_0 = const()[name = tensor("input_67_end_mask_0"), val = tensor([true, true, false])]; + tensor input_67 = slice_by_index(begin = input_67_begin_0, end = input_67_end_0, end_mask = input_67_end_mask_0, x = outputs_aug_3)[name = tensor("input_67")]; + tensor inputs_17 = batch_norm(beta = inner_encoder_conv_module_1_sequential_5_bias, epsilon = var_45, gamma = inner_encoder_conv_module_1_sequential_5_weight, mean = inner_encoder_conv_module_1_sequential_5_running_mean, variance = inner_encoder_conv_module_1_sequential_5_running_var, x = input_67)[name = tensor("inputs_17")]; + tensor input_69 = silu(x = inputs_17)[name = tensor("input_69")]; + tensor input_71_pad_type_0 = const()[name = tensor("input_71_pad_type_0"), val = tensor("valid")]; + tensor input_71_strides_0 = const()[name = tensor("input_71_strides_0"), val = tensor([1])]; + tensor input_71_pad_0 = const()[name = tensor("input_71_pad_0"), val = tensor([0, 0])]; + tensor input_71_dilations_0 = const()[name = tensor("input_71_dilations_0"), val = tensor([1])]; + tensor input_71_groups_0 = const()[name = tensor("input_71_groups_0"), val = tensor(1)]; + tensor input_71 = conv(bias = inner_encoder_conv_module_1_sequential_7_conv_bias, dilations = input_71_dilations_0, groups = input_71_groups_0, pad = input_71_pad_0, pad_type = input_71_pad_type_0, strides = input_71_strides_0, weight = inner_encoder_conv_module_1_sequential_7_conv_weight, x = input_69)[name = tensor("input_71")]; tensor conv_out_3_perm_0 = const()[name = tensor("conv_out_3_perm_0"), val = tensor([0, 2, 1])]; - tensor var_488_begin_0 = const()[name = tensor("op_488_begin_0"), val = tensor([0, -1, 0])]; - tensor var_488_end_0 = const()[name = tensor("op_488_end_0"), val = tensor([2, 16, 256])]; - tensor var_488_end_mask_0 = const()[name = tensor("op_488_end_mask_0"), val = tensor([true, true, true])]; - tensor conv_out_3 = transpose(perm = conv_out_3_perm_0, x = input_69)[name = tensor("transpose_46")]; - tensor var_488 = slice_by_index(begin = var_488_begin_0, end = var_488_end_0, end_mask = var_488_end_mask_0, x = conv_out_3)[name = tensor("op_488")]; - tensor var_490_perm_0 = const()[name = tensor("op_490_perm_0"), val = tensor([1, 0, 2])]; - tensor var_490 = transpose(perm = var_490_perm_0, x = var_488)[name = tensor("transpose_45")]; - tensor input_71 = add(x = x_9, y = var_490)[name = tensor("input_71")]; - tensor input_73_axes_0 = const()[name = tensor("input_73_axes_0"), val = tensor([-1])]; - tensor input_73 = layer_norm(axes = input_73_axes_0, beta = encoder_ffn2_1_module_sequential_0_bias, epsilon = var_28, gamma = encoder_ffn2_1_module_sequential_0_weight, x = input_71)[name = tensor("input_73")]; - tensor inputs_19 = linear(bias = encoder_ffn2_1_module_sequential_1_linear_bias, weight = encoder_ffn2_1_module_sequential_1_linear_weight, x = input_73)[name = tensor("linear_17")]; - tensor input_75 = silu(x = inputs_19)[name = tensor("input_75")]; - tensor input_79 = linear(bias = encoder_ffn2_1_module_sequential_4_linear_bias, weight = encoder_ffn2_1_module_sequential_4_linear_weight, x = input_75)[name = tensor("linear_18")]; - tensor var_513 = const()[name = tensor("op_513"), val = tensor(0x1p-1)]; - tensor var_514 = mul(x = input_79, y = var_513)[name = tensor("op_514")]; - tensor input_81 = add(x = var_514, y = input_71)[name = tensor("input_81")]; - tensor input_83_axes_0 = const()[name = tensor("input_83_axes_0"), val = tensor([-1])]; - tensor input_83 = layer_norm(axes = input_83_axes_0, beta = encoder_layer_norm_1_bias, epsilon = var_28, gamma = encoder_layer_norm_1_weight, x = input_81)[name = tensor("input_83")]; + tensor var_524_begin_0 = const()[name = tensor("op_524_begin_0"), val = tensor([0, -1, 0])]; + tensor var_524_end_0 = const()[name = tensor("op_524_end_0"), val = tensor([2, 16, 256])]; + tensor var_524_end_mask_0 = const()[name = tensor("op_524_end_mask_0"), val = tensor([true, true, true])]; + tensor conv_out_3 = transpose(perm = conv_out_3_perm_0, x = input_71)[name = tensor("transpose_46")]; + tensor var_524 = slice_by_index(begin = var_524_begin_0, end = var_524_end_0, end_mask = var_524_end_mask_0, x = conv_out_3)[name = tensor("op_524")]; + tensor var_526_perm_0 = const()[name = tensor("op_526_perm_0"), val = tensor([1, 0, 2])]; + tensor var_526 = transpose(perm = var_526_perm_0, x = var_524)[name = tensor("transpose_45")]; + tensor input_73 = add(x = x_9, y = var_526)[name = tensor("input_73")]; + tensor input_75_axes_0 = const()[name = tensor("input_75_axes_0"), val = tensor([-1])]; + tensor input_75 = layer_norm(axes = input_75_axes_0, beta = inner_encoder_ffn2_1_module_sequential_0_bias, epsilon = var_45, gamma = inner_encoder_ffn2_1_module_sequential_0_weight, x = input_73)[name = tensor("input_75")]; + tensor inputs_19 = linear(bias = inner_encoder_ffn2_1_module_sequential_1_linear_bias, weight = inner_encoder_ffn2_1_module_sequential_1_linear_weight, x = input_75)[name = tensor("linear_17")]; + tensor input_77 = silu(x = inputs_19)[name = tensor("input_77")]; + tensor input_81 = linear(bias = inner_encoder_ffn2_1_module_sequential_4_linear_bias, weight = inner_encoder_ffn2_1_module_sequential_4_linear_weight, x = input_77)[name = tensor("linear_18")]; + tensor var_549 = const()[name = tensor("op_549"), val = tensor(0x1p-1)]; + tensor var_550 = mul(x = input_81, y = var_549)[name = tensor("op_550")]; + tensor input_83 = add(x = var_550, y = input_73)[name = tensor("input_83")]; tensor input_85_axes_0 = const()[name = tensor("input_85_axes_0"), val = tensor([-1])]; - tensor input_85 = layer_norm(axes = input_85_axes_0, beta = encoder_ffn1_2_module_sequential_0_bias, epsilon = var_28, gamma = encoder_ffn1_2_module_sequential_0_weight, x = input_83)[name = tensor("input_85")]; - tensor inputs_21 = linear(bias = encoder_ffn1_2_module_sequential_1_linear_bias, weight = encoder_ffn1_2_module_sequential_1_linear_weight, x = input_85)[name = tensor("linear_19")]; - tensor input_87 = silu(x = inputs_21)[name = tensor("input_87")]; - tensor input_91 = linear(bias = encoder_ffn1_2_module_sequential_4_linear_bias, weight = encoder_ffn1_2_module_sequential_4_linear_weight, x = input_87)[name = tensor("linear_20")]; - tensor var_543 = const()[name = tensor("op_543"), val = tensor(0x1p-1)]; - tensor var_544 = mul(x = input_91, y = var_543)[name = tensor("op_544")]; - tensor input_93 = add(x = var_544, y = input_83)[name = tensor("input_93")]; + tensor input_85 = layer_norm(axes = input_85_axes_0, beta = inner_encoder_layer_norm_1_bias, epsilon = var_45, gamma = inner_encoder_layer_norm_1_weight, x = input_83)[name = tensor("input_85")]; + tensor input_87_axes_0 = const()[name = tensor("input_87_axes_0"), val = tensor([-1])]; + tensor input_87 = layer_norm(axes = input_87_axes_0, beta = inner_encoder_ffn1_2_module_sequential_0_bias, epsilon = var_45, gamma = inner_encoder_ffn1_2_module_sequential_0_weight, x = input_85)[name = tensor("input_87")]; + tensor inputs_21 = linear(bias = inner_encoder_ffn1_2_module_sequential_1_linear_bias, weight = inner_encoder_ffn1_2_module_sequential_1_linear_weight, x = input_87)[name = tensor("linear_19")]; + tensor input_89 = silu(x = inputs_21)[name = tensor("input_89")]; + tensor input_93 = linear(bias = inner_encoder_ffn1_2_module_sequential_4_linear_bias, weight = inner_encoder_ffn1_2_module_sequential_4_linear_weight, x = input_89)[name = tensor("linear_20")]; + tensor var_579 = const()[name = tensor("op_579"), val = tensor(0x1p-1)]; + tensor var_580 = mul(x = input_93, y = var_579)[name = tensor("op_580")]; + tensor input_95 = add(x = var_580, y = input_85)[name = tensor("input_95")]; tensor x_13_axes_0 = const()[name = tensor("x_13_axes_0"), val = tensor([-1])]; - tensor x_13 = layer_norm(axes = x_13_axes_0, beta = encoder_ret_lns_2_bias, epsilon = var_28, gamma = encoder_ret_lns_2_weight, x = input_93)[name = tensor("x_13")]; + tensor x_13 = layer_norm(axes = x_13_axes_0, beta = inner_encoder_ret_lns_2_bias, epsilon = var_45, gamma = inner_encoder_ret_lns_2_weight, x = input_95)[name = tensor("x_13")]; tensor prev_kv_5_begin_0 = const()[name = tensor("prev_kv_5_begin_0"), val = tensor([2, 0, 0, 0, 0])]; tensor prev_kv_5_end_0 = const()[name = tensor("prev_kv_5_end_0"), val = tensor([3, 1, 4, 64, 64])]; tensor prev_kv_5_end_mask_0 = const()[name = tensor("prev_kv_5_end_mask_0"), val = tensor([false, true, true, true, true])]; @@ -553,153 +567,153 @@ program(1.0) tensor prev_scale_5_end_mask_0 = const()[name = tensor("prev_scale_5_end_mask_0"), val = tensor([false, true])]; tensor prev_scale_5_squeeze_mask_0 = const()[name = tensor("prev_scale_5_squeeze_mask_0"), val = tensor([true, false])]; tensor prev_scale_5 = slice_by_index(begin = prev_scale_5_begin_0, end = prev_scale_5_end_0, end_mask = prev_scale_5_end_mask_0, squeeze_mask = prev_scale_5_squeeze_mask_0, x = enc_scale)[name = tensor("prev_scale_5")]; - tensor var_558 = linear(bias = encoder_q_proj_2_bias, weight = encoder_q_proj_2_weight, x = x_13)[name = tensor("linear_21")]; - tensor var_559 = const()[name = tensor("op_559"), val = tensor([1, 2, 4, 64])]; - tensor var_560 = reshape(shape = var_559, x = var_558)[name = tensor("op_560")]; + tensor var_594 = linear(bias = inner_encoder_q_proj_2_bias, weight = inner_encoder_q_proj_2_weight, x = x_13)[name = tensor("linear_21")]; + tensor var_595 = const()[name = tensor("op_595"), val = tensor([1, 2, 4, 64])]; + tensor var_596 = reshape(shape = var_595, x = var_594)[name = tensor("op_596")]; tensor q_5_perm_0 = const()[name = tensor("q_5_perm_0"), val = tensor([0, 2, 1, 3])]; - tensor var_564 = linear(bias = encoder_k_proj_2_bias, weight = encoder_k_proj_2_weight, x = x_13)[name = tensor("linear_22")]; - tensor var_565 = const()[name = tensor("op_565"), val = tensor(0x1p-3)]; - tensor var_566 = mul(x = var_564, y = var_565)[name = tensor("op_566")]; - tensor var_567 = const()[name = tensor("op_567"), val = tensor([1, 2, 4, 64])]; - tensor var_568 = reshape(shape = var_567, x = var_566)[name = tensor("op_568")]; + tensor var_600 = linear(bias = inner_encoder_k_proj_2_bias, weight = inner_encoder_k_proj_2_weight, x = x_13)[name = tensor("linear_22")]; + tensor var_601 = const()[name = tensor("op_601"), val = tensor(0x1p-3)]; + tensor var_602 = mul(x = var_600, y = var_601)[name = tensor("op_602")]; + tensor var_603 = const()[name = tensor("op_603"), val = tensor([1, 2, 4, 64])]; + tensor var_604 = reshape(shape = var_603, x = var_602)[name = tensor("op_604")]; tensor k_5_perm_0 = const()[name = tensor("k_5_perm_0"), val = tensor([0, 2, 1, 3])]; - tensor var_572 = linear(bias = encoder_v_proj_2_bias, weight = encoder_v_proj_2_weight, x = x_13)[name = tensor("linear_23")]; - tensor var_573 = const()[name = tensor("op_573"), val = tensor([1, 2, 4, 64])]; - tensor var_574 = reshape(shape = var_573, x = var_572)[name = tensor("op_574")]; + tensor var_608 = linear(bias = inner_encoder_v_proj_2_bias, weight = inner_encoder_v_proj_2_weight, x = x_13)[name = tensor("linear_23")]; + tensor var_609 = const()[name = tensor("op_609"), val = tensor([1, 2, 4, 64])]; + tensor var_610 = reshape(shape = var_609, x = var_608)[name = tensor("op_610")]; tensor v_5_perm_0 = const()[name = tensor("v_5_perm_0"), val = tensor([0, 2, 1, 3])]; - tensor input_97 = linear(bias = encoder_g_proj_2_bias, weight = encoder_g_proj_2_weight, x = x_13)[name = tensor("linear_24")]; + tensor input_99 = linear(bias = inner_encoder_g_proj_2_bias, weight = inner_encoder_g_proj_2_weight, x = x_13)[name = tensor("linear_24")]; tensor sqrt_s0_5 = sqrt(x = prev_scale_5)[name = tensor("sqrt_s0_5")]; - tensor s_t_5 = add(x = prev_scale_5, y = encoder__t_index)[name = tensor("s_t_5")]; + tensor s_t_5 = add(x = prev_scale_5, y = inner_encoder__t_index)[name = tensor("s_t_5")]; tensor sqrt_s_t_5 = sqrt(x = s_t_5)[name = tensor("sqrt_s_t_5")]; tensor qk_5_transpose_x_1 = const()[name = tensor("qk_5_transpose_x_1"), val = tensor(false)]; tensor qk_5_transpose_y_1 = const()[name = tensor("qk_5_transpose_y_1"), val = tensor(true)]; - tensor k_5 = transpose(perm = k_5_perm_0, x = var_568)[name = tensor("transpose_43")]; - tensor q_5 = transpose(perm = q_5_perm_0, x = var_560)[name = tensor("transpose_44")]; + tensor k_5 = transpose(perm = k_5_perm_0, x = var_604)[name = tensor("transpose_43")]; + tensor q_5 = transpose(perm = q_5_perm_0, x = var_596)[name = tensor("transpose_44")]; tensor qk_5 = matmul(transpose_x = qk_5_transpose_x_1, transpose_y = qk_5_transpose_y_1, x = q_5, y = k_5)[name = tensor("qk_5")]; - tensor var_584 = const()[name = tensor("op_584"), val = tensor([2, 1])]; - tensor var_585 = reshape(shape = var_584, x = sqrt_s_t_5)[name = tensor("op_585")]; - tensor M_5 = real_div(x = encoder__causal_mask, y = var_585)[name = tensor("M_5")]; - tensor var_587 = mul(x = qk_5, y = M_5)[name = tensor("op_587")]; + tensor var_620 = const()[name = tensor("op_620"), val = tensor([2, 1])]; + tensor var_621 = reshape(shape = var_620, x = sqrt_s_t_5)[name = tensor("op_621")]; + tensor M_5 = real_div(x = inner_encoder__causal_mask, y = var_621)[name = tensor("M_5")]; + tensor var_623 = mul(x = qk_5, y = M_5)[name = tensor("op_623")]; tensor inner_5_transpose_x_0 = const()[name = tensor("inner_5_transpose_x_0"), val = tensor(false)]; tensor inner_5_transpose_y_0 = const()[name = tensor("inner_5_transpose_y_0"), val = tensor(false)]; - tensor v_5 = transpose(perm = v_5_perm_0, x = var_574)[name = tensor("transpose_42")]; - tensor inner_5 = matmul(transpose_x = inner_5_transpose_x_0, transpose_y = inner_5_transpose_y_0, x = var_587, y = v_5)[name = tensor("inner_5")]; - tensor var_589_transpose_x_0 = const()[name = tensor("op_589_transpose_x_0"), val = tensor(false)]; - tensor var_589_transpose_y_0 = const()[name = tensor("op_589_transpose_y_0"), val = tensor(false)]; - tensor var_589 = matmul(transpose_x = var_589_transpose_x_0, transpose_y = var_589_transpose_y_0, x = q_5, y = prev_kv_5)[name = tensor("op_589")]; - tensor var_590 = real_div(x = sqrt_s0_5, y = sqrt_s_t_5)[name = tensor("op_590")]; - tensor var_591 = const()[name = tensor("op_591"), val = tensor([1, 1, 2, 1])]; - tensor var_592 = reshape(shape = var_591, x = var_590)[name = tensor("op_592")]; - tensor cross_5 = mul(x = var_589, y = var_592)[name = tensor("cross_5")]; + tensor v_5 = transpose(perm = v_5_perm_0, x = var_610)[name = tensor("transpose_42")]; + tensor inner_5 = matmul(transpose_x = inner_5_transpose_x_0, transpose_y = inner_5_transpose_y_0, x = var_623, y = v_5)[name = tensor("inner_5")]; + tensor var_625_transpose_x_0 = const()[name = tensor("op_625_transpose_x_0"), val = tensor(false)]; + tensor var_625_transpose_y_0 = const()[name = tensor("op_625_transpose_y_0"), val = tensor(false)]; + tensor var_625 = matmul(transpose_x = var_625_transpose_x_0, transpose_y = var_625_transpose_y_0, x = q_5, y = prev_kv_5)[name = tensor("op_625")]; + tensor var_626 = real_div(x = sqrt_s0_5, y = sqrt_s_t_5)[name = tensor("op_626")]; + tensor var_627 = const()[name = tensor("op_627"), val = tensor([1, 1, 2, 1])]; + tensor var_628 = reshape(shape = var_627, x = var_626)[name = tensor("op_628")]; + tensor cross_5 = mul(x = var_625, y = var_628)[name = tensor("cross_5")]; tensor out_13 = add(x = inner_5, y = cross_5)[name = tensor("out_13")]; - tensor var_595 = mul(x = prev_kv_5, y = sqrt_s0_5)[name = tensor("op_595")]; - tensor var_597_transpose_x_1 = const()[name = tensor("op_597_transpose_x_1"), val = tensor(true)]; - tensor var_597_transpose_y_1 = const()[name = tensor("op_597_transpose_y_1"), val = tensor(false)]; - tensor var_597 = matmul(transpose_x = var_597_transpose_x_1, transpose_y = var_597_transpose_y_1, x = k_5, y = v_5)[name = tensor("op_597")]; - tensor new_kv_unnorm_5 = add(x = var_595, y = var_597)[name = tensor("new_kv_unnorm_5")]; - tensor var_599 = const()[name = tensor("op_599"), val = tensor(0x1p+1)]; - tensor new_scale_5 = add(x = prev_scale_5, y = var_599)[name = tensor("new_scale_5")]; - tensor var_601 = sqrt(x = new_scale_5)[name = tensor("op_601")]; - tensor var_602 = real_div(x = new_kv_unnorm_5, y = var_601)[name = tensor("op_602")]; - tensor var_603_perm_0 = const()[name = tensor("op_603_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_631 = mul(x = prev_kv_5, y = sqrt_s0_5)[name = tensor("op_631")]; + tensor var_633_transpose_x_1 = const()[name = tensor("op_633_transpose_x_1"), val = tensor(true)]; + tensor var_633_transpose_y_1 = const()[name = tensor("op_633_transpose_y_1"), val = tensor(false)]; + tensor var_633 = matmul(transpose_x = var_633_transpose_x_1, transpose_y = var_633_transpose_y_1, x = k_5, y = v_5)[name = tensor("op_633")]; + tensor new_kv_unnorm_5 = add(x = var_631, y = var_633)[name = tensor("new_kv_unnorm_5")]; + tensor var_635 = const()[name = tensor("op_635"), val = tensor(0x1p+1)]; + tensor new_scale_5 = add(x = prev_scale_5, y = var_635)[name = tensor("new_scale_5")]; + tensor var_637 = sqrt(x = new_scale_5)[name = tensor("op_637")]; + tensor var_638 = real_div(x = new_kv_unnorm_5, y = var_637)[name = tensor("op_638")]; + tensor var_639_perm_0 = const()[name = tensor("op_639_perm_0"), val = tensor([0, 2, 1, 3])]; tensor out_15_axes_0 = const()[name = tensor("out_15_axes_0"), val = tensor([-1])]; - tensor var_603 = transpose(perm = var_603_perm_0, x = out_13)[name = tensor("transpose_41")]; - tensor out_15 = layer_norm(axes = out_15_axes_0, epsilon = var_17, x = var_603)[name = tensor("out_15")]; - tensor var_607 = const()[name = tensor("op_607"), val = tensor([1, 2, 256])]; - tensor out_17 = reshape(shape = var_607, x = out_15)[name = tensor("out_17")]; - tensor var_609 = silu(x = input_97)[name = tensor("op_609")]; - tensor input_99 = mul(x = var_609, y = out_17)[name = tensor("input_99")]; - tensor ret_out_5 = linear(bias = encoder_out_proj_2_bias, weight = encoder_out_proj_2_weight, x = input_99)[name = tensor("linear_25")]; - tensor x_15 = add(x = input_93, y = ret_out_5)[name = tensor("x_15")]; + tensor var_639 = transpose(perm = var_639_perm_0, x = out_13)[name = tensor("transpose_41")]; + tensor out_15 = layer_norm(axes = out_15_axes_0, epsilon = var_53, x = var_639)[name = tensor("out_15")]; + tensor var_643 = const()[name = tensor("op_643"), val = tensor([1, 2, 256])]; + tensor out_17 = reshape(shape = var_643, x = out_15)[name = tensor("out_17")]; + tensor var_645 = silu(x = input_99)[name = tensor("op_645")]; + tensor input_101 = mul(x = var_645, y = out_17)[name = tensor("input_101")]; + tensor ret_out_5 = linear(bias = inner_encoder_out_proj_2_bias, weight = inner_encoder_out_proj_2_weight, x = input_101)[name = tensor("linear_25")]; + tensor x_15 = add(x = input_95, y = ret_out_5)[name = tensor("x_15")]; tensor window_13_begin_0 = const()[name = tensor("window_13_begin_0"), val = tensor([2, 0, 0, 0])]; tensor window_13_end_0 = const()[name = tensor("window_13_end_0"), val = tensor([3, 1, 16, 256])]; tensor window_13_end_mask_0 = const()[name = tensor("window_13_end_mask_0"), val = tensor([false, true, true, true])]; tensor window_13_squeeze_mask_0 = const()[name = tensor("window_13_squeeze_mask_0"), val = tensor([true, false, false, false])]; tensor window_13 = slice_by_index(begin = window_13_begin_0, end = window_13_end_0, end_mask = window_13_end_mask_0, squeeze_mask = window_13_squeeze_mask_0, x = enc_conv_cache)[name = tensor("window_13")]; - tensor var_617_begin_0 = const()[name = tensor("op_617_begin_0"), val = tensor([0, 0, 0])]; - tensor var_617_end_0 = const()[name = tensor("op_617_end_0"), val = tensor([1, 1, 256])]; - tensor var_617_end_mask_0 = const()[name = tensor("op_617_end_mask_0"), val = tensor([true, false, true])]; - tensor var_617 = slice_by_index(begin = var_617_begin_0, end = var_617_end_0, end_mask = var_617_end_mask_0, x = x_15)[name = tensor("op_617")]; - tensor var_620_begin_0 = const()[name = tensor("op_620_begin_0"), val = tensor([0, 1, 0])]; - tensor var_620_end_0 = const()[name = tensor("op_620_end_0"), val = tensor([1, 16, 256])]; - tensor var_620_end_mask_0 = const()[name = tensor("op_620_end_mask_0"), val = tensor([true, true, true])]; - tensor var_620 = slice_by_index(begin = var_620_begin_0, end = var_620_end_0, end_mask = var_620_end_mask_0, x = window_13)[name = tensor("op_620")]; + tensor var_653_begin_0 = const()[name = tensor("op_653_begin_0"), val = tensor([0, 0, 0])]; + tensor var_653_end_0 = const()[name = tensor("op_653_end_0"), val = tensor([1, 1, 256])]; + tensor var_653_end_mask_0 = const()[name = tensor("op_653_end_mask_0"), val = tensor([true, false, true])]; + tensor var_653 = slice_by_index(begin = var_653_begin_0, end = var_653_end_0, end_mask = var_653_end_mask_0, x = x_15)[name = tensor("op_653")]; + tensor var_656_begin_0 = const()[name = tensor("op_656_begin_0"), val = tensor([0, 1, 0])]; + tensor var_656_end_0 = const()[name = tensor("op_656_end_0"), val = tensor([1, 16, 256])]; + tensor var_656_end_mask_0 = const()[name = tensor("op_656_end_mask_0"), val = tensor([true, true, true])]; + tensor var_656 = slice_by_index(begin = var_656_begin_0, end = var_656_end_0, end_mask = var_656_end_mask_0, x = window_13)[name = tensor("op_656")]; tensor window_15_interleave_0 = const()[name = tensor("window_15_interleave_0"), val = tensor(false)]; - tensor window_15 = concat(axis = var_26, interleave = window_15_interleave_0, values = (var_620, var_617))[name = tensor("window_15")]; - tensor var_625_begin_0 = const()[name = tensor("op_625_begin_0"), val = tensor([0, 1, 0])]; - tensor var_625_end_0 = const()[name = tensor("op_625_end_0"), val = tensor([1, 1, 256])]; - tensor var_625_end_mask_0 = const()[name = tensor("op_625_end_mask_0"), val = tensor([true, true, true])]; - tensor var_625 = slice_by_index(begin = var_625_begin_0, end = var_625_end_0, end_mask = var_625_end_mask_0, x = x_15)[name = tensor("op_625")]; - tensor var_628_begin_0 = const()[name = tensor("op_628_begin_0"), val = tensor([0, 1, 0])]; - tensor var_628_end_0 = const()[name = tensor("op_628_end_0"), val = tensor([1, 16, 256])]; - tensor var_628_end_mask_0 = const()[name = tensor("op_628_end_mask_0"), val = tensor([true, true, true])]; - tensor var_628 = slice_by_index(begin = var_628_begin_0, end = var_628_end_0, end_mask = var_628_end_mask_0, x = window_15)[name = tensor("op_628")]; + tensor window_15 = concat(axis = var_62, interleave = window_15_interleave_0, values = (var_656, var_653))[name = tensor("window_15")]; + tensor var_661_begin_0 = const()[name = tensor("op_661_begin_0"), val = tensor([0, 1, 0])]; + tensor var_661_end_0 = const()[name = tensor("op_661_end_0"), val = tensor([1, 1, 256])]; + tensor var_661_end_mask_0 = const()[name = tensor("op_661_end_mask_0"), val = tensor([true, true, true])]; + tensor var_661 = slice_by_index(begin = var_661_begin_0, end = var_661_end_0, end_mask = var_661_end_mask_0, x = x_15)[name = tensor("op_661")]; + tensor var_664_begin_0 = const()[name = tensor("op_664_begin_0"), val = tensor([0, 1, 0])]; + tensor var_664_end_0 = const()[name = tensor("op_664_end_0"), val = tensor([1, 16, 256])]; + tensor var_664_end_mask_0 = const()[name = tensor("op_664_end_mask_0"), val = tensor([true, true, true])]; + tensor var_664 = slice_by_index(begin = var_664_begin_0, end = var_664_end_0, end_mask = var_664_end_mask_0, x = window_15)[name = tensor("op_664")]; tensor window_17_interleave_0 = const()[name = tensor("window_17_interleave_0"), val = tensor(false)]; - tensor window_17 = concat(axis = var_26, interleave = window_17_interleave_0, values = (var_628, var_625))[name = tensor("window_17")]; - tensor input_101_interleave_0 = const()[name = tensor("input_101_interleave_0"), val = tensor(false)]; - tensor input_101 = concat(axis = var_23, interleave = input_101_interleave_0, values = (window_15, window_17))[name = tensor("input_101")]; + tensor window_17 = concat(axis = var_62, interleave = window_17_interleave_0, values = (var_664, var_661))[name = tensor("window_17")]; + tensor input_103_interleave_0 = const()[name = tensor("input_103_interleave_0"), val = tensor(false)]; + tensor input_103 = concat(axis = var_48, interleave = input_103_interleave_0, values = (window_15, window_17))[name = tensor("input_103")]; tensor x_17_axes_0 = const()[name = tensor("x_17_axes_0"), val = tensor([-1])]; - tensor x_17 = layer_norm(axes = x_17_axes_0, beta = encoder_conv_module_2_sequential_0_bias, epsilon = var_28, gamma = encoder_conv_module_2_sequential_0_weight, x = input_101)[name = tensor("x_17")]; - tensor input_103_perm_0 = const()[name = tensor("input_103_perm_0"), val = tensor([0, 2, 1])]; + tensor x_17 = layer_norm(axes = x_17_axes_0, beta = inner_encoder_conv_module_2_sequential_0_bias, epsilon = var_45, gamma = inner_encoder_conv_module_2_sequential_0_weight, x = input_103)[name = tensor("x_17")]; + tensor input_105_perm_0 = const()[name = tensor("input_105_perm_0"), val = tensor([0, 2, 1])]; tensor inputs_23_pad_type_0 = const()[name = tensor("inputs_23_pad_type_0"), val = tensor("valid")]; tensor inputs_23_strides_0 = const()[name = tensor("inputs_23_strides_0"), val = tensor([1])]; tensor inputs_23_pad_0 = const()[name = tensor("inputs_23_pad_0"), val = tensor([0, 0])]; tensor inputs_23_dilations_0 = const()[name = tensor("inputs_23_dilations_0"), val = tensor([1])]; tensor inputs_23_groups_0 = const()[name = tensor("inputs_23_groups_0"), val = tensor(1)]; - tensor input_103 = transpose(perm = input_103_perm_0, x = x_17)[name = tensor("transpose_40")]; - tensor inputs_23 = conv(bias = encoder_conv_module_2_sequential_2_conv_bias, dilations = inputs_23_dilations_0, groups = inputs_23_groups_0, pad = inputs_23_pad_0, pad_type = inputs_23_pad_type_0, strides = inputs_23_strides_0, weight = encoder_conv_module_2_sequential_2_conv_weight, x = input_103)[name = tensor("inputs_23")]; - tensor var_653_split_sizes_0 = const()[name = tensor("op_653_split_sizes_0"), val = tensor([256, 256])]; - tensor var_653_axis_0 = const()[name = tensor("op_653_axis_0"), val = tensor(1)]; - tensor var_653_0, tensor var_653_1 = split(axis = var_653_axis_0, split_sizes = var_653_split_sizes_0, x = inputs_23)[name = tensor("op_653")]; - tensor var_655 = sigmoid(x = var_653_1)[name = tensor("op_655")]; - tensor inputs_25 = mul(x = var_653_0, y = var_655)[name = tensor("inputs_25")]; + tensor input_105 = transpose(perm = input_105_perm_0, x = x_17)[name = tensor("transpose_40")]; + tensor inputs_23 = conv(bias = inner_encoder_conv_module_2_sequential_2_conv_bias, dilations = inputs_23_dilations_0, groups = inputs_23_groups_0, pad = inputs_23_pad_0, pad_type = inputs_23_pad_type_0, strides = inputs_23_strides_0, weight = inner_encoder_conv_module_2_sequential_2_conv_weight, x = input_105)[name = tensor("inputs_23")]; + tensor var_689_split_sizes_0 = const()[name = tensor("op_689_split_sizes_0"), val = tensor([256, 256])]; + tensor var_689_axis_0 = const()[name = tensor("op_689_axis_0"), val = tensor(1)]; + tensor var_689_0, tensor var_689_1 = split(axis = var_689_axis_0, split_sizes = var_689_split_sizes_0, x = inputs_23)[name = tensor("op_689")]; + tensor var_691 = sigmoid(x = var_689_1)[name = tensor("op_691")]; + tensor inputs_25 = mul(x = var_689_0, y = var_691)[name = tensor("inputs_25")]; tensor outputs_aug_5_pad_type_0 = const()[name = tensor("outputs_aug_5_pad_type_0"), val = tensor("custom")]; tensor outputs_aug_5_pad_0 = const()[name = tensor("outputs_aug_5_pad_0"), val = tensor([15, 15])]; tensor outputs_aug_5_groups_0 = const()[name = tensor("outputs_aug_5_groups_0"), val = tensor(256)]; tensor outputs_aug_5_strides_0 = const()[name = tensor("outputs_aug_5_strides_0"), val = tensor([1])]; tensor outputs_aug_5_dilations_0 = const()[name = tensor("outputs_aug_5_dilations_0"), val = tensor([1])]; - tensor outputs_aug_5 = conv(dilations = outputs_aug_5_dilations_0, groups = outputs_aug_5_groups_0, pad = outputs_aug_5_pad_0, pad_type = outputs_aug_5_pad_type_0, strides = outputs_aug_5_strides_0, weight = encoder_conv_module_2_sequential_4_conv_weight, x = inputs_25)[name = tensor("outputs_aug_5")]; - tensor input_105_begin_0 = const()[name = tensor("input_105_begin_0"), val = tensor([0, 0, 0])]; - tensor input_105_end_0 = const()[name = tensor("input_105_end_0"), val = tensor([2, 256, 16])]; - tensor input_105_end_mask_0 = const()[name = tensor("input_105_end_mask_0"), val = tensor([true, true, false])]; - tensor input_105 = slice_by_index(begin = input_105_begin_0, end = input_105_end_0, end_mask = input_105_end_mask_0, x = outputs_aug_5)[name = tensor("input_105")]; - tensor inputs_27 = batch_norm(beta = encoder_conv_module_2_sequential_5_bias, epsilon = var_28, gamma = encoder_conv_module_2_sequential_5_weight, mean = encoder_conv_module_2_sequential_5_running_mean, variance = encoder_conv_module_2_sequential_5_running_var, x = input_105)[name = tensor("inputs_27")]; - tensor input_107 = silu(x = inputs_27)[name = tensor("input_107")]; - tensor input_109_pad_type_0 = const()[name = tensor("input_109_pad_type_0"), val = tensor("valid")]; - tensor input_109_strides_0 = const()[name = tensor("input_109_strides_0"), val = tensor([1])]; - tensor input_109_pad_0 = const()[name = tensor("input_109_pad_0"), val = tensor([0, 0])]; - tensor input_109_dilations_0 = const()[name = tensor("input_109_dilations_0"), val = tensor([1])]; - tensor input_109_groups_0 = const()[name = tensor("input_109_groups_0"), val = tensor(1)]; - tensor input_109 = conv(bias = encoder_conv_module_2_sequential_7_conv_bias, dilations = input_109_dilations_0, groups = input_109_groups_0, pad = input_109_pad_0, pad_type = input_109_pad_type_0, strides = input_109_strides_0, weight = encoder_conv_module_2_sequential_7_conv_weight, x = input_107)[name = tensor("input_109")]; + tensor outputs_aug_5 = conv(dilations = outputs_aug_5_dilations_0, groups = outputs_aug_5_groups_0, pad = outputs_aug_5_pad_0, pad_type = outputs_aug_5_pad_type_0, strides = outputs_aug_5_strides_0, weight = inner_encoder_conv_module_2_sequential_4_conv_weight, x = inputs_25)[name = tensor("outputs_aug_5")]; + tensor input_107_begin_0 = const()[name = tensor("input_107_begin_0"), val = tensor([0, 0, 0])]; + tensor input_107_end_0 = const()[name = tensor("input_107_end_0"), val = tensor([2, 256, 16])]; + tensor input_107_end_mask_0 = const()[name = tensor("input_107_end_mask_0"), val = tensor([true, true, false])]; + tensor input_107 = slice_by_index(begin = input_107_begin_0, end = input_107_end_0, end_mask = input_107_end_mask_0, x = outputs_aug_5)[name = tensor("input_107")]; + tensor inputs_27 = batch_norm(beta = inner_encoder_conv_module_2_sequential_5_bias, epsilon = var_45, gamma = inner_encoder_conv_module_2_sequential_5_weight, mean = inner_encoder_conv_module_2_sequential_5_running_mean, variance = inner_encoder_conv_module_2_sequential_5_running_var, x = input_107)[name = tensor("inputs_27")]; + tensor input_109 = silu(x = inputs_27)[name = tensor("input_109")]; + tensor input_111_pad_type_0 = const()[name = tensor("input_111_pad_type_0"), val = tensor("valid")]; + tensor input_111_strides_0 = const()[name = tensor("input_111_strides_0"), val = tensor([1])]; + tensor input_111_pad_0 = const()[name = tensor("input_111_pad_0"), val = tensor([0, 0])]; + tensor input_111_dilations_0 = const()[name = tensor("input_111_dilations_0"), val = tensor([1])]; + tensor input_111_groups_0 = const()[name = tensor("input_111_groups_0"), val = tensor(1)]; + tensor input_111 = conv(bias = inner_encoder_conv_module_2_sequential_7_conv_bias, dilations = input_111_dilations_0, groups = input_111_groups_0, pad = input_111_pad_0, pad_type = input_111_pad_type_0, strides = input_111_strides_0, weight = inner_encoder_conv_module_2_sequential_7_conv_weight, x = input_109)[name = tensor("input_111")]; tensor conv_out_5_perm_0 = const()[name = tensor("conv_out_5_perm_0"), val = tensor([0, 2, 1])]; - tensor var_686_begin_0 = const()[name = tensor("op_686_begin_0"), val = tensor([0, -1, 0])]; - tensor var_686_end_0 = const()[name = tensor("op_686_end_0"), val = tensor([2, 16, 256])]; - tensor var_686_end_mask_0 = const()[name = tensor("op_686_end_mask_0"), val = tensor([true, true, true])]; - tensor conv_out_5 = transpose(perm = conv_out_5_perm_0, x = input_109)[name = tensor("transpose_39")]; - tensor var_686 = slice_by_index(begin = var_686_begin_0, end = var_686_end_0, end_mask = var_686_end_mask_0, x = conv_out_5)[name = tensor("op_686")]; - tensor var_688_perm_0 = const()[name = tensor("op_688_perm_0"), val = tensor([1, 0, 2])]; - tensor var_688 = transpose(perm = var_688_perm_0, x = var_686)[name = tensor("transpose_38")]; - tensor input_111 = add(x = x_15, y = var_688)[name = tensor("input_111")]; - tensor input_113_axes_0 = const()[name = tensor("input_113_axes_0"), val = tensor([-1])]; - tensor input_113 = layer_norm(axes = input_113_axes_0, beta = encoder_ffn2_2_module_sequential_0_bias, epsilon = var_28, gamma = encoder_ffn2_2_module_sequential_0_weight, x = input_111)[name = tensor("input_113")]; - tensor inputs_29 = linear(bias = encoder_ffn2_2_module_sequential_1_linear_bias, weight = encoder_ffn2_2_module_sequential_1_linear_weight, x = input_113)[name = tensor("linear_26")]; - tensor input_115 = silu(x = inputs_29)[name = tensor("input_115")]; - tensor input_119 = linear(bias = encoder_ffn2_2_module_sequential_4_linear_bias, weight = encoder_ffn2_2_module_sequential_4_linear_weight, x = input_115)[name = tensor("linear_27")]; - tensor var_711 = const()[name = tensor("op_711"), val = tensor(0x1p-1)]; - tensor var_712 = mul(x = input_119, y = var_711)[name = tensor("op_712")]; - tensor input_121 = add(x = var_712, y = input_111)[name = tensor("input_121")]; - tensor input_123_axes_0 = const()[name = tensor("input_123_axes_0"), val = tensor([-1])]; - tensor input_123 = layer_norm(axes = input_123_axes_0, beta = encoder_layer_norm_2_bias, epsilon = var_28, gamma = encoder_layer_norm_2_weight, x = input_121)[name = tensor("input_123")]; + tensor var_722_begin_0 = const()[name = tensor("op_722_begin_0"), val = tensor([0, -1, 0])]; + tensor var_722_end_0 = const()[name = tensor("op_722_end_0"), val = tensor([2, 16, 256])]; + tensor var_722_end_mask_0 = const()[name = tensor("op_722_end_mask_0"), val = tensor([true, true, true])]; + tensor conv_out_5 = transpose(perm = conv_out_5_perm_0, x = input_111)[name = tensor("transpose_39")]; + tensor var_722 = slice_by_index(begin = var_722_begin_0, end = var_722_end_0, end_mask = var_722_end_mask_0, x = conv_out_5)[name = tensor("op_722")]; + tensor var_724_perm_0 = const()[name = tensor("op_724_perm_0"), val = tensor([1, 0, 2])]; + tensor var_724 = transpose(perm = var_724_perm_0, x = var_722)[name = tensor("transpose_38")]; + tensor input_113 = add(x = x_15, y = var_724)[name = tensor("input_113")]; + tensor input_115_axes_0 = const()[name = tensor("input_115_axes_0"), val = tensor([-1])]; + tensor input_115 = layer_norm(axes = input_115_axes_0, beta = inner_encoder_ffn2_2_module_sequential_0_bias, epsilon = var_45, gamma = inner_encoder_ffn2_2_module_sequential_0_weight, x = input_113)[name = tensor("input_115")]; + tensor inputs_29 = linear(bias = inner_encoder_ffn2_2_module_sequential_1_linear_bias, weight = inner_encoder_ffn2_2_module_sequential_1_linear_weight, x = input_115)[name = tensor("linear_26")]; + tensor input_117 = silu(x = inputs_29)[name = tensor("input_117")]; + tensor input_121 = linear(bias = inner_encoder_ffn2_2_module_sequential_4_linear_bias, weight = inner_encoder_ffn2_2_module_sequential_4_linear_weight, x = input_117)[name = tensor("linear_27")]; + tensor var_747 = const()[name = tensor("op_747"), val = tensor(0x1p-1)]; + tensor var_748 = mul(x = input_121, y = var_747)[name = tensor("op_748")]; + tensor input_123 = add(x = var_748, y = input_113)[name = tensor("input_123")]; tensor input_125_axes_0 = const()[name = tensor("input_125_axes_0"), val = tensor([-1])]; - tensor input_125 = layer_norm(axes = input_125_axes_0, beta = encoder_ffn1_3_module_sequential_0_bias, epsilon = var_28, gamma = encoder_ffn1_3_module_sequential_0_weight, x = input_123)[name = tensor("input_125")]; - tensor inputs_31 = linear(bias = encoder_ffn1_3_module_sequential_1_linear_bias, weight = encoder_ffn1_3_module_sequential_1_linear_weight, x = input_125)[name = tensor("linear_28")]; - tensor input_127 = silu(x = inputs_31)[name = tensor("input_127")]; - tensor input_131 = linear(bias = encoder_ffn1_3_module_sequential_4_linear_bias, weight = encoder_ffn1_3_module_sequential_4_linear_weight, x = input_127)[name = tensor("linear_29")]; - tensor var_741 = const()[name = tensor("op_741"), val = tensor(0x1p-1)]; - tensor var_742 = mul(x = input_131, y = var_741)[name = tensor("op_742")]; - tensor input_133 = add(x = var_742, y = input_123)[name = tensor("input_133")]; + tensor input_125 = layer_norm(axes = input_125_axes_0, beta = inner_encoder_layer_norm_2_bias, epsilon = var_45, gamma = inner_encoder_layer_norm_2_weight, x = input_123)[name = tensor("input_125")]; + tensor input_127_axes_0 = const()[name = tensor("input_127_axes_0"), val = tensor([-1])]; + tensor input_127 = layer_norm(axes = input_127_axes_0, beta = inner_encoder_ffn1_3_module_sequential_0_bias, epsilon = var_45, gamma = inner_encoder_ffn1_3_module_sequential_0_weight, x = input_125)[name = tensor("input_127")]; + tensor inputs_31 = linear(bias = inner_encoder_ffn1_3_module_sequential_1_linear_bias, weight = inner_encoder_ffn1_3_module_sequential_1_linear_weight, x = input_127)[name = tensor("linear_28")]; + tensor input_129 = silu(x = inputs_31)[name = tensor("input_129")]; + tensor input_133 = linear(bias = inner_encoder_ffn1_3_module_sequential_4_linear_bias, weight = inner_encoder_ffn1_3_module_sequential_4_linear_weight, x = input_129)[name = tensor("linear_29")]; + tensor var_777 = const()[name = tensor("op_777"), val = tensor(0x1p-1)]; + tensor var_778 = mul(x = input_133, y = var_777)[name = tensor("op_778")]; + tensor input_135 = add(x = var_778, y = input_125)[name = tensor("input_135")]; tensor x_19_axes_0 = const()[name = tensor("x_19_axes_0"), val = tensor([-1])]; - tensor x_19 = layer_norm(axes = x_19_axes_0, beta = encoder_ret_lns_3_bias, epsilon = var_28, gamma = encoder_ret_lns_3_weight, x = input_133)[name = tensor("x_19")]; + tensor x_19 = layer_norm(axes = x_19_axes_0, beta = inner_encoder_ret_lns_3_bias, epsilon = var_45, gamma = inner_encoder_ret_lns_3_weight, x = input_135)[name = tensor("x_19")]; tensor prev_kv_7_begin_0 = const()[name = tensor("prev_kv_7_begin_0"), val = tensor([3, 0, 0, 0, 0])]; tensor prev_kv_7_end_0 = const()[name = tensor("prev_kv_7_end_0"), val = tensor([4, 1, 4, 64, 64])]; tensor prev_kv_7_end_mask_0 = const()[name = tensor("prev_kv_7_end_mask_0"), val = tensor([false, true, true, true, true])]; @@ -710,189 +724,182 @@ program(1.0) tensor prev_scale_7_end_mask_0 = const()[name = tensor("prev_scale_7_end_mask_0"), val = tensor([false, true])]; tensor prev_scale_7_squeeze_mask_0 = const()[name = tensor("prev_scale_7_squeeze_mask_0"), val = tensor([true, false])]; tensor prev_scale_7 = slice_by_index(begin = prev_scale_7_begin_0, end = prev_scale_7_end_0, end_mask = prev_scale_7_end_mask_0, squeeze_mask = prev_scale_7_squeeze_mask_0, x = enc_scale)[name = tensor("prev_scale_7")]; - tensor var_756 = linear(bias = encoder_q_proj_3_bias, weight = encoder_q_proj_3_weight, x = x_19)[name = tensor("linear_30")]; - tensor var_757 = const()[name = tensor("op_757"), val = tensor([1, 2, 4, 64])]; - tensor var_758 = reshape(shape = var_757, x = var_756)[name = tensor("op_758")]; + tensor var_792 = linear(bias = inner_encoder_q_proj_3_bias, weight = inner_encoder_q_proj_3_weight, x = x_19)[name = tensor("linear_30")]; + tensor var_793 = const()[name = tensor("op_793"), val = tensor([1, 2, 4, 64])]; + tensor var_794 = reshape(shape = var_793, x = var_792)[name = tensor("op_794")]; tensor q_7_perm_0 = const()[name = tensor("q_7_perm_0"), val = tensor([0, 2, 1, 3])]; - tensor var_762 = linear(bias = encoder_k_proj_3_bias, weight = encoder_k_proj_3_weight, x = x_19)[name = tensor("linear_31")]; - tensor var_763 = const()[name = tensor("op_763"), val = tensor(0x1p-3)]; - tensor var_764 = mul(x = var_762, y = var_763)[name = tensor("op_764")]; - tensor var_765 = const()[name = tensor("op_765"), val = tensor([1, 2, 4, 64])]; - tensor var_766 = reshape(shape = var_765, x = var_764)[name = tensor("op_766")]; + tensor var_798 = linear(bias = inner_encoder_k_proj_3_bias, weight = inner_encoder_k_proj_3_weight, x = x_19)[name = tensor("linear_31")]; + tensor var_799 = const()[name = tensor("op_799"), val = tensor(0x1p-3)]; + tensor var_800 = mul(x = var_798, y = var_799)[name = tensor("op_800")]; + tensor var_801 = const()[name = tensor("op_801"), val = tensor([1, 2, 4, 64])]; + tensor var_802 = reshape(shape = var_801, x = var_800)[name = tensor("op_802")]; tensor k_7_perm_0 = const()[name = tensor("k_7_perm_0"), val = tensor([0, 2, 1, 3])]; - tensor var_770 = linear(bias = encoder_v_proj_3_bias, weight = encoder_v_proj_3_weight, x = x_19)[name = tensor("linear_32")]; - tensor var_771 = const()[name = tensor("op_771"), val = tensor([1, 2, 4, 64])]; - tensor var_772 = reshape(shape = var_771, x = var_770)[name = tensor("op_772")]; + tensor var_806 = linear(bias = inner_encoder_v_proj_3_bias, weight = inner_encoder_v_proj_3_weight, x = x_19)[name = tensor("linear_32")]; + tensor var_807 = const()[name = tensor("op_807"), val = tensor([1, 2, 4, 64])]; + tensor var_808 = reshape(shape = var_807, x = var_806)[name = tensor("op_808")]; tensor v_7_perm_0 = const()[name = tensor("v_7_perm_0"), val = tensor([0, 2, 1, 3])]; - tensor input_137 = linear(bias = encoder_g_proj_3_bias, weight = encoder_g_proj_3_weight, x = x_19)[name = tensor("linear_33")]; + tensor input_139 = linear(bias = inner_encoder_g_proj_3_bias, weight = inner_encoder_g_proj_3_weight, x = x_19)[name = tensor("linear_33")]; tensor sqrt_s0_7 = sqrt(x = prev_scale_7)[name = tensor("sqrt_s0_7")]; - tensor s_t_7 = add(x = prev_scale_7, y = encoder__t_index)[name = tensor("s_t_7")]; + tensor s_t_7 = add(x = prev_scale_7, y = inner_encoder__t_index)[name = tensor("s_t_7")]; tensor sqrt_s_t_7 = sqrt(x = s_t_7)[name = tensor("sqrt_s_t_7")]; tensor qk_7_transpose_x_1 = const()[name = tensor("qk_7_transpose_x_1"), val = tensor(false)]; tensor qk_7_transpose_y_1 = const()[name = tensor("qk_7_transpose_y_1"), val = tensor(true)]; - tensor k_7 = transpose(perm = k_7_perm_0, x = var_766)[name = tensor("transpose_36")]; - tensor q_7 = transpose(perm = q_7_perm_0, x = var_758)[name = tensor("transpose_37")]; + tensor k_7 = transpose(perm = k_7_perm_0, x = var_802)[name = tensor("transpose_36")]; + tensor q_7 = transpose(perm = q_7_perm_0, x = var_794)[name = tensor("transpose_37")]; tensor qk_7 = matmul(transpose_x = qk_7_transpose_x_1, transpose_y = qk_7_transpose_y_1, x = q_7, y = k_7)[name = tensor("qk_7")]; - tensor var_782 = const()[name = tensor("op_782"), val = tensor([2, 1])]; - tensor var_783 = reshape(shape = var_782, x = sqrt_s_t_7)[name = tensor("op_783")]; - tensor M_7 = real_div(x = encoder__causal_mask, y = var_783)[name = tensor("M_7")]; - tensor var_785 = mul(x = qk_7, y = M_7)[name = tensor("op_785")]; + tensor var_818 = const()[name = tensor("op_818"), val = tensor([2, 1])]; + tensor var_819 = reshape(shape = var_818, x = sqrt_s_t_7)[name = tensor("op_819")]; + tensor M_7 = real_div(x = inner_encoder__causal_mask, y = var_819)[name = tensor("M_7")]; + tensor var_821 = mul(x = qk_7, y = M_7)[name = tensor("op_821")]; tensor inner_7_transpose_x_0 = const()[name = tensor("inner_7_transpose_x_0"), val = tensor(false)]; tensor inner_7_transpose_y_0 = const()[name = tensor("inner_7_transpose_y_0"), val = tensor(false)]; - tensor v_7 = transpose(perm = v_7_perm_0, x = var_772)[name = tensor("transpose_35")]; - tensor inner_7 = matmul(transpose_x = inner_7_transpose_x_0, transpose_y = inner_7_transpose_y_0, x = var_785, y = v_7)[name = tensor("inner_7")]; - tensor var_787_transpose_x_0 = const()[name = tensor("op_787_transpose_x_0"), val = tensor(false)]; - tensor var_787_transpose_y_0 = const()[name = tensor("op_787_transpose_y_0"), val = tensor(false)]; - tensor var_787 = matmul(transpose_x = var_787_transpose_x_0, transpose_y = var_787_transpose_y_0, x = q_7, y = prev_kv_7)[name = tensor("op_787")]; - tensor var_788 = real_div(x = sqrt_s0_7, y = sqrt_s_t_7)[name = tensor("op_788")]; - tensor var_789 = const()[name = tensor("op_789"), val = tensor([1, 1, 2, 1])]; - tensor var_790 = reshape(shape = var_789, x = var_788)[name = tensor("op_790")]; - tensor cross_7 = mul(x = var_787, y = var_790)[name = tensor("cross_7")]; + tensor v_7 = transpose(perm = v_7_perm_0, x = var_808)[name = tensor("transpose_35")]; + tensor inner_7 = matmul(transpose_x = inner_7_transpose_x_0, transpose_y = inner_7_transpose_y_0, x = var_821, y = v_7)[name = tensor("inner_7")]; + tensor var_823_transpose_x_0 = const()[name = tensor("op_823_transpose_x_0"), val = tensor(false)]; + tensor var_823_transpose_y_0 = const()[name = tensor("op_823_transpose_y_0"), val = tensor(false)]; + tensor var_823 = matmul(transpose_x = var_823_transpose_x_0, transpose_y = var_823_transpose_y_0, x = q_7, y = prev_kv_7)[name = tensor("op_823")]; + tensor var_824 = real_div(x = sqrt_s0_7, y = sqrt_s_t_7)[name = tensor("op_824")]; + tensor var_825 = const()[name = tensor("op_825"), val = tensor([1, 1, 2, 1])]; + tensor var_826 = reshape(shape = var_825, x = var_824)[name = tensor("op_826")]; + tensor cross_7 = mul(x = var_823, y = var_826)[name = tensor("cross_7")]; tensor out_19 = add(x = inner_7, y = cross_7)[name = tensor("out_19")]; - tensor var_793 = mul(x = prev_kv_7, y = sqrt_s0_7)[name = tensor("op_793")]; - tensor var_795_transpose_x_1 = const()[name = tensor("op_795_transpose_x_1"), val = tensor(true)]; - tensor var_795_transpose_y_1 = const()[name = tensor("op_795_transpose_y_1"), val = tensor(false)]; - tensor var_795 = matmul(transpose_x = var_795_transpose_x_1, transpose_y = var_795_transpose_y_1, x = k_7, y = v_7)[name = tensor("op_795")]; - tensor new_kv_unnorm_7 = add(x = var_793, y = var_795)[name = tensor("new_kv_unnorm_7")]; - tensor var_797 = const()[name = tensor("op_797"), val = tensor(0x1p+1)]; - tensor new_scale_7 = add(x = prev_scale_7, y = var_797)[name = tensor("new_scale_7")]; - tensor var_799 = sqrt(x = new_scale_7)[name = tensor("op_799")]; - tensor nkv_1 = real_div(x = new_kv_unnorm_7, y = var_799)[name = tensor("nkv_1")]; - tensor var_801_perm_0 = const()[name = tensor("op_801_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_829 = mul(x = prev_kv_7, y = sqrt_s0_7)[name = tensor("op_829")]; + tensor var_831_transpose_x_1 = const()[name = tensor("op_831_transpose_x_1"), val = tensor(true)]; + tensor var_831_transpose_y_1 = const()[name = tensor("op_831_transpose_y_1"), val = tensor(false)]; + tensor var_831 = matmul(transpose_x = var_831_transpose_x_1, transpose_y = var_831_transpose_y_1, x = k_7, y = v_7)[name = tensor("op_831")]; + tensor new_kv_unnorm_7 = add(x = var_829, y = var_831)[name = tensor("new_kv_unnorm_7")]; + tensor var_833 = const()[name = tensor("op_833"), val = tensor(0x1p+1)]; + tensor new_scale_7 = add(x = prev_scale_7, y = var_833)[name = tensor("new_scale_7")]; + tensor var_835 = sqrt(x = new_scale_7)[name = tensor("op_835")]; + tensor nkv_1 = real_div(x = new_kv_unnorm_7, y = var_835)[name = tensor("nkv_1")]; + tensor var_837_perm_0 = const()[name = tensor("op_837_perm_0"), val = tensor([0, 2, 1, 3])]; tensor out_21_axes_0 = const()[name = tensor("out_21_axes_0"), val = tensor([-1])]; - tensor var_801 = transpose(perm = var_801_perm_0, x = out_19)[name = tensor("transpose_34")]; - tensor out_21 = layer_norm(axes = out_21_axes_0, epsilon = var_17, x = var_801)[name = tensor("out_21")]; - tensor var_805 = const()[name = tensor("op_805"), val = tensor([1, 2, 256])]; - tensor out_23 = reshape(shape = var_805, x = out_21)[name = tensor("out_23")]; - tensor var_807 = silu(x = input_137)[name = tensor("op_807")]; - tensor input_139 = mul(x = var_807, y = out_23)[name = tensor("input_139")]; - tensor ret_out_7 = linear(bias = encoder_out_proj_3_bias, weight = encoder_out_proj_3_weight, x = input_139)[name = tensor("linear_34")]; - tensor x_21 = add(x = input_133, y = ret_out_7)[name = tensor("x_21")]; + tensor var_837 = transpose(perm = var_837_perm_0, x = out_19)[name = tensor("transpose_34")]; + tensor out_21 = layer_norm(axes = out_21_axes_0, epsilon = var_53, x = var_837)[name = tensor("out_21")]; + tensor var_841 = const()[name = tensor("op_841"), val = tensor([1, 2, 256])]; + tensor out_23 = reshape(shape = var_841, x = out_21)[name = tensor("out_23")]; + tensor var_843 = silu(x = input_139)[name = tensor("op_843")]; + tensor input_141 = mul(x = var_843, y = out_23)[name = tensor("input_141")]; + tensor ret_out_7 = linear(bias = inner_encoder_out_proj_3_bias, weight = inner_encoder_out_proj_3_weight, x = input_141)[name = tensor("linear_34")]; + tensor x_21 = add(x = input_135, y = ret_out_7)[name = tensor("x_21")]; tensor window_19_begin_0 = const()[name = tensor("window_19_begin_0"), val = tensor([3, 0, 0, 0])]; tensor window_19_end_0 = const()[name = tensor("window_19_end_0"), val = tensor([4, 1, 16, 256])]; tensor window_19_end_mask_0 = const()[name = tensor("window_19_end_mask_0"), val = tensor([false, true, true, true])]; tensor window_19_squeeze_mask_0 = const()[name = tensor("window_19_squeeze_mask_0"), val = tensor([true, false, false, false])]; tensor window_19 = slice_by_index(begin = window_19_begin_0, end = window_19_end_0, end_mask = window_19_end_mask_0, squeeze_mask = window_19_squeeze_mask_0, x = enc_conv_cache)[name = tensor("window_19")]; - tensor var_815_begin_0 = const()[name = tensor("op_815_begin_0"), val = tensor([0, 0, 0])]; - tensor var_815_end_0 = const()[name = tensor("op_815_end_0"), val = tensor([1, 1, 256])]; - tensor var_815_end_mask_0 = const()[name = tensor("op_815_end_mask_0"), val = tensor([true, false, true])]; - tensor var_815 = slice_by_index(begin = var_815_begin_0, end = var_815_end_0, end_mask = var_815_end_mask_0, x = x_21)[name = tensor("op_815")]; - tensor var_818_begin_0 = const()[name = tensor("op_818_begin_0"), val = tensor([0, 1, 0])]; - tensor var_818_end_0 = const()[name = tensor("op_818_end_0"), val = tensor([1, 16, 256])]; - tensor var_818_end_mask_0 = const()[name = tensor("op_818_end_mask_0"), val = tensor([true, true, true])]; - tensor var_818 = slice_by_index(begin = var_818_begin_0, end = var_818_end_0, end_mask = var_818_end_mask_0, x = window_19)[name = tensor("op_818")]; + tensor var_851_begin_0 = const()[name = tensor("op_851_begin_0"), val = tensor([0, 0, 0])]; + tensor var_851_end_0 = const()[name = tensor("op_851_end_0"), val = tensor([1, 1, 256])]; + tensor var_851_end_mask_0 = const()[name = tensor("op_851_end_mask_0"), val = tensor([true, false, true])]; + tensor var_851 = slice_by_index(begin = var_851_begin_0, end = var_851_end_0, end_mask = var_851_end_mask_0, x = x_21)[name = tensor("op_851")]; + tensor var_854_begin_0 = const()[name = tensor("op_854_begin_0"), val = tensor([0, 1, 0])]; + tensor var_854_end_0 = const()[name = tensor("op_854_end_0"), val = tensor([1, 16, 256])]; + tensor var_854_end_mask_0 = const()[name = tensor("op_854_end_mask_0"), val = tensor([true, true, true])]; + tensor var_854 = slice_by_index(begin = var_854_begin_0, end = var_854_end_0, end_mask = var_854_end_mask_0, x = window_19)[name = tensor("op_854")]; tensor window_21_interleave_0 = const()[name = tensor("window_21_interleave_0"), val = tensor(false)]; - tensor window_21 = concat(axis = var_26, interleave = window_21_interleave_0, values = (var_818, var_815))[name = tensor("window_21")]; - tensor var_823_begin_0 = const()[name = tensor("op_823_begin_0"), val = tensor([0, 1, 0])]; - tensor var_823_end_0 = const()[name = tensor("op_823_end_0"), val = tensor([1, 1, 256])]; - tensor var_823_end_mask_0 = const()[name = tensor("op_823_end_mask_0"), val = tensor([true, true, true])]; - tensor var_823 = slice_by_index(begin = var_823_begin_0, end = var_823_end_0, end_mask = var_823_end_mask_0, x = x_21)[name = tensor("op_823")]; - tensor var_826_begin_0 = const()[name = tensor("op_826_begin_0"), val = tensor([0, 1, 0])]; - tensor var_826_end_0 = const()[name = tensor("op_826_end_0"), val = tensor([1, 16, 256])]; - tensor var_826_end_mask_0 = const()[name = tensor("op_826_end_mask_0"), val = tensor([true, true, true])]; - tensor var_826 = slice_by_index(begin = var_826_begin_0, end = var_826_end_0, end_mask = var_826_end_mask_0, x = window_21)[name = tensor("op_826")]; + tensor window_21 = concat(axis = var_62, interleave = window_21_interleave_0, values = (var_854, var_851))[name = tensor("window_21")]; + tensor var_859_begin_0 = const()[name = tensor("op_859_begin_0"), val = tensor([0, 1, 0])]; + tensor var_859_end_0 = const()[name = tensor("op_859_end_0"), val = tensor([1, 1, 256])]; + tensor var_859_end_mask_0 = const()[name = tensor("op_859_end_mask_0"), val = tensor([true, true, true])]; + tensor var_859 = slice_by_index(begin = var_859_begin_0, end = var_859_end_0, end_mask = var_859_end_mask_0, x = x_21)[name = tensor("op_859")]; + tensor var_862_begin_0 = const()[name = tensor("op_862_begin_0"), val = tensor([0, 1, 0])]; + tensor var_862_end_0 = const()[name = tensor("op_862_end_0"), val = tensor([1, 16, 256])]; + tensor var_862_end_mask_0 = const()[name = tensor("op_862_end_mask_0"), val = tensor([true, true, true])]; + tensor var_862 = slice_by_index(begin = var_862_begin_0, end = var_862_end_0, end_mask = var_862_end_mask_0, x = window_21)[name = tensor("op_862")]; tensor window_interleave_0 = const()[name = tensor("window_interleave_0"), val = tensor(false)]; - tensor window = concat(axis = var_26, interleave = window_interleave_0, values = (var_826, var_823))[name = tensor("window")]; - tensor input_141_interleave_0 = const()[name = tensor("input_141_interleave_0"), val = tensor(false)]; - tensor input_141 = concat(axis = var_23, interleave = input_141_interleave_0, values = (window_21, window))[name = tensor("input_141")]; + tensor window = concat(axis = var_62, interleave = window_interleave_0, values = (var_862, var_859))[name = tensor("window")]; + tensor input_143_interleave_0 = const()[name = tensor("input_143_interleave_0"), val = tensor(false)]; + tensor input_143 = concat(axis = var_48, interleave = input_143_interleave_0, values = (window_21, window))[name = tensor("input_143")]; tensor x_23_axes_0 = const()[name = tensor("x_23_axes_0"), val = tensor([-1])]; - tensor x_23 = layer_norm(axes = x_23_axes_0, beta = encoder_conv_module_3_sequential_0_bias, epsilon = var_28, gamma = encoder_conv_module_3_sequential_0_weight, x = input_141)[name = tensor("x_23")]; - tensor input_143_perm_0 = const()[name = tensor("input_143_perm_0"), val = tensor([0, 2, 1])]; + tensor x_23 = layer_norm(axes = x_23_axes_0, beta = inner_encoder_conv_module_3_sequential_0_bias, epsilon = var_45, gamma = inner_encoder_conv_module_3_sequential_0_weight, x = input_143)[name = tensor("x_23")]; + tensor input_145_perm_0 = const()[name = tensor("input_145_perm_0"), val = tensor([0, 2, 1])]; tensor inputs_33_pad_type_0 = const()[name = tensor("inputs_33_pad_type_0"), val = tensor("valid")]; tensor inputs_33_strides_0 = const()[name = tensor("inputs_33_strides_0"), val = tensor([1])]; tensor inputs_33_pad_0 = const()[name = tensor("inputs_33_pad_0"), val = tensor([0, 0])]; tensor inputs_33_dilations_0 = const()[name = tensor("inputs_33_dilations_0"), val = tensor([1])]; tensor inputs_33_groups_0 = const()[name = tensor("inputs_33_groups_0"), val = tensor(1)]; - tensor input_143 = transpose(perm = input_143_perm_0, x = x_23)[name = tensor("transpose_33")]; - tensor inputs_33 = conv(bias = encoder_conv_module_3_sequential_2_conv_bias, dilations = inputs_33_dilations_0, groups = inputs_33_groups_0, pad = inputs_33_pad_0, pad_type = inputs_33_pad_type_0, strides = inputs_33_strides_0, weight = encoder_conv_module_3_sequential_2_conv_weight, x = input_143)[name = tensor("inputs_33")]; - tensor var_851_split_sizes_0 = const()[name = tensor("op_851_split_sizes_0"), val = tensor([256, 256])]; - tensor var_851_axis_0 = const()[name = tensor("op_851_axis_0"), val = tensor(1)]; - tensor var_851_0, tensor var_851_1 = split(axis = var_851_axis_0, split_sizes = var_851_split_sizes_0, x = inputs_33)[name = tensor("op_851")]; - tensor var_853 = sigmoid(x = var_851_1)[name = tensor("op_853")]; - tensor inputs_35 = mul(x = var_851_0, y = var_853)[name = tensor("inputs_35")]; + tensor input_145 = transpose(perm = input_145_perm_0, x = x_23)[name = tensor("transpose_33")]; + tensor inputs_33 = conv(bias = inner_encoder_conv_module_3_sequential_2_conv_bias, dilations = inputs_33_dilations_0, groups = inputs_33_groups_0, pad = inputs_33_pad_0, pad_type = inputs_33_pad_type_0, strides = inputs_33_strides_0, weight = inner_encoder_conv_module_3_sequential_2_conv_weight, x = input_145)[name = tensor("inputs_33")]; + tensor var_887_split_sizes_0 = const()[name = tensor("op_887_split_sizes_0"), val = tensor([256, 256])]; + tensor var_887_axis_0 = const()[name = tensor("op_887_axis_0"), val = tensor(1)]; + tensor var_887_0, tensor var_887_1 = split(axis = var_887_axis_0, split_sizes = var_887_split_sizes_0, x = inputs_33)[name = tensor("op_887")]; + tensor var_889 = sigmoid(x = var_887_1)[name = tensor("op_889")]; + tensor inputs_35 = mul(x = var_887_0, y = var_889)[name = tensor("inputs_35")]; tensor outputs_aug_pad_type_0 = const()[name = tensor("outputs_aug_pad_type_0"), val = tensor("custom")]; tensor outputs_aug_pad_0 = const()[name = tensor("outputs_aug_pad_0"), val = tensor([15, 15])]; tensor outputs_aug_groups_0 = const()[name = tensor("outputs_aug_groups_0"), val = tensor(256)]; tensor outputs_aug_strides_0 = const()[name = tensor("outputs_aug_strides_0"), val = tensor([1])]; tensor outputs_aug_dilations_0 = const()[name = tensor("outputs_aug_dilations_0"), val = tensor([1])]; - tensor outputs_aug = conv(dilations = outputs_aug_dilations_0, groups = outputs_aug_groups_0, pad = outputs_aug_pad_0, pad_type = outputs_aug_pad_type_0, strides = outputs_aug_strides_0, weight = encoder_conv_module_3_sequential_4_conv_weight, x = inputs_35)[name = tensor("outputs_aug")]; - tensor input_145_begin_0 = const()[name = tensor("input_145_begin_0"), val = tensor([0, 0, 0])]; - tensor input_145_end_0 = const()[name = tensor("input_145_end_0"), val = tensor([2, 256, 16])]; - tensor input_145_end_mask_0 = const()[name = tensor("input_145_end_mask_0"), val = tensor([true, true, false])]; - tensor input_145 = slice_by_index(begin = input_145_begin_0, end = input_145_end_0, end_mask = input_145_end_mask_0, x = outputs_aug)[name = tensor("input_145")]; - tensor inputs_37 = batch_norm(beta = encoder_conv_module_3_sequential_5_bias, epsilon = var_28, gamma = encoder_conv_module_3_sequential_5_weight, mean = encoder_conv_module_3_sequential_5_running_mean, variance = encoder_conv_module_3_sequential_5_running_var, x = input_145)[name = tensor("inputs_37")]; - tensor input_147 = silu(x = inputs_37)[name = tensor("input_147")]; - tensor input_149_pad_type_0 = const()[name = tensor("input_149_pad_type_0"), val = tensor("valid")]; - tensor input_149_strides_0 = const()[name = tensor("input_149_strides_0"), val = tensor([1])]; - tensor input_149_pad_0 = const()[name = tensor("input_149_pad_0"), val = tensor([0, 0])]; - tensor input_149_dilations_0 = const()[name = tensor("input_149_dilations_0"), val = tensor([1])]; - tensor input_149_groups_0 = const()[name = tensor("input_149_groups_0"), val = tensor(1)]; - tensor input_149 = conv(bias = encoder_conv_module_3_sequential_7_conv_bias, dilations = input_149_dilations_0, groups = input_149_groups_0, pad = input_149_pad_0, pad_type = input_149_pad_type_0, strides = input_149_strides_0, weight = encoder_conv_module_3_sequential_7_conv_weight, x = input_147)[name = tensor("input_149")]; + tensor outputs_aug = conv(dilations = outputs_aug_dilations_0, groups = outputs_aug_groups_0, pad = outputs_aug_pad_0, pad_type = outputs_aug_pad_type_0, strides = outputs_aug_strides_0, weight = inner_encoder_conv_module_3_sequential_4_conv_weight, x = inputs_35)[name = tensor("outputs_aug")]; + tensor input_147_begin_0 = const()[name = tensor("input_147_begin_0"), val = tensor([0, 0, 0])]; + tensor input_147_end_0 = const()[name = tensor("input_147_end_0"), val = tensor([2, 256, 16])]; + tensor input_147_end_mask_0 = const()[name = tensor("input_147_end_mask_0"), val = tensor([true, true, false])]; + tensor input_147 = slice_by_index(begin = input_147_begin_0, end = input_147_end_0, end_mask = input_147_end_mask_0, x = outputs_aug)[name = tensor("input_147")]; + tensor inputs_37 = batch_norm(beta = inner_encoder_conv_module_3_sequential_5_bias, epsilon = var_45, gamma = inner_encoder_conv_module_3_sequential_5_weight, mean = inner_encoder_conv_module_3_sequential_5_running_mean, variance = inner_encoder_conv_module_3_sequential_5_running_var, x = input_147)[name = tensor("inputs_37")]; + tensor input_149 = silu(x = inputs_37)[name = tensor("input_149")]; + tensor input_151_pad_type_0 = const()[name = tensor("input_151_pad_type_0"), val = tensor("valid")]; + tensor input_151_strides_0 = const()[name = tensor("input_151_strides_0"), val = tensor([1])]; + tensor input_151_pad_0 = const()[name = tensor("input_151_pad_0"), val = tensor([0, 0])]; + tensor input_151_dilations_0 = const()[name = tensor("input_151_dilations_0"), val = tensor([1])]; + tensor input_151_groups_0 = const()[name = tensor("input_151_groups_0"), val = tensor(1)]; + tensor input_151 = conv(bias = inner_encoder_conv_module_3_sequential_7_conv_bias, dilations = input_151_dilations_0, groups = input_151_groups_0, pad = input_151_pad_0, pad_type = input_151_pad_type_0, strides = input_151_strides_0, weight = inner_encoder_conv_module_3_sequential_7_conv_weight, x = input_149)[name = tensor("input_151")]; tensor conv_out_7_perm_0 = const()[name = tensor("conv_out_7_perm_0"), val = tensor([0, 2, 1])]; - tensor var_884_begin_0 = const()[name = tensor("op_884_begin_0"), val = tensor([0, -1, 0])]; - tensor var_884_end_0 = const()[name = tensor("op_884_end_0"), val = tensor([2, 16, 256])]; - tensor var_884_end_mask_0 = const()[name = tensor("op_884_end_mask_0"), val = tensor([true, true, true])]; - tensor conv_out_7 = transpose(perm = conv_out_7_perm_0, x = input_149)[name = tensor("transpose_32")]; - tensor var_884 = slice_by_index(begin = var_884_begin_0, end = var_884_end_0, end_mask = var_884_end_mask_0, x = conv_out_7)[name = tensor("op_884")]; - tensor var_886_perm_0 = const()[name = tensor("op_886_perm_0"), val = tensor([1, 0, 2])]; - tensor var_886 = transpose(perm = var_886_perm_0, x = var_884)[name = tensor("transpose_31")]; - tensor input_151 = add(x = x_21, y = var_886)[name = tensor("input_151")]; - tensor input_153_axes_0 = const()[name = tensor("input_153_axes_0"), val = tensor([-1])]; - tensor input_153 = layer_norm(axes = input_153_axes_0, beta = encoder_ffn2_3_module_sequential_0_bias, epsilon = var_28, gamma = encoder_ffn2_3_module_sequential_0_weight, x = input_151)[name = tensor("input_153")]; - tensor inputs = linear(bias = encoder_ffn2_3_module_sequential_1_linear_bias, weight = encoder_ffn2_3_module_sequential_1_linear_weight, x = input_153)[name = tensor("linear_35")]; - tensor input_155 = silu(x = inputs)[name = tensor("input_155")]; - tensor input_159 = linear(bias = encoder_ffn2_3_module_sequential_4_linear_bias, weight = encoder_ffn2_3_module_sequential_4_linear_weight, x = input_155)[name = tensor("linear_36")]; - tensor var_909 = const()[name = tensor("op_909"), val = tensor(0x1p-1)]; - tensor var_910 = mul(x = input_159, y = var_909)[name = tensor("op_910")]; - tensor input_161 = add(x = var_910, y = input_151)[name = tensor("input_161")]; + tensor var_920_begin_0 = const()[name = tensor("op_920_begin_0"), val = tensor([0, -1, 0])]; + tensor var_920_end_0 = const()[name = tensor("op_920_end_0"), val = tensor([2, 16, 256])]; + tensor var_920_end_mask_0 = const()[name = tensor("op_920_end_mask_0"), val = tensor([true, true, true])]; + tensor conv_out_7 = transpose(perm = conv_out_7_perm_0, x = input_151)[name = tensor("transpose_32")]; + tensor var_920 = slice_by_index(begin = var_920_begin_0, end = var_920_end_0, end_mask = var_920_end_mask_0, x = conv_out_7)[name = tensor("op_920")]; + tensor var_922_perm_0 = const()[name = tensor("op_922_perm_0"), val = tensor([1, 0, 2])]; + tensor var_922 = transpose(perm = var_922_perm_0, x = var_920)[name = tensor("transpose_31")]; + tensor input_153 = add(x = x_21, y = var_922)[name = tensor("input_153")]; + tensor input_155_axes_0 = const()[name = tensor("input_155_axes_0"), val = tensor([-1])]; + tensor input_155 = layer_norm(axes = input_155_axes_0, beta = inner_encoder_ffn2_3_module_sequential_0_bias, epsilon = var_45, gamma = inner_encoder_ffn2_3_module_sequential_0_weight, x = input_153)[name = tensor("input_155")]; + tensor inputs = linear(bias = inner_encoder_ffn2_3_module_sequential_1_linear_bias, weight = inner_encoder_ffn2_3_module_sequential_1_linear_weight, x = input_155)[name = tensor("linear_35")]; + tensor input_157 = silu(x = inputs)[name = tensor("input_157")]; + tensor input_161 = linear(bias = inner_encoder_ffn2_3_module_sequential_4_linear_bias, weight = inner_encoder_ffn2_3_module_sequential_4_linear_weight, x = input_157)[name = tensor("linear_36")]; + tensor var_945 = const()[name = tensor("op_945"), val = tensor(0x1p-1)]; + tensor var_946 = mul(x = input_161, y = var_945)[name = tensor("op_946")]; + tensor input_163 = add(x = var_946, y = input_153)[name = tensor("input_163")]; tensor x_25_axes_0 = const()[name = tensor("x_25_axes_0"), val = tensor([-1])]; - tensor x_25 = layer_norm(axes = x_25_axes_0, beta = encoder_layer_norm_3_bias, epsilon = var_28, gamma = encoder_layer_norm_3_weight, x = input_161)[name = tensor("x_25")]; + tensor x_25 = layer_norm(axes = x_25_axes_0, beta = inner_encoder_layer_norm_3_bias, epsilon = var_45, gamma = inner_encoder_layer_norm_3_weight, x = input_163)[name = tensor("x_25")]; tensor x_ct_perm_0 = const()[name = tensor("x_ct_perm_0"), val = tensor([0, 2, 1])]; tensor cat_interleave_0 = const()[name = tensor("cat_interleave_0"), val = tensor(false)]; tensor x_ct = transpose(perm = x_ct_perm_0, x = x_25)[name = tensor("transpose_30")]; - tensor cat = concat(axis = var_20, interleave = cat_interleave_0, values = (cnn_window, x_ct))[name = tensor("cat")]; + tensor cat = concat(axis = var_50, interleave = cat_interleave_0, values = (cnn_window, x_ct))[name = tensor("cat")]; tensor conv_out_pad_type_0 = const()[name = tensor("conv_out_pad_type_0"), val = tensor("valid")]; tensor conv_out_strides_0 = const()[name = tensor("conv_out_strides_0"), val = tensor([1])]; tensor conv_out_pad_0 = const()[name = tensor("conv_out_pad_0"), val = tensor([0, 0])]; tensor conv_out_dilations_0 = const()[name = tensor("conv_out_dilations_0"), val = tensor([1])]; tensor conv_out_groups_0 = const()[name = tensor("conv_out_groups_0"), val = tensor(1)]; - tensor conv_out = conv(bias = encoder_cnn_bias, dilations = conv_out_dilations_0, groups = conv_out_groups_0, pad = conv_out_pad_0, pad_type = conv_out_pad_type_0, strides = conv_out_strides_0, weight = encoder_cnn_weight, x = cat)[name = tensor("conv_out")]; - tensor var_928_begin_0 = const()[name = tensor("op_928_begin_0"), val = tensor([0, 0, 2])]; - tensor var_928_end_0 = const()[name = tensor("op_928_end_0"), val = tensor([1, 256, 20])]; - tensor var_928_end_mask_0 = const()[name = tensor("op_928_end_mask_0"), val = tensor([true, true, true])]; - tensor cnn_window_new = slice_by_index(begin = var_928_begin_0, end = var_928_end_0, end_mask = var_928_end_mask_0, x = cat)[name = tensor("op_928")]; - tensor input_163_perm_0 = const()[name = tensor("input_163_perm_0"), val = tensor([0, 2, 1])]; - tensor var_930 = const()[name = tensor("op_930"), val = tensor([-1])]; - tensor input_163 = transpose(perm = input_163_perm_0, x = conv_out)[name = tensor("transpose_29")]; - tensor var_931 = reduce_l2_norm(axes = var_930, keep_dims = var_29, x = input_163)[name = tensor("op_931")]; + tensor conv_out = conv(bias = inner_encoder_cnn_bias, dilations = conv_out_dilations_0, groups = conv_out_groups_0, pad = conv_out_pad_0, pad_type = conv_out_pad_type_0, strides = conv_out_strides_0, weight = inner_encoder_cnn_weight, x = cat)[name = tensor("conv_out")]; + tensor var_964_begin_0 = const()[name = tensor("op_964_begin_0"), val = tensor([0, 0, 2])]; + tensor var_964_end_0 = const()[name = tensor("op_964_end_0"), val = tensor([1, 256, 20])]; + tensor var_964_end_mask_0 = const()[name = tensor("op_964_end_mask_0"), val = tensor([true, true, true])]; + tensor cnn_window_new = slice_by_index(begin = var_964_begin_0, end = var_964_end_0, end_mask = var_964_end_mask_0, x = cat)[name = tensor("op_964")]; + tensor input_165_perm_0 = const()[name = tensor("input_165_perm_0"), val = tensor([0, 2, 1])]; + tensor var_966 = const()[name = tensor("op_966"), val = tensor([-1])]; + tensor input_165 = transpose(perm = input_165_perm_0, x = conv_out)[name = tensor("transpose_29")]; + tensor var_967 = reduce_l2_norm(axes = var_966, keep_dims = var_44, x = input_165)[name = tensor("op_967")]; tensor const_12 = const()[name = tensor("const_12"), val = tensor(0x1.fffffep+127)]; - tensor clip_0 = clip(alpha = var_11, beta = const_12, x = var_931)[name = tensor("clip_0")]; - tensor emb = real_div(x = input_163, y = clip_0)[name = tensor("emb")]; - tensor var_935_axis_0 = const()[name = tensor("op_935_axis_0"), val = tensor(0)]; - tensor enc_kv_new = stack(axis = var_935_axis_0, values = (var_206, var_404, var_602, nkv_1))[name = tensor("op_935")]; - tensor var_937_axis_0 = const()[name = tensor("op_937_axis_0"), val = tensor(0)]; - tensor enc_scale_new = stack(axis = var_937_axis_0, values = (new_scale_1, new_scale_3, new_scale_5, new_scale_7))[name = tensor("op_937")]; - tensor var_939_axis_0 = const()[name = tensor("op_939_axis_0"), val = tensor(0)]; - tensor enc_conv_cache_new = stack(axis = var_939_axis_0, values = (window_5, window_11, window_17, window))[name = tensor("op_939")]; - tensor var_948 = const()[name = tensor("op_948"), val = tensor(0x1.5798eep-27)]; - tensor var_953 = const()[name = tensor("op_953"), val = tensor(0x1.4f8b58p-17)]; - tensor var_955 = const()[name = tensor("op_955"), val = tensor(0x1.0c6f7ap-20)]; - tensor var_956 = const()[name = tensor("op_956"), val = tensor(true)]; - tensor var_958 = const()[name = tensor("op_958"), val = tensor(0x1p+0)]; - tensor var_962 = const()[name = tensor("op_962"), val = tensor(-1)]; - tensor var_968 = const()[name = tensor("op_968"), val = tensor(0)]; + tensor clip_0 = clip(alpha = var_58, beta = const_12, x = var_967)[name = tensor("clip_0")]; + tensor emb = real_div(x = input_165, y = clip_0)[name = tensor("emb")]; + tensor var_971_axis_0 = const()[name = tensor("op_971_axis_0"), val = tensor(0)]; + tensor enc_kv_new = stack(axis = var_971_axis_0, values = (var_242, var_440, var_638, nkv_1))[name = tensor("op_971")]; + tensor var_973_axis_0 = const()[name = tensor("op_973_axis_0"), val = tensor(0)]; + tensor enc_scale_new = stack(axis = var_973_axis_0, values = (new_scale_1, new_scale_3, new_scale_5, new_scale_7))[name = tensor("op_973")]; + tensor var_975_axis_0 = const()[name = tensor("op_975_axis_0"), val = tensor(0)]; + tensor enc_conv_cache_new = stack(axis = var_975_axis_0, values = (window_5, window_11, window_17, window))[name = tensor("op_975")]; tensor pos = const()[name = tensor("pos"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(44395328)))]; - tensor var_1030_axes_0 = const()[name = tensor("op_1030_axes_0"), val = tensor([2])]; - tensor var_1030 = expand_dims(axes = var_1030_axes_0, x = emb)[name = tensor("op_1030")]; + tensor var_1043_axes_0 = const()[name = tensor("op_1043_axes_0"), val = tensor([2])]; + tensor var_1043 = expand_dims(axes = var_1043_axes_0, x = emb)[name = tensor("op_1043")]; tensor emb_exp_reps_0 = const()[name = tensor("emb_exp_reps_0"), val = tensor([1, 1, 6, 1])]; - tensor emb_exp = tile(reps = emb_exp_reps_0, x = var_1030)[name = tensor("emb_exp")]; - tensor input_165_interleave_0 = const()[name = tensor("input_165_interleave_0"), val = tensor(false)]; - tensor input_165 = concat(axis = var_962, interleave = input_165_interleave_0, values = (emb_exp, pos))[name = tensor("input_165")]; - tensor x_27 = linear(bias = decoder_convert_bias, weight = decoder_convert_weight, x = input_165)[name = tensor("linear_37")]; - tensor var_1038_perm_0 = const()[name = tensor("op_1038_perm_0"), val = tensor([0, 2, 1, 3])]; - tensor var_1042 = const()[name = tensor("op_1042"), val = tensor([6, 2, 256])]; - tensor var_1038 = transpose(perm = var_1038_perm_0, x = x_27)[name = tensor("transpose_28")]; - tensor x_29 = reshape(shape = var_1042, x = var_1038)[name = tensor("x_29")]; + tensor emb_exp = tile(reps = emb_exp_reps_0, x = var_1043)[name = tensor("emb_exp")]; + tensor input_167_interleave_0 = const()[name = tensor("input_167_interleave_0"), val = tensor(false)]; + tensor input_167 = concat(axis = var_51, interleave = input_167_interleave_0, values = (emb_exp, pos))[name = tensor("input_167")]; + tensor x_27 = linear(bias = inner_decoder_convert_bias, weight = inner_decoder_convert_weight, x = input_167)[name = tensor("linear_37")]; + tensor var_1051_perm_0 = const()[name = tensor("op_1051_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_1055 = const()[name = tensor("op_1055"), val = tensor([6, 2, 256])]; + tensor var_1051 = transpose(perm = var_1051_perm_0, x = x_27)[name = tensor("transpose_28")]; + tensor x_29 = reshape(shape = var_1055, x = var_1051)[name = tensor("x_29")]; tensor prev_kv_9_begin_0 = const()[name = tensor("prev_kv_9_begin_0"), val = tensor([0, 0, 0, 0, 0])]; tensor prev_kv_9_end_0 = const()[name = tensor("prev_kv_9_end_0"), val = tensor([1, 6, 4, 64, 64])]; tensor prev_kv_9_end_mask_0 = const()[name = tensor("prev_kv_9_end_mask_0"), val = tensor([false, true, true, true, true])]; @@ -903,132 +910,132 @@ program(1.0) tensor prev_scale_9_end_mask_0 = const()[name = tensor("prev_scale_9_end_mask_0"), val = tensor([false, true])]; tensor prev_scale_9_squeeze_mask_0 = const()[name = tensor("prev_scale_9_squeeze_mask_0"), val = tensor([true, false])]; tensor prev_scale_9 = slice_by_index(begin = prev_scale_9_begin_0, end = prev_scale_9_end_0, end_mask = prev_scale_9_end_mask_0, squeeze_mask = prev_scale_9_squeeze_mask_0, x = dec_scale)[name = tensor("prev_scale_9")]; - tensor var_1050 = linear(bias = decoder_q_proj_0_bias, weight = decoder_q_proj_0_weight, x = x_29)[name = tensor("linear_38")]; - tensor var_1051 = const()[name = tensor("op_1051"), val = tensor([6, 2, 4, 64])]; - tensor var_1052 = reshape(shape = var_1051, x = var_1050)[name = tensor("op_1052")]; + tensor var_1063 = linear(bias = inner_decoder_q_proj_0_bias, weight = inner_decoder_q_proj_0_weight, x = x_29)[name = tensor("linear_38")]; + tensor var_1064 = const()[name = tensor("op_1064"), val = tensor([6, 2, 4, 64])]; + tensor var_1065 = reshape(shape = var_1064, x = var_1063)[name = tensor("op_1065")]; tensor q_9_perm_0 = const()[name = tensor("q_9_perm_0"), val = tensor([0, 2, 1, 3])]; - tensor var_1056 = linear(bias = decoder_k_proj_0_bias, weight = decoder_k_proj_0_weight, x = x_29)[name = tensor("linear_39")]; - tensor var_1057 = const()[name = tensor("op_1057"), val = tensor(0x1p-3)]; - tensor var_1058 = mul(x = var_1056, y = var_1057)[name = tensor("op_1058")]; - tensor var_1059 = const()[name = tensor("op_1059"), val = tensor([6, 2, 4, 64])]; - tensor var_1060 = reshape(shape = var_1059, x = var_1058)[name = tensor("op_1060")]; + tensor var_1069 = linear(bias = inner_decoder_k_proj_0_bias, weight = inner_decoder_k_proj_0_weight, x = x_29)[name = tensor("linear_39")]; + tensor var_1070 = const()[name = tensor("op_1070"), val = tensor(0x1p-3)]; + tensor var_1071 = mul(x = var_1069, y = var_1070)[name = tensor("op_1071")]; + tensor var_1072 = const()[name = tensor("op_1072"), val = tensor([6, 2, 4, 64])]; + tensor var_1073 = reshape(shape = var_1072, x = var_1071)[name = tensor("op_1073")]; tensor k_9_perm_0 = const()[name = tensor("k_9_perm_0"), val = tensor([0, 2, 1, 3])]; - tensor var_1064 = linear(bias = decoder_v_proj_0_bias, weight = decoder_v_proj_0_weight, x = x_29)[name = tensor("linear_40")]; - tensor var_1065 = const()[name = tensor("op_1065"), val = tensor([6, 2, 4, 64])]; - tensor var_1066 = reshape(shape = var_1065, x = var_1064)[name = tensor("op_1066")]; + tensor var_1077 = linear(bias = inner_decoder_v_proj_0_bias, weight = inner_decoder_v_proj_0_weight, x = x_29)[name = tensor("linear_40")]; + tensor var_1078 = const()[name = tensor("op_1078"), val = tensor([6, 2, 4, 64])]; + tensor var_1079 = reshape(shape = var_1078, x = var_1077)[name = tensor("op_1079")]; tensor v_9_perm_0 = const()[name = tensor("v_9_perm_0"), val = tensor([0, 2, 1, 3])]; - tensor input_169 = linear(bias = decoder_g_proj_0_bias, weight = decoder_g_proj_0_weight, x = x_29)[name = tensor("linear_41")]; + tensor input_171 = linear(bias = inner_decoder_g_proj_0_bias, weight = inner_decoder_g_proj_0_weight, x = x_29)[name = tensor("linear_41")]; tensor cumsum_mask_1_exclusive_0 = const()[name = tensor("cumsum_mask_1_exclusive_0"), val = tensor(false)]; tensor cumsum_mask_1_reverse_0 = const()[name = tensor("cumsum_mask_1_reverse_0"), val = tensor(false)]; - tensor cumsum_mask_1 = cumsum(axis = var_968, exclusive = cumsum_mask_1_exclusive_0, reverse = cumsum_mask_1_reverse_0, x = valid_mask)[name = tensor("cumsum_mask_1")]; + tensor cumsum_mask_1 = cumsum(axis = var_48, exclusive = cumsum_mask_1_exclusive_0, reverse = cumsum_mask_1_reverse_0, x = valid_mask)[name = tensor("cumsum_mask_1")]; tensor sqrt_s0_9 = sqrt(x = prev_scale_9)[name = tensor("sqrt_s0_9")]; tensor s_t_9 = add(x = prev_scale_9, y = cumsum_mask_1)[name = tensor("s_t_9")]; tensor const_20 = const()[name = tensor("const_20"), val = tensor(0x1.fffffep+127)]; - tensor clip_1 = clip(alpha = var_958, beta = const_20, x = s_t_9)[name = tensor("clip_1")]; + tensor clip_1 = clip(alpha = var_38, beta = const_20, x = s_t_9)[name = tensor("clip_1")]; tensor sqrt_s_t_9 = sqrt(x = clip_1)[name = tensor("sqrt_s_t_9")]; tensor qk_9_transpose_x_1 = const()[name = tensor("qk_9_transpose_x_1"), val = tensor(false)]; tensor qk_9_transpose_y_1 = const()[name = tensor("qk_9_transpose_y_1"), val = tensor(true)]; - tensor k_9 = transpose(perm = k_9_perm_0, x = var_1060)[name = tensor("transpose_26")]; - tensor q_9 = transpose(perm = q_9_perm_0, x = var_1052)[name = tensor("transpose_27")]; + tensor k_9 = transpose(perm = k_9_perm_0, x = var_1073)[name = tensor("transpose_26")]; + tensor q_9 = transpose(perm = q_9_perm_0, x = var_1065)[name = tensor("transpose_27")]; tensor qk_9 = matmul(transpose_x = qk_9_transpose_x_1, transpose_y = qk_9_transpose_y_1, x = q_9, y = k_9)[name = tensor("qk_9")]; - tensor var_1078 = const()[name = tensor("op_1078"), val = tensor([1, 2])]; - tensor var_1079 = reshape(shape = var_1078, x = valid_mask)[name = tensor("op_1079")]; - tensor causal_with_valid_1 = mul(x = decoder__causal_mask, y = var_1079)[name = tensor("causal_with_valid_1")]; - tensor var_1081 = const()[name = tensor("op_1081"), val = tensor([2, 1])]; - tensor var_1082 = reshape(shape = var_1081, x = sqrt_s_t_9)[name = tensor("op_1082")]; - tensor M_9 = real_div(x = causal_with_valid_1, y = var_1082)[name = tensor("M_9")]; - tensor var_1084 = mul(x = qk_9, y = M_9)[name = tensor("op_1084")]; + tensor var_1091 = const()[name = tensor("op_1091"), val = tensor([1, 2])]; + tensor var_1092 = reshape(shape = var_1091, x = valid_mask)[name = tensor("op_1092")]; + tensor causal_with_valid_1 = mul(x = inner_decoder__causal_mask, y = var_1092)[name = tensor("causal_with_valid_1")]; + tensor var_1094 = const()[name = tensor("op_1094"), val = tensor([2, 1])]; + tensor var_1095 = reshape(shape = var_1094, x = sqrt_s_t_9)[name = tensor("op_1095")]; + tensor M_9 = real_div(x = causal_with_valid_1, y = var_1095)[name = tensor("M_9")]; + tensor var_1097 = mul(x = qk_9, y = M_9)[name = tensor("op_1097")]; tensor inner_9_transpose_x_0 = const()[name = tensor("inner_9_transpose_x_0"), val = tensor(false)]; tensor inner_9_transpose_y_0 = const()[name = tensor("inner_9_transpose_y_0"), val = tensor(false)]; - tensor v_9 = transpose(perm = v_9_perm_0, x = var_1066)[name = tensor("transpose_25")]; - tensor inner_9 = matmul(transpose_x = inner_9_transpose_x_0, transpose_y = inner_9_transpose_y_0, x = var_1084, y = v_9)[name = tensor("inner_9")]; - tensor var_1086_transpose_x_0 = const()[name = tensor("op_1086_transpose_x_0"), val = tensor(false)]; - tensor var_1086_transpose_y_0 = const()[name = tensor("op_1086_transpose_y_0"), val = tensor(false)]; - tensor var_1086 = matmul(transpose_x = var_1086_transpose_x_0, transpose_y = var_1086_transpose_y_0, x = q_9, y = prev_kv_9)[name = tensor("op_1086")]; - tensor var_1087 = real_div(x = sqrt_s0_9, y = sqrt_s_t_9)[name = tensor("op_1087")]; - tensor var_1088 = const()[name = tensor("op_1088"), val = tensor([1, 1, 2, 1])]; - tensor var_1089 = reshape(shape = var_1088, x = var_1087)[name = tensor("op_1089")]; - tensor cross_9 = mul(x = var_1086, y = var_1089)[name = tensor("cross_9")]; + tensor v_9 = transpose(perm = v_9_perm_0, x = var_1079)[name = tensor("transpose_25")]; + tensor inner_9 = matmul(transpose_x = inner_9_transpose_x_0, transpose_y = inner_9_transpose_y_0, x = var_1097, y = v_9)[name = tensor("inner_9")]; + tensor var_1099_transpose_x_0 = const()[name = tensor("op_1099_transpose_x_0"), val = tensor(false)]; + tensor var_1099_transpose_y_0 = const()[name = tensor("op_1099_transpose_y_0"), val = tensor(false)]; + tensor var_1099 = matmul(transpose_x = var_1099_transpose_x_0, transpose_y = var_1099_transpose_y_0, x = q_9, y = prev_kv_9)[name = tensor("op_1099")]; + tensor var_1100 = real_div(x = sqrt_s0_9, y = sqrt_s_t_9)[name = tensor("op_1100")]; + tensor var_1101 = const()[name = tensor("op_1101"), val = tensor([1, 1, 2, 1])]; + tensor var_1102 = reshape(shape = var_1101, x = var_1100)[name = tensor("op_1102")]; + tensor cross_9 = mul(x = var_1099, y = var_1102)[name = tensor("cross_9")]; tensor out_25 = add(x = inner_9, y = cross_9)[name = tensor("out_25")]; - tensor var_1092 = const()[name = tensor("op_1092"), val = tensor([1, 1, 2, 1])]; - tensor var_1093 = reshape(shape = var_1092, x = valid_mask)[name = tensor("op_1093")]; - tensor v_masked_1 = mul(x = v_9, y = var_1093)[name = tensor("v_masked_1")]; - tensor var_1095 = mul(x = prev_kv_9, y = sqrt_s0_9)[name = tensor("op_1095")]; - tensor var_1097_transpose_x_1 = const()[name = tensor("op_1097_transpose_x_1"), val = tensor(true)]; - tensor var_1097_transpose_y_1 = const()[name = tensor("op_1097_transpose_y_1"), val = tensor(false)]; - tensor var_1097 = matmul(transpose_x = var_1097_transpose_x_1, transpose_y = var_1097_transpose_y_1, x = k_9, y = v_masked_1)[name = tensor("op_1097")]; - tensor new_kv_unnorm_9 = add(x = var_1095, y = var_1097)[name = tensor("new_kv_unnorm_9")]; - tensor var_1099_keep_dims_0 = const()[name = tensor("op_1099_keep_dims_0"), val = tensor(false)]; - tensor var_1099 = reduce_sum(keep_dims = var_1099_keep_dims_0, x = valid_mask)[name = tensor("op_1099")]; - tensor var_1100 = const()[name = tensor("op_1100"), val = tensor([1])]; - tensor var_1101 = reshape(shape = var_1100, x = var_1099)[name = tensor("op_1101")]; - tensor new_scale_9 = add(x = prev_scale_9, y = var_1101)[name = tensor("new_scale_9")]; + tensor var_1105 = const()[name = tensor("op_1105"), val = tensor([1, 1, 2, 1])]; + tensor var_1106 = reshape(shape = var_1105, x = valid_mask)[name = tensor("op_1106")]; + tensor v_masked_1 = mul(x = v_9, y = var_1106)[name = tensor("v_masked_1")]; + tensor var_1108 = mul(x = prev_kv_9, y = sqrt_s0_9)[name = tensor("op_1108")]; + tensor var_1110_transpose_x_1 = const()[name = tensor("op_1110_transpose_x_1"), val = tensor(true)]; + tensor var_1110_transpose_y_1 = const()[name = tensor("op_1110_transpose_y_1"), val = tensor(false)]; + tensor var_1110 = matmul(transpose_x = var_1110_transpose_x_1, transpose_y = var_1110_transpose_y_1, x = k_9, y = v_masked_1)[name = tensor("op_1110")]; + tensor new_kv_unnorm_9 = add(x = var_1108, y = var_1110)[name = tensor("new_kv_unnorm_9")]; + tensor var_1112_keep_dims_0 = const()[name = tensor("op_1112_keep_dims_0"), val = tensor(false)]; + tensor var_1112 = reduce_sum(keep_dims = var_1112_keep_dims_0, x = valid_mask)[name = tensor("op_1112")]; + tensor var_1113 = const()[name = tensor("op_1113"), val = tensor([1])]; + tensor var_1114 = reshape(shape = var_1113, x = var_1112)[name = tensor("op_1114")]; + tensor new_scale_9 = add(x = prev_scale_9, y = var_1114)[name = tensor("new_scale_9")]; tensor const_21 = const()[name = tensor("const_21"), val = tensor(0x1.fffffep+127)]; - tensor clip_2 = clip(alpha = var_958, beta = const_21, x = new_scale_9)[name = tensor("clip_2")]; + tensor clip_2 = clip(alpha = var_38, beta = const_21, x = new_scale_9)[name = tensor("clip_2")]; tensor sqrt_new_scale_1 = sqrt(x = clip_2)[name = tensor("sqrt_new_scale_1")]; - tensor var_1105 = real_div(x = new_kv_unnorm_9, y = sqrt_new_scale_1)[name = tensor("op_1105")]; - tensor var_1106_perm_0 = const()[name = tensor("op_1106_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_1118 = real_div(x = new_kv_unnorm_9, y = sqrt_new_scale_1)[name = tensor("op_1118")]; + tensor var_1119_perm_0 = const()[name = tensor("op_1119_perm_0"), val = tensor([0, 2, 1, 3])]; tensor out_27_axes_0 = const()[name = tensor("out_27_axes_0"), val = tensor([-1])]; - tensor var_1106 = transpose(perm = var_1106_perm_0, x = out_25)[name = tensor("transpose_24")]; - tensor out_27 = layer_norm(axes = out_27_axes_0, epsilon = var_955, x = var_1106)[name = tensor("out_27")]; - tensor var_1110 = const()[name = tensor("op_1110"), val = tensor([6, 2, 256])]; - tensor out_29 = reshape(shape = var_1110, x = out_27)[name = tensor("out_29")]; - tensor var_1112 = silu(x = input_169)[name = tensor("op_1112")]; - tensor input_171 = mul(x = var_1112, y = out_29)[name = tensor("input_171")]; - tensor ret_out_9 = linear(bias = decoder_out_proj_0_bias, weight = decoder_out_proj_0_weight, x = input_171)[name = tensor("linear_42")]; - tensor input_173 = add(x = x_29, y = ret_out_9)[name = tensor("input_173")]; + tensor var_1119 = transpose(perm = var_1119_perm_0, x = out_25)[name = tensor("transpose_24")]; + tensor out_27 = layer_norm(axes = out_27_axes_0, epsilon = var_53, x = var_1119)[name = tensor("out_27")]; + tensor var_1123 = const()[name = tensor("op_1123"), val = tensor([6, 2, 256])]; + tensor out_29 = reshape(shape = var_1123, x = out_27)[name = tensor("out_29")]; + tensor var_1125 = silu(x = input_171)[name = tensor("op_1125")]; + tensor input_173 = mul(x = var_1125, y = out_29)[name = tensor("input_173")]; + tensor ret_out_9 = linear(bias = inner_decoder_out_proj_0_bias, weight = inner_decoder_out_proj_0_weight, x = input_173)[name = tensor("linear_42")]; + tensor input_175 = add(x = x_29, y = ret_out_9)[name = tensor("input_175")]; tensor xt_1_axes_0 = const()[name = tensor("xt_1_axes_0"), val = tensor([-1])]; - tensor xt_1 = layer_norm(axes = xt_1_axes_0, beta = decoder_norm11_0_bias, epsilon = var_953, gamma = decoder_norm11_0_weight, x = input_173)[name = tensor("xt_1")]; - tensor var_1122 = const()[name = tensor("op_1122"), val = tensor([1, 6, 2, 256])]; - tensor var_1123 = reshape(shape = var_1122, x = xt_1)[name = tensor("op_1123")]; - tensor var_1124_perm_0 = const()[name = tensor("op_1124_perm_0"), val = tensor([0, 2, 1, 3])]; - tensor var_1127 = const()[name = tensor("op_1127"), val = tensor([2, 6, 256])]; - tensor var_1124 = transpose(perm = var_1124_perm_0, x = var_1123)[name = tensor("transpose_23")]; - tensor query_1 = reshape(shape = var_1127, x = var_1124)[name = tensor("query_1")]; + tensor xt_1 = layer_norm(axes = xt_1_axes_0, beta = inner_decoder_norm11_0_bias, epsilon = var_45, gamma = inner_decoder_norm11_0_weight, x = input_175)[name = tensor("xt_1")]; + tensor var_1135 = const()[name = tensor("op_1135"), val = tensor([1, 6, 2, 256])]; + tensor var_1136 = reshape(shape = var_1135, x = xt_1)[name = tensor("op_1136")]; + tensor var_1137_perm_0 = const()[name = tensor("op_1137_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_1140 = const()[name = tensor("op_1140"), val = tensor([2, 6, 256])]; + tensor var_1137 = transpose(perm = var_1137_perm_0, x = var_1136)[name = tensor("transpose_23")]; + tensor query_1 = reshape(shape = var_1140, x = var_1137)[name = tensor("query_1")]; tensor query_3_perm_0 = const()[name = tensor("query_3_perm_0"), val = tensor([1, 0, 2])]; tensor query_3 = transpose(perm = query_3_perm_0, x = query_1)[name = tensor("transpose_22")]; - tensor var_1150 = linear(bias = decoder_self_attn2_0_in_proj_bias, weight = decoder_self_attn2_0_in_proj_weight, x = query_3)[name = tensor("linear_43")]; + tensor var_1163 = linear(bias = inner_decoder_self_attn2_0_in_proj_bias, weight = inner_decoder_self_attn2_0_in_proj_weight, x = query_3)[name = tensor("linear_43")]; tensor concat_1 = const()[name = tensor("concat_1"), val = tensor([6, 2, 3, 256])]; - tensor var_1152 = reshape(shape = concat_1, x = var_1150)[name = tensor("op_1152")]; - tensor var_1153_axes_0 = const()[name = tensor("op_1153_axes_0"), val = tensor([0])]; - tensor var_1153 = expand_dims(axes = var_1153_axes_0, x = var_1152)[name = tensor("op_1153")]; - tensor var_1154_perm_0 = const()[name = tensor("op_1154_perm_0"), val = tensor([-2, 1, 2, 0, 4])]; - tensor var_1155_axes_0 = const()[name = tensor("op_1155_axes_0"), val = tensor([-2])]; - tensor var_1154 = transpose(perm = var_1154_perm_0, x = var_1153)[name = tensor("transpose_21")]; - tensor var_1155 = squeeze(axes = var_1155_axes_0, x = var_1154)[name = tensor("op_1155")]; + tensor var_1165 = reshape(shape = concat_1, x = var_1163)[name = tensor("op_1165")]; + tensor var_1166_axes_0 = const()[name = tensor("op_1166_axes_0"), val = tensor([0])]; + tensor var_1166 = expand_dims(axes = var_1166_axes_0, x = var_1165)[name = tensor("op_1166")]; + tensor var_1167_perm_0 = const()[name = tensor("op_1167_perm_0"), val = tensor([-2, 1, 2, 0, 4])]; + tensor var_1168_axes_0 = const()[name = tensor("op_1168_axes_0"), val = tensor([-2])]; + tensor var_1167 = transpose(perm = var_1167_perm_0, x = var_1166)[name = tensor("transpose_21")]; + tensor var_1168 = squeeze(axes = var_1168_axes_0, x = var_1167)[name = tensor("op_1168")]; tensor q_11_begin_0 = const()[name = tensor("q_11_begin_0"), val = tensor([0, 0, 0, 0])]; tensor q_11_end_0 = const()[name = tensor("q_11_end_0"), val = tensor([1, 6, 2, 256])]; tensor q_11_end_mask_0 = const()[name = tensor("q_11_end_mask_0"), val = tensor([false, true, true, true])]; tensor q_11_squeeze_mask_0 = const()[name = tensor("q_11_squeeze_mask_0"), val = tensor([true, false, false, false])]; - tensor q_11 = slice_by_index(begin = q_11_begin_0, end = q_11_end_0, end_mask = q_11_end_mask_0, squeeze_mask = q_11_squeeze_mask_0, x = var_1155)[name = tensor("q_11")]; + tensor q_11 = slice_by_index(begin = q_11_begin_0, end = q_11_end_0, end_mask = q_11_end_mask_0, squeeze_mask = q_11_squeeze_mask_0, x = var_1168)[name = tensor("q_11")]; tensor k_11_begin_0 = const()[name = tensor("k_11_begin_0"), val = tensor([1, 0, 0, 0])]; tensor k_11_end_0 = const()[name = tensor("k_11_end_0"), val = tensor([2, 6, 2, 256])]; tensor k_11_end_mask_0 = const()[name = tensor("k_11_end_mask_0"), val = tensor([false, true, true, true])]; tensor k_11_squeeze_mask_0 = const()[name = tensor("k_11_squeeze_mask_0"), val = tensor([true, false, false, false])]; - tensor k_11 = slice_by_index(begin = k_11_begin_0, end = k_11_end_0, end_mask = k_11_end_mask_0, squeeze_mask = k_11_squeeze_mask_0, x = var_1155)[name = tensor("k_11")]; + tensor k_11 = slice_by_index(begin = k_11_begin_0, end = k_11_end_0, end_mask = k_11_end_mask_0, squeeze_mask = k_11_squeeze_mask_0, x = var_1168)[name = tensor("k_11")]; tensor v_11_begin_0 = const()[name = tensor("v_11_begin_0"), val = tensor([2, 0, 0, 0])]; tensor v_11_end_0 = const()[name = tensor("v_11_end_0"), val = tensor([3, 6, 2, 256])]; tensor v_11_end_mask_0 = const()[name = tensor("v_11_end_mask_0"), val = tensor([false, true, true, true])]; tensor v_11_squeeze_mask_0 = const()[name = tensor("v_11_squeeze_mask_0"), val = tensor([true, false, false, false])]; - tensor v_11 = slice_by_index(begin = v_11_begin_0, end = v_11_end_0, end_mask = v_11_end_mask_0, squeeze_mask = v_11_squeeze_mask_0, x = var_1155)[name = tensor("v_11")]; - tensor var_1163 = const()[name = tensor("op_1163"), val = tensor([6, 8, 64])]; - tensor var_1164 = reshape(shape = var_1163, x = q_11)[name = tensor("op_1164")]; + tensor v_11 = slice_by_index(begin = v_11_begin_0, end = v_11_end_0, end_mask = v_11_end_mask_0, squeeze_mask = v_11_squeeze_mask_0, x = var_1168)[name = tensor("v_11")]; + tensor var_1176 = const()[name = tensor("op_1176"), val = tensor([6, 8, 64])]; + tensor var_1177 = reshape(shape = var_1176, x = q_11)[name = tensor("op_1177")]; tensor q_13_perm_0 = const()[name = tensor("q_13_perm_0"), val = tensor([1, 0, 2])]; - tensor var_1170 = const()[name = tensor("op_1170"), val = tensor([6, 8, 64])]; - tensor var_1171 = reshape(shape = var_1170, x = k_11)[name = tensor("op_1171")]; + tensor var_1183 = const()[name = tensor("op_1183"), val = tensor([6, 8, 64])]; + tensor var_1184 = reshape(shape = var_1183, x = k_11)[name = tensor("op_1184")]; tensor k_13_perm_0 = const()[name = tensor("k_13_perm_0"), val = tensor([1, 0, 2])]; - tensor var_1177 = const()[name = tensor("op_1177"), val = tensor([6, 8, 64])]; - tensor var_1178 = reshape(shape = var_1177, x = v_11)[name = tensor("op_1178")]; + tensor var_1190 = const()[name = tensor("op_1190"), val = tensor([6, 8, 64])]; + tensor var_1191 = reshape(shape = var_1190, x = v_11)[name = tensor("op_1191")]; tensor v_13_perm_0 = const()[name = tensor("v_13_perm_0"), val = tensor([1, 0, 2])]; - tensor var_1181 = const()[name = tensor("op_1181"), val = tensor([2, 4, 6, 64])]; - tensor q_13 = transpose(perm = q_13_perm_0, x = var_1164)[name = tensor("transpose_20")]; - tensor q_15 = reshape(shape = var_1181, x = q_13)[name = tensor("q_15")]; - tensor var_1183 = const()[name = tensor("op_1183"), val = tensor([2, 4, 6, 64])]; - tensor k_13 = transpose(perm = k_13_perm_0, x = var_1171)[name = tensor("transpose_19")]; - tensor k_15 = reshape(shape = var_1183, x = k_13)[name = tensor("k_15")]; - tensor var_1185 = const()[name = tensor("op_1185"), val = tensor([2, 4, 6, 64])]; - tensor v_13 = transpose(perm = v_13_perm_0, x = var_1178)[name = tensor("transpose_18")]; - tensor v_15 = reshape(shape = var_1185, x = v_13)[name = tensor("v_15")]; + tensor var_1194 = const()[name = tensor("op_1194"), val = tensor([2, 4, 6, 64])]; + tensor q_13 = transpose(perm = q_13_perm_0, x = var_1177)[name = tensor("transpose_20")]; + tensor q_15 = reshape(shape = var_1194, x = q_13)[name = tensor("q_15")]; + tensor var_1196 = const()[name = tensor("op_1196"), val = tensor([2, 4, 6, 64])]; + tensor k_13 = transpose(perm = k_13_perm_0, x = var_1184)[name = tensor("transpose_19")]; + tensor k_15 = reshape(shape = var_1196, x = k_13)[name = tensor("k_15")]; + tensor var_1198 = const()[name = tensor("op_1198"), val = tensor([2, 4, 6, 64])]; + tensor v_13 = transpose(perm = v_13_perm_0, x = var_1191)[name = tensor("transpose_18")]; + tensor v_15 = reshape(shape = var_1198, x = v_13)[name = tensor("v_15")]; tensor mul_1_y_0 = const()[name = tensor("mul_1_y_0"), val = tensor(0x1p-3)]; tensor mul_1 = mul(x = q_15, y = mul_1_y_0)[name = tensor("mul_1")]; tensor matmul_0_transpose_y_0 = const()[name = tensor("matmul_0_transpose_y_0"), val = tensor(true)]; @@ -1039,30 +1046,30 @@ program(1.0) tensor attn_output_1_transpose_x_0 = const()[name = tensor("attn_output_1_transpose_x_0"), val = tensor(false)]; tensor attn_output_1_transpose_y_0 = const()[name = tensor("attn_output_1_transpose_y_0"), val = tensor(false)]; tensor attn_output_1 = matmul(transpose_x = attn_output_1_transpose_x_0, transpose_y = attn_output_1_transpose_y_0, x = softmax_0, y = v_15)[name = tensor("attn_output_1")]; - tensor var_1188 = const()[name = tensor("op_1188"), val = tensor([2, 0, 1, 3])]; - tensor var_1193 = const()[name = tensor("op_1193"), val = tensor([12, 256])]; - tensor var_1189 = transpose(perm = var_1188, x = attn_output_1)[name = tensor("transpose_17")]; - tensor attn_output_3 = reshape(shape = var_1193, x = var_1189)[name = tensor("attn_output_3")]; - tensor attn_output_5 = linear(bias = decoder_self_attn2_0_out_proj_bias, weight = decoder_self_attn2_0_out_proj_weight, x = attn_output_3)[name = tensor("linear_44")]; - tensor var_1197 = const()[name = tensor("op_1197"), val = tensor([6, 2, 256])]; - tensor attn_output_7 = reshape(shape = var_1197, x = attn_output_5)[name = tensor("attn_output_7")]; + tensor var_1201 = const()[name = tensor("op_1201"), val = tensor([2, 0, 1, 3])]; + tensor var_1206 = const()[name = tensor("op_1206"), val = tensor([12, 256])]; + tensor var_1202 = transpose(perm = var_1201, x = attn_output_1)[name = tensor("transpose_17")]; + tensor attn_output_3 = reshape(shape = var_1206, x = var_1202)[name = tensor("attn_output_3")]; + tensor attn_output_5 = linear(bias = inner_decoder_self_attn2_0_out_proj_bias, weight = inner_decoder_self_attn2_0_out_proj_weight, x = attn_output_3)[name = tensor("linear_44")]; + tensor var_1210 = const()[name = tensor("op_1210"), val = tensor([6, 2, 256])]; + tensor attn_output_7 = reshape(shape = var_1210, x = attn_output_5)[name = tensor("attn_output_7")]; tensor sa2_out_1_perm_0 = const()[name = tensor("sa2_out_1_perm_0"), val = tensor([1, 0, 2])]; tensor sa2_out_1 = transpose(perm = sa2_out_1_perm_0, x = attn_output_7)[name = tensor("transpose_16")]; - tensor input_175 = add(x = query_1, y = sa2_out_1)[name = tensor("input_175")]; - tensor input_177_axes_0 = const()[name = tensor("input_177_axes_0"), val = tensor([-1])]; - tensor input_177 = layer_norm(axes = input_177_axes_0, beta = decoder_norm21_0_bias, epsilon = var_953, gamma = decoder_norm21_0_weight, x = input_175)[name = tensor("input_177")]; - tensor input_179 = linear(bias = decoder_linear1_0_bias, weight = decoder_linear1_0_weight, x = input_177)[name = tensor("linear_45")]; - tensor input_181 = relu(x = input_179)[name = tensor("input_181")]; - tensor ffn_out_1 = linear(bias = decoder_linear2_0_bias, weight = decoder_linear2_0_weight, x = input_181)[name = tensor("linear_46")]; - tensor input_183 = add(x = input_177, y = ffn_out_1)[name = tensor("input_183")]; + tensor input_177 = add(x = query_1, y = sa2_out_1)[name = tensor("input_177")]; + tensor input_179_axes_0 = const()[name = tensor("input_179_axes_0"), val = tensor([-1])]; + tensor input_179 = layer_norm(axes = input_179_axes_0, beta = inner_decoder_norm21_0_bias, epsilon = var_45, gamma = inner_decoder_norm21_0_weight, x = input_177)[name = tensor("input_179")]; + tensor input_181 = linear(bias = inner_decoder_linear1_0_bias, weight = inner_decoder_linear1_0_weight, x = input_179)[name = tensor("linear_45")]; + tensor input_183 = relu(x = input_181)[name = tensor("input_183")]; + tensor ffn_out_1 = linear(bias = inner_decoder_linear2_0_bias, weight = inner_decoder_linear2_0_weight, x = input_183)[name = tensor("linear_46")]; + tensor input_185 = add(x = input_179, y = ffn_out_1)[name = tensor("input_185")]; tensor xt_3_axes_0 = const()[name = tensor("xt_3_axes_0"), val = tensor([-1])]; - tensor xt_3 = layer_norm(axes = xt_3_axes_0, beta = decoder_norm22_0_bias, epsilon = var_953, gamma = decoder_norm22_0_weight, x = input_183)[name = tensor("xt_3")]; - tensor var_1217 = const()[name = tensor("op_1217"), val = tensor([1, 2, 6, 256])]; - tensor x_31 = reshape(shape = var_1217, x = xt_3)[name = tensor("x_31")]; - tensor var_1219_perm_0 = const()[name = tensor("op_1219_perm_0"), val = tensor([0, 2, 1, 3])]; - tensor var_1223 = const()[name = tensor("op_1223"), val = tensor([6, 2, 256])]; - tensor var_1219 = transpose(perm = var_1219_perm_0, x = x_31)[name = tensor("transpose_15")]; - tensor x = reshape(shape = var_1223, x = var_1219)[name = tensor("x")]; + tensor xt_3 = layer_norm(axes = xt_3_axes_0, beta = inner_decoder_norm22_0_bias, epsilon = var_45, gamma = inner_decoder_norm22_0_weight, x = input_185)[name = tensor("xt_3")]; + tensor var_1230 = const()[name = tensor("op_1230"), val = tensor([1, 2, 6, 256])]; + tensor x_31 = reshape(shape = var_1230, x = xt_3)[name = tensor("x_31")]; + tensor var_1232_perm_0 = const()[name = tensor("op_1232_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_1236 = const()[name = tensor("op_1236"), val = tensor([6, 2, 256])]; + tensor var_1232 = transpose(perm = var_1232_perm_0, x = x_31)[name = tensor("transpose_15")]; + tensor x = reshape(shape = var_1236, x = var_1232)[name = tensor("x")]; tensor prev_kv_begin_0 = const()[name = tensor("prev_kv_begin_0"), val = tensor([1, 0, 0, 0, 0])]; tensor prev_kv_end_0 = const()[name = tensor("prev_kv_end_0"), val = tensor([2, 6, 4, 64, 64])]; tensor prev_kv_end_mask_0 = const()[name = tensor("prev_kv_end_mask_0"), val = tensor([false, true, true, true, true])]; @@ -1073,120 +1080,120 @@ program(1.0) tensor prev_scale_end_mask_0 = const()[name = tensor("prev_scale_end_mask_0"), val = tensor([false, true])]; tensor prev_scale_squeeze_mask_0 = const()[name = tensor("prev_scale_squeeze_mask_0"), val = tensor([true, false])]; tensor prev_scale = slice_by_index(begin = prev_scale_begin_0, end = prev_scale_end_0, end_mask = prev_scale_end_mask_0, squeeze_mask = prev_scale_squeeze_mask_0, x = dec_scale)[name = tensor("prev_scale")]; - tensor var_1231 = linear(bias = decoder_q_proj_1_bias, weight = decoder_q_proj_1_weight, x = x)[name = tensor("linear_47")]; - tensor var_1232 = const()[name = tensor("op_1232"), val = tensor([6, 2, 4, 64])]; - tensor var_1233 = reshape(shape = var_1232, x = var_1231)[name = tensor("op_1233")]; + tensor var_1244 = linear(bias = inner_decoder_q_proj_1_bias, weight = inner_decoder_q_proj_1_weight, x = x)[name = tensor("linear_47")]; + tensor var_1245 = const()[name = tensor("op_1245"), val = tensor([6, 2, 4, 64])]; + tensor var_1246 = reshape(shape = var_1245, x = var_1244)[name = tensor("op_1246")]; tensor q_17_perm_0 = const()[name = tensor("q_17_perm_0"), val = tensor([0, 2, 1, 3])]; - tensor var_1237 = linear(bias = decoder_k_proj_1_bias, weight = decoder_k_proj_1_weight, x = x)[name = tensor("linear_48")]; - tensor var_1238 = const()[name = tensor("op_1238"), val = tensor(0x1p-3)]; - tensor var_1239 = mul(x = var_1237, y = var_1238)[name = tensor("op_1239")]; - tensor var_1240 = const()[name = tensor("op_1240"), val = tensor([6, 2, 4, 64])]; - tensor var_1241 = reshape(shape = var_1240, x = var_1239)[name = tensor("op_1241")]; + tensor var_1250 = linear(bias = inner_decoder_k_proj_1_bias, weight = inner_decoder_k_proj_1_weight, x = x)[name = tensor("linear_48")]; + tensor var_1251 = const()[name = tensor("op_1251"), val = tensor(0x1p-3)]; + tensor var_1252 = mul(x = var_1250, y = var_1251)[name = tensor("op_1252")]; + tensor var_1253 = const()[name = tensor("op_1253"), val = tensor([6, 2, 4, 64])]; + tensor var_1254 = reshape(shape = var_1253, x = var_1252)[name = tensor("op_1254")]; tensor k_17_perm_0 = const()[name = tensor("k_17_perm_0"), val = tensor([0, 2, 1, 3])]; - tensor var_1245 = linear(bias = decoder_v_proj_1_bias, weight = decoder_v_proj_1_weight, x = x)[name = tensor("linear_49")]; - tensor var_1246 = const()[name = tensor("op_1246"), val = tensor([6, 2, 4, 64])]; - tensor var_1247 = reshape(shape = var_1246, x = var_1245)[name = tensor("op_1247")]; + tensor var_1258 = linear(bias = inner_decoder_v_proj_1_bias, weight = inner_decoder_v_proj_1_weight, x = x)[name = tensor("linear_49")]; + tensor var_1259 = const()[name = tensor("op_1259"), val = tensor([6, 2, 4, 64])]; + tensor var_1260 = reshape(shape = var_1259, x = var_1258)[name = tensor("op_1260")]; tensor v_17_perm_0 = const()[name = tensor("v_17_perm_0"), val = tensor([0, 2, 1, 3])]; - tensor input_187 = linear(bias = decoder_g_proj_1_bias, weight = decoder_g_proj_1_weight, x = x)[name = tensor("linear_50")]; + tensor input_189 = linear(bias = inner_decoder_g_proj_1_bias, weight = inner_decoder_g_proj_1_weight, x = x)[name = tensor("linear_50")]; tensor sqrt_s0 = sqrt(x = prev_scale)[name = tensor("sqrt_s0")]; tensor s_t = add(x = prev_scale, y = cumsum_mask_1)[name = tensor("s_t")]; tensor const_32 = const()[name = tensor("const_32"), val = tensor(0x1.fffffep+127)]; - tensor clip_3 = clip(alpha = var_958, beta = const_32, x = s_t)[name = tensor("clip_3")]; + tensor clip_3 = clip(alpha = var_38, beta = const_32, x = s_t)[name = tensor("clip_3")]; tensor sqrt_s_t = sqrt(x = clip_3)[name = tensor("sqrt_s_t")]; tensor qk_transpose_x_1 = const()[name = tensor("qk_transpose_x_1"), val = tensor(false)]; tensor qk_transpose_y_1 = const()[name = tensor("qk_transpose_y_1"), val = tensor(true)]; - tensor k_17 = transpose(perm = k_17_perm_0, x = var_1241)[name = tensor("transpose_13")]; - tensor q_17 = transpose(perm = q_17_perm_0, x = var_1233)[name = tensor("transpose_14")]; + tensor k_17 = transpose(perm = k_17_perm_0, x = var_1254)[name = tensor("transpose_13")]; + tensor q_17 = transpose(perm = q_17_perm_0, x = var_1246)[name = tensor("transpose_14")]; tensor qk = matmul(transpose_x = qk_transpose_x_1, transpose_y = qk_transpose_y_1, x = q_17, y = k_17)[name = tensor("qk")]; - tensor var_1262 = const()[name = tensor("op_1262"), val = tensor([2, 1])]; - tensor var_1263 = reshape(shape = var_1262, x = sqrt_s_t)[name = tensor("op_1263")]; - tensor M = real_div(x = causal_with_valid_1, y = var_1263)[name = tensor("M")]; - tensor var_1265 = mul(x = qk, y = M)[name = tensor("op_1265")]; - tensor inner_transpose_x_0 = const()[name = tensor("inner_transpose_x_0"), val = tensor(false)]; - tensor inner_transpose_y_0 = const()[name = tensor("inner_transpose_y_0"), val = tensor(false)]; - tensor v_17 = transpose(perm = v_17_perm_0, x = var_1247)[name = tensor("transpose_12")]; - tensor inner = matmul(transpose_x = inner_transpose_x_0, transpose_y = inner_transpose_y_0, x = var_1265, y = v_17)[name = tensor("inner")]; - tensor var_1267_transpose_x_0 = const()[name = tensor("op_1267_transpose_x_0"), val = tensor(false)]; - tensor var_1267_transpose_y_0 = const()[name = tensor("op_1267_transpose_y_0"), val = tensor(false)]; - tensor var_1267 = matmul(transpose_x = var_1267_transpose_x_0, transpose_y = var_1267_transpose_y_0, x = q_17, y = prev_kv)[name = tensor("op_1267")]; - tensor var_1268 = real_div(x = sqrt_s0, y = sqrt_s_t)[name = tensor("op_1268")]; - tensor var_1269 = const()[name = tensor("op_1269"), val = tensor([1, 1, 2, 1])]; - tensor var_1270 = reshape(shape = var_1269, x = var_1268)[name = tensor("op_1270")]; - tensor cross = mul(x = var_1267, y = var_1270)[name = tensor("cross")]; - tensor out_31 = add(x = inner, y = cross)[name = tensor("out_31")]; - tensor v_masked = mul(x = v_17, y = var_1093)[name = tensor("v_masked")]; - tensor var_1276 = mul(x = prev_kv, y = sqrt_s0)[name = tensor("op_1276")]; - tensor var_1278_transpose_x_1 = const()[name = tensor("op_1278_transpose_x_1"), val = tensor(true)]; - tensor var_1278_transpose_y_1 = const()[name = tensor("op_1278_transpose_y_1"), val = tensor(false)]; - tensor var_1278 = matmul(transpose_x = var_1278_transpose_x_1, transpose_y = var_1278_transpose_y_1, x = k_17, y = v_masked)[name = tensor("op_1278")]; - tensor new_kv_unnorm = add(x = var_1276, y = var_1278)[name = tensor("new_kv_unnorm")]; - tensor new_scale = add(x = prev_scale, y = var_1101)[name = tensor("new_scale")]; + tensor var_1275 = const()[name = tensor("op_1275"), val = tensor([2, 1])]; + tensor var_1276 = reshape(shape = var_1275, x = sqrt_s_t)[name = tensor("op_1276")]; + tensor M = real_div(x = causal_with_valid_1, y = var_1276)[name = tensor("M")]; + tensor var_1278 = mul(x = qk, y = M)[name = tensor("op_1278")]; + tensor inner_11_transpose_x_0 = const()[name = tensor("inner_11_transpose_x_0"), val = tensor(false)]; + tensor inner_11_transpose_y_0 = const()[name = tensor("inner_11_transpose_y_0"), val = tensor(false)]; + tensor v_17 = transpose(perm = v_17_perm_0, x = var_1260)[name = tensor("transpose_12")]; + tensor inner_11 = matmul(transpose_x = inner_11_transpose_x_0, transpose_y = inner_11_transpose_y_0, x = var_1278, y = v_17)[name = tensor("inner_11")]; + tensor var_1280_transpose_x_0 = const()[name = tensor("op_1280_transpose_x_0"), val = tensor(false)]; + tensor var_1280_transpose_y_0 = const()[name = tensor("op_1280_transpose_y_0"), val = tensor(false)]; + tensor var_1280 = matmul(transpose_x = var_1280_transpose_x_0, transpose_y = var_1280_transpose_y_0, x = q_17, y = prev_kv)[name = tensor("op_1280")]; + tensor var_1281 = real_div(x = sqrt_s0, y = sqrt_s_t)[name = tensor("op_1281")]; + tensor var_1282 = const()[name = tensor("op_1282"), val = tensor([1, 1, 2, 1])]; + tensor var_1283 = reshape(shape = var_1282, x = var_1281)[name = tensor("op_1283")]; + tensor cross = mul(x = var_1280, y = var_1283)[name = tensor("cross")]; + tensor out_31 = add(x = inner_11, y = cross)[name = tensor("out_31")]; + tensor v_masked = mul(x = v_17, y = var_1106)[name = tensor("v_masked")]; + tensor var_1289 = mul(x = prev_kv, y = sqrt_s0)[name = tensor("op_1289")]; + tensor var_1291_transpose_x_1 = const()[name = tensor("op_1291_transpose_x_1"), val = tensor(true)]; + tensor var_1291_transpose_y_1 = const()[name = tensor("op_1291_transpose_y_1"), val = tensor(false)]; + tensor var_1291 = matmul(transpose_x = var_1291_transpose_x_1, transpose_y = var_1291_transpose_y_1, x = k_17, y = v_masked)[name = tensor("op_1291")]; + tensor new_kv_unnorm = add(x = var_1289, y = var_1291)[name = tensor("new_kv_unnorm")]; + tensor new_scale = add(x = prev_scale, y = var_1114)[name = tensor("new_scale")]; tensor const_33 = const()[name = tensor("const_33"), val = tensor(0x1.fffffep+127)]; - tensor clip_4 = clip(alpha = var_958, beta = const_33, x = new_scale)[name = tensor("clip_4")]; + tensor clip_4 = clip(alpha = var_38, beta = const_33, x = new_scale)[name = tensor("clip_4")]; tensor sqrt_new_scale = sqrt(x = clip_4)[name = tensor("sqrt_new_scale")]; tensor nkv = real_div(x = new_kv_unnorm, y = sqrt_new_scale)[name = tensor("nkv")]; - tensor var_1287_perm_0 = const()[name = tensor("op_1287_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_1300_perm_0 = const()[name = tensor("op_1300_perm_0"), val = tensor([0, 2, 1, 3])]; tensor out_33_axes_0 = const()[name = tensor("out_33_axes_0"), val = tensor([-1])]; - tensor var_1287 = transpose(perm = var_1287_perm_0, x = out_31)[name = tensor("transpose_11")]; - tensor out_33 = layer_norm(axes = out_33_axes_0, epsilon = var_955, x = var_1287)[name = tensor("out_33")]; - tensor var_1291 = const()[name = tensor("op_1291"), val = tensor([6, 2, 256])]; - tensor out = reshape(shape = var_1291, x = out_33)[name = tensor("out")]; - tensor var_1293 = silu(x = input_187)[name = tensor("op_1293")]; - tensor input_189 = mul(x = var_1293, y = out)[name = tensor("input_189")]; - tensor ret_out = linear(bias = decoder_out_proj_1_bias, weight = decoder_out_proj_1_weight, x = input_189)[name = tensor("linear_51")]; - tensor input_191 = add(x = x, y = ret_out)[name = tensor("input_191")]; + tensor var_1300 = transpose(perm = var_1300_perm_0, x = out_31)[name = tensor("transpose_11")]; + tensor out_33 = layer_norm(axes = out_33_axes_0, epsilon = var_53, x = var_1300)[name = tensor("out_33")]; + tensor var_1304 = const()[name = tensor("op_1304"), val = tensor([6, 2, 256])]; + tensor out = reshape(shape = var_1304, x = out_33)[name = tensor("out")]; + tensor var_1306 = silu(x = input_189)[name = tensor("op_1306")]; + tensor input_191 = mul(x = var_1306, y = out)[name = tensor("input_191")]; + tensor ret_out = linear(bias = inner_decoder_out_proj_1_bias, weight = inner_decoder_out_proj_1_weight, x = input_191)[name = tensor("linear_51")]; + tensor input_193 = add(x = x, y = ret_out)[name = tensor("input_193")]; tensor xt_5_axes_0 = const()[name = tensor("xt_5_axes_0"), val = tensor([-1])]; - tensor xt_5 = layer_norm(axes = xt_5_axes_0, beta = decoder_norm11_1_bias, epsilon = var_953, gamma = decoder_norm11_1_weight, x = input_191)[name = tensor("xt_5")]; - tensor var_1303 = const()[name = tensor("op_1303"), val = tensor([1, 6, 2, 256])]; - tensor var_1304 = reshape(shape = var_1303, x = xt_5)[name = tensor("op_1304")]; - tensor var_1305_perm_0 = const()[name = tensor("op_1305_perm_0"), val = tensor([0, 2, 1, 3])]; - tensor var_1308 = const()[name = tensor("op_1308"), val = tensor([2, 6, 256])]; - tensor var_1305 = transpose(perm = var_1305_perm_0, x = var_1304)[name = tensor("transpose_10")]; - tensor query_5 = reshape(shape = var_1308, x = var_1305)[name = tensor("query_5")]; + tensor xt_5 = layer_norm(axes = xt_5_axes_0, beta = inner_decoder_norm11_1_bias, epsilon = var_45, gamma = inner_decoder_norm11_1_weight, x = input_193)[name = tensor("xt_5")]; + tensor var_1316 = const()[name = tensor("op_1316"), val = tensor([1, 6, 2, 256])]; + tensor var_1317 = reshape(shape = var_1316, x = xt_5)[name = tensor("op_1317")]; + tensor var_1318_perm_0 = const()[name = tensor("op_1318_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_1321 = const()[name = tensor("op_1321"), val = tensor([2, 6, 256])]; + tensor var_1318 = transpose(perm = var_1318_perm_0, x = var_1317)[name = tensor("transpose_10")]; + tensor query_5 = reshape(shape = var_1321, x = var_1318)[name = tensor("query_5")]; tensor query_perm_0 = const()[name = tensor("query_perm_0"), val = tensor([1, 0, 2])]; tensor query = transpose(perm = query_perm_0, x = query_5)[name = tensor("transpose_9")]; - tensor var_1331 = linear(bias = decoder_self_attn2_1_in_proj_bias, weight = decoder_self_attn2_1_in_proj_weight, x = query)[name = tensor("linear_52")]; + tensor var_1344 = linear(bias = inner_decoder_self_attn2_1_in_proj_bias, weight = inner_decoder_self_attn2_1_in_proj_weight, x = query)[name = tensor("linear_52")]; tensor concat_2 = const()[name = tensor("concat_2"), val = tensor([6, 2, 3, 256])]; - tensor var_1333 = reshape(shape = concat_2, x = var_1331)[name = tensor("op_1333")]; - tensor var_1334_axes_0 = const()[name = tensor("op_1334_axes_0"), val = tensor([0])]; - tensor var_1334 = expand_dims(axes = var_1334_axes_0, x = var_1333)[name = tensor("op_1334")]; - tensor var_1335_perm_0 = const()[name = tensor("op_1335_perm_0"), val = tensor([-2, 1, 2, 0, 4])]; - tensor var_1336_axes_0 = const()[name = tensor("op_1336_axes_0"), val = tensor([-2])]; - tensor var_1335 = transpose(perm = var_1335_perm_0, x = var_1334)[name = tensor("transpose_8")]; - tensor var_1336 = squeeze(axes = var_1336_axes_0, x = var_1335)[name = tensor("op_1336")]; + tensor var_1346 = reshape(shape = concat_2, x = var_1344)[name = tensor("op_1346")]; + tensor var_1347_axes_0 = const()[name = tensor("op_1347_axes_0"), val = tensor([0])]; + tensor var_1347 = expand_dims(axes = var_1347_axes_0, x = var_1346)[name = tensor("op_1347")]; + tensor var_1348_perm_0 = const()[name = tensor("op_1348_perm_0"), val = tensor([-2, 1, 2, 0, 4])]; + tensor var_1349_axes_0 = const()[name = tensor("op_1349_axes_0"), val = tensor([-2])]; + tensor var_1348 = transpose(perm = var_1348_perm_0, x = var_1347)[name = tensor("transpose_8")]; + tensor var_1349 = squeeze(axes = var_1349_axes_0, x = var_1348)[name = tensor("op_1349")]; tensor q_19_begin_0 = const()[name = tensor("q_19_begin_0"), val = tensor([0, 0, 0, 0])]; tensor q_19_end_0 = const()[name = tensor("q_19_end_0"), val = tensor([1, 6, 2, 256])]; tensor q_19_end_mask_0 = const()[name = tensor("q_19_end_mask_0"), val = tensor([false, true, true, true])]; tensor q_19_squeeze_mask_0 = const()[name = tensor("q_19_squeeze_mask_0"), val = tensor([true, false, false, false])]; - tensor q_19 = slice_by_index(begin = q_19_begin_0, end = q_19_end_0, end_mask = q_19_end_mask_0, squeeze_mask = q_19_squeeze_mask_0, x = var_1336)[name = tensor("q_19")]; + tensor q_19 = slice_by_index(begin = q_19_begin_0, end = q_19_end_0, end_mask = q_19_end_mask_0, squeeze_mask = q_19_squeeze_mask_0, x = var_1349)[name = tensor("q_19")]; tensor k_19_begin_0 = const()[name = tensor("k_19_begin_0"), val = tensor([1, 0, 0, 0])]; tensor k_19_end_0 = const()[name = tensor("k_19_end_0"), val = tensor([2, 6, 2, 256])]; tensor k_19_end_mask_0 = const()[name = tensor("k_19_end_mask_0"), val = tensor([false, true, true, true])]; tensor k_19_squeeze_mask_0 = const()[name = tensor("k_19_squeeze_mask_0"), val = tensor([true, false, false, false])]; - tensor k_19 = slice_by_index(begin = k_19_begin_0, end = k_19_end_0, end_mask = k_19_end_mask_0, squeeze_mask = k_19_squeeze_mask_0, x = var_1336)[name = tensor("k_19")]; + tensor k_19 = slice_by_index(begin = k_19_begin_0, end = k_19_end_0, end_mask = k_19_end_mask_0, squeeze_mask = k_19_squeeze_mask_0, x = var_1349)[name = tensor("k_19")]; tensor v_19_begin_0 = const()[name = tensor("v_19_begin_0"), val = tensor([2, 0, 0, 0])]; tensor v_19_end_0 = const()[name = tensor("v_19_end_0"), val = tensor([3, 6, 2, 256])]; tensor v_19_end_mask_0 = const()[name = tensor("v_19_end_mask_0"), val = tensor([false, true, true, true])]; tensor v_19_squeeze_mask_0 = const()[name = tensor("v_19_squeeze_mask_0"), val = tensor([true, false, false, false])]; - tensor v_19 = slice_by_index(begin = v_19_begin_0, end = v_19_end_0, end_mask = v_19_end_mask_0, squeeze_mask = v_19_squeeze_mask_0, x = var_1336)[name = tensor("v_19")]; - tensor var_1344 = const()[name = tensor("op_1344"), val = tensor([6, 8, 64])]; - tensor var_1345 = reshape(shape = var_1344, x = q_19)[name = tensor("op_1345")]; + tensor v_19 = slice_by_index(begin = v_19_begin_0, end = v_19_end_0, end_mask = v_19_end_mask_0, squeeze_mask = v_19_squeeze_mask_0, x = var_1349)[name = tensor("v_19")]; + tensor var_1357 = const()[name = tensor("op_1357"), val = tensor([6, 8, 64])]; + tensor var_1358 = reshape(shape = var_1357, x = q_19)[name = tensor("op_1358")]; tensor q_21_perm_0 = const()[name = tensor("q_21_perm_0"), val = tensor([1, 0, 2])]; - tensor var_1351 = const()[name = tensor("op_1351"), val = tensor([6, 8, 64])]; - tensor var_1352 = reshape(shape = var_1351, x = k_19)[name = tensor("op_1352")]; + tensor var_1364 = const()[name = tensor("op_1364"), val = tensor([6, 8, 64])]; + tensor var_1365 = reshape(shape = var_1364, x = k_19)[name = tensor("op_1365")]; tensor k_21_perm_0 = const()[name = tensor("k_21_perm_0"), val = tensor([1, 0, 2])]; - tensor var_1358 = const()[name = tensor("op_1358"), val = tensor([6, 8, 64])]; - tensor var_1359 = reshape(shape = var_1358, x = v_19)[name = tensor("op_1359")]; + tensor var_1371 = const()[name = tensor("op_1371"), val = tensor([6, 8, 64])]; + tensor var_1372 = reshape(shape = var_1371, x = v_19)[name = tensor("op_1372")]; tensor v_21_perm_0 = const()[name = tensor("v_21_perm_0"), val = tensor([1, 0, 2])]; - tensor var_1362 = const()[name = tensor("op_1362"), val = tensor([2, 4, 6, 64])]; - tensor q_21 = transpose(perm = q_21_perm_0, x = var_1345)[name = tensor("transpose_7")]; - tensor q = reshape(shape = var_1362, x = q_21)[name = tensor("q")]; - tensor var_1364 = const()[name = tensor("op_1364"), val = tensor([2, 4, 6, 64])]; - tensor k_21 = transpose(perm = k_21_perm_0, x = var_1352)[name = tensor("transpose_6")]; - tensor k = reshape(shape = var_1364, x = k_21)[name = tensor("k")]; - tensor var_1366 = const()[name = tensor("op_1366"), val = tensor([2, 4, 6, 64])]; - tensor v_21 = transpose(perm = v_21_perm_0, x = var_1359)[name = tensor("transpose_5")]; - tensor v = reshape(shape = var_1366, x = v_21)[name = tensor("v")]; + tensor var_1375 = const()[name = tensor("op_1375"), val = tensor([2, 4, 6, 64])]; + tensor q_21 = transpose(perm = q_21_perm_0, x = var_1358)[name = tensor("transpose_7")]; + tensor q = reshape(shape = var_1375, x = q_21)[name = tensor("q")]; + tensor var_1377 = const()[name = tensor("op_1377"), val = tensor([2, 4, 6, 64])]; + tensor k_21 = transpose(perm = k_21_perm_0, x = var_1365)[name = tensor("transpose_6")]; + tensor k = reshape(shape = var_1377, x = k_21)[name = tensor("k")]; + tensor var_1379 = const()[name = tensor("op_1379"), val = tensor([2, 4, 6, 64])]; + tensor v_21 = transpose(perm = v_21_perm_0, x = var_1372)[name = tensor("transpose_5")]; + tensor v = reshape(shape = var_1379, x = v_21)[name = tensor("v")]; tensor mul_3_y_0 = const()[name = tensor("mul_3_y_0"), val = tensor(0x1p-3)]; tensor mul_3 = mul(x = q, y = mul_3_y_0)[name = tensor("mul_3")]; tensor matmul_1_transpose_y_0 = const()[name = tensor("matmul_1_transpose_y_0"), val = tensor(true)]; @@ -1197,36 +1204,36 @@ program(1.0) tensor attn_output_9_transpose_x_0 = const()[name = tensor("attn_output_9_transpose_x_0"), val = tensor(false)]; tensor attn_output_9_transpose_y_0 = const()[name = tensor("attn_output_9_transpose_y_0"), val = tensor(false)]; tensor attn_output_9 = matmul(transpose_x = attn_output_9_transpose_x_0, transpose_y = attn_output_9_transpose_y_0, x = softmax_1, y = v)[name = tensor("attn_output_9")]; - tensor var_1369 = const()[name = tensor("op_1369"), val = tensor([2, 0, 1, 3])]; - tensor var_1374 = const()[name = tensor("op_1374"), val = tensor([12, 256])]; - tensor var_1370 = transpose(perm = var_1369, x = attn_output_9)[name = tensor("transpose_4")]; - tensor attn_output_11 = reshape(shape = var_1374, x = var_1370)[name = tensor("attn_output_11")]; - tensor attn_output_13 = linear(bias = decoder_self_attn2_1_out_proj_bias, weight = decoder_self_attn2_1_out_proj_weight, x = attn_output_11)[name = tensor("linear_53")]; - tensor var_1378 = const()[name = tensor("op_1378"), val = tensor([6, 2, 256])]; - tensor attn_output = reshape(shape = var_1378, x = attn_output_13)[name = tensor("attn_output")]; + tensor var_1382 = const()[name = tensor("op_1382"), val = tensor([2, 0, 1, 3])]; + tensor var_1387 = const()[name = tensor("op_1387"), val = tensor([12, 256])]; + tensor var_1383 = transpose(perm = var_1382, x = attn_output_9)[name = tensor("transpose_4")]; + tensor attn_output_11 = reshape(shape = var_1387, x = var_1383)[name = tensor("attn_output_11")]; + tensor attn_output_13 = linear(bias = inner_decoder_self_attn2_1_out_proj_bias, weight = inner_decoder_self_attn2_1_out_proj_weight, x = attn_output_11)[name = tensor("linear_53")]; + tensor var_1391 = const()[name = tensor("op_1391"), val = tensor([6, 2, 256])]; + tensor attn_output = reshape(shape = var_1391, x = attn_output_13)[name = tensor("attn_output")]; tensor sa2_out_perm_0 = const()[name = tensor("sa2_out_perm_0"), val = tensor([1, 0, 2])]; tensor sa2_out = transpose(perm = sa2_out_perm_0, x = attn_output)[name = tensor("transpose_3")]; - tensor input_193 = add(x = query_5, y = sa2_out)[name = tensor("input_193")]; - tensor input_195_axes_0 = const()[name = tensor("input_195_axes_0"), val = tensor([-1])]; - tensor input_195 = layer_norm(axes = input_195_axes_0, beta = decoder_norm21_1_bias, epsilon = var_953, gamma = decoder_norm21_1_weight, x = input_193)[name = tensor("input_195")]; - tensor input_197 = linear(bias = decoder_linear1_1_bias, weight = decoder_linear1_1_weight, x = input_195)[name = tensor("linear_54")]; - tensor input_199 = relu(x = input_197)[name = tensor("input_199")]; - tensor ffn_out = linear(bias = decoder_linear2_1_bias, weight = decoder_linear2_1_weight, x = input_199)[name = tensor("linear_55")]; - tensor input_201 = add(x = input_195, y = ffn_out)[name = tensor("input_201")]; + tensor input_195 = add(x = query_5, y = sa2_out)[name = tensor("input_195")]; + tensor input_197_axes_0 = const()[name = tensor("input_197_axes_0"), val = tensor([-1])]; + tensor input_197 = layer_norm(axes = input_197_axes_0, beta = inner_decoder_norm21_1_bias, epsilon = var_45, gamma = inner_decoder_norm21_1_weight, x = input_195)[name = tensor("input_197")]; + tensor input_199 = linear(bias = inner_decoder_linear1_1_bias, weight = inner_decoder_linear1_1_weight, x = input_197)[name = tensor("linear_54")]; + tensor input_201 = relu(x = input_199)[name = tensor("input_201")]; + tensor ffn_out = linear(bias = inner_decoder_linear2_1_bias, weight = inner_decoder_linear2_1_weight, x = input_201)[name = tensor("linear_55")]; + tensor input_203 = add(x = input_197, y = ffn_out)[name = tensor("input_203")]; tensor xt_axes_0 = const()[name = tensor("xt_axes_0"), val = tensor([-1])]; - tensor xt = layer_norm(axes = xt_axes_0, beta = decoder_norm22_1_bias, epsilon = var_953, gamma = decoder_norm22_1_weight, x = input_201)[name = tensor("xt")]; - tensor var_1398 = const()[name = tensor("op_1398"), val = tensor([1, 2, 6, 256])]; - tensor input = reshape(shape = var_1398, x = xt)[name = tensor("input")]; - tensor var_1400 = const()[name = tensor("op_1400"), val = tensor([-1])]; - tensor var_1401 = reduce_l2_norm(axes = var_1400, keep_dims = var_956, x = input)[name = tensor("op_1401")]; + tensor xt = layer_norm(axes = xt_axes_0, beta = inner_decoder_norm22_1_bias, epsilon = var_45, gamma = inner_decoder_norm22_1_weight, x = input_203)[name = tensor("xt")]; + tensor var_1411 = const()[name = tensor("op_1411"), val = tensor([1, 2, 6, 256])]; + tensor input = reshape(shape = var_1411, x = xt)[name = tensor("input")]; + tensor var_1413 = const()[name = tensor("op_1413"), val = tensor([-1])]; + tensor var_1414 = reduce_l2_norm(axes = var_1413, keep_dims = var_44, x = input)[name = tensor("op_1414")]; tensor const_42 = const()[name = tensor("const_42"), val = tensor(0x1.fffffep+127)]; - tensor clip_5 = clip(alpha = var_948, beta = const_42, x = var_1401)[name = tensor("clip_5")]; - tensor var_1403 = real_div(x = input, y = clip_5)[name = tensor("op_1403")]; + tensor clip_5 = clip(alpha = var_58, beta = const_42, x = var_1414)[name = tensor("clip_5")]; + tensor var_1416 = real_div(x = input, y = clip_5)[name = tensor("op_1416")]; tensor concat_6 = const()[name = tensor("concat_6"), val = tensor([2, 1, 256])]; tensor reshape_0 = reshape(shape = concat_6, x = emb)[name = tensor("reshape_0")]; tensor transpose_1_perm_0 = const()[name = tensor("transpose_1_perm_0"), val = tensor([0, 1, 3, 2])]; tensor concat_7 = const()[name = tensor("concat_7"), val = tensor([2, 256, 6])]; - tensor transpose_1 = transpose(perm = transpose_1_perm_0, x = var_1403)[name = tensor("transpose_2")]; + tensor transpose_1 = transpose(perm = transpose_1_perm_0, x = var_1416)[name = tensor("transpose_2")]; tensor reshape_1 = reshape(shape = concat_7, x = transpose_1)[name = tensor("reshape_1")]; tensor matmul_2_transpose_x_0 = const()[name = tensor("matmul_2_transpose_x_0"), val = tensor(false)]; tensor matmul_2_transpose_y_0 = const()[name = tensor("matmul_2_transpose_y_0"), val = tensor(false)]; @@ -1237,10 +1244,10 @@ program(1.0) tensor output_end_0 = const()[name = tensor("output_end_0"), val = tensor([1, 2, 5])]; tensor output_end_mask_0 = const()[name = tensor("output_end_mask_0"), val = tensor([true, true, false])]; tensor output = slice_by_index(begin = output_begin_0, end = output_end_0, end_mask = output_end_mask_0, x = reshape_2)[name = tensor("output")]; - tensor probs = sigmoid(x = output)[name = tensor("op_1407")]; - tensor var_1409_axis_0 = const()[name = tensor("op_1409_axis_0"), val = tensor(0)]; - tensor dec_kv_new = stack(axis = var_1409_axis_0, values = (var_1105, nkv))[name = tensor("op_1409")]; - tensor var_1411_axis_0 = const()[name = tensor("op_1411_axis_0"), val = tensor(0)]; - tensor dec_scale_new = stack(axis = var_1411_axis_0, values = (new_scale_9, new_scale))[name = tensor("op_1411")]; + tensor probs = sigmoid(x = output)[name = tensor("op_1420")]; + tensor var_1422_axis_0 = const()[name = tensor("op_1422_axis_0"), val = tensor(0)]; + tensor dec_kv_new = stack(axis = var_1422_axis_0, values = (var_1118, nkv))[name = tensor("op_1422")]; + tensor var_1424_axis_0 = const()[name = tensor("op_1424_axis_0"), val = tensor(0)]; + tensor dec_scale_new = stack(axis = var_1424_axis_0, values = (new_scale_9, new_scale))[name = tensor("op_1424")]; } -> (probs, enc_kv_new, enc_scale_new, enc_conv_cache_new, cnn_window_new, dec_kv_new, dec_scale_new); } \ No newline at end of file diff --git a/optimized/ami/200ms/ls_eend_ami_200ms.mlpackage/Data/com.apple.CoreML/model.mlmodel b/optimized/ami/200ms/ls_eend_ami_200ms.mlpackage/Data/com.apple.CoreML/model.mlmodel index 4c678f86789d8217225a774ec066ff29f1c6adcc..e7ff7887f06048798363799e4ecb439139f1e774 100644 --- a/optimized/ami/200ms/ls_eend_ami_200ms.mlpackage/Data/com.apple.CoreML/model.mlmodel +++ b/optimized/ami/200ms/ls_eend_ami_200ms.mlpackage/Data/com.apple.CoreML/model.mlmodel @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:286832ef1477a85722d0e808565e32c2edf2c7abbbb9f2b3447253ac30f39155 -size 179858 +oid sha256:7a616be3cfa5c13828e61ddda409e43a20eb8f48c2893530948c7d073a6c87b3 +size 184838 diff --git a/optimized/ami/200ms/ls_eend_ami_200ms.mlpackage/Manifest.json b/optimized/ami/200ms/ls_eend_ami_200ms.mlpackage/Manifest.json index 612bcaf8e0b488f282381fd935e727601819bea6..579432c763882788ec14fcf33d6cef3c9fad713b 100644 --- a/optimized/ami/200ms/ls_eend_ami_200ms.mlpackage/Manifest.json +++ b/optimized/ami/200ms/ls_eend_ami_200ms.mlpackage/Manifest.json @@ -1,18 +1,18 @@ { "fileFormatVersion": "1.0.0", "itemInfoEntries": { - "9AA855BB-15E6-4440-985F-9D01D6A86BD7": { + "1F7F287A-3524-40B3-B470-5E617233CC0A": { "author": "com.apple.CoreML", "description": "CoreML Model Weights", "name": "weights", "path": "com.apple.CoreML/weights" }, - "FE3BD696-8DA2-4220-BA67-6DFD3C1F41A7": { + "D525D189-4337-40E9-852C-1E328106320D": { "author": "com.apple.CoreML", "description": "CoreML Model Specification", "name": "model.mlmodel", "path": "com.apple.CoreML/model.mlmodel" } }, - "rootModelIdentifier": "FE3BD696-8DA2-4220-BA67-6DFD3C1F41A7" + "rootModelIdentifier": "D525D189-4337-40E9-852C-1E328106320D" } diff --git a/optimized/ami/300ms/ls_eend_ami_300ms.mlmodelc/analytics/coremldata.bin b/optimized/ami/300ms/ls_eend_ami_300ms.mlmodelc/analytics/coremldata.bin index 3f6639b57a269384ee54f9f1660607c52c504864..a13d866bed0a243a078e8be8aba58b6889b0eb4e 100644 --- a/optimized/ami/300ms/ls_eend_ami_300ms.mlmodelc/analytics/coremldata.bin +++ b/optimized/ami/300ms/ls_eend_ami_300ms.mlmodelc/analytics/coremldata.bin @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:a6f56029ce7301e97b0cefc55ff6f7d7d8a47604b531cac5a18717b8934c6c12 +oid sha256:1440d924c2808e18a0ec1b1d62ad4cd68275891ba9c6688a69241a12558f5ae5 size 243 diff --git a/optimized/ami/300ms/ls_eend_ami_300ms.mlmodelc/coremldata.bin b/optimized/ami/300ms/ls_eend_ami_300ms.mlmodelc/coremldata.bin index 0adcb3383c298e39175a387aa8c5ed93b77e17d8..0b6e249fee2052fd2894882289efbf14210186c6 100644 --- a/optimized/ami/300ms/ls_eend_ami_300ms.mlmodelc/coremldata.bin +++ b/optimized/ami/300ms/ls_eend_ami_300ms.mlmodelc/coremldata.bin @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:e5272855c6f2aed3afbb9d4479de3da636f7780e628ff7ad02dc5bbc5f5d7894 -size 1292 +oid sha256:44f6bb31efc91684a91067fdc08db6715c7bcac6d9358cca9e7ff435eaa3e4a1 +size 1395 diff --git a/optimized/ami/300ms/ls_eend_ami_300ms.mlmodelc/metadata.json b/optimized/ami/300ms/ls_eend_ami_300ms.mlmodelc/metadata.json index 24752739dc034358ad51ac236bb8ea00423b379e..c0b4784fccc4747491b550cb4a84e30f4973d127 100644 --- a/optimized/ami/300ms/ls_eend_ami_300ms.mlmodelc/metadata.json +++ b/optimized/ami/300ms/ls_eend_ami_300ms.mlmodelc/metadata.json @@ -1,7 +1,7 @@ [ { "metadataOutputVersion" : "3.0", - "shortDescription" : "LS-EEND AMI streaming diarizer (pipeline, T=3, max_speakers=4)", + "shortDescription" : "LS-EEND AMI streaming diarizer (pipeline, T=3, max_speakers=4, layout=raw_mel, cu=all)", "outputSchema" : [ { "hasShapeFlexibility" : "0", @@ -81,7 +81,7 @@ "specificationVersion" : 8, "mlProgramOperationTypeHistogram" : { "Ios16.reduceL2Norm" : 2, - "Ios17.reshape" : 68, + "Ios17.reshape" : 69, "Ios16.softmax" : 2, "Ios17.matmul" : 29, "Ios17.transpose" : 57, @@ -89,7 +89,7 @@ "Ios17.expandDims" : 3, "Ios17.add" : 46, "Ios16.sigmoid" : 5, - "Ios17.sliceByIndex" : 56, + "Ios17.sliceByIndex" : 59, "Tile" : 1, "Ios16.reduceSum" : 1, "Ios17.squeeze" : 2, @@ -101,7 +101,7 @@ "Ios16.silu" : 18, "Ios17.realDiv" : 20, "Ios17.linear" : 56, - "Stack" : 5, + "Stack" : 6, "Ios17.concat" : 18, "Ios16.relu" : 2, "Ios16.cumsum" : 1, @@ -128,9 +128,9 @@ "hasShapeFlexibility" : "0", "isOptional" : "0", "dataType" : "Float32", - "formattedType" : "MultiArray (Float32 1 × 3 × 345)", + "formattedType" : "MultiArray (Float32 1 × 35 × 23)", "shortDescription" : "", - "shape" : "[1, 3, 345]", + "shape" : "[1, 35, 23]", "name" : "features", "type" : "MultiArray" }, @@ -206,8 +206,8 @@ } ], "userDefinedMetadata" : { - "com.github.apple.coremltools.conversion_date" : "2026-04-16", - "config" : "{\"model_name\": \"ami\", \"model_label\": \"AMI\", \"variant\": \"pipeline\", \"chunk_size\": 3, \"step_duration_ms\": 300, \"frame_duration_ms\": 100, \"frame_duration_seconds\": 0.1, \"max_speakers\": 4, \"max_nspks\": 6, \"n_units\": 256, \"n_heads\": 4, \"enc_n_layers\": 4, \"dec_n_layers\": 2, \"conv_kernel_size\": 16, \"conv_delay\": 9, \"sample_rate\": 8000, \"win_length\": 200, \"hop_length\": 80, \"n_mels\": 23, \"context_size\": 7, \"subsampling\": 10, \"feat_type\": \"logmel23_cummn\", \"pure_roll\": true}", + "com.github.apple.coremltools.conversion_date" : "2026-04-18", + "config" : "{\"model_name\": \"ami\", \"model_label\": \"AMI\", \"variant\": \"pipeline\", \"chunk_size\": 3, \"step_duration_ms\": 300, \"frame_duration_ms\": 100, \"frame_duration_seconds\": 0.1, \"max_speakers\": 4, \"max_nspks\": 6, \"n_units\": 256, \"n_heads\": 4, \"enc_n_layers\": 4, \"dec_n_layers\": 2, \"conv_kernel_size\": 16, \"conv_delay\": 9, \"sample_rate\": 8000, \"win_length\": 200, \"hop_length\": 80, \"n_mels\": 23, \"context_size\": 7, \"subsampling\": 10, \"feat_type\": \"logmel23_cummn\", \"pure_roll\": true, \"input_layout\": \"raw_mel\", \"compute_units_export\": \"all\", \"raw_mel_length\": 35}", "com.github.apple.coremltools.source" : "torch==2.6.0", "com.github.apple.coremltools.version" : "9.0", "com.github.apple.coremltools.source_dialect" : "TorchScript" diff --git a/optimized/ami/300ms/ls_eend_ami_300ms.mlmodelc/model.mil b/optimized/ami/300ms/ls_eend_ami_300ms.mlmodelc/model.mil index 970596a9f4827af95d9232b6600321f0fda4a1c0..0b1302f5a5fd1e848cfab7ab40d920e3e8992726 100644 --- a/optimized/ami/300ms/ls_eend_ami_300ms.mlmodelc/model.mil +++ b/optimized/ami/300ms/ls_eend_ami_300ms.mlmodelc/model.mil @@ -1,234 +1,252 @@ program(1.0) [buildInfo = dict, tensor>({{"coremlc-component-MIL", "3510.2.1"}, {"coremlc-version", "3500.32.1"}, {"coremltools-component-torch", "2.6.0"}, {"coremltools-source-dialect", "TorchScript"}, {"coremltools-version", "9.0"}})] { - func main(tensor cnn_window, tensor dec_kv, tensor dec_scale, tensor enc_conv_cache, tensor enc_kv, tensor enc_scale, tensor features, tensor valid_mask) { - tensor encoder_cnn_bias = const()[name = tensor("encoder_cnn_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(64)))]; - tensor encoder_cnn_weight = const()[name = tensor("encoder_cnn_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1152)))]; - tensor encoder__causal_mask = const()[name = tensor("encoder__causal_mask"), val = tensor([[0x1p+0, 0x0p+0, 0x0p+0], [0x1p+0, 0x1p+0, 0x0p+0], [0x1p+0, 0x1p+0, 0x1p+0]])]; - tensor encoder__t_index = const()[name = tensor("encoder__t_index"), val = tensor([0x1p+0, 0x1p+1, 0x1.8p+1])]; - tensor encoder_input_projection_linear_bias = const()[name = tensor("encoder_input_projection_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4981952)))]; - tensor encoder_input_projection_linear_weight = const()[name = tensor("encoder_input_projection_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4983040)))]; - tensor encoder_pre_layer_norm_bias = const()[name = tensor("encoder_pre_layer_norm_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(5336384)))]; - tensor encoder_pre_layer_norm_weight = const()[name = tensor("encoder_pre_layer_norm_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(5337472)))]; - tensor encoder_ffn1_0_module_sequential_0_bias = const()[name = tensor("encoder_ffn1_0_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(5338560)))]; - tensor encoder_ffn1_0_module_sequential_0_weight = const()[name = tensor("encoder_ffn1_0_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(5339648)))]; - tensor encoder_ffn1_0_module_sequential_1_linear_bias = const()[name = tensor("encoder_ffn1_0_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(5340736)))]; - tensor encoder_ffn1_0_module_sequential_1_linear_weight = const()[name = tensor("encoder_ffn1_0_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(5344896)))]; - tensor encoder_ffn1_0_module_sequential_4_linear_bias = const()[name = tensor("encoder_ffn1_0_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(6393536)))]; - tensor encoder_ffn1_0_module_sequential_4_linear_weight = const()[name = tensor("encoder_ffn1_0_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(6394624)))]; - tensor encoder_ret_lns_0_bias = const()[name = tensor("encoder_ret_lns_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7443264)))]; - tensor encoder_ret_lns_0_weight = const()[name = tensor("encoder_ret_lns_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7444352)))]; - tensor encoder_q_proj_0_bias = const()[name = tensor("encoder_q_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7445440)))]; - tensor encoder_q_proj_0_weight = const()[name = tensor("encoder_q_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7446528)))]; - tensor encoder_k_proj_0_bias = const()[name = tensor("encoder_k_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7708736)))]; - tensor encoder_k_proj_0_weight = const()[name = tensor("encoder_k_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7709824)))]; - tensor encoder_v_proj_0_bias = const()[name = tensor("encoder_v_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7972032)))]; - tensor encoder_v_proj_0_weight = const()[name = tensor("encoder_v_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7973120)))]; - tensor encoder_g_proj_0_bias = const()[name = tensor("encoder_g_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8235328)))]; - tensor encoder_g_proj_0_weight = const()[name = tensor("encoder_g_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8236416)))]; - tensor encoder_out_proj_0_bias = const()[name = tensor("encoder_out_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8498624)))]; - tensor encoder_out_proj_0_weight = const()[name = tensor("encoder_out_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8499712)))]; - tensor encoder_conv_module_0_sequential_0_bias = const()[name = tensor("encoder_conv_module_0_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8761920)))]; - tensor encoder_conv_module_0_sequential_0_weight = const()[name = tensor("encoder_conv_module_0_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8763008)))]; - tensor encoder_conv_module_0_sequential_2_conv_bias = const()[name = tensor("encoder_conv_module_0_sequential_2_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8764096)))]; - tensor encoder_conv_module_0_sequential_2_conv_weight = const()[name = tensor("encoder_conv_module_0_sequential_2_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8766208)))]; - tensor encoder_conv_module_0_sequential_4_conv_weight = const()[name = tensor("encoder_conv_module_0_sequential_4_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9290560)))]; - tensor encoder_conv_module_0_sequential_5_running_var = const()[name = tensor("encoder_conv_module_0_sequential_5_running_var"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9307008)))]; - tensor encoder_conv_module_0_sequential_5_running_mean = const()[name = tensor("encoder_conv_module_0_sequential_5_running_mean"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9308096)))]; - tensor encoder_conv_module_0_sequential_5_bias = const()[name = tensor("encoder_conv_module_0_sequential_5_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9309184)))]; - tensor encoder_conv_module_0_sequential_5_weight = const()[name = tensor("encoder_conv_module_0_sequential_5_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9310272)))]; - tensor encoder_conv_module_0_sequential_7_conv_bias = const()[name = tensor("encoder_conv_module_0_sequential_7_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9311360)))]; - tensor encoder_conv_module_0_sequential_7_conv_weight = const()[name = tensor("encoder_conv_module_0_sequential_7_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9312448)))]; - tensor encoder_ffn2_0_module_sequential_0_bias = const()[name = tensor("encoder_ffn2_0_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9574656)))]; - tensor encoder_ffn2_0_module_sequential_0_weight = const()[name = tensor("encoder_ffn2_0_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9575744)))]; - tensor encoder_ffn2_0_module_sequential_1_linear_bias = const()[name = tensor("encoder_ffn2_0_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9576832)))]; - tensor encoder_ffn2_0_module_sequential_1_linear_weight = const()[name = tensor("encoder_ffn2_0_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9580992)))]; - tensor encoder_ffn2_0_module_sequential_4_linear_bias = const()[name = tensor("encoder_ffn2_0_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(10629632)))]; - tensor encoder_ffn2_0_module_sequential_4_linear_weight = const()[name = tensor("encoder_ffn2_0_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(10630720)))]; - tensor encoder_layer_norm_0_bias = const()[name = tensor("encoder_layer_norm_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11679360)))]; - tensor encoder_layer_norm_0_weight = const()[name = tensor("encoder_layer_norm_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11680448)))]; - tensor encoder_ffn1_1_module_sequential_0_bias = const()[name = tensor("encoder_ffn1_1_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11681536)))]; - tensor encoder_ffn1_1_module_sequential_0_weight = const()[name = tensor("encoder_ffn1_1_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11682624)))]; - tensor encoder_ffn1_1_module_sequential_1_linear_bias = const()[name = tensor("encoder_ffn1_1_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11683712)))]; - tensor encoder_ffn1_1_module_sequential_1_linear_weight = const()[name = tensor("encoder_ffn1_1_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11687872)))]; - tensor encoder_ffn1_1_module_sequential_4_linear_bias = const()[name = tensor("encoder_ffn1_1_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(12736512)))]; - tensor encoder_ffn1_1_module_sequential_4_linear_weight = const()[name = tensor("encoder_ffn1_1_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(12737600)))]; - tensor encoder_ret_lns_1_bias = const()[name = tensor("encoder_ret_lns_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(13786240)))]; - tensor encoder_ret_lns_1_weight = const()[name = tensor("encoder_ret_lns_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(13787328)))]; - tensor encoder_q_proj_1_bias = const()[name = tensor("encoder_q_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(13788416)))]; - tensor encoder_q_proj_1_weight = const()[name = tensor("encoder_q_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(13789504)))]; - tensor encoder_k_proj_1_bias = const()[name = tensor("encoder_k_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14051712)))]; - tensor encoder_k_proj_1_weight = const()[name = tensor("encoder_k_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14052800)))]; - tensor encoder_v_proj_1_bias = const()[name = tensor("encoder_v_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14315008)))]; - tensor encoder_v_proj_1_weight = const()[name = tensor("encoder_v_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14316096)))]; - tensor encoder_g_proj_1_bias = const()[name = tensor("encoder_g_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14578304)))]; - tensor encoder_g_proj_1_weight = const()[name = tensor("encoder_g_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14579392)))]; - tensor encoder_out_proj_1_bias = const()[name = tensor("encoder_out_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14841600)))]; - tensor encoder_out_proj_1_weight = const()[name = tensor("encoder_out_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14842688)))]; - tensor encoder_conv_module_1_sequential_0_bias = const()[name = tensor("encoder_conv_module_1_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15104896)))]; - tensor encoder_conv_module_1_sequential_0_weight = const()[name = tensor("encoder_conv_module_1_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15105984)))]; - tensor encoder_conv_module_1_sequential_2_conv_bias = const()[name = tensor("encoder_conv_module_1_sequential_2_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15107072)))]; - tensor encoder_conv_module_1_sequential_2_conv_weight = const()[name = tensor("encoder_conv_module_1_sequential_2_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15109184)))]; - tensor encoder_conv_module_1_sequential_4_conv_weight = const()[name = tensor("encoder_conv_module_1_sequential_4_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15633536)))]; - tensor encoder_conv_module_1_sequential_5_running_var = const()[name = tensor("encoder_conv_module_1_sequential_5_running_var"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15649984)))]; - tensor encoder_conv_module_1_sequential_5_running_mean = const()[name = tensor("encoder_conv_module_1_sequential_5_running_mean"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15651072)))]; - tensor encoder_conv_module_1_sequential_5_bias = const()[name = tensor("encoder_conv_module_1_sequential_5_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15652160)))]; - tensor encoder_conv_module_1_sequential_5_weight = const()[name = tensor("encoder_conv_module_1_sequential_5_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15653248)))]; - tensor encoder_conv_module_1_sequential_7_conv_bias = const()[name = tensor("encoder_conv_module_1_sequential_7_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15654336)))]; - tensor encoder_conv_module_1_sequential_7_conv_weight = const()[name = tensor("encoder_conv_module_1_sequential_7_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15655424)))]; - tensor encoder_ffn2_1_module_sequential_0_bias = const()[name = tensor("encoder_ffn2_1_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15917632)))]; - tensor encoder_ffn2_1_module_sequential_0_weight = const()[name = tensor("encoder_ffn2_1_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15918720)))]; - tensor encoder_ffn2_1_module_sequential_1_linear_bias = const()[name = tensor("encoder_ffn2_1_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15919808)))]; - tensor encoder_ffn2_1_module_sequential_1_linear_weight = const()[name = tensor("encoder_ffn2_1_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15923968)))]; - tensor encoder_ffn2_1_module_sequential_4_linear_bias = const()[name = tensor("encoder_ffn2_1_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(16972608)))]; - tensor encoder_ffn2_1_module_sequential_4_linear_weight = const()[name = tensor("encoder_ffn2_1_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(16973696)))]; - tensor encoder_layer_norm_1_bias = const()[name = tensor("encoder_layer_norm_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18022336)))]; - tensor encoder_layer_norm_1_weight = const()[name = tensor("encoder_layer_norm_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18023424)))]; - tensor encoder_ffn1_2_module_sequential_0_bias = const()[name = tensor("encoder_ffn1_2_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18024512)))]; - tensor encoder_ffn1_2_module_sequential_0_weight = const()[name = tensor("encoder_ffn1_2_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18025600)))]; - tensor encoder_ffn1_2_module_sequential_1_linear_bias = const()[name = tensor("encoder_ffn1_2_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18026688)))]; - tensor encoder_ffn1_2_module_sequential_1_linear_weight = const()[name = tensor("encoder_ffn1_2_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18030848)))]; - tensor encoder_ffn1_2_module_sequential_4_linear_bias = const()[name = tensor("encoder_ffn1_2_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(19079488)))]; - tensor encoder_ffn1_2_module_sequential_4_linear_weight = const()[name = tensor("encoder_ffn1_2_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(19080576)))]; - tensor encoder_ret_lns_2_bias = const()[name = tensor("encoder_ret_lns_2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20129216)))]; - tensor encoder_ret_lns_2_weight = const()[name = tensor("encoder_ret_lns_2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20130304)))]; - tensor encoder_q_proj_2_bias = const()[name = tensor("encoder_q_proj_2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20131392)))]; - tensor encoder_q_proj_2_weight = const()[name = tensor("encoder_q_proj_2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20132480)))]; - tensor encoder_k_proj_2_bias = const()[name = tensor("encoder_k_proj_2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20394688)))]; - tensor encoder_k_proj_2_weight = const()[name = tensor("encoder_k_proj_2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20395776)))]; - tensor encoder_v_proj_2_bias = const()[name = tensor("encoder_v_proj_2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20657984)))]; - tensor encoder_v_proj_2_weight = const()[name = tensor("encoder_v_proj_2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20659072)))]; - tensor encoder_g_proj_2_bias = const()[name = tensor("encoder_g_proj_2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20921280)))]; - tensor encoder_g_proj_2_weight = const()[name = tensor("encoder_g_proj_2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20922368)))]; - tensor encoder_out_proj_2_bias = const()[name = tensor("encoder_out_proj_2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21184576)))]; - tensor encoder_out_proj_2_weight = const()[name = tensor("encoder_out_proj_2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21185664)))]; - tensor encoder_conv_module_2_sequential_0_bias = const()[name = tensor("encoder_conv_module_2_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21447872)))]; - tensor encoder_conv_module_2_sequential_0_weight = const()[name = tensor("encoder_conv_module_2_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21448960)))]; - tensor encoder_conv_module_2_sequential_2_conv_bias = const()[name = tensor("encoder_conv_module_2_sequential_2_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21450048)))]; - tensor encoder_conv_module_2_sequential_2_conv_weight = const()[name = tensor("encoder_conv_module_2_sequential_2_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21452160)))]; - tensor encoder_conv_module_2_sequential_4_conv_weight = const()[name = tensor("encoder_conv_module_2_sequential_4_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21976512)))]; - tensor encoder_conv_module_2_sequential_5_running_var = const()[name = tensor("encoder_conv_module_2_sequential_5_running_var"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21992960)))]; - tensor encoder_conv_module_2_sequential_5_running_mean = const()[name = tensor("encoder_conv_module_2_sequential_5_running_mean"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21994048)))]; - tensor encoder_conv_module_2_sequential_5_bias = const()[name = tensor("encoder_conv_module_2_sequential_5_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21995136)))]; - tensor encoder_conv_module_2_sequential_5_weight = const()[name = tensor("encoder_conv_module_2_sequential_5_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21996224)))]; - tensor encoder_conv_module_2_sequential_7_conv_bias = const()[name = tensor("encoder_conv_module_2_sequential_7_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21997312)))]; - tensor encoder_conv_module_2_sequential_7_conv_weight = const()[name = tensor("encoder_conv_module_2_sequential_7_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21998400)))]; - tensor encoder_ffn2_2_module_sequential_0_bias = const()[name = tensor("encoder_ffn2_2_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(22260608)))]; - tensor encoder_ffn2_2_module_sequential_0_weight = const()[name = tensor("encoder_ffn2_2_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(22261696)))]; - tensor encoder_ffn2_2_module_sequential_1_linear_bias = const()[name = tensor("encoder_ffn2_2_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(22262784)))]; - tensor encoder_ffn2_2_module_sequential_1_linear_weight = const()[name = tensor("encoder_ffn2_2_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(22266944)))]; - tensor encoder_ffn2_2_module_sequential_4_linear_bias = const()[name = tensor("encoder_ffn2_2_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(23315584)))]; - tensor encoder_ffn2_2_module_sequential_4_linear_weight = const()[name = tensor("encoder_ffn2_2_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(23316672)))]; - tensor encoder_layer_norm_2_bias = const()[name = tensor("encoder_layer_norm_2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(24365312)))]; - tensor encoder_layer_norm_2_weight = const()[name = tensor("encoder_layer_norm_2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(24366400)))]; - tensor encoder_ffn1_3_module_sequential_0_bias = const()[name = tensor("encoder_ffn1_3_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(24367488)))]; - tensor encoder_ffn1_3_module_sequential_0_weight = const()[name = tensor("encoder_ffn1_3_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(24368576)))]; - tensor encoder_ffn1_3_module_sequential_1_linear_bias = const()[name = tensor("encoder_ffn1_3_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(24369664)))]; - tensor encoder_ffn1_3_module_sequential_1_linear_weight = const()[name = tensor("encoder_ffn1_3_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(24373824)))]; - tensor encoder_ffn1_3_module_sequential_4_linear_bias = const()[name = tensor("encoder_ffn1_3_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(25422464)))]; - tensor encoder_ffn1_3_module_sequential_4_linear_weight = const()[name = tensor("encoder_ffn1_3_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(25423552)))]; - tensor encoder_ret_lns_3_bias = const()[name = tensor("encoder_ret_lns_3_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(26472192)))]; - tensor encoder_ret_lns_3_weight = const()[name = tensor("encoder_ret_lns_3_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(26473280)))]; - tensor encoder_q_proj_3_bias = const()[name = tensor("encoder_q_proj_3_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(26474368)))]; - tensor encoder_q_proj_3_weight = const()[name = tensor("encoder_q_proj_3_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(26475456)))]; - tensor encoder_k_proj_3_bias = const()[name = tensor("encoder_k_proj_3_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(26737664)))]; - tensor encoder_k_proj_3_weight = const()[name = tensor("encoder_k_proj_3_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(26738752)))]; - tensor encoder_v_proj_3_bias = const()[name = tensor("encoder_v_proj_3_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27000960)))]; - tensor encoder_v_proj_3_weight = const()[name = tensor("encoder_v_proj_3_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27002048)))]; - tensor encoder_g_proj_3_bias = const()[name = tensor("encoder_g_proj_3_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27264256)))]; - tensor encoder_g_proj_3_weight = const()[name = tensor("encoder_g_proj_3_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27265344)))]; - tensor encoder_out_proj_3_bias = const()[name = tensor("encoder_out_proj_3_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27527552)))]; - tensor encoder_out_proj_3_weight = const()[name = tensor("encoder_out_proj_3_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27528640)))]; - tensor encoder_conv_module_3_sequential_0_bias = const()[name = tensor("encoder_conv_module_3_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27790848)))]; - tensor encoder_conv_module_3_sequential_0_weight = const()[name = tensor("encoder_conv_module_3_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27791936)))]; - tensor encoder_conv_module_3_sequential_2_conv_bias = const()[name = tensor("encoder_conv_module_3_sequential_2_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27793024)))]; - tensor encoder_conv_module_3_sequential_2_conv_weight = const()[name = tensor("encoder_conv_module_3_sequential_2_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27795136)))]; - tensor encoder_conv_module_3_sequential_4_conv_weight = const()[name = tensor("encoder_conv_module_3_sequential_4_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28319488)))]; - tensor encoder_conv_module_3_sequential_5_running_var = const()[name = tensor("encoder_conv_module_3_sequential_5_running_var"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28335936)))]; - tensor encoder_conv_module_3_sequential_5_running_mean = const()[name = tensor("encoder_conv_module_3_sequential_5_running_mean"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28337024)))]; - tensor encoder_conv_module_3_sequential_5_bias = const()[name = tensor("encoder_conv_module_3_sequential_5_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28338112)))]; - tensor encoder_conv_module_3_sequential_5_weight = const()[name = tensor("encoder_conv_module_3_sequential_5_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28339200)))]; - tensor encoder_conv_module_3_sequential_7_conv_bias = const()[name = tensor("encoder_conv_module_3_sequential_7_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28340288)))]; - tensor encoder_conv_module_3_sequential_7_conv_weight = const()[name = tensor("encoder_conv_module_3_sequential_7_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28341376)))]; - tensor encoder_ffn2_3_module_sequential_0_bias = const()[name = tensor("encoder_ffn2_3_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28603584)))]; - tensor encoder_ffn2_3_module_sequential_0_weight = const()[name = tensor("encoder_ffn2_3_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28604672)))]; - tensor encoder_ffn2_3_module_sequential_1_linear_bias = const()[name = tensor("encoder_ffn2_3_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28605760)))]; - tensor encoder_ffn2_3_module_sequential_1_linear_weight = const()[name = tensor("encoder_ffn2_3_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28609920)))]; - tensor encoder_ffn2_3_module_sequential_4_linear_bias = const()[name = tensor("encoder_ffn2_3_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(29658560)))]; - tensor encoder_ffn2_3_module_sequential_4_linear_weight = const()[name = tensor("encoder_ffn2_3_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(29659648)))]; - tensor encoder_layer_norm_3_bias = const()[name = tensor("encoder_layer_norm_3_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(30708288)))]; - tensor encoder_layer_norm_3_weight = const()[name = tensor("encoder_layer_norm_3_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(30709376)))]; - tensor decoder__causal_mask = const()[name = tensor("decoder__causal_mask"), val = tensor([[0x1p+0, 0x0p+0, 0x0p+0], [0x1p+0, 0x1p+0, 0x0p+0], [0x1p+0, 0x1p+0, 0x1p+0]])]; - tensor decoder_convert_bias = const()[name = tensor("decoder_convert_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(30710464)))]; - tensor decoder_convert_weight = const()[name = tensor("decoder_convert_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(30711552)))]; - tensor decoder_q_proj_0_bias = const()[name = tensor("decoder_q_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(31235904)))]; - tensor decoder_q_proj_0_weight = const()[name = tensor("decoder_q_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(31236992)))]; - tensor decoder_k_proj_0_bias = const()[name = tensor("decoder_k_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(31499200)))]; - tensor decoder_k_proj_0_weight = const()[name = tensor("decoder_k_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(31500288)))]; - tensor decoder_v_proj_0_bias = const()[name = tensor("decoder_v_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(31762496)))]; - tensor decoder_v_proj_0_weight = const()[name = tensor("decoder_v_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(31763584)))]; - tensor decoder_g_proj_0_bias = const()[name = tensor("decoder_g_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32025792)))]; - tensor decoder_g_proj_0_weight = const()[name = tensor("decoder_g_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32026880)))]; - tensor decoder_out_proj_0_bias = const()[name = tensor("decoder_out_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32289088)))]; - tensor decoder_out_proj_0_weight = const()[name = tensor("decoder_out_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32290176)))]; - tensor decoder_norm11_0_bias = const()[name = tensor("decoder_norm11_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32552384)))]; - tensor decoder_norm11_0_weight = const()[name = tensor("decoder_norm11_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32553472)))]; - tensor decoder_self_attn2_0_out_proj_bias = const()[name = tensor("decoder_self_attn2_0_out_proj_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32554560)))]; - tensor decoder_self_attn2_0_out_proj_weight = const()[name = tensor("decoder_self_attn2_0_out_proj_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32555648)))]; - tensor decoder_self_attn2_0_in_proj_bias = const()[name = tensor("decoder_self_attn2_0_in_proj_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32817856)))]; - tensor decoder_self_attn2_0_in_proj_weight = const()[name = tensor("decoder_self_attn2_0_in_proj_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32820992)))]; - tensor decoder_norm21_0_bias = const()[name = tensor("decoder_norm21_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(33607488)))]; - tensor decoder_norm21_0_weight = const()[name = tensor("decoder_norm21_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(33608576)))]; - tensor decoder_linear1_0_bias = const()[name = tensor("decoder_linear1_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(33609664)))]; - tensor decoder_linear1_0_weight = const()[name = tensor("decoder_linear1_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(33617920)))]; - tensor decoder_linear2_0_bias = const()[name = tensor("decoder_linear2_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(35715136)))]; - tensor decoder_linear2_0_weight = const()[name = tensor("decoder_linear2_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(35716224)))]; - tensor decoder_norm22_0_bias = const()[name = tensor("decoder_norm22_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(37813440)))]; - tensor decoder_norm22_0_weight = const()[name = tensor("decoder_norm22_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(37814528)))]; - tensor decoder_q_proj_1_bias = const()[name = tensor("decoder_q_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(37815616)))]; - tensor decoder_q_proj_1_weight = const()[name = tensor("decoder_q_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(37816704)))]; - tensor decoder_k_proj_1_bias = const()[name = tensor("decoder_k_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38078912)))]; - tensor decoder_k_proj_1_weight = const()[name = tensor("decoder_k_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38080000)))]; - tensor decoder_v_proj_1_bias = const()[name = tensor("decoder_v_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38342208)))]; - tensor decoder_v_proj_1_weight = const()[name = tensor("decoder_v_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38343296)))]; - tensor decoder_g_proj_1_bias = const()[name = tensor("decoder_g_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38605504)))]; - tensor decoder_g_proj_1_weight = const()[name = tensor("decoder_g_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38606592)))]; - tensor decoder_out_proj_1_bias = const()[name = tensor("decoder_out_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38868800)))]; - tensor decoder_out_proj_1_weight = const()[name = tensor("decoder_out_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38869888)))]; - tensor decoder_norm11_1_bias = const()[name = tensor("decoder_norm11_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(39132096)))]; - tensor decoder_norm11_1_weight = const()[name = tensor("decoder_norm11_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(39133184)))]; - tensor decoder_self_attn2_1_out_proj_bias = const()[name = tensor("decoder_self_attn2_1_out_proj_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(39134272)))]; - tensor decoder_self_attn2_1_out_proj_weight = const()[name = tensor("decoder_self_attn2_1_out_proj_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(39135360)))]; - tensor decoder_self_attn2_1_in_proj_bias = const()[name = tensor("decoder_self_attn2_1_in_proj_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(39397568)))]; - tensor decoder_self_attn2_1_in_proj_weight = const()[name = tensor("decoder_self_attn2_1_in_proj_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(39400704)))]; - tensor decoder_norm21_1_bias = const()[name = tensor("decoder_norm21_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(40187200)))]; - tensor decoder_norm21_1_weight = const()[name = tensor("decoder_norm21_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(40188288)))]; - tensor decoder_linear1_1_bias = const()[name = tensor("decoder_linear1_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(40189376)))]; - tensor decoder_linear1_1_weight = const()[name = tensor("decoder_linear1_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(40197632)))]; - tensor decoder_linear2_1_bias = const()[name = tensor("decoder_linear2_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(42294848)))]; - tensor decoder_linear2_1_weight = const()[name = tensor("decoder_linear2_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(42295936)))]; - tensor decoder_norm22_1_bias = const()[name = tensor("decoder_norm22_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(44393152)))]; - tensor decoder_norm22_1_weight = const()[name = tensor("decoder_norm22_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(44394240)))]; - tensor var_11 = const()[name = tensor("op_11"), val = tensor(0x1.5798eep-27)]; - tensor var_17 = const()[name = tensor("op_17"), val = tensor(0x1.0c6f7ap-20)]; - tensor var_20 = const()[name = tensor("op_20"), val = tensor(2)]; - tensor var_23 = const()[name = tensor("op_23"), val = tensor(0)]; - tensor var_26 = const()[name = tensor("op_26"), val = tensor(1)]; - tensor var_28 = const()[name = tensor("op_28"), val = tensor(0x1.4f8b58p-17)]; - tensor var_29 = const()[name = tensor("op_29"), val = tensor(true)]; - tensor input_1 = linear(bias = encoder_input_projection_linear_bias, weight = encoder_input_projection_linear_weight, x = features)[name = tensor("linear_0")]; - tensor input_3_axes_0 = const()[name = tensor("input_3_axes_0"), val = tensor([-1])]; - tensor input_3 = layer_norm(axes = input_3_axes_0, beta = encoder_pre_layer_norm_bias, epsilon = var_28, gamma = encoder_pre_layer_norm_weight, x = input_1)[name = tensor("input_3")]; + func main(tensor cnn_window, tensor dec_kv, tensor dec_scale, tensor enc_conv_cache, tensor enc_kv, tensor enc_scale, tensor features, tensor valid_mask) { + tensor inner_encoder_cnn_bias = const()[name = tensor("inner_encoder_cnn_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(64)))]; + tensor inner_encoder_cnn_weight = const()[name = tensor("inner_encoder_cnn_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1152)))]; + tensor inner_encoder__causal_mask = const()[name = tensor("inner_encoder__causal_mask"), val = tensor([[0x1p+0, 0x0p+0, 0x0p+0], [0x1p+0, 0x1p+0, 0x0p+0], [0x1p+0, 0x1p+0, 0x1p+0]])]; + tensor inner_encoder__t_index = const()[name = tensor("inner_encoder__t_index"), val = tensor([0x1p+0, 0x1p+1, 0x1.8p+1])]; + tensor inner_encoder_input_projection_linear_bias = const()[name = tensor("inner_encoder_input_projection_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4981952)))]; + tensor inner_encoder_input_projection_linear_weight = const()[name = tensor("inner_encoder_input_projection_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4983040)))]; + tensor inner_encoder_pre_layer_norm_bias = const()[name = tensor("inner_encoder_pre_layer_norm_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(5336384)))]; + tensor inner_encoder_pre_layer_norm_weight = const()[name = tensor("inner_encoder_pre_layer_norm_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(5337472)))]; + tensor inner_encoder_ffn1_0_module_sequential_0_bias = const()[name = tensor("inner_encoder_ffn1_0_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(5338560)))]; + tensor inner_encoder_ffn1_0_module_sequential_0_weight = const()[name = tensor("inner_encoder_ffn1_0_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(5339648)))]; + tensor inner_encoder_ffn1_0_module_sequential_1_linear_bias = const()[name = tensor("inner_encoder_ffn1_0_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(5340736)))]; + tensor inner_encoder_ffn1_0_module_sequential_1_linear_weight = const()[name = tensor("inner_encoder_ffn1_0_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(5344896)))]; + tensor inner_encoder_ffn1_0_module_sequential_4_linear_bias = const()[name = tensor("inner_encoder_ffn1_0_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(6393536)))]; + tensor inner_encoder_ffn1_0_module_sequential_4_linear_weight = const()[name = tensor("inner_encoder_ffn1_0_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(6394624)))]; + tensor inner_encoder_ret_lns_0_bias = const()[name = tensor("inner_encoder_ret_lns_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7443264)))]; + tensor inner_encoder_ret_lns_0_weight = const()[name = tensor("inner_encoder_ret_lns_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7444352)))]; + tensor inner_encoder_q_proj_0_bias = const()[name = tensor("inner_encoder_q_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7445440)))]; + tensor inner_encoder_q_proj_0_weight = const()[name = tensor("inner_encoder_q_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7446528)))]; + tensor inner_encoder_k_proj_0_bias = const()[name = tensor("inner_encoder_k_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7708736)))]; + tensor inner_encoder_k_proj_0_weight = const()[name = tensor("inner_encoder_k_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7709824)))]; + tensor inner_encoder_v_proj_0_bias = const()[name = tensor("inner_encoder_v_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7972032)))]; + tensor inner_encoder_v_proj_0_weight = const()[name = tensor("inner_encoder_v_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7973120)))]; + tensor inner_encoder_g_proj_0_bias = const()[name = tensor("inner_encoder_g_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8235328)))]; + tensor inner_encoder_g_proj_0_weight = const()[name = tensor("inner_encoder_g_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8236416)))]; + tensor inner_encoder_out_proj_0_bias = const()[name = tensor("inner_encoder_out_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8498624)))]; + tensor inner_encoder_out_proj_0_weight = const()[name = tensor("inner_encoder_out_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8499712)))]; + tensor inner_encoder_conv_module_0_sequential_0_bias = const()[name = tensor("inner_encoder_conv_module_0_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8761920)))]; + tensor inner_encoder_conv_module_0_sequential_0_weight = const()[name = tensor("inner_encoder_conv_module_0_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8763008)))]; + tensor inner_encoder_conv_module_0_sequential_2_conv_bias = const()[name = tensor("inner_encoder_conv_module_0_sequential_2_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8764096)))]; + tensor inner_encoder_conv_module_0_sequential_2_conv_weight = const()[name = tensor("inner_encoder_conv_module_0_sequential_2_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8766208)))]; + tensor inner_encoder_conv_module_0_sequential_4_conv_weight = const()[name = tensor("inner_encoder_conv_module_0_sequential_4_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9290560)))]; + tensor inner_encoder_conv_module_0_sequential_5_running_var = const()[name = tensor("inner_encoder_conv_module_0_sequential_5_running_var"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9307008)))]; + tensor inner_encoder_conv_module_0_sequential_5_running_mean = const()[name = tensor("inner_encoder_conv_module_0_sequential_5_running_mean"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9308096)))]; + tensor inner_encoder_conv_module_0_sequential_5_bias = const()[name = tensor("inner_encoder_conv_module_0_sequential_5_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9309184)))]; + tensor inner_encoder_conv_module_0_sequential_5_weight = const()[name = tensor("inner_encoder_conv_module_0_sequential_5_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9310272)))]; + tensor inner_encoder_conv_module_0_sequential_7_conv_bias = const()[name = tensor("inner_encoder_conv_module_0_sequential_7_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9311360)))]; + tensor inner_encoder_conv_module_0_sequential_7_conv_weight = const()[name = tensor("inner_encoder_conv_module_0_sequential_7_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9312448)))]; + tensor inner_encoder_ffn2_0_module_sequential_0_bias = const()[name = tensor("inner_encoder_ffn2_0_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9574656)))]; + tensor inner_encoder_ffn2_0_module_sequential_0_weight = const()[name = tensor("inner_encoder_ffn2_0_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9575744)))]; + tensor inner_encoder_ffn2_0_module_sequential_1_linear_bias = const()[name = tensor("inner_encoder_ffn2_0_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9576832)))]; + tensor inner_encoder_ffn2_0_module_sequential_1_linear_weight = const()[name = tensor("inner_encoder_ffn2_0_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9580992)))]; + tensor inner_encoder_ffn2_0_module_sequential_4_linear_bias = const()[name = tensor("inner_encoder_ffn2_0_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(10629632)))]; + tensor inner_encoder_ffn2_0_module_sequential_4_linear_weight = const()[name = tensor("inner_encoder_ffn2_0_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(10630720)))]; + tensor inner_encoder_layer_norm_0_bias = const()[name = tensor("inner_encoder_layer_norm_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11679360)))]; + tensor inner_encoder_layer_norm_0_weight = const()[name = tensor("inner_encoder_layer_norm_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11680448)))]; + tensor inner_encoder_ffn1_1_module_sequential_0_bias = const()[name = tensor("inner_encoder_ffn1_1_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11681536)))]; + tensor inner_encoder_ffn1_1_module_sequential_0_weight = const()[name = tensor("inner_encoder_ffn1_1_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11682624)))]; + tensor inner_encoder_ffn1_1_module_sequential_1_linear_bias = const()[name = tensor("inner_encoder_ffn1_1_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11683712)))]; + tensor inner_encoder_ffn1_1_module_sequential_1_linear_weight = const()[name = tensor("inner_encoder_ffn1_1_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11687872)))]; + tensor inner_encoder_ffn1_1_module_sequential_4_linear_bias = const()[name = tensor("inner_encoder_ffn1_1_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(12736512)))]; + tensor inner_encoder_ffn1_1_module_sequential_4_linear_weight = const()[name = tensor("inner_encoder_ffn1_1_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(12737600)))]; + tensor inner_encoder_ret_lns_1_bias = const()[name = tensor("inner_encoder_ret_lns_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(13786240)))]; + tensor inner_encoder_ret_lns_1_weight = const()[name = tensor("inner_encoder_ret_lns_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(13787328)))]; + tensor inner_encoder_q_proj_1_bias = const()[name = tensor("inner_encoder_q_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(13788416)))]; + tensor inner_encoder_q_proj_1_weight = const()[name = tensor("inner_encoder_q_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(13789504)))]; + tensor inner_encoder_k_proj_1_bias = const()[name = tensor("inner_encoder_k_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14051712)))]; + tensor inner_encoder_k_proj_1_weight = const()[name = tensor("inner_encoder_k_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14052800)))]; + tensor inner_encoder_v_proj_1_bias = const()[name = tensor("inner_encoder_v_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14315008)))]; + tensor inner_encoder_v_proj_1_weight = const()[name = tensor("inner_encoder_v_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14316096)))]; + tensor inner_encoder_g_proj_1_bias = const()[name = tensor("inner_encoder_g_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14578304)))]; + tensor inner_encoder_g_proj_1_weight = const()[name = tensor("inner_encoder_g_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14579392)))]; + tensor inner_encoder_out_proj_1_bias = const()[name = tensor("inner_encoder_out_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14841600)))]; + tensor inner_encoder_out_proj_1_weight = const()[name = tensor("inner_encoder_out_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14842688)))]; + tensor inner_encoder_conv_module_1_sequential_0_bias = const()[name = tensor("inner_encoder_conv_module_1_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15104896)))]; + tensor inner_encoder_conv_module_1_sequential_0_weight = const()[name = tensor("inner_encoder_conv_module_1_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15105984)))]; + tensor inner_encoder_conv_module_1_sequential_2_conv_bias = const()[name = tensor("inner_encoder_conv_module_1_sequential_2_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15107072)))]; + tensor inner_encoder_conv_module_1_sequential_2_conv_weight = const()[name = tensor("inner_encoder_conv_module_1_sequential_2_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15109184)))]; + tensor inner_encoder_conv_module_1_sequential_4_conv_weight = const()[name = tensor("inner_encoder_conv_module_1_sequential_4_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15633536)))]; + tensor inner_encoder_conv_module_1_sequential_5_running_var = const()[name = tensor("inner_encoder_conv_module_1_sequential_5_running_var"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15649984)))]; + tensor inner_encoder_conv_module_1_sequential_5_running_mean = const()[name = tensor("inner_encoder_conv_module_1_sequential_5_running_mean"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15651072)))]; + tensor inner_encoder_conv_module_1_sequential_5_bias = const()[name = tensor("inner_encoder_conv_module_1_sequential_5_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15652160)))]; + tensor inner_encoder_conv_module_1_sequential_5_weight = const()[name = tensor("inner_encoder_conv_module_1_sequential_5_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15653248)))]; + tensor inner_encoder_conv_module_1_sequential_7_conv_bias = const()[name = tensor("inner_encoder_conv_module_1_sequential_7_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15654336)))]; + tensor inner_encoder_conv_module_1_sequential_7_conv_weight = const()[name = tensor("inner_encoder_conv_module_1_sequential_7_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15655424)))]; + tensor inner_encoder_ffn2_1_module_sequential_0_bias = const()[name = tensor("inner_encoder_ffn2_1_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15917632)))]; + tensor inner_encoder_ffn2_1_module_sequential_0_weight = const()[name = tensor("inner_encoder_ffn2_1_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15918720)))]; + tensor inner_encoder_ffn2_1_module_sequential_1_linear_bias = const()[name = tensor("inner_encoder_ffn2_1_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15919808)))]; + tensor inner_encoder_ffn2_1_module_sequential_1_linear_weight = const()[name = tensor("inner_encoder_ffn2_1_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15923968)))]; + tensor inner_encoder_ffn2_1_module_sequential_4_linear_bias = const()[name = tensor("inner_encoder_ffn2_1_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(16972608)))]; + tensor inner_encoder_ffn2_1_module_sequential_4_linear_weight = const()[name = tensor("inner_encoder_ffn2_1_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(16973696)))]; + tensor inner_encoder_layer_norm_1_bias = const()[name = tensor("inner_encoder_layer_norm_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18022336)))]; + tensor inner_encoder_layer_norm_1_weight = const()[name = tensor("inner_encoder_layer_norm_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18023424)))]; + tensor inner_encoder_ffn1_2_module_sequential_0_bias = const()[name = tensor("inner_encoder_ffn1_2_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18024512)))]; + tensor inner_encoder_ffn1_2_module_sequential_0_weight = const()[name = tensor("inner_encoder_ffn1_2_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18025600)))]; + tensor inner_encoder_ffn1_2_module_sequential_1_linear_bias = const()[name = tensor("inner_encoder_ffn1_2_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18026688)))]; + tensor inner_encoder_ffn1_2_module_sequential_1_linear_weight = const()[name = tensor("inner_encoder_ffn1_2_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18030848)))]; + tensor inner_encoder_ffn1_2_module_sequential_4_linear_bias = const()[name = tensor("inner_encoder_ffn1_2_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(19079488)))]; + tensor inner_encoder_ffn1_2_module_sequential_4_linear_weight = const()[name = tensor("inner_encoder_ffn1_2_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(19080576)))]; + tensor inner_encoder_ret_lns_2_bias = const()[name = tensor("inner_encoder_ret_lns_2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20129216)))]; + tensor inner_encoder_ret_lns_2_weight = const()[name = tensor("inner_encoder_ret_lns_2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20130304)))]; + tensor inner_encoder_q_proj_2_bias = const()[name = tensor("inner_encoder_q_proj_2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20131392)))]; + tensor inner_encoder_q_proj_2_weight = const()[name = tensor("inner_encoder_q_proj_2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20132480)))]; + tensor inner_encoder_k_proj_2_bias = const()[name = tensor("inner_encoder_k_proj_2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20394688)))]; + tensor inner_encoder_k_proj_2_weight = const()[name = tensor("inner_encoder_k_proj_2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20395776)))]; + tensor inner_encoder_v_proj_2_bias = const()[name = tensor("inner_encoder_v_proj_2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20657984)))]; + tensor inner_encoder_v_proj_2_weight = const()[name = tensor("inner_encoder_v_proj_2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20659072)))]; + tensor inner_encoder_g_proj_2_bias = const()[name = tensor("inner_encoder_g_proj_2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20921280)))]; + tensor inner_encoder_g_proj_2_weight = const()[name = tensor("inner_encoder_g_proj_2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20922368)))]; + tensor inner_encoder_out_proj_2_bias = const()[name = tensor("inner_encoder_out_proj_2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21184576)))]; + tensor inner_encoder_out_proj_2_weight = const()[name = tensor("inner_encoder_out_proj_2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21185664)))]; + tensor inner_encoder_conv_module_2_sequential_0_bias = const()[name = tensor("inner_encoder_conv_module_2_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21447872)))]; + tensor inner_encoder_conv_module_2_sequential_0_weight = const()[name = tensor("inner_encoder_conv_module_2_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21448960)))]; + tensor inner_encoder_conv_module_2_sequential_2_conv_bias = const()[name = tensor("inner_encoder_conv_module_2_sequential_2_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21450048)))]; + tensor inner_encoder_conv_module_2_sequential_2_conv_weight = const()[name = tensor("inner_encoder_conv_module_2_sequential_2_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21452160)))]; + tensor inner_encoder_conv_module_2_sequential_4_conv_weight = const()[name = tensor("inner_encoder_conv_module_2_sequential_4_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21976512)))]; + tensor inner_encoder_conv_module_2_sequential_5_running_var = const()[name = tensor("inner_encoder_conv_module_2_sequential_5_running_var"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21992960)))]; + tensor inner_encoder_conv_module_2_sequential_5_running_mean = const()[name = tensor("inner_encoder_conv_module_2_sequential_5_running_mean"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21994048)))]; + tensor inner_encoder_conv_module_2_sequential_5_bias = const()[name = tensor("inner_encoder_conv_module_2_sequential_5_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21995136)))]; + tensor inner_encoder_conv_module_2_sequential_5_weight = const()[name = tensor("inner_encoder_conv_module_2_sequential_5_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21996224)))]; + tensor inner_encoder_conv_module_2_sequential_7_conv_bias = const()[name = tensor("inner_encoder_conv_module_2_sequential_7_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21997312)))]; + tensor inner_encoder_conv_module_2_sequential_7_conv_weight = const()[name = tensor("inner_encoder_conv_module_2_sequential_7_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21998400)))]; + tensor inner_encoder_ffn2_2_module_sequential_0_bias = const()[name = tensor("inner_encoder_ffn2_2_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(22260608)))]; + tensor inner_encoder_ffn2_2_module_sequential_0_weight = const()[name = tensor("inner_encoder_ffn2_2_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(22261696)))]; + tensor inner_encoder_ffn2_2_module_sequential_1_linear_bias = const()[name = tensor("inner_encoder_ffn2_2_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(22262784)))]; + tensor inner_encoder_ffn2_2_module_sequential_1_linear_weight = const()[name = tensor("inner_encoder_ffn2_2_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(22266944)))]; + tensor inner_encoder_ffn2_2_module_sequential_4_linear_bias = const()[name = tensor("inner_encoder_ffn2_2_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(23315584)))]; + tensor inner_encoder_ffn2_2_module_sequential_4_linear_weight = const()[name = tensor("inner_encoder_ffn2_2_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(23316672)))]; + tensor inner_encoder_layer_norm_2_bias = const()[name = tensor("inner_encoder_layer_norm_2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(24365312)))]; + tensor inner_encoder_layer_norm_2_weight = const()[name = tensor("inner_encoder_layer_norm_2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(24366400)))]; + tensor inner_encoder_ffn1_3_module_sequential_0_bias = const()[name = tensor("inner_encoder_ffn1_3_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(24367488)))]; + tensor inner_encoder_ffn1_3_module_sequential_0_weight = const()[name = tensor("inner_encoder_ffn1_3_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(24368576)))]; + tensor inner_encoder_ffn1_3_module_sequential_1_linear_bias = const()[name = tensor("inner_encoder_ffn1_3_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(24369664)))]; + tensor inner_encoder_ffn1_3_module_sequential_1_linear_weight = const()[name = tensor("inner_encoder_ffn1_3_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(24373824)))]; + tensor inner_encoder_ffn1_3_module_sequential_4_linear_bias = const()[name = tensor("inner_encoder_ffn1_3_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(25422464)))]; + tensor inner_encoder_ffn1_3_module_sequential_4_linear_weight = const()[name = tensor("inner_encoder_ffn1_3_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(25423552)))]; + tensor inner_encoder_ret_lns_3_bias = const()[name = tensor("inner_encoder_ret_lns_3_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(26472192)))]; + tensor inner_encoder_ret_lns_3_weight = const()[name = tensor("inner_encoder_ret_lns_3_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(26473280)))]; + tensor inner_encoder_q_proj_3_bias = const()[name = tensor("inner_encoder_q_proj_3_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(26474368)))]; + tensor inner_encoder_q_proj_3_weight = const()[name = tensor("inner_encoder_q_proj_3_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(26475456)))]; + tensor inner_encoder_k_proj_3_bias = const()[name = tensor("inner_encoder_k_proj_3_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(26737664)))]; + tensor inner_encoder_k_proj_3_weight = const()[name = tensor("inner_encoder_k_proj_3_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(26738752)))]; + tensor inner_encoder_v_proj_3_bias = const()[name = tensor("inner_encoder_v_proj_3_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27000960)))]; + tensor inner_encoder_v_proj_3_weight = const()[name = tensor("inner_encoder_v_proj_3_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27002048)))]; + tensor inner_encoder_g_proj_3_bias = const()[name = tensor("inner_encoder_g_proj_3_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27264256)))]; + tensor inner_encoder_g_proj_3_weight = const()[name = tensor("inner_encoder_g_proj_3_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27265344)))]; + tensor inner_encoder_out_proj_3_bias = const()[name = tensor("inner_encoder_out_proj_3_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27527552)))]; + tensor inner_encoder_out_proj_3_weight = const()[name = tensor("inner_encoder_out_proj_3_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27528640)))]; + tensor inner_encoder_conv_module_3_sequential_0_bias = const()[name = tensor("inner_encoder_conv_module_3_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27790848)))]; + tensor inner_encoder_conv_module_3_sequential_0_weight = const()[name = tensor("inner_encoder_conv_module_3_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27791936)))]; + tensor inner_encoder_conv_module_3_sequential_2_conv_bias = const()[name = tensor("inner_encoder_conv_module_3_sequential_2_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27793024)))]; + tensor inner_encoder_conv_module_3_sequential_2_conv_weight = const()[name = tensor("inner_encoder_conv_module_3_sequential_2_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27795136)))]; + tensor inner_encoder_conv_module_3_sequential_4_conv_weight = const()[name = tensor("inner_encoder_conv_module_3_sequential_4_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28319488)))]; + tensor inner_encoder_conv_module_3_sequential_5_running_var = const()[name = tensor("inner_encoder_conv_module_3_sequential_5_running_var"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28335936)))]; + tensor inner_encoder_conv_module_3_sequential_5_running_mean = const()[name = tensor("inner_encoder_conv_module_3_sequential_5_running_mean"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28337024)))]; + tensor inner_encoder_conv_module_3_sequential_5_bias = const()[name = tensor("inner_encoder_conv_module_3_sequential_5_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28338112)))]; + tensor inner_encoder_conv_module_3_sequential_5_weight = const()[name = tensor("inner_encoder_conv_module_3_sequential_5_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28339200)))]; + tensor inner_encoder_conv_module_3_sequential_7_conv_bias = const()[name = tensor("inner_encoder_conv_module_3_sequential_7_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28340288)))]; + tensor inner_encoder_conv_module_3_sequential_7_conv_weight = const()[name = tensor("inner_encoder_conv_module_3_sequential_7_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28341376)))]; + tensor inner_encoder_ffn2_3_module_sequential_0_bias = const()[name = tensor("inner_encoder_ffn2_3_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28603584)))]; + tensor inner_encoder_ffn2_3_module_sequential_0_weight = const()[name = tensor("inner_encoder_ffn2_3_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28604672)))]; + tensor inner_encoder_ffn2_3_module_sequential_1_linear_bias = const()[name = tensor("inner_encoder_ffn2_3_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28605760)))]; + tensor inner_encoder_ffn2_3_module_sequential_1_linear_weight = const()[name = tensor("inner_encoder_ffn2_3_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28609920)))]; + tensor inner_encoder_ffn2_3_module_sequential_4_linear_bias = const()[name = tensor("inner_encoder_ffn2_3_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(29658560)))]; + tensor inner_encoder_ffn2_3_module_sequential_4_linear_weight = const()[name = tensor("inner_encoder_ffn2_3_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(29659648)))]; + tensor inner_encoder_layer_norm_3_bias = const()[name = tensor("inner_encoder_layer_norm_3_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(30708288)))]; + tensor inner_encoder_layer_norm_3_weight = const()[name = tensor("inner_encoder_layer_norm_3_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(30709376)))]; + tensor inner_decoder__causal_mask = const()[name = tensor("inner_decoder__causal_mask"), val = tensor([[0x1p+0, 0x0p+0, 0x0p+0], [0x1p+0, 0x1p+0, 0x0p+0], [0x1p+0, 0x1p+0, 0x1p+0]])]; + tensor inner_decoder_convert_bias = const()[name = tensor("inner_decoder_convert_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(30710464)))]; + tensor inner_decoder_convert_weight = const()[name = tensor("inner_decoder_convert_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(30711552)))]; + tensor inner_decoder_q_proj_0_bias = const()[name = tensor("inner_decoder_q_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(31235904)))]; + tensor inner_decoder_q_proj_0_weight = const()[name = tensor("inner_decoder_q_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(31236992)))]; + tensor inner_decoder_k_proj_0_bias = const()[name = tensor("inner_decoder_k_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(31499200)))]; + tensor inner_decoder_k_proj_0_weight = const()[name = tensor("inner_decoder_k_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(31500288)))]; + tensor inner_decoder_v_proj_0_bias = const()[name = tensor("inner_decoder_v_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(31762496)))]; + tensor inner_decoder_v_proj_0_weight = const()[name = tensor("inner_decoder_v_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(31763584)))]; + tensor inner_decoder_g_proj_0_bias = const()[name = tensor("inner_decoder_g_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32025792)))]; + tensor inner_decoder_g_proj_0_weight = const()[name = tensor("inner_decoder_g_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32026880)))]; + tensor inner_decoder_out_proj_0_bias = const()[name = tensor("inner_decoder_out_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32289088)))]; + tensor inner_decoder_out_proj_0_weight = const()[name = tensor("inner_decoder_out_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32290176)))]; + tensor inner_decoder_norm11_0_bias = const()[name = tensor("inner_decoder_norm11_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32552384)))]; + tensor inner_decoder_norm11_0_weight = const()[name = tensor("inner_decoder_norm11_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32553472)))]; + tensor inner_decoder_self_attn2_0_out_proj_bias = const()[name = tensor("inner_decoder_self_attn2_0_out_proj_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32554560)))]; + tensor inner_decoder_self_attn2_0_out_proj_weight = const()[name = tensor("inner_decoder_self_attn2_0_out_proj_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32555648)))]; + tensor inner_decoder_self_attn2_0_in_proj_bias = const()[name = tensor("inner_decoder_self_attn2_0_in_proj_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32817856)))]; + tensor inner_decoder_self_attn2_0_in_proj_weight = const()[name = tensor("inner_decoder_self_attn2_0_in_proj_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32820992)))]; + tensor inner_decoder_norm21_0_bias = const()[name = tensor("inner_decoder_norm21_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(33607488)))]; + tensor inner_decoder_norm21_0_weight = const()[name = tensor("inner_decoder_norm21_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(33608576)))]; + tensor inner_decoder_linear1_0_bias = const()[name = tensor("inner_decoder_linear1_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(33609664)))]; + tensor inner_decoder_linear1_0_weight = const()[name = tensor("inner_decoder_linear1_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(33617920)))]; + tensor inner_decoder_linear2_0_bias = const()[name = tensor("inner_decoder_linear2_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(35715136)))]; + tensor inner_decoder_linear2_0_weight = const()[name = tensor("inner_decoder_linear2_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(35716224)))]; + tensor inner_decoder_norm22_0_bias = const()[name = tensor("inner_decoder_norm22_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(37813440)))]; + tensor inner_decoder_norm22_0_weight = const()[name = tensor("inner_decoder_norm22_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(37814528)))]; + tensor inner_decoder_q_proj_1_bias = const()[name = tensor("inner_decoder_q_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(37815616)))]; + tensor inner_decoder_q_proj_1_weight = const()[name = tensor("inner_decoder_q_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(37816704)))]; + tensor inner_decoder_k_proj_1_bias = const()[name = tensor("inner_decoder_k_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38078912)))]; + tensor inner_decoder_k_proj_1_weight = const()[name = tensor("inner_decoder_k_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38080000)))]; + tensor inner_decoder_v_proj_1_bias = const()[name = tensor("inner_decoder_v_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38342208)))]; + tensor inner_decoder_v_proj_1_weight = const()[name = tensor("inner_decoder_v_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38343296)))]; + tensor inner_decoder_g_proj_1_bias = const()[name = tensor("inner_decoder_g_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38605504)))]; + tensor inner_decoder_g_proj_1_weight = const()[name = tensor("inner_decoder_g_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38606592)))]; + tensor inner_decoder_out_proj_1_bias = const()[name = tensor("inner_decoder_out_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38868800)))]; + tensor inner_decoder_out_proj_1_weight = const()[name = tensor("inner_decoder_out_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38869888)))]; + tensor inner_decoder_norm11_1_bias = const()[name = tensor("inner_decoder_norm11_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(39132096)))]; + tensor inner_decoder_norm11_1_weight = const()[name = tensor("inner_decoder_norm11_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(39133184)))]; + tensor inner_decoder_self_attn2_1_out_proj_bias = const()[name = tensor("inner_decoder_self_attn2_1_out_proj_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(39134272)))]; + tensor inner_decoder_self_attn2_1_out_proj_weight = const()[name = tensor("inner_decoder_self_attn2_1_out_proj_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(39135360)))]; + tensor inner_decoder_self_attn2_1_in_proj_bias = const()[name = tensor("inner_decoder_self_attn2_1_in_proj_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(39397568)))]; + tensor inner_decoder_self_attn2_1_in_proj_weight = const()[name = tensor("inner_decoder_self_attn2_1_in_proj_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(39400704)))]; + tensor inner_decoder_norm21_1_bias = const()[name = tensor("inner_decoder_norm21_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(40187200)))]; + tensor inner_decoder_norm21_1_weight = const()[name = tensor("inner_decoder_norm21_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(40188288)))]; + tensor inner_decoder_linear1_1_bias = const()[name = tensor("inner_decoder_linear1_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(40189376)))]; + tensor inner_decoder_linear1_1_weight = const()[name = tensor("inner_decoder_linear1_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(40197632)))]; + tensor inner_decoder_linear2_1_bias = const()[name = tensor("inner_decoder_linear2_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(42294848)))]; + tensor inner_decoder_linear2_1_weight = const()[name = tensor("inner_decoder_linear2_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(42295936)))]; + tensor inner_decoder_norm22_1_bias = const()[name = tensor("inner_decoder_norm22_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(44393152)))]; + tensor inner_decoder_norm22_1_weight = const()[name = tensor("inner_decoder_norm22_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(44394240)))]; + tensor var_19_begin_0 = const()[name = tensor("op_19_begin_0"), val = tensor([0, 0, 0])]; + tensor var_19_end_0 = const()[name = tensor("op_19_end_0"), val = tensor([1, 15, 23])]; + tensor var_19_end_mask_0 = const()[name = tensor("op_19_end_mask_0"), val = tensor([true, false, true])]; + tensor var_19 = slice_by_index(begin = var_19_begin_0, end = var_19_end_0, end_mask = var_19_end_mask_0, x = features)[name = tensor("op_19")]; + tensor var_29_begin_0 = const()[name = tensor("op_29_begin_0"), val = tensor([0, 10, 0])]; + tensor var_29_end_0 = const()[name = tensor("op_29_end_0"), val = tensor([1, 25, 23])]; + tensor var_29_end_mask_0 = const()[name = tensor("op_29_end_mask_0"), val = tensor([true, false, true])]; + tensor var_29 = slice_by_index(begin = var_29_begin_0, end = var_29_end_0, end_mask = var_29_end_mask_0, x = features)[name = tensor("op_29")]; + tensor var_39_begin_0 = const()[name = tensor("op_39_begin_0"), val = tensor([0, 20, 0])]; + tensor var_39_end_0 = const()[name = tensor("op_39_end_0"), val = tensor([1, 1, 23])]; + tensor var_39_end_mask_0 = const()[name = tensor("op_39_end_mask_0"), val = tensor([true, true, true])]; + tensor var_39 = slice_by_index(begin = var_39_begin_0, end = var_39_end_0, end_mask = var_39_end_mask_0, x = features)[name = tensor("op_39")]; + tensor stacked_axis_0 = const()[name = tensor("stacked_axis_0"), val = tensor(1)]; + tensor stacked = stack(axis = stacked_axis_0, values = (var_19, var_29, var_39))[name = tensor("stacked")]; + tensor var_46 = const()[name = tensor("op_46"), val = tensor([1, 3, 345])]; + tensor input_1 = reshape(shape = var_46, x = stacked)[name = tensor("input_1")]; + tensor var_48 = const()[name = tensor("op_48"), val = tensor(0x1p+0)]; + tensor var_54 = const()[name = tensor("op_54"), val = tensor(true)]; + tensor var_55 = const()[name = tensor("op_55"), val = tensor(0x1.4f8b58p-17)]; + tensor var_58 = const()[name = tensor("op_58"), val = tensor(0)]; + tensor var_60 = const()[name = tensor("op_60"), val = tensor(2)]; + tensor var_61 = const()[name = tensor("op_61"), val = tensor(-1)]; + tensor var_63 = const()[name = tensor("op_63"), val = tensor(0x1.0c6f7ap-20)]; + tensor var_68 = const()[name = tensor("op_68"), val = tensor(0x1.5798eep-27)]; + tensor var_72 = const()[name = tensor("op_72"), val = tensor(1)]; + tensor input_3 = linear(bias = inner_encoder_input_projection_linear_bias, weight = inner_encoder_input_projection_linear_weight, x = input_1)[name = tensor("linear_0")]; tensor input_5_axes_0 = const()[name = tensor("input_5_axes_0"), val = tensor([-1])]; - tensor input_5 = layer_norm(axes = input_5_axes_0, beta = encoder_ffn1_0_module_sequential_0_bias, epsilon = var_28, gamma = encoder_ffn1_0_module_sequential_0_weight, x = input_3)[name = tensor("input_5")]; - tensor inputs_1 = linear(bias = encoder_ffn1_0_module_sequential_1_linear_bias, weight = encoder_ffn1_0_module_sequential_1_linear_weight, x = input_5)[name = tensor("linear_1")]; - tensor input_7 = silu(x = inputs_1)[name = tensor("input_7")]; - tensor input_11 = linear(bias = encoder_ffn1_0_module_sequential_4_linear_bias, weight = encoder_ffn1_0_module_sequential_4_linear_weight, x = input_7)[name = tensor("linear_2")]; - tensor var_147 = const()[name = tensor("op_147"), val = tensor(0x1p-1)]; - tensor var_148 = mul(x = input_11, y = var_147)[name = tensor("op_148")]; - tensor input_13 = add(x = var_148, y = input_3)[name = tensor("input_13")]; + tensor input_5 = layer_norm(axes = input_5_axes_0, beta = inner_encoder_pre_layer_norm_bias, epsilon = var_55, gamma = inner_encoder_pre_layer_norm_weight, x = input_3)[name = tensor("input_5")]; + tensor input_7_axes_0 = const()[name = tensor("input_7_axes_0"), val = tensor([-1])]; + tensor input_7 = layer_norm(axes = input_7_axes_0, beta = inner_encoder_ffn1_0_module_sequential_0_bias, epsilon = var_55, gamma = inner_encoder_ffn1_0_module_sequential_0_weight, x = input_5)[name = tensor("input_7")]; + tensor inputs_1 = linear(bias = inner_encoder_ffn1_0_module_sequential_1_linear_bias, weight = inner_encoder_ffn1_0_module_sequential_1_linear_weight, x = input_7)[name = tensor("linear_1")]; + tensor input_9 = silu(x = inputs_1)[name = tensor("input_9")]; + tensor input_13 = linear(bias = inner_encoder_ffn1_0_module_sequential_4_linear_bias, weight = inner_encoder_ffn1_0_module_sequential_4_linear_weight, x = input_9)[name = tensor("linear_2")]; + tensor var_193 = const()[name = tensor("op_193"), val = tensor(0x1p-1)]; + tensor var_194 = mul(x = input_13, y = var_193)[name = tensor("op_194")]; + tensor input_15 = add(x = var_194, y = input_5)[name = tensor("input_15")]; tensor x_1_axes_0 = const()[name = tensor("x_1_axes_0"), val = tensor([-1])]; - tensor x_1 = layer_norm(axes = x_1_axes_0, beta = encoder_ret_lns_0_bias, epsilon = var_28, gamma = encoder_ret_lns_0_weight, x = input_13)[name = tensor("x_1")]; + tensor x_1 = layer_norm(axes = x_1_axes_0, beta = inner_encoder_ret_lns_0_bias, epsilon = var_55, gamma = inner_encoder_ret_lns_0_weight, x = input_15)[name = tensor("x_1")]; tensor prev_kv_1_begin_0 = const()[name = tensor("prev_kv_1_begin_0"), val = tensor([0, 0, 0, 0, 0])]; tensor prev_kv_1_end_0 = const()[name = tensor("prev_kv_1_end_0"), val = tensor([1, 1, 4, 64, 64])]; tensor prev_kv_1_end_mask_0 = const()[name = tensor("prev_kv_1_end_mask_0"), val = tensor([false, true, true, true, true])]; @@ -239,163 +257,163 @@ program(1.0) tensor prev_scale_1_end_mask_0 = const()[name = tensor("prev_scale_1_end_mask_0"), val = tensor([false, true])]; tensor prev_scale_1_squeeze_mask_0 = const()[name = tensor("prev_scale_1_squeeze_mask_0"), val = tensor([true, false])]; tensor prev_scale_1 = slice_by_index(begin = prev_scale_1_begin_0, end = prev_scale_1_end_0, end_mask = prev_scale_1_end_mask_0, squeeze_mask = prev_scale_1_squeeze_mask_0, x = enc_scale)[name = tensor("prev_scale_1")]; - tensor var_162 = linear(bias = encoder_q_proj_0_bias, weight = encoder_q_proj_0_weight, x = x_1)[name = tensor("linear_3")]; - tensor var_163 = const()[name = tensor("op_163"), val = tensor([1, 3, 4, 64])]; - tensor var_164 = reshape(shape = var_163, x = var_162)[name = tensor("op_164")]; + tensor var_208 = linear(bias = inner_encoder_q_proj_0_bias, weight = inner_encoder_q_proj_0_weight, x = x_1)[name = tensor("linear_3")]; + tensor var_209 = const()[name = tensor("op_209"), val = tensor([1, 3, 4, 64])]; + tensor var_210 = reshape(shape = var_209, x = var_208)[name = tensor("op_210")]; tensor q_1_perm_0 = const()[name = tensor("q_1_perm_0"), val = tensor([0, 2, 1, 3])]; - tensor var_168 = linear(bias = encoder_k_proj_0_bias, weight = encoder_k_proj_0_weight, x = x_1)[name = tensor("linear_4")]; - tensor var_169 = const()[name = tensor("op_169"), val = tensor(0x1p-3)]; - tensor var_170 = mul(x = var_168, y = var_169)[name = tensor("op_170")]; - tensor var_171 = const()[name = tensor("op_171"), val = tensor([1, 3, 4, 64])]; - tensor var_172 = reshape(shape = var_171, x = var_170)[name = tensor("op_172")]; + tensor var_214 = linear(bias = inner_encoder_k_proj_0_bias, weight = inner_encoder_k_proj_0_weight, x = x_1)[name = tensor("linear_4")]; + tensor var_215 = const()[name = tensor("op_215"), val = tensor(0x1p-3)]; + tensor var_216 = mul(x = var_214, y = var_215)[name = tensor("op_216")]; + tensor var_217 = const()[name = tensor("op_217"), val = tensor([1, 3, 4, 64])]; + tensor var_218 = reshape(shape = var_217, x = var_216)[name = tensor("op_218")]; tensor k_1_perm_0 = const()[name = tensor("k_1_perm_0"), val = tensor([0, 2, 1, 3])]; - tensor var_176 = linear(bias = encoder_v_proj_0_bias, weight = encoder_v_proj_0_weight, x = x_1)[name = tensor("linear_5")]; - tensor var_177 = const()[name = tensor("op_177"), val = tensor([1, 3, 4, 64])]; - tensor var_178 = reshape(shape = var_177, x = var_176)[name = tensor("op_178")]; + tensor var_222 = linear(bias = inner_encoder_v_proj_0_bias, weight = inner_encoder_v_proj_0_weight, x = x_1)[name = tensor("linear_5")]; + tensor var_223 = const()[name = tensor("op_223"), val = tensor([1, 3, 4, 64])]; + tensor var_224 = reshape(shape = var_223, x = var_222)[name = tensor("op_224")]; tensor v_1_perm_0 = const()[name = tensor("v_1_perm_0"), val = tensor([0, 2, 1, 3])]; - tensor input_17 = linear(bias = encoder_g_proj_0_bias, weight = encoder_g_proj_0_weight, x = x_1)[name = tensor("linear_6")]; + tensor input_19 = linear(bias = inner_encoder_g_proj_0_bias, weight = inner_encoder_g_proj_0_weight, x = x_1)[name = tensor("linear_6")]; tensor sqrt_s0_1 = sqrt(x = prev_scale_1)[name = tensor("sqrt_s0_1")]; - tensor s_t_1 = add(x = prev_scale_1, y = encoder__t_index)[name = tensor("s_t_1")]; + tensor s_t_1 = add(x = prev_scale_1, y = inner_encoder__t_index)[name = tensor("s_t_1")]; tensor sqrt_s_t_1 = sqrt(x = s_t_1)[name = tensor("sqrt_s_t_1")]; tensor qk_1_transpose_x_1 = const()[name = tensor("qk_1_transpose_x_1"), val = tensor(false)]; tensor qk_1_transpose_y_1 = const()[name = tensor("qk_1_transpose_y_1"), val = tensor(true)]; - tensor k_1 = transpose(perm = k_1_perm_0, x = var_172)[name = tensor("transpose_57")]; - tensor q_1 = transpose(perm = q_1_perm_0, x = var_164)[name = tensor("transpose_58")]; + tensor k_1 = transpose(perm = k_1_perm_0, x = var_218)[name = tensor("transpose_57")]; + tensor q_1 = transpose(perm = q_1_perm_0, x = var_210)[name = tensor("transpose_58")]; tensor qk_1 = matmul(transpose_x = qk_1_transpose_x_1, transpose_y = qk_1_transpose_y_1, x = q_1, y = k_1)[name = tensor("qk_1")]; - tensor var_188 = const()[name = tensor("op_188"), val = tensor([3, 1])]; - tensor var_189 = reshape(shape = var_188, x = sqrt_s_t_1)[name = tensor("op_189")]; - tensor M_1 = real_div(x = encoder__causal_mask, y = var_189)[name = tensor("M_1")]; - tensor var_191 = mul(x = qk_1, y = M_1)[name = tensor("op_191")]; + tensor var_234 = const()[name = tensor("op_234"), val = tensor([3, 1])]; + tensor var_235 = reshape(shape = var_234, x = sqrt_s_t_1)[name = tensor("op_235")]; + tensor M_1 = real_div(x = inner_encoder__causal_mask, y = var_235)[name = tensor("M_1")]; + tensor var_237 = mul(x = qk_1, y = M_1)[name = tensor("op_237")]; tensor inner_1_transpose_x_0 = const()[name = tensor("inner_1_transpose_x_0"), val = tensor(false)]; tensor inner_1_transpose_y_0 = const()[name = tensor("inner_1_transpose_y_0"), val = tensor(false)]; - tensor v_1 = transpose(perm = v_1_perm_0, x = var_178)[name = tensor("transpose_56")]; - tensor inner_1 = matmul(transpose_x = inner_1_transpose_x_0, transpose_y = inner_1_transpose_y_0, x = var_191, y = v_1)[name = tensor("inner_1")]; - tensor var_193_transpose_x_0 = const()[name = tensor("op_193_transpose_x_0"), val = tensor(false)]; - tensor var_193_transpose_y_0 = const()[name = tensor("op_193_transpose_y_0"), val = tensor(false)]; - tensor var_193 = matmul(transpose_x = var_193_transpose_x_0, transpose_y = var_193_transpose_y_0, x = q_1, y = prev_kv_1)[name = tensor("op_193")]; - tensor var_194 = real_div(x = sqrt_s0_1, y = sqrt_s_t_1)[name = tensor("op_194")]; - tensor var_195 = const()[name = tensor("op_195"), val = tensor([1, 1, 3, 1])]; - tensor var_196 = reshape(shape = var_195, x = var_194)[name = tensor("op_196")]; - tensor cross_1 = mul(x = var_193, y = var_196)[name = tensor("cross_1")]; + tensor v_1 = transpose(perm = v_1_perm_0, x = var_224)[name = tensor("transpose_56")]; + tensor inner_1 = matmul(transpose_x = inner_1_transpose_x_0, transpose_y = inner_1_transpose_y_0, x = var_237, y = v_1)[name = tensor("inner_1")]; + tensor var_239_transpose_x_0 = const()[name = tensor("op_239_transpose_x_0"), val = tensor(false)]; + tensor var_239_transpose_y_0 = const()[name = tensor("op_239_transpose_y_0"), val = tensor(false)]; + tensor var_239 = matmul(transpose_x = var_239_transpose_x_0, transpose_y = var_239_transpose_y_0, x = q_1, y = prev_kv_1)[name = tensor("op_239")]; + tensor var_240 = real_div(x = sqrt_s0_1, y = sqrt_s_t_1)[name = tensor("op_240")]; + tensor var_241 = const()[name = tensor("op_241"), val = tensor([1, 1, 3, 1])]; + tensor var_242 = reshape(shape = var_241, x = var_240)[name = tensor("op_242")]; + tensor cross_1 = mul(x = var_239, y = var_242)[name = tensor("cross_1")]; tensor out_1 = add(x = inner_1, y = cross_1)[name = tensor("out_1")]; - tensor var_199 = mul(x = prev_kv_1, y = sqrt_s0_1)[name = tensor("op_199")]; - tensor var_201_transpose_x_1 = const()[name = tensor("op_201_transpose_x_1"), val = tensor(true)]; - tensor var_201_transpose_y_1 = const()[name = tensor("op_201_transpose_y_1"), val = tensor(false)]; - tensor var_201 = matmul(transpose_x = var_201_transpose_x_1, transpose_y = var_201_transpose_y_1, x = k_1, y = v_1)[name = tensor("op_201")]; - tensor new_kv_unnorm_1 = add(x = var_199, y = var_201)[name = tensor("new_kv_unnorm_1")]; - tensor var_203 = const()[name = tensor("op_203"), val = tensor(0x1.8p+1)]; - tensor new_scale_1 = add(x = prev_scale_1, y = var_203)[name = tensor("new_scale_1")]; - tensor var_205 = sqrt(x = new_scale_1)[name = tensor("op_205")]; - tensor var_206 = real_div(x = new_kv_unnorm_1, y = var_205)[name = tensor("op_206")]; - tensor var_207_perm_0 = const()[name = tensor("op_207_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_245 = mul(x = prev_kv_1, y = sqrt_s0_1)[name = tensor("op_245")]; + tensor var_247_transpose_x_1 = const()[name = tensor("op_247_transpose_x_1"), val = tensor(true)]; + tensor var_247_transpose_y_1 = const()[name = tensor("op_247_transpose_y_1"), val = tensor(false)]; + tensor var_247 = matmul(transpose_x = var_247_transpose_x_1, transpose_y = var_247_transpose_y_1, x = k_1, y = v_1)[name = tensor("op_247")]; + tensor new_kv_unnorm_1 = add(x = var_245, y = var_247)[name = tensor("new_kv_unnorm_1")]; + tensor var_249 = const()[name = tensor("op_249"), val = tensor(0x1.8p+1)]; + tensor new_scale_1 = add(x = prev_scale_1, y = var_249)[name = tensor("new_scale_1")]; + tensor var_251 = sqrt(x = new_scale_1)[name = tensor("op_251")]; + tensor var_252 = real_div(x = new_kv_unnorm_1, y = var_251)[name = tensor("op_252")]; + tensor var_253_perm_0 = const()[name = tensor("op_253_perm_0"), val = tensor([0, 2, 1, 3])]; tensor out_3_axes_0 = const()[name = tensor("out_3_axes_0"), val = tensor([-1])]; - tensor var_207 = transpose(perm = var_207_perm_0, x = out_1)[name = tensor("transpose_55")]; - tensor out_3 = layer_norm(axes = out_3_axes_0, epsilon = var_17, x = var_207)[name = tensor("out_3")]; - tensor var_211 = const()[name = tensor("op_211"), val = tensor([1, 3, 256])]; - tensor out_5 = reshape(shape = var_211, x = out_3)[name = tensor("out_5")]; - tensor var_213 = silu(x = input_17)[name = tensor("op_213")]; - tensor input_19 = mul(x = var_213, y = out_5)[name = tensor("input_19")]; - tensor ret_out_1 = linear(bias = encoder_out_proj_0_bias, weight = encoder_out_proj_0_weight, x = input_19)[name = tensor("linear_7")]; - tensor x_3 = add(x = input_13, y = ret_out_1)[name = tensor("x_3")]; + tensor var_253 = transpose(perm = var_253_perm_0, x = out_1)[name = tensor("transpose_55")]; + tensor out_3 = layer_norm(axes = out_3_axes_0, epsilon = var_63, x = var_253)[name = tensor("out_3")]; + tensor var_257 = const()[name = tensor("op_257"), val = tensor([1, 3, 256])]; + tensor out_5 = reshape(shape = var_257, x = out_3)[name = tensor("out_5")]; + tensor var_259 = silu(x = input_19)[name = tensor("op_259")]; + tensor input_21 = mul(x = var_259, y = out_5)[name = tensor("input_21")]; + tensor ret_out_1 = linear(bias = inner_encoder_out_proj_0_bias, weight = inner_encoder_out_proj_0_weight, x = input_21)[name = tensor("linear_7")]; + tensor x_3 = add(x = input_15, y = ret_out_1)[name = tensor("x_3")]; tensor window_1_begin_0 = const()[name = tensor("window_1_begin_0"), val = tensor([0, 0, 0, 0])]; tensor window_1_end_0 = const()[name = tensor("window_1_end_0"), val = tensor([1, 1, 16, 256])]; tensor window_1_end_mask_0 = const()[name = tensor("window_1_end_mask_0"), val = tensor([false, true, true, true])]; tensor window_1_squeeze_mask_0 = const()[name = tensor("window_1_squeeze_mask_0"), val = tensor([true, false, false, false])]; tensor window_1 = slice_by_index(begin = window_1_begin_0, end = window_1_end_0, end_mask = window_1_end_mask_0, squeeze_mask = window_1_squeeze_mask_0, x = enc_conv_cache)[name = tensor("window_1")]; - tensor var_221_begin_0 = const()[name = tensor("op_221_begin_0"), val = tensor([0, 0, 0])]; - tensor var_221_end_0 = const()[name = tensor("op_221_end_0"), val = tensor([1, 1, 256])]; - tensor var_221_end_mask_0 = const()[name = tensor("op_221_end_mask_0"), val = tensor([true, false, true])]; - tensor var_221 = slice_by_index(begin = var_221_begin_0, end = var_221_end_0, end_mask = var_221_end_mask_0, x = x_3)[name = tensor("op_221")]; - tensor var_224_begin_0 = const()[name = tensor("op_224_begin_0"), val = tensor([0, 1, 0])]; - tensor var_224_end_0 = const()[name = tensor("op_224_end_0"), val = tensor([1, 16, 256])]; - tensor var_224_end_mask_0 = const()[name = tensor("op_224_end_mask_0"), val = tensor([true, true, true])]; - tensor var_224 = slice_by_index(begin = var_224_begin_0, end = var_224_end_0, end_mask = var_224_end_mask_0, x = window_1)[name = tensor("op_224")]; + tensor var_267_begin_0 = const()[name = tensor("op_267_begin_0"), val = tensor([0, 0, 0])]; + tensor var_267_end_0 = const()[name = tensor("op_267_end_0"), val = tensor([1, 1, 256])]; + tensor var_267_end_mask_0 = const()[name = tensor("op_267_end_mask_0"), val = tensor([true, false, true])]; + tensor var_267 = slice_by_index(begin = var_267_begin_0, end = var_267_end_0, end_mask = var_267_end_mask_0, x = x_3)[name = tensor("op_267")]; + tensor var_270_begin_0 = const()[name = tensor("op_270_begin_0"), val = tensor([0, 1, 0])]; + tensor var_270_end_0 = const()[name = tensor("op_270_end_0"), val = tensor([1, 16, 256])]; + tensor var_270_end_mask_0 = const()[name = tensor("op_270_end_mask_0"), val = tensor([true, true, true])]; + tensor var_270 = slice_by_index(begin = var_270_begin_0, end = var_270_end_0, end_mask = var_270_end_mask_0, x = window_1)[name = tensor("op_270")]; tensor window_3_interleave_0 = const()[name = tensor("window_3_interleave_0"), val = tensor(false)]; - tensor window_3 = concat(axis = var_26, interleave = window_3_interleave_0, values = (var_224, var_221))[name = tensor("window_3")]; - tensor var_229_begin_0 = const()[name = tensor("op_229_begin_0"), val = tensor([0, 1, 0])]; - tensor var_229_end_0 = const()[name = tensor("op_229_end_0"), val = tensor([1, 2, 256])]; - tensor var_229_end_mask_0 = const()[name = tensor("op_229_end_mask_0"), val = tensor([true, false, true])]; - tensor var_229 = slice_by_index(begin = var_229_begin_0, end = var_229_end_0, end_mask = var_229_end_mask_0, x = x_3)[name = tensor("op_229")]; - tensor var_232_begin_0 = const()[name = tensor("op_232_begin_0"), val = tensor([0, 1, 0])]; - tensor var_232_end_0 = const()[name = tensor("op_232_end_0"), val = tensor([1, 16, 256])]; - tensor var_232_end_mask_0 = const()[name = tensor("op_232_end_mask_0"), val = tensor([true, true, true])]; - tensor var_232 = slice_by_index(begin = var_232_begin_0, end = var_232_end_0, end_mask = var_232_end_mask_0, x = window_3)[name = tensor("op_232")]; + tensor window_3 = concat(axis = var_72, interleave = window_3_interleave_0, values = (var_270, var_267))[name = tensor("window_3")]; + tensor var_275_begin_0 = const()[name = tensor("op_275_begin_0"), val = tensor([0, 1, 0])]; + tensor var_275_end_0 = const()[name = tensor("op_275_end_0"), val = tensor([1, 2, 256])]; + tensor var_275_end_mask_0 = const()[name = tensor("op_275_end_mask_0"), val = tensor([true, false, true])]; + tensor var_275 = slice_by_index(begin = var_275_begin_0, end = var_275_end_0, end_mask = var_275_end_mask_0, x = x_3)[name = tensor("op_275")]; + tensor var_278_begin_0 = const()[name = tensor("op_278_begin_0"), val = tensor([0, 1, 0])]; + tensor var_278_end_0 = const()[name = tensor("op_278_end_0"), val = tensor([1, 16, 256])]; + tensor var_278_end_mask_0 = const()[name = tensor("op_278_end_mask_0"), val = tensor([true, true, true])]; + tensor var_278 = slice_by_index(begin = var_278_begin_0, end = var_278_end_0, end_mask = var_278_end_mask_0, x = window_3)[name = tensor("op_278")]; tensor window_5_interleave_0 = const()[name = tensor("window_5_interleave_0"), val = tensor(false)]; - tensor window_5 = concat(axis = var_26, interleave = window_5_interleave_0, values = (var_232, var_229))[name = tensor("window_5")]; - tensor var_237_begin_0 = const()[name = tensor("op_237_begin_0"), val = tensor([0, 2, 0])]; - tensor var_237_end_0 = const()[name = tensor("op_237_end_0"), val = tensor([1, 1, 256])]; - tensor var_237_end_mask_0 = const()[name = tensor("op_237_end_mask_0"), val = tensor([true, true, true])]; - tensor var_237 = slice_by_index(begin = var_237_begin_0, end = var_237_end_0, end_mask = var_237_end_mask_0, x = x_3)[name = tensor("op_237")]; - tensor var_240_begin_0 = const()[name = tensor("op_240_begin_0"), val = tensor([0, 1, 0])]; - tensor var_240_end_0 = const()[name = tensor("op_240_end_0"), val = tensor([1, 16, 256])]; - tensor var_240_end_mask_0 = const()[name = tensor("op_240_end_mask_0"), val = tensor([true, true, true])]; - tensor var_240 = slice_by_index(begin = var_240_begin_0, end = var_240_end_0, end_mask = var_240_end_mask_0, x = window_5)[name = tensor("op_240")]; + tensor window_5 = concat(axis = var_72, interleave = window_5_interleave_0, values = (var_278, var_275))[name = tensor("window_5")]; + tensor var_283_begin_0 = const()[name = tensor("op_283_begin_0"), val = tensor([0, 2, 0])]; + tensor var_283_end_0 = const()[name = tensor("op_283_end_0"), val = tensor([1, 1, 256])]; + tensor var_283_end_mask_0 = const()[name = tensor("op_283_end_mask_0"), val = tensor([true, true, true])]; + tensor var_283 = slice_by_index(begin = var_283_begin_0, end = var_283_end_0, end_mask = var_283_end_mask_0, x = x_3)[name = tensor("op_283")]; + tensor var_286_begin_0 = const()[name = tensor("op_286_begin_0"), val = tensor([0, 1, 0])]; + tensor var_286_end_0 = const()[name = tensor("op_286_end_0"), val = tensor([1, 16, 256])]; + tensor var_286_end_mask_0 = const()[name = tensor("op_286_end_mask_0"), val = tensor([true, true, true])]; + tensor var_286 = slice_by_index(begin = var_286_begin_0, end = var_286_end_0, end_mask = var_286_end_mask_0, x = window_5)[name = tensor("op_286")]; tensor window_7_interleave_0 = const()[name = tensor("window_7_interleave_0"), val = tensor(false)]; - tensor window_7 = concat(axis = var_26, interleave = window_7_interleave_0, values = (var_240, var_237))[name = tensor("window_7")]; - tensor input_21_interleave_0 = const()[name = tensor("input_21_interleave_0"), val = tensor(false)]; - tensor input_21 = concat(axis = var_23, interleave = input_21_interleave_0, values = (window_3, window_5, window_7))[name = tensor("input_21")]; + tensor window_7 = concat(axis = var_72, interleave = window_7_interleave_0, values = (var_286, var_283))[name = tensor("window_7")]; + tensor input_23_interleave_0 = const()[name = tensor("input_23_interleave_0"), val = tensor(false)]; + tensor input_23 = concat(axis = var_58, interleave = input_23_interleave_0, values = (window_3, window_5, window_7))[name = tensor("input_23")]; tensor x_5_axes_0 = const()[name = tensor("x_5_axes_0"), val = tensor([-1])]; - tensor x_5 = layer_norm(axes = x_5_axes_0, beta = encoder_conv_module_0_sequential_0_bias, epsilon = var_28, gamma = encoder_conv_module_0_sequential_0_weight, x = input_21)[name = tensor("x_5")]; - tensor input_23_perm_0 = const()[name = tensor("input_23_perm_0"), val = tensor([0, 2, 1])]; + tensor x_5 = layer_norm(axes = x_5_axes_0, beta = inner_encoder_conv_module_0_sequential_0_bias, epsilon = var_55, gamma = inner_encoder_conv_module_0_sequential_0_weight, x = input_23)[name = tensor("x_5")]; + tensor input_25_perm_0 = const()[name = tensor("input_25_perm_0"), val = tensor([0, 2, 1])]; tensor inputs_3_pad_type_0 = const()[name = tensor("inputs_3_pad_type_0"), val = tensor("valid")]; tensor inputs_3_strides_0 = const()[name = tensor("inputs_3_strides_0"), val = tensor([1])]; tensor inputs_3_pad_0 = const()[name = tensor("inputs_3_pad_0"), val = tensor([0, 0])]; tensor inputs_3_dilations_0 = const()[name = tensor("inputs_3_dilations_0"), val = tensor([1])]; tensor inputs_3_groups_0 = const()[name = tensor("inputs_3_groups_0"), val = tensor(1)]; - tensor input_23 = transpose(perm = input_23_perm_0, x = x_5)[name = tensor("transpose_54")]; - tensor inputs_3 = conv(bias = encoder_conv_module_0_sequential_2_conv_bias, dilations = inputs_3_dilations_0, groups = inputs_3_groups_0, pad = inputs_3_pad_0, pad_type = inputs_3_pad_type_0, strides = inputs_3_strides_0, weight = encoder_conv_module_0_sequential_2_conv_weight, x = input_23)[name = tensor("inputs_3")]; - tensor var_265_split_sizes_0 = const()[name = tensor("op_265_split_sizes_0"), val = tensor([256, 256])]; - tensor var_265_axis_0 = const()[name = tensor("op_265_axis_0"), val = tensor(1)]; - tensor var_265_0, tensor var_265_1 = split(axis = var_265_axis_0, split_sizes = var_265_split_sizes_0, x = inputs_3)[name = tensor("op_265")]; - tensor var_267 = sigmoid(x = var_265_1)[name = tensor("op_267")]; - tensor inputs_5 = mul(x = var_265_0, y = var_267)[name = tensor("inputs_5")]; + tensor input_25 = transpose(perm = input_25_perm_0, x = x_5)[name = tensor("transpose_54")]; + tensor inputs_3 = conv(bias = inner_encoder_conv_module_0_sequential_2_conv_bias, dilations = inputs_3_dilations_0, groups = inputs_3_groups_0, pad = inputs_3_pad_0, pad_type = inputs_3_pad_type_0, strides = inputs_3_strides_0, weight = inner_encoder_conv_module_0_sequential_2_conv_weight, x = input_25)[name = tensor("inputs_3")]; + tensor var_311_split_sizes_0 = const()[name = tensor("op_311_split_sizes_0"), val = tensor([256, 256])]; + tensor var_311_axis_0 = const()[name = tensor("op_311_axis_0"), val = tensor(1)]; + tensor var_311_0, tensor var_311_1 = split(axis = var_311_axis_0, split_sizes = var_311_split_sizes_0, x = inputs_3)[name = tensor("op_311")]; + tensor var_313 = sigmoid(x = var_311_1)[name = tensor("op_313")]; + tensor inputs_5 = mul(x = var_311_0, y = var_313)[name = tensor("inputs_5")]; tensor outputs_aug_1_pad_type_0 = const()[name = tensor("outputs_aug_1_pad_type_0"), val = tensor("custom")]; tensor outputs_aug_1_pad_0 = const()[name = tensor("outputs_aug_1_pad_0"), val = tensor([15, 15])]; tensor outputs_aug_1_groups_0 = const()[name = tensor("outputs_aug_1_groups_0"), val = tensor(256)]; tensor outputs_aug_1_strides_0 = const()[name = tensor("outputs_aug_1_strides_0"), val = tensor([1])]; tensor outputs_aug_1_dilations_0 = const()[name = tensor("outputs_aug_1_dilations_0"), val = tensor([1])]; - tensor outputs_aug_1 = conv(dilations = outputs_aug_1_dilations_0, groups = outputs_aug_1_groups_0, pad = outputs_aug_1_pad_0, pad_type = outputs_aug_1_pad_type_0, strides = outputs_aug_1_strides_0, weight = encoder_conv_module_0_sequential_4_conv_weight, x = inputs_5)[name = tensor("outputs_aug_1")]; - tensor input_25_begin_0 = const()[name = tensor("input_25_begin_0"), val = tensor([0, 0, 0])]; - tensor input_25_end_0 = const()[name = tensor("input_25_end_0"), val = tensor([3, 256, 16])]; - tensor input_25_end_mask_0 = const()[name = tensor("input_25_end_mask_0"), val = tensor([true, true, false])]; - tensor input_25 = slice_by_index(begin = input_25_begin_0, end = input_25_end_0, end_mask = input_25_end_mask_0, x = outputs_aug_1)[name = tensor("input_25")]; - tensor inputs_7 = batch_norm(beta = encoder_conv_module_0_sequential_5_bias, epsilon = var_28, gamma = encoder_conv_module_0_sequential_5_weight, mean = encoder_conv_module_0_sequential_5_running_mean, variance = encoder_conv_module_0_sequential_5_running_var, x = input_25)[name = tensor("inputs_7")]; - tensor input_27 = silu(x = inputs_7)[name = tensor("input_27")]; - tensor input_29_pad_type_0 = const()[name = tensor("input_29_pad_type_0"), val = tensor("valid")]; - tensor input_29_strides_0 = const()[name = tensor("input_29_strides_0"), val = tensor([1])]; - tensor input_29_pad_0 = const()[name = tensor("input_29_pad_0"), val = tensor([0, 0])]; - tensor input_29_dilations_0 = const()[name = tensor("input_29_dilations_0"), val = tensor([1])]; - tensor input_29_groups_0 = const()[name = tensor("input_29_groups_0"), val = tensor(1)]; - tensor input_29 = conv(bias = encoder_conv_module_0_sequential_7_conv_bias, dilations = input_29_dilations_0, groups = input_29_groups_0, pad = input_29_pad_0, pad_type = input_29_pad_type_0, strides = input_29_strides_0, weight = encoder_conv_module_0_sequential_7_conv_weight, x = input_27)[name = tensor("input_29")]; + tensor outputs_aug_1 = conv(dilations = outputs_aug_1_dilations_0, groups = outputs_aug_1_groups_0, pad = outputs_aug_1_pad_0, pad_type = outputs_aug_1_pad_type_0, strides = outputs_aug_1_strides_0, weight = inner_encoder_conv_module_0_sequential_4_conv_weight, x = inputs_5)[name = tensor("outputs_aug_1")]; + tensor input_27_begin_0 = const()[name = tensor("input_27_begin_0"), val = tensor([0, 0, 0])]; + tensor input_27_end_0 = const()[name = tensor("input_27_end_0"), val = tensor([3, 256, 16])]; + tensor input_27_end_mask_0 = const()[name = tensor("input_27_end_mask_0"), val = tensor([true, true, false])]; + tensor input_27 = slice_by_index(begin = input_27_begin_0, end = input_27_end_0, end_mask = input_27_end_mask_0, x = outputs_aug_1)[name = tensor("input_27")]; + tensor inputs_7 = batch_norm(beta = inner_encoder_conv_module_0_sequential_5_bias, epsilon = var_55, gamma = inner_encoder_conv_module_0_sequential_5_weight, mean = inner_encoder_conv_module_0_sequential_5_running_mean, variance = inner_encoder_conv_module_0_sequential_5_running_var, x = input_27)[name = tensor("inputs_7")]; + tensor input_29 = silu(x = inputs_7)[name = tensor("input_29")]; + tensor input_31_pad_type_0 = const()[name = tensor("input_31_pad_type_0"), val = tensor("valid")]; + tensor input_31_strides_0 = const()[name = tensor("input_31_strides_0"), val = tensor([1])]; + tensor input_31_pad_0 = const()[name = tensor("input_31_pad_0"), val = tensor([0, 0])]; + tensor input_31_dilations_0 = const()[name = tensor("input_31_dilations_0"), val = tensor([1])]; + tensor input_31_groups_0 = const()[name = tensor("input_31_groups_0"), val = tensor(1)]; + tensor input_31 = conv(bias = inner_encoder_conv_module_0_sequential_7_conv_bias, dilations = input_31_dilations_0, groups = input_31_groups_0, pad = input_31_pad_0, pad_type = input_31_pad_type_0, strides = input_31_strides_0, weight = inner_encoder_conv_module_0_sequential_7_conv_weight, x = input_29)[name = tensor("input_31")]; tensor conv_out_1_perm_0 = const()[name = tensor("conv_out_1_perm_0"), val = tensor([0, 2, 1])]; - tensor var_298_begin_0 = const()[name = tensor("op_298_begin_0"), val = tensor([0, -1, 0])]; - tensor var_298_end_0 = const()[name = tensor("op_298_end_0"), val = tensor([3, 16, 256])]; - tensor var_298_end_mask_0 = const()[name = tensor("op_298_end_mask_0"), val = tensor([true, true, true])]; - tensor conv_out_1 = transpose(perm = conv_out_1_perm_0, x = input_29)[name = tensor("transpose_53")]; - tensor var_298 = slice_by_index(begin = var_298_begin_0, end = var_298_end_0, end_mask = var_298_end_mask_0, x = conv_out_1)[name = tensor("op_298")]; - tensor var_300_perm_0 = const()[name = tensor("op_300_perm_0"), val = tensor([1, 0, 2])]; - tensor var_300 = transpose(perm = var_300_perm_0, x = var_298)[name = tensor("transpose_52")]; - tensor input_31 = add(x = x_3, y = var_300)[name = tensor("input_31")]; - tensor input_33_axes_0 = const()[name = tensor("input_33_axes_0"), val = tensor([-1])]; - tensor input_33 = layer_norm(axes = input_33_axes_0, beta = encoder_ffn2_0_module_sequential_0_bias, epsilon = var_28, gamma = encoder_ffn2_0_module_sequential_0_weight, x = input_31)[name = tensor("input_33")]; - tensor inputs_9 = linear(bias = encoder_ffn2_0_module_sequential_1_linear_bias, weight = encoder_ffn2_0_module_sequential_1_linear_weight, x = input_33)[name = tensor("linear_8")]; - tensor input_35 = silu(x = inputs_9)[name = tensor("input_35")]; - tensor input_39 = linear(bias = encoder_ffn2_0_module_sequential_4_linear_bias, weight = encoder_ffn2_0_module_sequential_4_linear_weight, x = input_35)[name = tensor("linear_9")]; - tensor var_323 = const()[name = tensor("op_323"), val = tensor(0x1p-1)]; - tensor var_324 = mul(x = input_39, y = var_323)[name = tensor("op_324")]; - tensor input_41 = add(x = var_324, y = input_31)[name = tensor("input_41")]; - tensor input_43_axes_0 = const()[name = tensor("input_43_axes_0"), val = tensor([-1])]; - tensor input_43 = layer_norm(axes = input_43_axes_0, beta = encoder_layer_norm_0_bias, epsilon = var_28, gamma = encoder_layer_norm_0_weight, x = input_41)[name = tensor("input_43")]; + tensor var_344_begin_0 = const()[name = tensor("op_344_begin_0"), val = tensor([0, -1, 0])]; + tensor var_344_end_0 = const()[name = tensor("op_344_end_0"), val = tensor([3, 16, 256])]; + tensor var_344_end_mask_0 = const()[name = tensor("op_344_end_mask_0"), val = tensor([true, true, true])]; + tensor conv_out_1 = transpose(perm = conv_out_1_perm_0, x = input_31)[name = tensor("transpose_53")]; + tensor var_344 = slice_by_index(begin = var_344_begin_0, end = var_344_end_0, end_mask = var_344_end_mask_0, x = conv_out_1)[name = tensor("op_344")]; + tensor var_346_perm_0 = const()[name = tensor("op_346_perm_0"), val = tensor([1, 0, 2])]; + tensor var_346 = transpose(perm = var_346_perm_0, x = var_344)[name = tensor("transpose_52")]; + tensor input_33 = add(x = x_3, y = var_346)[name = tensor("input_33")]; + tensor input_35_axes_0 = const()[name = tensor("input_35_axes_0"), val = tensor([-1])]; + tensor input_35 = layer_norm(axes = input_35_axes_0, beta = inner_encoder_ffn2_0_module_sequential_0_bias, epsilon = var_55, gamma = inner_encoder_ffn2_0_module_sequential_0_weight, x = input_33)[name = tensor("input_35")]; + tensor inputs_9 = linear(bias = inner_encoder_ffn2_0_module_sequential_1_linear_bias, weight = inner_encoder_ffn2_0_module_sequential_1_linear_weight, x = input_35)[name = tensor("linear_8")]; + tensor input_37 = silu(x = inputs_9)[name = tensor("input_37")]; + tensor input_41 = linear(bias = inner_encoder_ffn2_0_module_sequential_4_linear_bias, weight = inner_encoder_ffn2_0_module_sequential_4_linear_weight, x = input_37)[name = tensor("linear_9")]; + tensor var_369 = const()[name = tensor("op_369"), val = tensor(0x1p-1)]; + tensor var_370 = mul(x = input_41, y = var_369)[name = tensor("op_370")]; + tensor input_43 = add(x = var_370, y = input_33)[name = tensor("input_43")]; tensor input_45_axes_0 = const()[name = tensor("input_45_axes_0"), val = tensor([-1])]; - tensor input_45 = layer_norm(axes = input_45_axes_0, beta = encoder_ffn1_1_module_sequential_0_bias, epsilon = var_28, gamma = encoder_ffn1_1_module_sequential_0_weight, x = input_43)[name = tensor("input_45")]; - tensor inputs_11 = linear(bias = encoder_ffn1_1_module_sequential_1_linear_bias, weight = encoder_ffn1_1_module_sequential_1_linear_weight, x = input_45)[name = tensor("linear_10")]; - tensor input_47 = silu(x = inputs_11)[name = tensor("input_47")]; - tensor input_51 = linear(bias = encoder_ffn1_1_module_sequential_4_linear_bias, weight = encoder_ffn1_1_module_sequential_4_linear_weight, x = input_47)[name = tensor("linear_11")]; - tensor var_353 = const()[name = tensor("op_353"), val = tensor(0x1p-1)]; - tensor var_354 = mul(x = input_51, y = var_353)[name = tensor("op_354")]; - tensor input_53 = add(x = var_354, y = input_43)[name = tensor("input_53")]; + tensor input_45 = layer_norm(axes = input_45_axes_0, beta = inner_encoder_layer_norm_0_bias, epsilon = var_55, gamma = inner_encoder_layer_norm_0_weight, x = input_43)[name = tensor("input_45")]; + tensor input_47_axes_0 = const()[name = tensor("input_47_axes_0"), val = tensor([-1])]; + tensor input_47 = layer_norm(axes = input_47_axes_0, beta = inner_encoder_ffn1_1_module_sequential_0_bias, epsilon = var_55, gamma = inner_encoder_ffn1_1_module_sequential_0_weight, x = input_45)[name = tensor("input_47")]; + tensor inputs_11 = linear(bias = inner_encoder_ffn1_1_module_sequential_1_linear_bias, weight = inner_encoder_ffn1_1_module_sequential_1_linear_weight, x = input_47)[name = tensor("linear_10")]; + tensor input_49 = silu(x = inputs_11)[name = tensor("input_49")]; + tensor input_53 = linear(bias = inner_encoder_ffn1_1_module_sequential_4_linear_bias, weight = inner_encoder_ffn1_1_module_sequential_4_linear_weight, x = input_49)[name = tensor("linear_11")]; + tensor var_399 = const()[name = tensor("op_399"), val = tensor(0x1p-1)]; + tensor var_400 = mul(x = input_53, y = var_399)[name = tensor("op_400")]; + tensor input_55 = add(x = var_400, y = input_45)[name = tensor("input_55")]; tensor x_7_axes_0 = const()[name = tensor("x_7_axes_0"), val = tensor([-1])]; - tensor x_7 = layer_norm(axes = x_7_axes_0, beta = encoder_ret_lns_1_bias, epsilon = var_28, gamma = encoder_ret_lns_1_weight, x = input_53)[name = tensor("x_7")]; + tensor x_7 = layer_norm(axes = x_7_axes_0, beta = inner_encoder_ret_lns_1_bias, epsilon = var_55, gamma = inner_encoder_ret_lns_1_weight, x = input_55)[name = tensor("x_7")]; tensor prev_kv_3_begin_0 = const()[name = tensor("prev_kv_3_begin_0"), val = tensor([1, 0, 0, 0, 0])]; tensor prev_kv_3_end_0 = const()[name = tensor("prev_kv_3_end_0"), val = tensor([2, 1, 4, 64, 64])]; tensor prev_kv_3_end_mask_0 = const()[name = tensor("prev_kv_3_end_mask_0"), val = tensor([false, true, true, true, true])]; @@ -406,163 +424,163 @@ program(1.0) tensor prev_scale_3_end_mask_0 = const()[name = tensor("prev_scale_3_end_mask_0"), val = tensor([false, true])]; tensor prev_scale_3_squeeze_mask_0 = const()[name = tensor("prev_scale_3_squeeze_mask_0"), val = tensor([true, false])]; tensor prev_scale_3 = slice_by_index(begin = prev_scale_3_begin_0, end = prev_scale_3_end_0, end_mask = prev_scale_3_end_mask_0, squeeze_mask = prev_scale_3_squeeze_mask_0, x = enc_scale)[name = tensor("prev_scale_3")]; - tensor var_368 = linear(bias = encoder_q_proj_1_bias, weight = encoder_q_proj_1_weight, x = x_7)[name = tensor("linear_12")]; - tensor var_369 = const()[name = tensor("op_369"), val = tensor([1, 3, 4, 64])]; - tensor var_370 = reshape(shape = var_369, x = var_368)[name = tensor("op_370")]; + tensor var_414 = linear(bias = inner_encoder_q_proj_1_bias, weight = inner_encoder_q_proj_1_weight, x = x_7)[name = tensor("linear_12")]; + tensor var_415 = const()[name = tensor("op_415"), val = tensor([1, 3, 4, 64])]; + tensor var_416 = reshape(shape = var_415, x = var_414)[name = tensor("op_416")]; tensor q_3_perm_0 = const()[name = tensor("q_3_perm_0"), val = tensor([0, 2, 1, 3])]; - tensor var_374 = linear(bias = encoder_k_proj_1_bias, weight = encoder_k_proj_1_weight, x = x_7)[name = tensor("linear_13")]; - tensor var_375 = const()[name = tensor("op_375"), val = tensor(0x1p-3)]; - tensor var_376 = mul(x = var_374, y = var_375)[name = tensor("op_376")]; - tensor var_377 = const()[name = tensor("op_377"), val = tensor([1, 3, 4, 64])]; - tensor var_378 = reshape(shape = var_377, x = var_376)[name = tensor("op_378")]; + tensor var_420 = linear(bias = inner_encoder_k_proj_1_bias, weight = inner_encoder_k_proj_1_weight, x = x_7)[name = tensor("linear_13")]; + tensor var_421 = const()[name = tensor("op_421"), val = tensor(0x1p-3)]; + tensor var_422 = mul(x = var_420, y = var_421)[name = tensor("op_422")]; + tensor var_423 = const()[name = tensor("op_423"), val = tensor([1, 3, 4, 64])]; + tensor var_424 = reshape(shape = var_423, x = var_422)[name = tensor("op_424")]; tensor k_3_perm_0 = const()[name = tensor("k_3_perm_0"), val = tensor([0, 2, 1, 3])]; - tensor var_382 = linear(bias = encoder_v_proj_1_bias, weight = encoder_v_proj_1_weight, x = x_7)[name = tensor("linear_14")]; - tensor var_383 = const()[name = tensor("op_383"), val = tensor([1, 3, 4, 64])]; - tensor var_384 = reshape(shape = var_383, x = var_382)[name = tensor("op_384")]; + tensor var_428 = linear(bias = inner_encoder_v_proj_1_bias, weight = inner_encoder_v_proj_1_weight, x = x_7)[name = tensor("linear_14")]; + tensor var_429 = const()[name = tensor("op_429"), val = tensor([1, 3, 4, 64])]; + tensor var_430 = reshape(shape = var_429, x = var_428)[name = tensor("op_430")]; tensor v_3_perm_0 = const()[name = tensor("v_3_perm_0"), val = tensor([0, 2, 1, 3])]; - tensor input_57 = linear(bias = encoder_g_proj_1_bias, weight = encoder_g_proj_1_weight, x = x_7)[name = tensor("linear_15")]; + tensor input_59 = linear(bias = inner_encoder_g_proj_1_bias, weight = inner_encoder_g_proj_1_weight, x = x_7)[name = tensor("linear_15")]; tensor sqrt_s0_3 = sqrt(x = prev_scale_3)[name = tensor("sqrt_s0_3")]; - tensor s_t_3 = add(x = prev_scale_3, y = encoder__t_index)[name = tensor("s_t_3")]; + tensor s_t_3 = add(x = prev_scale_3, y = inner_encoder__t_index)[name = tensor("s_t_3")]; tensor sqrt_s_t_3 = sqrt(x = s_t_3)[name = tensor("sqrt_s_t_3")]; tensor qk_3_transpose_x_1 = const()[name = tensor("qk_3_transpose_x_1"), val = tensor(false)]; tensor qk_3_transpose_y_1 = const()[name = tensor("qk_3_transpose_y_1"), val = tensor(true)]; - tensor k_3 = transpose(perm = k_3_perm_0, x = var_378)[name = tensor("transpose_50")]; - tensor q_3 = transpose(perm = q_3_perm_0, x = var_370)[name = tensor("transpose_51")]; + tensor k_3 = transpose(perm = k_3_perm_0, x = var_424)[name = tensor("transpose_50")]; + tensor q_3 = transpose(perm = q_3_perm_0, x = var_416)[name = tensor("transpose_51")]; tensor qk_3 = matmul(transpose_x = qk_3_transpose_x_1, transpose_y = qk_3_transpose_y_1, x = q_3, y = k_3)[name = tensor("qk_3")]; - tensor var_394 = const()[name = tensor("op_394"), val = tensor([3, 1])]; - tensor var_395 = reshape(shape = var_394, x = sqrt_s_t_3)[name = tensor("op_395")]; - tensor M_3 = real_div(x = encoder__causal_mask, y = var_395)[name = tensor("M_3")]; - tensor var_397 = mul(x = qk_3, y = M_3)[name = tensor("op_397")]; + tensor var_440 = const()[name = tensor("op_440"), val = tensor([3, 1])]; + tensor var_441 = reshape(shape = var_440, x = sqrt_s_t_3)[name = tensor("op_441")]; + tensor M_3 = real_div(x = inner_encoder__causal_mask, y = var_441)[name = tensor("M_3")]; + tensor var_443 = mul(x = qk_3, y = M_3)[name = tensor("op_443")]; tensor inner_3_transpose_x_0 = const()[name = tensor("inner_3_transpose_x_0"), val = tensor(false)]; tensor inner_3_transpose_y_0 = const()[name = tensor("inner_3_transpose_y_0"), val = tensor(false)]; - tensor v_3 = transpose(perm = v_3_perm_0, x = var_384)[name = tensor("transpose_49")]; - tensor inner_3 = matmul(transpose_x = inner_3_transpose_x_0, transpose_y = inner_3_transpose_y_0, x = var_397, y = v_3)[name = tensor("inner_3")]; - tensor var_399_transpose_x_0 = const()[name = tensor("op_399_transpose_x_0"), val = tensor(false)]; - tensor var_399_transpose_y_0 = const()[name = tensor("op_399_transpose_y_0"), val = tensor(false)]; - tensor var_399 = matmul(transpose_x = var_399_transpose_x_0, transpose_y = var_399_transpose_y_0, x = q_3, y = prev_kv_3)[name = tensor("op_399")]; - tensor var_400 = real_div(x = sqrt_s0_3, y = sqrt_s_t_3)[name = tensor("op_400")]; - tensor var_401 = const()[name = tensor("op_401"), val = tensor([1, 1, 3, 1])]; - tensor var_402 = reshape(shape = var_401, x = var_400)[name = tensor("op_402")]; - tensor cross_3 = mul(x = var_399, y = var_402)[name = tensor("cross_3")]; + tensor v_3 = transpose(perm = v_3_perm_0, x = var_430)[name = tensor("transpose_49")]; + tensor inner_3 = matmul(transpose_x = inner_3_transpose_x_0, transpose_y = inner_3_transpose_y_0, x = var_443, y = v_3)[name = tensor("inner_3")]; + tensor var_445_transpose_x_0 = const()[name = tensor("op_445_transpose_x_0"), val = tensor(false)]; + tensor var_445_transpose_y_0 = const()[name = tensor("op_445_transpose_y_0"), val = tensor(false)]; + tensor var_445 = matmul(transpose_x = var_445_transpose_x_0, transpose_y = var_445_transpose_y_0, x = q_3, y = prev_kv_3)[name = tensor("op_445")]; + tensor var_446 = real_div(x = sqrt_s0_3, y = sqrt_s_t_3)[name = tensor("op_446")]; + tensor var_447 = const()[name = tensor("op_447"), val = tensor([1, 1, 3, 1])]; + tensor var_448 = reshape(shape = var_447, x = var_446)[name = tensor("op_448")]; + tensor cross_3 = mul(x = var_445, y = var_448)[name = tensor("cross_3")]; tensor out_7 = add(x = inner_3, y = cross_3)[name = tensor("out_7")]; - tensor var_405 = mul(x = prev_kv_3, y = sqrt_s0_3)[name = tensor("op_405")]; - tensor var_407_transpose_x_1 = const()[name = tensor("op_407_transpose_x_1"), val = tensor(true)]; - tensor var_407_transpose_y_1 = const()[name = tensor("op_407_transpose_y_1"), val = tensor(false)]; - tensor var_407 = matmul(transpose_x = var_407_transpose_x_1, transpose_y = var_407_transpose_y_1, x = k_3, y = v_3)[name = tensor("op_407")]; - tensor new_kv_unnorm_3 = add(x = var_405, y = var_407)[name = tensor("new_kv_unnorm_3")]; - tensor var_409 = const()[name = tensor("op_409"), val = tensor(0x1.8p+1)]; - tensor new_scale_3 = add(x = prev_scale_3, y = var_409)[name = tensor("new_scale_3")]; - tensor var_411 = sqrt(x = new_scale_3)[name = tensor("op_411")]; - tensor var_412 = real_div(x = new_kv_unnorm_3, y = var_411)[name = tensor("op_412")]; - tensor var_413_perm_0 = const()[name = tensor("op_413_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_451 = mul(x = prev_kv_3, y = sqrt_s0_3)[name = tensor("op_451")]; + tensor var_453_transpose_x_1 = const()[name = tensor("op_453_transpose_x_1"), val = tensor(true)]; + tensor var_453_transpose_y_1 = const()[name = tensor("op_453_transpose_y_1"), val = tensor(false)]; + tensor var_453 = matmul(transpose_x = var_453_transpose_x_1, transpose_y = var_453_transpose_y_1, x = k_3, y = v_3)[name = tensor("op_453")]; + tensor new_kv_unnorm_3 = add(x = var_451, y = var_453)[name = tensor("new_kv_unnorm_3")]; + tensor var_455 = const()[name = tensor("op_455"), val = tensor(0x1.8p+1)]; + tensor new_scale_3 = add(x = prev_scale_3, y = var_455)[name = tensor("new_scale_3")]; + tensor var_457 = sqrt(x = new_scale_3)[name = tensor("op_457")]; + tensor var_458 = real_div(x = new_kv_unnorm_3, y = var_457)[name = tensor("op_458")]; + tensor var_459_perm_0 = const()[name = tensor("op_459_perm_0"), val = tensor([0, 2, 1, 3])]; tensor out_9_axes_0 = const()[name = tensor("out_9_axes_0"), val = tensor([-1])]; - tensor var_413 = transpose(perm = var_413_perm_0, x = out_7)[name = tensor("transpose_48")]; - tensor out_9 = layer_norm(axes = out_9_axes_0, epsilon = var_17, x = var_413)[name = tensor("out_9")]; - tensor var_417 = const()[name = tensor("op_417"), val = tensor([1, 3, 256])]; - tensor out_11 = reshape(shape = var_417, x = out_9)[name = tensor("out_11")]; - tensor var_419 = silu(x = input_57)[name = tensor("op_419")]; - tensor input_59 = mul(x = var_419, y = out_11)[name = tensor("input_59")]; - tensor ret_out_3 = linear(bias = encoder_out_proj_1_bias, weight = encoder_out_proj_1_weight, x = input_59)[name = tensor("linear_16")]; - tensor x_9 = add(x = input_53, y = ret_out_3)[name = tensor("x_9")]; + tensor var_459 = transpose(perm = var_459_perm_0, x = out_7)[name = tensor("transpose_48")]; + tensor out_9 = layer_norm(axes = out_9_axes_0, epsilon = var_63, x = var_459)[name = tensor("out_9")]; + tensor var_463 = const()[name = tensor("op_463"), val = tensor([1, 3, 256])]; + tensor out_11 = reshape(shape = var_463, x = out_9)[name = tensor("out_11")]; + tensor var_465 = silu(x = input_59)[name = tensor("op_465")]; + tensor input_61 = mul(x = var_465, y = out_11)[name = tensor("input_61")]; + tensor ret_out_3 = linear(bias = inner_encoder_out_proj_1_bias, weight = inner_encoder_out_proj_1_weight, x = input_61)[name = tensor("linear_16")]; + tensor x_9 = add(x = input_55, y = ret_out_3)[name = tensor("x_9")]; tensor window_9_begin_0 = const()[name = tensor("window_9_begin_0"), val = tensor([1, 0, 0, 0])]; tensor window_9_end_0 = const()[name = tensor("window_9_end_0"), val = tensor([2, 1, 16, 256])]; tensor window_9_end_mask_0 = const()[name = tensor("window_9_end_mask_0"), val = tensor([false, true, true, true])]; tensor window_9_squeeze_mask_0 = const()[name = tensor("window_9_squeeze_mask_0"), val = tensor([true, false, false, false])]; tensor window_9 = slice_by_index(begin = window_9_begin_0, end = window_9_end_0, end_mask = window_9_end_mask_0, squeeze_mask = window_9_squeeze_mask_0, x = enc_conv_cache)[name = tensor("window_9")]; - tensor var_427_begin_0 = const()[name = tensor("op_427_begin_0"), val = tensor([0, 0, 0])]; - tensor var_427_end_0 = const()[name = tensor("op_427_end_0"), val = tensor([1, 1, 256])]; - tensor var_427_end_mask_0 = const()[name = tensor("op_427_end_mask_0"), val = tensor([true, false, true])]; - tensor var_427 = slice_by_index(begin = var_427_begin_0, end = var_427_end_0, end_mask = var_427_end_mask_0, x = x_9)[name = tensor("op_427")]; - tensor var_430_begin_0 = const()[name = tensor("op_430_begin_0"), val = tensor([0, 1, 0])]; - tensor var_430_end_0 = const()[name = tensor("op_430_end_0"), val = tensor([1, 16, 256])]; - tensor var_430_end_mask_0 = const()[name = tensor("op_430_end_mask_0"), val = tensor([true, true, true])]; - tensor var_430 = slice_by_index(begin = var_430_begin_0, end = var_430_end_0, end_mask = var_430_end_mask_0, x = window_9)[name = tensor("op_430")]; + tensor var_473_begin_0 = const()[name = tensor("op_473_begin_0"), val = tensor([0, 0, 0])]; + tensor var_473_end_0 = const()[name = tensor("op_473_end_0"), val = tensor([1, 1, 256])]; + tensor var_473_end_mask_0 = const()[name = tensor("op_473_end_mask_0"), val = tensor([true, false, true])]; + tensor var_473 = slice_by_index(begin = var_473_begin_0, end = var_473_end_0, end_mask = var_473_end_mask_0, x = x_9)[name = tensor("op_473")]; + tensor var_476_begin_0 = const()[name = tensor("op_476_begin_0"), val = tensor([0, 1, 0])]; + tensor var_476_end_0 = const()[name = tensor("op_476_end_0"), val = tensor([1, 16, 256])]; + tensor var_476_end_mask_0 = const()[name = tensor("op_476_end_mask_0"), val = tensor([true, true, true])]; + tensor var_476 = slice_by_index(begin = var_476_begin_0, end = var_476_end_0, end_mask = var_476_end_mask_0, x = window_9)[name = tensor("op_476")]; tensor window_11_interleave_0 = const()[name = tensor("window_11_interleave_0"), val = tensor(false)]; - tensor window_11 = concat(axis = var_26, interleave = window_11_interleave_0, values = (var_430, var_427))[name = tensor("window_11")]; - tensor var_435_begin_0 = const()[name = tensor("op_435_begin_0"), val = tensor([0, 1, 0])]; - tensor var_435_end_0 = const()[name = tensor("op_435_end_0"), val = tensor([1, 2, 256])]; - tensor var_435_end_mask_0 = const()[name = tensor("op_435_end_mask_0"), val = tensor([true, false, true])]; - tensor var_435 = slice_by_index(begin = var_435_begin_0, end = var_435_end_0, end_mask = var_435_end_mask_0, x = x_9)[name = tensor("op_435")]; - tensor var_438_begin_0 = const()[name = tensor("op_438_begin_0"), val = tensor([0, 1, 0])]; - tensor var_438_end_0 = const()[name = tensor("op_438_end_0"), val = tensor([1, 16, 256])]; - tensor var_438_end_mask_0 = const()[name = tensor("op_438_end_mask_0"), val = tensor([true, true, true])]; - tensor var_438 = slice_by_index(begin = var_438_begin_0, end = var_438_end_0, end_mask = var_438_end_mask_0, x = window_11)[name = tensor("op_438")]; + tensor window_11 = concat(axis = var_72, interleave = window_11_interleave_0, values = (var_476, var_473))[name = tensor("window_11")]; + tensor var_481_begin_0 = const()[name = tensor("op_481_begin_0"), val = tensor([0, 1, 0])]; + tensor var_481_end_0 = const()[name = tensor("op_481_end_0"), val = tensor([1, 2, 256])]; + tensor var_481_end_mask_0 = const()[name = tensor("op_481_end_mask_0"), val = tensor([true, false, true])]; + tensor var_481 = slice_by_index(begin = var_481_begin_0, end = var_481_end_0, end_mask = var_481_end_mask_0, x = x_9)[name = tensor("op_481")]; + tensor var_484_begin_0 = const()[name = tensor("op_484_begin_0"), val = tensor([0, 1, 0])]; + tensor var_484_end_0 = const()[name = tensor("op_484_end_0"), val = tensor([1, 16, 256])]; + tensor var_484_end_mask_0 = const()[name = tensor("op_484_end_mask_0"), val = tensor([true, true, true])]; + tensor var_484 = slice_by_index(begin = var_484_begin_0, end = var_484_end_0, end_mask = var_484_end_mask_0, x = window_11)[name = tensor("op_484")]; tensor window_13_interleave_0 = const()[name = tensor("window_13_interleave_0"), val = tensor(false)]; - tensor window_13 = concat(axis = var_26, interleave = window_13_interleave_0, values = (var_438, var_435))[name = tensor("window_13")]; - tensor var_443_begin_0 = const()[name = tensor("op_443_begin_0"), val = tensor([0, 2, 0])]; - tensor var_443_end_0 = const()[name = tensor("op_443_end_0"), val = tensor([1, 1, 256])]; - tensor var_443_end_mask_0 = const()[name = tensor("op_443_end_mask_0"), val = tensor([true, true, true])]; - tensor var_443 = slice_by_index(begin = var_443_begin_0, end = var_443_end_0, end_mask = var_443_end_mask_0, x = x_9)[name = tensor("op_443")]; - tensor var_446_begin_0 = const()[name = tensor("op_446_begin_0"), val = tensor([0, 1, 0])]; - tensor var_446_end_0 = const()[name = tensor("op_446_end_0"), val = tensor([1, 16, 256])]; - tensor var_446_end_mask_0 = const()[name = tensor("op_446_end_mask_0"), val = tensor([true, true, true])]; - tensor var_446 = slice_by_index(begin = var_446_begin_0, end = var_446_end_0, end_mask = var_446_end_mask_0, x = window_13)[name = tensor("op_446")]; + tensor window_13 = concat(axis = var_72, interleave = window_13_interleave_0, values = (var_484, var_481))[name = tensor("window_13")]; + tensor var_489_begin_0 = const()[name = tensor("op_489_begin_0"), val = tensor([0, 2, 0])]; + tensor var_489_end_0 = const()[name = tensor("op_489_end_0"), val = tensor([1, 1, 256])]; + tensor var_489_end_mask_0 = const()[name = tensor("op_489_end_mask_0"), val = tensor([true, true, true])]; + tensor var_489 = slice_by_index(begin = var_489_begin_0, end = var_489_end_0, end_mask = var_489_end_mask_0, x = x_9)[name = tensor("op_489")]; + tensor var_492_begin_0 = const()[name = tensor("op_492_begin_0"), val = tensor([0, 1, 0])]; + tensor var_492_end_0 = const()[name = tensor("op_492_end_0"), val = tensor([1, 16, 256])]; + tensor var_492_end_mask_0 = const()[name = tensor("op_492_end_mask_0"), val = tensor([true, true, true])]; + tensor var_492 = slice_by_index(begin = var_492_begin_0, end = var_492_end_0, end_mask = var_492_end_mask_0, x = window_13)[name = tensor("op_492")]; tensor window_15_interleave_0 = const()[name = tensor("window_15_interleave_0"), val = tensor(false)]; - tensor window_15 = concat(axis = var_26, interleave = window_15_interleave_0, values = (var_446, var_443))[name = tensor("window_15")]; - tensor input_61_interleave_0 = const()[name = tensor("input_61_interleave_0"), val = tensor(false)]; - tensor input_61 = concat(axis = var_23, interleave = input_61_interleave_0, values = (window_11, window_13, window_15))[name = tensor("input_61")]; + tensor window_15 = concat(axis = var_72, interleave = window_15_interleave_0, values = (var_492, var_489))[name = tensor("window_15")]; + tensor input_63_interleave_0 = const()[name = tensor("input_63_interleave_0"), val = tensor(false)]; + tensor input_63 = concat(axis = var_58, interleave = input_63_interleave_0, values = (window_11, window_13, window_15))[name = tensor("input_63")]; tensor x_11_axes_0 = const()[name = tensor("x_11_axes_0"), val = tensor([-1])]; - tensor x_11 = layer_norm(axes = x_11_axes_0, beta = encoder_conv_module_1_sequential_0_bias, epsilon = var_28, gamma = encoder_conv_module_1_sequential_0_weight, x = input_61)[name = tensor("x_11")]; - tensor input_63_perm_0 = const()[name = tensor("input_63_perm_0"), val = tensor([0, 2, 1])]; + tensor x_11 = layer_norm(axes = x_11_axes_0, beta = inner_encoder_conv_module_1_sequential_0_bias, epsilon = var_55, gamma = inner_encoder_conv_module_1_sequential_0_weight, x = input_63)[name = tensor("x_11")]; + tensor input_65_perm_0 = const()[name = tensor("input_65_perm_0"), val = tensor([0, 2, 1])]; tensor inputs_13_pad_type_0 = const()[name = tensor("inputs_13_pad_type_0"), val = tensor("valid")]; tensor inputs_13_strides_0 = const()[name = tensor("inputs_13_strides_0"), val = tensor([1])]; tensor inputs_13_pad_0 = const()[name = tensor("inputs_13_pad_0"), val = tensor([0, 0])]; tensor inputs_13_dilations_0 = const()[name = tensor("inputs_13_dilations_0"), val = tensor([1])]; tensor inputs_13_groups_0 = const()[name = tensor("inputs_13_groups_0"), val = tensor(1)]; - tensor input_63 = transpose(perm = input_63_perm_0, x = x_11)[name = tensor("transpose_47")]; - tensor inputs_13 = conv(bias = encoder_conv_module_1_sequential_2_conv_bias, dilations = inputs_13_dilations_0, groups = inputs_13_groups_0, pad = inputs_13_pad_0, pad_type = inputs_13_pad_type_0, strides = inputs_13_strides_0, weight = encoder_conv_module_1_sequential_2_conv_weight, x = input_63)[name = tensor("inputs_13")]; - tensor var_471_split_sizes_0 = const()[name = tensor("op_471_split_sizes_0"), val = tensor([256, 256])]; - tensor var_471_axis_0 = const()[name = tensor("op_471_axis_0"), val = tensor(1)]; - tensor var_471_0, tensor var_471_1 = split(axis = var_471_axis_0, split_sizes = var_471_split_sizes_0, x = inputs_13)[name = tensor("op_471")]; - tensor var_473 = sigmoid(x = var_471_1)[name = tensor("op_473")]; - tensor inputs_15 = mul(x = var_471_0, y = var_473)[name = tensor("inputs_15")]; + tensor input_65 = transpose(perm = input_65_perm_0, x = x_11)[name = tensor("transpose_47")]; + tensor inputs_13 = conv(bias = inner_encoder_conv_module_1_sequential_2_conv_bias, dilations = inputs_13_dilations_0, groups = inputs_13_groups_0, pad = inputs_13_pad_0, pad_type = inputs_13_pad_type_0, strides = inputs_13_strides_0, weight = inner_encoder_conv_module_1_sequential_2_conv_weight, x = input_65)[name = tensor("inputs_13")]; + tensor var_517_split_sizes_0 = const()[name = tensor("op_517_split_sizes_0"), val = tensor([256, 256])]; + tensor var_517_axis_0 = const()[name = tensor("op_517_axis_0"), val = tensor(1)]; + tensor var_517_0, tensor var_517_1 = split(axis = var_517_axis_0, split_sizes = var_517_split_sizes_0, x = inputs_13)[name = tensor("op_517")]; + tensor var_519 = sigmoid(x = var_517_1)[name = tensor("op_519")]; + tensor inputs_15 = mul(x = var_517_0, y = var_519)[name = tensor("inputs_15")]; tensor outputs_aug_3_pad_type_0 = const()[name = tensor("outputs_aug_3_pad_type_0"), val = tensor("custom")]; tensor outputs_aug_3_pad_0 = const()[name = tensor("outputs_aug_3_pad_0"), val = tensor([15, 15])]; tensor outputs_aug_3_groups_0 = const()[name = tensor("outputs_aug_3_groups_0"), val = tensor(256)]; tensor outputs_aug_3_strides_0 = const()[name = tensor("outputs_aug_3_strides_0"), val = tensor([1])]; tensor outputs_aug_3_dilations_0 = const()[name = tensor("outputs_aug_3_dilations_0"), val = tensor([1])]; - tensor outputs_aug_3 = conv(dilations = outputs_aug_3_dilations_0, groups = outputs_aug_3_groups_0, pad = outputs_aug_3_pad_0, pad_type = outputs_aug_3_pad_type_0, strides = outputs_aug_3_strides_0, weight = encoder_conv_module_1_sequential_4_conv_weight, x = inputs_15)[name = tensor("outputs_aug_3")]; - tensor input_65_begin_0 = const()[name = tensor("input_65_begin_0"), val = tensor([0, 0, 0])]; - tensor input_65_end_0 = const()[name = tensor("input_65_end_0"), val = tensor([3, 256, 16])]; - tensor input_65_end_mask_0 = const()[name = tensor("input_65_end_mask_0"), val = tensor([true, true, false])]; - tensor input_65 = slice_by_index(begin = input_65_begin_0, end = input_65_end_0, end_mask = input_65_end_mask_0, x = outputs_aug_3)[name = tensor("input_65")]; - tensor inputs_17 = batch_norm(beta = encoder_conv_module_1_sequential_5_bias, epsilon = var_28, gamma = encoder_conv_module_1_sequential_5_weight, mean = encoder_conv_module_1_sequential_5_running_mean, variance = encoder_conv_module_1_sequential_5_running_var, x = input_65)[name = tensor("inputs_17")]; - tensor input_67 = silu(x = inputs_17)[name = tensor("input_67")]; - tensor input_69_pad_type_0 = const()[name = tensor("input_69_pad_type_0"), val = tensor("valid")]; - tensor input_69_strides_0 = const()[name = tensor("input_69_strides_0"), val = tensor([1])]; - tensor input_69_pad_0 = const()[name = tensor("input_69_pad_0"), val = tensor([0, 0])]; - tensor input_69_dilations_0 = const()[name = tensor("input_69_dilations_0"), val = tensor([1])]; - tensor input_69_groups_0 = const()[name = tensor("input_69_groups_0"), val = tensor(1)]; - tensor input_69 = conv(bias = encoder_conv_module_1_sequential_7_conv_bias, dilations = input_69_dilations_0, groups = input_69_groups_0, pad = input_69_pad_0, pad_type = input_69_pad_type_0, strides = input_69_strides_0, weight = encoder_conv_module_1_sequential_7_conv_weight, x = input_67)[name = tensor("input_69")]; + tensor outputs_aug_3 = conv(dilations = outputs_aug_3_dilations_0, groups = outputs_aug_3_groups_0, pad = outputs_aug_3_pad_0, pad_type = outputs_aug_3_pad_type_0, strides = outputs_aug_3_strides_0, weight = inner_encoder_conv_module_1_sequential_4_conv_weight, x = inputs_15)[name = tensor("outputs_aug_3")]; + tensor input_67_begin_0 = const()[name = tensor("input_67_begin_0"), val = tensor([0, 0, 0])]; + tensor input_67_end_0 = const()[name = tensor("input_67_end_0"), val = tensor([3, 256, 16])]; + tensor input_67_end_mask_0 = const()[name = tensor("input_67_end_mask_0"), val = tensor([true, true, false])]; + tensor input_67 = slice_by_index(begin = input_67_begin_0, end = input_67_end_0, end_mask = input_67_end_mask_0, x = outputs_aug_3)[name = tensor("input_67")]; + tensor inputs_17 = batch_norm(beta = inner_encoder_conv_module_1_sequential_5_bias, epsilon = var_55, gamma = inner_encoder_conv_module_1_sequential_5_weight, mean = inner_encoder_conv_module_1_sequential_5_running_mean, variance = inner_encoder_conv_module_1_sequential_5_running_var, x = input_67)[name = tensor("inputs_17")]; + tensor input_69 = silu(x = inputs_17)[name = tensor("input_69")]; + tensor input_71_pad_type_0 = const()[name = tensor("input_71_pad_type_0"), val = tensor("valid")]; + tensor input_71_strides_0 = const()[name = tensor("input_71_strides_0"), val = tensor([1])]; + tensor input_71_pad_0 = const()[name = tensor("input_71_pad_0"), val = tensor([0, 0])]; + tensor input_71_dilations_0 = const()[name = tensor("input_71_dilations_0"), val = tensor([1])]; + tensor input_71_groups_0 = const()[name = tensor("input_71_groups_0"), val = tensor(1)]; + tensor input_71 = conv(bias = inner_encoder_conv_module_1_sequential_7_conv_bias, dilations = input_71_dilations_0, groups = input_71_groups_0, pad = input_71_pad_0, pad_type = input_71_pad_type_0, strides = input_71_strides_0, weight = inner_encoder_conv_module_1_sequential_7_conv_weight, x = input_69)[name = tensor("input_71")]; tensor conv_out_3_perm_0 = const()[name = tensor("conv_out_3_perm_0"), val = tensor([0, 2, 1])]; - tensor var_504_begin_0 = const()[name = tensor("op_504_begin_0"), val = tensor([0, -1, 0])]; - tensor var_504_end_0 = const()[name = tensor("op_504_end_0"), val = tensor([3, 16, 256])]; - tensor var_504_end_mask_0 = const()[name = tensor("op_504_end_mask_0"), val = tensor([true, true, true])]; - tensor conv_out_3 = transpose(perm = conv_out_3_perm_0, x = input_69)[name = tensor("transpose_46")]; - tensor var_504 = slice_by_index(begin = var_504_begin_0, end = var_504_end_0, end_mask = var_504_end_mask_0, x = conv_out_3)[name = tensor("op_504")]; - tensor var_506_perm_0 = const()[name = tensor("op_506_perm_0"), val = tensor([1, 0, 2])]; - tensor var_506 = transpose(perm = var_506_perm_0, x = var_504)[name = tensor("transpose_45")]; - tensor input_71 = add(x = x_9, y = var_506)[name = tensor("input_71")]; - tensor input_73_axes_0 = const()[name = tensor("input_73_axes_0"), val = tensor([-1])]; - tensor input_73 = layer_norm(axes = input_73_axes_0, beta = encoder_ffn2_1_module_sequential_0_bias, epsilon = var_28, gamma = encoder_ffn2_1_module_sequential_0_weight, x = input_71)[name = tensor("input_73")]; - tensor inputs_19 = linear(bias = encoder_ffn2_1_module_sequential_1_linear_bias, weight = encoder_ffn2_1_module_sequential_1_linear_weight, x = input_73)[name = tensor("linear_17")]; - tensor input_75 = silu(x = inputs_19)[name = tensor("input_75")]; - tensor input_79 = linear(bias = encoder_ffn2_1_module_sequential_4_linear_bias, weight = encoder_ffn2_1_module_sequential_4_linear_weight, x = input_75)[name = tensor("linear_18")]; - tensor var_529 = const()[name = tensor("op_529"), val = tensor(0x1p-1)]; - tensor var_530 = mul(x = input_79, y = var_529)[name = tensor("op_530")]; - tensor input_81 = add(x = var_530, y = input_71)[name = tensor("input_81")]; - tensor input_83_axes_0 = const()[name = tensor("input_83_axes_0"), val = tensor([-1])]; - tensor input_83 = layer_norm(axes = input_83_axes_0, beta = encoder_layer_norm_1_bias, epsilon = var_28, gamma = encoder_layer_norm_1_weight, x = input_81)[name = tensor("input_83")]; + tensor var_550_begin_0 = const()[name = tensor("op_550_begin_0"), val = tensor([0, -1, 0])]; + tensor var_550_end_0 = const()[name = tensor("op_550_end_0"), val = tensor([3, 16, 256])]; + tensor var_550_end_mask_0 = const()[name = tensor("op_550_end_mask_0"), val = tensor([true, true, true])]; + tensor conv_out_3 = transpose(perm = conv_out_3_perm_0, x = input_71)[name = tensor("transpose_46")]; + tensor var_550 = slice_by_index(begin = var_550_begin_0, end = var_550_end_0, end_mask = var_550_end_mask_0, x = conv_out_3)[name = tensor("op_550")]; + tensor var_552_perm_0 = const()[name = tensor("op_552_perm_0"), val = tensor([1, 0, 2])]; + tensor var_552 = transpose(perm = var_552_perm_0, x = var_550)[name = tensor("transpose_45")]; + tensor input_73 = add(x = x_9, y = var_552)[name = tensor("input_73")]; + tensor input_75_axes_0 = const()[name = tensor("input_75_axes_0"), val = tensor([-1])]; + tensor input_75 = layer_norm(axes = input_75_axes_0, beta = inner_encoder_ffn2_1_module_sequential_0_bias, epsilon = var_55, gamma = inner_encoder_ffn2_1_module_sequential_0_weight, x = input_73)[name = tensor("input_75")]; + tensor inputs_19 = linear(bias = inner_encoder_ffn2_1_module_sequential_1_linear_bias, weight = inner_encoder_ffn2_1_module_sequential_1_linear_weight, x = input_75)[name = tensor("linear_17")]; + tensor input_77 = silu(x = inputs_19)[name = tensor("input_77")]; + tensor input_81 = linear(bias = inner_encoder_ffn2_1_module_sequential_4_linear_bias, weight = inner_encoder_ffn2_1_module_sequential_4_linear_weight, x = input_77)[name = tensor("linear_18")]; + tensor var_575 = const()[name = tensor("op_575"), val = tensor(0x1p-1)]; + tensor var_576 = mul(x = input_81, y = var_575)[name = tensor("op_576")]; + tensor input_83 = add(x = var_576, y = input_73)[name = tensor("input_83")]; tensor input_85_axes_0 = const()[name = tensor("input_85_axes_0"), val = tensor([-1])]; - tensor input_85 = layer_norm(axes = input_85_axes_0, beta = encoder_ffn1_2_module_sequential_0_bias, epsilon = var_28, gamma = encoder_ffn1_2_module_sequential_0_weight, x = input_83)[name = tensor("input_85")]; - tensor inputs_21 = linear(bias = encoder_ffn1_2_module_sequential_1_linear_bias, weight = encoder_ffn1_2_module_sequential_1_linear_weight, x = input_85)[name = tensor("linear_19")]; - tensor input_87 = silu(x = inputs_21)[name = tensor("input_87")]; - tensor input_91 = linear(bias = encoder_ffn1_2_module_sequential_4_linear_bias, weight = encoder_ffn1_2_module_sequential_4_linear_weight, x = input_87)[name = tensor("linear_20")]; - tensor var_559 = const()[name = tensor("op_559"), val = tensor(0x1p-1)]; - tensor var_560 = mul(x = input_91, y = var_559)[name = tensor("op_560")]; - tensor input_93 = add(x = var_560, y = input_83)[name = tensor("input_93")]; + tensor input_85 = layer_norm(axes = input_85_axes_0, beta = inner_encoder_layer_norm_1_bias, epsilon = var_55, gamma = inner_encoder_layer_norm_1_weight, x = input_83)[name = tensor("input_85")]; + tensor input_87_axes_0 = const()[name = tensor("input_87_axes_0"), val = tensor([-1])]; + tensor input_87 = layer_norm(axes = input_87_axes_0, beta = inner_encoder_ffn1_2_module_sequential_0_bias, epsilon = var_55, gamma = inner_encoder_ffn1_2_module_sequential_0_weight, x = input_85)[name = tensor("input_87")]; + tensor inputs_21 = linear(bias = inner_encoder_ffn1_2_module_sequential_1_linear_bias, weight = inner_encoder_ffn1_2_module_sequential_1_linear_weight, x = input_87)[name = tensor("linear_19")]; + tensor input_89 = silu(x = inputs_21)[name = tensor("input_89")]; + tensor input_93 = linear(bias = inner_encoder_ffn1_2_module_sequential_4_linear_bias, weight = inner_encoder_ffn1_2_module_sequential_4_linear_weight, x = input_89)[name = tensor("linear_20")]; + tensor var_605 = const()[name = tensor("op_605"), val = tensor(0x1p-1)]; + tensor var_606 = mul(x = input_93, y = var_605)[name = tensor("op_606")]; + tensor input_95 = add(x = var_606, y = input_85)[name = tensor("input_95")]; tensor x_13_axes_0 = const()[name = tensor("x_13_axes_0"), val = tensor([-1])]; - tensor x_13 = layer_norm(axes = x_13_axes_0, beta = encoder_ret_lns_2_bias, epsilon = var_28, gamma = encoder_ret_lns_2_weight, x = input_93)[name = tensor("x_13")]; + tensor x_13 = layer_norm(axes = x_13_axes_0, beta = inner_encoder_ret_lns_2_bias, epsilon = var_55, gamma = inner_encoder_ret_lns_2_weight, x = input_95)[name = tensor("x_13")]; tensor prev_kv_5_begin_0 = const()[name = tensor("prev_kv_5_begin_0"), val = tensor([2, 0, 0, 0, 0])]; tensor prev_kv_5_end_0 = const()[name = tensor("prev_kv_5_end_0"), val = tensor([3, 1, 4, 64, 64])]; tensor prev_kv_5_end_mask_0 = const()[name = tensor("prev_kv_5_end_mask_0"), val = tensor([false, true, true, true, true])]; @@ -573,163 +591,163 @@ program(1.0) tensor prev_scale_5_end_mask_0 = const()[name = tensor("prev_scale_5_end_mask_0"), val = tensor([false, true])]; tensor prev_scale_5_squeeze_mask_0 = const()[name = tensor("prev_scale_5_squeeze_mask_0"), val = tensor([true, false])]; tensor prev_scale_5 = slice_by_index(begin = prev_scale_5_begin_0, end = prev_scale_5_end_0, end_mask = prev_scale_5_end_mask_0, squeeze_mask = prev_scale_5_squeeze_mask_0, x = enc_scale)[name = tensor("prev_scale_5")]; - tensor var_574 = linear(bias = encoder_q_proj_2_bias, weight = encoder_q_proj_2_weight, x = x_13)[name = tensor("linear_21")]; - tensor var_575 = const()[name = tensor("op_575"), val = tensor([1, 3, 4, 64])]; - tensor var_576 = reshape(shape = var_575, x = var_574)[name = tensor("op_576")]; + tensor var_620 = linear(bias = inner_encoder_q_proj_2_bias, weight = inner_encoder_q_proj_2_weight, x = x_13)[name = tensor("linear_21")]; + tensor var_621 = const()[name = tensor("op_621"), val = tensor([1, 3, 4, 64])]; + tensor var_622 = reshape(shape = var_621, x = var_620)[name = tensor("op_622")]; tensor q_5_perm_0 = const()[name = tensor("q_5_perm_0"), val = tensor([0, 2, 1, 3])]; - tensor var_580 = linear(bias = encoder_k_proj_2_bias, weight = encoder_k_proj_2_weight, x = x_13)[name = tensor("linear_22")]; - tensor var_581 = const()[name = tensor("op_581"), val = tensor(0x1p-3)]; - tensor var_582 = mul(x = var_580, y = var_581)[name = tensor("op_582")]; - tensor var_583 = const()[name = tensor("op_583"), val = tensor([1, 3, 4, 64])]; - tensor var_584 = reshape(shape = var_583, x = var_582)[name = tensor("op_584")]; + tensor var_626 = linear(bias = inner_encoder_k_proj_2_bias, weight = inner_encoder_k_proj_2_weight, x = x_13)[name = tensor("linear_22")]; + tensor var_627 = const()[name = tensor("op_627"), val = tensor(0x1p-3)]; + tensor var_628 = mul(x = var_626, y = var_627)[name = tensor("op_628")]; + tensor var_629 = const()[name = tensor("op_629"), val = tensor([1, 3, 4, 64])]; + tensor var_630 = reshape(shape = var_629, x = var_628)[name = tensor("op_630")]; tensor k_5_perm_0 = const()[name = tensor("k_5_perm_0"), val = tensor([0, 2, 1, 3])]; - tensor var_588 = linear(bias = encoder_v_proj_2_bias, weight = encoder_v_proj_2_weight, x = x_13)[name = tensor("linear_23")]; - tensor var_589 = const()[name = tensor("op_589"), val = tensor([1, 3, 4, 64])]; - tensor var_590 = reshape(shape = var_589, x = var_588)[name = tensor("op_590")]; + tensor var_634 = linear(bias = inner_encoder_v_proj_2_bias, weight = inner_encoder_v_proj_2_weight, x = x_13)[name = tensor("linear_23")]; + tensor var_635 = const()[name = tensor("op_635"), val = tensor([1, 3, 4, 64])]; + tensor var_636 = reshape(shape = var_635, x = var_634)[name = tensor("op_636")]; tensor v_5_perm_0 = const()[name = tensor("v_5_perm_0"), val = tensor([0, 2, 1, 3])]; - tensor input_97 = linear(bias = encoder_g_proj_2_bias, weight = encoder_g_proj_2_weight, x = x_13)[name = tensor("linear_24")]; + tensor input_99 = linear(bias = inner_encoder_g_proj_2_bias, weight = inner_encoder_g_proj_2_weight, x = x_13)[name = tensor("linear_24")]; tensor sqrt_s0_5 = sqrt(x = prev_scale_5)[name = tensor("sqrt_s0_5")]; - tensor s_t_5 = add(x = prev_scale_5, y = encoder__t_index)[name = tensor("s_t_5")]; + tensor s_t_5 = add(x = prev_scale_5, y = inner_encoder__t_index)[name = tensor("s_t_5")]; tensor sqrt_s_t_5 = sqrt(x = s_t_5)[name = tensor("sqrt_s_t_5")]; tensor qk_5_transpose_x_1 = const()[name = tensor("qk_5_transpose_x_1"), val = tensor(false)]; tensor qk_5_transpose_y_1 = const()[name = tensor("qk_5_transpose_y_1"), val = tensor(true)]; - tensor k_5 = transpose(perm = k_5_perm_0, x = var_584)[name = tensor("transpose_43")]; - tensor q_5 = transpose(perm = q_5_perm_0, x = var_576)[name = tensor("transpose_44")]; + tensor k_5 = transpose(perm = k_5_perm_0, x = var_630)[name = tensor("transpose_43")]; + tensor q_5 = transpose(perm = q_5_perm_0, x = var_622)[name = tensor("transpose_44")]; tensor qk_5 = matmul(transpose_x = qk_5_transpose_x_1, transpose_y = qk_5_transpose_y_1, x = q_5, y = k_5)[name = tensor("qk_5")]; - tensor var_600 = const()[name = tensor("op_600"), val = tensor([3, 1])]; - tensor var_601 = reshape(shape = var_600, x = sqrt_s_t_5)[name = tensor("op_601")]; - tensor M_5 = real_div(x = encoder__causal_mask, y = var_601)[name = tensor("M_5")]; - tensor var_603 = mul(x = qk_5, y = M_5)[name = tensor("op_603")]; + tensor var_646 = const()[name = tensor("op_646"), val = tensor([3, 1])]; + tensor var_647 = reshape(shape = var_646, x = sqrt_s_t_5)[name = tensor("op_647")]; + tensor M_5 = real_div(x = inner_encoder__causal_mask, y = var_647)[name = tensor("M_5")]; + tensor var_649 = mul(x = qk_5, y = M_5)[name = tensor("op_649")]; tensor inner_5_transpose_x_0 = const()[name = tensor("inner_5_transpose_x_0"), val = tensor(false)]; tensor inner_5_transpose_y_0 = const()[name = tensor("inner_5_transpose_y_0"), val = tensor(false)]; - tensor v_5 = transpose(perm = v_5_perm_0, x = var_590)[name = tensor("transpose_42")]; - tensor inner_5 = matmul(transpose_x = inner_5_transpose_x_0, transpose_y = inner_5_transpose_y_0, x = var_603, y = v_5)[name = tensor("inner_5")]; - tensor var_605_transpose_x_0 = const()[name = tensor("op_605_transpose_x_0"), val = tensor(false)]; - tensor var_605_transpose_y_0 = const()[name = tensor("op_605_transpose_y_0"), val = tensor(false)]; - tensor var_605 = matmul(transpose_x = var_605_transpose_x_0, transpose_y = var_605_transpose_y_0, x = q_5, y = prev_kv_5)[name = tensor("op_605")]; - tensor var_606 = real_div(x = sqrt_s0_5, y = sqrt_s_t_5)[name = tensor("op_606")]; - tensor var_607 = const()[name = tensor("op_607"), val = tensor([1, 1, 3, 1])]; - tensor var_608 = reshape(shape = var_607, x = var_606)[name = tensor("op_608")]; - tensor cross_5 = mul(x = var_605, y = var_608)[name = tensor("cross_5")]; + tensor v_5 = transpose(perm = v_5_perm_0, x = var_636)[name = tensor("transpose_42")]; + tensor inner_5 = matmul(transpose_x = inner_5_transpose_x_0, transpose_y = inner_5_transpose_y_0, x = var_649, y = v_5)[name = tensor("inner_5")]; + tensor var_651_transpose_x_0 = const()[name = tensor("op_651_transpose_x_0"), val = tensor(false)]; + tensor var_651_transpose_y_0 = const()[name = tensor("op_651_transpose_y_0"), val = tensor(false)]; + tensor var_651 = matmul(transpose_x = var_651_transpose_x_0, transpose_y = var_651_transpose_y_0, x = q_5, y = prev_kv_5)[name = tensor("op_651")]; + tensor var_652 = real_div(x = sqrt_s0_5, y = sqrt_s_t_5)[name = tensor("op_652")]; + tensor var_653 = const()[name = tensor("op_653"), val = tensor([1, 1, 3, 1])]; + tensor var_654 = reshape(shape = var_653, x = var_652)[name = tensor("op_654")]; + tensor cross_5 = mul(x = var_651, y = var_654)[name = tensor("cross_5")]; tensor out_13 = add(x = inner_5, y = cross_5)[name = tensor("out_13")]; - tensor var_611 = mul(x = prev_kv_5, y = sqrt_s0_5)[name = tensor("op_611")]; - tensor var_613_transpose_x_1 = const()[name = tensor("op_613_transpose_x_1"), val = tensor(true)]; - tensor var_613_transpose_y_1 = const()[name = tensor("op_613_transpose_y_1"), val = tensor(false)]; - tensor var_613 = matmul(transpose_x = var_613_transpose_x_1, transpose_y = var_613_transpose_y_1, x = k_5, y = v_5)[name = tensor("op_613")]; - tensor new_kv_unnorm_5 = add(x = var_611, y = var_613)[name = tensor("new_kv_unnorm_5")]; - tensor var_615 = const()[name = tensor("op_615"), val = tensor(0x1.8p+1)]; - tensor new_scale_5 = add(x = prev_scale_5, y = var_615)[name = tensor("new_scale_5")]; - tensor var_617 = sqrt(x = new_scale_5)[name = tensor("op_617")]; - tensor var_618 = real_div(x = new_kv_unnorm_5, y = var_617)[name = tensor("op_618")]; - tensor var_619_perm_0 = const()[name = tensor("op_619_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_657 = mul(x = prev_kv_5, y = sqrt_s0_5)[name = tensor("op_657")]; + tensor var_659_transpose_x_1 = const()[name = tensor("op_659_transpose_x_1"), val = tensor(true)]; + tensor var_659_transpose_y_1 = const()[name = tensor("op_659_transpose_y_1"), val = tensor(false)]; + tensor var_659 = matmul(transpose_x = var_659_transpose_x_1, transpose_y = var_659_transpose_y_1, x = k_5, y = v_5)[name = tensor("op_659")]; + tensor new_kv_unnorm_5 = add(x = var_657, y = var_659)[name = tensor("new_kv_unnorm_5")]; + tensor var_661 = const()[name = tensor("op_661"), val = tensor(0x1.8p+1)]; + tensor new_scale_5 = add(x = prev_scale_5, y = var_661)[name = tensor("new_scale_5")]; + tensor var_663 = sqrt(x = new_scale_5)[name = tensor("op_663")]; + tensor var_664 = real_div(x = new_kv_unnorm_5, y = var_663)[name = tensor("op_664")]; + tensor var_665_perm_0 = const()[name = tensor("op_665_perm_0"), val = tensor([0, 2, 1, 3])]; tensor out_15_axes_0 = const()[name = tensor("out_15_axes_0"), val = tensor([-1])]; - tensor var_619 = transpose(perm = var_619_perm_0, x = out_13)[name = tensor("transpose_41")]; - tensor out_15 = layer_norm(axes = out_15_axes_0, epsilon = var_17, x = var_619)[name = tensor("out_15")]; - tensor var_623 = const()[name = tensor("op_623"), val = tensor([1, 3, 256])]; - tensor out_17 = reshape(shape = var_623, x = out_15)[name = tensor("out_17")]; - tensor var_625 = silu(x = input_97)[name = tensor("op_625")]; - tensor input_99 = mul(x = var_625, y = out_17)[name = tensor("input_99")]; - tensor ret_out_5 = linear(bias = encoder_out_proj_2_bias, weight = encoder_out_proj_2_weight, x = input_99)[name = tensor("linear_25")]; - tensor x_15 = add(x = input_93, y = ret_out_5)[name = tensor("x_15")]; + tensor var_665 = transpose(perm = var_665_perm_0, x = out_13)[name = tensor("transpose_41")]; + tensor out_15 = layer_norm(axes = out_15_axes_0, epsilon = var_63, x = var_665)[name = tensor("out_15")]; + tensor var_669 = const()[name = tensor("op_669"), val = tensor([1, 3, 256])]; + tensor out_17 = reshape(shape = var_669, x = out_15)[name = tensor("out_17")]; + tensor var_671 = silu(x = input_99)[name = tensor("op_671")]; + tensor input_101 = mul(x = var_671, y = out_17)[name = tensor("input_101")]; + tensor ret_out_5 = linear(bias = inner_encoder_out_proj_2_bias, weight = inner_encoder_out_proj_2_weight, x = input_101)[name = tensor("linear_25")]; + tensor x_15 = add(x = input_95, y = ret_out_5)[name = tensor("x_15")]; tensor window_17_begin_0 = const()[name = tensor("window_17_begin_0"), val = tensor([2, 0, 0, 0])]; tensor window_17_end_0 = const()[name = tensor("window_17_end_0"), val = tensor([3, 1, 16, 256])]; tensor window_17_end_mask_0 = const()[name = tensor("window_17_end_mask_0"), val = tensor([false, true, true, true])]; tensor window_17_squeeze_mask_0 = const()[name = tensor("window_17_squeeze_mask_0"), val = tensor([true, false, false, false])]; tensor window_17 = slice_by_index(begin = window_17_begin_0, end = window_17_end_0, end_mask = window_17_end_mask_0, squeeze_mask = window_17_squeeze_mask_0, x = enc_conv_cache)[name = tensor("window_17")]; - tensor var_633_begin_0 = const()[name = tensor("op_633_begin_0"), val = tensor([0, 0, 0])]; - tensor var_633_end_0 = const()[name = tensor("op_633_end_0"), val = tensor([1, 1, 256])]; - tensor var_633_end_mask_0 = const()[name = tensor("op_633_end_mask_0"), val = tensor([true, false, true])]; - tensor var_633 = slice_by_index(begin = var_633_begin_0, end = var_633_end_0, end_mask = var_633_end_mask_0, x = x_15)[name = tensor("op_633")]; - tensor var_636_begin_0 = const()[name = tensor("op_636_begin_0"), val = tensor([0, 1, 0])]; - tensor var_636_end_0 = const()[name = tensor("op_636_end_0"), val = tensor([1, 16, 256])]; - tensor var_636_end_mask_0 = const()[name = tensor("op_636_end_mask_0"), val = tensor([true, true, true])]; - tensor var_636 = slice_by_index(begin = var_636_begin_0, end = var_636_end_0, end_mask = var_636_end_mask_0, x = window_17)[name = tensor("op_636")]; + tensor var_679_begin_0 = const()[name = tensor("op_679_begin_0"), val = tensor([0, 0, 0])]; + tensor var_679_end_0 = const()[name = tensor("op_679_end_0"), val = tensor([1, 1, 256])]; + tensor var_679_end_mask_0 = const()[name = tensor("op_679_end_mask_0"), val = tensor([true, false, true])]; + tensor var_679 = slice_by_index(begin = var_679_begin_0, end = var_679_end_0, end_mask = var_679_end_mask_0, x = x_15)[name = tensor("op_679")]; + tensor var_682_begin_0 = const()[name = tensor("op_682_begin_0"), val = tensor([0, 1, 0])]; + tensor var_682_end_0 = const()[name = tensor("op_682_end_0"), val = tensor([1, 16, 256])]; + tensor var_682_end_mask_0 = const()[name = tensor("op_682_end_mask_0"), val = tensor([true, true, true])]; + tensor var_682 = slice_by_index(begin = var_682_begin_0, end = var_682_end_0, end_mask = var_682_end_mask_0, x = window_17)[name = tensor("op_682")]; tensor window_19_interleave_0 = const()[name = tensor("window_19_interleave_0"), val = tensor(false)]; - tensor window_19 = concat(axis = var_26, interleave = window_19_interleave_0, values = (var_636, var_633))[name = tensor("window_19")]; - tensor var_641_begin_0 = const()[name = tensor("op_641_begin_0"), val = tensor([0, 1, 0])]; - tensor var_641_end_0 = const()[name = tensor("op_641_end_0"), val = tensor([1, 2, 256])]; - tensor var_641_end_mask_0 = const()[name = tensor("op_641_end_mask_0"), val = tensor([true, false, true])]; - tensor var_641 = slice_by_index(begin = var_641_begin_0, end = var_641_end_0, end_mask = var_641_end_mask_0, x = x_15)[name = tensor("op_641")]; - tensor var_644_begin_0 = const()[name = tensor("op_644_begin_0"), val = tensor([0, 1, 0])]; - tensor var_644_end_0 = const()[name = tensor("op_644_end_0"), val = tensor([1, 16, 256])]; - tensor var_644_end_mask_0 = const()[name = tensor("op_644_end_mask_0"), val = tensor([true, true, true])]; - tensor var_644 = slice_by_index(begin = var_644_begin_0, end = var_644_end_0, end_mask = var_644_end_mask_0, x = window_19)[name = tensor("op_644")]; + tensor window_19 = concat(axis = var_72, interleave = window_19_interleave_0, values = (var_682, var_679))[name = tensor("window_19")]; + tensor var_687_begin_0 = const()[name = tensor("op_687_begin_0"), val = tensor([0, 1, 0])]; + tensor var_687_end_0 = const()[name = tensor("op_687_end_0"), val = tensor([1, 2, 256])]; + tensor var_687_end_mask_0 = const()[name = tensor("op_687_end_mask_0"), val = tensor([true, false, true])]; + tensor var_687 = slice_by_index(begin = var_687_begin_0, end = var_687_end_0, end_mask = var_687_end_mask_0, x = x_15)[name = tensor("op_687")]; + tensor var_690_begin_0 = const()[name = tensor("op_690_begin_0"), val = tensor([0, 1, 0])]; + tensor var_690_end_0 = const()[name = tensor("op_690_end_0"), val = tensor([1, 16, 256])]; + tensor var_690_end_mask_0 = const()[name = tensor("op_690_end_mask_0"), val = tensor([true, true, true])]; + tensor var_690 = slice_by_index(begin = var_690_begin_0, end = var_690_end_0, end_mask = var_690_end_mask_0, x = window_19)[name = tensor("op_690")]; tensor window_21_interleave_0 = const()[name = tensor("window_21_interleave_0"), val = tensor(false)]; - tensor window_21 = concat(axis = var_26, interleave = window_21_interleave_0, values = (var_644, var_641))[name = tensor("window_21")]; - tensor var_649_begin_0 = const()[name = tensor("op_649_begin_0"), val = tensor([0, 2, 0])]; - tensor var_649_end_0 = const()[name = tensor("op_649_end_0"), val = tensor([1, 1, 256])]; - tensor var_649_end_mask_0 = const()[name = tensor("op_649_end_mask_0"), val = tensor([true, true, true])]; - tensor var_649 = slice_by_index(begin = var_649_begin_0, end = var_649_end_0, end_mask = var_649_end_mask_0, x = x_15)[name = tensor("op_649")]; - tensor var_652_begin_0 = const()[name = tensor("op_652_begin_0"), val = tensor([0, 1, 0])]; - tensor var_652_end_0 = const()[name = tensor("op_652_end_0"), val = tensor([1, 16, 256])]; - tensor var_652_end_mask_0 = const()[name = tensor("op_652_end_mask_0"), val = tensor([true, true, true])]; - tensor var_652 = slice_by_index(begin = var_652_begin_0, end = var_652_end_0, end_mask = var_652_end_mask_0, x = window_21)[name = tensor("op_652")]; + tensor window_21 = concat(axis = var_72, interleave = window_21_interleave_0, values = (var_690, var_687))[name = tensor("window_21")]; + tensor var_695_begin_0 = const()[name = tensor("op_695_begin_0"), val = tensor([0, 2, 0])]; + tensor var_695_end_0 = const()[name = tensor("op_695_end_0"), val = tensor([1, 1, 256])]; + tensor var_695_end_mask_0 = const()[name = tensor("op_695_end_mask_0"), val = tensor([true, true, true])]; + tensor var_695 = slice_by_index(begin = var_695_begin_0, end = var_695_end_0, end_mask = var_695_end_mask_0, x = x_15)[name = tensor("op_695")]; + tensor var_698_begin_0 = const()[name = tensor("op_698_begin_0"), val = tensor([0, 1, 0])]; + tensor var_698_end_0 = const()[name = tensor("op_698_end_0"), val = tensor([1, 16, 256])]; + tensor var_698_end_mask_0 = const()[name = tensor("op_698_end_mask_0"), val = tensor([true, true, true])]; + tensor var_698 = slice_by_index(begin = var_698_begin_0, end = var_698_end_0, end_mask = var_698_end_mask_0, x = window_21)[name = tensor("op_698")]; tensor window_23_interleave_0 = const()[name = tensor("window_23_interleave_0"), val = tensor(false)]; - tensor window_23 = concat(axis = var_26, interleave = window_23_interleave_0, values = (var_652, var_649))[name = tensor("window_23")]; - tensor input_101_interleave_0 = const()[name = tensor("input_101_interleave_0"), val = tensor(false)]; - tensor input_101 = concat(axis = var_23, interleave = input_101_interleave_0, values = (window_19, window_21, window_23))[name = tensor("input_101")]; + tensor window_23 = concat(axis = var_72, interleave = window_23_interleave_0, values = (var_698, var_695))[name = tensor("window_23")]; + tensor input_103_interleave_0 = const()[name = tensor("input_103_interleave_0"), val = tensor(false)]; + tensor input_103 = concat(axis = var_58, interleave = input_103_interleave_0, values = (window_19, window_21, window_23))[name = tensor("input_103")]; tensor x_17_axes_0 = const()[name = tensor("x_17_axes_0"), val = tensor([-1])]; - tensor x_17 = layer_norm(axes = x_17_axes_0, beta = encoder_conv_module_2_sequential_0_bias, epsilon = var_28, gamma = encoder_conv_module_2_sequential_0_weight, x = input_101)[name = tensor("x_17")]; - tensor input_103_perm_0 = const()[name = tensor("input_103_perm_0"), val = tensor([0, 2, 1])]; + tensor x_17 = layer_norm(axes = x_17_axes_0, beta = inner_encoder_conv_module_2_sequential_0_bias, epsilon = var_55, gamma = inner_encoder_conv_module_2_sequential_0_weight, x = input_103)[name = tensor("x_17")]; + tensor input_105_perm_0 = const()[name = tensor("input_105_perm_0"), val = tensor([0, 2, 1])]; tensor inputs_23_pad_type_0 = const()[name = tensor("inputs_23_pad_type_0"), val = tensor("valid")]; tensor inputs_23_strides_0 = const()[name = tensor("inputs_23_strides_0"), val = tensor([1])]; tensor inputs_23_pad_0 = const()[name = tensor("inputs_23_pad_0"), val = tensor([0, 0])]; tensor inputs_23_dilations_0 = const()[name = tensor("inputs_23_dilations_0"), val = tensor([1])]; tensor inputs_23_groups_0 = const()[name = tensor("inputs_23_groups_0"), val = tensor(1)]; - tensor input_103 = transpose(perm = input_103_perm_0, x = x_17)[name = tensor("transpose_40")]; - tensor inputs_23 = conv(bias = encoder_conv_module_2_sequential_2_conv_bias, dilations = inputs_23_dilations_0, groups = inputs_23_groups_0, pad = inputs_23_pad_0, pad_type = inputs_23_pad_type_0, strides = inputs_23_strides_0, weight = encoder_conv_module_2_sequential_2_conv_weight, x = input_103)[name = tensor("inputs_23")]; - tensor var_677_split_sizes_0 = const()[name = tensor("op_677_split_sizes_0"), val = tensor([256, 256])]; - tensor var_677_axis_0 = const()[name = tensor("op_677_axis_0"), val = tensor(1)]; - tensor var_677_0, tensor var_677_1 = split(axis = var_677_axis_0, split_sizes = var_677_split_sizes_0, x = inputs_23)[name = tensor("op_677")]; - tensor var_679 = sigmoid(x = var_677_1)[name = tensor("op_679")]; - tensor inputs_25 = mul(x = var_677_0, y = var_679)[name = tensor("inputs_25")]; + tensor input_105 = transpose(perm = input_105_perm_0, x = x_17)[name = tensor("transpose_40")]; + tensor inputs_23 = conv(bias = inner_encoder_conv_module_2_sequential_2_conv_bias, dilations = inputs_23_dilations_0, groups = inputs_23_groups_0, pad = inputs_23_pad_0, pad_type = inputs_23_pad_type_0, strides = inputs_23_strides_0, weight = inner_encoder_conv_module_2_sequential_2_conv_weight, x = input_105)[name = tensor("inputs_23")]; + tensor var_723_split_sizes_0 = const()[name = tensor("op_723_split_sizes_0"), val = tensor([256, 256])]; + tensor var_723_axis_0 = const()[name = tensor("op_723_axis_0"), val = tensor(1)]; + tensor var_723_0, tensor var_723_1 = split(axis = var_723_axis_0, split_sizes = var_723_split_sizes_0, x = inputs_23)[name = tensor("op_723")]; + tensor var_725 = sigmoid(x = var_723_1)[name = tensor("op_725")]; + tensor inputs_25 = mul(x = var_723_0, y = var_725)[name = tensor("inputs_25")]; tensor outputs_aug_5_pad_type_0 = const()[name = tensor("outputs_aug_5_pad_type_0"), val = tensor("custom")]; tensor outputs_aug_5_pad_0 = const()[name = tensor("outputs_aug_5_pad_0"), val = tensor([15, 15])]; tensor outputs_aug_5_groups_0 = const()[name = tensor("outputs_aug_5_groups_0"), val = tensor(256)]; tensor outputs_aug_5_strides_0 = const()[name = tensor("outputs_aug_5_strides_0"), val = tensor([1])]; tensor outputs_aug_5_dilations_0 = const()[name = tensor("outputs_aug_5_dilations_0"), val = tensor([1])]; - tensor outputs_aug_5 = conv(dilations = outputs_aug_5_dilations_0, groups = outputs_aug_5_groups_0, pad = outputs_aug_5_pad_0, pad_type = outputs_aug_5_pad_type_0, strides = outputs_aug_5_strides_0, weight = encoder_conv_module_2_sequential_4_conv_weight, x = inputs_25)[name = tensor("outputs_aug_5")]; - tensor input_105_begin_0 = const()[name = tensor("input_105_begin_0"), val = tensor([0, 0, 0])]; - tensor input_105_end_0 = const()[name = tensor("input_105_end_0"), val = tensor([3, 256, 16])]; - tensor input_105_end_mask_0 = const()[name = tensor("input_105_end_mask_0"), val = tensor([true, true, false])]; - tensor input_105 = slice_by_index(begin = input_105_begin_0, end = input_105_end_0, end_mask = input_105_end_mask_0, x = outputs_aug_5)[name = tensor("input_105")]; - tensor inputs_27 = batch_norm(beta = encoder_conv_module_2_sequential_5_bias, epsilon = var_28, gamma = encoder_conv_module_2_sequential_5_weight, mean = encoder_conv_module_2_sequential_5_running_mean, variance = encoder_conv_module_2_sequential_5_running_var, x = input_105)[name = tensor("inputs_27")]; - tensor input_107 = silu(x = inputs_27)[name = tensor("input_107")]; - tensor input_109_pad_type_0 = const()[name = tensor("input_109_pad_type_0"), val = tensor("valid")]; - tensor input_109_strides_0 = const()[name = tensor("input_109_strides_0"), val = tensor([1])]; - tensor input_109_pad_0 = const()[name = tensor("input_109_pad_0"), val = tensor([0, 0])]; - tensor input_109_dilations_0 = const()[name = tensor("input_109_dilations_0"), val = tensor([1])]; - tensor input_109_groups_0 = const()[name = tensor("input_109_groups_0"), val = tensor(1)]; - tensor input_109 = conv(bias = encoder_conv_module_2_sequential_7_conv_bias, dilations = input_109_dilations_0, groups = input_109_groups_0, pad = input_109_pad_0, pad_type = input_109_pad_type_0, strides = input_109_strides_0, weight = encoder_conv_module_2_sequential_7_conv_weight, x = input_107)[name = tensor("input_109")]; + tensor outputs_aug_5 = conv(dilations = outputs_aug_5_dilations_0, groups = outputs_aug_5_groups_0, pad = outputs_aug_5_pad_0, pad_type = outputs_aug_5_pad_type_0, strides = outputs_aug_5_strides_0, weight = inner_encoder_conv_module_2_sequential_4_conv_weight, x = inputs_25)[name = tensor("outputs_aug_5")]; + tensor input_107_begin_0 = const()[name = tensor("input_107_begin_0"), val = tensor([0, 0, 0])]; + tensor input_107_end_0 = const()[name = tensor("input_107_end_0"), val = tensor([3, 256, 16])]; + tensor input_107_end_mask_0 = const()[name = tensor("input_107_end_mask_0"), val = tensor([true, true, false])]; + tensor input_107 = slice_by_index(begin = input_107_begin_0, end = input_107_end_0, end_mask = input_107_end_mask_0, x = outputs_aug_5)[name = tensor("input_107")]; + tensor inputs_27 = batch_norm(beta = inner_encoder_conv_module_2_sequential_5_bias, epsilon = var_55, gamma = inner_encoder_conv_module_2_sequential_5_weight, mean = inner_encoder_conv_module_2_sequential_5_running_mean, variance = inner_encoder_conv_module_2_sequential_5_running_var, x = input_107)[name = tensor("inputs_27")]; + tensor input_109 = silu(x = inputs_27)[name = tensor("input_109")]; + tensor input_111_pad_type_0 = const()[name = tensor("input_111_pad_type_0"), val = tensor("valid")]; + tensor input_111_strides_0 = const()[name = tensor("input_111_strides_0"), val = tensor([1])]; + tensor input_111_pad_0 = const()[name = tensor("input_111_pad_0"), val = tensor([0, 0])]; + tensor input_111_dilations_0 = const()[name = tensor("input_111_dilations_0"), val = tensor([1])]; + tensor input_111_groups_0 = const()[name = tensor("input_111_groups_0"), val = tensor(1)]; + tensor input_111 = conv(bias = inner_encoder_conv_module_2_sequential_7_conv_bias, dilations = input_111_dilations_0, groups = input_111_groups_0, pad = input_111_pad_0, pad_type = input_111_pad_type_0, strides = input_111_strides_0, weight = inner_encoder_conv_module_2_sequential_7_conv_weight, x = input_109)[name = tensor("input_111")]; tensor conv_out_5_perm_0 = const()[name = tensor("conv_out_5_perm_0"), val = tensor([0, 2, 1])]; - tensor var_710_begin_0 = const()[name = tensor("op_710_begin_0"), val = tensor([0, -1, 0])]; - tensor var_710_end_0 = const()[name = tensor("op_710_end_0"), val = tensor([3, 16, 256])]; - tensor var_710_end_mask_0 = const()[name = tensor("op_710_end_mask_0"), val = tensor([true, true, true])]; - tensor conv_out_5 = transpose(perm = conv_out_5_perm_0, x = input_109)[name = tensor("transpose_39")]; - tensor var_710 = slice_by_index(begin = var_710_begin_0, end = var_710_end_0, end_mask = var_710_end_mask_0, x = conv_out_5)[name = tensor("op_710")]; - tensor var_712_perm_0 = const()[name = tensor("op_712_perm_0"), val = tensor([1, 0, 2])]; - tensor var_712 = transpose(perm = var_712_perm_0, x = var_710)[name = tensor("transpose_38")]; - tensor input_111 = add(x = x_15, y = var_712)[name = tensor("input_111")]; - tensor input_113_axes_0 = const()[name = tensor("input_113_axes_0"), val = tensor([-1])]; - tensor input_113 = layer_norm(axes = input_113_axes_0, beta = encoder_ffn2_2_module_sequential_0_bias, epsilon = var_28, gamma = encoder_ffn2_2_module_sequential_0_weight, x = input_111)[name = tensor("input_113")]; - tensor inputs_29 = linear(bias = encoder_ffn2_2_module_sequential_1_linear_bias, weight = encoder_ffn2_2_module_sequential_1_linear_weight, x = input_113)[name = tensor("linear_26")]; - tensor input_115 = silu(x = inputs_29)[name = tensor("input_115")]; - tensor input_119 = linear(bias = encoder_ffn2_2_module_sequential_4_linear_bias, weight = encoder_ffn2_2_module_sequential_4_linear_weight, x = input_115)[name = tensor("linear_27")]; - tensor var_735 = const()[name = tensor("op_735"), val = tensor(0x1p-1)]; - tensor var_736 = mul(x = input_119, y = var_735)[name = tensor("op_736")]; - tensor input_121 = add(x = var_736, y = input_111)[name = tensor("input_121")]; - tensor input_123_axes_0 = const()[name = tensor("input_123_axes_0"), val = tensor([-1])]; - tensor input_123 = layer_norm(axes = input_123_axes_0, beta = encoder_layer_norm_2_bias, epsilon = var_28, gamma = encoder_layer_norm_2_weight, x = input_121)[name = tensor("input_123")]; + tensor var_756_begin_0 = const()[name = tensor("op_756_begin_0"), val = tensor([0, -1, 0])]; + tensor var_756_end_0 = const()[name = tensor("op_756_end_0"), val = tensor([3, 16, 256])]; + tensor var_756_end_mask_0 = const()[name = tensor("op_756_end_mask_0"), val = tensor([true, true, true])]; + tensor conv_out_5 = transpose(perm = conv_out_5_perm_0, x = input_111)[name = tensor("transpose_39")]; + tensor var_756 = slice_by_index(begin = var_756_begin_0, end = var_756_end_0, end_mask = var_756_end_mask_0, x = conv_out_5)[name = tensor("op_756")]; + tensor var_758_perm_0 = const()[name = tensor("op_758_perm_0"), val = tensor([1, 0, 2])]; + tensor var_758 = transpose(perm = var_758_perm_0, x = var_756)[name = tensor("transpose_38")]; + tensor input_113 = add(x = x_15, y = var_758)[name = tensor("input_113")]; + tensor input_115_axes_0 = const()[name = tensor("input_115_axes_0"), val = tensor([-1])]; + tensor input_115 = layer_norm(axes = input_115_axes_0, beta = inner_encoder_ffn2_2_module_sequential_0_bias, epsilon = var_55, gamma = inner_encoder_ffn2_2_module_sequential_0_weight, x = input_113)[name = tensor("input_115")]; + tensor inputs_29 = linear(bias = inner_encoder_ffn2_2_module_sequential_1_linear_bias, weight = inner_encoder_ffn2_2_module_sequential_1_linear_weight, x = input_115)[name = tensor("linear_26")]; + tensor input_117 = silu(x = inputs_29)[name = tensor("input_117")]; + tensor input_121 = linear(bias = inner_encoder_ffn2_2_module_sequential_4_linear_bias, weight = inner_encoder_ffn2_2_module_sequential_4_linear_weight, x = input_117)[name = tensor("linear_27")]; + tensor var_781 = const()[name = tensor("op_781"), val = tensor(0x1p-1)]; + tensor var_782 = mul(x = input_121, y = var_781)[name = tensor("op_782")]; + tensor input_123 = add(x = var_782, y = input_113)[name = tensor("input_123")]; tensor input_125_axes_0 = const()[name = tensor("input_125_axes_0"), val = tensor([-1])]; - tensor input_125 = layer_norm(axes = input_125_axes_0, beta = encoder_ffn1_3_module_sequential_0_bias, epsilon = var_28, gamma = encoder_ffn1_3_module_sequential_0_weight, x = input_123)[name = tensor("input_125")]; - tensor inputs_31 = linear(bias = encoder_ffn1_3_module_sequential_1_linear_bias, weight = encoder_ffn1_3_module_sequential_1_linear_weight, x = input_125)[name = tensor("linear_28")]; - tensor input_127 = silu(x = inputs_31)[name = tensor("input_127")]; - tensor input_131 = linear(bias = encoder_ffn1_3_module_sequential_4_linear_bias, weight = encoder_ffn1_3_module_sequential_4_linear_weight, x = input_127)[name = tensor("linear_29")]; - tensor var_765 = const()[name = tensor("op_765"), val = tensor(0x1p-1)]; - tensor var_766 = mul(x = input_131, y = var_765)[name = tensor("op_766")]; - tensor input_133 = add(x = var_766, y = input_123)[name = tensor("input_133")]; + tensor input_125 = layer_norm(axes = input_125_axes_0, beta = inner_encoder_layer_norm_2_bias, epsilon = var_55, gamma = inner_encoder_layer_norm_2_weight, x = input_123)[name = tensor("input_125")]; + tensor input_127_axes_0 = const()[name = tensor("input_127_axes_0"), val = tensor([-1])]; + tensor input_127 = layer_norm(axes = input_127_axes_0, beta = inner_encoder_ffn1_3_module_sequential_0_bias, epsilon = var_55, gamma = inner_encoder_ffn1_3_module_sequential_0_weight, x = input_125)[name = tensor("input_127")]; + tensor inputs_31 = linear(bias = inner_encoder_ffn1_3_module_sequential_1_linear_bias, weight = inner_encoder_ffn1_3_module_sequential_1_linear_weight, x = input_127)[name = tensor("linear_28")]; + tensor input_129 = silu(x = inputs_31)[name = tensor("input_129")]; + tensor input_133 = linear(bias = inner_encoder_ffn1_3_module_sequential_4_linear_bias, weight = inner_encoder_ffn1_3_module_sequential_4_linear_weight, x = input_129)[name = tensor("linear_29")]; + tensor var_811 = const()[name = tensor("op_811"), val = tensor(0x1p-1)]; + tensor var_812 = mul(x = input_133, y = var_811)[name = tensor("op_812")]; + tensor input_135 = add(x = var_812, y = input_125)[name = tensor("input_135")]; tensor x_19_axes_0 = const()[name = tensor("x_19_axes_0"), val = tensor([-1])]; - tensor x_19 = layer_norm(axes = x_19_axes_0, beta = encoder_ret_lns_3_bias, epsilon = var_28, gamma = encoder_ret_lns_3_weight, x = input_133)[name = tensor("x_19")]; + tensor x_19 = layer_norm(axes = x_19_axes_0, beta = inner_encoder_ret_lns_3_bias, epsilon = var_55, gamma = inner_encoder_ret_lns_3_weight, x = input_135)[name = tensor("x_19")]; tensor prev_kv_7_begin_0 = const()[name = tensor("prev_kv_7_begin_0"), val = tensor([3, 0, 0, 0, 0])]; tensor prev_kv_7_end_0 = const()[name = tensor("prev_kv_7_end_0"), val = tensor([4, 1, 4, 64, 64])]; tensor prev_kv_7_end_mask_0 = const()[name = tensor("prev_kv_7_end_mask_0"), val = tensor([false, true, true, true, true])]; @@ -740,199 +758,192 @@ program(1.0) tensor prev_scale_7_end_mask_0 = const()[name = tensor("prev_scale_7_end_mask_0"), val = tensor([false, true])]; tensor prev_scale_7_squeeze_mask_0 = const()[name = tensor("prev_scale_7_squeeze_mask_0"), val = tensor([true, false])]; tensor prev_scale_7 = slice_by_index(begin = prev_scale_7_begin_0, end = prev_scale_7_end_0, end_mask = prev_scale_7_end_mask_0, squeeze_mask = prev_scale_7_squeeze_mask_0, x = enc_scale)[name = tensor("prev_scale_7")]; - tensor var_780 = linear(bias = encoder_q_proj_3_bias, weight = encoder_q_proj_3_weight, x = x_19)[name = tensor("linear_30")]; - tensor var_781 = const()[name = tensor("op_781"), val = tensor([1, 3, 4, 64])]; - tensor var_782 = reshape(shape = var_781, x = var_780)[name = tensor("op_782")]; + tensor var_826 = linear(bias = inner_encoder_q_proj_3_bias, weight = inner_encoder_q_proj_3_weight, x = x_19)[name = tensor("linear_30")]; + tensor var_827 = const()[name = tensor("op_827"), val = tensor([1, 3, 4, 64])]; + tensor var_828 = reshape(shape = var_827, x = var_826)[name = tensor("op_828")]; tensor q_7_perm_0 = const()[name = tensor("q_7_perm_0"), val = tensor([0, 2, 1, 3])]; - tensor var_786 = linear(bias = encoder_k_proj_3_bias, weight = encoder_k_proj_3_weight, x = x_19)[name = tensor("linear_31")]; - tensor var_787 = const()[name = tensor("op_787"), val = tensor(0x1p-3)]; - tensor var_788 = mul(x = var_786, y = var_787)[name = tensor("op_788")]; - tensor var_789 = const()[name = tensor("op_789"), val = tensor([1, 3, 4, 64])]; - tensor var_790 = reshape(shape = var_789, x = var_788)[name = tensor("op_790")]; + tensor var_832 = linear(bias = inner_encoder_k_proj_3_bias, weight = inner_encoder_k_proj_3_weight, x = x_19)[name = tensor("linear_31")]; + tensor var_833 = const()[name = tensor("op_833"), val = tensor(0x1p-3)]; + tensor var_834 = mul(x = var_832, y = var_833)[name = tensor("op_834")]; + tensor var_835 = const()[name = tensor("op_835"), val = tensor([1, 3, 4, 64])]; + tensor var_836 = reshape(shape = var_835, x = var_834)[name = tensor("op_836")]; tensor k_7_perm_0 = const()[name = tensor("k_7_perm_0"), val = tensor([0, 2, 1, 3])]; - tensor var_794 = linear(bias = encoder_v_proj_3_bias, weight = encoder_v_proj_3_weight, x = x_19)[name = tensor("linear_32")]; - tensor var_795 = const()[name = tensor("op_795"), val = tensor([1, 3, 4, 64])]; - tensor var_796 = reshape(shape = var_795, x = var_794)[name = tensor("op_796")]; + tensor var_840 = linear(bias = inner_encoder_v_proj_3_bias, weight = inner_encoder_v_proj_3_weight, x = x_19)[name = tensor("linear_32")]; + tensor var_841 = const()[name = tensor("op_841"), val = tensor([1, 3, 4, 64])]; + tensor var_842 = reshape(shape = var_841, x = var_840)[name = tensor("op_842")]; tensor v_7_perm_0 = const()[name = tensor("v_7_perm_0"), val = tensor([0, 2, 1, 3])]; - tensor input_137 = linear(bias = encoder_g_proj_3_bias, weight = encoder_g_proj_3_weight, x = x_19)[name = tensor("linear_33")]; + tensor input_139 = linear(bias = inner_encoder_g_proj_3_bias, weight = inner_encoder_g_proj_3_weight, x = x_19)[name = tensor("linear_33")]; tensor sqrt_s0_7 = sqrt(x = prev_scale_7)[name = tensor("sqrt_s0_7")]; - tensor s_t_7 = add(x = prev_scale_7, y = encoder__t_index)[name = tensor("s_t_7")]; + tensor s_t_7 = add(x = prev_scale_7, y = inner_encoder__t_index)[name = tensor("s_t_7")]; tensor sqrt_s_t_7 = sqrt(x = s_t_7)[name = tensor("sqrt_s_t_7")]; tensor qk_7_transpose_x_1 = const()[name = tensor("qk_7_transpose_x_1"), val = tensor(false)]; tensor qk_7_transpose_y_1 = const()[name = tensor("qk_7_transpose_y_1"), val = tensor(true)]; - tensor k_7 = transpose(perm = k_7_perm_0, x = var_790)[name = tensor("transpose_36")]; - tensor q_7 = transpose(perm = q_7_perm_0, x = var_782)[name = tensor("transpose_37")]; + tensor k_7 = transpose(perm = k_7_perm_0, x = var_836)[name = tensor("transpose_36")]; + tensor q_7 = transpose(perm = q_7_perm_0, x = var_828)[name = tensor("transpose_37")]; tensor qk_7 = matmul(transpose_x = qk_7_transpose_x_1, transpose_y = qk_7_transpose_y_1, x = q_7, y = k_7)[name = tensor("qk_7")]; - tensor var_806 = const()[name = tensor("op_806"), val = tensor([3, 1])]; - tensor var_807 = reshape(shape = var_806, x = sqrt_s_t_7)[name = tensor("op_807")]; - tensor M_7 = real_div(x = encoder__causal_mask, y = var_807)[name = tensor("M_7")]; - tensor var_809 = mul(x = qk_7, y = M_7)[name = tensor("op_809")]; + tensor var_852 = const()[name = tensor("op_852"), val = tensor([3, 1])]; + tensor var_853 = reshape(shape = var_852, x = sqrt_s_t_7)[name = tensor("op_853")]; + tensor M_7 = real_div(x = inner_encoder__causal_mask, y = var_853)[name = tensor("M_7")]; + tensor var_855 = mul(x = qk_7, y = M_7)[name = tensor("op_855")]; tensor inner_7_transpose_x_0 = const()[name = tensor("inner_7_transpose_x_0"), val = tensor(false)]; tensor inner_7_transpose_y_0 = const()[name = tensor("inner_7_transpose_y_0"), val = tensor(false)]; - tensor v_7 = transpose(perm = v_7_perm_0, x = var_796)[name = tensor("transpose_35")]; - tensor inner_7 = matmul(transpose_x = inner_7_transpose_x_0, transpose_y = inner_7_transpose_y_0, x = var_809, y = v_7)[name = tensor("inner_7")]; - tensor var_811_transpose_x_0 = const()[name = tensor("op_811_transpose_x_0"), val = tensor(false)]; - tensor var_811_transpose_y_0 = const()[name = tensor("op_811_transpose_y_0"), val = tensor(false)]; - tensor var_811 = matmul(transpose_x = var_811_transpose_x_0, transpose_y = var_811_transpose_y_0, x = q_7, y = prev_kv_7)[name = tensor("op_811")]; - tensor var_812 = real_div(x = sqrt_s0_7, y = sqrt_s_t_7)[name = tensor("op_812")]; - tensor var_813 = const()[name = tensor("op_813"), val = tensor([1, 1, 3, 1])]; - tensor var_814 = reshape(shape = var_813, x = var_812)[name = tensor("op_814")]; - tensor cross_7 = mul(x = var_811, y = var_814)[name = tensor("cross_7")]; + tensor v_7 = transpose(perm = v_7_perm_0, x = var_842)[name = tensor("transpose_35")]; + tensor inner_7 = matmul(transpose_x = inner_7_transpose_x_0, transpose_y = inner_7_transpose_y_0, x = var_855, y = v_7)[name = tensor("inner_7")]; + tensor var_857_transpose_x_0 = const()[name = tensor("op_857_transpose_x_0"), val = tensor(false)]; + tensor var_857_transpose_y_0 = const()[name = tensor("op_857_transpose_y_0"), val = tensor(false)]; + tensor var_857 = matmul(transpose_x = var_857_transpose_x_0, transpose_y = var_857_transpose_y_0, x = q_7, y = prev_kv_7)[name = tensor("op_857")]; + tensor var_858 = real_div(x = sqrt_s0_7, y = sqrt_s_t_7)[name = tensor("op_858")]; + tensor var_859 = const()[name = tensor("op_859"), val = tensor([1, 1, 3, 1])]; + tensor var_860 = reshape(shape = var_859, x = var_858)[name = tensor("op_860")]; + tensor cross_7 = mul(x = var_857, y = var_860)[name = tensor("cross_7")]; tensor out_19 = add(x = inner_7, y = cross_7)[name = tensor("out_19")]; - tensor var_817 = mul(x = prev_kv_7, y = sqrt_s0_7)[name = tensor("op_817")]; - tensor var_819_transpose_x_1 = const()[name = tensor("op_819_transpose_x_1"), val = tensor(true)]; - tensor var_819_transpose_y_1 = const()[name = tensor("op_819_transpose_y_1"), val = tensor(false)]; - tensor var_819 = matmul(transpose_x = var_819_transpose_x_1, transpose_y = var_819_transpose_y_1, x = k_7, y = v_7)[name = tensor("op_819")]; - tensor new_kv_unnorm_7 = add(x = var_817, y = var_819)[name = tensor("new_kv_unnorm_7")]; - tensor var_821 = const()[name = tensor("op_821"), val = tensor(0x1.8p+1)]; - tensor new_scale_7 = add(x = prev_scale_7, y = var_821)[name = tensor("new_scale_7")]; - tensor var_823 = sqrt(x = new_scale_7)[name = tensor("op_823")]; - tensor nkv_1 = real_div(x = new_kv_unnorm_7, y = var_823)[name = tensor("nkv_1")]; - tensor var_825_perm_0 = const()[name = tensor("op_825_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_863 = mul(x = prev_kv_7, y = sqrt_s0_7)[name = tensor("op_863")]; + tensor var_865_transpose_x_1 = const()[name = tensor("op_865_transpose_x_1"), val = tensor(true)]; + tensor var_865_transpose_y_1 = const()[name = tensor("op_865_transpose_y_1"), val = tensor(false)]; + tensor var_865 = matmul(transpose_x = var_865_transpose_x_1, transpose_y = var_865_transpose_y_1, x = k_7, y = v_7)[name = tensor("op_865")]; + tensor new_kv_unnorm_7 = add(x = var_863, y = var_865)[name = tensor("new_kv_unnorm_7")]; + tensor var_867 = const()[name = tensor("op_867"), val = tensor(0x1.8p+1)]; + tensor new_scale_7 = add(x = prev_scale_7, y = var_867)[name = tensor("new_scale_7")]; + tensor var_869 = sqrt(x = new_scale_7)[name = tensor("op_869")]; + tensor nkv_1 = real_div(x = new_kv_unnorm_7, y = var_869)[name = tensor("nkv_1")]; + tensor var_871_perm_0 = const()[name = tensor("op_871_perm_0"), val = tensor([0, 2, 1, 3])]; tensor out_21_axes_0 = const()[name = tensor("out_21_axes_0"), val = tensor([-1])]; - tensor var_825 = transpose(perm = var_825_perm_0, x = out_19)[name = tensor("transpose_34")]; - tensor out_21 = layer_norm(axes = out_21_axes_0, epsilon = var_17, x = var_825)[name = tensor("out_21")]; - tensor var_829 = const()[name = tensor("op_829"), val = tensor([1, 3, 256])]; - tensor out_23 = reshape(shape = var_829, x = out_21)[name = tensor("out_23")]; - tensor var_831 = silu(x = input_137)[name = tensor("op_831")]; - tensor input_139 = mul(x = var_831, y = out_23)[name = tensor("input_139")]; - tensor ret_out_7 = linear(bias = encoder_out_proj_3_bias, weight = encoder_out_proj_3_weight, x = input_139)[name = tensor("linear_34")]; - tensor x_21 = add(x = input_133, y = ret_out_7)[name = tensor("x_21")]; + tensor var_871 = transpose(perm = var_871_perm_0, x = out_19)[name = tensor("transpose_34")]; + tensor out_21 = layer_norm(axes = out_21_axes_0, epsilon = var_63, x = var_871)[name = tensor("out_21")]; + tensor var_875 = const()[name = tensor("op_875"), val = tensor([1, 3, 256])]; + tensor out_23 = reshape(shape = var_875, x = out_21)[name = tensor("out_23")]; + tensor var_877 = silu(x = input_139)[name = tensor("op_877")]; + tensor input_141 = mul(x = var_877, y = out_23)[name = tensor("input_141")]; + tensor ret_out_7 = linear(bias = inner_encoder_out_proj_3_bias, weight = inner_encoder_out_proj_3_weight, x = input_141)[name = tensor("linear_34")]; + tensor x_21 = add(x = input_135, y = ret_out_7)[name = tensor("x_21")]; tensor window_25_begin_0 = const()[name = tensor("window_25_begin_0"), val = tensor([3, 0, 0, 0])]; tensor window_25_end_0 = const()[name = tensor("window_25_end_0"), val = tensor([4, 1, 16, 256])]; tensor window_25_end_mask_0 = const()[name = tensor("window_25_end_mask_0"), val = tensor([false, true, true, true])]; tensor window_25_squeeze_mask_0 = const()[name = tensor("window_25_squeeze_mask_0"), val = tensor([true, false, false, false])]; tensor window_25 = slice_by_index(begin = window_25_begin_0, end = window_25_end_0, end_mask = window_25_end_mask_0, squeeze_mask = window_25_squeeze_mask_0, x = enc_conv_cache)[name = tensor("window_25")]; - tensor var_839_begin_0 = const()[name = tensor("op_839_begin_0"), val = tensor([0, 0, 0])]; - tensor var_839_end_0 = const()[name = tensor("op_839_end_0"), val = tensor([1, 1, 256])]; - tensor var_839_end_mask_0 = const()[name = tensor("op_839_end_mask_0"), val = tensor([true, false, true])]; - tensor var_839 = slice_by_index(begin = var_839_begin_0, end = var_839_end_0, end_mask = var_839_end_mask_0, x = x_21)[name = tensor("op_839")]; - tensor var_842_begin_0 = const()[name = tensor("op_842_begin_0"), val = tensor([0, 1, 0])]; - tensor var_842_end_0 = const()[name = tensor("op_842_end_0"), val = tensor([1, 16, 256])]; - tensor var_842_end_mask_0 = const()[name = tensor("op_842_end_mask_0"), val = tensor([true, true, true])]; - tensor var_842 = slice_by_index(begin = var_842_begin_0, end = var_842_end_0, end_mask = var_842_end_mask_0, x = window_25)[name = tensor("op_842")]; + tensor var_885_begin_0 = const()[name = tensor("op_885_begin_0"), val = tensor([0, 0, 0])]; + tensor var_885_end_0 = const()[name = tensor("op_885_end_0"), val = tensor([1, 1, 256])]; + tensor var_885_end_mask_0 = const()[name = tensor("op_885_end_mask_0"), val = tensor([true, false, true])]; + tensor var_885 = slice_by_index(begin = var_885_begin_0, end = var_885_end_0, end_mask = var_885_end_mask_0, x = x_21)[name = tensor("op_885")]; + tensor var_888_begin_0 = const()[name = tensor("op_888_begin_0"), val = tensor([0, 1, 0])]; + tensor var_888_end_0 = const()[name = tensor("op_888_end_0"), val = tensor([1, 16, 256])]; + tensor var_888_end_mask_0 = const()[name = tensor("op_888_end_mask_0"), val = tensor([true, true, true])]; + tensor var_888 = slice_by_index(begin = var_888_begin_0, end = var_888_end_0, end_mask = var_888_end_mask_0, x = window_25)[name = tensor("op_888")]; tensor window_27_interleave_0 = const()[name = tensor("window_27_interleave_0"), val = tensor(false)]; - tensor window_27 = concat(axis = var_26, interleave = window_27_interleave_0, values = (var_842, var_839))[name = tensor("window_27")]; - tensor var_847_begin_0 = const()[name = tensor("op_847_begin_0"), val = tensor([0, 1, 0])]; - tensor var_847_end_0 = const()[name = tensor("op_847_end_0"), val = tensor([1, 2, 256])]; - tensor var_847_end_mask_0 = const()[name = tensor("op_847_end_mask_0"), val = tensor([true, false, true])]; - tensor var_847 = slice_by_index(begin = var_847_begin_0, end = var_847_end_0, end_mask = var_847_end_mask_0, x = x_21)[name = tensor("op_847")]; - tensor var_850_begin_0 = const()[name = tensor("op_850_begin_0"), val = tensor([0, 1, 0])]; - tensor var_850_end_0 = const()[name = tensor("op_850_end_0"), val = tensor([1, 16, 256])]; - tensor var_850_end_mask_0 = const()[name = tensor("op_850_end_mask_0"), val = tensor([true, true, true])]; - tensor var_850 = slice_by_index(begin = var_850_begin_0, end = var_850_end_0, end_mask = var_850_end_mask_0, x = window_27)[name = tensor("op_850")]; + tensor window_27 = concat(axis = var_72, interleave = window_27_interleave_0, values = (var_888, var_885))[name = tensor("window_27")]; + tensor var_893_begin_0 = const()[name = tensor("op_893_begin_0"), val = tensor([0, 1, 0])]; + tensor var_893_end_0 = const()[name = tensor("op_893_end_0"), val = tensor([1, 2, 256])]; + tensor var_893_end_mask_0 = const()[name = tensor("op_893_end_mask_0"), val = tensor([true, false, true])]; + tensor var_893 = slice_by_index(begin = var_893_begin_0, end = var_893_end_0, end_mask = var_893_end_mask_0, x = x_21)[name = tensor("op_893")]; + tensor var_896_begin_0 = const()[name = tensor("op_896_begin_0"), val = tensor([0, 1, 0])]; + tensor var_896_end_0 = const()[name = tensor("op_896_end_0"), val = tensor([1, 16, 256])]; + tensor var_896_end_mask_0 = const()[name = tensor("op_896_end_mask_0"), val = tensor([true, true, true])]; + tensor var_896 = slice_by_index(begin = var_896_begin_0, end = var_896_end_0, end_mask = var_896_end_mask_0, x = window_27)[name = tensor("op_896")]; tensor window_29_interleave_0 = const()[name = tensor("window_29_interleave_0"), val = tensor(false)]; - tensor window_29 = concat(axis = var_26, interleave = window_29_interleave_0, values = (var_850, var_847))[name = tensor("window_29")]; - tensor var_855_begin_0 = const()[name = tensor("op_855_begin_0"), val = tensor([0, 2, 0])]; - tensor var_855_end_0 = const()[name = tensor("op_855_end_0"), val = tensor([1, 1, 256])]; - tensor var_855_end_mask_0 = const()[name = tensor("op_855_end_mask_0"), val = tensor([true, true, true])]; - tensor var_855 = slice_by_index(begin = var_855_begin_0, end = var_855_end_0, end_mask = var_855_end_mask_0, x = x_21)[name = tensor("op_855")]; - tensor var_858_begin_0 = const()[name = tensor("op_858_begin_0"), val = tensor([0, 1, 0])]; - tensor var_858_end_0 = const()[name = tensor("op_858_end_0"), val = tensor([1, 16, 256])]; - tensor var_858_end_mask_0 = const()[name = tensor("op_858_end_mask_0"), val = tensor([true, true, true])]; - tensor var_858 = slice_by_index(begin = var_858_begin_0, end = var_858_end_0, end_mask = var_858_end_mask_0, x = window_29)[name = tensor("op_858")]; + tensor window_29 = concat(axis = var_72, interleave = window_29_interleave_0, values = (var_896, var_893))[name = tensor("window_29")]; + tensor var_901_begin_0 = const()[name = tensor("op_901_begin_0"), val = tensor([0, 2, 0])]; + tensor var_901_end_0 = const()[name = tensor("op_901_end_0"), val = tensor([1, 1, 256])]; + tensor var_901_end_mask_0 = const()[name = tensor("op_901_end_mask_0"), val = tensor([true, true, true])]; + tensor var_901 = slice_by_index(begin = var_901_begin_0, end = var_901_end_0, end_mask = var_901_end_mask_0, x = x_21)[name = tensor("op_901")]; + tensor var_904_begin_0 = const()[name = tensor("op_904_begin_0"), val = tensor([0, 1, 0])]; + tensor var_904_end_0 = const()[name = tensor("op_904_end_0"), val = tensor([1, 16, 256])]; + tensor var_904_end_mask_0 = const()[name = tensor("op_904_end_mask_0"), val = tensor([true, true, true])]; + tensor var_904 = slice_by_index(begin = var_904_begin_0, end = var_904_end_0, end_mask = var_904_end_mask_0, x = window_29)[name = tensor("op_904")]; tensor window_interleave_0 = const()[name = tensor("window_interleave_0"), val = tensor(false)]; - tensor window = concat(axis = var_26, interleave = window_interleave_0, values = (var_858, var_855))[name = tensor("window")]; - tensor input_141_interleave_0 = const()[name = tensor("input_141_interleave_0"), val = tensor(false)]; - tensor input_141 = concat(axis = var_23, interleave = input_141_interleave_0, values = (window_27, window_29, window))[name = tensor("input_141")]; + tensor window = concat(axis = var_72, interleave = window_interleave_0, values = (var_904, var_901))[name = tensor("window")]; + tensor input_143_interleave_0 = const()[name = tensor("input_143_interleave_0"), val = tensor(false)]; + tensor input_143 = concat(axis = var_58, interleave = input_143_interleave_0, values = (window_27, window_29, window))[name = tensor("input_143")]; tensor x_23_axes_0 = const()[name = tensor("x_23_axes_0"), val = tensor([-1])]; - tensor x_23 = layer_norm(axes = x_23_axes_0, beta = encoder_conv_module_3_sequential_0_bias, epsilon = var_28, gamma = encoder_conv_module_3_sequential_0_weight, x = input_141)[name = tensor("x_23")]; - tensor input_143_perm_0 = const()[name = tensor("input_143_perm_0"), val = tensor([0, 2, 1])]; + tensor x_23 = layer_norm(axes = x_23_axes_0, beta = inner_encoder_conv_module_3_sequential_0_bias, epsilon = var_55, gamma = inner_encoder_conv_module_3_sequential_0_weight, x = input_143)[name = tensor("x_23")]; + tensor input_145_perm_0 = const()[name = tensor("input_145_perm_0"), val = tensor([0, 2, 1])]; tensor inputs_33_pad_type_0 = const()[name = tensor("inputs_33_pad_type_0"), val = tensor("valid")]; tensor inputs_33_strides_0 = const()[name = tensor("inputs_33_strides_0"), val = tensor([1])]; tensor inputs_33_pad_0 = const()[name = tensor("inputs_33_pad_0"), val = tensor([0, 0])]; tensor inputs_33_dilations_0 = const()[name = tensor("inputs_33_dilations_0"), val = tensor([1])]; tensor inputs_33_groups_0 = const()[name = tensor("inputs_33_groups_0"), val = tensor(1)]; - tensor input_143 = transpose(perm = input_143_perm_0, x = x_23)[name = tensor("transpose_33")]; - tensor inputs_33 = conv(bias = encoder_conv_module_3_sequential_2_conv_bias, dilations = inputs_33_dilations_0, groups = inputs_33_groups_0, pad = inputs_33_pad_0, pad_type = inputs_33_pad_type_0, strides = inputs_33_strides_0, weight = encoder_conv_module_3_sequential_2_conv_weight, x = input_143)[name = tensor("inputs_33")]; - tensor var_883_split_sizes_0 = const()[name = tensor("op_883_split_sizes_0"), val = tensor([256, 256])]; - tensor var_883_axis_0 = const()[name = tensor("op_883_axis_0"), val = tensor(1)]; - tensor var_883_0, tensor var_883_1 = split(axis = var_883_axis_0, split_sizes = var_883_split_sizes_0, x = inputs_33)[name = tensor("op_883")]; - tensor var_885 = sigmoid(x = var_883_1)[name = tensor("op_885")]; - tensor inputs_35 = mul(x = var_883_0, y = var_885)[name = tensor("inputs_35")]; + tensor input_145 = transpose(perm = input_145_perm_0, x = x_23)[name = tensor("transpose_33")]; + tensor inputs_33 = conv(bias = inner_encoder_conv_module_3_sequential_2_conv_bias, dilations = inputs_33_dilations_0, groups = inputs_33_groups_0, pad = inputs_33_pad_0, pad_type = inputs_33_pad_type_0, strides = inputs_33_strides_0, weight = inner_encoder_conv_module_3_sequential_2_conv_weight, x = input_145)[name = tensor("inputs_33")]; + tensor var_929_split_sizes_0 = const()[name = tensor("op_929_split_sizes_0"), val = tensor([256, 256])]; + tensor var_929_axis_0 = const()[name = tensor("op_929_axis_0"), val = tensor(1)]; + tensor var_929_0, tensor var_929_1 = split(axis = var_929_axis_0, split_sizes = var_929_split_sizes_0, x = inputs_33)[name = tensor("op_929")]; + tensor var_931 = sigmoid(x = var_929_1)[name = tensor("op_931")]; + tensor inputs_35 = mul(x = var_929_0, y = var_931)[name = tensor("inputs_35")]; tensor outputs_aug_pad_type_0 = const()[name = tensor("outputs_aug_pad_type_0"), val = tensor("custom")]; tensor outputs_aug_pad_0 = const()[name = tensor("outputs_aug_pad_0"), val = tensor([15, 15])]; tensor outputs_aug_groups_0 = const()[name = tensor("outputs_aug_groups_0"), val = tensor(256)]; tensor outputs_aug_strides_0 = const()[name = tensor("outputs_aug_strides_0"), val = tensor([1])]; tensor outputs_aug_dilations_0 = const()[name = tensor("outputs_aug_dilations_0"), val = tensor([1])]; - tensor outputs_aug = conv(dilations = outputs_aug_dilations_0, groups = outputs_aug_groups_0, pad = outputs_aug_pad_0, pad_type = outputs_aug_pad_type_0, strides = outputs_aug_strides_0, weight = encoder_conv_module_3_sequential_4_conv_weight, x = inputs_35)[name = tensor("outputs_aug")]; - tensor input_145_begin_0 = const()[name = tensor("input_145_begin_0"), val = tensor([0, 0, 0])]; - tensor input_145_end_0 = const()[name = tensor("input_145_end_0"), val = tensor([3, 256, 16])]; - tensor input_145_end_mask_0 = const()[name = tensor("input_145_end_mask_0"), val = tensor([true, true, false])]; - tensor input_145 = slice_by_index(begin = input_145_begin_0, end = input_145_end_0, end_mask = input_145_end_mask_0, x = outputs_aug)[name = tensor("input_145")]; - tensor inputs_37 = batch_norm(beta = encoder_conv_module_3_sequential_5_bias, epsilon = var_28, gamma = encoder_conv_module_3_sequential_5_weight, mean = encoder_conv_module_3_sequential_5_running_mean, variance = encoder_conv_module_3_sequential_5_running_var, x = input_145)[name = tensor("inputs_37")]; - tensor input_147 = silu(x = inputs_37)[name = tensor("input_147")]; - tensor input_149_pad_type_0 = const()[name = tensor("input_149_pad_type_0"), val = tensor("valid")]; - tensor input_149_strides_0 = const()[name = tensor("input_149_strides_0"), val = tensor([1])]; - tensor input_149_pad_0 = const()[name = tensor("input_149_pad_0"), val = tensor([0, 0])]; - tensor input_149_dilations_0 = const()[name = tensor("input_149_dilations_0"), val = tensor([1])]; - tensor input_149_groups_0 = const()[name = tensor("input_149_groups_0"), val = tensor(1)]; - tensor input_149 = conv(bias = encoder_conv_module_3_sequential_7_conv_bias, dilations = input_149_dilations_0, groups = input_149_groups_0, pad = input_149_pad_0, pad_type = input_149_pad_type_0, strides = input_149_strides_0, weight = encoder_conv_module_3_sequential_7_conv_weight, x = input_147)[name = tensor("input_149")]; + tensor outputs_aug = conv(dilations = outputs_aug_dilations_0, groups = outputs_aug_groups_0, pad = outputs_aug_pad_0, pad_type = outputs_aug_pad_type_0, strides = outputs_aug_strides_0, weight = inner_encoder_conv_module_3_sequential_4_conv_weight, x = inputs_35)[name = tensor("outputs_aug")]; + tensor input_147_begin_0 = const()[name = tensor("input_147_begin_0"), val = tensor([0, 0, 0])]; + tensor input_147_end_0 = const()[name = tensor("input_147_end_0"), val = tensor([3, 256, 16])]; + tensor input_147_end_mask_0 = const()[name = tensor("input_147_end_mask_0"), val = tensor([true, true, false])]; + tensor input_147 = slice_by_index(begin = input_147_begin_0, end = input_147_end_0, end_mask = input_147_end_mask_0, x = outputs_aug)[name = tensor("input_147")]; + tensor inputs_37 = batch_norm(beta = inner_encoder_conv_module_3_sequential_5_bias, epsilon = var_55, gamma = inner_encoder_conv_module_3_sequential_5_weight, mean = inner_encoder_conv_module_3_sequential_5_running_mean, variance = inner_encoder_conv_module_3_sequential_5_running_var, x = input_147)[name = tensor("inputs_37")]; + tensor input_149 = silu(x = inputs_37)[name = tensor("input_149")]; + tensor input_151_pad_type_0 = const()[name = tensor("input_151_pad_type_0"), val = tensor("valid")]; + tensor input_151_strides_0 = const()[name = tensor("input_151_strides_0"), val = tensor([1])]; + tensor input_151_pad_0 = const()[name = tensor("input_151_pad_0"), val = tensor([0, 0])]; + tensor input_151_dilations_0 = const()[name = tensor("input_151_dilations_0"), val = tensor([1])]; + tensor input_151_groups_0 = const()[name = tensor("input_151_groups_0"), val = tensor(1)]; + tensor input_151 = conv(bias = inner_encoder_conv_module_3_sequential_7_conv_bias, dilations = input_151_dilations_0, groups = input_151_groups_0, pad = input_151_pad_0, pad_type = input_151_pad_type_0, strides = input_151_strides_0, weight = inner_encoder_conv_module_3_sequential_7_conv_weight, x = input_149)[name = tensor("input_151")]; tensor conv_out_7_perm_0 = const()[name = tensor("conv_out_7_perm_0"), val = tensor([0, 2, 1])]; - tensor var_916_begin_0 = const()[name = tensor("op_916_begin_0"), val = tensor([0, -1, 0])]; - tensor var_916_end_0 = const()[name = tensor("op_916_end_0"), val = tensor([3, 16, 256])]; - tensor var_916_end_mask_0 = const()[name = tensor("op_916_end_mask_0"), val = tensor([true, true, true])]; - tensor conv_out_7 = transpose(perm = conv_out_7_perm_0, x = input_149)[name = tensor("transpose_32")]; - tensor var_916 = slice_by_index(begin = var_916_begin_0, end = var_916_end_0, end_mask = var_916_end_mask_0, x = conv_out_7)[name = tensor("op_916")]; - tensor var_918_perm_0 = const()[name = tensor("op_918_perm_0"), val = tensor([1, 0, 2])]; - tensor var_918 = transpose(perm = var_918_perm_0, x = var_916)[name = tensor("transpose_31")]; - tensor input_151 = add(x = x_21, y = var_918)[name = tensor("input_151")]; - tensor input_153_axes_0 = const()[name = tensor("input_153_axes_0"), val = tensor([-1])]; - tensor input_153 = layer_norm(axes = input_153_axes_0, beta = encoder_ffn2_3_module_sequential_0_bias, epsilon = var_28, gamma = encoder_ffn2_3_module_sequential_0_weight, x = input_151)[name = tensor("input_153")]; - tensor inputs = linear(bias = encoder_ffn2_3_module_sequential_1_linear_bias, weight = encoder_ffn2_3_module_sequential_1_linear_weight, x = input_153)[name = tensor("linear_35")]; - tensor input_155 = silu(x = inputs)[name = tensor("input_155")]; - tensor input_159 = linear(bias = encoder_ffn2_3_module_sequential_4_linear_bias, weight = encoder_ffn2_3_module_sequential_4_linear_weight, x = input_155)[name = tensor("linear_36")]; - tensor var_941 = const()[name = tensor("op_941"), val = tensor(0x1p-1)]; - tensor var_942 = mul(x = input_159, y = var_941)[name = tensor("op_942")]; - tensor input_161 = add(x = var_942, y = input_151)[name = tensor("input_161")]; + tensor var_962_begin_0 = const()[name = tensor("op_962_begin_0"), val = tensor([0, -1, 0])]; + tensor var_962_end_0 = const()[name = tensor("op_962_end_0"), val = tensor([3, 16, 256])]; + tensor var_962_end_mask_0 = const()[name = tensor("op_962_end_mask_0"), val = tensor([true, true, true])]; + tensor conv_out_7 = transpose(perm = conv_out_7_perm_0, x = input_151)[name = tensor("transpose_32")]; + tensor var_962 = slice_by_index(begin = var_962_begin_0, end = var_962_end_0, end_mask = var_962_end_mask_0, x = conv_out_7)[name = tensor("op_962")]; + tensor var_964_perm_0 = const()[name = tensor("op_964_perm_0"), val = tensor([1, 0, 2])]; + tensor var_964 = transpose(perm = var_964_perm_0, x = var_962)[name = tensor("transpose_31")]; + tensor input_153 = add(x = x_21, y = var_964)[name = tensor("input_153")]; + tensor input_155_axes_0 = const()[name = tensor("input_155_axes_0"), val = tensor([-1])]; + tensor input_155 = layer_norm(axes = input_155_axes_0, beta = inner_encoder_ffn2_3_module_sequential_0_bias, epsilon = var_55, gamma = inner_encoder_ffn2_3_module_sequential_0_weight, x = input_153)[name = tensor("input_155")]; + tensor inputs = linear(bias = inner_encoder_ffn2_3_module_sequential_1_linear_bias, weight = inner_encoder_ffn2_3_module_sequential_1_linear_weight, x = input_155)[name = tensor("linear_35")]; + tensor input_157 = silu(x = inputs)[name = tensor("input_157")]; + tensor input_161 = linear(bias = inner_encoder_ffn2_3_module_sequential_4_linear_bias, weight = inner_encoder_ffn2_3_module_sequential_4_linear_weight, x = input_157)[name = tensor("linear_36")]; + tensor var_987 = const()[name = tensor("op_987"), val = tensor(0x1p-1)]; + tensor var_988 = mul(x = input_161, y = var_987)[name = tensor("op_988")]; + tensor input_163 = add(x = var_988, y = input_153)[name = tensor("input_163")]; tensor x_25_axes_0 = const()[name = tensor("x_25_axes_0"), val = tensor([-1])]; - tensor x_25 = layer_norm(axes = x_25_axes_0, beta = encoder_layer_norm_3_bias, epsilon = var_28, gamma = encoder_layer_norm_3_weight, x = input_161)[name = tensor("x_25")]; + tensor x_25 = layer_norm(axes = x_25_axes_0, beta = inner_encoder_layer_norm_3_bias, epsilon = var_55, gamma = inner_encoder_layer_norm_3_weight, x = input_163)[name = tensor("x_25")]; tensor x_ct_perm_0 = const()[name = tensor("x_ct_perm_0"), val = tensor([0, 2, 1])]; tensor cat_interleave_0 = const()[name = tensor("cat_interleave_0"), val = tensor(false)]; tensor x_ct = transpose(perm = x_ct_perm_0, x = x_25)[name = tensor("transpose_30")]; - tensor cat = concat(axis = var_20, interleave = cat_interleave_0, values = (cnn_window, x_ct))[name = tensor("cat")]; + tensor cat = concat(axis = var_60, interleave = cat_interleave_0, values = (cnn_window, x_ct))[name = tensor("cat")]; tensor conv_out_pad_type_0 = const()[name = tensor("conv_out_pad_type_0"), val = tensor("valid")]; tensor conv_out_strides_0 = const()[name = tensor("conv_out_strides_0"), val = tensor([1])]; tensor conv_out_pad_0 = const()[name = tensor("conv_out_pad_0"), val = tensor([0, 0])]; tensor conv_out_dilations_0 = const()[name = tensor("conv_out_dilations_0"), val = tensor([1])]; tensor conv_out_groups_0 = const()[name = tensor("conv_out_groups_0"), val = tensor(1)]; - tensor conv_out = conv(bias = encoder_cnn_bias, dilations = conv_out_dilations_0, groups = conv_out_groups_0, pad = conv_out_pad_0, pad_type = conv_out_pad_type_0, strides = conv_out_strides_0, weight = encoder_cnn_weight, x = cat)[name = tensor("conv_out")]; - tensor var_960_begin_0 = const()[name = tensor("op_960_begin_0"), val = tensor([0, 0, 3])]; - tensor var_960_end_0 = const()[name = tensor("op_960_end_0"), val = tensor([1, 256, 21])]; - tensor var_960_end_mask_0 = const()[name = tensor("op_960_end_mask_0"), val = tensor([true, true, true])]; - tensor cnn_window_new = slice_by_index(begin = var_960_begin_0, end = var_960_end_0, end_mask = var_960_end_mask_0, x = cat)[name = tensor("op_960")]; - tensor input_163_perm_0 = const()[name = tensor("input_163_perm_0"), val = tensor([0, 2, 1])]; - tensor var_962 = const()[name = tensor("op_962"), val = tensor([-1])]; - tensor input_163 = transpose(perm = input_163_perm_0, x = conv_out)[name = tensor("transpose_29")]; - tensor var_963 = reduce_l2_norm(axes = var_962, keep_dims = var_29, x = input_163)[name = tensor("op_963")]; + tensor conv_out = conv(bias = inner_encoder_cnn_bias, dilations = conv_out_dilations_0, groups = conv_out_groups_0, pad = conv_out_pad_0, pad_type = conv_out_pad_type_0, strides = conv_out_strides_0, weight = inner_encoder_cnn_weight, x = cat)[name = tensor("conv_out")]; + tensor var_1006_begin_0 = const()[name = tensor("op_1006_begin_0"), val = tensor([0, 0, 3])]; + tensor var_1006_end_0 = const()[name = tensor("op_1006_end_0"), val = tensor([1, 256, 21])]; + tensor var_1006_end_mask_0 = const()[name = tensor("op_1006_end_mask_0"), val = tensor([true, true, true])]; + tensor cnn_window_new = slice_by_index(begin = var_1006_begin_0, end = var_1006_end_0, end_mask = var_1006_end_mask_0, x = cat)[name = tensor("op_1006")]; + tensor input_165_perm_0 = const()[name = tensor("input_165_perm_0"), val = tensor([0, 2, 1])]; + tensor var_1008 = const()[name = tensor("op_1008"), val = tensor([-1])]; + tensor input_165 = transpose(perm = input_165_perm_0, x = conv_out)[name = tensor("transpose_29")]; + tensor var_1009 = reduce_l2_norm(axes = var_1008, keep_dims = var_54, x = input_165)[name = tensor("op_1009")]; tensor const_12 = const()[name = tensor("const_12"), val = tensor(0x1.fffffep+127)]; - tensor clip_0 = clip(alpha = var_11, beta = const_12, x = var_963)[name = tensor("clip_0")]; - tensor emb = real_div(x = input_163, y = clip_0)[name = tensor("emb")]; - tensor var_967_axis_0 = const()[name = tensor("op_967_axis_0"), val = tensor(0)]; - tensor enc_kv_new = stack(axis = var_967_axis_0, values = (var_206, var_412, var_618, nkv_1))[name = tensor("op_967")]; - tensor var_969_axis_0 = const()[name = tensor("op_969_axis_0"), val = tensor(0)]; - tensor enc_scale_new = stack(axis = var_969_axis_0, values = (new_scale_1, new_scale_3, new_scale_5, new_scale_7))[name = tensor("op_969")]; - tensor var_971_axis_0 = const()[name = tensor("op_971_axis_0"), val = tensor(0)]; - tensor enc_conv_cache_new = stack(axis = var_971_axis_0, values = (window_7, window_15, window_23, window))[name = tensor("op_971")]; - tensor var_980 = const()[name = tensor("op_980"), val = tensor(0x1.5798eep-27)]; - tensor var_985 = const()[name = tensor("op_985"), val = tensor(0x1.4f8b58p-17)]; - tensor var_987 = const()[name = tensor("op_987"), val = tensor(0x1.0c6f7ap-20)]; - tensor var_988 = const()[name = tensor("op_988"), val = tensor(true)]; - tensor var_990 = const()[name = tensor("op_990"), val = tensor(0x1p+0)]; - tensor var_994 = const()[name = tensor("op_994"), val = tensor(-1)]; - tensor var_1000 = const()[name = tensor("op_1000"), val = tensor(0)]; + tensor clip_0 = clip(alpha = var_68, beta = const_12, x = var_1009)[name = tensor("clip_0")]; + tensor emb = real_div(x = input_165, y = clip_0)[name = tensor("emb")]; + tensor var_1013_axis_0 = const()[name = tensor("op_1013_axis_0"), val = tensor(0)]; + tensor enc_kv_new = stack(axis = var_1013_axis_0, values = (var_252, var_458, var_664, nkv_1))[name = tensor("op_1013")]; + tensor var_1015_axis_0 = const()[name = tensor("op_1015_axis_0"), val = tensor(0)]; + tensor enc_scale_new = stack(axis = var_1015_axis_0, values = (new_scale_1, new_scale_3, new_scale_5, new_scale_7))[name = tensor("op_1015")]; + tensor var_1017_axis_0 = const()[name = tensor("op_1017_axis_0"), val = tensor(0)]; + tensor enc_conv_cache_new = stack(axis = var_1017_axis_0, values = (window_7, window_15, window_23, window))[name = tensor("op_1017")]; tensor pos = const()[name = tensor("pos"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(44395328)))]; - tensor var_1062_axes_0 = const()[name = tensor("op_1062_axes_0"), val = tensor([2])]; - tensor var_1062 = expand_dims(axes = var_1062_axes_0, x = emb)[name = tensor("op_1062")]; + tensor var_1085_axes_0 = const()[name = tensor("op_1085_axes_0"), val = tensor([2])]; + tensor var_1085 = expand_dims(axes = var_1085_axes_0, x = emb)[name = tensor("op_1085")]; tensor emb_exp_reps_0 = const()[name = tensor("emb_exp_reps_0"), val = tensor([1, 1, 6, 1])]; - tensor emb_exp = tile(reps = emb_exp_reps_0, x = var_1062)[name = tensor("emb_exp")]; - tensor input_165_interleave_0 = const()[name = tensor("input_165_interleave_0"), val = tensor(false)]; - tensor input_165 = concat(axis = var_994, interleave = input_165_interleave_0, values = (emb_exp, pos))[name = tensor("input_165")]; - tensor x_27 = linear(bias = decoder_convert_bias, weight = decoder_convert_weight, x = input_165)[name = tensor("linear_37")]; - tensor var_1070_perm_0 = const()[name = tensor("op_1070_perm_0"), val = tensor([0, 2, 1, 3])]; - tensor var_1074 = const()[name = tensor("op_1074"), val = tensor([6, 3, 256])]; - tensor var_1070 = transpose(perm = var_1070_perm_0, x = x_27)[name = tensor("transpose_28")]; - tensor x_29 = reshape(shape = var_1074, x = var_1070)[name = tensor("x_29")]; + tensor emb_exp = tile(reps = emb_exp_reps_0, x = var_1085)[name = tensor("emb_exp")]; + tensor input_167_interleave_0 = const()[name = tensor("input_167_interleave_0"), val = tensor(false)]; + tensor input_167 = concat(axis = var_61, interleave = input_167_interleave_0, values = (emb_exp, pos))[name = tensor("input_167")]; + tensor x_27 = linear(bias = inner_decoder_convert_bias, weight = inner_decoder_convert_weight, x = input_167)[name = tensor("linear_37")]; + tensor var_1093_perm_0 = const()[name = tensor("op_1093_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_1097 = const()[name = tensor("op_1097"), val = tensor([6, 3, 256])]; + tensor var_1093 = transpose(perm = var_1093_perm_0, x = x_27)[name = tensor("transpose_28")]; + tensor x_29 = reshape(shape = var_1097, x = var_1093)[name = tensor("x_29")]; tensor prev_kv_9_begin_0 = const()[name = tensor("prev_kv_9_begin_0"), val = tensor([0, 0, 0, 0, 0])]; tensor prev_kv_9_end_0 = const()[name = tensor("prev_kv_9_end_0"), val = tensor([1, 6, 4, 64, 64])]; tensor prev_kv_9_end_mask_0 = const()[name = tensor("prev_kv_9_end_mask_0"), val = tensor([false, true, true, true, true])]; @@ -943,132 +954,132 @@ program(1.0) tensor prev_scale_9_end_mask_0 = const()[name = tensor("prev_scale_9_end_mask_0"), val = tensor([false, true])]; tensor prev_scale_9_squeeze_mask_0 = const()[name = tensor("prev_scale_9_squeeze_mask_0"), val = tensor([true, false])]; tensor prev_scale_9 = slice_by_index(begin = prev_scale_9_begin_0, end = prev_scale_9_end_0, end_mask = prev_scale_9_end_mask_0, squeeze_mask = prev_scale_9_squeeze_mask_0, x = dec_scale)[name = tensor("prev_scale_9")]; - tensor var_1082 = linear(bias = decoder_q_proj_0_bias, weight = decoder_q_proj_0_weight, x = x_29)[name = tensor("linear_38")]; - tensor var_1083 = const()[name = tensor("op_1083"), val = tensor([6, 3, 4, 64])]; - tensor var_1084 = reshape(shape = var_1083, x = var_1082)[name = tensor("op_1084")]; + tensor var_1105 = linear(bias = inner_decoder_q_proj_0_bias, weight = inner_decoder_q_proj_0_weight, x = x_29)[name = tensor("linear_38")]; + tensor var_1106 = const()[name = tensor("op_1106"), val = tensor([6, 3, 4, 64])]; + tensor var_1107 = reshape(shape = var_1106, x = var_1105)[name = tensor("op_1107")]; tensor q_9_perm_0 = const()[name = tensor("q_9_perm_0"), val = tensor([0, 2, 1, 3])]; - tensor var_1088 = linear(bias = decoder_k_proj_0_bias, weight = decoder_k_proj_0_weight, x = x_29)[name = tensor("linear_39")]; - tensor var_1089 = const()[name = tensor("op_1089"), val = tensor(0x1p-3)]; - tensor var_1090 = mul(x = var_1088, y = var_1089)[name = tensor("op_1090")]; - tensor var_1091 = const()[name = tensor("op_1091"), val = tensor([6, 3, 4, 64])]; - tensor var_1092 = reshape(shape = var_1091, x = var_1090)[name = tensor("op_1092")]; + tensor var_1111 = linear(bias = inner_decoder_k_proj_0_bias, weight = inner_decoder_k_proj_0_weight, x = x_29)[name = tensor("linear_39")]; + tensor var_1112 = const()[name = tensor("op_1112"), val = tensor(0x1p-3)]; + tensor var_1113 = mul(x = var_1111, y = var_1112)[name = tensor("op_1113")]; + tensor var_1114 = const()[name = tensor("op_1114"), val = tensor([6, 3, 4, 64])]; + tensor var_1115 = reshape(shape = var_1114, x = var_1113)[name = tensor("op_1115")]; tensor k_9_perm_0 = const()[name = tensor("k_9_perm_0"), val = tensor([0, 2, 1, 3])]; - tensor var_1096 = linear(bias = decoder_v_proj_0_bias, weight = decoder_v_proj_0_weight, x = x_29)[name = tensor("linear_40")]; - tensor var_1097 = const()[name = tensor("op_1097"), val = tensor([6, 3, 4, 64])]; - tensor var_1098 = reshape(shape = var_1097, x = var_1096)[name = tensor("op_1098")]; + tensor var_1119 = linear(bias = inner_decoder_v_proj_0_bias, weight = inner_decoder_v_proj_0_weight, x = x_29)[name = tensor("linear_40")]; + tensor var_1120 = const()[name = tensor("op_1120"), val = tensor([6, 3, 4, 64])]; + tensor var_1121 = reshape(shape = var_1120, x = var_1119)[name = tensor("op_1121")]; tensor v_9_perm_0 = const()[name = tensor("v_9_perm_0"), val = tensor([0, 2, 1, 3])]; - tensor input_169 = linear(bias = decoder_g_proj_0_bias, weight = decoder_g_proj_0_weight, x = x_29)[name = tensor("linear_41")]; + tensor input_171 = linear(bias = inner_decoder_g_proj_0_bias, weight = inner_decoder_g_proj_0_weight, x = x_29)[name = tensor("linear_41")]; tensor cumsum_mask_1_exclusive_0 = const()[name = tensor("cumsum_mask_1_exclusive_0"), val = tensor(false)]; tensor cumsum_mask_1_reverse_0 = const()[name = tensor("cumsum_mask_1_reverse_0"), val = tensor(false)]; - tensor cumsum_mask_1 = cumsum(axis = var_1000, exclusive = cumsum_mask_1_exclusive_0, reverse = cumsum_mask_1_reverse_0, x = valid_mask)[name = tensor("cumsum_mask_1")]; + tensor cumsum_mask_1 = cumsum(axis = var_58, exclusive = cumsum_mask_1_exclusive_0, reverse = cumsum_mask_1_reverse_0, x = valid_mask)[name = tensor("cumsum_mask_1")]; tensor sqrt_s0_9 = sqrt(x = prev_scale_9)[name = tensor("sqrt_s0_9")]; tensor s_t_9 = add(x = prev_scale_9, y = cumsum_mask_1)[name = tensor("s_t_9")]; tensor const_20 = const()[name = tensor("const_20"), val = tensor(0x1.fffffep+127)]; - tensor clip_1 = clip(alpha = var_990, beta = const_20, x = s_t_9)[name = tensor("clip_1")]; + tensor clip_1 = clip(alpha = var_48, beta = const_20, x = s_t_9)[name = tensor("clip_1")]; tensor sqrt_s_t_9 = sqrt(x = clip_1)[name = tensor("sqrt_s_t_9")]; tensor qk_9_transpose_x_1 = const()[name = tensor("qk_9_transpose_x_1"), val = tensor(false)]; tensor qk_9_transpose_y_1 = const()[name = tensor("qk_9_transpose_y_1"), val = tensor(true)]; - tensor k_9 = transpose(perm = k_9_perm_0, x = var_1092)[name = tensor("transpose_26")]; - tensor q_9 = transpose(perm = q_9_perm_0, x = var_1084)[name = tensor("transpose_27")]; + tensor k_9 = transpose(perm = k_9_perm_0, x = var_1115)[name = tensor("transpose_26")]; + tensor q_9 = transpose(perm = q_9_perm_0, x = var_1107)[name = tensor("transpose_27")]; tensor qk_9 = matmul(transpose_x = qk_9_transpose_x_1, transpose_y = qk_9_transpose_y_1, x = q_9, y = k_9)[name = tensor("qk_9")]; - tensor var_1110 = const()[name = tensor("op_1110"), val = tensor([1, 3])]; - tensor var_1111 = reshape(shape = var_1110, x = valid_mask)[name = tensor("op_1111")]; - tensor causal_with_valid_1 = mul(x = decoder__causal_mask, y = var_1111)[name = tensor("causal_with_valid_1")]; - tensor var_1113 = const()[name = tensor("op_1113"), val = tensor([3, 1])]; - tensor var_1114 = reshape(shape = var_1113, x = sqrt_s_t_9)[name = tensor("op_1114")]; - tensor M_9 = real_div(x = causal_with_valid_1, y = var_1114)[name = tensor("M_9")]; - tensor var_1116 = mul(x = qk_9, y = M_9)[name = tensor("op_1116")]; + tensor var_1133 = const()[name = tensor("op_1133"), val = tensor([1, 3])]; + tensor var_1134 = reshape(shape = var_1133, x = valid_mask)[name = tensor("op_1134")]; + tensor causal_with_valid_1 = mul(x = inner_decoder__causal_mask, y = var_1134)[name = tensor("causal_with_valid_1")]; + tensor var_1136 = const()[name = tensor("op_1136"), val = tensor([3, 1])]; + tensor var_1137 = reshape(shape = var_1136, x = sqrt_s_t_9)[name = tensor("op_1137")]; + tensor M_9 = real_div(x = causal_with_valid_1, y = var_1137)[name = tensor("M_9")]; + tensor var_1139 = mul(x = qk_9, y = M_9)[name = tensor("op_1139")]; tensor inner_9_transpose_x_0 = const()[name = tensor("inner_9_transpose_x_0"), val = tensor(false)]; tensor inner_9_transpose_y_0 = const()[name = tensor("inner_9_transpose_y_0"), val = tensor(false)]; - tensor v_9 = transpose(perm = v_9_perm_0, x = var_1098)[name = tensor("transpose_25")]; - tensor inner_9 = matmul(transpose_x = inner_9_transpose_x_0, transpose_y = inner_9_transpose_y_0, x = var_1116, y = v_9)[name = tensor("inner_9")]; - tensor var_1118_transpose_x_0 = const()[name = tensor("op_1118_transpose_x_0"), val = tensor(false)]; - tensor var_1118_transpose_y_0 = const()[name = tensor("op_1118_transpose_y_0"), val = tensor(false)]; - tensor var_1118 = matmul(transpose_x = var_1118_transpose_x_0, transpose_y = var_1118_transpose_y_0, x = q_9, y = prev_kv_9)[name = tensor("op_1118")]; - tensor var_1119 = real_div(x = sqrt_s0_9, y = sqrt_s_t_9)[name = tensor("op_1119")]; - tensor var_1120 = const()[name = tensor("op_1120"), val = tensor([1, 1, 3, 1])]; - tensor var_1121 = reshape(shape = var_1120, x = var_1119)[name = tensor("op_1121")]; - tensor cross_9 = mul(x = var_1118, y = var_1121)[name = tensor("cross_9")]; + tensor v_9 = transpose(perm = v_9_perm_0, x = var_1121)[name = tensor("transpose_25")]; + tensor inner_9 = matmul(transpose_x = inner_9_transpose_x_0, transpose_y = inner_9_transpose_y_0, x = var_1139, y = v_9)[name = tensor("inner_9")]; + tensor var_1141_transpose_x_0 = const()[name = tensor("op_1141_transpose_x_0"), val = tensor(false)]; + tensor var_1141_transpose_y_0 = const()[name = tensor("op_1141_transpose_y_0"), val = tensor(false)]; + tensor var_1141 = matmul(transpose_x = var_1141_transpose_x_0, transpose_y = var_1141_transpose_y_0, x = q_9, y = prev_kv_9)[name = tensor("op_1141")]; + tensor var_1142 = real_div(x = sqrt_s0_9, y = sqrt_s_t_9)[name = tensor("op_1142")]; + tensor var_1143 = const()[name = tensor("op_1143"), val = tensor([1, 1, 3, 1])]; + tensor var_1144 = reshape(shape = var_1143, x = var_1142)[name = tensor("op_1144")]; + tensor cross_9 = mul(x = var_1141, y = var_1144)[name = tensor("cross_9")]; tensor out_25 = add(x = inner_9, y = cross_9)[name = tensor("out_25")]; - tensor var_1124 = const()[name = tensor("op_1124"), val = tensor([1, 1, 3, 1])]; - tensor var_1125 = reshape(shape = var_1124, x = valid_mask)[name = tensor("op_1125")]; - tensor v_masked_1 = mul(x = v_9, y = var_1125)[name = tensor("v_masked_1")]; - tensor var_1127 = mul(x = prev_kv_9, y = sqrt_s0_9)[name = tensor("op_1127")]; - tensor var_1129_transpose_x_1 = const()[name = tensor("op_1129_transpose_x_1"), val = tensor(true)]; - tensor var_1129_transpose_y_1 = const()[name = tensor("op_1129_transpose_y_1"), val = tensor(false)]; - tensor var_1129 = matmul(transpose_x = var_1129_transpose_x_1, transpose_y = var_1129_transpose_y_1, x = k_9, y = v_masked_1)[name = tensor("op_1129")]; - tensor new_kv_unnorm_9 = add(x = var_1127, y = var_1129)[name = tensor("new_kv_unnorm_9")]; - tensor var_1131_keep_dims_0 = const()[name = tensor("op_1131_keep_dims_0"), val = tensor(false)]; - tensor var_1131 = reduce_sum(keep_dims = var_1131_keep_dims_0, x = valid_mask)[name = tensor("op_1131")]; - tensor var_1132 = const()[name = tensor("op_1132"), val = tensor([1])]; - tensor var_1133 = reshape(shape = var_1132, x = var_1131)[name = tensor("op_1133")]; - tensor new_scale_9 = add(x = prev_scale_9, y = var_1133)[name = tensor("new_scale_9")]; + tensor var_1147 = const()[name = tensor("op_1147"), val = tensor([1, 1, 3, 1])]; + tensor var_1148 = reshape(shape = var_1147, x = valid_mask)[name = tensor("op_1148")]; + tensor v_masked_1 = mul(x = v_9, y = var_1148)[name = tensor("v_masked_1")]; + tensor var_1150 = mul(x = prev_kv_9, y = sqrt_s0_9)[name = tensor("op_1150")]; + tensor var_1152_transpose_x_1 = const()[name = tensor("op_1152_transpose_x_1"), val = tensor(true)]; + tensor var_1152_transpose_y_1 = const()[name = tensor("op_1152_transpose_y_1"), val = tensor(false)]; + tensor var_1152 = matmul(transpose_x = var_1152_transpose_x_1, transpose_y = var_1152_transpose_y_1, x = k_9, y = v_masked_1)[name = tensor("op_1152")]; + tensor new_kv_unnorm_9 = add(x = var_1150, y = var_1152)[name = tensor("new_kv_unnorm_9")]; + tensor var_1154_keep_dims_0 = const()[name = tensor("op_1154_keep_dims_0"), val = tensor(false)]; + tensor var_1154 = reduce_sum(keep_dims = var_1154_keep_dims_0, x = valid_mask)[name = tensor("op_1154")]; + tensor var_1155 = const()[name = tensor("op_1155"), val = tensor([1])]; + tensor var_1156 = reshape(shape = var_1155, x = var_1154)[name = tensor("op_1156")]; + tensor new_scale_9 = add(x = prev_scale_9, y = var_1156)[name = tensor("new_scale_9")]; tensor const_21 = const()[name = tensor("const_21"), val = tensor(0x1.fffffep+127)]; - tensor clip_2 = clip(alpha = var_990, beta = const_21, x = new_scale_9)[name = tensor("clip_2")]; + tensor clip_2 = clip(alpha = var_48, beta = const_21, x = new_scale_9)[name = tensor("clip_2")]; tensor sqrt_new_scale_1 = sqrt(x = clip_2)[name = tensor("sqrt_new_scale_1")]; - tensor var_1137 = real_div(x = new_kv_unnorm_9, y = sqrt_new_scale_1)[name = tensor("op_1137")]; - tensor var_1138_perm_0 = const()[name = tensor("op_1138_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_1160 = real_div(x = new_kv_unnorm_9, y = sqrt_new_scale_1)[name = tensor("op_1160")]; + tensor var_1161_perm_0 = const()[name = tensor("op_1161_perm_0"), val = tensor([0, 2, 1, 3])]; tensor out_27_axes_0 = const()[name = tensor("out_27_axes_0"), val = tensor([-1])]; - tensor var_1138 = transpose(perm = var_1138_perm_0, x = out_25)[name = tensor("transpose_24")]; - tensor out_27 = layer_norm(axes = out_27_axes_0, epsilon = var_987, x = var_1138)[name = tensor("out_27")]; - tensor var_1142 = const()[name = tensor("op_1142"), val = tensor([6, 3, 256])]; - tensor out_29 = reshape(shape = var_1142, x = out_27)[name = tensor("out_29")]; - tensor var_1144 = silu(x = input_169)[name = tensor("op_1144")]; - tensor input_171 = mul(x = var_1144, y = out_29)[name = tensor("input_171")]; - tensor ret_out_9 = linear(bias = decoder_out_proj_0_bias, weight = decoder_out_proj_0_weight, x = input_171)[name = tensor("linear_42")]; - tensor input_173 = add(x = x_29, y = ret_out_9)[name = tensor("input_173")]; + tensor var_1161 = transpose(perm = var_1161_perm_0, x = out_25)[name = tensor("transpose_24")]; + tensor out_27 = layer_norm(axes = out_27_axes_0, epsilon = var_63, x = var_1161)[name = tensor("out_27")]; + tensor var_1165 = const()[name = tensor("op_1165"), val = tensor([6, 3, 256])]; + tensor out_29 = reshape(shape = var_1165, x = out_27)[name = tensor("out_29")]; + tensor var_1167 = silu(x = input_171)[name = tensor("op_1167")]; + tensor input_173 = mul(x = var_1167, y = out_29)[name = tensor("input_173")]; + tensor ret_out_9 = linear(bias = inner_decoder_out_proj_0_bias, weight = inner_decoder_out_proj_0_weight, x = input_173)[name = tensor("linear_42")]; + tensor input_175 = add(x = x_29, y = ret_out_9)[name = tensor("input_175")]; tensor xt_1_axes_0 = const()[name = tensor("xt_1_axes_0"), val = tensor([-1])]; - tensor xt_1 = layer_norm(axes = xt_1_axes_0, beta = decoder_norm11_0_bias, epsilon = var_985, gamma = decoder_norm11_0_weight, x = input_173)[name = tensor("xt_1")]; - tensor var_1154 = const()[name = tensor("op_1154"), val = tensor([1, 6, 3, 256])]; - tensor var_1155 = reshape(shape = var_1154, x = xt_1)[name = tensor("op_1155")]; - tensor var_1156_perm_0 = const()[name = tensor("op_1156_perm_0"), val = tensor([0, 2, 1, 3])]; - tensor var_1159 = const()[name = tensor("op_1159"), val = tensor([3, 6, 256])]; - tensor var_1156 = transpose(perm = var_1156_perm_0, x = var_1155)[name = tensor("transpose_23")]; - tensor query_1 = reshape(shape = var_1159, x = var_1156)[name = tensor("query_1")]; + tensor xt_1 = layer_norm(axes = xt_1_axes_0, beta = inner_decoder_norm11_0_bias, epsilon = var_55, gamma = inner_decoder_norm11_0_weight, x = input_175)[name = tensor("xt_1")]; + tensor var_1177 = const()[name = tensor("op_1177"), val = tensor([1, 6, 3, 256])]; + tensor var_1178 = reshape(shape = var_1177, x = xt_1)[name = tensor("op_1178")]; + tensor var_1179_perm_0 = const()[name = tensor("op_1179_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_1182 = const()[name = tensor("op_1182"), val = tensor([3, 6, 256])]; + tensor var_1179 = transpose(perm = var_1179_perm_0, x = var_1178)[name = tensor("transpose_23")]; + tensor query_1 = reshape(shape = var_1182, x = var_1179)[name = tensor("query_1")]; tensor query_3_perm_0 = const()[name = tensor("query_3_perm_0"), val = tensor([1, 0, 2])]; tensor query_3 = transpose(perm = query_3_perm_0, x = query_1)[name = tensor("transpose_22")]; - tensor var_1182 = linear(bias = decoder_self_attn2_0_in_proj_bias, weight = decoder_self_attn2_0_in_proj_weight, x = query_3)[name = tensor("linear_43")]; + tensor var_1205 = linear(bias = inner_decoder_self_attn2_0_in_proj_bias, weight = inner_decoder_self_attn2_0_in_proj_weight, x = query_3)[name = tensor("linear_43")]; tensor concat_1 = const()[name = tensor("concat_1"), val = tensor([6, 3, 3, 256])]; - tensor var_1184 = reshape(shape = concat_1, x = var_1182)[name = tensor("op_1184")]; - tensor var_1185_axes_0 = const()[name = tensor("op_1185_axes_0"), val = tensor([0])]; - tensor var_1185 = expand_dims(axes = var_1185_axes_0, x = var_1184)[name = tensor("op_1185")]; - tensor var_1186_perm_0 = const()[name = tensor("op_1186_perm_0"), val = tensor([-2, 1, 2, 0, 4])]; - tensor var_1187_axes_0 = const()[name = tensor("op_1187_axes_0"), val = tensor([-2])]; - tensor var_1186 = transpose(perm = var_1186_perm_0, x = var_1185)[name = tensor("transpose_21")]; - tensor var_1187 = squeeze(axes = var_1187_axes_0, x = var_1186)[name = tensor("op_1187")]; + tensor var_1207 = reshape(shape = concat_1, x = var_1205)[name = tensor("op_1207")]; + tensor var_1208_axes_0 = const()[name = tensor("op_1208_axes_0"), val = tensor([0])]; + tensor var_1208 = expand_dims(axes = var_1208_axes_0, x = var_1207)[name = tensor("op_1208")]; + tensor var_1209_perm_0 = const()[name = tensor("op_1209_perm_0"), val = tensor([-2, 1, 2, 0, 4])]; + tensor var_1210_axes_0 = const()[name = tensor("op_1210_axes_0"), val = tensor([-2])]; + tensor var_1209 = transpose(perm = var_1209_perm_0, x = var_1208)[name = tensor("transpose_21")]; + tensor var_1210 = squeeze(axes = var_1210_axes_0, x = var_1209)[name = tensor("op_1210")]; tensor q_11_begin_0 = const()[name = tensor("q_11_begin_0"), val = tensor([0, 0, 0, 0])]; tensor q_11_end_0 = const()[name = tensor("q_11_end_0"), val = tensor([1, 6, 3, 256])]; tensor q_11_end_mask_0 = const()[name = tensor("q_11_end_mask_0"), val = tensor([false, true, true, true])]; tensor q_11_squeeze_mask_0 = const()[name = tensor("q_11_squeeze_mask_0"), val = tensor([true, false, false, false])]; - tensor q_11 = slice_by_index(begin = q_11_begin_0, end = q_11_end_0, end_mask = q_11_end_mask_0, squeeze_mask = q_11_squeeze_mask_0, x = var_1187)[name = tensor("q_11")]; + tensor q_11 = slice_by_index(begin = q_11_begin_0, end = q_11_end_0, end_mask = q_11_end_mask_0, squeeze_mask = q_11_squeeze_mask_0, x = var_1210)[name = tensor("q_11")]; tensor k_11_begin_0 = const()[name = tensor("k_11_begin_0"), val = tensor([1, 0, 0, 0])]; tensor k_11_end_0 = const()[name = tensor("k_11_end_0"), val = tensor([2, 6, 3, 256])]; tensor k_11_end_mask_0 = const()[name = tensor("k_11_end_mask_0"), val = tensor([false, true, true, true])]; tensor k_11_squeeze_mask_0 = const()[name = tensor("k_11_squeeze_mask_0"), val = tensor([true, false, false, false])]; - tensor k_11 = slice_by_index(begin = k_11_begin_0, end = k_11_end_0, end_mask = k_11_end_mask_0, squeeze_mask = k_11_squeeze_mask_0, x = var_1187)[name = tensor("k_11")]; + tensor k_11 = slice_by_index(begin = k_11_begin_0, end = k_11_end_0, end_mask = k_11_end_mask_0, squeeze_mask = k_11_squeeze_mask_0, x = var_1210)[name = tensor("k_11")]; tensor v_11_begin_0 = const()[name = tensor("v_11_begin_0"), val = tensor([2, 0, 0, 0])]; tensor v_11_end_0 = const()[name = tensor("v_11_end_0"), val = tensor([3, 6, 3, 256])]; tensor v_11_end_mask_0 = const()[name = tensor("v_11_end_mask_0"), val = tensor([false, true, true, true])]; tensor v_11_squeeze_mask_0 = const()[name = tensor("v_11_squeeze_mask_0"), val = tensor([true, false, false, false])]; - tensor v_11 = slice_by_index(begin = v_11_begin_0, end = v_11_end_0, end_mask = v_11_end_mask_0, squeeze_mask = v_11_squeeze_mask_0, x = var_1187)[name = tensor("v_11")]; - tensor var_1195 = const()[name = tensor("op_1195"), val = tensor([6, 12, 64])]; - tensor var_1196 = reshape(shape = var_1195, x = q_11)[name = tensor("op_1196")]; + tensor v_11 = slice_by_index(begin = v_11_begin_0, end = v_11_end_0, end_mask = v_11_end_mask_0, squeeze_mask = v_11_squeeze_mask_0, x = var_1210)[name = tensor("v_11")]; + tensor var_1218 = const()[name = tensor("op_1218"), val = tensor([6, 12, 64])]; + tensor var_1219 = reshape(shape = var_1218, x = q_11)[name = tensor("op_1219")]; tensor q_13_perm_0 = const()[name = tensor("q_13_perm_0"), val = tensor([1, 0, 2])]; - tensor var_1202 = const()[name = tensor("op_1202"), val = tensor([6, 12, 64])]; - tensor var_1203 = reshape(shape = var_1202, x = k_11)[name = tensor("op_1203")]; + tensor var_1225 = const()[name = tensor("op_1225"), val = tensor([6, 12, 64])]; + tensor var_1226 = reshape(shape = var_1225, x = k_11)[name = tensor("op_1226")]; tensor k_13_perm_0 = const()[name = tensor("k_13_perm_0"), val = tensor([1, 0, 2])]; - tensor var_1209 = const()[name = tensor("op_1209"), val = tensor([6, 12, 64])]; - tensor var_1210 = reshape(shape = var_1209, x = v_11)[name = tensor("op_1210")]; + tensor var_1232 = const()[name = tensor("op_1232"), val = tensor([6, 12, 64])]; + tensor var_1233 = reshape(shape = var_1232, x = v_11)[name = tensor("op_1233")]; tensor v_13_perm_0 = const()[name = tensor("v_13_perm_0"), val = tensor([1, 0, 2])]; - tensor var_1213 = const()[name = tensor("op_1213"), val = tensor([3, 4, 6, 64])]; - tensor q_13 = transpose(perm = q_13_perm_0, x = var_1196)[name = tensor("transpose_20")]; - tensor q_15 = reshape(shape = var_1213, x = q_13)[name = tensor("q_15")]; - tensor var_1215 = const()[name = tensor("op_1215"), val = tensor([3, 4, 6, 64])]; - tensor k_13 = transpose(perm = k_13_perm_0, x = var_1203)[name = tensor("transpose_19")]; - tensor k_15 = reshape(shape = var_1215, x = k_13)[name = tensor("k_15")]; - tensor var_1217 = const()[name = tensor("op_1217"), val = tensor([3, 4, 6, 64])]; - tensor v_13 = transpose(perm = v_13_perm_0, x = var_1210)[name = tensor("transpose_18")]; - tensor v_15 = reshape(shape = var_1217, x = v_13)[name = tensor("v_15")]; + tensor var_1236 = const()[name = tensor("op_1236"), val = tensor([3, 4, 6, 64])]; + tensor q_13 = transpose(perm = q_13_perm_0, x = var_1219)[name = tensor("transpose_20")]; + tensor q_15 = reshape(shape = var_1236, x = q_13)[name = tensor("q_15")]; + tensor var_1238 = const()[name = tensor("op_1238"), val = tensor([3, 4, 6, 64])]; + tensor k_13 = transpose(perm = k_13_perm_0, x = var_1226)[name = tensor("transpose_19")]; + tensor k_15 = reshape(shape = var_1238, x = k_13)[name = tensor("k_15")]; + tensor var_1240 = const()[name = tensor("op_1240"), val = tensor([3, 4, 6, 64])]; + tensor v_13 = transpose(perm = v_13_perm_0, x = var_1233)[name = tensor("transpose_18")]; + tensor v_15 = reshape(shape = var_1240, x = v_13)[name = tensor("v_15")]; tensor mul_1_y_0 = const()[name = tensor("mul_1_y_0"), val = tensor(0x1p-3)]; tensor mul_1 = mul(x = q_15, y = mul_1_y_0)[name = tensor("mul_1")]; tensor matmul_0_transpose_y_0 = const()[name = tensor("matmul_0_transpose_y_0"), val = tensor(true)]; @@ -1079,30 +1090,30 @@ program(1.0) tensor attn_output_1_transpose_x_0 = const()[name = tensor("attn_output_1_transpose_x_0"), val = tensor(false)]; tensor attn_output_1_transpose_y_0 = const()[name = tensor("attn_output_1_transpose_y_0"), val = tensor(false)]; tensor attn_output_1 = matmul(transpose_x = attn_output_1_transpose_x_0, transpose_y = attn_output_1_transpose_y_0, x = softmax_0, y = v_15)[name = tensor("attn_output_1")]; - tensor var_1220 = const()[name = tensor("op_1220"), val = tensor([2, 0, 1, 3])]; - tensor var_1225 = const()[name = tensor("op_1225"), val = tensor([18, 256])]; - tensor var_1221 = transpose(perm = var_1220, x = attn_output_1)[name = tensor("transpose_17")]; - tensor attn_output_3 = reshape(shape = var_1225, x = var_1221)[name = tensor("attn_output_3")]; - tensor attn_output_5 = linear(bias = decoder_self_attn2_0_out_proj_bias, weight = decoder_self_attn2_0_out_proj_weight, x = attn_output_3)[name = tensor("linear_44")]; - tensor var_1229 = const()[name = tensor("op_1229"), val = tensor([6, 3, 256])]; - tensor attn_output_7 = reshape(shape = var_1229, x = attn_output_5)[name = tensor("attn_output_7")]; + tensor var_1243 = const()[name = tensor("op_1243"), val = tensor([2, 0, 1, 3])]; + tensor var_1248 = const()[name = tensor("op_1248"), val = tensor([18, 256])]; + tensor var_1244 = transpose(perm = var_1243, x = attn_output_1)[name = tensor("transpose_17")]; + tensor attn_output_3 = reshape(shape = var_1248, x = var_1244)[name = tensor("attn_output_3")]; + tensor attn_output_5 = linear(bias = inner_decoder_self_attn2_0_out_proj_bias, weight = inner_decoder_self_attn2_0_out_proj_weight, x = attn_output_3)[name = tensor("linear_44")]; + tensor var_1252 = const()[name = tensor("op_1252"), val = tensor([6, 3, 256])]; + tensor attn_output_7 = reshape(shape = var_1252, x = attn_output_5)[name = tensor("attn_output_7")]; tensor sa2_out_1_perm_0 = const()[name = tensor("sa2_out_1_perm_0"), val = tensor([1, 0, 2])]; tensor sa2_out_1 = transpose(perm = sa2_out_1_perm_0, x = attn_output_7)[name = tensor("transpose_16")]; - tensor input_175 = add(x = query_1, y = sa2_out_1)[name = tensor("input_175")]; - tensor input_177_axes_0 = const()[name = tensor("input_177_axes_0"), val = tensor([-1])]; - tensor input_177 = layer_norm(axes = input_177_axes_0, beta = decoder_norm21_0_bias, epsilon = var_985, gamma = decoder_norm21_0_weight, x = input_175)[name = tensor("input_177")]; - tensor input_179 = linear(bias = decoder_linear1_0_bias, weight = decoder_linear1_0_weight, x = input_177)[name = tensor("linear_45")]; - tensor input_181 = relu(x = input_179)[name = tensor("input_181")]; - tensor ffn_out_1 = linear(bias = decoder_linear2_0_bias, weight = decoder_linear2_0_weight, x = input_181)[name = tensor("linear_46")]; - tensor input_183 = add(x = input_177, y = ffn_out_1)[name = tensor("input_183")]; + tensor input_177 = add(x = query_1, y = sa2_out_1)[name = tensor("input_177")]; + tensor input_179_axes_0 = const()[name = tensor("input_179_axes_0"), val = tensor([-1])]; + tensor input_179 = layer_norm(axes = input_179_axes_0, beta = inner_decoder_norm21_0_bias, epsilon = var_55, gamma = inner_decoder_norm21_0_weight, x = input_177)[name = tensor("input_179")]; + tensor input_181 = linear(bias = inner_decoder_linear1_0_bias, weight = inner_decoder_linear1_0_weight, x = input_179)[name = tensor("linear_45")]; + tensor input_183 = relu(x = input_181)[name = tensor("input_183")]; + tensor ffn_out_1 = linear(bias = inner_decoder_linear2_0_bias, weight = inner_decoder_linear2_0_weight, x = input_183)[name = tensor("linear_46")]; + tensor input_185 = add(x = input_179, y = ffn_out_1)[name = tensor("input_185")]; tensor xt_3_axes_0 = const()[name = tensor("xt_3_axes_0"), val = tensor([-1])]; - tensor xt_3 = layer_norm(axes = xt_3_axes_0, beta = decoder_norm22_0_bias, epsilon = var_985, gamma = decoder_norm22_0_weight, x = input_183)[name = tensor("xt_3")]; - tensor var_1249 = const()[name = tensor("op_1249"), val = tensor([1, 3, 6, 256])]; - tensor x_31 = reshape(shape = var_1249, x = xt_3)[name = tensor("x_31")]; - tensor var_1251_perm_0 = const()[name = tensor("op_1251_perm_0"), val = tensor([0, 2, 1, 3])]; - tensor var_1255 = const()[name = tensor("op_1255"), val = tensor([6, 3, 256])]; - tensor var_1251 = transpose(perm = var_1251_perm_0, x = x_31)[name = tensor("transpose_15")]; - tensor x = reshape(shape = var_1255, x = var_1251)[name = tensor("x")]; + tensor xt_3 = layer_norm(axes = xt_3_axes_0, beta = inner_decoder_norm22_0_bias, epsilon = var_55, gamma = inner_decoder_norm22_0_weight, x = input_185)[name = tensor("xt_3")]; + tensor var_1272 = const()[name = tensor("op_1272"), val = tensor([1, 3, 6, 256])]; + tensor x_31 = reshape(shape = var_1272, x = xt_3)[name = tensor("x_31")]; + tensor var_1274_perm_0 = const()[name = tensor("op_1274_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_1278 = const()[name = tensor("op_1278"), val = tensor([6, 3, 256])]; + tensor var_1274 = transpose(perm = var_1274_perm_0, x = x_31)[name = tensor("transpose_15")]; + tensor x = reshape(shape = var_1278, x = var_1274)[name = tensor("x")]; tensor prev_kv_begin_0 = const()[name = tensor("prev_kv_begin_0"), val = tensor([1, 0, 0, 0, 0])]; tensor prev_kv_end_0 = const()[name = tensor("prev_kv_end_0"), val = tensor([2, 6, 4, 64, 64])]; tensor prev_kv_end_mask_0 = const()[name = tensor("prev_kv_end_mask_0"), val = tensor([false, true, true, true, true])]; @@ -1113,120 +1124,120 @@ program(1.0) tensor prev_scale_end_mask_0 = const()[name = tensor("prev_scale_end_mask_0"), val = tensor([false, true])]; tensor prev_scale_squeeze_mask_0 = const()[name = tensor("prev_scale_squeeze_mask_0"), val = tensor([true, false])]; tensor prev_scale = slice_by_index(begin = prev_scale_begin_0, end = prev_scale_end_0, end_mask = prev_scale_end_mask_0, squeeze_mask = prev_scale_squeeze_mask_0, x = dec_scale)[name = tensor("prev_scale")]; - tensor var_1263 = linear(bias = decoder_q_proj_1_bias, weight = decoder_q_proj_1_weight, x = x)[name = tensor("linear_47")]; - tensor var_1264 = const()[name = tensor("op_1264"), val = tensor([6, 3, 4, 64])]; - tensor var_1265 = reshape(shape = var_1264, x = var_1263)[name = tensor("op_1265")]; + tensor var_1286 = linear(bias = inner_decoder_q_proj_1_bias, weight = inner_decoder_q_proj_1_weight, x = x)[name = tensor("linear_47")]; + tensor var_1287 = const()[name = tensor("op_1287"), val = tensor([6, 3, 4, 64])]; + tensor var_1288 = reshape(shape = var_1287, x = var_1286)[name = tensor("op_1288")]; tensor q_17_perm_0 = const()[name = tensor("q_17_perm_0"), val = tensor([0, 2, 1, 3])]; - tensor var_1269 = linear(bias = decoder_k_proj_1_bias, weight = decoder_k_proj_1_weight, x = x)[name = tensor("linear_48")]; - tensor var_1270 = const()[name = tensor("op_1270"), val = tensor(0x1p-3)]; - tensor var_1271 = mul(x = var_1269, y = var_1270)[name = tensor("op_1271")]; - tensor var_1272 = const()[name = tensor("op_1272"), val = tensor([6, 3, 4, 64])]; - tensor var_1273 = reshape(shape = var_1272, x = var_1271)[name = tensor("op_1273")]; + tensor var_1292 = linear(bias = inner_decoder_k_proj_1_bias, weight = inner_decoder_k_proj_1_weight, x = x)[name = tensor("linear_48")]; + tensor var_1293 = const()[name = tensor("op_1293"), val = tensor(0x1p-3)]; + tensor var_1294 = mul(x = var_1292, y = var_1293)[name = tensor("op_1294")]; + tensor var_1295 = const()[name = tensor("op_1295"), val = tensor([6, 3, 4, 64])]; + tensor var_1296 = reshape(shape = var_1295, x = var_1294)[name = tensor("op_1296")]; tensor k_17_perm_0 = const()[name = tensor("k_17_perm_0"), val = tensor([0, 2, 1, 3])]; - tensor var_1277 = linear(bias = decoder_v_proj_1_bias, weight = decoder_v_proj_1_weight, x = x)[name = tensor("linear_49")]; - tensor var_1278 = const()[name = tensor("op_1278"), val = tensor([6, 3, 4, 64])]; - tensor var_1279 = reshape(shape = var_1278, x = var_1277)[name = tensor("op_1279")]; + tensor var_1300 = linear(bias = inner_decoder_v_proj_1_bias, weight = inner_decoder_v_proj_1_weight, x = x)[name = tensor("linear_49")]; + tensor var_1301 = const()[name = tensor("op_1301"), val = tensor([6, 3, 4, 64])]; + tensor var_1302 = reshape(shape = var_1301, x = var_1300)[name = tensor("op_1302")]; tensor v_17_perm_0 = const()[name = tensor("v_17_perm_0"), val = tensor([0, 2, 1, 3])]; - tensor input_187 = linear(bias = decoder_g_proj_1_bias, weight = decoder_g_proj_1_weight, x = x)[name = tensor("linear_50")]; + tensor input_189 = linear(bias = inner_decoder_g_proj_1_bias, weight = inner_decoder_g_proj_1_weight, x = x)[name = tensor("linear_50")]; tensor sqrt_s0 = sqrt(x = prev_scale)[name = tensor("sqrt_s0")]; tensor s_t = add(x = prev_scale, y = cumsum_mask_1)[name = tensor("s_t")]; tensor const_32 = const()[name = tensor("const_32"), val = tensor(0x1.fffffep+127)]; - tensor clip_3 = clip(alpha = var_990, beta = const_32, x = s_t)[name = tensor("clip_3")]; + tensor clip_3 = clip(alpha = var_48, beta = const_32, x = s_t)[name = tensor("clip_3")]; tensor sqrt_s_t = sqrt(x = clip_3)[name = tensor("sqrt_s_t")]; tensor qk_transpose_x_1 = const()[name = tensor("qk_transpose_x_1"), val = tensor(false)]; tensor qk_transpose_y_1 = const()[name = tensor("qk_transpose_y_1"), val = tensor(true)]; - tensor k_17 = transpose(perm = k_17_perm_0, x = var_1273)[name = tensor("transpose_13")]; - tensor q_17 = transpose(perm = q_17_perm_0, x = var_1265)[name = tensor("transpose_14")]; + tensor k_17 = transpose(perm = k_17_perm_0, x = var_1296)[name = tensor("transpose_13")]; + tensor q_17 = transpose(perm = q_17_perm_0, x = var_1288)[name = tensor("transpose_14")]; tensor qk = matmul(transpose_x = qk_transpose_x_1, transpose_y = qk_transpose_y_1, x = q_17, y = k_17)[name = tensor("qk")]; - tensor var_1294 = const()[name = tensor("op_1294"), val = tensor([3, 1])]; - tensor var_1295 = reshape(shape = var_1294, x = sqrt_s_t)[name = tensor("op_1295")]; - tensor M = real_div(x = causal_with_valid_1, y = var_1295)[name = tensor("M")]; - tensor var_1297 = mul(x = qk, y = M)[name = tensor("op_1297")]; - tensor inner_transpose_x_0 = const()[name = tensor("inner_transpose_x_0"), val = tensor(false)]; - tensor inner_transpose_y_0 = const()[name = tensor("inner_transpose_y_0"), val = tensor(false)]; - tensor v_17 = transpose(perm = v_17_perm_0, x = var_1279)[name = tensor("transpose_12")]; - tensor inner = matmul(transpose_x = inner_transpose_x_0, transpose_y = inner_transpose_y_0, x = var_1297, y = v_17)[name = tensor("inner")]; - tensor var_1299_transpose_x_0 = const()[name = tensor("op_1299_transpose_x_0"), val = tensor(false)]; - tensor var_1299_transpose_y_0 = const()[name = tensor("op_1299_transpose_y_0"), val = tensor(false)]; - tensor var_1299 = matmul(transpose_x = var_1299_transpose_x_0, transpose_y = var_1299_transpose_y_0, x = q_17, y = prev_kv)[name = tensor("op_1299")]; - tensor var_1300 = real_div(x = sqrt_s0, y = sqrt_s_t)[name = tensor("op_1300")]; - tensor var_1301 = const()[name = tensor("op_1301"), val = tensor([1, 1, 3, 1])]; - tensor var_1302 = reshape(shape = var_1301, x = var_1300)[name = tensor("op_1302")]; - tensor cross = mul(x = var_1299, y = var_1302)[name = tensor("cross")]; - tensor out_31 = add(x = inner, y = cross)[name = tensor("out_31")]; - tensor v_masked = mul(x = v_17, y = var_1125)[name = tensor("v_masked")]; - tensor var_1308 = mul(x = prev_kv, y = sqrt_s0)[name = tensor("op_1308")]; - tensor var_1310_transpose_x_1 = const()[name = tensor("op_1310_transpose_x_1"), val = tensor(true)]; - tensor var_1310_transpose_y_1 = const()[name = tensor("op_1310_transpose_y_1"), val = tensor(false)]; - tensor var_1310 = matmul(transpose_x = var_1310_transpose_x_1, transpose_y = var_1310_transpose_y_1, x = k_17, y = v_masked)[name = tensor("op_1310")]; - tensor new_kv_unnorm = add(x = var_1308, y = var_1310)[name = tensor("new_kv_unnorm")]; - tensor new_scale = add(x = prev_scale, y = var_1133)[name = tensor("new_scale")]; + tensor var_1317 = const()[name = tensor("op_1317"), val = tensor([3, 1])]; + tensor var_1318 = reshape(shape = var_1317, x = sqrt_s_t)[name = tensor("op_1318")]; + tensor M = real_div(x = causal_with_valid_1, y = var_1318)[name = tensor("M")]; + tensor var_1320 = mul(x = qk, y = M)[name = tensor("op_1320")]; + tensor inner_11_transpose_x_0 = const()[name = tensor("inner_11_transpose_x_0"), val = tensor(false)]; + tensor inner_11_transpose_y_0 = const()[name = tensor("inner_11_transpose_y_0"), val = tensor(false)]; + tensor v_17 = transpose(perm = v_17_perm_0, x = var_1302)[name = tensor("transpose_12")]; + tensor inner_11 = matmul(transpose_x = inner_11_transpose_x_0, transpose_y = inner_11_transpose_y_0, x = var_1320, y = v_17)[name = tensor("inner_11")]; + tensor var_1322_transpose_x_0 = const()[name = tensor("op_1322_transpose_x_0"), val = tensor(false)]; + tensor var_1322_transpose_y_0 = const()[name = tensor("op_1322_transpose_y_0"), val = tensor(false)]; + tensor var_1322 = matmul(transpose_x = var_1322_transpose_x_0, transpose_y = var_1322_transpose_y_0, x = q_17, y = prev_kv)[name = tensor("op_1322")]; + tensor var_1323 = real_div(x = sqrt_s0, y = sqrt_s_t)[name = tensor("op_1323")]; + tensor var_1324 = const()[name = tensor("op_1324"), val = tensor([1, 1, 3, 1])]; + tensor var_1325 = reshape(shape = var_1324, x = var_1323)[name = tensor("op_1325")]; + tensor cross = mul(x = var_1322, y = var_1325)[name = tensor("cross")]; + tensor out_31 = add(x = inner_11, y = cross)[name = tensor("out_31")]; + tensor v_masked = mul(x = v_17, y = var_1148)[name = tensor("v_masked")]; + tensor var_1331 = mul(x = prev_kv, y = sqrt_s0)[name = tensor("op_1331")]; + tensor var_1333_transpose_x_1 = const()[name = tensor("op_1333_transpose_x_1"), val = tensor(true)]; + tensor var_1333_transpose_y_1 = const()[name = tensor("op_1333_transpose_y_1"), val = tensor(false)]; + tensor var_1333 = matmul(transpose_x = var_1333_transpose_x_1, transpose_y = var_1333_transpose_y_1, x = k_17, y = v_masked)[name = tensor("op_1333")]; + tensor new_kv_unnorm = add(x = var_1331, y = var_1333)[name = tensor("new_kv_unnorm")]; + tensor new_scale = add(x = prev_scale, y = var_1156)[name = tensor("new_scale")]; tensor const_33 = const()[name = tensor("const_33"), val = tensor(0x1.fffffep+127)]; - tensor clip_4 = clip(alpha = var_990, beta = const_33, x = new_scale)[name = tensor("clip_4")]; + tensor clip_4 = clip(alpha = var_48, beta = const_33, x = new_scale)[name = tensor("clip_4")]; tensor sqrt_new_scale = sqrt(x = clip_4)[name = tensor("sqrt_new_scale")]; tensor nkv = real_div(x = new_kv_unnorm, y = sqrt_new_scale)[name = tensor("nkv")]; - tensor var_1319_perm_0 = const()[name = tensor("op_1319_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_1342_perm_0 = const()[name = tensor("op_1342_perm_0"), val = tensor([0, 2, 1, 3])]; tensor out_33_axes_0 = const()[name = tensor("out_33_axes_0"), val = tensor([-1])]; - tensor var_1319 = transpose(perm = var_1319_perm_0, x = out_31)[name = tensor("transpose_11")]; - tensor out_33 = layer_norm(axes = out_33_axes_0, epsilon = var_987, x = var_1319)[name = tensor("out_33")]; - tensor var_1323 = const()[name = tensor("op_1323"), val = tensor([6, 3, 256])]; - tensor out = reshape(shape = var_1323, x = out_33)[name = tensor("out")]; - tensor var_1325 = silu(x = input_187)[name = tensor("op_1325")]; - tensor input_189 = mul(x = var_1325, y = out)[name = tensor("input_189")]; - tensor ret_out = linear(bias = decoder_out_proj_1_bias, weight = decoder_out_proj_1_weight, x = input_189)[name = tensor("linear_51")]; - tensor input_191 = add(x = x, y = ret_out)[name = tensor("input_191")]; + tensor var_1342 = transpose(perm = var_1342_perm_0, x = out_31)[name = tensor("transpose_11")]; + tensor out_33 = layer_norm(axes = out_33_axes_0, epsilon = var_63, x = var_1342)[name = tensor("out_33")]; + tensor var_1346 = const()[name = tensor("op_1346"), val = tensor([6, 3, 256])]; + tensor out = reshape(shape = var_1346, x = out_33)[name = tensor("out")]; + tensor var_1348 = silu(x = input_189)[name = tensor("op_1348")]; + tensor input_191 = mul(x = var_1348, y = out)[name = tensor("input_191")]; + tensor ret_out = linear(bias = inner_decoder_out_proj_1_bias, weight = inner_decoder_out_proj_1_weight, x = input_191)[name = tensor("linear_51")]; + tensor input_193 = add(x = x, y = ret_out)[name = tensor("input_193")]; tensor xt_5_axes_0 = const()[name = tensor("xt_5_axes_0"), val = tensor([-1])]; - tensor xt_5 = layer_norm(axes = xt_5_axes_0, beta = decoder_norm11_1_bias, epsilon = var_985, gamma = decoder_norm11_1_weight, x = input_191)[name = tensor("xt_5")]; - tensor var_1335 = const()[name = tensor("op_1335"), val = tensor([1, 6, 3, 256])]; - tensor var_1336 = reshape(shape = var_1335, x = xt_5)[name = tensor("op_1336")]; - tensor var_1337_perm_0 = const()[name = tensor("op_1337_perm_0"), val = tensor([0, 2, 1, 3])]; - tensor var_1340 = const()[name = tensor("op_1340"), val = tensor([3, 6, 256])]; - tensor var_1337 = transpose(perm = var_1337_perm_0, x = var_1336)[name = tensor("transpose_10")]; - tensor query_5 = reshape(shape = var_1340, x = var_1337)[name = tensor("query_5")]; + tensor xt_5 = layer_norm(axes = xt_5_axes_0, beta = inner_decoder_norm11_1_bias, epsilon = var_55, gamma = inner_decoder_norm11_1_weight, x = input_193)[name = tensor("xt_5")]; + tensor var_1358 = const()[name = tensor("op_1358"), val = tensor([1, 6, 3, 256])]; + tensor var_1359 = reshape(shape = var_1358, x = xt_5)[name = tensor("op_1359")]; + tensor var_1360_perm_0 = const()[name = tensor("op_1360_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_1363 = const()[name = tensor("op_1363"), val = tensor([3, 6, 256])]; + tensor var_1360 = transpose(perm = var_1360_perm_0, x = var_1359)[name = tensor("transpose_10")]; + tensor query_5 = reshape(shape = var_1363, x = var_1360)[name = tensor("query_5")]; tensor query_perm_0 = const()[name = tensor("query_perm_0"), val = tensor([1, 0, 2])]; tensor query = transpose(perm = query_perm_0, x = query_5)[name = tensor("transpose_9")]; - tensor var_1363 = linear(bias = decoder_self_attn2_1_in_proj_bias, weight = decoder_self_attn2_1_in_proj_weight, x = query)[name = tensor("linear_52")]; + tensor var_1386 = linear(bias = inner_decoder_self_attn2_1_in_proj_bias, weight = inner_decoder_self_attn2_1_in_proj_weight, x = query)[name = tensor("linear_52")]; tensor concat_2 = const()[name = tensor("concat_2"), val = tensor([6, 3, 3, 256])]; - tensor var_1365 = reshape(shape = concat_2, x = var_1363)[name = tensor("op_1365")]; - tensor var_1366_axes_0 = const()[name = tensor("op_1366_axes_0"), val = tensor([0])]; - tensor var_1366 = expand_dims(axes = var_1366_axes_0, x = var_1365)[name = tensor("op_1366")]; - tensor var_1367_perm_0 = const()[name = tensor("op_1367_perm_0"), val = tensor([-2, 1, 2, 0, 4])]; - tensor var_1368_axes_0 = const()[name = tensor("op_1368_axes_0"), val = tensor([-2])]; - tensor var_1367 = transpose(perm = var_1367_perm_0, x = var_1366)[name = tensor("transpose_8")]; - tensor var_1368 = squeeze(axes = var_1368_axes_0, x = var_1367)[name = tensor("op_1368")]; + tensor var_1388 = reshape(shape = concat_2, x = var_1386)[name = tensor("op_1388")]; + tensor var_1389_axes_0 = const()[name = tensor("op_1389_axes_0"), val = tensor([0])]; + tensor var_1389 = expand_dims(axes = var_1389_axes_0, x = var_1388)[name = tensor("op_1389")]; + tensor var_1390_perm_0 = const()[name = tensor("op_1390_perm_0"), val = tensor([-2, 1, 2, 0, 4])]; + tensor var_1391_axes_0 = const()[name = tensor("op_1391_axes_0"), val = tensor([-2])]; + tensor var_1390 = transpose(perm = var_1390_perm_0, x = var_1389)[name = tensor("transpose_8")]; + tensor var_1391 = squeeze(axes = var_1391_axes_0, x = var_1390)[name = tensor("op_1391")]; tensor q_19_begin_0 = const()[name = tensor("q_19_begin_0"), val = tensor([0, 0, 0, 0])]; tensor q_19_end_0 = const()[name = tensor("q_19_end_0"), val = tensor([1, 6, 3, 256])]; tensor q_19_end_mask_0 = const()[name = tensor("q_19_end_mask_0"), val = tensor([false, true, true, true])]; tensor q_19_squeeze_mask_0 = const()[name = tensor("q_19_squeeze_mask_0"), val = tensor([true, false, false, false])]; - tensor q_19 = slice_by_index(begin = q_19_begin_0, end = q_19_end_0, end_mask = q_19_end_mask_0, squeeze_mask = q_19_squeeze_mask_0, x = var_1368)[name = tensor("q_19")]; + tensor q_19 = slice_by_index(begin = q_19_begin_0, end = q_19_end_0, end_mask = q_19_end_mask_0, squeeze_mask = q_19_squeeze_mask_0, x = var_1391)[name = tensor("q_19")]; tensor k_19_begin_0 = const()[name = tensor("k_19_begin_0"), val = tensor([1, 0, 0, 0])]; tensor k_19_end_0 = const()[name = tensor("k_19_end_0"), val = tensor([2, 6, 3, 256])]; tensor k_19_end_mask_0 = const()[name = tensor("k_19_end_mask_0"), val = tensor([false, true, true, true])]; tensor k_19_squeeze_mask_0 = const()[name = tensor("k_19_squeeze_mask_0"), val = tensor([true, false, false, false])]; - tensor k_19 = slice_by_index(begin = k_19_begin_0, end = k_19_end_0, end_mask = k_19_end_mask_0, squeeze_mask = k_19_squeeze_mask_0, x = var_1368)[name = tensor("k_19")]; + tensor k_19 = slice_by_index(begin = k_19_begin_0, end = k_19_end_0, end_mask = k_19_end_mask_0, squeeze_mask = k_19_squeeze_mask_0, x = var_1391)[name = tensor("k_19")]; tensor v_19_begin_0 = const()[name = tensor("v_19_begin_0"), val = tensor([2, 0, 0, 0])]; tensor v_19_end_0 = const()[name = tensor("v_19_end_0"), val = tensor([3, 6, 3, 256])]; tensor v_19_end_mask_0 = const()[name = tensor("v_19_end_mask_0"), val = tensor([false, true, true, true])]; tensor v_19_squeeze_mask_0 = const()[name = tensor("v_19_squeeze_mask_0"), val = tensor([true, false, false, false])]; - tensor v_19 = slice_by_index(begin = v_19_begin_0, end = v_19_end_0, end_mask = v_19_end_mask_0, squeeze_mask = v_19_squeeze_mask_0, x = var_1368)[name = tensor("v_19")]; - tensor var_1376 = const()[name = tensor("op_1376"), val = tensor([6, 12, 64])]; - tensor var_1377 = reshape(shape = var_1376, x = q_19)[name = tensor("op_1377")]; + tensor v_19 = slice_by_index(begin = v_19_begin_0, end = v_19_end_0, end_mask = v_19_end_mask_0, squeeze_mask = v_19_squeeze_mask_0, x = var_1391)[name = tensor("v_19")]; + tensor var_1399 = const()[name = tensor("op_1399"), val = tensor([6, 12, 64])]; + tensor var_1400 = reshape(shape = var_1399, x = q_19)[name = tensor("op_1400")]; tensor q_21_perm_0 = const()[name = tensor("q_21_perm_0"), val = tensor([1, 0, 2])]; - tensor var_1383 = const()[name = tensor("op_1383"), val = tensor([6, 12, 64])]; - tensor var_1384 = reshape(shape = var_1383, x = k_19)[name = tensor("op_1384")]; + tensor var_1406 = const()[name = tensor("op_1406"), val = tensor([6, 12, 64])]; + tensor var_1407 = reshape(shape = var_1406, x = k_19)[name = tensor("op_1407")]; tensor k_21_perm_0 = const()[name = tensor("k_21_perm_0"), val = tensor([1, 0, 2])]; - tensor var_1390 = const()[name = tensor("op_1390"), val = tensor([6, 12, 64])]; - tensor var_1391 = reshape(shape = var_1390, x = v_19)[name = tensor("op_1391")]; + tensor var_1413 = const()[name = tensor("op_1413"), val = tensor([6, 12, 64])]; + tensor var_1414 = reshape(shape = var_1413, x = v_19)[name = tensor("op_1414")]; tensor v_21_perm_0 = const()[name = tensor("v_21_perm_0"), val = tensor([1, 0, 2])]; - tensor var_1394 = const()[name = tensor("op_1394"), val = tensor([3, 4, 6, 64])]; - tensor q_21 = transpose(perm = q_21_perm_0, x = var_1377)[name = tensor("transpose_7")]; - tensor q = reshape(shape = var_1394, x = q_21)[name = tensor("q")]; - tensor var_1396 = const()[name = tensor("op_1396"), val = tensor([3, 4, 6, 64])]; - tensor k_21 = transpose(perm = k_21_perm_0, x = var_1384)[name = tensor("transpose_6")]; - tensor k = reshape(shape = var_1396, x = k_21)[name = tensor("k")]; - tensor var_1398 = const()[name = tensor("op_1398"), val = tensor([3, 4, 6, 64])]; - tensor v_21 = transpose(perm = v_21_perm_0, x = var_1391)[name = tensor("transpose_5")]; - tensor v = reshape(shape = var_1398, x = v_21)[name = tensor("v")]; + tensor var_1417 = const()[name = tensor("op_1417"), val = tensor([3, 4, 6, 64])]; + tensor q_21 = transpose(perm = q_21_perm_0, x = var_1400)[name = tensor("transpose_7")]; + tensor q = reshape(shape = var_1417, x = q_21)[name = tensor("q")]; + tensor var_1419 = const()[name = tensor("op_1419"), val = tensor([3, 4, 6, 64])]; + tensor k_21 = transpose(perm = k_21_perm_0, x = var_1407)[name = tensor("transpose_6")]; + tensor k = reshape(shape = var_1419, x = k_21)[name = tensor("k")]; + tensor var_1421 = const()[name = tensor("op_1421"), val = tensor([3, 4, 6, 64])]; + tensor v_21 = transpose(perm = v_21_perm_0, x = var_1414)[name = tensor("transpose_5")]; + tensor v = reshape(shape = var_1421, x = v_21)[name = tensor("v")]; tensor mul_3_y_0 = const()[name = tensor("mul_3_y_0"), val = tensor(0x1p-3)]; tensor mul_3 = mul(x = q, y = mul_3_y_0)[name = tensor("mul_3")]; tensor matmul_1_transpose_y_0 = const()[name = tensor("matmul_1_transpose_y_0"), val = tensor(true)]; @@ -1237,36 +1248,36 @@ program(1.0) tensor attn_output_9_transpose_x_0 = const()[name = tensor("attn_output_9_transpose_x_0"), val = tensor(false)]; tensor attn_output_9_transpose_y_0 = const()[name = tensor("attn_output_9_transpose_y_0"), val = tensor(false)]; tensor attn_output_9 = matmul(transpose_x = attn_output_9_transpose_x_0, transpose_y = attn_output_9_transpose_y_0, x = softmax_1, y = v)[name = tensor("attn_output_9")]; - tensor var_1401 = const()[name = tensor("op_1401"), val = tensor([2, 0, 1, 3])]; - tensor var_1406 = const()[name = tensor("op_1406"), val = tensor([18, 256])]; - tensor var_1402 = transpose(perm = var_1401, x = attn_output_9)[name = tensor("transpose_4")]; - tensor attn_output_11 = reshape(shape = var_1406, x = var_1402)[name = tensor("attn_output_11")]; - tensor attn_output_13 = linear(bias = decoder_self_attn2_1_out_proj_bias, weight = decoder_self_attn2_1_out_proj_weight, x = attn_output_11)[name = tensor("linear_53")]; - tensor var_1410 = const()[name = tensor("op_1410"), val = tensor([6, 3, 256])]; - tensor attn_output = reshape(shape = var_1410, x = attn_output_13)[name = tensor("attn_output")]; + tensor var_1424 = const()[name = tensor("op_1424"), val = tensor([2, 0, 1, 3])]; + tensor var_1429 = const()[name = tensor("op_1429"), val = tensor([18, 256])]; + tensor var_1425 = transpose(perm = var_1424, x = attn_output_9)[name = tensor("transpose_4")]; + tensor attn_output_11 = reshape(shape = var_1429, x = var_1425)[name = tensor("attn_output_11")]; + tensor attn_output_13 = linear(bias = inner_decoder_self_attn2_1_out_proj_bias, weight = inner_decoder_self_attn2_1_out_proj_weight, x = attn_output_11)[name = tensor("linear_53")]; + tensor var_1433 = const()[name = tensor("op_1433"), val = tensor([6, 3, 256])]; + tensor attn_output = reshape(shape = var_1433, x = attn_output_13)[name = tensor("attn_output")]; tensor sa2_out_perm_0 = const()[name = tensor("sa2_out_perm_0"), val = tensor([1, 0, 2])]; tensor sa2_out = transpose(perm = sa2_out_perm_0, x = attn_output)[name = tensor("transpose_3")]; - tensor input_193 = add(x = query_5, y = sa2_out)[name = tensor("input_193")]; - tensor input_195_axes_0 = const()[name = tensor("input_195_axes_0"), val = tensor([-1])]; - tensor input_195 = layer_norm(axes = input_195_axes_0, beta = decoder_norm21_1_bias, epsilon = var_985, gamma = decoder_norm21_1_weight, x = input_193)[name = tensor("input_195")]; - tensor input_197 = linear(bias = decoder_linear1_1_bias, weight = decoder_linear1_1_weight, x = input_195)[name = tensor("linear_54")]; - tensor input_199 = relu(x = input_197)[name = tensor("input_199")]; - tensor ffn_out = linear(bias = decoder_linear2_1_bias, weight = decoder_linear2_1_weight, x = input_199)[name = tensor("linear_55")]; - tensor input_201 = add(x = input_195, y = ffn_out)[name = tensor("input_201")]; + tensor input_195 = add(x = query_5, y = sa2_out)[name = tensor("input_195")]; + tensor input_197_axes_0 = const()[name = tensor("input_197_axes_0"), val = tensor([-1])]; + tensor input_197 = layer_norm(axes = input_197_axes_0, beta = inner_decoder_norm21_1_bias, epsilon = var_55, gamma = inner_decoder_norm21_1_weight, x = input_195)[name = tensor("input_197")]; + tensor input_199 = linear(bias = inner_decoder_linear1_1_bias, weight = inner_decoder_linear1_1_weight, x = input_197)[name = tensor("linear_54")]; + tensor input_201 = relu(x = input_199)[name = tensor("input_201")]; + tensor ffn_out = linear(bias = inner_decoder_linear2_1_bias, weight = inner_decoder_linear2_1_weight, x = input_201)[name = tensor("linear_55")]; + tensor input_203 = add(x = input_197, y = ffn_out)[name = tensor("input_203")]; tensor xt_axes_0 = const()[name = tensor("xt_axes_0"), val = tensor([-1])]; - tensor xt = layer_norm(axes = xt_axes_0, beta = decoder_norm22_1_bias, epsilon = var_985, gamma = decoder_norm22_1_weight, x = input_201)[name = tensor("xt")]; - tensor var_1430 = const()[name = tensor("op_1430"), val = tensor([1, 3, 6, 256])]; - tensor input = reshape(shape = var_1430, x = xt)[name = tensor("input")]; - tensor var_1432 = const()[name = tensor("op_1432"), val = tensor([-1])]; - tensor var_1433 = reduce_l2_norm(axes = var_1432, keep_dims = var_988, x = input)[name = tensor("op_1433")]; + tensor xt = layer_norm(axes = xt_axes_0, beta = inner_decoder_norm22_1_bias, epsilon = var_55, gamma = inner_decoder_norm22_1_weight, x = input_203)[name = tensor("xt")]; + tensor var_1453 = const()[name = tensor("op_1453"), val = tensor([1, 3, 6, 256])]; + tensor input = reshape(shape = var_1453, x = xt)[name = tensor("input")]; + tensor var_1455 = const()[name = tensor("op_1455"), val = tensor([-1])]; + tensor var_1456 = reduce_l2_norm(axes = var_1455, keep_dims = var_54, x = input)[name = tensor("op_1456")]; tensor const_42 = const()[name = tensor("const_42"), val = tensor(0x1.fffffep+127)]; - tensor clip_5 = clip(alpha = var_980, beta = const_42, x = var_1433)[name = tensor("clip_5")]; - tensor var_1435 = real_div(x = input, y = clip_5)[name = tensor("op_1435")]; + tensor clip_5 = clip(alpha = var_68, beta = const_42, x = var_1456)[name = tensor("clip_5")]; + tensor var_1458 = real_div(x = input, y = clip_5)[name = tensor("op_1458")]; tensor concat_6 = const()[name = tensor("concat_6"), val = tensor([3, 1, 256])]; tensor reshape_0 = reshape(shape = concat_6, x = emb)[name = tensor("reshape_0")]; tensor transpose_1_perm_0 = const()[name = tensor("transpose_1_perm_0"), val = tensor([0, 1, 3, 2])]; tensor concat_7 = const()[name = tensor("concat_7"), val = tensor([3, 256, 6])]; - tensor transpose_1 = transpose(perm = transpose_1_perm_0, x = var_1435)[name = tensor("transpose_2")]; + tensor transpose_1 = transpose(perm = transpose_1_perm_0, x = var_1458)[name = tensor("transpose_2")]; tensor reshape_1 = reshape(shape = concat_7, x = transpose_1)[name = tensor("reshape_1")]; tensor matmul_2_transpose_x_0 = const()[name = tensor("matmul_2_transpose_x_0"), val = tensor(false)]; tensor matmul_2_transpose_y_0 = const()[name = tensor("matmul_2_transpose_y_0"), val = tensor(false)]; @@ -1277,10 +1288,10 @@ program(1.0) tensor output_end_0 = const()[name = tensor("output_end_0"), val = tensor([1, 3, 5])]; tensor output_end_mask_0 = const()[name = tensor("output_end_mask_0"), val = tensor([true, true, false])]; tensor output = slice_by_index(begin = output_begin_0, end = output_end_0, end_mask = output_end_mask_0, x = reshape_2)[name = tensor("output")]; - tensor probs = sigmoid(x = output)[name = tensor("op_1439")]; - tensor var_1441_axis_0 = const()[name = tensor("op_1441_axis_0"), val = tensor(0)]; - tensor dec_kv_new = stack(axis = var_1441_axis_0, values = (var_1137, nkv))[name = tensor("op_1441")]; - tensor var_1443_axis_0 = const()[name = tensor("op_1443_axis_0"), val = tensor(0)]; - tensor dec_scale_new = stack(axis = var_1443_axis_0, values = (new_scale_9, new_scale))[name = tensor("op_1443")]; + tensor probs = sigmoid(x = output)[name = tensor("op_1462")]; + tensor var_1464_axis_0 = const()[name = tensor("op_1464_axis_0"), val = tensor(0)]; + tensor dec_kv_new = stack(axis = var_1464_axis_0, values = (var_1160, nkv))[name = tensor("op_1464")]; + tensor var_1466_axis_0 = const()[name = tensor("op_1466_axis_0"), val = tensor(0)]; + tensor dec_scale_new = stack(axis = var_1466_axis_0, values = (new_scale_9, new_scale))[name = tensor("op_1466")]; } -> (probs, enc_kv_new, enc_scale_new, enc_conv_cache_new, cnn_window_new, dec_kv_new, dec_scale_new); } \ No newline at end of file diff --git a/optimized/ami/300ms/ls_eend_ami_300ms.mlpackage/Data/com.apple.CoreML/model.mlmodel b/optimized/ami/300ms/ls_eend_ami_300ms.mlpackage/Data/com.apple.CoreML/model.mlmodel index 3ee6a266410b4c8f967985e6dd4010971e326dda..bce4e082d0c99aa965c5b822189ac93ab0474b36 100644 --- a/optimized/ami/300ms/ls_eend_ami_300ms.mlpackage/Data/com.apple.CoreML/model.mlmodel +++ b/optimized/ami/300ms/ls_eend_ami_300ms.mlpackage/Data/com.apple.CoreML/model.mlmodel @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:7c02e235e84b720c6aaf09d92bbac369c38d556327dd8f5f72579d9e76c86d43 -size 185451 +oid sha256:7b50c82d16896cf3aa0109f4b74206ce7f7fdcbdee5fe341dfa6406075be97d3 +size 190996 diff --git a/optimized/ami/300ms/ls_eend_ami_300ms.mlpackage/Manifest.json b/optimized/ami/300ms/ls_eend_ami_300ms.mlpackage/Manifest.json index c9c2cd6e7172b08333ab9eafcbbc60ac652c7e02..2020c11765ac2eddf1513882c9c54dcecf8ce40d 100644 --- a/optimized/ami/300ms/ls_eend_ami_300ms.mlpackage/Manifest.json +++ b/optimized/ami/300ms/ls_eend_ami_300ms.mlpackage/Manifest.json @@ -1,18 +1,18 @@ { "fileFormatVersion": "1.0.0", "itemInfoEntries": { - "68330719-A46B-4379-8B85-C9B4E8FD52AF": { + "8152B853-43F4-4A6B-92C5-99387E9814E3": { "author": "com.apple.CoreML", "description": "CoreML Model Specification", "name": "model.mlmodel", "path": "com.apple.CoreML/model.mlmodel" }, - "E31DDEE0-26D5-4CE8-AD79-31130F2DF498": { + "F91AC4AF-14C0-45F7-BDD2-A116D8D48885": { "author": "com.apple.CoreML", "description": "CoreML Model Weights", "name": "weights", "path": "com.apple.CoreML/weights" } }, - "rootModelIdentifier": "68330719-A46B-4379-8B85-C9B4E8FD52AF" + "rootModelIdentifier": "8152B853-43F4-4A6B-92C5-99387E9814E3" } diff --git a/optimized/ami/400ms/ls_eend_ami_400ms.mlmodelc/analytics/coremldata.bin b/optimized/ami/400ms/ls_eend_ami_400ms.mlmodelc/analytics/coremldata.bin index 991ad5aa06a48e1d0f11baf09e0e8defe19c30ba..1f275b6c49d69319653d30ab214c476112d49cd3 100644 --- a/optimized/ami/400ms/ls_eend_ami_400ms.mlmodelc/analytics/coremldata.bin +++ b/optimized/ami/400ms/ls_eend_ami_400ms.mlmodelc/analytics/coremldata.bin @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:b5f8f8f57d8e374772e2d54b8950c8de8c6a3a8cc10450c3832b9bf5e75ce240 +oid sha256:a6758d5ecdb39569a10f08b51c266206c2824dc825070864428f42ace8434465 size 243 diff --git a/optimized/ami/400ms/ls_eend_ami_400ms.mlmodelc/coremldata.bin b/optimized/ami/400ms/ls_eend_ami_400ms.mlmodelc/coremldata.bin index 2e9e19cf1f89eb088959423ab8f66c6e00895ceb..8d4b5942ba1b7b92e69b60b50b2b17899354b405 100644 --- a/optimized/ami/400ms/ls_eend_ami_400ms.mlmodelc/coremldata.bin +++ b/optimized/ami/400ms/ls_eend_ami_400ms.mlmodelc/coremldata.bin @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:71e5da37ce4f16ebb4f5b2a2ef320e4f17aee294a47d49fa31f021a549ed1af3 -size 1292 +oid sha256:976e4dbdf87c668f484460fb519558b7c5d6040a00e786a40be1c8ba08e57ac9 +size 1395 diff --git a/optimized/ami/400ms/ls_eend_ami_400ms.mlmodelc/metadata.json b/optimized/ami/400ms/ls_eend_ami_400ms.mlmodelc/metadata.json index 02cd719d1f593b541368c72808de5e7b68e572ad..789e43d6c81056b77a84b77dc3a1249d69757faa 100644 --- a/optimized/ami/400ms/ls_eend_ami_400ms.mlmodelc/metadata.json +++ b/optimized/ami/400ms/ls_eend_ami_400ms.mlmodelc/metadata.json @@ -1,7 +1,7 @@ [ { "metadataOutputVersion" : "3.0", - "shortDescription" : "LS-EEND AMI streaming diarizer (pipeline, T=4, max_speakers=4)", + "shortDescription" : "LS-EEND AMI streaming diarizer (pipeline, T=4, max_speakers=4, layout=raw_mel, cu=all)", "outputSchema" : [ { "hasShapeFlexibility" : "0", @@ -81,7 +81,7 @@ "specificationVersion" : 8, "mlProgramOperationTypeHistogram" : { "Ios16.reduceL2Norm" : 2, - "Ios17.reshape" : 68, + "Ios17.reshape" : 69, "Ios16.softmax" : 2, "Ios17.matmul" : 29, "Ios17.transpose" : 57, @@ -89,7 +89,7 @@ "Ios17.expandDims" : 3, "Ios17.add" : 46, "Ios16.sigmoid" : 5, - "Ios17.sliceByIndex" : 64, + "Ios17.sliceByIndex" : 68, "Tile" : 1, "Ios16.reduceSum" : 1, "Ios17.squeeze" : 2, @@ -101,7 +101,7 @@ "Ios16.silu" : 18, "Ios17.realDiv" : 20, "Ios17.linear" : 56, - "Stack" : 5, + "Stack" : 6, "Ios17.concat" : 22, "Ios16.relu" : 2, "Ios16.cumsum" : 1, @@ -128,9 +128,9 @@ "hasShapeFlexibility" : "0", "isOptional" : "0", "dataType" : "Float32", - "formattedType" : "MultiArray (Float32 1 × 4 × 345)", + "formattedType" : "MultiArray (Float32 1 × 45 × 23)", "shortDescription" : "", - "shape" : "[1, 4, 345]", + "shape" : "[1, 45, 23]", "name" : "features", "type" : "MultiArray" }, @@ -206,8 +206,8 @@ } ], "userDefinedMetadata" : { - "com.github.apple.coremltools.conversion_date" : "2026-04-16", - "config" : "{\"model_name\": \"ami\", \"model_label\": \"AMI\", \"variant\": \"pipeline\", \"chunk_size\": 4, \"step_duration_ms\": 400, \"frame_duration_ms\": 100, \"frame_duration_seconds\": 0.1, \"max_speakers\": 4, \"max_nspks\": 6, \"n_units\": 256, \"n_heads\": 4, \"enc_n_layers\": 4, \"dec_n_layers\": 2, \"conv_kernel_size\": 16, \"conv_delay\": 9, \"sample_rate\": 8000, \"win_length\": 200, \"hop_length\": 80, \"n_mels\": 23, \"context_size\": 7, \"subsampling\": 10, \"feat_type\": \"logmel23_cummn\", \"pure_roll\": true}", + "com.github.apple.coremltools.conversion_date" : "2026-04-18", + "config" : "{\"model_name\": \"ami\", \"model_label\": \"AMI\", \"variant\": \"pipeline\", \"chunk_size\": 4, \"step_duration_ms\": 400, \"frame_duration_ms\": 100, \"frame_duration_seconds\": 0.1, \"max_speakers\": 4, \"max_nspks\": 6, \"n_units\": 256, \"n_heads\": 4, \"enc_n_layers\": 4, \"dec_n_layers\": 2, \"conv_kernel_size\": 16, \"conv_delay\": 9, \"sample_rate\": 8000, \"win_length\": 200, \"hop_length\": 80, \"n_mels\": 23, \"context_size\": 7, \"subsampling\": 10, \"feat_type\": \"logmel23_cummn\", \"pure_roll\": true, \"input_layout\": \"raw_mel\", \"compute_units_export\": \"all\", \"raw_mel_length\": 45}", "com.github.apple.coremltools.source" : "torch==2.6.0", "com.github.apple.coremltools.version" : "9.0", "com.github.apple.coremltools.source_dialect" : "TorchScript" diff --git a/optimized/ami/400ms/ls_eend_ami_400ms.mlmodelc/model.mil b/optimized/ami/400ms/ls_eend_ami_400ms.mlmodelc/model.mil index 189cd5c18852bac2bde041df8ebc35de2ca1a491..9e45f1e77725011414fcf74aa7144ae43d4ba02e 100644 --- a/optimized/ami/400ms/ls_eend_ami_400ms.mlmodelc/model.mil +++ b/optimized/ami/400ms/ls_eend_ami_400ms.mlmodelc/model.mil @@ -1,234 +1,256 @@ program(1.0) [buildInfo = dict, tensor>({{"coremlc-component-MIL", "3510.2.1"}, {"coremlc-version", "3500.32.1"}, {"coremltools-component-torch", "2.6.0"}, {"coremltools-source-dialect", "TorchScript"}, {"coremltools-version", "9.0"}})] { - func main(tensor cnn_window, tensor dec_kv, tensor dec_scale, tensor enc_conv_cache, tensor enc_kv, tensor enc_scale, tensor features, tensor valid_mask) { - tensor encoder_cnn_bias = const()[name = tensor("encoder_cnn_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(64)))]; - tensor encoder_cnn_weight = const()[name = tensor("encoder_cnn_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1152)))]; - tensor encoder__causal_mask = const()[name = tensor("encoder__causal_mask"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4981952)))]; - tensor encoder__t_index = const()[name = tensor("encoder__t_index"), val = tensor([0x1p+0, 0x1p+1, 0x1.8p+1, 0x1p+2])]; - tensor encoder_input_projection_linear_bias = const()[name = tensor("encoder_input_projection_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4982080)))]; - tensor encoder_input_projection_linear_weight = const()[name = tensor("encoder_input_projection_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4983168)))]; - tensor encoder_pre_layer_norm_bias = const()[name = tensor("encoder_pre_layer_norm_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(5336512)))]; - tensor encoder_pre_layer_norm_weight = const()[name = tensor("encoder_pre_layer_norm_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(5337600)))]; - tensor encoder_ffn1_0_module_sequential_0_bias = const()[name = tensor("encoder_ffn1_0_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(5338688)))]; - tensor encoder_ffn1_0_module_sequential_0_weight = const()[name = tensor("encoder_ffn1_0_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(5339776)))]; - tensor encoder_ffn1_0_module_sequential_1_linear_bias = const()[name = tensor("encoder_ffn1_0_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(5340864)))]; - tensor encoder_ffn1_0_module_sequential_1_linear_weight = const()[name = tensor("encoder_ffn1_0_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(5345024)))]; - tensor encoder_ffn1_0_module_sequential_4_linear_bias = const()[name = tensor("encoder_ffn1_0_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(6393664)))]; - tensor encoder_ffn1_0_module_sequential_4_linear_weight = const()[name = tensor("encoder_ffn1_0_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(6394752)))]; - tensor encoder_ret_lns_0_bias = const()[name = tensor("encoder_ret_lns_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7443392)))]; - tensor encoder_ret_lns_0_weight = const()[name = tensor("encoder_ret_lns_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7444480)))]; - tensor encoder_q_proj_0_bias = const()[name = tensor("encoder_q_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7445568)))]; - tensor encoder_q_proj_0_weight = const()[name = tensor("encoder_q_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7446656)))]; - tensor encoder_k_proj_0_bias = const()[name = tensor("encoder_k_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7708864)))]; - tensor encoder_k_proj_0_weight = const()[name = tensor("encoder_k_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7709952)))]; - tensor encoder_v_proj_0_bias = const()[name = tensor("encoder_v_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7972160)))]; - tensor encoder_v_proj_0_weight = const()[name = tensor("encoder_v_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7973248)))]; - tensor encoder_g_proj_0_bias = const()[name = tensor("encoder_g_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8235456)))]; - tensor encoder_g_proj_0_weight = const()[name = tensor("encoder_g_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8236544)))]; - tensor encoder_out_proj_0_bias = const()[name = tensor("encoder_out_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8498752)))]; - tensor encoder_out_proj_0_weight = const()[name = tensor("encoder_out_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8499840)))]; - tensor encoder_conv_module_0_sequential_0_bias = const()[name = tensor("encoder_conv_module_0_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8762048)))]; - tensor encoder_conv_module_0_sequential_0_weight = const()[name = tensor("encoder_conv_module_0_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8763136)))]; - tensor encoder_conv_module_0_sequential_2_conv_bias = const()[name = tensor("encoder_conv_module_0_sequential_2_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8764224)))]; - tensor encoder_conv_module_0_sequential_2_conv_weight = const()[name = tensor("encoder_conv_module_0_sequential_2_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8766336)))]; - tensor encoder_conv_module_0_sequential_4_conv_weight = const()[name = tensor("encoder_conv_module_0_sequential_4_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9290688)))]; - tensor encoder_conv_module_0_sequential_5_running_var = const()[name = tensor("encoder_conv_module_0_sequential_5_running_var"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9307136)))]; - tensor encoder_conv_module_0_sequential_5_running_mean = const()[name = tensor("encoder_conv_module_0_sequential_5_running_mean"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9308224)))]; - tensor encoder_conv_module_0_sequential_5_bias = const()[name = tensor("encoder_conv_module_0_sequential_5_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9309312)))]; - tensor encoder_conv_module_0_sequential_5_weight = const()[name = tensor("encoder_conv_module_0_sequential_5_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9310400)))]; - tensor encoder_conv_module_0_sequential_7_conv_bias = const()[name = tensor("encoder_conv_module_0_sequential_7_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9311488)))]; - tensor encoder_conv_module_0_sequential_7_conv_weight = const()[name = tensor("encoder_conv_module_0_sequential_7_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9312576)))]; - tensor encoder_ffn2_0_module_sequential_0_bias = const()[name = tensor("encoder_ffn2_0_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9574784)))]; - tensor encoder_ffn2_0_module_sequential_0_weight = const()[name = tensor("encoder_ffn2_0_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9575872)))]; - tensor encoder_ffn2_0_module_sequential_1_linear_bias = const()[name = tensor("encoder_ffn2_0_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9576960)))]; - tensor encoder_ffn2_0_module_sequential_1_linear_weight = const()[name = tensor("encoder_ffn2_0_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9581120)))]; - tensor encoder_ffn2_0_module_sequential_4_linear_bias = const()[name = tensor("encoder_ffn2_0_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(10629760)))]; - tensor encoder_ffn2_0_module_sequential_4_linear_weight = const()[name = tensor("encoder_ffn2_0_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(10630848)))]; - tensor encoder_layer_norm_0_bias = const()[name = tensor("encoder_layer_norm_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11679488)))]; - tensor encoder_layer_norm_0_weight = const()[name = tensor("encoder_layer_norm_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11680576)))]; - tensor encoder_ffn1_1_module_sequential_0_bias = const()[name = tensor("encoder_ffn1_1_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11681664)))]; - tensor encoder_ffn1_1_module_sequential_0_weight = const()[name = tensor("encoder_ffn1_1_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11682752)))]; - tensor encoder_ffn1_1_module_sequential_1_linear_bias = const()[name = tensor("encoder_ffn1_1_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11683840)))]; - tensor encoder_ffn1_1_module_sequential_1_linear_weight = const()[name = tensor("encoder_ffn1_1_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11688000)))]; - tensor encoder_ffn1_1_module_sequential_4_linear_bias = const()[name = tensor("encoder_ffn1_1_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(12736640)))]; - tensor encoder_ffn1_1_module_sequential_4_linear_weight = const()[name = tensor("encoder_ffn1_1_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(12737728)))]; - tensor encoder_ret_lns_1_bias = const()[name = tensor("encoder_ret_lns_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(13786368)))]; - tensor encoder_ret_lns_1_weight = const()[name = tensor("encoder_ret_lns_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(13787456)))]; - tensor encoder_q_proj_1_bias = const()[name = tensor("encoder_q_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(13788544)))]; - tensor encoder_q_proj_1_weight = const()[name = tensor("encoder_q_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(13789632)))]; - tensor encoder_k_proj_1_bias = const()[name = tensor("encoder_k_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14051840)))]; - tensor encoder_k_proj_1_weight = const()[name = tensor("encoder_k_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14052928)))]; - tensor encoder_v_proj_1_bias = const()[name = tensor("encoder_v_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14315136)))]; - tensor encoder_v_proj_1_weight = const()[name = tensor("encoder_v_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14316224)))]; - tensor encoder_g_proj_1_bias = const()[name = tensor("encoder_g_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14578432)))]; - tensor encoder_g_proj_1_weight = const()[name = tensor("encoder_g_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14579520)))]; - tensor encoder_out_proj_1_bias = const()[name = tensor("encoder_out_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14841728)))]; - tensor encoder_out_proj_1_weight = const()[name = tensor("encoder_out_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14842816)))]; - tensor encoder_conv_module_1_sequential_0_bias = const()[name = tensor("encoder_conv_module_1_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15105024)))]; - tensor encoder_conv_module_1_sequential_0_weight = const()[name = tensor("encoder_conv_module_1_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15106112)))]; - tensor encoder_conv_module_1_sequential_2_conv_bias = const()[name = tensor("encoder_conv_module_1_sequential_2_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15107200)))]; - tensor encoder_conv_module_1_sequential_2_conv_weight = const()[name = tensor("encoder_conv_module_1_sequential_2_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15109312)))]; - tensor encoder_conv_module_1_sequential_4_conv_weight = const()[name = tensor("encoder_conv_module_1_sequential_4_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15633664)))]; - tensor encoder_conv_module_1_sequential_5_running_var = const()[name = tensor("encoder_conv_module_1_sequential_5_running_var"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15650112)))]; - tensor encoder_conv_module_1_sequential_5_running_mean = const()[name = tensor("encoder_conv_module_1_sequential_5_running_mean"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15651200)))]; - tensor encoder_conv_module_1_sequential_5_bias = const()[name = tensor("encoder_conv_module_1_sequential_5_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15652288)))]; - tensor encoder_conv_module_1_sequential_5_weight = const()[name = tensor("encoder_conv_module_1_sequential_5_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15653376)))]; - tensor encoder_conv_module_1_sequential_7_conv_bias = const()[name = tensor("encoder_conv_module_1_sequential_7_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15654464)))]; - tensor encoder_conv_module_1_sequential_7_conv_weight = const()[name = tensor("encoder_conv_module_1_sequential_7_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15655552)))]; - tensor encoder_ffn2_1_module_sequential_0_bias = const()[name = tensor("encoder_ffn2_1_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15917760)))]; - tensor encoder_ffn2_1_module_sequential_0_weight = const()[name = tensor("encoder_ffn2_1_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15918848)))]; - tensor encoder_ffn2_1_module_sequential_1_linear_bias = const()[name = tensor("encoder_ffn2_1_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15919936)))]; - tensor encoder_ffn2_1_module_sequential_1_linear_weight = const()[name = tensor("encoder_ffn2_1_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15924096)))]; - tensor encoder_ffn2_1_module_sequential_4_linear_bias = const()[name = tensor("encoder_ffn2_1_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(16972736)))]; - tensor encoder_ffn2_1_module_sequential_4_linear_weight = const()[name = tensor("encoder_ffn2_1_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(16973824)))]; - tensor encoder_layer_norm_1_bias = const()[name = tensor("encoder_layer_norm_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18022464)))]; - tensor encoder_layer_norm_1_weight = const()[name = tensor("encoder_layer_norm_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18023552)))]; - tensor encoder_ffn1_2_module_sequential_0_bias = const()[name = tensor("encoder_ffn1_2_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18024640)))]; - tensor encoder_ffn1_2_module_sequential_0_weight = const()[name = tensor("encoder_ffn1_2_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18025728)))]; - tensor encoder_ffn1_2_module_sequential_1_linear_bias = const()[name = tensor("encoder_ffn1_2_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18026816)))]; - tensor encoder_ffn1_2_module_sequential_1_linear_weight = const()[name = tensor("encoder_ffn1_2_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18030976)))]; - tensor encoder_ffn1_2_module_sequential_4_linear_bias = const()[name = tensor("encoder_ffn1_2_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(19079616)))]; - tensor encoder_ffn1_2_module_sequential_4_linear_weight = const()[name = tensor("encoder_ffn1_2_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(19080704)))]; - tensor encoder_ret_lns_2_bias = const()[name = tensor("encoder_ret_lns_2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20129344)))]; - tensor encoder_ret_lns_2_weight = const()[name = tensor("encoder_ret_lns_2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20130432)))]; - tensor encoder_q_proj_2_bias = const()[name = tensor("encoder_q_proj_2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20131520)))]; - tensor encoder_q_proj_2_weight = const()[name = tensor("encoder_q_proj_2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20132608)))]; - tensor encoder_k_proj_2_bias = const()[name = tensor("encoder_k_proj_2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20394816)))]; - tensor encoder_k_proj_2_weight = const()[name = tensor("encoder_k_proj_2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20395904)))]; - tensor encoder_v_proj_2_bias = const()[name = tensor("encoder_v_proj_2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20658112)))]; - tensor encoder_v_proj_2_weight = const()[name = tensor("encoder_v_proj_2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20659200)))]; - tensor encoder_g_proj_2_bias = const()[name = tensor("encoder_g_proj_2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20921408)))]; - tensor encoder_g_proj_2_weight = const()[name = tensor("encoder_g_proj_2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20922496)))]; - tensor encoder_out_proj_2_bias = const()[name = tensor("encoder_out_proj_2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21184704)))]; - tensor encoder_out_proj_2_weight = const()[name = tensor("encoder_out_proj_2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21185792)))]; - tensor encoder_conv_module_2_sequential_0_bias = const()[name = tensor("encoder_conv_module_2_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21448000)))]; - tensor encoder_conv_module_2_sequential_0_weight = const()[name = tensor("encoder_conv_module_2_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21449088)))]; - tensor encoder_conv_module_2_sequential_2_conv_bias = const()[name = tensor("encoder_conv_module_2_sequential_2_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21450176)))]; - tensor encoder_conv_module_2_sequential_2_conv_weight = const()[name = tensor("encoder_conv_module_2_sequential_2_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21452288)))]; - tensor encoder_conv_module_2_sequential_4_conv_weight = const()[name = tensor("encoder_conv_module_2_sequential_4_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21976640)))]; - tensor encoder_conv_module_2_sequential_5_running_var = const()[name = tensor("encoder_conv_module_2_sequential_5_running_var"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21993088)))]; - tensor encoder_conv_module_2_sequential_5_running_mean = const()[name = tensor("encoder_conv_module_2_sequential_5_running_mean"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21994176)))]; - tensor encoder_conv_module_2_sequential_5_bias = const()[name = tensor("encoder_conv_module_2_sequential_5_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21995264)))]; - tensor encoder_conv_module_2_sequential_5_weight = const()[name = tensor("encoder_conv_module_2_sequential_5_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21996352)))]; - tensor encoder_conv_module_2_sequential_7_conv_bias = const()[name = tensor("encoder_conv_module_2_sequential_7_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21997440)))]; - tensor encoder_conv_module_2_sequential_7_conv_weight = const()[name = tensor("encoder_conv_module_2_sequential_7_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21998528)))]; - tensor encoder_ffn2_2_module_sequential_0_bias = const()[name = tensor("encoder_ffn2_2_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(22260736)))]; - tensor encoder_ffn2_2_module_sequential_0_weight = const()[name = tensor("encoder_ffn2_2_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(22261824)))]; - tensor encoder_ffn2_2_module_sequential_1_linear_bias = const()[name = tensor("encoder_ffn2_2_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(22262912)))]; - tensor encoder_ffn2_2_module_sequential_1_linear_weight = const()[name = tensor("encoder_ffn2_2_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(22267072)))]; - tensor encoder_ffn2_2_module_sequential_4_linear_bias = const()[name = tensor("encoder_ffn2_2_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(23315712)))]; - tensor encoder_ffn2_2_module_sequential_4_linear_weight = const()[name = tensor("encoder_ffn2_2_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(23316800)))]; - tensor encoder_layer_norm_2_bias = const()[name = tensor("encoder_layer_norm_2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(24365440)))]; - tensor encoder_layer_norm_2_weight = const()[name = tensor("encoder_layer_norm_2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(24366528)))]; - tensor encoder_ffn1_3_module_sequential_0_bias = const()[name = tensor("encoder_ffn1_3_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(24367616)))]; - tensor encoder_ffn1_3_module_sequential_0_weight = const()[name = tensor("encoder_ffn1_3_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(24368704)))]; - tensor encoder_ffn1_3_module_sequential_1_linear_bias = const()[name = tensor("encoder_ffn1_3_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(24369792)))]; - tensor encoder_ffn1_3_module_sequential_1_linear_weight = const()[name = tensor("encoder_ffn1_3_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(24373952)))]; - tensor encoder_ffn1_3_module_sequential_4_linear_bias = const()[name = tensor("encoder_ffn1_3_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(25422592)))]; - tensor encoder_ffn1_3_module_sequential_4_linear_weight = const()[name = tensor("encoder_ffn1_3_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(25423680)))]; - tensor encoder_ret_lns_3_bias = const()[name = tensor("encoder_ret_lns_3_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(26472320)))]; - tensor encoder_ret_lns_3_weight = const()[name = tensor("encoder_ret_lns_3_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(26473408)))]; - tensor encoder_q_proj_3_bias = const()[name = tensor("encoder_q_proj_3_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(26474496)))]; - tensor encoder_q_proj_3_weight = const()[name = tensor("encoder_q_proj_3_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(26475584)))]; - tensor encoder_k_proj_3_bias = const()[name = tensor("encoder_k_proj_3_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(26737792)))]; - tensor encoder_k_proj_3_weight = const()[name = tensor("encoder_k_proj_3_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(26738880)))]; - tensor encoder_v_proj_3_bias = const()[name = tensor("encoder_v_proj_3_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27001088)))]; - tensor encoder_v_proj_3_weight = const()[name = tensor("encoder_v_proj_3_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27002176)))]; - tensor encoder_g_proj_3_bias = const()[name = tensor("encoder_g_proj_3_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27264384)))]; - tensor encoder_g_proj_3_weight = const()[name = tensor("encoder_g_proj_3_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27265472)))]; - tensor encoder_out_proj_3_bias = const()[name = tensor("encoder_out_proj_3_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27527680)))]; - tensor encoder_out_proj_3_weight = const()[name = tensor("encoder_out_proj_3_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27528768)))]; - tensor encoder_conv_module_3_sequential_0_bias = const()[name = tensor("encoder_conv_module_3_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27790976)))]; - tensor encoder_conv_module_3_sequential_0_weight = const()[name = tensor("encoder_conv_module_3_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27792064)))]; - tensor encoder_conv_module_3_sequential_2_conv_bias = const()[name = tensor("encoder_conv_module_3_sequential_2_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27793152)))]; - tensor encoder_conv_module_3_sequential_2_conv_weight = const()[name = tensor("encoder_conv_module_3_sequential_2_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27795264)))]; - tensor encoder_conv_module_3_sequential_4_conv_weight = const()[name = tensor("encoder_conv_module_3_sequential_4_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28319616)))]; - tensor encoder_conv_module_3_sequential_5_running_var = const()[name = tensor("encoder_conv_module_3_sequential_5_running_var"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28336064)))]; - tensor encoder_conv_module_3_sequential_5_running_mean = const()[name = tensor("encoder_conv_module_3_sequential_5_running_mean"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28337152)))]; - tensor encoder_conv_module_3_sequential_5_bias = const()[name = tensor("encoder_conv_module_3_sequential_5_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28338240)))]; - tensor encoder_conv_module_3_sequential_5_weight = const()[name = tensor("encoder_conv_module_3_sequential_5_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28339328)))]; - tensor encoder_conv_module_3_sequential_7_conv_bias = const()[name = tensor("encoder_conv_module_3_sequential_7_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28340416)))]; - tensor encoder_conv_module_3_sequential_7_conv_weight = const()[name = tensor("encoder_conv_module_3_sequential_7_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28341504)))]; - tensor encoder_ffn2_3_module_sequential_0_bias = const()[name = tensor("encoder_ffn2_3_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28603712)))]; - tensor encoder_ffn2_3_module_sequential_0_weight = const()[name = tensor("encoder_ffn2_3_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28604800)))]; - tensor encoder_ffn2_3_module_sequential_1_linear_bias = const()[name = tensor("encoder_ffn2_3_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28605888)))]; - tensor encoder_ffn2_3_module_sequential_1_linear_weight = const()[name = tensor("encoder_ffn2_3_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28610048)))]; - tensor encoder_ffn2_3_module_sequential_4_linear_bias = const()[name = tensor("encoder_ffn2_3_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(29658688)))]; - tensor encoder_ffn2_3_module_sequential_4_linear_weight = const()[name = tensor("encoder_ffn2_3_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(29659776)))]; - tensor encoder_layer_norm_3_bias = const()[name = tensor("encoder_layer_norm_3_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(30708416)))]; - tensor encoder_layer_norm_3_weight = const()[name = tensor("encoder_layer_norm_3_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(30709504)))]; - tensor decoder__causal_mask = const()[name = tensor("decoder__causal_mask"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(30710592)))]; - tensor decoder_convert_bias = const()[name = tensor("decoder_convert_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(30710720)))]; - tensor decoder_convert_weight = const()[name = tensor("decoder_convert_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(30711808)))]; - tensor decoder_q_proj_0_bias = const()[name = tensor("decoder_q_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(31236160)))]; - tensor decoder_q_proj_0_weight = const()[name = tensor("decoder_q_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(31237248)))]; - tensor decoder_k_proj_0_bias = const()[name = tensor("decoder_k_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(31499456)))]; - tensor decoder_k_proj_0_weight = const()[name = tensor("decoder_k_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(31500544)))]; - tensor decoder_v_proj_0_bias = const()[name = tensor("decoder_v_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(31762752)))]; - tensor decoder_v_proj_0_weight = const()[name = tensor("decoder_v_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(31763840)))]; - tensor decoder_g_proj_0_bias = const()[name = tensor("decoder_g_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32026048)))]; - tensor decoder_g_proj_0_weight = const()[name = tensor("decoder_g_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32027136)))]; - tensor decoder_out_proj_0_bias = const()[name = tensor("decoder_out_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32289344)))]; - tensor decoder_out_proj_0_weight = const()[name = tensor("decoder_out_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32290432)))]; - tensor decoder_norm11_0_bias = const()[name = tensor("decoder_norm11_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32552640)))]; - tensor decoder_norm11_0_weight = const()[name = tensor("decoder_norm11_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32553728)))]; - tensor decoder_self_attn2_0_out_proj_bias = const()[name = tensor("decoder_self_attn2_0_out_proj_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32554816)))]; - tensor decoder_self_attn2_0_out_proj_weight = const()[name = tensor("decoder_self_attn2_0_out_proj_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32555904)))]; - tensor decoder_self_attn2_0_in_proj_bias = const()[name = tensor("decoder_self_attn2_0_in_proj_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32818112)))]; - tensor decoder_self_attn2_0_in_proj_weight = const()[name = tensor("decoder_self_attn2_0_in_proj_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32821248)))]; - tensor decoder_norm21_0_bias = const()[name = tensor("decoder_norm21_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(33607744)))]; - tensor decoder_norm21_0_weight = const()[name = tensor("decoder_norm21_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(33608832)))]; - tensor decoder_linear1_0_bias = const()[name = tensor("decoder_linear1_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(33609920)))]; - tensor decoder_linear1_0_weight = const()[name = tensor("decoder_linear1_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(33618176)))]; - tensor decoder_linear2_0_bias = const()[name = tensor("decoder_linear2_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(35715392)))]; - tensor decoder_linear2_0_weight = const()[name = tensor("decoder_linear2_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(35716480)))]; - tensor decoder_norm22_0_bias = const()[name = tensor("decoder_norm22_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(37813696)))]; - tensor decoder_norm22_0_weight = const()[name = tensor("decoder_norm22_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(37814784)))]; - tensor decoder_q_proj_1_bias = const()[name = tensor("decoder_q_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(37815872)))]; - tensor decoder_q_proj_1_weight = const()[name = tensor("decoder_q_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(37816960)))]; - tensor decoder_k_proj_1_bias = const()[name = tensor("decoder_k_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38079168)))]; - tensor decoder_k_proj_1_weight = const()[name = tensor("decoder_k_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38080256)))]; - tensor decoder_v_proj_1_bias = const()[name = tensor("decoder_v_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38342464)))]; - tensor decoder_v_proj_1_weight = const()[name = tensor("decoder_v_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38343552)))]; - tensor decoder_g_proj_1_bias = const()[name = tensor("decoder_g_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38605760)))]; - tensor decoder_g_proj_1_weight = const()[name = tensor("decoder_g_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38606848)))]; - tensor decoder_out_proj_1_bias = const()[name = tensor("decoder_out_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38869056)))]; - tensor decoder_out_proj_1_weight = const()[name = tensor("decoder_out_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38870144)))]; - tensor decoder_norm11_1_bias = const()[name = tensor("decoder_norm11_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(39132352)))]; - tensor decoder_norm11_1_weight = const()[name = tensor("decoder_norm11_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(39133440)))]; - tensor decoder_self_attn2_1_out_proj_bias = const()[name = tensor("decoder_self_attn2_1_out_proj_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(39134528)))]; - tensor decoder_self_attn2_1_out_proj_weight = const()[name = tensor("decoder_self_attn2_1_out_proj_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(39135616)))]; - tensor decoder_self_attn2_1_in_proj_bias = const()[name = tensor("decoder_self_attn2_1_in_proj_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(39397824)))]; - tensor decoder_self_attn2_1_in_proj_weight = const()[name = tensor("decoder_self_attn2_1_in_proj_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(39400960)))]; - tensor decoder_norm21_1_bias = const()[name = tensor("decoder_norm21_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(40187456)))]; - tensor decoder_norm21_1_weight = const()[name = tensor("decoder_norm21_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(40188544)))]; - tensor decoder_linear1_1_bias = const()[name = tensor("decoder_linear1_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(40189632)))]; - tensor decoder_linear1_1_weight = const()[name = tensor("decoder_linear1_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(40197888)))]; - tensor decoder_linear2_1_bias = const()[name = tensor("decoder_linear2_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(42295104)))]; - tensor decoder_linear2_1_weight = const()[name = tensor("decoder_linear2_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(42296192)))]; - tensor decoder_norm22_1_bias = const()[name = tensor("decoder_norm22_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(44393408)))]; - tensor decoder_norm22_1_weight = const()[name = tensor("decoder_norm22_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(44394496)))]; - tensor var_11 = const()[name = tensor("op_11"), val = tensor(0x1.5798eep-27)]; - tensor var_17 = const()[name = tensor("op_17"), val = tensor(0x1.0c6f7ap-20)]; - tensor var_20 = const()[name = tensor("op_20"), val = tensor(2)]; - tensor var_23 = const()[name = tensor("op_23"), val = tensor(0)]; - tensor var_26 = const()[name = tensor("op_26"), val = tensor(1)]; - tensor var_28 = const()[name = tensor("op_28"), val = tensor(0x1.4f8b58p-17)]; - tensor var_29 = const()[name = tensor("op_29"), val = tensor(true)]; - tensor input_1 = linear(bias = encoder_input_projection_linear_bias, weight = encoder_input_projection_linear_weight, x = features)[name = tensor("linear_0")]; - tensor input_3_axes_0 = const()[name = tensor("input_3_axes_0"), val = tensor([-1])]; - tensor input_3 = layer_norm(axes = input_3_axes_0, beta = encoder_pre_layer_norm_bias, epsilon = var_28, gamma = encoder_pre_layer_norm_weight, x = input_1)[name = tensor("input_3")]; + func main(tensor cnn_window, tensor dec_kv, tensor dec_scale, tensor enc_conv_cache, tensor enc_kv, tensor enc_scale, tensor features, tensor valid_mask) { + tensor inner_encoder_cnn_bias = const()[name = tensor("inner_encoder_cnn_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(64)))]; + tensor inner_encoder_cnn_weight = const()[name = tensor("inner_encoder_cnn_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1152)))]; + tensor inner_encoder__causal_mask = const()[name = tensor("inner_encoder__causal_mask"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4981952)))]; + tensor inner_encoder__t_index = const()[name = tensor("inner_encoder__t_index"), val = tensor([0x1p+0, 0x1p+1, 0x1.8p+1, 0x1p+2])]; + tensor inner_encoder_input_projection_linear_bias = const()[name = tensor("inner_encoder_input_projection_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4982080)))]; + tensor inner_encoder_input_projection_linear_weight = const()[name = tensor("inner_encoder_input_projection_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4983168)))]; + tensor inner_encoder_pre_layer_norm_bias = const()[name = tensor("inner_encoder_pre_layer_norm_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(5336512)))]; + tensor inner_encoder_pre_layer_norm_weight = const()[name = tensor("inner_encoder_pre_layer_norm_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(5337600)))]; + tensor inner_encoder_ffn1_0_module_sequential_0_bias = const()[name = tensor("inner_encoder_ffn1_0_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(5338688)))]; + tensor inner_encoder_ffn1_0_module_sequential_0_weight = const()[name = tensor("inner_encoder_ffn1_0_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(5339776)))]; + tensor inner_encoder_ffn1_0_module_sequential_1_linear_bias = const()[name = tensor("inner_encoder_ffn1_0_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(5340864)))]; + tensor inner_encoder_ffn1_0_module_sequential_1_linear_weight = const()[name = tensor("inner_encoder_ffn1_0_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(5345024)))]; + tensor inner_encoder_ffn1_0_module_sequential_4_linear_bias = const()[name = tensor("inner_encoder_ffn1_0_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(6393664)))]; + tensor inner_encoder_ffn1_0_module_sequential_4_linear_weight = const()[name = tensor("inner_encoder_ffn1_0_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(6394752)))]; + tensor inner_encoder_ret_lns_0_bias = const()[name = tensor("inner_encoder_ret_lns_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7443392)))]; + tensor inner_encoder_ret_lns_0_weight = const()[name = tensor("inner_encoder_ret_lns_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7444480)))]; + tensor inner_encoder_q_proj_0_bias = const()[name = tensor("inner_encoder_q_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7445568)))]; + tensor inner_encoder_q_proj_0_weight = const()[name = tensor("inner_encoder_q_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7446656)))]; + tensor inner_encoder_k_proj_0_bias = const()[name = tensor("inner_encoder_k_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7708864)))]; + tensor inner_encoder_k_proj_0_weight = const()[name = tensor("inner_encoder_k_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7709952)))]; + tensor inner_encoder_v_proj_0_bias = const()[name = tensor("inner_encoder_v_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7972160)))]; + tensor inner_encoder_v_proj_0_weight = const()[name = tensor("inner_encoder_v_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7973248)))]; + tensor inner_encoder_g_proj_0_bias = const()[name = tensor("inner_encoder_g_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8235456)))]; + tensor inner_encoder_g_proj_0_weight = const()[name = tensor("inner_encoder_g_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8236544)))]; + tensor inner_encoder_out_proj_0_bias = const()[name = tensor("inner_encoder_out_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8498752)))]; + tensor inner_encoder_out_proj_0_weight = const()[name = tensor("inner_encoder_out_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8499840)))]; + tensor inner_encoder_conv_module_0_sequential_0_bias = const()[name = tensor("inner_encoder_conv_module_0_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8762048)))]; + tensor inner_encoder_conv_module_0_sequential_0_weight = const()[name = tensor("inner_encoder_conv_module_0_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8763136)))]; + tensor inner_encoder_conv_module_0_sequential_2_conv_bias = const()[name = tensor("inner_encoder_conv_module_0_sequential_2_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8764224)))]; + tensor inner_encoder_conv_module_0_sequential_2_conv_weight = const()[name = tensor("inner_encoder_conv_module_0_sequential_2_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8766336)))]; + tensor inner_encoder_conv_module_0_sequential_4_conv_weight = const()[name = tensor("inner_encoder_conv_module_0_sequential_4_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9290688)))]; + tensor inner_encoder_conv_module_0_sequential_5_running_var = const()[name = tensor("inner_encoder_conv_module_0_sequential_5_running_var"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9307136)))]; + tensor inner_encoder_conv_module_0_sequential_5_running_mean = const()[name = tensor("inner_encoder_conv_module_0_sequential_5_running_mean"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9308224)))]; + tensor inner_encoder_conv_module_0_sequential_5_bias = const()[name = tensor("inner_encoder_conv_module_0_sequential_5_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9309312)))]; + tensor inner_encoder_conv_module_0_sequential_5_weight = const()[name = tensor("inner_encoder_conv_module_0_sequential_5_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9310400)))]; + tensor inner_encoder_conv_module_0_sequential_7_conv_bias = const()[name = tensor("inner_encoder_conv_module_0_sequential_7_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9311488)))]; + tensor inner_encoder_conv_module_0_sequential_7_conv_weight = const()[name = tensor("inner_encoder_conv_module_0_sequential_7_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9312576)))]; + tensor inner_encoder_ffn2_0_module_sequential_0_bias = const()[name = tensor("inner_encoder_ffn2_0_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9574784)))]; + tensor inner_encoder_ffn2_0_module_sequential_0_weight = const()[name = tensor("inner_encoder_ffn2_0_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9575872)))]; + tensor inner_encoder_ffn2_0_module_sequential_1_linear_bias = const()[name = tensor("inner_encoder_ffn2_0_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9576960)))]; + tensor inner_encoder_ffn2_0_module_sequential_1_linear_weight = const()[name = tensor("inner_encoder_ffn2_0_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9581120)))]; + tensor inner_encoder_ffn2_0_module_sequential_4_linear_bias = const()[name = tensor("inner_encoder_ffn2_0_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(10629760)))]; + tensor inner_encoder_ffn2_0_module_sequential_4_linear_weight = const()[name = tensor("inner_encoder_ffn2_0_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(10630848)))]; + tensor inner_encoder_layer_norm_0_bias = const()[name = tensor("inner_encoder_layer_norm_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11679488)))]; + tensor inner_encoder_layer_norm_0_weight = const()[name = tensor("inner_encoder_layer_norm_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11680576)))]; + tensor inner_encoder_ffn1_1_module_sequential_0_bias = const()[name = tensor("inner_encoder_ffn1_1_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11681664)))]; + tensor inner_encoder_ffn1_1_module_sequential_0_weight = const()[name = tensor("inner_encoder_ffn1_1_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11682752)))]; + tensor inner_encoder_ffn1_1_module_sequential_1_linear_bias = const()[name = tensor("inner_encoder_ffn1_1_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11683840)))]; + tensor inner_encoder_ffn1_1_module_sequential_1_linear_weight = const()[name = tensor("inner_encoder_ffn1_1_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11688000)))]; + tensor inner_encoder_ffn1_1_module_sequential_4_linear_bias = const()[name = tensor("inner_encoder_ffn1_1_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(12736640)))]; + tensor inner_encoder_ffn1_1_module_sequential_4_linear_weight = const()[name = tensor("inner_encoder_ffn1_1_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(12737728)))]; + tensor inner_encoder_ret_lns_1_bias = const()[name = tensor("inner_encoder_ret_lns_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(13786368)))]; + tensor inner_encoder_ret_lns_1_weight = const()[name = tensor("inner_encoder_ret_lns_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(13787456)))]; + tensor inner_encoder_q_proj_1_bias = const()[name = tensor("inner_encoder_q_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(13788544)))]; + tensor inner_encoder_q_proj_1_weight = const()[name = tensor("inner_encoder_q_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(13789632)))]; + tensor inner_encoder_k_proj_1_bias = const()[name = tensor("inner_encoder_k_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14051840)))]; + tensor inner_encoder_k_proj_1_weight = const()[name = tensor("inner_encoder_k_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14052928)))]; + tensor inner_encoder_v_proj_1_bias = const()[name = tensor("inner_encoder_v_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14315136)))]; + tensor inner_encoder_v_proj_1_weight = const()[name = tensor("inner_encoder_v_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14316224)))]; + tensor inner_encoder_g_proj_1_bias = const()[name = tensor("inner_encoder_g_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14578432)))]; + tensor inner_encoder_g_proj_1_weight = const()[name = tensor("inner_encoder_g_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14579520)))]; + tensor inner_encoder_out_proj_1_bias = const()[name = tensor("inner_encoder_out_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14841728)))]; + tensor inner_encoder_out_proj_1_weight = const()[name = tensor("inner_encoder_out_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14842816)))]; + tensor inner_encoder_conv_module_1_sequential_0_bias = const()[name = tensor("inner_encoder_conv_module_1_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15105024)))]; + tensor inner_encoder_conv_module_1_sequential_0_weight = const()[name = tensor("inner_encoder_conv_module_1_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15106112)))]; + tensor inner_encoder_conv_module_1_sequential_2_conv_bias = const()[name = tensor("inner_encoder_conv_module_1_sequential_2_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15107200)))]; + tensor inner_encoder_conv_module_1_sequential_2_conv_weight = const()[name = tensor("inner_encoder_conv_module_1_sequential_2_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15109312)))]; + tensor inner_encoder_conv_module_1_sequential_4_conv_weight = const()[name = tensor("inner_encoder_conv_module_1_sequential_4_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15633664)))]; + tensor inner_encoder_conv_module_1_sequential_5_running_var = const()[name = tensor("inner_encoder_conv_module_1_sequential_5_running_var"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15650112)))]; + tensor inner_encoder_conv_module_1_sequential_5_running_mean = const()[name = tensor("inner_encoder_conv_module_1_sequential_5_running_mean"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15651200)))]; + tensor inner_encoder_conv_module_1_sequential_5_bias = const()[name = tensor("inner_encoder_conv_module_1_sequential_5_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15652288)))]; + tensor inner_encoder_conv_module_1_sequential_5_weight = const()[name = tensor("inner_encoder_conv_module_1_sequential_5_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15653376)))]; + tensor inner_encoder_conv_module_1_sequential_7_conv_bias = const()[name = tensor("inner_encoder_conv_module_1_sequential_7_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15654464)))]; + tensor inner_encoder_conv_module_1_sequential_7_conv_weight = const()[name = tensor("inner_encoder_conv_module_1_sequential_7_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15655552)))]; + tensor inner_encoder_ffn2_1_module_sequential_0_bias = const()[name = tensor("inner_encoder_ffn2_1_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15917760)))]; + tensor inner_encoder_ffn2_1_module_sequential_0_weight = const()[name = tensor("inner_encoder_ffn2_1_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15918848)))]; + tensor inner_encoder_ffn2_1_module_sequential_1_linear_bias = const()[name = tensor("inner_encoder_ffn2_1_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15919936)))]; + tensor inner_encoder_ffn2_1_module_sequential_1_linear_weight = const()[name = tensor("inner_encoder_ffn2_1_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15924096)))]; + tensor inner_encoder_ffn2_1_module_sequential_4_linear_bias = const()[name = tensor("inner_encoder_ffn2_1_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(16972736)))]; + tensor inner_encoder_ffn2_1_module_sequential_4_linear_weight = const()[name = tensor("inner_encoder_ffn2_1_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(16973824)))]; + tensor inner_encoder_layer_norm_1_bias = const()[name = tensor("inner_encoder_layer_norm_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18022464)))]; + tensor inner_encoder_layer_norm_1_weight = const()[name = tensor("inner_encoder_layer_norm_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18023552)))]; + tensor inner_encoder_ffn1_2_module_sequential_0_bias = const()[name = tensor("inner_encoder_ffn1_2_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18024640)))]; + tensor inner_encoder_ffn1_2_module_sequential_0_weight = const()[name = tensor("inner_encoder_ffn1_2_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18025728)))]; + tensor inner_encoder_ffn1_2_module_sequential_1_linear_bias = const()[name = tensor("inner_encoder_ffn1_2_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18026816)))]; + tensor inner_encoder_ffn1_2_module_sequential_1_linear_weight = const()[name = tensor("inner_encoder_ffn1_2_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18030976)))]; + tensor inner_encoder_ffn1_2_module_sequential_4_linear_bias = const()[name = tensor("inner_encoder_ffn1_2_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(19079616)))]; + tensor inner_encoder_ffn1_2_module_sequential_4_linear_weight = const()[name = tensor("inner_encoder_ffn1_2_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(19080704)))]; + tensor inner_encoder_ret_lns_2_bias = const()[name = tensor("inner_encoder_ret_lns_2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20129344)))]; + tensor inner_encoder_ret_lns_2_weight = const()[name = tensor("inner_encoder_ret_lns_2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20130432)))]; + tensor inner_encoder_q_proj_2_bias = const()[name = tensor("inner_encoder_q_proj_2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20131520)))]; + tensor inner_encoder_q_proj_2_weight = const()[name = tensor("inner_encoder_q_proj_2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20132608)))]; + tensor inner_encoder_k_proj_2_bias = const()[name = tensor("inner_encoder_k_proj_2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20394816)))]; + tensor inner_encoder_k_proj_2_weight = const()[name = tensor("inner_encoder_k_proj_2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20395904)))]; + tensor inner_encoder_v_proj_2_bias = const()[name = tensor("inner_encoder_v_proj_2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20658112)))]; + tensor inner_encoder_v_proj_2_weight = const()[name = tensor("inner_encoder_v_proj_2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20659200)))]; + tensor inner_encoder_g_proj_2_bias = const()[name = tensor("inner_encoder_g_proj_2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20921408)))]; + tensor inner_encoder_g_proj_2_weight = const()[name = tensor("inner_encoder_g_proj_2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20922496)))]; + tensor inner_encoder_out_proj_2_bias = const()[name = tensor("inner_encoder_out_proj_2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21184704)))]; + tensor inner_encoder_out_proj_2_weight = const()[name = tensor("inner_encoder_out_proj_2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21185792)))]; + tensor inner_encoder_conv_module_2_sequential_0_bias = const()[name = tensor("inner_encoder_conv_module_2_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21448000)))]; + tensor inner_encoder_conv_module_2_sequential_0_weight = const()[name = tensor("inner_encoder_conv_module_2_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21449088)))]; + tensor inner_encoder_conv_module_2_sequential_2_conv_bias = const()[name = tensor("inner_encoder_conv_module_2_sequential_2_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21450176)))]; + tensor inner_encoder_conv_module_2_sequential_2_conv_weight = const()[name = tensor("inner_encoder_conv_module_2_sequential_2_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21452288)))]; + tensor inner_encoder_conv_module_2_sequential_4_conv_weight = const()[name = tensor("inner_encoder_conv_module_2_sequential_4_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21976640)))]; + tensor inner_encoder_conv_module_2_sequential_5_running_var = const()[name = tensor("inner_encoder_conv_module_2_sequential_5_running_var"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21993088)))]; + tensor inner_encoder_conv_module_2_sequential_5_running_mean = const()[name = tensor("inner_encoder_conv_module_2_sequential_5_running_mean"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21994176)))]; + tensor inner_encoder_conv_module_2_sequential_5_bias = const()[name = tensor("inner_encoder_conv_module_2_sequential_5_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21995264)))]; + tensor inner_encoder_conv_module_2_sequential_5_weight = const()[name = tensor("inner_encoder_conv_module_2_sequential_5_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21996352)))]; + tensor inner_encoder_conv_module_2_sequential_7_conv_bias = const()[name = tensor("inner_encoder_conv_module_2_sequential_7_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21997440)))]; + tensor inner_encoder_conv_module_2_sequential_7_conv_weight = const()[name = tensor("inner_encoder_conv_module_2_sequential_7_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21998528)))]; + tensor inner_encoder_ffn2_2_module_sequential_0_bias = const()[name = tensor("inner_encoder_ffn2_2_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(22260736)))]; + tensor inner_encoder_ffn2_2_module_sequential_0_weight = const()[name = tensor("inner_encoder_ffn2_2_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(22261824)))]; + tensor inner_encoder_ffn2_2_module_sequential_1_linear_bias = const()[name = tensor("inner_encoder_ffn2_2_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(22262912)))]; + tensor inner_encoder_ffn2_2_module_sequential_1_linear_weight = const()[name = tensor("inner_encoder_ffn2_2_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(22267072)))]; + tensor inner_encoder_ffn2_2_module_sequential_4_linear_bias = const()[name = tensor("inner_encoder_ffn2_2_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(23315712)))]; + tensor inner_encoder_ffn2_2_module_sequential_4_linear_weight = const()[name = tensor("inner_encoder_ffn2_2_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(23316800)))]; + tensor inner_encoder_layer_norm_2_bias = const()[name = tensor("inner_encoder_layer_norm_2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(24365440)))]; + tensor inner_encoder_layer_norm_2_weight = const()[name = tensor("inner_encoder_layer_norm_2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(24366528)))]; + tensor inner_encoder_ffn1_3_module_sequential_0_bias = const()[name = tensor("inner_encoder_ffn1_3_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(24367616)))]; + tensor inner_encoder_ffn1_3_module_sequential_0_weight = const()[name = tensor("inner_encoder_ffn1_3_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(24368704)))]; + tensor inner_encoder_ffn1_3_module_sequential_1_linear_bias = const()[name = tensor("inner_encoder_ffn1_3_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(24369792)))]; + tensor inner_encoder_ffn1_3_module_sequential_1_linear_weight = const()[name = tensor("inner_encoder_ffn1_3_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(24373952)))]; + tensor inner_encoder_ffn1_3_module_sequential_4_linear_bias = const()[name = tensor("inner_encoder_ffn1_3_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(25422592)))]; + tensor inner_encoder_ffn1_3_module_sequential_4_linear_weight = const()[name = tensor("inner_encoder_ffn1_3_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(25423680)))]; + tensor inner_encoder_ret_lns_3_bias = const()[name = tensor("inner_encoder_ret_lns_3_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(26472320)))]; + tensor inner_encoder_ret_lns_3_weight = const()[name = tensor("inner_encoder_ret_lns_3_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(26473408)))]; + tensor inner_encoder_q_proj_3_bias = const()[name = tensor("inner_encoder_q_proj_3_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(26474496)))]; + tensor inner_encoder_q_proj_3_weight = const()[name = tensor("inner_encoder_q_proj_3_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(26475584)))]; + tensor inner_encoder_k_proj_3_bias = const()[name = tensor("inner_encoder_k_proj_3_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(26737792)))]; + tensor inner_encoder_k_proj_3_weight = const()[name = tensor("inner_encoder_k_proj_3_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(26738880)))]; + tensor inner_encoder_v_proj_3_bias = const()[name = tensor("inner_encoder_v_proj_3_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27001088)))]; + tensor inner_encoder_v_proj_3_weight = const()[name = tensor("inner_encoder_v_proj_3_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27002176)))]; + tensor inner_encoder_g_proj_3_bias = const()[name = tensor("inner_encoder_g_proj_3_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27264384)))]; + tensor inner_encoder_g_proj_3_weight = const()[name = tensor("inner_encoder_g_proj_3_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27265472)))]; + tensor inner_encoder_out_proj_3_bias = const()[name = tensor("inner_encoder_out_proj_3_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27527680)))]; + tensor inner_encoder_out_proj_3_weight = const()[name = tensor("inner_encoder_out_proj_3_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27528768)))]; + tensor inner_encoder_conv_module_3_sequential_0_bias = const()[name = tensor("inner_encoder_conv_module_3_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27790976)))]; + tensor inner_encoder_conv_module_3_sequential_0_weight = const()[name = tensor("inner_encoder_conv_module_3_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27792064)))]; + tensor inner_encoder_conv_module_3_sequential_2_conv_bias = const()[name = tensor("inner_encoder_conv_module_3_sequential_2_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27793152)))]; + tensor inner_encoder_conv_module_3_sequential_2_conv_weight = const()[name = tensor("inner_encoder_conv_module_3_sequential_2_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27795264)))]; + tensor inner_encoder_conv_module_3_sequential_4_conv_weight = const()[name = tensor("inner_encoder_conv_module_3_sequential_4_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28319616)))]; + tensor inner_encoder_conv_module_3_sequential_5_running_var = const()[name = tensor("inner_encoder_conv_module_3_sequential_5_running_var"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28336064)))]; + tensor inner_encoder_conv_module_3_sequential_5_running_mean = const()[name = tensor("inner_encoder_conv_module_3_sequential_5_running_mean"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28337152)))]; + tensor inner_encoder_conv_module_3_sequential_5_bias = const()[name = tensor("inner_encoder_conv_module_3_sequential_5_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28338240)))]; + tensor inner_encoder_conv_module_3_sequential_5_weight = const()[name = tensor("inner_encoder_conv_module_3_sequential_5_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28339328)))]; + tensor inner_encoder_conv_module_3_sequential_7_conv_bias = const()[name = tensor("inner_encoder_conv_module_3_sequential_7_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28340416)))]; + tensor inner_encoder_conv_module_3_sequential_7_conv_weight = const()[name = tensor("inner_encoder_conv_module_3_sequential_7_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28341504)))]; + tensor inner_encoder_ffn2_3_module_sequential_0_bias = const()[name = tensor("inner_encoder_ffn2_3_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28603712)))]; + tensor inner_encoder_ffn2_3_module_sequential_0_weight = const()[name = tensor("inner_encoder_ffn2_3_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28604800)))]; + tensor inner_encoder_ffn2_3_module_sequential_1_linear_bias = const()[name = tensor("inner_encoder_ffn2_3_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28605888)))]; + tensor inner_encoder_ffn2_3_module_sequential_1_linear_weight = const()[name = tensor("inner_encoder_ffn2_3_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28610048)))]; + tensor inner_encoder_ffn2_3_module_sequential_4_linear_bias = const()[name = tensor("inner_encoder_ffn2_3_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(29658688)))]; + tensor inner_encoder_ffn2_3_module_sequential_4_linear_weight = const()[name = tensor("inner_encoder_ffn2_3_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(29659776)))]; + tensor inner_encoder_layer_norm_3_bias = const()[name = tensor("inner_encoder_layer_norm_3_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(30708416)))]; + tensor inner_encoder_layer_norm_3_weight = const()[name = tensor("inner_encoder_layer_norm_3_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(30709504)))]; + tensor inner_decoder__causal_mask = const()[name = tensor("inner_decoder__causal_mask"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(30710592)))]; + tensor inner_decoder_convert_bias = const()[name = tensor("inner_decoder_convert_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(30710720)))]; + tensor inner_decoder_convert_weight = const()[name = tensor("inner_decoder_convert_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(30711808)))]; + tensor inner_decoder_q_proj_0_bias = const()[name = tensor("inner_decoder_q_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(31236160)))]; + tensor inner_decoder_q_proj_0_weight = const()[name = tensor("inner_decoder_q_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(31237248)))]; + tensor inner_decoder_k_proj_0_bias = const()[name = tensor("inner_decoder_k_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(31499456)))]; + tensor inner_decoder_k_proj_0_weight = const()[name = tensor("inner_decoder_k_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(31500544)))]; + tensor inner_decoder_v_proj_0_bias = const()[name = tensor("inner_decoder_v_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(31762752)))]; + tensor inner_decoder_v_proj_0_weight = const()[name = tensor("inner_decoder_v_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(31763840)))]; + tensor inner_decoder_g_proj_0_bias = const()[name = tensor("inner_decoder_g_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32026048)))]; + tensor inner_decoder_g_proj_0_weight = const()[name = tensor("inner_decoder_g_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32027136)))]; + tensor inner_decoder_out_proj_0_bias = const()[name = tensor("inner_decoder_out_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32289344)))]; + tensor inner_decoder_out_proj_0_weight = const()[name = tensor("inner_decoder_out_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32290432)))]; + tensor inner_decoder_norm11_0_bias = const()[name = tensor("inner_decoder_norm11_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32552640)))]; + tensor inner_decoder_norm11_0_weight = const()[name = tensor("inner_decoder_norm11_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32553728)))]; + tensor inner_decoder_self_attn2_0_out_proj_bias = const()[name = tensor("inner_decoder_self_attn2_0_out_proj_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32554816)))]; + tensor inner_decoder_self_attn2_0_out_proj_weight = const()[name = tensor("inner_decoder_self_attn2_0_out_proj_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32555904)))]; + tensor inner_decoder_self_attn2_0_in_proj_bias = const()[name = tensor("inner_decoder_self_attn2_0_in_proj_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32818112)))]; + tensor inner_decoder_self_attn2_0_in_proj_weight = const()[name = tensor("inner_decoder_self_attn2_0_in_proj_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32821248)))]; + tensor inner_decoder_norm21_0_bias = const()[name = tensor("inner_decoder_norm21_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(33607744)))]; + tensor inner_decoder_norm21_0_weight = const()[name = tensor("inner_decoder_norm21_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(33608832)))]; + tensor inner_decoder_linear1_0_bias = const()[name = tensor("inner_decoder_linear1_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(33609920)))]; + tensor inner_decoder_linear1_0_weight = const()[name = tensor("inner_decoder_linear1_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(33618176)))]; + tensor inner_decoder_linear2_0_bias = const()[name = tensor("inner_decoder_linear2_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(35715392)))]; + tensor inner_decoder_linear2_0_weight = const()[name = tensor("inner_decoder_linear2_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(35716480)))]; + tensor inner_decoder_norm22_0_bias = const()[name = tensor("inner_decoder_norm22_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(37813696)))]; + tensor inner_decoder_norm22_0_weight = const()[name = tensor("inner_decoder_norm22_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(37814784)))]; + tensor inner_decoder_q_proj_1_bias = const()[name = tensor("inner_decoder_q_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(37815872)))]; + tensor inner_decoder_q_proj_1_weight = const()[name = tensor("inner_decoder_q_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(37816960)))]; + tensor inner_decoder_k_proj_1_bias = const()[name = tensor("inner_decoder_k_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38079168)))]; + tensor inner_decoder_k_proj_1_weight = const()[name = tensor("inner_decoder_k_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38080256)))]; + tensor inner_decoder_v_proj_1_bias = const()[name = tensor("inner_decoder_v_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38342464)))]; + tensor inner_decoder_v_proj_1_weight = const()[name = tensor("inner_decoder_v_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38343552)))]; + tensor inner_decoder_g_proj_1_bias = const()[name = tensor("inner_decoder_g_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38605760)))]; + tensor inner_decoder_g_proj_1_weight = const()[name = tensor("inner_decoder_g_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38606848)))]; + tensor inner_decoder_out_proj_1_bias = const()[name = tensor("inner_decoder_out_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38869056)))]; + tensor inner_decoder_out_proj_1_weight = const()[name = tensor("inner_decoder_out_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38870144)))]; + tensor inner_decoder_norm11_1_bias = const()[name = tensor("inner_decoder_norm11_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(39132352)))]; + tensor inner_decoder_norm11_1_weight = const()[name = tensor("inner_decoder_norm11_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(39133440)))]; + tensor inner_decoder_self_attn2_1_out_proj_bias = const()[name = tensor("inner_decoder_self_attn2_1_out_proj_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(39134528)))]; + tensor inner_decoder_self_attn2_1_out_proj_weight = const()[name = tensor("inner_decoder_self_attn2_1_out_proj_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(39135616)))]; + tensor inner_decoder_self_attn2_1_in_proj_bias = const()[name = tensor("inner_decoder_self_attn2_1_in_proj_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(39397824)))]; + tensor inner_decoder_self_attn2_1_in_proj_weight = const()[name = tensor("inner_decoder_self_attn2_1_in_proj_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(39400960)))]; + tensor inner_decoder_norm21_1_bias = const()[name = tensor("inner_decoder_norm21_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(40187456)))]; + tensor inner_decoder_norm21_1_weight = const()[name = tensor("inner_decoder_norm21_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(40188544)))]; + tensor inner_decoder_linear1_1_bias = const()[name = tensor("inner_decoder_linear1_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(40189632)))]; + tensor inner_decoder_linear1_1_weight = const()[name = tensor("inner_decoder_linear1_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(40197888)))]; + tensor inner_decoder_linear2_1_bias = const()[name = tensor("inner_decoder_linear2_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(42295104)))]; + tensor inner_decoder_linear2_1_weight = const()[name = tensor("inner_decoder_linear2_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(42296192)))]; + tensor inner_decoder_norm22_1_bias = const()[name = tensor("inner_decoder_norm22_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(44393408)))]; + tensor inner_decoder_norm22_1_weight = const()[name = tensor("inner_decoder_norm22_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(44394496)))]; + tensor var_19_begin_0 = const()[name = tensor("op_19_begin_0"), val = tensor([0, 0, 0])]; + tensor var_19_end_0 = const()[name = tensor("op_19_end_0"), val = tensor([1, 15, 23])]; + tensor var_19_end_mask_0 = const()[name = tensor("op_19_end_mask_0"), val = tensor([true, false, true])]; + tensor var_19 = slice_by_index(begin = var_19_begin_0, end = var_19_end_0, end_mask = var_19_end_mask_0, x = features)[name = tensor("op_19")]; + tensor var_29_begin_0 = const()[name = tensor("op_29_begin_0"), val = tensor([0, 10, 0])]; + tensor var_29_end_0 = const()[name = tensor("op_29_end_0"), val = tensor([1, 25, 23])]; + tensor var_29_end_mask_0 = const()[name = tensor("op_29_end_mask_0"), val = tensor([true, false, true])]; + tensor var_29 = slice_by_index(begin = var_29_begin_0, end = var_29_end_0, end_mask = var_29_end_mask_0, x = features)[name = tensor("op_29")]; + tensor var_39_begin_0 = const()[name = tensor("op_39_begin_0"), val = tensor([0, 20, 0])]; + tensor var_39_end_0 = const()[name = tensor("op_39_end_0"), val = tensor([1, 35, 23])]; + tensor var_39_end_mask_0 = const()[name = tensor("op_39_end_mask_0"), val = tensor([true, false, true])]; + tensor var_39 = slice_by_index(begin = var_39_begin_0, end = var_39_end_0, end_mask = var_39_end_mask_0, x = features)[name = tensor("op_39")]; + tensor var_49_begin_0 = const()[name = tensor("op_49_begin_0"), val = tensor([0, 30, 0])]; + tensor var_49_end_0 = const()[name = tensor("op_49_end_0"), val = tensor([1, 1, 23])]; + tensor var_49_end_mask_0 = const()[name = tensor("op_49_end_mask_0"), val = tensor([true, true, true])]; + tensor var_49 = slice_by_index(begin = var_49_begin_0, end = var_49_end_0, end_mask = var_49_end_mask_0, x = features)[name = tensor("op_49")]; + tensor stacked_axis_0 = const()[name = tensor("stacked_axis_0"), val = tensor(1)]; + tensor stacked = stack(axis = stacked_axis_0, values = (var_19, var_29, var_39, var_49))[name = tensor("stacked")]; + tensor var_56 = const()[name = tensor("op_56"), val = tensor([1, 4, 345])]; + tensor input_1 = reshape(shape = var_56, x = stacked)[name = tensor("input_1")]; + tensor var_58 = const()[name = tensor("op_58"), val = tensor(0x1p+0)]; + tensor var_64 = const()[name = tensor("op_64"), val = tensor(true)]; + tensor var_65 = const()[name = tensor("op_65"), val = tensor(0x1.4f8b58p-17)]; + tensor var_68 = const()[name = tensor("op_68"), val = tensor(0)]; + tensor var_70 = const()[name = tensor("op_70"), val = tensor(2)]; + tensor var_71 = const()[name = tensor("op_71"), val = tensor(-1)]; + tensor var_73 = const()[name = tensor("op_73"), val = tensor(0x1.0c6f7ap-20)]; + tensor var_78 = const()[name = tensor("op_78"), val = tensor(0x1.5798eep-27)]; + tensor var_82 = const()[name = tensor("op_82"), val = tensor(1)]; + tensor input_3 = linear(bias = inner_encoder_input_projection_linear_bias, weight = inner_encoder_input_projection_linear_weight, x = input_1)[name = tensor("linear_0")]; tensor input_5_axes_0 = const()[name = tensor("input_5_axes_0"), val = tensor([-1])]; - tensor input_5 = layer_norm(axes = input_5_axes_0, beta = encoder_ffn1_0_module_sequential_0_bias, epsilon = var_28, gamma = encoder_ffn1_0_module_sequential_0_weight, x = input_3)[name = tensor("input_5")]; - tensor inputs_1 = linear(bias = encoder_ffn1_0_module_sequential_1_linear_bias, weight = encoder_ffn1_0_module_sequential_1_linear_weight, x = input_5)[name = tensor("linear_1")]; - tensor input_7 = silu(x = inputs_1)[name = tensor("input_7")]; - tensor input_11 = linear(bias = encoder_ffn1_0_module_sequential_4_linear_bias, weight = encoder_ffn1_0_module_sequential_4_linear_weight, x = input_7)[name = tensor("linear_2")]; - tensor var_147 = const()[name = tensor("op_147"), val = tensor(0x1p-1)]; - tensor var_148 = mul(x = input_11, y = var_147)[name = tensor("op_148")]; - tensor input_13 = add(x = var_148, y = input_3)[name = tensor("input_13")]; + tensor input_5 = layer_norm(axes = input_5_axes_0, beta = inner_encoder_pre_layer_norm_bias, epsilon = var_65, gamma = inner_encoder_pre_layer_norm_weight, x = input_3)[name = tensor("input_5")]; + tensor input_7_axes_0 = const()[name = tensor("input_7_axes_0"), val = tensor([-1])]; + tensor input_7 = layer_norm(axes = input_7_axes_0, beta = inner_encoder_ffn1_0_module_sequential_0_bias, epsilon = var_65, gamma = inner_encoder_ffn1_0_module_sequential_0_weight, x = input_5)[name = tensor("input_7")]; + tensor inputs_1 = linear(bias = inner_encoder_ffn1_0_module_sequential_1_linear_bias, weight = inner_encoder_ffn1_0_module_sequential_1_linear_weight, x = input_7)[name = tensor("linear_1")]; + tensor input_9 = silu(x = inputs_1)[name = tensor("input_9")]; + tensor input_13 = linear(bias = inner_encoder_ffn1_0_module_sequential_4_linear_bias, weight = inner_encoder_ffn1_0_module_sequential_4_linear_weight, x = input_9)[name = tensor("linear_2")]; + tensor var_203 = const()[name = tensor("op_203"), val = tensor(0x1p-1)]; + tensor var_204 = mul(x = input_13, y = var_203)[name = tensor("op_204")]; + tensor input_15 = add(x = var_204, y = input_5)[name = tensor("input_15")]; tensor x_1_axes_0 = const()[name = tensor("x_1_axes_0"), val = tensor([-1])]; - tensor x_1 = layer_norm(axes = x_1_axes_0, beta = encoder_ret_lns_0_bias, epsilon = var_28, gamma = encoder_ret_lns_0_weight, x = input_13)[name = tensor("x_1")]; + tensor x_1 = layer_norm(axes = x_1_axes_0, beta = inner_encoder_ret_lns_0_bias, epsilon = var_65, gamma = inner_encoder_ret_lns_0_weight, x = input_15)[name = tensor("x_1")]; tensor prev_kv_1_begin_0 = const()[name = tensor("prev_kv_1_begin_0"), val = tensor([0, 0, 0, 0, 0])]; tensor prev_kv_1_end_0 = const()[name = tensor("prev_kv_1_end_0"), val = tensor([1, 1, 4, 64, 64])]; tensor prev_kv_1_end_mask_0 = const()[name = tensor("prev_kv_1_end_mask_0"), val = tensor([false, true, true, true, true])]; @@ -239,173 +261,173 @@ program(1.0) tensor prev_scale_1_end_mask_0 = const()[name = tensor("prev_scale_1_end_mask_0"), val = tensor([false, true])]; tensor prev_scale_1_squeeze_mask_0 = const()[name = tensor("prev_scale_1_squeeze_mask_0"), val = tensor([true, false])]; tensor prev_scale_1 = slice_by_index(begin = prev_scale_1_begin_0, end = prev_scale_1_end_0, end_mask = prev_scale_1_end_mask_0, squeeze_mask = prev_scale_1_squeeze_mask_0, x = enc_scale)[name = tensor("prev_scale_1")]; - tensor var_162 = linear(bias = encoder_q_proj_0_bias, weight = encoder_q_proj_0_weight, x = x_1)[name = tensor("linear_3")]; - tensor var_163 = const()[name = tensor("op_163"), val = tensor([1, 4, 4, 64])]; - tensor var_164 = reshape(shape = var_163, x = var_162)[name = tensor("op_164")]; + tensor var_218 = linear(bias = inner_encoder_q_proj_0_bias, weight = inner_encoder_q_proj_0_weight, x = x_1)[name = tensor("linear_3")]; + tensor var_219 = const()[name = tensor("op_219"), val = tensor([1, 4, 4, 64])]; + tensor var_220 = reshape(shape = var_219, x = var_218)[name = tensor("op_220")]; tensor q_1_perm_0 = const()[name = tensor("q_1_perm_0"), val = tensor([0, 2, 1, 3])]; - tensor var_168 = linear(bias = encoder_k_proj_0_bias, weight = encoder_k_proj_0_weight, x = x_1)[name = tensor("linear_4")]; - tensor var_169 = const()[name = tensor("op_169"), val = tensor(0x1p-3)]; - tensor var_170 = mul(x = var_168, y = var_169)[name = tensor("op_170")]; - tensor var_171 = const()[name = tensor("op_171"), val = tensor([1, 4, 4, 64])]; - tensor var_172 = reshape(shape = var_171, x = var_170)[name = tensor("op_172")]; + tensor var_224 = linear(bias = inner_encoder_k_proj_0_bias, weight = inner_encoder_k_proj_0_weight, x = x_1)[name = tensor("linear_4")]; + tensor var_225 = const()[name = tensor("op_225"), val = tensor(0x1p-3)]; + tensor var_226 = mul(x = var_224, y = var_225)[name = tensor("op_226")]; + tensor var_227 = const()[name = tensor("op_227"), val = tensor([1, 4, 4, 64])]; + tensor var_228 = reshape(shape = var_227, x = var_226)[name = tensor("op_228")]; tensor k_1_perm_0 = const()[name = tensor("k_1_perm_0"), val = tensor([0, 2, 1, 3])]; - tensor var_176 = linear(bias = encoder_v_proj_0_bias, weight = encoder_v_proj_0_weight, x = x_1)[name = tensor("linear_5")]; - tensor var_177 = const()[name = tensor("op_177"), val = tensor([1, 4, 4, 64])]; - tensor var_178 = reshape(shape = var_177, x = var_176)[name = tensor("op_178")]; + tensor var_232 = linear(bias = inner_encoder_v_proj_0_bias, weight = inner_encoder_v_proj_0_weight, x = x_1)[name = tensor("linear_5")]; + tensor var_233 = const()[name = tensor("op_233"), val = tensor([1, 4, 4, 64])]; + tensor var_234 = reshape(shape = var_233, x = var_232)[name = tensor("op_234")]; tensor v_1_perm_0 = const()[name = tensor("v_1_perm_0"), val = tensor([0, 2, 1, 3])]; - tensor input_17 = linear(bias = encoder_g_proj_0_bias, weight = encoder_g_proj_0_weight, x = x_1)[name = tensor("linear_6")]; + tensor input_19 = linear(bias = inner_encoder_g_proj_0_bias, weight = inner_encoder_g_proj_0_weight, x = x_1)[name = tensor("linear_6")]; tensor sqrt_s0_1 = sqrt(x = prev_scale_1)[name = tensor("sqrt_s0_1")]; - tensor s_t_1 = add(x = prev_scale_1, y = encoder__t_index)[name = tensor("s_t_1")]; + tensor s_t_1 = add(x = prev_scale_1, y = inner_encoder__t_index)[name = tensor("s_t_1")]; tensor sqrt_s_t_1 = sqrt(x = s_t_1)[name = tensor("sqrt_s_t_1")]; tensor qk_1_transpose_x_1 = const()[name = tensor("qk_1_transpose_x_1"), val = tensor(false)]; tensor qk_1_transpose_y_1 = const()[name = tensor("qk_1_transpose_y_1"), val = tensor(true)]; - tensor k_1 = transpose(perm = k_1_perm_0, x = var_172)[name = tensor("transpose_57")]; - tensor q_1 = transpose(perm = q_1_perm_0, x = var_164)[name = tensor("transpose_58")]; + tensor k_1 = transpose(perm = k_1_perm_0, x = var_228)[name = tensor("transpose_57")]; + tensor q_1 = transpose(perm = q_1_perm_0, x = var_220)[name = tensor("transpose_58")]; tensor qk_1 = matmul(transpose_x = qk_1_transpose_x_1, transpose_y = qk_1_transpose_y_1, x = q_1, y = k_1)[name = tensor("qk_1")]; - tensor var_188 = const()[name = tensor("op_188"), val = tensor([4, 1])]; - tensor var_189 = reshape(shape = var_188, x = sqrt_s_t_1)[name = tensor("op_189")]; - tensor M_1 = real_div(x = encoder__causal_mask, y = var_189)[name = tensor("M_1")]; - tensor var_191 = mul(x = qk_1, y = M_1)[name = tensor("op_191")]; + tensor var_244 = const()[name = tensor("op_244"), val = tensor([4, 1])]; + tensor var_245 = reshape(shape = var_244, x = sqrt_s_t_1)[name = tensor("op_245")]; + tensor M_1 = real_div(x = inner_encoder__causal_mask, y = var_245)[name = tensor("M_1")]; + tensor var_247 = mul(x = qk_1, y = M_1)[name = tensor("op_247")]; tensor inner_1_transpose_x_0 = const()[name = tensor("inner_1_transpose_x_0"), val = tensor(false)]; tensor inner_1_transpose_y_0 = const()[name = tensor("inner_1_transpose_y_0"), val = tensor(false)]; - tensor v_1 = transpose(perm = v_1_perm_0, x = var_178)[name = tensor("transpose_56")]; - tensor inner_1 = matmul(transpose_x = inner_1_transpose_x_0, transpose_y = inner_1_transpose_y_0, x = var_191, y = v_1)[name = tensor("inner_1")]; - tensor var_193_transpose_x_0 = const()[name = tensor("op_193_transpose_x_0"), val = tensor(false)]; - tensor var_193_transpose_y_0 = const()[name = tensor("op_193_transpose_y_0"), val = tensor(false)]; - tensor var_193 = matmul(transpose_x = var_193_transpose_x_0, transpose_y = var_193_transpose_y_0, x = q_1, y = prev_kv_1)[name = tensor("op_193")]; - tensor var_194 = real_div(x = sqrt_s0_1, y = sqrt_s_t_1)[name = tensor("op_194")]; - tensor var_195 = const()[name = tensor("op_195"), val = tensor([1, 1, 4, 1])]; - tensor var_196 = reshape(shape = var_195, x = var_194)[name = tensor("op_196")]; - tensor cross_1 = mul(x = var_193, y = var_196)[name = tensor("cross_1")]; + tensor v_1 = transpose(perm = v_1_perm_0, x = var_234)[name = tensor("transpose_56")]; + tensor inner_1 = matmul(transpose_x = inner_1_transpose_x_0, transpose_y = inner_1_transpose_y_0, x = var_247, y = v_1)[name = tensor("inner_1")]; + tensor var_249_transpose_x_0 = const()[name = tensor("op_249_transpose_x_0"), val = tensor(false)]; + tensor var_249_transpose_y_0 = const()[name = tensor("op_249_transpose_y_0"), val = tensor(false)]; + tensor var_249 = matmul(transpose_x = var_249_transpose_x_0, transpose_y = var_249_transpose_y_0, x = q_1, y = prev_kv_1)[name = tensor("op_249")]; + tensor var_250 = real_div(x = sqrt_s0_1, y = sqrt_s_t_1)[name = tensor("op_250")]; + tensor var_251 = const()[name = tensor("op_251"), val = tensor([1, 1, 4, 1])]; + tensor var_252 = reshape(shape = var_251, x = var_250)[name = tensor("op_252")]; + tensor cross_1 = mul(x = var_249, y = var_252)[name = tensor("cross_1")]; tensor out_1 = add(x = inner_1, y = cross_1)[name = tensor("out_1")]; - tensor var_199 = mul(x = prev_kv_1, y = sqrt_s0_1)[name = tensor("op_199")]; - tensor var_201_transpose_x_1 = const()[name = tensor("op_201_transpose_x_1"), val = tensor(true)]; - tensor var_201_transpose_y_1 = const()[name = tensor("op_201_transpose_y_1"), val = tensor(false)]; - tensor var_201 = matmul(transpose_x = var_201_transpose_x_1, transpose_y = var_201_transpose_y_1, x = k_1, y = v_1)[name = tensor("op_201")]; - tensor new_kv_unnorm_1 = add(x = var_199, y = var_201)[name = tensor("new_kv_unnorm_1")]; - tensor var_203 = const()[name = tensor("op_203"), val = tensor(0x1p+2)]; - tensor new_scale_1 = add(x = prev_scale_1, y = var_203)[name = tensor("new_scale_1")]; - tensor var_205 = sqrt(x = new_scale_1)[name = tensor("op_205")]; - tensor var_206 = real_div(x = new_kv_unnorm_1, y = var_205)[name = tensor("op_206")]; - tensor var_207_perm_0 = const()[name = tensor("op_207_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_255 = mul(x = prev_kv_1, y = sqrt_s0_1)[name = tensor("op_255")]; + tensor var_257_transpose_x_1 = const()[name = tensor("op_257_transpose_x_1"), val = tensor(true)]; + tensor var_257_transpose_y_1 = const()[name = tensor("op_257_transpose_y_1"), val = tensor(false)]; + tensor var_257 = matmul(transpose_x = var_257_transpose_x_1, transpose_y = var_257_transpose_y_1, x = k_1, y = v_1)[name = tensor("op_257")]; + tensor new_kv_unnorm_1 = add(x = var_255, y = var_257)[name = tensor("new_kv_unnorm_1")]; + tensor var_259 = const()[name = tensor("op_259"), val = tensor(0x1p+2)]; + tensor new_scale_1 = add(x = prev_scale_1, y = var_259)[name = tensor("new_scale_1")]; + tensor var_261 = sqrt(x = new_scale_1)[name = tensor("op_261")]; + tensor var_262 = real_div(x = new_kv_unnorm_1, y = var_261)[name = tensor("op_262")]; + tensor var_263_perm_0 = const()[name = tensor("op_263_perm_0"), val = tensor([0, 2, 1, 3])]; tensor out_3_axes_0 = const()[name = tensor("out_3_axes_0"), val = tensor([-1])]; - tensor var_207 = transpose(perm = var_207_perm_0, x = out_1)[name = tensor("transpose_55")]; - tensor out_3 = layer_norm(axes = out_3_axes_0, epsilon = var_17, x = var_207)[name = tensor("out_3")]; - tensor var_211 = const()[name = tensor("op_211"), val = tensor([1, 4, 256])]; - tensor out_5 = reshape(shape = var_211, x = out_3)[name = tensor("out_5")]; - tensor var_213 = silu(x = input_17)[name = tensor("op_213")]; - tensor input_19 = mul(x = var_213, y = out_5)[name = tensor("input_19")]; - tensor ret_out_1 = linear(bias = encoder_out_proj_0_bias, weight = encoder_out_proj_0_weight, x = input_19)[name = tensor("linear_7")]; - tensor x_3 = add(x = input_13, y = ret_out_1)[name = tensor("x_3")]; + tensor var_263 = transpose(perm = var_263_perm_0, x = out_1)[name = tensor("transpose_55")]; + tensor out_3 = layer_norm(axes = out_3_axes_0, epsilon = var_73, x = var_263)[name = tensor("out_3")]; + tensor var_267 = const()[name = tensor("op_267"), val = tensor([1, 4, 256])]; + tensor out_5 = reshape(shape = var_267, x = out_3)[name = tensor("out_5")]; + tensor var_269 = silu(x = input_19)[name = tensor("op_269")]; + tensor input_21 = mul(x = var_269, y = out_5)[name = tensor("input_21")]; + tensor ret_out_1 = linear(bias = inner_encoder_out_proj_0_bias, weight = inner_encoder_out_proj_0_weight, x = input_21)[name = tensor("linear_7")]; + tensor x_3 = add(x = input_15, y = ret_out_1)[name = tensor("x_3")]; tensor window_1_begin_0 = const()[name = tensor("window_1_begin_0"), val = tensor([0, 0, 0, 0])]; tensor window_1_end_0 = const()[name = tensor("window_1_end_0"), val = tensor([1, 1, 16, 256])]; tensor window_1_end_mask_0 = const()[name = tensor("window_1_end_mask_0"), val = tensor([false, true, true, true])]; tensor window_1_squeeze_mask_0 = const()[name = tensor("window_1_squeeze_mask_0"), val = tensor([true, false, false, false])]; tensor window_1 = slice_by_index(begin = window_1_begin_0, end = window_1_end_0, end_mask = window_1_end_mask_0, squeeze_mask = window_1_squeeze_mask_0, x = enc_conv_cache)[name = tensor("window_1")]; - tensor var_221_begin_0 = const()[name = tensor("op_221_begin_0"), val = tensor([0, 0, 0])]; - tensor var_221_end_0 = const()[name = tensor("op_221_end_0"), val = tensor([1, 1, 256])]; - tensor var_221_end_mask_0 = const()[name = tensor("op_221_end_mask_0"), val = tensor([true, false, true])]; - tensor var_221 = slice_by_index(begin = var_221_begin_0, end = var_221_end_0, end_mask = var_221_end_mask_0, x = x_3)[name = tensor("op_221")]; - tensor var_224_begin_0 = const()[name = tensor("op_224_begin_0"), val = tensor([0, 1, 0])]; - tensor var_224_end_0 = const()[name = tensor("op_224_end_0"), val = tensor([1, 16, 256])]; - tensor var_224_end_mask_0 = const()[name = tensor("op_224_end_mask_0"), val = tensor([true, true, true])]; - tensor var_224 = slice_by_index(begin = var_224_begin_0, end = var_224_end_0, end_mask = var_224_end_mask_0, x = window_1)[name = tensor("op_224")]; + tensor var_277_begin_0 = const()[name = tensor("op_277_begin_0"), val = tensor([0, 0, 0])]; + tensor var_277_end_0 = const()[name = tensor("op_277_end_0"), val = tensor([1, 1, 256])]; + tensor var_277_end_mask_0 = const()[name = tensor("op_277_end_mask_0"), val = tensor([true, false, true])]; + tensor var_277 = slice_by_index(begin = var_277_begin_0, end = var_277_end_0, end_mask = var_277_end_mask_0, x = x_3)[name = tensor("op_277")]; + tensor var_280_begin_0 = const()[name = tensor("op_280_begin_0"), val = tensor([0, 1, 0])]; + tensor var_280_end_0 = const()[name = tensor("op_280_end_0"), val = tensor([1, 16, 256])]; + tensor var_280_end_mask_0 = const()[name = tensor("op_280_end_mask_0"), val = tensor([true, true, true])]; + tensor var_280 = slice_by_index(begin = var_280_begin_0, end = var_280_end_0, end_mask = var_280_end_mask_0, x = window_1)[name = tensor("op_280")]; tensor window_3_interleave_0 = const()[name = tensor("window_3_interleave_0"), val = tensor(false)]; - tensor window_3 = concat(axis = var_26, interleave = window_3_interleave_0, values = (var_224, var_221))[name = tensor("window_3")]; - tensor var_229_begin_0 = const()[name = tensor("op_229_begin_0"), val = tensor([0, 1, 0])]; - tensor var_229_end_0 = const()[name = tensor("op_229_end_0"), val = tensor([1, 2, 256])]; - tensor var_229_end_mask_0 = const()[name = tensor("op_229_end_mask_0"), val = tensor([true, false, true])]; - tensor var_229 = slice_by_index(begin = var_229_begin_0, end = var_229_end_0, end_mask = var_229_end_mask_0, x = x_3)[name = tensor("op_229")]; - tensor var_232_begin_0 = const()[name = tensor("op_232_begin_0"), val = tensor([0, 1, 0])]; - tensor var_232_end_0 = const()[name = tensor("op_232_end_0"), val = tensor([1, 16, 256])]; - tensor var_232_end_mask_0 = const()[name = tensor("op_232_end_mask_0"), val = tensor([true, true, true])]; - tensor var_232 = slice_by_index(begin = var_232_begin_0, end = var_232_end_0, end_mask = var_232_end_mask_0, x = window_3)[name = tensor("op_232")]; + tensor window_3 = concat(axis = var_82, interleave = window_3_interleave_0, values = (var_280, var_277))[name = tensor("window_3")]; + tensor var_285_begin_0 = const()[name = tensor("op_285_begin_0"), val = tensor([0, 1, 0])]; + tensor var_285_end_0 = const()[name = tensor("op_285_end_0"), val = tensor([1, 2, 256])]; + tensor var_285_end_mask_0 = const()[name = tensor("op_285_end_mask_0"), val = tensor([true, false, true])]; + tensor var_285 = slice_by_index(begin = var_285_begin_0, end = var_285_end_0, end_mask = var_285_end_mask_0, x = x_3)[name = tensor("op_285")]; + tensor var_288_begin_0 = const()[name = tensor("op_288_begin_0"), val = tensor([0, 1, 0])]; + tensor var_288_end_0 = const()[name = tensor("op_288_end_0"), val = tensor([1, 16, 256])]; + tensor var_288_end_mask_0 = const()[name = tensor("op_288_end_mask_0"), val = tensor([true, true, true])]; + tensor var_288 = slice_by_index(begin = var_288_begin_0, end = var_288_end_0, end_mask = var_288_end_mask_0, x = window_3)[name = tensor("op_288")]; tensor window_5_interleave_0 = const()[name = tensor("window_5_interleave_0"), val = tensor(false)]; - tensor window_5 = concat(axis = var_26, interleave = window_5_interleave_0, values = (var_232, var_229))[name = tensor("window_5")]; - tensor var_237_begin_0 = const()[name = tensor("op_237_begin_0"), val = tensor([0, 2, 0])]; - tensor var_237_end_0 = const()[name = tensor("op_237_end_0"), val = tensor([1, 3, 256])]; - tensor var_237_end_mask_0 = const()[name = tensor("op_237_end_mask_0"), val = tensor([true, false, true])]; - tensor var_237 = slice_by_index(begin = var_237_begin_0, end = var_237_end_0, end_mask = var_237_end_mask_0, x = x_3)[name = tensor("op_237")]; - tensor var_240_begin_0 = const()[name = tensor("op_240_begin_0"), val = tensor([0, 1, 0])]; - tensor var_240_end_0 = const()[name = tensor("op_240_end_0"), val = tensor([1, 16, 256])]; - tensor var_240_end_mask_0 = const()[name = tensor("op_240_end_mask_0"), val = tensor([true, true, true])]; - tensor var_240 = slice_by_index(begin = var_240_begin_0, end = var_240_end_0, end_mask = var_240_end_mask_0, x = window_5)[name = tensor("op_240")]; + tensor window_5 = concat(axis = var_82, interleave = window_5_interleave_0, values = (var_288, var_285))[name = tensor("window_5")]; + tensor var_293_begin_0 = const()[name = tensor("op_293_begin_0"), val = tensor([0, 2, 0])]; + tensor var_293_end_0 = const()[name = tensor("op_293_end_0"), val = tensor([1, 3, 256])]; + tensor var_293_end_mask_0 = const()[name = tensor("op_293_end_mask_0"), val = tensor([true, false, true])]; + tensor var_293 = slice_by_index(begin = var_293_begin_0, end = var_293_end_0, end_mask = var_293_end_mask_0, x = x_3)[name = tensor("op_293")]; + tensor var_296_begin_0 = const()[name = tensor("op_296_begin_0"), val = tensor([0, 1, 0])]; + tensor var_296_end_0 = const()[name = tensor("op_296_end_0"), val = tensor([1, 16, 256])]; + tensor var_296_end_mask_0 = const()[name = tensor("op_296_end_mask_0"), val = tensor([true, true, true])]; + tensor var_296 = slice_by_index(begin = var_296_begin_0, end = var_296_end_0, end_mask = var_296_end_mask_0, x = window_5)[name = tensor("op_296")]; tensor window_7_interleave_0 = const()[name = tensor("window_7_interleave_0"), val = tensor(false)]; - tensor window_7 = concat(axis = var_26, interleave = window_7_interleave_0, values = (var_240, var_237))[name = tensor("window_7")]; - tensor var_245_begin_0 = const()[name = tensor("op_245_begin_0"), val = tensor([0, 3, 0])]; - tensor var_245_end_0 = const()[name = tensor("op_245_end_0"), val = tensor([1, 1, 256])]; - tensor var_245_end_mask_0 = const()[name = tensor("op_245_end_mask_0"), val = tensor([true, true, true])]; - tensor var_245 = slice_by_index(begin = var_245_begin_0, end = var_245_end_0, end_mask = var_245_end_mask_0, x = x_3)[name = tensor("op_245")]; - tensor var_248_begin_0 = const()[name = tensor("op_248_begin_0"), val = tensor([0, 1, 0])]; - tensor var_248_end_0 = const()[name = tensor("op_248_end_0"), val = tensor([1, 16, 256])]; - tensor var_248_end_mask_0 = const()[name = tensor("op_248_end_mask_0"), val = tensor([true, true, true])]; - tensor var_248 = slice_by_index(begin = var_248_begin_0, end = var_248_end_0, end_mask = var_248_end_mask_0, x = window_7)[name = tensor("op_248")]; + tensor window_7 = concat(axis = var_82, interleave = window_7_interleave_0, values = (var_296, var_293))[name = tensor("window_7")]; + tensor var_301_begin_0 = const()[name = tensor("op_301_begin_0"), val = tensor([0, 3, 0])]; + tensor var_301_end_0 = const()[name = tensor("op_301_end_0"), val = tensor([1, 1, 256])]; + tensor var_301_end_mask_0 = const()[name = tensor("op_301_end_mask_0"), val = tensor([true, true, true])]; + tensor var_301 = slice_by_index(begin = var_301_begin_0, end = var_301_end_0, end_mask = var_301_end_mask_0, x = x_3)[name = tensor("op_301")]; + tensor var_304_begin_0 = const()[name = tensor("op_304_begin_0"), val = tensor([0, 1, 0])]; + tensor var_304_end_0 = const()[name = tensor("op_304_end_0"), val = tensor([1, 16, 256])]; + tensor var_304_end_mask_0 = const()[name = tensor("op_304_end_mask_0"), val = tensor([true, true, true])]; + tensor var_304 = slice_by_index(begin = var_304_begin_0, end = var_304_end_0, end_mask = var_304_end_mask_0, x = window_7)[name = tensor("op_304")]; tensor window_9_interleave_0 = const()[name = tensor("window_9_interleave_0"), val = tensor(false)]; - tensor window_9 = concat(axis = var_26, interleave = window_9_interleave_0, values = (var_248, var_245))[name = tensor("window_9")]; - tensor input_21_interleave_0 = const()[name = tensor("input_21_interleave_0"), val = tensor(false)]; - tensor input_21 = concat(axis = var_23, interleave = input_21_interleave_0, values = (window_3, window_5, window_7, window_9))[name = tensor("input_21")]; + tensor window_9 = concat(axis = var_82, interleave = window_9_interleave_0, values = (var_304, var_301))[name = tensor("window_9")]; + tensor input_23_interleave_0 = const()[name = tensor("input_23_interleave_0"), val = tensor(false)]; + tensor input_23 = concat(axis = var_68, interleave = input_23_interleave_0, values = (window_3, window_5, window_7, window_9))[name = tensor("input_23")]; tensor x_5_axes_0 = const()[name = tensor("x_5_axes_0"), val = tensor([-1])]; - tensor x_5 = layer_norm(axes = x_5_axes_0, beta = encoder_conv_module_0_sequential_0_bias, epsilon = var_28, gamma = encoder_conv_module_0_sequential_0_weight, x = input_21)[name = tensor("x_5")]; - tensor input_23_perm_0 = const()[name = tensor("input_23_perm_0"), val = tensor([0, 2, 1])]; + tensor x_5 = layer_norm(axes = x_5_axes_0, beta = inner_encoder_conv_module_0_sequential_0_bias, epsilon = var_65, gamma = inner_encoder_conv_module_0_sequential_0_weight, x = input_23)[name = tensor("x_5")]; + tensor input_25_perm_0 = const()[name = tensor("input_25_perm_0"), val = tensor([0, 2, 1])]; tensor inputs_3_pad_type_0 = const()[name = tensor("inputs_3_pad_type_0"), val = tensor("valid")]; tensor inputs_3_strides_0 = const()[name = tensor("inputs_3_strides_0"), val = tensor([1])]; tensor inputs_3_pad_0 = const()[name = tensor("inputs_3_pad_0"), val = tensor([0, 0])]; tensor inputs_3_dilations_0 = const()[name = tensor("inputs_3_dilations_0"), val = tensor([1])]; tensor inputs_3_groups_0 = const()[name = tensor("inputs_3_groups_0"), val = tensor(1)]; - tensor input_23 = transpose(perm = input_23_perm_0, x = x_5)[name = tensor("transpose_54")]; - tensor inputs_3 = conv(bias = encoder_conv_module_0_sequential_2_conv_bias, dilations = inputs_3_dilations_0, groups = inputs_3_groups_0, pad = inputs_3_pad_0, pad_type = inputs_3_pad_type_0, strides = inputs_3_strides_0, weight = encoder_conv_module_0_sequential_2_conv_weight, x = input_23)[name = tensor("inputs_3")]; - tensor var_273_split_sizes_0 = const()[name = tensor("op_273_split_sizes_0"), val = tensor([256, 256])]; - tensor var_273_axis_0 = const()[name = tensor("op_273_axis_0"), val = tensor(1)]; - tensor var_273_0, tensor var_273_1 = split(axis = var_273_axis_0, split_sizes = var_273_split_sizes_0, x = inputs_3)[name = tensor("op_273")]; - tensor var_275 = sigmoid(x = var_273_1)[name = tensor("op_275")]; - tensor inputs_5 = mul(x = var_273_0, y = var_275)[name = tensor("inputs_5")]; + tensor input_25 = transpose(perm = input_25_perm_0, x = x_5)[name = tensor("transpose_54")]; + tensor inputs_3 = conv(bias = inner_encoder_conv_module_0_sequential_2_conv_bias, dilations = inputs_3_dilations_0, groups = inputs_3_groups_0, pad = inputs_3_pad_0, pad_type = inputs_3_pad_type_0, strides = inputs_3_strides_0, weight = inner_encoder_conv_module_0_sequential_2_conv_weight, x = input_25)[name = tensor("inputs_3")]; + tensor var_329_split_sizes_0 = const()[name = tensor("op_329_split_sizes_0"), val = tensor([256, 256])]; + tensor var_329_axis_0 = const()[name = tensor("op_329_axis_0"), val = tensor(1)]; + tensor var_329_0, tensor var_329_1 = split(axis = var_329_axis_0, split_sizes = var_329_split_sizes_0, x = inputs_3)[name = tensor("op_329")]; + tensor var_331 = sigmoid(x = var_329_1)[name = tensor("op_331")]; + tensor inputs_5 = mul(x = var_329_0, y = var_331)[name = tensor("inputs_5")]; tensor outputs_aug_1_pad_type_0 = const()[name = tensor("outputs_aug_1_pad_type_0"), val = tensor("custom")]; tensor outputs_aug_1_pad_0 = const()[name = tensor("outputs_aug_1_pad_0"), val = tensor([15, 15])]; tensor outputs_aug_1_groups_0 = const()[name = tensor("outputs_aug_1_groups_0"), val = tensor(256)]; tensor outputs_aug_1_strides_0 = const()[name = tensor("outputs_aug_1_strides_0"), val = tensor([1])]; tensor outputs_aug_1_dilations_0 = const()[name = tensor("outputs_aug_1_dilations_0"), val = tensor([1])]; - tensor outputs_aug_1 = conv(dilations = outputs_aug_1_dilations_0, groups = outputs_aug_1_groups_0, pad = outputs_aug_1_pad_0, pad_type = outputs_aug_1_pad_type_0, strides = outputs_aug_1_strides_0, weight = encoder_conv_module_0_sequential_4_conv_weight, x = inputs_5)[name = tensor("outputs_aug_1")]; - tensor input_25_begin_0 = const()[name = tensor("input_25_begin_0"), val = tensor([0, 0, 0])]; - tensor input_25_end_0 = const()[name = tensor("input_25_end_0"), val = tensor([4, 256, 16])]; - tensor input_25_end_mask_0 = const()[name = tensor("input_25_end_mask_0"), val = tensor([true, true, false])]; - tensor input_25 = slice_by_index(begin = input_25_begin_0, end = input_25_end_0, end_mask = input_25_end_mask_0, x = outputs_aug_1)[name = tensor("input_25")]; - tensor inputs_7 = batch_norm(beta = encoder_conv_module_0_sequential_5_bias, epsilon = var_28, gamma = encoder_conv_module_0_sequential_5_weight, mean = encoder_conv_module_0_sequential_5_running_mean, variance = encoder_conv_module_0_sequential_5_running_var, x = input_25)[name = tensor("inputs_7")]; - tensor input_27 = silu(x = inputs_7)[name = tensor("input_27")]; - tensor input_29_pad_type_0 = const()[name = tensor("input_29_pad_type_0"), val = tensor("valid")]; - tensor input_29_strides_0 = const()[name = tensor("input_29_strides_0"), val = tensor([1])]; - tensor input_29_pad_0 = const()[name = tensor("input_29_pad_0"), val = tensor([0, 0])]; - tensor input_29_dilations_0 = const()[name = tensor("input_29_dilations_0"), val = tensor([1])]; - tensor input_29_groups_0 = const()[name = tensor("input_29_groups_0"), val = tensor(1)]; - tensor input_29 = conv(bias = encoder_conv_module_0_sequential_7_conv_bias, dilations = input_29_dilations_0, groups = input_29_groups_0, pad = input_29_pad_0, pad_type = input_29_pad_type_0, strides = input_29_strides_0, weight = encoder_conv_module_0_sequential_7_conv_weight, x = input_27)[name = tensor("input_29")]; + tensor outputs_aug_1 = conv(dilations = outputs_aug_1_dilations_0, groups = outputs_aug_1_groups_0, pad = outputs_aug_1_pad_0, pad_type = outputs_aug_1_pad_type_0, strides = outputs_aug_1_strides_0, weight = inner_encoder_conv_module_0_sequential_4_conv_weight, x = inputs_5)[name = tensor("outputs_aug_1")]; + tensor input_27_begin_0 = const()[name = tensor("input_27_begin_0"), val = tensor([0, 0, 0])]; + tensor input_27_end_0 = const()[name = tensor("input_27_end_0"), val = tensor([4, 256, 16])]; + tensor input_27_end_mask_0 = const()[name = tensor("input_27_end_mask_0"), val = tensor([true, true, false])]; + tensor input_27 = slice_by_index(begin = input_27_begin_0, end = input_27_end_0, end_mask = input_27_end_mask_0, x = outputs_aug_1)[name = tensor("input_27")]; + tensor inputs_7 = batch_norm(beta = inner_encoder_conv_module_0_sequential_5_bias, epsilon = var_65, gamma = inner_encoder_conv_module_0_sequential_5_weight, mean = inner_encoder_conv_module_0_sequential_5_running_mean, variance = inner_encoder_conv_module_0_sequential_5_running_var, x = input_27)[name = tensor("inputs_7")]; + tensor input_29 = silu(x = inputs_7)[name = tensor("input_29")]; + tensor input_31_pad_type_0 = const()[name = tensor("input_31_pad_type_0"), val = tensor("valid")]; + tensor input_31_strides_0 = const()[name = tensor("input_31_strides_0"), val = tensor([1])]; + tensor input_31_pad_0 = const()[name = tensor("input_31_pad_0"), val = tensor([0, 0])]; + tensor input_31_dilations_0 = const()[name = tensor("input_31_dilations_0"), val = tensor([1])]; + tensor input_31_groups_0 = const()[name = tensor("input_31_groups_0"), val = tensor(1)]; + tensor input_31 = conv(bias = inner_encoder_conv_module_0_sequential_7_conv_bias, dilations = input_31_dilations_0, groups = input_31_groups_0, pad = input_31_pad_0, pad_type = input_31_pad_type_0, strides = input_31_strides_0, weight = inner_encoder_conv_module_0_sequential_7_conv_weight, x = input_29)[name = tensor("input_31")]; tensor conv_out_1_perm_0 = const()[name = tensor("conv_out_1_perm_0"), val = tensor([0, 2, 1])]; - tensor var_306_begin_0 = const()[name = tensor("op_306_begin_0"), val = tensor([0, -1, 0])]; - tensor var_306_end_0 = const()[name = tensor("op_306_end_0"), val = tensor([4, 16, 256])]; - tensor var_306_end_mask_0 = const()[name = tensor("op_306_end_mask_0"), val = tensor([true, true, true])]; - tensor conv_out_1 = transpose(perm = conv_out_1_perm_0, x = input_29)[name = tensor("transpose_53")]; - tensor var_306 = slice_by_index(begin = var_306_begin_0, end = var_306_end_0, end_mask = var_306_end_mask_0, x = conv_out_1)[name = tensor("op_306")]; - tensor var_308_perm_0 = const()[name = tensor("op_308_perm_0"), val = tensor([1, 0, 2])]; - tensor var_308 = transpose(perm = var_308_perm_0, x = var_306)[name = tensor("transpose_52")]; - tensor input_31 = add(x = x_3, y = var_308)[name = tensor("input_31")]; - tensor input_33_axes_0 = const()[name = tensor("input_33_axes_0"), val = tensor([-1])]; - tensor input_33 = layer_norm(axes = input_33_axes_0, beta = encoder_ffn2_0_module_sequential_0_bias, epsilon = var_28, gamma = encoder_ffn2_0_module_sequential_0_weight, x = input_31)[name = tensor("input_33")]; - tensor inputs_9 = linear(bias = encoder_ffn2_0_module_sequential_1_linear_bias, weight = encoder_ffn2_0_module_sequential_1_linear_weight, x = input_33)[name = tensor("linear_8")]; - tensor input_35 = silu(x = inputs_9)[name = tensor("input_35")]; - tensor input_39 = linear(bias = encoder_ffn2_0_module_sequential_4_linear_bias, weight = encoder_ffn2_0_module_sequential_4_linear_weight, x = input_35)[name = tensor("linear_9")]; - tensor var_331 = const()[name = tensor("op_331"), val = tensor(0x1p-1)]; - tensor var_332 = mul(x = input_39, y = var_331)[name = tensor("op_332")]; - tensor input_41 = add(x = var_332, y = input_31)[name = tensor("input_41")]; - tensor input_43_axes_0 = const()[name = tensor("input_43_axes_0"), val = tensor([-1])]; - tensor input_43 = layer_norm(axes = input_43_axes_0, beta = encoder_layer_norm_0_bias, epsilon = var_28, gamma = encoder_layer_norm_0_weight, x = input_41)[name = tensor("input_43")]; + tensor var_362_begin_0 = const()[name = tensor("op_362_begin_0"), val = tensor([0, -1, 0])]; + tensor var_362_end_0 = const()[name = tensor("op_362_end_0"), val = tensor([4, 16, 256])]; + tensor var_362_end_mask_0 = const()[name = tensor("op_362_end_mask_0"), val = tensor([true, true, true])]; + tensor conv_out_1 = transpose(perm = conv_out_1_perm_0, x = input_31)[name = tensor("transpose_53")]; + tensor var_362 = slice_by_index(begin = var_362_begin_0, end = var_362_end_0, end_mask = var_362_end_mask_0, x = conv_out_1)[name = tensor("op_362")]; + tensor var_364_perm_0 = const()[name = tensor("op_364_perm_0"), val = tensor([1, 0, 2])]; + tensor var_364 = transpose(perm = var_364_perm_0, x = var_362)[name = tensor("transpose_52")]; + tensor input_33 = add(x = x_3, y = var_364)[name = tensor("input_33")]; + tensor input_35_axes_0 = const()[name = tensor("input_35_axes_0"), val = tensor([-1])]; + tensor input_35 = layer_norm(axes = input_35_axes_0, beta = inner_encoder_ffn2_0_module_sequential_0_bias, epsilon = var_65, gamma = inner_encoder_ffn2_0_module_sequential_0_weight, x = input_33)[name = tensor("input_35")]; + tensor inputs_9 = linear(bias = inner_encoder_ffn2_0_module_sequential_1_linear_bias, weight = inner_encoder_ffn2_0_module_sequential_1_linear_weight, x = input_35)[name = tensor("linear_8")]; + tensor input_37 = silu(x = inputs_9)[name = tensor("input_37")]; + tensor input_41 = linear(bias = inner_encoder_ffn2_0_module_sequential_4_linear_bias, weight = inner_encoder_ffn2_0_module_sequential_4_linear_weight, x = input_37)[name = tensor("linear_9")]; + tensor var_387 = const()[name = tensor("op_387"), val = tensor(0x1p-1)]; + tensor var_388 = mul(x = input_41, y = var_387)[name = tensor("op_388")]; + tensor input_43 = add(x = var_388, y = input_33)[name = tensor("input_43")]; tensor input_45_axes_0 = const()[name = tensor("input_45_axes_0"), val = tensor([-1])]; - tensor input_45 = layer_norm(axes = input_45_axes_0, beta = encoder_ffn1_1_module_sequential_0_bias, epsilon = var_28, gamma = encoder_ffn1_1_module_sequential_0_weight, x = input_43)[name = tensor("input_45")]; - tensor inputs_11 = linear(bias = encoder_ffn1_1_module_sequential_1_linear_bias, weight = encoder_ffn1_1_module_sequential_1_linear_weight, x = input_45)[name = tensor("linear_10")]; - tensor input_47 = silu(x = inputs_11)[name = tensor("input_47")]; - tensor input_51 = linear(bias = encoder_ffn1_1_module_sequential_4_linear_bias, weight = encoder_ffn1_1_module_sequential_4_linear_weight, x = input_47)[name = tensor("linear_11")]; - tensor var_361 = const()[name = tensor("op_361"), val = tensor(0x1p-1)]; - tensor var_362 = mul(x = input_51, y = var_361)[name = tensor("op_362")]; - tensor input_53 = add(x = var_362, y = input_43)[name = tensor("input_53")]; + tensor input_45 = layer_norm(axes = input_45_axes_0, beta = inner_encoder_layer_norm_0_bias, epsilon = var_65, gamma = inner_encoder_layer_norm_0_weight, x = input_43)[name = tensor("input_45")]; + tensor input_47_axes_0 = const()[name = tensor("input_47_axes_0"), val = tensor([-1])]; + tensor input_47 = layer_norm(axes = input_47_axes_0, beta = inner_encoder_ffn1_1_module_sequential_0_bias, epsilon = var_65, gamma = inner_encoder_ffn1_1_module_sequential_0_weight, x = input_45)[name = tensor("input_47")]; + tensor inputs_11 = linear(bias = inner_encoder_ffn1_1_module_sequential_1_linear_bias, weight = inner_encoder_ffn1_1_module_sequential_1_linear_weight, x = input_47)[name = tensor("linear_10")]; + tensor input_49 = silu(x = inputs_11)[name = tensor("input_49")]; + tensor input_53 = linear(bias = inner_encoder_ffn1_1_module_sequential_4_linear_bias, weight = inner_encoder_ffn1_1_module_sequential_4_linear_weight, x = input_49)[name = tensor("linear_11")]; + tensor var_417 = const()[name = tensor("op_417"), val = tensor(0x1p-1)]; + tensor var_418 = mul(x = input_53, y = var_417)[name = tensor("op_418")]; + tensor input_55 = add(x = var_418, y = input_45)[name = tensor("input_55")]; tensor x_7_axes_0 = const()[name = tensor("x_7_axes_0"), val = tensor([-1])]; - tensor x_7 = layer_norm(axes = x_7_axes_0, beta = encoder_ret_lns_1_bias, epsilon = var_28, gamma = encoder_ret_lns_1_weight, x = input_53)[name = tensor("x_7")]; + tensor x_7 = layer_norm(axes = x_7_axes_0, beta = inner_encoder_ret_lns_1_bias, epsilon = var_65, gamma = inner_encoder_ret_lns_1_weight, x = input_55)[name = tensor("x_7")]; tensor prev_kv_3_begin_0 = const()[name = tensor("prev_kv_3_begin_0"), val = tensor([1, 0, 0, 0, 0])]; tensor prev_kv_3_end_0 = const()[name = tensor("prev_kv_3_end_0"), val = tensor([2, 1, 4, 64, 64])]; tensor prev_kv_3_end_mask_0 = const()[name = tensor("prev_kv_3_end_mask_0"), val = tensor([false, true, true, true, true])]; @@ -416,173 +438,173 @@ program(1.0) tensor prev_scale_3_end_mask_0 = const()[name = tensor("prev_scale_3_end_mask_0"), val = tensor([false, true])]; tensor prev_scale_3_squeeze_mask_0 = const()[name = tensor("prev_scale_3_squeeze_mask_0"), val = tensor([true, false])]; tensor prev_scale_3 = slice_by_index(begin = prev_scale_3_begin_0, end = prev_scale_3_end_0, end_mask = prev_scale_3_end_mask_0, squeeze_mask = prev_scale_3_squeeze_mask_0, x = enc_scale)[name = tensor("prev_scale_3")]; - tensor var_376 = linear(bias = encoder_q_proj_1_bias, weight = encoder_q_proj_1_weight, x = x_7)[name = tensor("linear_12")]; - tensor var_377 = const()[name = tensor("op_377"), val = tensor([1, 4, 4, 64])]; - tensor var_378 = reshape(shape = var_377, x = var_376)[name = tensor("op_378")]; + tensor var_432 = linear(bias = inner_encoder_q_proj_1_bias, weight = inner_encoder_q_proj_1_weight, x = x_7)[name = tensor("linear_12")]; + tensor var_433 = const()[name = tensor("op_433"), val = tensor([1, 4, 4, 64])]; + tensor var_434 = reshape(shape = var_433, x = var_432)[name = tensor("op_434")]; tensor q_3_perm_0 = const()[name = tensor("q_3_perm_0"), val = tensor([0, 2, 1, 3])]; - tensor var_382 = linear(bias = encoder_k_proj_1_bias, weight = encoder_k_proj_1_weight, x = x_7)[name = tensor("linear_13")]; - tensor var_383 = const()[name = tensor("op_383"), val = tensor(0x1p-3)]; - tensor var_384 = mul(x = var_382, y = var_383)[name = tensor("op_384")]; - tensor var_385 = const()[name = tensor("op_385"), val = tensor([1, 4, 4, 64])]; - tensor var_386 = reshape(shape = var_385, x = var_384)[name = tensor("op_386")]; + tensor var_438 = linear(bias = inner_encoder_k_proj_1_bias, weight = inner_encoder_k_proj_1_weight, x = x_7)[name = tensor("linear_13")]; + tensor var_439 = const()[name = tensor("op_439"), val = tensor(0x1p-3)]; + tensor var_440 = mul(x = var_438, y = var_439)[name = tensor("op_440")]; + tensor var_441 = const()[name = tensor("op_441"), val = tensor([1, 4, 4, 64])]; + tensor var_442 = reshape(shape = var_441, x = var_440)[name = tensor("op_442")]; tensor k_3_perm_0 = const()[name = tensor("k_3_perm_0"), val = tensor([0, 2, 1, 3])]; - tensor var_390 = linear(bias = encoder_v_proj_1_bias, weight = encoder_v_proj_1_weight, x = x_7)[name = tensor("linear_14")]; - tensor var_391 = const()[name = tensor("op_391"), val = tensor([1, 4, 4, 64])]; - tensor var_392 = reshape(shape = var_391, x = var_390)[name = tensor("op_392")]; + tensor var_446 = linear(bias = inner_encoder_v_proj_1_bias, weight = inner_encoder_v_proj_1_weight, x = x_7)[name = tensor("linear_14")]; + tensor var_447 = const()[name = tensor("op_447"), val = tensor([1, 4, 4, 64])]; + tensor var_448 = reshape(shape = var_447, x = var_446)[name = tensor("op_448")]; tensor v_3_perm_0 = const()[name = tensor("v_3_perm_0"), val = tensor([0, 2, 1, 3])]; - tensor input_57 = linear(bias = encoder_g_proj_1_bias, weight = encoder_g_proj_1_weight, x = x_7)[name = tensor("linear_15")]; + tensor input_59 = linear(bias = inner_encoder_g_proj_1_bias, weight = inner_encoder_g_proj_1_weight, x = x_7)[name = tensor("linear_15")]; tensor sqrt_s0_3 = sqrt(x = prev_scale_3)[name = tensor("sqrt_s0_3")]; - tensor s_t_3 = add(x = prev_scale_3, y = encoder__t_index)[name = tensor("s_t_3")]; + tensor s_t_3 = add(x = prev_scale_3, y = inner_encoder__t_index)[name = tensor("s_t_3")]; tensor sqrt_s_t_3 = sqrt(x = s_t_3)[name = tensor("sqrt_s_t_3")]; tensor qk_3_transpose_x_1 = const()[name = tensor("qk_3_transpose_x_1"), val = tensor(false)]; tensor qk_3_transpose_y_1 = const()[name = tensor("qk_3_transpose_y_1"), val = tensor(true)]; - tensor k_3 = transpose(perm = k_3_perm_0, x = var_386)[name = tensor("transpose_50")]; - tensor q_3 = transpose(perm = q_3_perm_0, x = var_378)[name = tensor("transpose_51")]; + tensor k_3 = transpose(perm = k_3_perm_0, x = var_442)[name = tensor("transpose_50")]; + tensor q_3 = transpose(perm = q_3_perm_0, x = var_434)[name = tensor("transpose_51")]; tensor qk_3 = matmul(transpose_x = qk_3_transpose_x_1, transpose_y = qk_3_transpose_y_1, x = q_3, y = k_3)[name = tensor("qk_3")]; - tensor var_402 = const()[name = tensor("op_402"), val = tensor([4, 1])]; - tensor var_403 = reshape(shape = var_402, x = sqrt_s_t_3)[name = tensor("op_403")]; - tensor M_3 = real_div(x = encoder__causal_mask, y = var_403)[name = tensor("M_3")]; - tensor var_405 = mul(x = qk_3, y = M_3)[name = tensor("op_405")]; + tensor var_458 = const()[name = tensor("op_458"), val = tensor([4, 1])]; + tensor var_459 = reshape(shape = var_458, x = sqrt_s_t_3)[name = tensor("op_459")]; + tensor M_3 = real_div(x = inner_encoder__causal_mask, y = var_459)[name = tensor("M_3")]; + tensor var_461 = mul(x = qk_3, y = M_3)[name = tensor("op_461")]; tensor inner_3_transpose_x_0 = const()[name = tensor("inner_3_transpose_x_0"), val = tensor(false)]; tensor inner_3_transpose_y_0 = const()[name = tensor("inner_3_transpose_y_0"), val = tensor(false)]; - tensor v_3 = transpose(perm = v_3_perm_0, x = var_392)[name = tensor("transpose_49")]; - tensor inner_3 = matmul(transpose_x = inner_3_transpose_x_0, transpose_y = inner_3_transpose_y_0, x = var_405, y = v_3)[name = tensor("inner_3")]; - tensor var_407_transpose_x_0 = const()[name = tensor("op_407_transpose_x_0"), val = tensor(false)]; - tensor var_407_transpose_y_0 = const()[name = tensor("op_407_transpose_y_0"), val = tensor(false)]; - tensor var_407 = matmul(transpose_x = var_407_transpose_x_0, transpose_y = var_407_transpose_y_0, x = q_3, y = prev_kv_3)[name = tensor("op_407")]; - tensor var_408 = real_div(x = sqrt_s0_3, y = sqrt_s_t_3)[name = tensor("op_408")]; - tensor var_409 = const()[name = tensor("op_409"), val = tensor([1, 1, 4, 1])]; - tensor var_410 = reshape(shape = var_409, x = var_408)[name = tensor("op_410")]; - tensor cross_3 = mul(x = var_407, y = var_410)[name = tensor("cross_3")]; + tensor v_3 = transpose(perm = v_3_perm_0, x = var_448)[name = tensor("transpose_49")]; + tensor inner_3 = matmul(transpose_x = inner_3_transpose_x_0, transpose_y = inner_3_transpose_y_0, x = var_461, y = v_3)[name = tensor("inner_3")]; + tensor var_463_transpose_x_0 = const()[name = tensor("op_463_transpose_x_0"), val = tensor(false)]; + tensor var_463_transpose_y_0 = const()[name = tensor("op_463_transpose_y_0"), val = tensor(false)]; + tensor var_463 = matmul(transpose_x = var_463_transpose_x_0, transpose_y = var_463_transpose_y_0, x = q_3, y = prev_kv_3)[name = tensor("op_463")]; + tensor var_464 = real_div(x = sqrt_s0_3, y = sqrt_s_t_3)[name = tensor("op_464")]; + tensor var_465 = const()[name = tensor("op_465"), val = tensor([1, 1, 4, 1])]; + tensor var_466 = reshape(shape = var_465, x = var_464)[name = tensor("op_466")]; + tensor cross_3 = mul(x = var_463, y = var_466)[name = tensor("cross_3")]; tensor out_7 = add(x = inner_3, y = cross_3)[name = tensor("out_7")]; - tensor var_413 = mul(x = prev_kv_3, y = sqrt_s0_3)[name = tensor("op_413")]; - tensor var_415_transpose_x_1 = const()[name = tensor("op_415_transpose_x_1"), val = tensor(true)]; - tensor var_415_transpose_y_1 = const()[name = tensor("op_415_transpose_y_1"), val = tensor(false)]; - tensor var_415 = matmul(transpose_x = var_415_transpose_x_1, transpose_y = var_415_transpose_y_1, x = k_3, y = v_3)[name = tensor("op_415")]; - tensor new_kv_unnorm_3 = add(x = var_413, y = var_415)[name = tensor("new_kv_unnorm_3")]; - tensor var_417 = const()[name = tensor("op_417"), val = tensor(0x1p+2)]; - tensor new_scale_3 = add(x = prev_scale_3, y = var_417)[name = tensor("new_scale_3")]; - tensor var_419 = sqrt(x = new_scale_3)[name = tensor("op_419")]; - tensor var_420 = real_div(x = new_kv_unnorm_3, y = var_419)[name = tensor("op_420")]; - tensor var_421_perm_0 = const()[name = tensor("op_421_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_469 = mul(x = prev_kv_3, y = sqrt_s0_3)[name = tensor("op_469")]; + tensor var_471_transpose_x_1 = const()[name = tensor("op_471_transpose_x_1"), val = tensor(true)]; + tensor var_471_transpose_y_1 = const()[name = tensor("op_471_transpose_y_1"), val = tensor(false)]; + tensor var_471 = matmul(transpose_x = var_471_transpose_x_1, transpose_y = var_471_transpose_y_1, x = k_3, y = v_3)[name = tensor("op_471")]; + tensor new_kv_unnorm_3 = add(x = var_469, y = var_471)[name = tensor("new_kv_unnorm_3")]; + tensor var_473 = const()[name = tensor("op_473"), val = tensor(0x1p+2)]; + tensor new_scale_3 = add(x = prev_scale_3, y = var_473)[name = tensor("new_scale_3")]; + tensor var_475 = sqrt(x = new_scale_3)[name = tensor("op_475")]; + tensor var_476 = real_div(x = new_kv_unnorm_3, y = var_475)[name = tensor("op_476")]; + tensor var_477_perm_0 = const()[name = tensor("op_477_perm_0"), val = tensor([0, 2, 1, 3])]; tensor out_9_axes_0 = const()[name = tensor("out_9_axes_0"), val = tensor([-1])]; - tensor var_421 = transpose(perm = var_421_perm_0, x = out_7)[name = tensor("transpose_48")]; - tensor out_9 = layer_norm(axes = out_9_axes_0, epsilon = var_17, x = var_421)[name = tensor("out_9")]; - tensor var_425 = const()[name = tensor("op_425"), val = tensor([1, 4, 256])]; - tensor out_11 = reshape(shape = var_425, x = out_9)[name = tensor("out_11")]; - tensor var_427 = silu(x = input_57)[name = tensor("op_427")]; - tensor input_59 = mul(x = var_427, y = out_11)[name = tensor("input_59")]; - tensor ret_out_3 = linear(bias = encoder_out_proj_1_bias, weight = encoder_out_proj_1_weight, x = input_59)[name = tensor("linear_16")]; - tensor x_9 = add(x = input_53, y = ret_out_3)[name = tensor("x_9")]; + tensor var_477 = transpose(perm = var_477_perm_0, x = out_7)[name = tensor("transpose_48")]; + tensor out_9 = layer_norm(axes = out_9_axes_0, epsilon = var_73, x = var_477)[name = tensor("out_9")]; + tensor var_481 = const()[name = tensor("op_481"), val = tensor([1, 4, 256])]; + tensor out_11 = reshape(shape = var_481, x = out_9)[name = tensor("out_11")]; + tensor var_483 = silu(x = input_59)[name = tensor("op_483")]; + tensor input_61 = mul(x = var_483, y = out_11)[name = tensor("input_61")]; + tensor ret_out_3 = linear(bias = inner_encoder_out_proj_1_bias, weight = inner_encoder_out_proj_1_weight, x = input_61)[name = tensor("linear_16")]; + tensor x_9 = add(x = input_55, y = ret_out_3)[name = tensor("x_9")]; tensor window_11_begin_0 = const()[name = tensor("window_11_begin_0"), val = tensor([1, 0, 0, 0])]; tensor window_11_end_0 = const()[name = tensor("window_11_end_0"), val = tensor([2, 1, 16, 256])]; tensor window_11_end_mask_0 = const()[name = tensor("window_11_end_mask_0"), val = tensor([false, true, true, true])]; tensor window_11_squeeze_mask_0 = const()[name = tensor("window_11_squeeze_mask_0"), val = tensor([true, false, false, false])]; tensor window_11 = slice_by_index(begin = window_11_begin_0, end = window_11_end_0, end_mask = window_11_end_mask_0, squeeze_mask = window_11_squeeze_mask_0, x = enc_conv_cache)[name = tensor("window_11")]; - tensor var_435_begin_0 = const()[name = tensor("op_435_begin_0"), val = tensor([0, 0, 0])]; - tensor var_435_end_0 = const()[name = tensor("op_435_end_0"), val = tensor([1, 1, 256])]; - tensor var_435_end_mask_0 = const()[name = tensor("op_435_end_mask_0"), val = tensor([true, false, true])]; - tensor var_435 = slice_by_index(begin = var_435_begin_0, end = var_435_end_0, end_mask = var_435_end_mask_0, x = x_9)[name = tensor("op_435")]; - tensor var_438_begin_0 = const()[name = tensor("op_438_begin_0"), val = tensor([0, 1, 0])]; - tensor var_438_end_0 = const()[name = tensor("op_438_end_0"), val = tensor([1, 16, 256])]; - tensor var_438_end_mask_0 = const()[name = tensor("op_438_end_mask_0"), val = tensor([true, true, true])]; - tensor var_438 = slice_by_index(begin = var_438_begin_0, end = var_438_end_0, end_mask = var_438_end_mask_0, x = window_11)[name = tensor("op_438")]; + tensor var_491_begin_0 = const()[name = tensor("op_491_begin_0"), val = tensor([0, 0, 0])]; + tensor var_491_end_0 = const()[name = tensor("op_491_end_0"), val = tensor([1, 1, 256])]; + tensor var_491_end_mask_0 = const()[name = tensor("op_491_end_mask_0"), val = tensor([true, false, true])]; + tensor var_491 = slice_by_index(begin = var_491_begin_0, end = var_491_end_0, end_mask = var_491_end_mask_0, x = x_9)[name = tensor("op_491")]; + tensor var_494_begin_0 = const()[name = tensor("op_494_begin_0"), val = tensor([0, 1, 0])]; + tensor var_494_end_0 = const()[name = tensor("op_494_end_0"), val = tensor([1, 16, 256])]; + tensor var_494_end_mask_0 = const()[name = tensor("op_494_end_mask_0"), val = tensor([true, true, true])]; + tensor var_494 = slice_by_index(begin = var_494_begin_0, end = var_494_end_0, end_mask = var_494_end_mask_0, x = window_11)[name = tensor("op_494")]; tensor window_13_interleave_0 = const()[name = tensor("window_13_interleave_0"), val = tensor(false)]; - tensor window_13 = concat(axis = var_26, interleave = window_13_interleave_0, values = (var_438, var_435))[name = tensor("window_13")]; - tensor var_443_begin_0 = const()[name = tensor("op_443_begin_0"), val = tensor([0, 1, 0])]; - tensor var_443_end_0 = const()[name = tensor("op_443_end_0"), val = tensor([1, 2, 256])]; - tensor var_443_end_mask_0 = const()[name = tensor("op_443_end_mask_0"), val = tensor([true, false, true])]; - tensor var_443 = slice_by_index(begin = var_443_begin_0, end = var_443_end_0, end_mask = var_443_end_mask_0, x = x_9)[name = tensor("op_443")]; - tensor var_446_begin_0 = const()[name = tensor("op_446_begin_0"), val = tensor([0, 1, 0])]; - tensor var_446_end_0 = const()[name = tensor("op_446_end_0"), val = tensor([1, 16, 256])]; - tensor var_446_end_mask_0 = const()[name = tensor("op_446_end_mask_0"), val = tensor([true, true, true])]; - tensor var_446 = slice_by_index(begin = var_446_begin_0, end = var_446_end_0, end_mask = var_446_end_mask_0, x = window_13)[name = tensor("op_446")]; + tensor window_13 = concat(axis = var_82, interleave = window_13_interleave_0, values = (var_494, var_491))[name = tensor("window_13")]; + tensor var_499_begin_0 = const()[name = tensor("op_499_begin_0"), val = tensor([0, 1, 0])]; + tensor var_499_end_0 = const()[name = tensor("op_499_end_0"), val = tensor([1, 2, 256])]; + tensor var_499_end_mask_0 = const()[name = tensor("op_499_end_mask_0"), val = tensor([true, false, true])]; + tensor var_499 = slice_by_index(begin = var_499_begin_0, end = var_499_end_0, end_mask = var_499_end_mask_0, x = x_9)[name = tensor("op_499")]; + tensor var_502_begin_0 = const()[name = tensor("op_502_begin_0"), val = tensor([0, 1, 0])]; + tensor var_502_end_0 = const()[name = tensor("op_502_end_0"), val = tensor([1, 16, 256])]; + tensor var_502_end_mask_0 = const()[name = tensor("op_502_end_mask_0"), val = tensor([true, true, true])]; + tensor var_502 = slice_by_index(begin = var_502_begin_0, end = var_502_end_0, end_mask = var_502_end_mask_0, x = window_13)[name = tensor("op_502")]; tensor window_15_interleave_0 = const()[name = tensor("window_15_interleave_0"), val = tensor(false)]; - tensor window_15 = concat(axis = var_26, interleave = window_15_interleave_0, values = (var_446, var_443))[name = tensor("window_15")]; - tensor var_451_begin_0 = const()[name = tensor("op_451_begin_0"), val = tensor([0, 2, 0])]; - tensor var_451_end_0 = const()[name = tensor("op_451_end_0"), val = tensor([1, 3, 256])]; - tensor var_451_end_mask_0 = const()[name = tensor("op_451_end_mask_0"), val = tensor([true, false, true])]; - tensor var_451 = slice_by_index(begin = var_451_begin_0, end = var_451_end_0, end_mask = var_451_end_mask_0, x = x_9)[name = tensor("op_451")]; - tensor var_454_begin_0 = const()[name = tensor("op_454_begin_0"), val = tensor([0, 1, 0])]; - tensor var_454_end_0 = const()[name = tensor("op_454_end_0"), val = tensor([1, 16, 256])]; - tensor var_454_end_mask_0 = const()[name = tensor("op_454_end_mask_0"), val = tensor([true, true, true])]; - tensor var_454 = slice_by_index(begin = var_454_begin_0, end = var_454_end_0, end_mask = var_454_end_mask_0, x = window_15)[name = tensor("op_454")]; + tensor window_15 = concat(axis = var_82, interleave = window_15_interleave_0, values = (var_502, var_499))[name = tensor("window_15")]; + tensor var_507_begin_0 = const()[name = tensor("op_507_begin_0"), val = tensor([0, 2, 0])]; + tensor var_507_end_0 = const()[name = tensor("op_507_end_0"), val = tensor([1, 3, 256])]; + tensor var_507_end_mask_0 = const()[name = tensor("op_507_end_mask_0"), val = tensor([true, false, true])]; + tensor var_507 = slice_by_index(begin = var_507_begin_0, end = var_507_end_0, end_mask = var_507_end_mask_0, x = x_9)[name = tensor("op_507")]; + tensor var_510_begin_0 = const()[name = tensor("op_510_begin_0"), val = tensor([0, 1, 0])]; + tensor var_510_end_0 = const()[name = tensor("op_510_end_0"), val = tensor([1, 16, 256])]; + tensor var_510_end_mask_0 = const()[name = tensor("op_510_end_mask_0"), val = tensor([true, true, true])]; + tensor var_510 = slice_by_index(begin = var_510_begin_0, end = var_510_end_0, end_mask = var_510_end_mask_0, x = window_15)[name = tensor("op_510")]; tensor window_17_interleave_0 = const()[name = tensor("window_17_interleave_0"), val = tensor(false)]; - tensor window_17 = concat(axis = var_26, interleave = window_17_interleave_0, values = (var_454, var_451))[name = tensor("window_17")]; - tensor var_459_begin_0 = const()[name = tensor("op_459_begin_0"), val = tensor([0, 3, 0])]; - tensor var_459_end_0 = const()[name = tensor("op_459_end_0"), val = tensor([1, 1, 256])]; - tensor var_459_end_mask_0 = const()[name = tensor("op_459_end_mask_0"), val = tensor([true, true, true])]; - tensor var_459 = slice_by_index(begin = var_459_begin_0, end = var_459_end_0, end_mask = var_459_end_mask_0, x = x_9)[name = tensor("op_459")]; - tensor var_462_begin_0 = const()[name = tensor("op_462_begin_0"), val = tensor([0, 1, 0])]; - tensor var_462_end_0 = const()[name = tensor("op_462_end_0"), val = tensor([1, 16, 256])]; - tensor var_462_end_mask_0 = const()[name = tensor("op_462_end_mask_0"), val = tensor([true, true, true])]; - tensor var_462 = slice_by_index(begin = var_462_begin_0, end = var_462_end_0, end_mask = var_462_end_mask_0, x = window_17)[name = tensor("op_462")]; + tensor window_17 = concat(axis = var_82, interleave = window_17_interleave_0, values = (var_510, var_507))[name = tensor("window_17")]; + tensor var_515_begin_0 = const()[name = tensor("op_515_begin_0"), val = tensor([0, 3, 0])]; + tensor var_515_end_0 = const()[name = tensor("op_515_end_0"), val = tensor([1, 1, 256])]; + tensor var_515_end_mask_0 = const()[name = tensor("op_515_end_mask_0"), val = tensor([true, true, true])]; + tensor var_515 = slice_by_index(begin = var_515_begin_0, end = var_515_end_0, end_mask = var_515_end_mask_0, x = x_9)[name = tensor("op_515")]; + tensor var_518_begin_0 = const()[name = tensor("op_518_begin_0"), val = tensor([0, 1, 0])]; + tensor var_518_end_0 = const()[name = tensor("op_518_end_0"), val = tensor([1, 16, 256])]; + tensor var_518_end_mask_0 = const()[name = tensor("op_518_end_mask_0"), val = tensor([true, true, true])]; + tensor var_518 = slice_by_index(begin = var_518_begin_0, end = var_518_end_0, end_mask = var_518_end_mask_0, x = window_17)[name = tensor("op_518")]; tensor window_19_interleave_0 = const()[name = tensor("window_19_interleave_0"), val = tensor(false)]; - tensor window_19 = concat(axis = var_26, interleave = window_19_interleave_0, values = (var_462, var_459))[name = tensor("window_19")]; - tensor input_61_interleave_0 = const()[name = tensor("input_61_interleave_0"), val = tensor(false)]; - tensor input_61 = concat(axis = var_23, interleave = input_61_interleave_0, values = (window_13, window_15, window_17, window_19))[name = tensor("input_61")]; + tensor window_19 = concat(axis = var_82, interleave = window_19_interleave_0, values = (var_518, var_515))[name = tensor("window_19")]; + tensor input_63_interleave_0 = const()[name = tensor("input_63_interleave_0"), val = tensor(false)]; + tensor input_63 = concat(axis = var_68, interleave = input_63_interleave_0, values = (window_13, window_15, window_17, window_19))[name = tensor("input_63")]; tensor x_11_axes_0 = const()[name = tensor("x_11_axes_0"), val = tensor([-1])]; - tensor x_11 = layer_norm(axes = x_11_axes_0, beta = encoder_conv_module_1_sequential_0_bias, epsilon = var_28, gamma = encoder_conv_module_1_sequential_0_weight, x = input_61)[name = tensor("x_11")]; - tensor input_63_perm_0 = const()[name = tensor("input_63_perm_0"), val = tensor([0, 2, 1])]; + tensor x_11 = layer_norm(axes = x_11_axes_0, beta = inner_encoder_conv_module_1_sequential_0_bias, epsilon = var_65, gamma = inner_encoder_conv_module_1_sequential_0_weight, x = input_63)[name = tensor("x_11")]; + tensor input_65_perm_0 = const()[name = tensor("input_65_perm_0"), val = tensor([0, 2, 1])]; tensor inputs_13_pad_type_0 = const()[name = tensor("inputs_13_pad_type_0"), val = tensor("valid")]; tensor inputs_13_strides_0 = const()[name = tensor("inputs_13_strides_0"), val = tensor([1])]; tensor inputs_13_pad_0 = const()[name = tensor("inputs_13_pad_0"), val = tensor([0, 0])]; tensor inputs_13_dilations_0 = const()[name = tensor("inputs_13_dilations_0"), val = tensor([1])]; tensor inputs_13_groups_0 = const()[name = tensor("inputs_13_groups_0"), val = tensor(1)]; - tensor input_63 = transpose(perm = input_63_perm_0, x = x_11)[name = tensor("transpose_47")]; - tensor inputs_13 = conv(bias = encoder_conv_module_1_sequential_2_conv_bias, dilations = inputs_13_dilations_0, groups = inputs_13_groups_0, pad = inputs_13_pad_0, pad_type = inputs_13_pad_type_0, strides = inputs_13_strides_0, weight = encoder_conv_module_1_sequential_2_conv_weight, x = input_63)[name = tensor("inputs_13")]; - tensor var_487_split_sizes_0 = const()[name = tensor("op_487_split_sizes_0"), val = tensor([256, 256])]; - tensor var_487_axis_0 = const()[name = tensor("op_487_axis_0"), val = tensor(1)]; - tensor var_487_0, tensor var_487_1 = split(axis = var_487_axis_0, split_sizes = var_487_split_sizes_0, x = inputs_13)[name = tensor("op_487")]; - tensor var_489 = sigmoid(x = var_487_1)[name = tensor("op_489")]; - tensor inputs_15 = mul(x = var_487_0, y = var_489)[name = tensor("inputs_15")]; + tensor input_65 = transpose(perm = input_65_perm_0, x = x_11)[name = tensor("transpose_47")]; + tensor inputs_13 = conv(bias = inner_encoder_conv_module_1_sequential_2_conv_bias, dilations = inputs_13_dilations_0, groups = inputs_13_groups_0, pad = inputs_13_pad_0, pad_type = inputs_13_pad_type_0, strides = inputs_13_strides_0, weight = inner_encoder_conv_module_1_sequential_2_conv_weight, x = input_65)[name = tensor("inputs_13")]; + tensor var_543_split_sizes_0 = const()[name = tensor("op_543_split_sizes_0"), val = tensor([256, 256])]; + tensor var_543_axis_0 = const()[name = tensor("op_543_axis_0"), val = tensor(1)]; + tensor var_543_0, tensor var_543_1 = split(axis = var_543_axis_0, split_sizes = var_543_split_sizes_0, x = inputs_13)[name = tensor("op_543")]; + tensor var_545 = sigmoid(x = var_543_1)[name = tensor("op_545")]; + tensor inputs_15 = mul(x = var_543_0, y = var_545)[name = tensor("inputs_15")]; tensor outputs_aug_3_pad_type_0 = const()[name = tensor("outputs_aug_3_pad_type_0"), val = tensor("custom")]; tensor outputs_aug_3_pad_0 = const()[name = tensor("outputs_aug_3_pad_0"), val = tensor([15, 15])]; tensor outputs_aug_3_groups_0 = const()[name = tensor("outputs_aug_3_groups_0"), val = tensor(256)]; tensor outputs_aug_3_strides_0 = const()[name = tensor("outputs_aug_3_strides_0"), val = tensor([1])]; tensor outputs_aug_3_dilations_0 = const()[name = tensor("outputs_aug_3_dilations_0"), val = tensor([1])]; - tensor outputs_aug_3 = conv(dilations = outputs_aug_3_dilations_0, groups = outputs_aug_3_groups_0, pad = outputs_aug_3_pad_0, pad_type = outputs_aug_3_pad_type_0, strides = outputs_aug_3_strides_0, weight = encoder_conv_module_1_sequential_4_conv_weight, x = inputs_15)[name = tensor("outputs_aug_3")]; - tensor input_65_begin_0 = const()[name = tensor("input_65_begin_0"), val = tensor([0, 0, 0])]; - tensor input_65_end_0 = const()[name = tensor("input_65_end_0"), val = tensor([4, 256, 16])]; - tensor input_65_end_mask_0 = const()[name = tensor("input_65_end_mask_0"), val = tensor([true, true, false])]; - tensor input_65 = slice_by_index(begin = input_65_begin_0, end = input_65_end_0, end_mask = input_65_end_mask_0, x = outputs_aug_3)[name = tensor("input_65")]; - tensor inputs_17 = batch_norm(beta = encoder_conv_module_1_sequential_5_bias, epsilon = var_28, gamma = encoder_conv_module_1_sequential_5_weight, mean = encoder_conv_module_1_sequential_5_running_mean, variance = encoder_conv_module_1_sequential_5_running_var, x = input_65)[name = tensor("inputs_17")]; - tensor input_67 = silu(x = inputs_17)[name = tensor("input_67")]; - tensor input_69_pad_type_0 = const()[name = tensor("input_69_pad_type_0"), val = tensor("valid")]; - tensor input_69_strides_0 = const()[name = tensor("input_69_strides_0"), val = tensor([1])]; - tensor input_69_pad_0 = const()[name = tensor("input_69_pad_0"), val = tensor([0, 0])]; - tensor input_69_dilations_0 = const()[name = tensor("input_69_dilations_0"), val = tensor([1])]; - tensor input_69_groups_0 = const()[name = tensor("input_69_groups_0"), val = tensor(1)]; - tensor input_69 = conv(bias = encoder_conv_module_1_sequential_7_conv_bias, dilations = input_69_dilations_0, groups = input_69_groups_0, pad = input_69_pad_0, pad_type = input_69_pad_type_0, strides = input_69_strides_0, weight = encoder_conv_module_1_sequential_7_conv_weight, x = input_67)[name = tensor("input_69")]; + tensor outputs_aug_3 = conv(dilations = outputs_aug_3_dilations_0, groups = outputs_aug_3_groups_0, pad = outputs_aug_3_pad_0, pad_type = outputs_aug_3_pad_type_0, strides = outputs_aug_3_strides_0, weight = inner_encoder_conv_module_1_sequential_4_conv_weight, x = inputs_15)[name = tensor("outputs_aug_3")]; + tensor input_67_begin_0 = const()[name = tensor("input_67_begin_0"), val = tensor([0, 0, 0])]; + tensor input_67_end_0 = const()[name = tensor("input_67_end_0"), val = tensor([4, 256, 16])]; + tensor input_67_end_mask_0 = const()[name = tensor("input_67_end_mask_0"), val = tensor([true, true, false])]; + tensor input_67 = slice_by_index(begin = input_67_begin_0, end = input_67_end_0, end_mask = input_67_end_mask_0, x = outputs_aug_3)[name = tensor("input_67")]; + tensor inputs_17 = batch_norm(beta = inner_encoder_conv_module_1_sequential_5_bias, epsilon = var_65, gamma = inner_encoder_conv_module_1_sequential_5_weight, mean = inner_encoder_conv_module_1_sequential_5_running_mean, variance = inner_encoder_conv_module_1_sequential_5_running_var, x = input_67)[name = tensor("inputs_17")]; + tensor input_69 = silu(x = inputs_17)[name = tensor("input_69")]; + tensor input_71_pad_type_0 = const()[name = tensor("input_71_pad_type_0"), val = tensor("valid")]; + tensor input_71_strides_0 = const()[name = tensor("input_71_strides_0"), val = tensor([1])]; + tensor input_71_pad_0 = const()[name = tensor("input_71_pad_0"), val = tensor([0, 0])]; + tensor input_71_dilations_0 = const()[name = tensor("input_71_dilations_0"), val = tensor([1])]; + tensor input_71_groups_0 = const()[name = tensor("input_71_groups_0"), val = tensor(1)]; + tensor input_71 = conv(bias = inner_encoder_conv_module_1_sequential_7_conv_bias, dilations = input_71_dilations_0, groups = input_71_groups_0, pad = input_71_pad_0, pad_type = input_71_pad_type_0, strides = input_71_strides_0, weight = inner_encoder_conv_module_1_sequential_7_conv_weight, x = input_69)[name = tensor("input_71")]; tensor conv_out_3_perm_0 = const()[name = tensor("conv_out_3_perm_0"), val = tensor([0, 2, 1])]; - tensor var_520_begin_0 = const()[name = tensor("op_520_begin_0"), val = tensor([0, -1, 0])]; - tensor var_520_end_0 = const()[name = tensor("op_520_end_0"), val = tensor([4, 16, 256])]; - tensor var_520_end_mask_0 = const()[name = tensor("op_520_end_mask_0"), val = tensor([true, true, true])]; - tensor conv_out_3 = transpose(perm = conv_out_3_perm_0, x = input_69)[name = tensor("transpose_46")]; - tensor var_520 = slice_by_index(begin = var_520_begin_0, end = var_520_end_0, end_mask = var_520_end_mask_0, x = conv_out_3)[name = tensor("op_520")]; - tensor var_522_perm_0 = const()[name = tensor("op_522_perm_0"), val = tensor([1, 0, 2])]; - tensor var_522 = transpose(perm = var_522_perm_0, x = var_520)[name = tensor("transpose_45")]; - tensor input_71 = add(x = x_9, y = var_522)[name = tensor("input_71")]; - tensor input_73_axes_0 = const()[name = tensor("input_73_axes_0"), val = tensor([-1])]; - tensor input_73 = layer_norm(axes = input_73_axes_0, beta = encoder_ffn2_1_module_sequential_0_bias, epsilon = var_28, gamma = encoder_ffn2_1_module_sequential_0_weight, x = input_71)[name = tensor("input_73")]; - tensor inputs_19 = linear(bias = encoder_ffn2_1_module_sequential_1_linear_bias, weight = encoder_ffn2_1_module_sequential_1_linear_weight, x = input_73)[name = tensor("linear_17")]; - tensor input_75 = silu(x = inputs_19)[name = tensor("input_75")]; - tensor input_79 = linear(bias = encoder_ffn2_1_module_sequential_4_linear_bias, weight = encoder_ffn2_1_module_sequential_4_linear_weight, x = input_75)[name = tensor("linear_18")]; - tensor var_545 = const()[name = tensor("op_545"), val = tensor(0x1p-1)]; - tensor var_546 = mul(x = input_79, y = var_545)[name = tensor("op_546")]; - tensor input_81 = add(x = var_546, y = input_71)[name = tensor("input_81")]; - tensor input_83_axes_0 = const()[name = tensor("input_83_axes_0"), val = tensor([-1])]; - tensor input_83 = layer_norm(axes = input_83_axes_0, beta = encoder_layer_norm_1_bias, epsilon = var_28, gamma = encoder_layer_norm_1_weight, x = input_81)[name = tensor("input_83")]; + tensor var_576_begin_0 = const()[name = tensor("op_576_begin_0"), val = tensor([0, -1, 0])]; + tensor var_576_end_0 = const()[name = tensor("op_576_end_0"), val = tensor([4, 16, 256])]; + tensor var_576_end_mask_0 = const()[name = tensor("op_576_end_mask_0"), val = tensor([true, true, true])]; + tensor conv_out_3 = transpose(perm = conv_out_3_perm_0, x = input_71)[name = tensor("transpose_46")]; + tensor var_576 = slice_by_index(begin = var_576_begin_0, end = var_576_end_0, end_mask = var_576_end_mask_0, x = conv_out_3)[name = tensor("op_576")]; + tensor var_578_perm_0 = const()[name = tensor("op_578_perm_0"), val = tensor([1, 0, 2])]; + tensor var_578 = transpose(perm = var_578_perm_0, x = var_576)[name = tensor("transpose_45")]; + tensor input_73 = add(x = x_9, y = var_578)[name = tensor("input_73")]; + tensor input_75_axes_0 = const()[name = tensor("input_75_axes_0"), val = tensor([-1])]; + tensor input_75 = layer_norm(axes = input_75_axes_0, beta = inner_encoder_ffn2_1_module_sequential_0_bias, epsilon = var_65, gamma = inner_encoder_ffn2_1_module_sequential_0_weight, x = input_73)[name = tensor("input_75")]; + tensor inputs_19 = linear(bias = inner_encoder_ffn2_1_module_sequential_1_linear_bias, weight = inner_encoder_ffn2_1_module_sequential_1_linear_weight, x = input_75)[name = tensor("linear_17")]; + tensor input_77 = silu(x = inputs_19)[name = tensor("input_77")]; + tensor input_81 = linear(bias = inner_encoder_ffn2_1_module_sequential_4_linear_bias, weight = inner_encoder_ffn2_1_module_sequential_4_linear_weight, x = input_77)[name = tensor("linear_18")]; + tensor var_601 = const()[name = tensor("op_601"), val = tensor(0x1p-1)]; + tensor var_602 = mul(x = input_81, y = var_601)[name = tensor("op_602")]; + tensor input_83 = add(x = var_602, y = input_73)[name = tensor("input_83")]; tensor input_85_axes_0 = const()[name = tensor("input_85_axes_0"), val = tensor([-1])]; - tensor input_85 = layer_norm(axes = input_85_axes_0, beta = encoder_ffn1_2_module_sequential_0_bias, epsilon = var_28, gamma = encoder_ffn1_2_module_sequential_0_weight, x = input_83)[name = tensor("input_85")]; - tensor inputs_21 = linear(bias = encoder_ffn1_2_module_sequential_1_linear_bias, weight = encoder_ffn1_2_module_sequential_1_linear_weight, x = input_85)[name = tensor("linear_19")]; - tensor input_87 = silu(x = inputs_21)[name = tensor("input_87")]; - tensor input_91 = linear(bias = encoder_ffn1_2_module_sequential_4_linear_bias, weight = encoder_ffn1_2_module_sequential_4_linear_weight, x = input_87)[name = tensor("linear_20")]; - tensor var_575 = const()[name = tensor("op_575"), val = tensor(0x1p-1)]; - tensor var_576 = mul(x = input_91, y = var_575)[name = tensor("op_576")]; - tensor input_93 = add(x = var_576, y = input_83)[name = tensor("input_93")]; + tensor input_85 = layer_norm(axes = input_85_axes_0, beta = inner_encoder_layer_norm_1_bias, epsilon = var_65, gamma = inner_encoder_layer_norm_1_weight, x = input_83)[name = tensor("input_85")]; + tensor input_87_axes_0 = const()[name = tensor("input_87_axes_0"), val = tensor([-1])]; + tensor input_87 = layer_norm(axes = input_87_axes_0, beta = inner_encoder_ffn1_2_module_sequential_0_bias, epsilon = var_65, gamma = inner_encoder_ffn1_2_module_sequential_0_weight, x = input_85)[name = tensor("input_87")]; + tensor inputs_21 = linear(bias = inner_encoder_ffn1_2_module_sequential_1_linear_bias, weight = inner_encoder_ffn1_2_module_sequential_1_linear_weight, x = input_87)[name = tensor("linear_19")]; + tensor input_89 = silu(x = inputs_21)[name = tensor("input_89")]; + tensor input_93 = linear(bias = inner_encoder_ffn1_2_module_sequential_4_linear_bias, weight = inner_encoder_ffn1_2_module_sequential_4_linear_weight, x = input_89)[name = tensor("linear_20")]; + tensor var_631 = const()[name = tensor("op_631"), val = tensor(0x1p-1)]; + tensor var_632 = mul(x = input_93, y = var_631)[name = tensor("op_632")]; + tensor input_95 = add(x = var_632, y = input_85)[name = tensor("input_95")]; tensor x_13_axes_0 = const()[name = tensor("x_13_axes_0"), val = tensor([-1])]; - tensor x_13 = layer_norm(axes = x_13_axes_0, beta = encoder_ret_lns_2_bias, epsilon = var_28, gamma = encoder_ret_lns_2_weight, x = input_93)[name = tensor("x_13")]; + tensor x_13 = layer_norm(axes = x_13_axes_0, beta = inner_encoder_ret_lns_2_bias, epsilon = var_65, gamma = inner_encoder_ret_lns_2_weight, x = input_95)[name = tensor("x_13")]; tensor prev_kv_5_begin_0 = const()[name = tensor("prev_kv_5_begin_0"), val = tensor([2, 0, 0, 0, 0])]; tensor prev_kv_5_end_0 = const()[name = tensor("prev_kv_5_end_0"), val = tensor([3, 1, 4, 64, 64])]; tensor prev_kv_5_end_mask_0 = const()[name = tensor("prev_kv_5_end_mask_0"), val = tensor([false, true, true, true, true])]; @@ -593,173 +615,173 @@ program(1.0) tensor prev_scale_5_end_mask_0 = const()[name = tensor("prev_scale_5_end_mask_0"), val = tensor([false, true])]; tensor prev_scale_5_squeeze_mask_0 = const()[name = tensor("prev_scale_5_squeeze_mask_0"), val = tensor([true, false])]; tensor prev_scale_5 = slice_by_index(begin = prev_scale_5_begin_0, end = prev_scale_5_end_0, end_mask = prev_scale_5_end_mask_0, squeeze_mask = prev_scale_5_squeeze_mask_0, x = enc_scale)[name = tensor("prev_scale_5")]; - tensor var_590 = linear(bias = encoder_q_proj_2_bias, weight = encoder_q_proj_2_weight, x = x_13)[name = tensor("linear_21")]; - tensor var_591 = const()[name = tensor("op_591"), val = tensor([1, 4, 4, 64])]; - tensor var_592 = reshape(shape = var_591, x = var_590)[name = tensor("op_592")]; + tensor var_646 = linear(bias = inner_encoder_q_proj_2_bias, weight = inner_encoder_q_proj_2_weight, x = x_13)[name = tensor("linear_21")]; + tensor var_647 = const()[name = tensor("op_647"), val = tensor([1, 4, 4, 64])]; + tensor var_648 = reshape(shape = var_647, x = var_646)[name = tensor("op_648")]; tensor q_5_perm_0 = const()[name = tensor("q_5_perm_0"), val = tensor([0, 2, 1, 3])]; - tensor var_596 = linear(bias = encoder_k_proj_2_bias, weight = encoder_k_proj_2_weight, x = x_13)[name = tensor("linear_22")]; - tensor var_597 = const()[name = tensor("op_597"), val = tensor(0x1p-3)]; - tensor var_598 = mul(x = var_596, y = var_597)[name = tensor("op_598")]; - tensor var_599 = const()[name = tensor("op_599"), val = tensor([1, 4, 4, 64])]; - tensor var_600 = reshape(shape = var_599, x = var_598)[name = tensor("op_600")]; + tensor var_652 = linear(bias = inner_encoder_k_proj_2_bias, weight = inner_encoder_k_proj_2_weight, x = x_13)[name = tensor("linear_22")]; + tensor var_653 = const()[name = tensor("op_653"), val = tensor(0x1p-3)]; + tensor var_654 = mul(x = var_652, y = var_653)[name = tensor("op_654")]; + tensor var_655 = const()[name = tensor("op_655"), val = tensor([1, 4, 4, 64])]; + tensor var_656 = reshape(shape = var_655, x = var_654)[name = tensor("op_656")]; tensor k_5_perm_0 = const()[name = tensor("k_5_perm_0"), val = tensor([0, 2, 1, 3])]; - tensor var_604 = linear(bias = encoder_v_proj_2_bias, weight = encoder_v_proj_2_weight, x = x_13)[name = tensor("linear_23")]; - tensor var_605 = const()[name = tensor("op_605"), val = tensor([1, 4, 4, 64])]; - tensor var_606 = reshape(shape = var_605, x = var_604)[name = tensor("op_606")]; + tensor var_660 = linear(bias = inner_encoder_v_proj_2_bias, weight = inner_encoder_v_proj_2_weight, x = x_13)[name = tensor("linear_23")]; + tensor var_661 = const()[name = tensor("op_661"), val = tensor([1, 4, 4, 64])]; + tensor var_662 = reshape(shape = var_661, x = var_660)[name = tensor("op_662")]; tensor v_5_perm_0 = const()[name = tensor("v_5_perm_0"), val = tensor([0, 2, 1, 3])]; - tensor input_97 = linear(bias = encoder_g_proj_2_bias, weight = encoder_g_proj_2_weight, x = x_13)[name = tensor("linear_24")]; + tensor input_99 = linear(bias = inner_encoder_g_proj_2_bias, weight = inner_encoder_g_proj_2_weight, x = x_13)[name = tensor("linear_24")]; tensor sqrt_s0_5 = sqrt(x = prev_scale_5)[name = tensor("sqrt_s0_5")]; - tensor s_t_5 = add(x = prev_scale_5, y = encoder__t_index)[name = tensor("s_t_5")]; + tensor s_t_5 = add(x = prev_scale_5, y = inner_encoder__t_index)[name = tensor("s_t_5")]; tensor sqrt_s_t_5 = sqrt(x = s_t_5)[name = tensor("sqrt_s_t_5")]; tensor qk_5_transpose_x_1 = const()[name = tensor("qk_5_transpose_x_1"), val = tensor(false)]; tensor qk_5_transpose_y_1 = const()[name = tensor("qk_5_transpose_y_1"), val = tensor(true)]; - tensor k_5 = transpose(perm = k_5_perm_0, x = var_600)[name = tensor("transpose_43")]; - tensor q_5 = transpose(perm = q_5_perm_0, x = var_592)[name = tensor("transpose_44")]; + tensor k_5 = transpose(perm = k_5_perm_0, x = var_656)[name = tensor("transpose_43")]; + tensor q_5 = transpose(perm = q_5_perm_0, x = var_648)[name = tensor("transpose_44")]; tensor qk_5 = matmul(transpose_x = qk_5_transpose_x_1, transpose_y = qk_5_transpose_y_1, x = q_5, y = k_5)[name = tensor("qk_5")]; - tensor var_616 = const()[name = tensor("op_616"), val = tensor([4, 1])]; - tensor var_617 = reshape(shape = var_616, x = sqrt_s_t_5)[name = tensor("op_617")]; - tensor M_5 = real_div(x = encoder__causal_mask, y = var_617)[name = tensor("M_5")]; - tensor var_619 = mul(x = qk_5, y = M_5)[name = tensor("op_619")]; + tensor var_672 = const()[name = tensor("op_672"), val = tensor([4, 1])]; + tensor var_673 = reshape(shape = var_672, x = sqrt_s_t_5)[name = tensor("op_673")]; + tensor M_5 = real_div(x = inner_encoder__causal_mask, y = var_673)[name = tensor("M_5")]; + tensor var_675 = mul(x = qk_5, y = M_5)[name = tensor("op_675")]; tensor inner_5_transpose_x_0 = const()[name = tensor("inner_5_transpose_x_0"), val = tensor(false)]; tensor inner_5_transpose_y_0 = const()[name = tensor("inner_5_transpose_y_0"), val = tensor(false)]; - tensor v_5 = transpose(perm = v_5_perm_0, x = var_606)[name = tensor("transpose_42")]; - tensor inner_5 = matmul(transpose_x = inner_5_transpose_x_0, transpose_y = inner_5_transpose_y_0, x = var_619, y = v_5)[name = tensor("inner_5")]; - tensor var_621_transpose_x_0 = const()[name = tensor("op_621_transpose_x_0"), val = tensor(false)]; - tensor var_621_transpose_y_0 = const()[name = tensor("op_621_transpose_y_0"), val = tensor(false)]; - tensor var_621 = matmul(transpose_x = var_621_transpose_x_0, transpose_y = var_621_transpose_y_0, x = q_5, y = prev_kv_5)[name = tensor("op_621")]; - tensor var_622 = real_div(x = sqrt_s0_5, y = sqrt_s_t_5)[name = tensor("op_622")]; - tensor var_623 = const()[name = tensor("op_623"), val = tensor([1, 1, 4, 1])]; - tensor var_624 = reshape(shape = var_623, x = var_622)[name = tensor("op_624")]; - tensor cross_5 = mul(x = var_621, y = var_624)[name = tensor("cross_5")]; + tensor v_5 = transpose(perm = v_5_perm_0, x = var_662)[name = tensor("transpose_42")]; + tensor inner_5 = matmul(transpose_x = inner_5_transpose_x_0, transpose_y = inner_5_transpose_y_0, x = var_675, y = v_5)[name = tensor("inner_5")]; + tensor var_677_transpose_x_0 = const()[name = tensor("op_677_transpose_x_0"), val = tensor(false)]; + tensor var_677_transpose_y_0 = const()[name = tensor("op_677_transpose_y_0"), val = tensor(false)]; + tensor var_677 = matmul(transpose_x = var_677_transpose_x_0, transpose_y = var_677_transpose_y_0, x = q_5, y = prev_kv_5)[name = tensor("op_677")]; + tensor var_678 = real_div(x = sqrt_s0_5, y = sqrt_s_t_5)[name = tensor("op_678")]; + tensor var_679 = const()[name = tensor("op_679"), val = tensor([1, 1, 4, 1])]; + tensor var_680 = reshape(shape = var_679, x = var_678)[name = tensor("op_680")]; + tensor cross_5 = mul(x = var_677, y = var_680)[name = tensor("cross_5")]; tensor out_13 = add(x = inner_5, y = cross_5)[name = tensor("out_13")]; - tensor var_627 = mul(x = prev_kv_5, y = sqrt_s0_5)[name = tensor("op_627")]; - tensor var_629_transpose_x_1 = const()[name = tensor("op_629_transpose_x_1"), val = tensor(true)]; - tensor var_629_transpose_y_1 = const()[name = tensor("op_629_transpose_y_1"), val = tensor(false)]; - tensor var_629 = matmul(transpose_x = var_629_transpose_x_1, transpose_y = var_629_transpose_y_1, x = k_5, y = v_5)[name = tensor("op_629")]; - tensor new_kv_unnorm_5 = add(x = var_627, y = var_629)[name = tensor("new_kv_unnorm_5")]; - tensor var_631 = const()[name = tensor("op_631"), val = tensor(0x1p+2)]; - tensor new_scale_5 = add(x = prev_scale_5, y = var_631)[name = tensor("new_scale_5")]; - tensor var_633 = sqrt(x = new_scale_5)[name = tensor("op_633")]; - tensor var_634 = real_div(x = new_kv_unnorm_5, y = var_633)[name = tensor("op_634")]; - tensor var_635_perm_0 = const()[name = tensor("op_635_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_683 = mul(x = prev_kv_5, y = sqrt_s0_5)[name = tensor("op_683")]; + tensor var_685_transpose_x_1 = const()[name = tensor("op_685_transpose_x_1"), val = tensor(true)]; + tensor var_685_transpose_y_1 = const()[name = tensor("op_685_transpose_y_1"), val = tensor(false)]; + tensor var_685 = matmul(transpose_x = var_685_transpose_x_1, transpose_y = var_685_transpose_y_1, x = k_5, y = v_5)[name = tensor("op_685")]; + tensor new_kv_unnorm_5 = add(x = var_683, y = var_685)[name = tensor("new_kv_unnorm_5")]; + tensor var_687 = const()[name = tensor("op_687"), val = tensor(0x1p+2)]; + tensor new_scale_5 = add(x = prev_scale_5, y = var_687)[name = tensor("new_scale_5")]; + tensor var_689 = sqrt(x = new_scale_5)[name = tensor("op_689")]; + tensor var_690 = real_div(x = new_kv_unnorm_5, y = var_689)[name = tensor("op_690")]; + tensor var_691_perm_0 = const()[name = tensor("op_691_perm_0"), val = tensor([0, 2, 1, 3])]; tensor out_15_axes_0 = const()[name = tensor("out_15_axes_0"), val = tensor([-1])]; - tensor var_635 = transpose(perm = var_635_perm_0, x = out_13)[name = tensor("transpose_41")]; - tensor out_15 = layer_norm(axes = out_15_axes_0, epsilon = var_17, x = var_635)[name = tensor("out_15")]; - tensor var_639 = const()[name = tensor("op_639"), val = tensor([1, 4, 256])]; - tensor out_17 = reshape(shape = var_639, x = out_15)[name = tensor("out_17")]; - tensor var_641 = silu(x = input_97)[name = tensor("op_641")]; - tensor input_99 = mul(x = var_641, y = out_17)[name = tensor("input_99")]; - tensor ret_out_5 = linear(bias = encoder_out_proj_2_bias, weight = encoder_out_proj_2_weight, x = input_99)[name = tensor("linear_25")]; - tensor x_15 = add(x = input_93, y = ret_out_5)[name = tensor("x_15")]; + tensor var_691 = transpose(perm = var_691_perm_0, x = out_13)[name = tensor("transpose_41")]; + tensor out_15 = layer_norm(axes = out_15_axes_0, epsilon = var_73, x = var_691)[name = tensor("out_15")]; + tensor var_695 = const()[name = tensor("op_695"), val = tensor([1, 4, 256])]; + tensor out_17 = reshape(shape = var_695, x = out_15)[name = tensor("out_17")]; + tensor var_697 = silu(x = input_99)[name = tensor("op_697")]; + tensor input_101 = mul(x = var_697, y = out_17)[name = tensor("input_101")]; + tensor ret_out_5 = linear(bias = inner_encoder_out_proj_2_bias, weight = inner_encoder_out_proj_2_weight, x = input_101)[name = tensor("linear_25")]; + tensor x_15 = add(x = input_95, y = ret_out_5)[name = tensor("x_15")]; tensor window_21_begin_0 = const()[name = tensor("window_21_begin_0"), val = tensor([2, 0, 0, 0])]; tensor window_21_end_0 = const()[name = tensor("window_21_end_0"), val = tensor([3, 1, 16, 256])]; tensor window_21_end_mask_0 = const()[name = tensor("window_21_end_mask_0"), val = tensor([false, true, true, true])]; tensor window_21_squeeze_mask_0 = const()[name = tensor("window_21_squeeze_mask_0"), val = tensor([true, false, false, false])]; tensor window_21 = slice_by_index(begin = window_21_begin_0, end = window_21_end_0, end_mask = window_21_end_mask_0, squeeze_mask = window_21_squeeze_mask_0, x = enc_conv_cache)[name = tensor("window_21")]; - tensor var_649_begin_0 = const()[name = tensor("op_649_begin_0"), val = tensor([0, 0, 0])]; - tensor var_649_end_0 = const()[name = tensor("op_649_end_0"), val = tensor([1, 1, 256])]; - tensor var_649_end_mask_0 = const()[name = tensor("op_649_end_mask_0"), val = tensor([true, false, true])]; - tensor var_649 = slice_by_index(begin = var_649_begin_0, end = var_649_end_0, end_mask = var_649_end_mask_0, x = x_15)[name = tensor("op_649")]; - tensor var_652_begin_0 = const()[name = tensor("op_652_begin_0"), val = tensor([0, 1, 0])]; - tensor var_652_end_0 = const()[name = tensor("op_652_end_0"), val = tensor([1, 16, 256])]; - tensor var_652_end_mask_0 = const()[name = tensor("op_652_end_mask_0"), val = tensor([true, true, true])]; - tensor var_652 = slice_by_index(begin = var_652_begin_0, end = var_652_end_0, end_mask = var_652_end_mask_0, x = window_21)[name = tensor("op_652")]; + tensor var_705_begin_0 = const()[name = tensor("op_705_begin_0"), val = tensor([0, 0, 0])]; + tensor var_705_end_0 = const()[name = tensor("op_705_end_0"), val = tensor([1, 1, 256])]; + tensor var_705_end_mask_0 = const()[name = tensor("op_705_end_mask_0"), val = tensor([true, false, true])]; + tensor var_705 = slice_by_index(begin = var_705_begin_0, end = var_705_end_0, end_mask = var_705_end_mask_0, x = x_15)[name = tensor("op_705")]; + tensor var_708_begin_0 = const()[name = tensor("op_708_begin_0"), val = tensor([0, 1, 0])]; + tensor var_708_end_0 = const()[name = tensor("op_708_end_0"), val = tensor([1, 16, 256])]; + tensor var_708_end_mask_0 = const()[name = tensor("op_708_end_mask_0"), val = tensor([true, true, true])]; + tensor var_708 = slice_by_index(begin = var_708_begin_0, end = var_708_end_0, end_mask = var_708_end_mask_0, x = window_21)[name = tensor("op_708")]; tensor window_23_interleave_0 = const()[name = tensor("window_23_interleave_0"), val = tensor(false)]; - tensor window_23 = concat(axis = var_26, interleave = window_23_interleave_0, values = (var_652, var_649))[name = tensor("window_23")]; - tensor var_657_begin_0 = const()[name = tensor("op_657_begin_0"), val = tensor([0, 1, 0])]; - tensor var_657_end_0 = const()[name = tensor("op_657_end_0"), val = tensor([1, 2, 256])]; - tensor var_657_end_mask_0 = const()[name = tensor("op_657_end_mask_0"), val = tensor([true, false, true])]; - tensor var_657 = slice_by_index(begin = var_657_begin_0, end = var_657_end_0, end_mask = var_657_end_mask_0, x = x_15)[name = tensor("op_657")]; - tensor var_660_begin_0 = const()[name = tensor("op_660_begin_0"), val = tensor([0, 1, 0])]; - tensor var_660_end_0 = const()[name = tensor("op_660_end_0"), val = tensor([1, 16, 256])]; - tensor var_660_end_mask_0 = const()[name = tensor("op_660_end_mask_0"), val = tensor([true, true, true])]; - tensor var_660 = slice_by_index(begin = var_660_begin_0, end = var_660_end_0, end_mask = var_660_end_mask_0, x = window_23)[name = tensor("op_660")]; + tensor window_23 = concat(axis = var_82, interleave = window_23_interleave_0, values = (var_708, var_705))[name = tensor("window_23")]; + tensor var_713_begin_0 = const()[name = tensor("op_713_begin_0"), val = tensor([0, 1, 0])]; + tensor var_713_end_0 = const()[name = tensor("op_713_end_0"), val = tensor([1, 2, 256])]; + tensor var_713_end_mask_0 = const()[name = tensor("op_713_end_mask_0"), val = tensor([true, false, true])]; + tensor var_713 = slice_by_index(begin = var_713_begin_0, end = var_713_end_0, end_mask = var_713_end_mask_0, x = x_15)[name = tensor("op_713")]; + tensor var_716_begin_0 = const()[name = tensor("op_716_begin_0"), val = tensor([0, 1, 0])]; + tensor var_716_end_0 = const()[name = tensor("op_716_end_0"), val = tensor([1, 16, 256])]; + tensor var_716_end_mask_0 = const()[name = tensor("op_716_end_mask_0"), val = tensor([true, true, true])]; + tensor var_716 = slice_by_index(begin = var_716_begin_0, end = var_716_end_0, end_mask = var_716_end_mask_0, x = window_23)[name = tensor("op_716")]; tensor window_25_interleave_0 = const()[name = tensor("window_25_interleave_0"), val = tensor(false)]; - tensor window_25 = concat(axis = var_26, interleave = window_25_interleave_0, values = (var_660, var_657))[name = tensor("window_25")]; - tensor var_665_begin_0 = const()[name = tensor("op_665_begin_0"), val = tensor([0, 2, 0])]; - tensor var_665_end_0 = const()[name = tensor("op_665_end_0"), val = tensor([1, 3, 256])]; - tensor var_665_end_mask_0 = const()[name = tensor("op_665_end_mask_0"), val = tensor([true, false, true])]; - tensor var_665 = slice_by_index(begin = var_665_begin_0, end = var_665_end_0, end_mask = var_665_end_mask_0, x = x_15)[name = tensor("op_665")]; - tensor var_668_begin_0 = const()[name = tensor("op_668_begin_0"), val = tensor([0, 1, 0])]; - tensor var_668_end_0 = const()[name = tensor("op_668_end_0"), val = tensor([1, 16, 256])]; - tensor var_668_end_mask_0 = const()[name = tensor("op_668_end_mask_0"), val = tensor([true, true, true])]; - tensor var_668 = slice_by_index(begin = var_668_begin_0, end = var_668_end_0, end_mask = var_668_end_mask_0, x = window_25)[name = tensor("op_668")]; + tensor window_25 = concat(axis = var_82, interleave = window_25_interleave_0, values = (var_716, var_713))[name = tensor("window_25")]; + tensor var_721_begin_0 = const()[name = tensor("op_721_begin_0"), val = tensor([0, 2, 0])]; + tensor var_721_end_0 = const()[name = tensor("op_721_end_0"), val = tensor([1, 3, 256])]; + tensor var_721_end_mask_0 = const()[name = tensor("op_721_end_mask_0"), val = tensor([true, false, true])]; + tensor var_721 = slice_by_index(begin = var_721_begin_0, end = var_721_end_0, end_mask = var_721_end_mask_0, x = x_15)[name = tensor("op_721")]; + tensor var_724_begin_0 = const()[name = tensor("op_724_begin_0"), val = tensor([0, 1, 0])]; + tensor var_724_end_0 = const()[name = tensor("op_724_end_0"), val = tensor([1, 16, 256])]; + tensor var_724_end_mask_0 = const()[name = tensor("op_724_end_mask_0"), val = tensor([true, true, true])]; + tensor var_724 = slice_by_index(begin = var_724_begin_0, end = var_724_end_0, end_mask = var_724_end_mask_0, x = window_25)[name = tensor("op_724")]; tensor window_27_interleave_0 = const()[name = tensor("window_27_interleave_0"), val = tensor(false)]; - tensor window_27 = concat(axis = var_26, interleave = window_27_interleave_0, values = (var_668, var_665))[name = tensor("window_27")]; - tensor var_673_begin_0 = const()[name = tensor("op_673_begin_0"), val = tensor([0, 3, 0])]; - tensor var_673_end_0 = const()[name = tensor("op_673_end_0"), val = tensor([1, 1, 256])]; - tensor var_673_end_mask_0 = const()[name = tensor("op_673_end_mask_0"), val = tensor([true, true, true])]; - tensor var_673 = slice_by_index(begin = var_673_begin_0, end = var_673_end_0, end_mask = var_673_end_mask_0, x = x_15)[name = tensor("op_673")]; - tensor var_676_begin_0 = const()[name = tensor("op_676_begin_0"), val = tensor([0, 1, 0])]; - tensor var_676_end_0 = const()[name = tensor("op_676_end_0"), val = tensor([1, 16, 256])]; - tensor var_676_end_mask_0 = const()[name = tensor("op_676_end_mask_0"), val = tensor([true, true, true])]; - tensor var_676 = slice_by_index(begin = var_676_begin_0, end = var_676_end_0, end_mask = var_676_end_mask_0, x = window_27)[name = tensor("op_676")]; + tensor window_27 = concat(axis = var_82, interleave = window_27_interleave_0, values = (var_724, var_721))[name = tensor("window_27")]; + tensor var_729_begin_0 = const()[name = tensor("op_729_begin_0"), val = tensor([0, 3, 0])]; + tensor var_729_end_0 = const()[name = tensor("op_729_end_0"), val = tensor([1, 1, 256])]; + tensor var_729_end_mask_0 = const()[name = tensor("op_729_end_mask_0"), val = tensor([true, true, true])]; + tensor var_729 = slice_by_index(begin = var_729_begin_0, end = var_729_end_0, end_mask = var_729_end_mask_0, x = x_15)[name = tensor("op_729")]; + tensor var_732_begin_0 = const()[name = tensor("op_732_begin_0"), val = tensor([0, 1, 0])]; + tensor var_732_end_0 = const()[name = tensor("op_732_end_0"), val = tensor([1, 16, 256])]; + tensor var_732_end_mask_0 = const()[name = tensor("op_732_end_mask_0"), val = tensor([true, true, true])]; + tensor var_732 = slice_by_index(begin = var_732_begin_0, end = var_732_end_0, end_mask = var_732_end_mask_0, x = window_27)[name = tensor("op_732")]; tensor window_29_interleave_0 = const()[name = tensor("window_29_interleave_0"), val = tensor(false)]; - tensor window_29 = concat(axis = var_26, interleave = window_29_interleave_0, values = (var_676, var_673))[name = tensor("window_29")]; - tensor input_101_interleave_0 = const()[name = tensor("input_101_interleave_0"), val = tensor(false)]; - tensor input_101 = concat(axis = var_23, interleave = input_101_interleave_0, values = (window_23, window_25, window_27, window_29))[name = tensor("input_101")]; + tensor window_29 = concat(axis = var_82, interleave = window_29_interleave_0, values = (var_732, var_729))[name = tensor("window_29")]; + tensor input_103_interleave_0 = const()[name = tensor("input_103_interleave_0"), val = tensor(false)]; + tensor input_103 = concat(axis = var_68, interleave = input_103_interleave_0, values = (window_23, window_25, window_27, window_29))[name = tensor("input_103")]; tensor x_17_axes_0 = const()[name = tensor("x_17_axes_0"), val = tensor([-1])]; - tensor x_17 = layer_norm(axes = x_17_axes_0, beta = encoder_conv_module_2_sequential_0_bias, epsilon = var_28, gamma = encoder_conv_module_2_sequential_0_weight, x = input_101)[name = tensor("x_17")]; - tensor input_103_perm_0 = const()[name = tensor("input_103_perm_0"), val = tensor([0, 2, 1])]; + tensor x_17 = layer_norm(axes = x_17_axes_0, beta = inner_encoder_conv_module_2_sequential_0_bias, epsilon = var_65, gamma = inner_encoder_conv_module_2_sequential_0_weight, x = input_103)[name = tensor("x_17")]; + tensor input_105_perm_0 = const()[name = tensor("input_105_perm_0"), val = tensor([0, 2, 1])]; tensor inputs_23_pad_type_0 = const()[name = tensor("inputs_23_pad_type_0"), val = tensor("valid")]; tensor inputs_23_strides_0 = const()[name = tensor("inputs_23_strides_0"), val = tensor([1])]; tensor inputs_23_pad_0 = const()[name = tensor("inputs_23_pad_0"), val = tensor([0, 0])]; tensor inputs_23_dilations_0 = const()[name = tensor("inputs_23_dilations_0"), val = tensor([1])]; tensor inputs_23_groups_0 = const()[name = tensor("inputs_23_groups_0"), val = tensor(1)]; - tensor input_103 = transpose(perm = input_103_perm_0, x = x_17)[name = tensor("transpose_40")]; - tensor inputs_23 = conv(bias = encoder_conv_module_2_sequential_2_conv_bias, dilations = inputs_23_dilations_0, groups = inputs_23_groups_0, pad = inputs_23_pad_0, pad_type = inputs_23_pad_type_0, strides = inputs_23_strides_0, weight = encoder_conv_module_2_sequential_2_conv_weight, x = input_103)[name = tensor("inputs_23")]; - tensor var_701_split_sizes_0 = const()[name = tensor("op_701_split_sizes_0"), val = tensor([256, 256])]; - tensor var_701_axis_0 = const()[name = tensor("op_701_axis_0"), val = tensor(1)]; - tensor var_701_0, tensor var_701_1 = split(axis = var_701_axis_0, split_sizes = var_701_split_sizes_0, x = inputs_23)[name = tensor("op_701")]; - tensor var_703 = sigmoid(x = var_701_1)[name = tensor("op_703")]; - tensor inputs_25 = mul(x = var_701_0, y = var_703)[name = tensor("inputs_25")]; + tensor input_105 = transpose(perm = input_105_perm_0, x = x_17)[name = tensor("transpose_40")]; + tensor inputs_23 = conv(bias = inner_encoder_conv_module_2_sequential_2_conv_bias, dilations = inputs_23_dilations_0, groups = inputs_23_groups_0, pad = inputs_23_pad_0, pad_type = inputs_23_pad_type_0, strides = inputs_23_strides_0, weight = inner_encoder_conv_module_2_sequential_2_conv_weight, x = input_105)[name = tensor("inputs_23")]; + tensor var_757_split_sizes_0 = const()[name = tensor("op_757_split_sizes_0"), val = tensor([256, 256])]; + tensor var_757_axis_0 = const()[name = tensor("op_757_axis_0"), val = tensor(1)]; + tensor var_757_0, tensor var_757_1 = split(axis = var_757_axis_0, split_sizes = var_757_split_sizes_0, x = inputs_23)[name = tensor("op_757")]; + tensor var_759 = sigmoid(x = var_757_1)[name = tensor("op_759")]; + tensor inputs_25 = mul(x = var_757_0, y = var_759)[name = tensor("inputs_25")]; tensor outputs_aug_5_pad_type_0 = const()[name = tensor("outputs_aug_5_pad_type_0"), val = tensor("custom")]; tensor outputs_aug_5_pad_0 = const()[name = tensor("outputs_aug_5_pad_0"), val = tensor([15, 15])]; tensor outputs_aug_5_groups_0 = const()[name = tensor("outputs_aug_5_groups_0"), val = tensor(256)]; tensor outputs_aug_5_strides_0 = const()[name = tensor("outputs_aug_5_strides_0"), val = tensor([1])]; tensor outputs_aug_5_dilations_0 = const()[name = tensor("outputs_aug_5_dilations_0"), val = tensor([1])]; - tensor outputs_aug_5 = conv(dilations = outputs_aug_5_dilations_0, groups = outputs_aug_5_groups_0, pad = outputs_aug_5_pad_0, pad_type = outputs_aug_5_pad_type_0, strides = outputs_aug_5_strides_0, weight = encoder_conv_module_2_sequential_4_conv_weight, x = inputs_25)[name = tensor("outputs_aug_5")]; - tensor input_105_begin_0 = const()[name = tensor("input_105_begin_0"), val = tensor([0, 0, 0])]; - tensor input_105_end_0 = const()[name = tensor("input_105_end_0"), val = tensor([4, 256, 16])]; - tensor input_105_end_mask_0 = const()[name = tensor("input_105_end_mask_0"), val = tensor([true, true, false])]; - tensor input_105 = slice_by_index(begin = input_105_begin_0, end = input_105_end_0, end_mask = input_105_end_mask_0, x = outputs_aug_5)[name = tensor("input_105")]; - tensor inputs_27 = batch_norm(beta = encoder_conv_module_2_sequential_5_bias, epsilon = var_28, gamma = encoder_conv_module_2_sequential_5_weight, mean = encoder_conv_module_2_sequential_5_running_mean, variance = encoder_conv_module_2_sequential_5_running_var, x = input_105)[name = tensor("inputs_27")]; - tensor input_107 = silu(x = inputs_27)[name = tensor("input_107")]; - tensor input_109_pad_type_0 = const()[name = tensor("input_109_pad_type_0"), val = tensor("valid")]; - tensor input_109_strides_0 = const()[name = tensor("input_109_strides_0"), val = tensor([1])]; - tensor input_109_pad_0 = const()[name = tensor("input_109_pad_0"), val = tensor([0, 0])]; - tensor input_109_dilations_0 = const()[name = tensor("input_109_dilations_0"), val = tensor([1])]; - tensor input_109_groups_0 = const()[name = tensor("input_109_groups_0"), val = tensor(1)]; - tensor input_109 = conv(bias = encoder_conv_module_2_sequential_7_conv_bias, dilations = input_109_dilations_0, groups = input_109_groups_0, pad = input_109_pad_0, pad_type = input_109_pad_type_0, strides = input_109_strides_0, weight = encoder_conv_module_2_sequential_7_conv_weight, x = input_107)[name = tensor("input_109")]; + tensor outputs_aug_5 = conv(dilations = outputs_aug_5_dilations_0, groups = outputs_aug_5_groups_0, pad = outputs_aug_5_pad_0, pad_type = outputs_aug_5_pad_type_0, strides = outputs_aug_5_strides_0, weight = inner_encoder_conv_module_2_sequential_4_conv_weight, x = inputs_25)[name = tensor("outputs_aug_5")]; + tensor input_107_begin_0 = const()[name = tensor("input_107_begin_0"), val = tensor([0, 0, 0])]; + tensor input_107_end_0 = const()[name = tensor("input_107_end_0"), val = tensor([4, 256, 16])]; + tensor input_107_end_mask_0 = const()[name = tensor("input_107_end_mask_0"), val = tensor([true, true, false])]; + tensor input_107 = slice_by_index(begin = input_107_begin_0, end = input_107_end_0, end_mask = input_107_end_mask_0, x = outputs_aug_5)[name = tensor("input_107")]; + tensor inputs_27 = batch_norm(beta = inner_encoder_conv_module_2_sequential_5_bias, epsilon = var_65, gamma = inner_encoder_conv_module_2_sequential_5_weight, mean = inner_encoder_conv_module_2_sequential_5_running_mean, variance = inner_encoder_conv_module_2_sequential_5_running_var, x = input_107)[name = tensor("inputs_27")]; + tensor input_109 = silu(x = inputs_27)[name = tensor("input_109")]; + tensor input_111_pad_type_0 = const()[name = tensor("input_111_pad_type_0"), val = tensor("valid")]; + tensor input_111_strides_0 = const()[name = tensor("input_111_strides_0"), val = tensor([1])]; + tensor input_111_pad_0 = const()[name = tensor("input_111_pad_0"), val = tensor([0, 0])]; + tensor input_111_dilations_0 = const()[name = tensor("input_111_dilations_0"), val = tensor([1])]; + tensor input_111_groups_0 = const()[name = tensor("input_111_groups_0"), val = tensor(1)]; + tensor input_111 = conv(bias = inner_encoder_conv_module_2_sequential_7_conv_bias, dilations = input_111_dilations_0, groups = input_111_groups_0, pad = input_111_pad_0, pad_type = input_111_pad_type_0, strides = input_111_strides_0, weight = inner_encoder_conv_module_2_sequential_7_conv_weight, x = input_109)[name = tensor("input_111")]; tensor conv_out_5_perm_0 = const()[name = tensor("conv_out_5_perm_0"), val = tensor([0, 2, 1])]; - tensor var_734_begin_0 = const()[name = tensor("op_734_begin_0"), val = tensor([0, -1, 0])]; - tensor var_734_end_0 = const()[name = tensor("op_734_end_0"), val = tensor([4, 16, 256])]; - tensor var_734_end_mask_0 = const()[name = tensor("op_734_end_mask_0"), val = tensor([true, true, true])]; - tensor conv_out_5 = transpose(perm = conv_out_5_perm_0, x = input_109)[name = tensor("transpose_39")]; - tensor var_734 = slice_by_index(begin = var_734_begin_0, end = var_734_end_0, end_mask = var_734_end_mask_0, x = conv_out_5)[name = tensor("op_734")]; - tensor var_736_perm_0 = const()[name = tensor("op_736_perm_0"), val = tensor([1, 0, 2])]; - tensor var_736 = transpose(perm = var_736_perm_0, x = var_734)[name = tensor("transpose_38")]; - tensor input_111 = add(x = x_15, y = var_736)[name = tensor("input_111")]; - tensor input_113_axes_0 = const()[name = tensor("input_113_axes_0"), val = tensor([-1])]; - tensor input_113 = layer_norm(axes = input_113_axes_0, beta = encoder_ffn2_2_module_sequential_0_bias, epsilon = var_28, gamma = encoder_ffn2_2_module_sequential_0_weight, x = input_111)[name = tensor("input_113")]; - tensor inputs_29 = linear(bias = encoder_ffn2_2_module_sequential_1_linear_bias, weight = encoder_ffn2_2_module_sequential_1_linear_weight, x = input_113)[name = tensor("linear_26")]; - tensor input_115 = silu(x = inputs_29)[name = tensor("input_115")]; - tensor input_119 = linear(bias = encoder_ffn2_2_module_sequential_4_linear_bias, weight = encoder_ffn2_2_module_sequential_4_linear_weight, x = input_115)[name = tensor("linear_27")]; - tensor var_759 = const()[name = tensor("op_759"), val = tensor(0x1p-1)]; - tensor var_760 = mul(x = input_119, y = var_759)[name = tensor("op_760")]; - tensor input_121 = add(x = var_760, y = input_111)[name = tensor("input_121")]; - tensor input_123_axes_0 = const()[name = tensor("input_123_axes_0"), val = tensor([-1])]; - tensor input_123 = layer_norm(axes = input_123_axes_0, beta = encoder_layer_norm_2_bias, epsilon = var_28, gamma = encoder_layer_norm_2_weight, x = input_121)[name = tensor("input_123")]; + tensor var_790_begin_0 = const()[name = tensor("op_790_begin_0"), val = tensor([0, -1, 0])]; + tensor var_790_end_0 = const()[name = tensor("op_790_end_0"), val = tensor([4, 16, 256])]; + tensor var_790_end_mask_0 = const()[name = tensor("op_790_end_mask_0"), val = tensor([true, true, true])]; + tensor conv_out_5 = transpose(perm = conv_out_5_perm_0, x = input_111)[name = tensor("transpose_39")]; + tensor var_790 = slice_by_index(begin = var_790_begin_0, end = var_790_end_0, end_mask = var_790_end_mask_0, x = conv_out_5)[name = tensor("op_790")]; + tensor var_792_perm_0 = const()[name = tensor("op_792_perm_0"), val = tensor([1, 0, 2])]; + tensor var_792 = transpose(perm = var_792_perm_0, x = var_790)[name = tensor("transpose_38")]; + tensor input_113 = add(x = x_15, y = var_792)[name = tensor("input_113")]; + tensor input_115_axes_0 = const()[name = tensor("input_115_axes_0"), val = tensor([-1])]; + tensor input_115 = layer_norm(axes = input_115_axes_0, beta = inner_encoder_ffn2_2_module_sequential_0_bias, epsilon = var_65, gamma = inner_encoder_ffn2_2_module_sequential_0_weight, x = input_113)[name = tensor("input_115")]; + tensor inputs_29 = linear(bias = inner_encoder_ffn2_2_module_sequential_1_linear_bias, weight = inner_encoder_ffn2_2_module_sequential_1_linear_weight, x = input_115)[name = tensor("linear_26")]; + tensor input_117 = silu(x = inputs_29)[name = tensor("input_117")]; + tensor input_121 = linear(bias = inner_encoder_ffn2_2_module_sequential_4_linear_bias, weight = inner_encoder_ffn2_2_module_sequential_4_linear_weight, x = input_117)[name = tensor("linear_27")]; + tensor var_815 = const()[name = tensor("op_815"), val = tensor(0x1p-1)]; + tensor var_816 = mul(x = input_121, y = var_815)[name = tensor("op_816")]; + tensor input_123 = add(x = var_816, y = input_113)[name = tensor("input_123")]; tensor input_125_axes_0 = const()[name = tensor("input_125_axes_0"), val = tensor([-1])]; - tensor input_125 = layer_norm(axes = input_125_axes_0, beta = encoder_ffn1_3_module_sequential_0_bias, epsilon = var_28, gamma = encoder_ffn1_3_module_sequential_0_weight, x = input_123)[name = tensor("input_125")]; - tensor inputs_31 = linear(bias = encoder_ffn1_3_module_sequential_1_linear_bias, weight = encoder_ffn1_3_module_sequential_1_linear_weight, x = input_125)[name = tensor("linear_28")]; - tensor input_127 = silu(x = inputs_31)[name = tensor("input_127")]; - tensor input_131 = linear(bias = encoder_ffn1_3_module_sequential_4_linear_bias, weight = encoder_ffn1_3_module_sequential_4_linear_weight, x = input_127)[name = tensor("linear_29")]; - tensor var_789 = const()[name = tensor("op_789"), val = tensor(0x1p-1)]; - tensor var_790 = mul(x = input_131, y = var_789)[name = tensor("op_790")]; - tensor input_133 = add(x = var_790, y = input_123)[name = tensor("input_133")]; + tensor input_125 = layer_norm(axes = input_125_axes_0, beta = inner_encoder_layer_norm_2_bias, epsilon = var_65, gamma = inner_encoder_layer_norm_2_weight, x = input_123)[name = tensor("input_125")]; + tensor input_127_axes_0 = const()[name = tensor("input_127_axes_0"), val = tensor([-1])]; + tensor input_127 = layer_norm(axes = input_127_axes_0, beta = inner_encoder_ffn1_3_module_sequential_0_bias, epsilon = var_65, gamma = inner_encoder_ffn1_3_module_sequential_0_weight, x = input_125)[name = tensor("input_127")]; + tensor inputs_31 = linear(bias = inner_encoder_ffn1_3_module_sequential_1_linear_bias, weight = inner_encoder_ffn1_3_module_sequential_1_linear_weight, x = input_127)[name = tensor("linear_28")]; + tensor input_129 = silu(x = inputs_31)[name = tensor("input_129")]; + tensor input_133 = linear(bias = inner_encoder_ffn1_3_module_sequential_4_linear_bias, weight = inner_encoder_ffn1_3_module_sequential_4_linear_weight, x = input_129)[name = tensor("linear_29")]; + tensor var_845 = const()[name = tensor("op_845"), val = tensor(0x1p-1)]; + tensor var_846 = mul(x = input_133, y = var_845)[name = tensor("op_846")]; + tensor input_135 = add(x = var_846, y = input_125)[name = tensor("input_135")]; tensor x_19_axes_0 = const()[name = tensor("x_19_axes_0"), val = tensor([-1])]; - tensor x_19 = layer_norm(axes = x_19_axes_0, beta = encoder_ret_lns_3_bias, epsilon = var_28, gamma = encoder_ret_lns_3_weight, x = input_133)[name = tensor("x_19")]; + tensor x_19 = layer_norm(axes = x_19_axes_0, beta = inner_encoder_ret_lns_3_bias, epsilon = var_65, gamma = inner_encoder_ret_lns_3_weight, x = input_135)[name = tensor("x_19")]; tensor prev_kv_7_begin_0 = const()[name = tensor("prev_kv_7_begin_0"), val = tensor([3, 0, 0, 0, 0])]; tensor prev_kv_7_end_0 = const()[name = tensor("prev_kv_7_end_0"), val = tensor([4, 1, 4, 64, 64])]; tensor prev_kv_7_end_mask_0 = const()[name = tensor("prev_kv_7_end_mask_0"), val = tensor([false, true, true, true, true])]; @@ -770,209 +792,202 @@ program(1.0) tensor prev_scale_7_end_mask_0 = const()[name = tensor("prev_scale_7_end_mask_0"), val = tensor([false, true])]; tensor prev_scale_7_squeeze_mask_0 = const()[name = tensor("prev_scale_7_squeeze_mask_0"), val = tensor([true, false])]; tensor prev_scale_7 = slice_by_index(begin = prev_scale_7_begin_0, end = prev_scale_7_end_0, end_mask = prev_scale_7_end_mask_0, squeeze_mask = prev_scale_7_squeeze_mask_0, x = enc_scale)[name = tensor("prev_scale_7")]; - tensor var_804 = linear(bias = encoder_q_proj_3_bias, weight = encoder_q_proj_3_weight, x = x_19)[name = tensor("linear_30")]; - tensor var_805 = const()[name = tensor("op_805"), val = tensor([1, 4, 4, 64])]; - tensor var_806 = reshape(shape = var_805, x = var_804)[name = tensor("op_806")]; + tensor var_860 = linear(bias = inner_encoder_q_proj_3_bias, weight = inner_encoder_q_proj_3_weight, x = x_19)[name = tensor("linear_30")]; + tensor var_861 = const()[name = tensor("op_861"), val = tensor([1, 4, 4, 64])]; + tensor var_862 = reshape(shape = var_861, x = var_860)[name = tensor("op_862")]; tensor q_7_perm_0 = const()[name = tensor("q_7_perm_0"), val = tensor([0, 2, 1, 3])]; - tensor var_810 = linear(bias = encoder_k_proj_3_bias, weight = encoder_k_proj_3_weight, x = x_19)[name = tensor("linear_31")]; - tensor var_811 = const()[name = tensor("op_811"), val = tensor(0x1p-3)]; - tensor var_812 = mul(x = var_810, y = var_811)[name = tensor("op_812")]; - tensor var_813 = const()[name = tensor("op_813"), val = tensor([1, 4, 4, 64])]; - tensor var_814 = reshape(shape = var_813, x = var_812)[name = tensor("op_814")]; + tensor var_866 = linear(bias = inner_encoder_k_proj_3_bias, weight = inner_encoder_k_proj_3_weight, x = x_19)[name = tensor("linear_31")]; + tensor var_867 = const()[name = tensor("op_867"), val = tensor(0x1p-3)]; + tensor var_868 = mul(x = var_866, y = var_867)[name = tensor("op_868")]; + tensor var_869 = const()[name = tensor("op_869"), val = tensor([1, 4, 4, 64])]; + tensor var_870 = reshape(shape = var_869, x = var_868)[name = tensor("op_870")]; tensor k_7_perm_0 = const()[name = tensor("k_7_perm_0"), val = tensor([0, 2, 1, 3])]; - tensor var_818 = linear(bias = encoder_v_proj_3_bias, weight = encoder_v_proj_3_weight, x = x_19)[name = tensor("linear_32")]; - tensor var_819 = const()[name = tensor("op_819"), val = tensor([1, 4, 4, 64])]; - tensor var_820 = reshape(shape = var_819, x = var_818)[name = tensor("op_820")]; + tensor var_874 = linear(bias = inner_encoder_v_proj_3_bias, weight = inner_encoder_v_proj_3_weight, x = x_19)[name = tensor("linear_32")]; + tensor var_875 = const()[name = tensor("op_875"), val = tensor([1, 4, 4, 64])]; + tensor var_876 = reshape(shape = var_875, x = var_874)[name = tensor("op_876")]; tensor v_7_perm_0 = const()[name = tensor("v_7_perm_0"), val = tensor([0, 2, 1, 3])]; - tensor input_137 = linear(bias = encoder_g_proj_3_bias, weight = encoder_g_proj_3_weight, x = x_19)[name = tensor("linear_33")]; + tensor input_139 = linear(bias = inner_encoder_g_proj_3_bias, weight = inner_encoder_g_proj_3_weight, x = x_19)[name = tensor("linear_33")]; tensor sqrt_s0_7 = sqrt(x = prev_scale_7)[name = tensor("sqrt_s0_7")]; - tensor s_t_7 = add(x = prev_scale_7, y = encoder__t_index)[name = tensor("s_t_7")]; + tensor s_t_7 = add(x = prev_scale_7, y = inner_encoder__t_index)[name = tensor("s_t_7")]; tensor sqrt_s_t_7 = sqrt(x = s_t_7)[name = tensor("sqrt_s_t_7")]; tensor qk_7_transpose_x_1 = const()[name = tensor("qk_7_transpose_x_1"), val = tensor(false)]; tensor qk_7_transpose_y_1 = const()[name = tensor("qk_7_transpose_y_1"), val = tensor(true)]; - tensor k_7 = transpose(perm = k_7_perm_0, x = var_814)[name = tensor("transpose_36")]; - tensor q_7 = transpose(perm = q_7_perm_0, x = var_806)[name = tensor("transpose_37")]; + tensor k_7 = transpose(perm = k_7_perm_0, x = var_870)[name = tensor("transpose_36")]; + tensor q_7 = transpose(perm = q_7_perm_0, x = var_862)[name = tensor("transpose_37")]; tensor qk_7 = matmul(transpose_x = qk_7_transpose_x_1, transpose_y = qk_7_transpose_y_1, x = q_7, y = k_7)[name = tensor("qk_7")]; - tensor var_830 = const()[name = tensor("op_830"), val = tensor([4, 1])]; - tensor var_831 = reshape(shape = var_830, x = sqrt_s_t_7)[name = tensor("op_831")]; - tensor M_7 = real_div(x = encoder__causal_mask, y = var_831)[name = tensor("M_7")]; - tensor var_833 = mul(x = qk_7, y = M_7)[name = tensor("op_833")]; + tensor var_886 = const()[name = tensor("op_886"), val = tensor([4, 1])]; + tensor var_887 = reshape(shape = var_886, x = sqrt_s_t_7)[name = tensor("op_887")]; + tensor M_7 = real_div(x = inner_encoder__causal_mask, y = var_887)[name = tensor("M_7")]; + tensor var_889 = mul(x = qk_7, y = M_7)[name = tensor("op_889")]; tensor inner_7_transpose_x_0 = const()[name = tensor("inner_7_transpose_x_0"), val = tensor(false)]; tensor inner_7_transpose_y_0 = const()[name = tensor("inner_7_transpose_y_0"), val = tensor(false)]; - tensor v_7 = transpose(perm = v_7_perm_0, x = var_820)[name = tensor("transpose_35")]; - tensor inner_7 = matmul(transpose_x = inner_7_transpose_x_0, transpose_y = inner_7_transpose_y_0, x = var_833, y = v_7)[name = tensor("inner_7")]; - tensor var_835_transpose_x_0 = const()[name = tensor("op_835_transpose_x_0"), val = tensor(false)]; - tensor var_835_transpose_y_0 = const()[name = tensor("op_835_transpose_y_0"), val = tensor(false)]; - tensor var_835 = matmul(transpose_x = var_835_transpose_x_0, transpose_y = var_835_transpose_y_0, x = q_7, y = prev_kv_7)[name = tensor("op_835")]; - tensor var_836 = real_div(x = sqrt_s0_7, y = sqrt_s_t_7)[name = tensor("op_836")]; - tensor var_837 = const()[name = tensor("op_837"), val = tensor([1, 1, 4, 1])]; - tensor var_838 = reshape(shape = var_837, x = var_836)[name = tensor("op_838")]; - tensor cross_7 = mul(x = var_835, y = var_838)[name = tensor("cross_7")]; + tensor v_7 = transpose(perm = v_7_perm_0, x = var_876)[name = tensor("transpose_35")]; + tensor inner_7 = matmul(transpose_x = inner_7_transpose_x_0, transpose_y = inner_7_transpose_y_0, x = var_889, y = v_7)[name = tensor("inner_7")]; + tensor var_891_transpose_x_0 = const()[name = tensor("op_891_transpose_x_0"), val = tensor(false)]; + tensor var_891_transpose_y_0 = const()[name = tensor("op_891_transpose_y_0"), val = tensor(false)]; + tensor var_891 = matmul(transpose_x = var_891_transpose_x_0, transpose_y = var_891_transpose_y_0, x = q_7, y = prev_kv_7)[name = tensor("op_891")]; + tensor var_892 = real_div(x = sqrt_s0_7, y = sqrt_s_t_7)[name = tensor("op_892")]; + tensor var_893 = const()[name = tensor("op_893"), val = tensor([1, 1, 4, 1])]; + tensor var_894 = reshape(shape = var_893, x = var_892)[name = tensor("op_894")]; + tensor cross_7 = mul(x = var_891, y = var_894)[name = tensor("cross_7")]; tensor out_19 = add(x = inner_7, y = cross_7)[name = tensor("out_19")]; - tensor var_841 = mul(x = prev_kv_7, y = sqrt_s0_7)[name = tensor("op_841")]; - tensor var_843_transpose_x_1 = const()[name = tensor("op_843_transpose_x_1"), val = tensor(true)]; - tensor var_843_transpose_y_1 = const()[name = tensor("op_843_transpose_y_1"), val = tensor(false)]; - tensor var_843 = matmul(transpose_x = var_843_transpose_x_1, transpose_y = var_843_transpose_y_1, x = k_7, y = v_7)[name = tensor("op_843")]; - tensor new_kv_unnorm_7 = add(x = var_841, y = var_843)[name = tensor("new_kv_unnorm_7")]; - tensor var_845 = const()[name = tensor("op_845"), val = tensor(0x1p+2)]; - tensor new_scale_7 = add(x = prev_scale_7, y = var_845)[name = tensor("new_scale_7")]; - tensor var_847 = sqrt(x = new_scale_7)[name = tensor("op_847")]; - tensor nkv_1 = real_div(x = new_kv_unnorm_7, y = var_847)[name = tensor("nkv_1")]; - tensor var_849_perm_0 = const()[name = tensor("op_849_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_897 = mul(x = prev_kv_7, y = sqrt_s0_7)[name = tensor("op_897")]; + tensor var_899_transpose_x_1 = const()[name = tensor("op_899_transpose_x_1"), val = tensor(true)]; + tensor var_899_transpose_y_1 = const()[name = tensor("op_899_transpose_y_1"), val = tensor(false)]; + tensor var_899 = matmul(transpose_x = var_899_transpose_x_1, transpose_y = var_899_transpose_y_1, x = k_7, y = v_7)[name = tensor("op_899")]; + tensor new_kv_unnorm_7 = add(x = var_897, y = var_899)[name = tensor("new_kv_unnorm_7")]; + tensor var_901 = const()[name = tensor("op_901"), val = tensor(0x1p+2)]; + tensor new_scale_7 = add(x = prev_scale_7, y = var_901)[name = tensor("new_scale_7")]; + tensor var_903 = sqrt(x = new_scale_7)[name = tensor("op_903")]; + tensor nkv_1 = real_div(x = new_kv_unnorm_7, y = var_903)[name = tensor("nkv_1")]; + tensor var_905_perm_0 = const()[name = tensor("op_905_perm_0"), val = tensor([0, 2, 1, 3])]; tensor out_21_axes_0 = const()[name = tensor("out_21_axes_0"), val = tensor([-1])]; - tensor var_849 = transpose(perm = var_849_perm_0, x = out_19)[name = tensor("transpose_34")]; - tensor out_21 = layer_norm(axes = out_21_axes_0, epsilon = var_17, x = var_849)[name = tensor("out_21")]; - tensor var_853 = const()[name = tensor("op_853"), val = tensor([1, 4, 256])]; - tensor out_23 = reshape(shape = var_853, x = out_21)[name = tensor("out_23")]; - tensor var_855 = silu(x = input_137)[name = tensor("op_855")]; - tensor input_139 = mul(x = var_855, y = out_23)[name = tensor("input_139")]; - tensor ret_out_7 = linear(bias = encoder_out_proj_3_bias, weight = encoder_out_proj_3_weight, x = input_139)[name = tensor("linear_34")]; - tensor x_21 = add(x = input_133, y = ret_out_7)[name = tensor("x_21")]; + tensor var_905 = transpose(perm = var_905_perm_0, x = out_19)[name = tensor("transpose_34")]; + tensor out_21 = layer_norm(axes = out_21_axes_0, epsilon = var_73, x = var_905)[name = tensor("out_21")]; + tensor var_909 = const()[name = tensor("op_909"), val = tensor([1, 4, 256])]; + tensor out_23 = reshape(shape = var_909, x = out_21)[name = tensor("out_23")]; + tensor var_911 = silu(x = input_139)[name = tensor("op_911")]; + tensor input_141 = mul(x = var_911, y = out_23)[name = tensor("input_141")]; + tensor ret_out_7 = linear(bias = inner_encoder_out_proj_3_bias, weight = inner_encoder_out_proj_3_weight, x = input_141)[name = tensor("linear_34")]; + tensor x_21 = add(x = input_135, y = ret_out_7)[name = tensor("x_21")]; tensor window_31_begin_0 = const()[name = tensor("window_31_begin_0"), val = tensor([3, 0, 0, 0])]; tensor window_31_end_0 = const()[name = tensor("window_31_end_0"), val = tensor([4, 1, 16, 256])]; tensor window_31_end_mask_0 = const()[name = tensor("window_31_end_mask_0"), val = tensor([false, true, true, true])]; tensor window_31_squeeze_mask_0 = const()[name = tensor("window_31_squeeze_mask_0"), val = tensor([true, false, false, false])]; tensor window_31 = slice_by_index(begin = window_31_begin_0, end = window_31_end_0, end_mask = window_31_end_mask_0, squeeze_mask = window_31_squeeze_mask_0, x = enc_conv_cache)[name = tensor("window_31")]; - tensor var_863_begin_0 = const()[name = tensor("op_863_begin_0"), val = tensor([0, 0, 0])]; - tensor var_863_end_0 = const()[name = tensor("op_863_end_0"), val = tensor([1, 1, 256])]; - tensor var_863_end_mask_0 = const()[name = tensor("op_863_end_mask_0"), val = tensor([true, false, true])]; - tensor var_863 = slice_by_index(begin = var_863_begin_0, end = var_863_end_0, end_mask = var_863_end_mask_0, x = x_21)[name = tensor("op_863")]; - tensor var_866_begin_0 = const()[name = tensor("op_866_begin_0"), val = tensor([0, 1, 0])]; - tensor var_866_end_0 = const()[name = tensor("op_866_end_0"), val = tensor([1, 16, 256])]; - tensor var_866_end_mask_0 = const()[name = tensor("op_866_end_mask_0"), val = tensor([true, true, true])]; - tensor var_866 = slice_by_index(begin = var_866_begin_0, end = var_866_end_0, end_mask = var_866_end_mask_0, x = window_31)[name = tensor("op_866")]; + tensor var_919_begin_0 = const()[name = tensor("op_919_begin_0"), val = tensor([0, 0, 0])]; + tensor var_919_end_0 = const()[name = tensor("op_919_end_0"), val = tensor([1, 1, 256])]; + tensor var_919_end_mask_0 = const()[name = tensor("op_919_end_mask_0"), val = tensor([true, false, true])]; + tensor var_919 = slice_by_index(begin = var_919_begin_0, end = var_919_end_0, end_mask = var_919_end_mask_0, x = x_21)[name = tensor("op_919")]; + tensor var_922_begin_0 = const()[name = tensor("op_922_begin_0"), val = tensor([0, 1, 0])]; + tensor var_922_end_0 = const()[name = tensor("op_922_end_0"), val = tensor([1, 16, 256])]; + tensor var_922_end_mask_0 = const()[name = tensor("op_922_end_mask_0"), val = tensor([true, true, true])]; + tensor var_922 = slice_by_index(begin = var_922_begin_0, end = var_922_end_0, end_mask = var_922_end_mask_0, x = window_31)[name = tensor("op_922")]; tensor window_33_interleave_0 = const()[name = tensor("window_33_interleave_0"), val = tensor(false)]; - tensor window_33 = concat(axis = var_26, interleave = window_33_interleave_0, values = (var_866, var_863))[name = tensor("window_33")]; - tensor var_871_begin_0 = const()[name = tensor("op_871_begin_0"), val = tensor([0, 1, 0])]; - tensor var_871_end_0 = const()[name = tensor("op_871_end_0"), val = tensor([1, 2, 256])]; - tensor var_871_end_mask_0 = const()[name = tensor("op_871_end_mask_0"), val = tensor([true, false, true])]; - tensor var_871 = slice_by_index(begin = var_871_begin_0, end = var_871_end_0, end_mask = var_871_end_mask_0, x = x_21)[name = tensor("op_871")]; - tensor var_874_begin_0 = const()[name = tensor("op_874_begin_0"), val = tensor([0, 1, 0])]; - tensor var_874_end_0 = const()[name = tensor("op_874_end_0"), val = tensor([1, 16, 256])]; - tensor var_874_end_mask_0 = const()[name = tensor("op_874_end_mask_0"), val = tensor([true, true, true])]; - tensor var_874 = slice_by_index(begin = var_874_begin_0, end = var_874_end_0, end_mask = var_874_end_mask_0, x = window_33)[name = tensor("op_874")]; + tensor window_33 = concat(axis = var_82, interleave = window_33_interleave_0, values = (var_922, var_919))[name = tensor("window_33")]; + tensor var_927_begin_0 = const()[name = tensor("op_927_begin_0"), val = tensor([0, 1, 0])]; + tensor var_927_end_0 = const()[name = tensor("op_927_end_0"), val = tensor([1, 2, 256])]; + tensor var_927_end_mask_0 = const()[name = tensor("op_927_end_mask_0"), val = tensor([true, false, true])]; + tensor var_927 = slice_by_index(begin = var_927_begin_0, end = var_927_end_0, end_mask = var_927_end_mask_0, x = x_21)[name = tensor("op_927")]; + tensor var_930_begin_0 = const()[name = tensor("op_930_begin_0"), val = tensor([0, 1, 0])]; + tensor var_930_end_0 = const()[name = tensor("op_930_end_0"), val = tensor([1, 16, 256])]; + tensor var_930_end_mask_0 = const()[name = tensor("op_930_end_mask_0"), val = tensor([true, true, true])]; + tensor var_930 = slice_by_index(begin = var_930_begin_0, end = var_930_end_0, end_mask = var_930_end_mask_0, x = window_33)[name = tensor("op_930")]; tensor window_35_interleave_0 = const()[name = tensor("window_35_interleave_0"), val = tensor(false)]; - tensor window_35 = concat(axis = var_26, interleave = window_35_interleave_0, values = (var_874, var_871))[name = tensor("window_35")]; - tensor var_879_begin_0 = const()[name = tensor("op_879_begin_0"), val = tensor([0, 2, 0])]; - tensor var_879_end_0 = const()[name = tensor("op_879_end_0"), val = tensor([1, 3, 256])]; - tensor var_879_end_mask_0 = const()[name = tensor("op_879_end_mask_0"), val = tensor([true, false, true])]; - tensor var_879 = slice_by_index(begin = var_879_begin_0, end = var_879_end_0, end_mask = var_879_end_mask_0, x = x_21)[name = tensor("op_879")]; - tensor var_882_begin_0 = const()[name = tensor("op_882_begin_0"), val = tensor([0, 1, 0])]; - tensor var_882_end_0 = const()[name = tensor("op_882_end_0"), val = tensor([1, 16, 256])]; - tensor var_882_end_mask_0 = const()[name = tensor("op_882_end_mask_0"), val = tensor([true, true, true])]; - tensor var_882 = slice_by_index(begin = var_882_begin_0, end = var_882_end_0, end_mask = var_882_end_mask_0, x = window_35)[name = tensor("op_882")]; + tensor window_35 = concat(axis = var_82, interleave = window_35_interleave_0, values = (var_930, var_927))[name = tensor("window_35")]; + tensor var_935_begin_0 = const()[name = tensor("op_935_begin_0"), val = tensor([0, 2, 0])]; + tensor var_935_end_0 = const()[name = tensor("op_935_end_0"), val = tensor([1, 3, 256])]; + tensor var_935_end_mask_0 = const()[name = tensor("op_935_end_mask_0"), val = tensor([true, false, true])]; + tensor var_935 = slice_by_index(begin = var_935_begin_0, end = var_935_end_0, end_mask = var_935_end_mask_0, x = x_21)[name = tensor("op_935")]; + tensor var_938_begin_0 = const()[name = tensor("op_938_begin_0"), val = tensor([0, 1, 0])]; + tensor var_938_end_0 = const()[name = tensor("op_938_end_0"), val = tensor([1, 16, 256])]; + tensor var_938_end_mask_0 = const()[name = tensor("op_938_end_mask_0"), val = tensor([true, true, true])]; + tensor var_938 = slice_by_index(begin = var_938_begin_0, end = var_938_end_0, end_mask = var_938_end_mask_0, x = window_35)[name = tensor("op_938")]; tensor window_37_interleave_0 = const()[name = tensor("window_37_interleave_0"), val = tensor(false)]; - tensor window_37 = concat(axis = var_26, interleave = window_37_interleave_0, values = (var_882, var_879))[name = tensor("window_37")]; - tensor var_887_begin_0 = const()[name = tensor("op_887_begin_0"), val = tensor([0, 3, 0])]; - tensor var_887_end_0 = const()[name = tensor("op_887_end_0"), val = tensor([1, 1, 256])]; - tensor var_887_end_mask_0 = const()[name = tensor("op_887_end_mask_0"), val = tensor([true, true, true])]; - tensor var_887 = slice_by_index(begin = var_887_begin_0, end = var_887_end_0, end_mask = var_887_end_mask_0, x = x_21)[name = tensor("op_887")]; - tensor var_890_begin_0 = const()[name = tensor("op_890_begin_0"), val = tensor([0, 1, 0])]; - tensor var_890_end_0 = const()[name = tensor("op_890_end_0"), val = tensor([1, 16, 256])]; - tensor var_890_end_mask_0 = const()[name = tensor("op_890_end_mask_0"), val = tensor([true, true, true])]; - tensor var_890 = slice_by_index(begin = var_890_begin_0, end = var_890_end_0, end_mask = var_890_end_mask_0, x = window_37)[name = tensor("op_890")]; + tensor window_37 = concat(axis = var_82, interleave = window_37_interleave_0, values = (var_938, var_935))[name = tensor("window_37")]; + tensor var_943_begin_0 = const()[name = tensor("op_943_begin_0"), val = tensor([0, 3, 0])]; + tensor var_943_end_0 = const()[name = tensor("op_943_end_0"), val = tensor([1, 1, 256])]; + tensor var_943_end_mask_0 = const()[name = tensor("op_943_end_mask_0"), val = tensor([true, true, true])]; + tensor var_943 = slice_by_index(begin = var_943_begin_0, end = var_943_end_0, end_mask = var_943_end_mask_0, x = x_21)[name = tensor("op_943")]; + tensor var_946_begin_0 = const()[name = tensor("op_946_begin_0"), val = tensor([0, 1, 0])]; + tensor var_946_end_0 = const()[name = tensor("op_946_end_0"), val = tensor([1, 16, 256])]; + tensor var_946_end_mask_0 = const()[name = tensor("op_946_end_mask_0"), val = tensor([true, true, true])]; + tensor var_946 = slice_by_index(begin = var_946_begin_0, end = var_946_end_0, end_mask = var_946_end_mask_0, x = window_37)[name = tensor("op_946")]; tensor window_interleave_0 = const()[name = tensor("window_interleave_0"), val = tensor(false)]; - tensor window = concat(axis = var_26, interleave = window_interleave_0, values = (var_890, var_887))[name = tensor("window")]; - tensor input_141_interleave_0 = const()[name = tensor("input_141_interleave_0"), val = tensor(false)]; - tensor input_141 = concat(axis = var_23, interleave = input_141_interleave_0, values = (window_33, window_35, window_37, window))[name = tensor("input_141")]; + tensor window = concat(axis = var_82, interleave = window_interleave_0, values = (var_946, var_943))[name = tensor("window")]; + tensor input_143_interleave_0 = const()[name = tensor("input_143_interleave_0"), val = tensor(false)]; + tensor input_143 = concat(axis = var_68, interleave = input_143_interleave_0, values = (window_33, window_35, window_37, window))[name = tensor("input_143")]; tensor x_23_axes_0 = const()[name = tensor("x_23_axes_0"), val = tensor([-1])]; - tensor x_23 = layer_norm(axes = x_23_axes_0, beta = encoder_conv_module_3_sequential_0_bias, epsilon = var_28, gamma = encoder_conv_module_3_sequential_0_weight, x = input_141)[name = tensor("x_23")]; - tensor input_143_perm_0 = const()[name = tensor("input_143_perm_0"), val = tensor([0, 2, 1])]; + tensor x_23 = layer_norm(axes = x_23_axes_0, beta = inner_encoder_conv_module_3_sequential_0_bias, epsilon = var_65, gamma = inner_encoder_conv_module_3_sequential_0_weight, x = input_143)[name = tensor("x_23")]; + tensor input_145_perm_0 = const()[name = tensor("input_145_perm_0"), val = tensor([0, 2, 1])]; tensor inputs_33_pad_type_0 = const()[name = tensor("inputs_33_pad_type_0"), val = tensor("valid")]; tensor inputs_33_strides_0 = const()[name = tensor("inputs_33_strides_0"), val = tensor([1])]; tensor inputs_33_pad_0 = const()[name = tensor("inputs_33_pad_0"), val = tensor([0, 0])]; tensor inputs_33_dilations_0 = const()[name = tensor("inputs_33_dilations_0"), val = tensor([1])]; tensor inputs_33_groups_0 = const()[name = tensor("inputs_33_groups_0"), val = tensor(1)]; - tensor input_143 = transpose(perm = input_143_perm_0, x = x_23)[name = tensor("transpose_33")]; - tensor inputs_33 = conv(bias = encoder_conv_module_3_sequential_2_conv_bias, dilations = inputs_33_dilations_0, groups = inputs_33_groups_0, pad = inputs_33_pad_0, pad_type = inputs_33_pad_type_0, strides = inputs_33_strides_0, weight = encoder_conv_module_3_sequential_2_conv_weight, x = input_143)[name = tensor("inputs_33")]; - tensor var_915_split_sizes_0 = const()[name = tensor("op_915_split_sizes_0"), val = tensor([256, 256])]; - tensor var_915_axis_0 = const()[name = tensor("op_915_axis_0"), val = tensor(1)]; - tensor var_915_0, tensor var_915_1 = split(axis = var_915_axis_0, split_sizes = var_915_split_sizes_0, x = inputs_33)[name = tensor("op_915")]; - tensor var_917 = sigmoid(x = var_915_1)[name = tensor("op_917")]; - tensor inputs_35 = mul(x = var_915_0, y = var_917)[name = tensor("inputs_35")]; + tensor input_145 = transpose(perm = input_145_perm_0, x = x_23)[name = tensor("transpose_33")]; + tensor inputs_33 = conv(bias = inner_encoder_conv_module_3_sequential_2_conv_bias, dilations = inputs_33_dilations_0, groups = inputs_33_groups_0, pad = inputs_33_pad_0, pad_type = inputs_33_pad_type_0, strides = inputs_33_strides_0, weight = inner_encoder_conv_module_3_sequential_2_conv_weight, x = input_145)[name = tensor("inputs_33")]; + tensor var_971_split_sizes_0 = const()[name = tensor("op_971_split_sizes_0"), val = tensor([256, 256])]; + tensor var_971_axis_0 = const()[name = tensor("op_971_axis_0"), val = tensor(1)]; + tensor var_971_0, tensor var_971_1 = split(axis = var_971_axis_0, split_sizes = var_971_split_sizes_0, x = inputs_33)[name = tensor("op_971")]; + tensor var_973 = sigmoid(x = var_971_1)[name = tensor("op_973")]; + tensor inputs_35 = mul(x = var_971_0, y = var_973)[name = tensor("inputs_35")]; tensor outputs_aug_pad_type_0 = const()[name = tensor("outputs_aug_pad_type_0"), val = tensor("custom")]; tensor outputs_aug_pad_0 = const()[name = tensor("outputs_aug_pad_0"), val = tensor([15, 15])]; tensor outputs_aug_groups_0 = const()[name = tensor("outputs_aug_groups_0"), val = tensor(256)]; tensor outputs_aug_strides_0 = const()[name = tensor("outputs_aug_strides_0"), val = tensor([1])]; tensor outputs_aug_dilations_0 = const()[name = tensor("outputs_aug_dilations_0"), val = tensor([1])]; - tensor outputs_aug = conv(dilations = outputs_aug_dilations_0, groups = outputs_aug_groups_0, pad = outputs_aug_pad_0, pad_type = outputs_aug_pad_type_0, strides = outputs_aug_strides_0, weight = encoder_conv_module_3_sequential_4_conv_weight, x = inputs_35)[name = tensor("outputs_aug")]; - tensor input_145_begin_0 = const()[name = tensor("input_145_begin_0"), val = tensor([0, 0, 0])]; - tensor input_145_end_0 = const()[name = tensor("input_145_end_0"), val = tensor([4, 256, 16])]; - tensor input_145_end_mask_0 = const()[name = tensor("input_145_end_mask_0"), val = tensor([true, true, false])]; - tensor input_145 = slice_by_index(begin = input_145_begin_0, end = input_145_end_0, end_mask = input_145_end_mask_0, x = outputs_aug)[name = tensor("input_145")]; - tensor inputs_37 = batch_norm(beta = encoder_conv_module_3_sequential_5_bias, epsilon = var_28, gamma = encoder_conv_module_3_sequential_5_weight, mean = encoder_conv_module_3_sequential_5_running_mean, variance = encoder_conv_module_3_sequential_5_running_var, x = input_145)[name = tensor("inputs_37")]; - tensor input_147 = silu(x = inputs_37)[name = tensor("input_147")]; - tensor input_149_pad_type_0 = const()[name = tensor("input_149_pad_type_0"), val = tensor("valid")]; - tensor input_149_strides_0 = const()[name = tensor("input_149_strides_0"), val = tensor([1])]; - tensor input_149_pad_0 = const()[name = tensor("input_149_pad_0"), val = tensor([0, 0])]; - tensor input_149_dilations_0 = const()[name = tensor("input_149_dilations_0"), val = tensor([1])]; - tensor input_149_groups_0 = const()[name = tensor("input_149_groups_0"), val = tensor(1)]; - tensor input_149 = conv(bias = encoder_conv_module_3_sequential_7_conv_bias, dilations = input_149_dilations_0, groups = input_149_groups_0, pad = input_149_pad_0, pad_type = input_149_pad_type_0, strides = input_149_strides_0, weight = encoder_conv_module_3_sequential_7_conv_weight, x = input_147)[name = tensor("input_149")]; + tensor outputs_aug = conv(dilations = outputs_aug_dilations_0, groups = outputs_aug_groups_0, pad = outputs_aug_pad_0, pad_type = outputs_aug_pad_type_0, strides = outputs_aug_strides_0, weight = inner_encoder_conv_module_3_sequential_4_conv_weight, x = inputs_35)[name = tensor("outputs_aug")]; + tensor input_147_begin_0 = const()[name = tensor("input_147_begin_0"), val = tensor([0, 0, 0])]; + tensor input_147_end_0 = const()[name = tensor("input_147_end_0"), val = tensor([4, 256, 16])]; + tensor input_147_end_mask_0 = const()[name = tensor("input_147_end_mask_0"), val = tensor([true, true, false])]; + tensor input_147 = slice_by_index(begin = input_147_begin_0, end = input_147_end_0, end_mask = input_147_end_mask_0, x = outputs_aug)[name = tensor("input_147")]; + tensor inputs_37 = batch_norm(beta = inner_encoder_conv_module_3_sequential_5_bias, epsilon = var_65, gamma = inner_encoder_conv_module_3_sequential_5_weight, mean = inner_encoder_conv_module_3_sequential_5_running_mean, variance = inner_encoder_conv_module_3_sequential_5_running_var, x = input_147)[name = tensor("inputs_37")]; + tensor input_149 = silu(x = inputs_37)[name = tensor("input_149")]; + tensor input_151_pad_type_0 = const()[name = tensor("input_151_pad_type_0"), val = tensor("valid")]; + tensor input_151_strides_0 = const()[name = tensor("input_151_strides_0"), val = tensor([1])]; + tensor input_151_pad_0 = const()[name = tensor("input_151_pad_0"), val = tensor([0, 0])]; + tensor input_151_dilations_0 = const()[name = tensor("input_151_dilations_0"), val = tensor([1])]; + tensor input_151_groups_0 = const()[name = tensor("input_151_groups_0"), val = tensor(1)]; + tensor input_151 = conv(bias = inner_encoder_conv_module_3_sequential_7_conv_bias, dilations = input_151_dilations_0, groups = input_151_groups_0, pad = input_151_pad_0, pad_type = input_151_pad_type_0, strides = input_151_strides_0, weight = inner_encoder_conv_module_3_sequential_7_conv_weight, x = input_149)[name = tensor("input_151")]; tensor conv_out_7_perm_0 = const()[name = tensor("conv_out_7_perm_0"), val = tensor([0, 2, 1])]; - tensor var_948_begin_0 = const()[name = tensor("op_948_begin_0"), val = tensor([0, -1, 0])]; - tensor var_948_end_0 = const()[name = tensor("op_948_end_0"), val = tensor([4, 16, 256])]; - tensor var_948_end_mask_0 = const()[name = tensor("op_948_end_mask_0"), val = tensor([true, true, true])]; - tensor conv_out_7 = transpose(perm = conv_out_7_perm_0, x = input_149)[name = tensor("transpose_32")]; - tensor var_948 = slice_by_index(begin = var_948_begin_0, end = var_948_end_0, end_mask = var_948_end_mask_0, x = conv_out_7)[name = tensor("op_948")]; - tensor var_950_perm_0 = const()[name = tensor("op_950_perm_0"), val = tensor([1, 0, 2])]; - tensor var_950 = transpose(perm = var_950_perm_0, x = var_948)[name = tensor("transpose_31")]; - tensor input_151 = add(x = x_21, y = var_950)[name = tensor("input_151")]; - tensor input_153_axes_0 = const()[name = tensor("input_153_axes_0"), val = tensor([-1])]; - tensor input_153 = layer_norm(axes = input_153_axes_0, beta = encoder_ffn2_3_module_sequential_0_bias, epsilon = var_28, gamma = encoder_ffn2_3_module_sequential_0_weight, x = input_151)[name = tensor("input_153")]; - tensor inputs = linear(bias = encoder_ffn2_3_module_sequential_1_linear_bias, weight = encoder_ffn2_3_module_sequential_1_linear_weight, x = input_153)[name = tensor("linear_35")]; - tensor input_155 = silu(x = inputs)[name = tensor("input_155")]; - tensor input_159 = linear(bias = encoder_ffn2_3_module_sequential_4_linear_bias, weight = encoder_ffn2_3_module_sequential_4_linear_weight, x = input_155)[name = tensor("linear_36")]; - tensor var_973 = const()[name = tensor("op_973"), val = tensor(0x1p-1)]; - tensor var_974 = mul(x = input_159, y = var_973)[name = tensor("op_974")]; - tensor input_161 = add(x = var_974, y = input_151)[name = tensor("input_161")]; + tensor var_1004_begin_0 = const()[name = tensor("op_1004_begin_0"), val = tensor([0, -1, 0])]; + tensor var_1004_end_0 = const()[name = tensor("op_1004_end_0"), val = tensor([4, 16, 256])]; + tensor var_1004_end_mask_0 = const()[name = tensor("op_1004_end_mask_0"), val = tensor([true, true, true])]; + tensor conv_out_7 = transpose(perm = conv_out_7_perm_0, x = input_151)[name = tensor("transpose_32")]; + tensor var_1004 = slice_by_index(begin = var_1004_begin_0, end = var_1004_end_0, end_mask = var_1004_end_mask_0, x = conv_out_7)[name = tensor("op_1004")]; + tensor var_1006_perm_0 = const()[name = tensor("op_1006_perm_0"), val = tensor([1, 0, 2])]; + tensor var_1006 = transpose(perm = var_1006_perm_0, x = var_1004)[name = tensor("transpose_31")]; + tensor input_153 = add(x = x_21, y = var_1006)[name = tensor("input_153")]; + tensor input_155_axes_0 = const()[name = tensor("input_155_axes_0"), val = tensor([-1])]; + tensor input_155 = layer_norm(axes = input_155_axes_0, beta = inner_encoder_ffn2_3_module_sequential_0_bias, epsilon = var_65, gamma = inner_encoder_ffn2_3_module_sequential_0_weight, x = input_153)[name = tensor("input_155")]; + tensor inputs = linear(bias = inner_encoder_ffn2_3_module_sequential_1_linear_bias, weight = inner_encoder_ffn2_3_module_sequential_1_linear_weight, x = input_155)[name = tensor("linear_35")]; + tensor input_157 = silu(x = inputs)[name = tensor("input_157")]; + tensor input_161 = linear(bias = inner_encoder_ffn2_3_module_sequential_4_linear_bias, weight = inner_encoder_ffn2_3_module_sequential_4_linear_weight, x = input_157)[name = tensor("linear_36")]; + tensor var_1029 = const()[name = tensor("op_1029"), val = tensor(0x1p-1)]; + tensor var_1030 = mul(x = input_161, y = var_1029)[name = tensor("op_1030")]; + tensor input_163 = add(x = var_1030, y = input_153)[name = tensor("input_163")]; tensor x_25_axes_0 = const()[name = tensor("x_25_axes_0"), val = tensor([-1])]; - tensor x_25 = layer_norm(axes = x_25_axes_0, beta = encoder_layer_norm_3_bias, epsilon = var_28, gamma = encoder_layer_norm_3_weight, x = input_161)[name = tensor("x_25")]; + tensor x_25 = layer_norm(axes = x_25_axes_0, beta = inner_encoder_layer_norm_3_bias, epsilon = var_65, gamma = inner_encoder_layer_norm_3_weight, x = input_163)[name = tensor("x_25")]; tensor x_ct_perm_0 = const()[name = tensor("x_ct_perm_0"), val = tensor([0, 2, 1])]; tensor cat_interleave_0 = const()[name = tensor("cat_interleave_0"), val = tensor(false)]; tensor x_ct = transpose(perm = x_ct_perm_0, x = x_25)[name = tensor("transpose_30")]; - tensor cat = concat(axis = var_20, interleave = cat_interleave_0, values = (cnn_window, x_ct))[name = tensor("cat")]; + tensor cat = concat(axis = var_70, interleave = cat_interleave_0, values = (cnn_window, x_ct))[name = tensor("cat")]; tensor conv_out_pad_type_0 = const()[name = tensor("conv_out_pad_type_0"), val = tensor("valid")]; tensor conv_out_strides_0 = const()[name = tensor("conv_out_strides_0"), val = tensor([1])]; tensor conv_out_pad_0 = const()[name = tensor("conv_out_pad_0"), val = tensor([0, 0])]; tensor conv_out_dilations_0 = const()[name = tensor("conv_out_dilations_0"), val = tensor([1])]; tensor conv_out_groups_0 = const()[name = tensor("conv_out_groups_0"), val = tensor(1)]; - tensor conv_out = conv(bias = encoder_cnn_bias, dilations = conv_out_dilations_0, groups = conv_out_groups_0, pad = conv_out_pad_0, pad_type = conv_out_pad_type_0, strides = conv_out_strides_0, weight = encoder_cnn_weight, x = cat)[name = tensor("conv_out")]; - tensor var_992_begin_0 = const()[name = tensor("op_992_begin_0"), val = tensor([0, 0, 4])]; - tensor var_992_end_0 = const()[name = tensor("op_992_end_0"), val = tensor([1, 256, 22])]; - tensor var_992_end_mask_0 = const()[name = tensor("op_992_end_mask_0"), val = tensor([true, true, true])]; - tensor cnn_window_new = slice_by_index(begin = var_992_begin_0, end = var_992_end_0, end_mask = var_992_end_mask_0, x = cat)[name = tensor("op_992")]; - tensor input_163_perm_0 = const()[name = tensor("input_163_perm_0"), val = tensor([0, 2, 1])]; - tensor var_994 = const()[name = tensor("op_994"), val = tensor([-1])]; - tensor input_163 = transpose(perm = input_163_perm_0, x = conv_out)[name = tensor("transpose_29")]; - tensor var_995 = reduce_l2_norm(axes = var_994, keep_dims = var_29, x = input_163)[name = tensor("op_995")]; + tensor conv_out = conv(bias = inner_encoder_cnn_bias, dilations = conv_out_dilations_0, groups = conv_out_groups_0, pad = conv_out_pad_0, pad_type = conv_out_pad_type_0, strides = conv_out_strides_0, weight = inner_encoder_cnn_weight, x = cat)[name = tensor("conv_out")]; + tensor var_1048_begin_0 = const()[name = tensor("op_1048_begin_0"), val = tensor([0, 0, 4])]; + tensor var_1048_end_0 = const()[name = tensor("op_1048_end_0"), val = tensor([1, 256, 22])]; + tensor var_1048_end_mask_0 = const()[name = tensor("op_1048_end_mask_0"), val = tensor([true, true, true])]; + tensor cnn_window_new = slice_by_index(begin = var_1048_begin_0, end = var_1048_end_0, end_mask = var_1048_end_mask_0, x = cat)[name = tensor("op_1048")]; + tensor input_165_perm_0 = const()[name = tensor("input_165_perm_0"), val = tensor([0, 2, 1])]; + tensor var_1050 = const()[name = tensor("op_1050"), val = tensor([-1])]; + tensor input_165 = transpose(perm = input_165_perm_0, x = conv_out)[name = tensor("transpose_29")]; + tensor var_1051 = reduce_l2_norm(axes = var_1050, keep_dims = var_64, x = input_165)[name = tensor("op_1051")]; tensor const_12 = const()[name = tensor("const_12"), val = tensor(0x1.fffffep+127)]; - tensor clip_0 = clip(alpha = var_11, beta = const_12, x = var_995)[name = tensor("clip_0")]; - tensor emb = real_div(x = input_163, y = clip_0)[name = tensor("emb")]; - tensor var_999_axis_0 = const()[name = tensor("op_999_axis_0"), val = tensor(0)]; - tensor enc_kv_new = stack(axis = var_999_axis_0, values = (var_206, var_420, var_634, nkv_1))[name = tensor("op_999")]; - tensor var_1001_axis_0 = const()[name = tensor("op_1001_axis_0"), val = tensor(0)]; - tensor enc_scale_new = stack(axis = var_1001_axis_0, values = (new_scale_1, new_scale_3, new_scale_5, new_scale_7))[name = tensor("op_1001")]; - tensor var_1003_axis_0 = const()[name = tensor("op_1003_axis_0"), val = tensor(0)]; - tensor enc_conv_cache_new = stack(axis = var_1003_axis_0, values = (window_9, window_19, window_29, window))[name = tensor("op_1003")]; - tensor var_1012 = const()[name = tensor("op_1012"), val = tensor(0x1.5798eep-27)]; - tensor var_1017 = const()[name = tensor("op_1017"), val = tensor(0x1.4f8b58p-17)]; - tensor var_1019 = const()[name = tensor("op_1019"), val = tensor(0x1.0c6f7ap-20)]; - tensor var_1020 = const()[name = tensor("op_1020"), val = tensor(true)]; - tensor var_1022 = const()[name = tensor("op_1022"), val = tensor(0x1p+0)]; - tensor var_1026 = const()[name = tensor("op_1026"), val = tensor(-1)]; - tensor var_1032 = const()[name = tensor("op_1032"), val = tensor(0)]; + tensor clip_0 = clip(alpha = var_78, beta = const_12, x = var_1051)[name = tensor("clip_0")]; + tensor emb = real_div(x = input_165, y = clip_0)[name = tensor("emb")]; + tensor var_1055_axis_0 = const()[name = tensor("op_1055_axis_0"), val = tensor(0)]; + tensor enc_kv_new = stack(axis = var_1055_axis_0, values = (var_262, var_476, var_690, nkv_1))[name = tensor("op_1055")]; + tensor var_1057_axis_0 = const()[name = tensor("op_1057_axis_0"), val = tensor(0)]; + tensor enc_scale_new = stack(axis = var_1057_axis_0, values = (new_scale_1, new_scale_3, new_scale_5, new_scale_7))[name = tensor("op_1057")]; + tensor var_1059_axis_0 = const()[name = tensor("op_1059_axis_0"), val = tensor(0)]; + tensor enc_conv_cache_new = stack(axis = var_1059_axis_0, values = (window_9, window_19, window_29, window))[name = tensor("op_1059")]; tensor pos = const()[name = tensor("pos"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(44395584)))]; - tensor var_1094_axes_0 = const()[name = tensor("op_1094_axes_0"), val = tensor([2])]; - tensor var_1094 = expand_dims(axes = var_1094_axes_0, x = emb)[name = tensor("op_1094")]; + tensor var_1127_axes_0 = const()[name = tensor("op_1127_axes_0"), val = tensor([2])]; + tensor var_1127 = expand_dims(axes = var_1127_axes_0, x = emb)[name = tensor("op_1127")]; tensor emb_exp_reps_0 = const()[name = tensor("emb_exp_reps_0"), val = tensor([1, 1, 6, 1])]; - tensor emb_exp = tile(reps = emb_exp_reps_0, x = var_1094)[name = tensor("emb_exp")]; - tensor input_165_interleave_0 = const()[name = tensor("input_165_interleave_0"), val = tensor(false)]; - tensor input_165 = concat(axis = var_1026, interleave = input_165_interleave_0, values = (emb_exp, pos))[name = tensor("input_165")]; - tensor x_27 = linear(bias = decoder_convert_bias, weight = decoder_convert_weight, x = input_165)[name = tensor("linear_37")]; - tensor var_1102_perm_0 = const()[name = tensor("op_1102_perm_0"), val = tensor([0, 2, 1, 3])]; - tensor var_1106 = const()[name = tensor("op_1106"), val = tensor([6, 4, 256])]; - tensor var_1102 = transpose(perm = var_1102_perm_0, x = x_27)[name = tensor("transpose_28")]; - tensor x_29 = reshape(shape = var_1106, x = var_1102)[name = tensor("x_29")]; + tensor emb_exp = tile(reps = emb_exp_reps_0, x = var_1127)[name = tensor("emb_exp")]; + tensor input_167_interleave_0 = const()[name = tensor("input_167_interleave_0"), val = tensor(false)]; + tensor input_167 = concat(axis = var_71, interleave = input_167_interleave_0, values = (emb_exp, pos))[name = tensor("input_167")]; + tensor x_27 = linear(bias = inner_decoder_convert_bias, weight = inner_decoder_convert_weight, x = input_167)[name = tensor("linear_37")]; + tensor var_1135_perm_0 = const()[name = tensor("op_1135_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_1139 = const()[name = tensor("op_1139"), val = tensor([6, 4, 256])]; + tensor var_1135 = transpose(perm = var_1135_perm_0, x = x_27)[name = tensor("transpose_28")]; + tensor x_29 = reshape(shape = var_1139, x = var_1135)[name = tensor("x_29")]; tensor prev_kv_9_begin_0 = const()[name = tensor("prev_kv_9_begin_0"), val = tensor([0, 0, 0, 0, 0])]; tensor prev_kv_9_end_0 = const()[name = tensor("prev_kv_9_end_0"), val = tensor([1, 6, 4, 64, 64])]; tensor prev_kv_9_end_mask_0 = const()[name = tensor("prev_kv_9_end_mask_0"), val = tensor([false, true, true, true, true])]; @@ -983,132 +998,132 @@ program(1.0) tensor prev_scale_9_end_mask_0 = const()[name = tensor("prev_scale_9_end_mask_0"), val = tensor([false, true])]; tensor prev_scale_9_squeeze_mask_0 = const()[name = tensor("prev_scale_9_squeeze_mask_0"), val = tensor([true, false])]; tensor prev_scale_9 = slice_by_index(begin = prev_scale_9_begin_0, end = prev_scale_9_end_0, end_mask = prev_scale_9_end_mask_0, squeeze_mask = prev_scale_9_squeeze_mask_0, x = dec_scale)[name = tensor("prev_scale_9")]; - tensor var_1114 = linear(bias = decoder_q_proj_0_bias, weight = decoder_q_proj_0_weight, x = x_29)[name = tensor("linear_38")]; - tensor var_1115 = const()[name = tensor("op_1115"), val = tensor([6, 4, 4, 64])]; - tensor var_1116 = reshape(shape = var_1115, x = var_1114)[name = tensor("op_1116")]; + tensor var_1147 = linear(bias = inner_decoder_q_proj_0_bias, weight = inner_decoder_q_proj_0_weight, x = x_29)[name = tensor("linear_38")]; + tensor var_1148 = const()[name = tensor("op_1148"), val = tensor([6, 4, 4, 64])]; + tensor var_1149 = reshape(shape = var_1148, x = var_1147)[name = tensor("op_1149")]; tensor q_9_perm_0 = const()[name = tensor("q_9_perm_0"), val = tensor([0, 2, 1, 3])]; - tensor var_1120 = linear(bias = decoder_k_proj_0_bias, weight = decoder_k_proj_0_weight, x = x_29)[name = tensor("linear_39")]; - tensor var_1121 = const()[name = tensor("op_1121"), val = tensor(0x1p-3)]; - tensor var_1122 = mul(x = var_1120, y = var_1121)[name = tensor("op_1122")]; - tensor var_1123 = const()[name = tensor("op_1123"), val = tensor([6, 4, 4, 64])]; - tensor var_1124 = reshape(shape = var_1123, x = var_1122)[name = tensor("op_1124")]; + tensor var_1153 = linear(bias = inner_decoder_k_proj_0_bias, weight = inner_decoder_k_proj_0_weight, x = x_29)[name = tensor("linear_39")]; + tensor var_1154 = const()[name = tensor("op_1154"), val = tensor(0x1p-3)]; + tensor var_1155 = mul(x = var_1153, y = var_1154)[name = tensor("op_1155")]; + tensor var_1156 = const()[name = tensor("op_1156"), val = tensor([6, 4, 4, 64])]; + tensor var_1157 = reshape(shape = var_1156, x = var_1155)[name = tensor("op_1157")]; tensor k_9_perm_0 = const()[name = tensor("k_9_perm_0"), val = tensor([0, 2, 1, 3])]; - tensor var_1128 = linear(bias = decoder_v_proj_0_bias, weight = decoder_v_proj_0_weight, x = x_29)[name = tensor("linear_40")]; - tensor var_1129 = const()[name = tensor("op_1129"), val = tensor([6, 4, 4, 64])]; - tensor var_1130 = reshape(shape = var_1129, x = var_1128)[name = tensor("op_1130")]; + tensor var_1161 = linear(bias = inner_decoder_v_proj_0_bias, weight = inner_decoder_v_proj_0_weight, x = x_29)[name = tensor("linear_40")]; + tensor var_1162 = const()[name = tensor("op_1162"), val = tensor([6, 4, 4, 64])]; + tensor var_1163 = reshape(shape = var_1162, x = var_1161)[name = tensor("op_1163")]; tensor v_9_perm_0 = const()[name = tensor("v_9_perm_0"), val = tensor([0, 2, 1, 3])]; - tensor input_169 = linear(bias = decoder_g_proj_0_bias, weight = decoder_g_proj_0_weight, x = x_29)[name = tensor("linear_41")]; + tensor input_171 = linear(bias = inner_decoder_g_proj_0_bias, weight = inner_decoder_g_proj_0_weight, x = x_29)[name = tensor("linear_41")]; tensor cumsum_mask_1_exclusive_0 = const()[name = tensor("cumsum_mask_1_exclusive_0"), val = tensor(false)]; tensor cumsum_mask_1_reverse_0 = const()[name = tensor("cumsum_mask_1_reverse_0"), val = tensor(false)]; - tensor cumsum_mask_1 = cumsum(axis = var_1032, exclusive = cumsum_mask_1_exclusive_0, reverse = cumsum_mask_1_reverse_0, x = valid_mask)[name = tensor("cumsum_mask_1")]; + tensor cumsum_mask_1 = cumsum(axis = var_68, exclusive = cumsum_mask_1_exclusive_0, reverse = cumsum_mask_1_reverse_0, x = valid_mask)[name = tensor("cumsum_mask_1")]; tensor sqrt_s0_9 = sqrt(x = prev_scale_9)[name = tensor("sqrt_s0_9")]; tensor s_t_9 = add(x = prev_scale_9, y = cumsum_mask_1)[name = tensor("s_t_9")]; tensor const_20 = const()[name = tensor("const_20"), val = tensor(0x1.fffffep+127)]; - tensor clip_1 = clip(alpha = var_1022, beta = const_20, x = s_t_9)[name = tensor("clip_1")]; + tensor clip_1 = clip(alpha = var_58, beta = const_20, x = s_t_9)[name = tensor("clip_1")]; tensor sqrt_s_t_9 = sqrt(x = clip_1)[name = tensor("sqrt_s_t_9")]; tensor qk_9_transpose_x_1 = const()[name = tensor("qk_9_transpose_x_1"), val = tensor(false)]; tensor qk_9_transpose_y_1 = const()[name = tensor("qk_9_transpose_y_1"), val = tensor(true)]; - tensor k_9 = transpose(perm = k_9_perm_0, x = var_1124)[name = tensor("transpose_26")]; - tensor q_9 = transpose(perm = q_9_perm_0, x = var_1116)[name = tensor("transpose_27")]; + tensor k_9 = transpose(perm = k_9_perm_0, x = var_1157)[name = tensor("transpose_26")]; + tensor q_9 = transpose(perm = q_9_perm_0, x = var_1149)[name = tensor("transpose_27")]; tensor qk_9 = matmul(transpose_x = qk_9_transpose_x_1, transpose_y = qk_9_transpose_y_1, x = q_9, y = k_9)[name = tensor("qk_9")]; - tensor var_1142 = const()[name = tensor("op_1142"), val = tensor([1, 4])]; - tensor var_1143 = reshape(shape = var_1142, x = valid_mask)[name = tensor("op_1143")]; - tensor causal_with_valid_1 = mul(x = decoder__causal_mask, y = var_1143)[name = tensor("causal_with_valid_1")]; - tensor var_1145 = const()[name = tensor("op_1145"), val = tensor([4, 1])]; - tensor var_1146 = reshape(shape = var_1145, x = sqrt_s_t_9)[name = tensor("op_1146")]; - tensor M_9 = real_div(x = causal_with_valid_1, y = var_1146)[name = tensor("M_9")]; - tensor var_1148 = mul(x = qk_9, y = M_9)[name = tensor("op_1148")]; + tensor var_1175 = const()[name = tensor("op_1175"), val = tensor([1, 4])]; + tensor var_1176 = reshape(shape = var_1175, x = valid_mask)[name = tensor("op_1176")]; + tensor causal_with_valid_1 = mul(x = inner_decoder__causal_mask, y = var_1176)[name = tensor("causal_with_valid_1")]; + tensor var_1178 = const()[name = tensor("op_1178"), val = tensor([4, 1])]; + tensor var_1179 = reshape(shape = var_1178, x = sqrt_s_t_9)[name = tensor("op_1179")]; + tensor M_9 = real_div(x = causal_with_valid_1, y = var_1179)[name = tensor("M_9")]; + tensor var_1181 = mul(x = qk_9, y = M_9)[name = tensor("op_1181")]; tensor inner_9_transpose_x_0 = const()[name = tensor("inner_9_transpose_x_0"), val = tensor(false)]; tensor inner_9_transpose_y_0 = const()[name = tensor("inner_9_transpose_y_0"), val = tensor(false)]; - tensor v_9 = transpose(perm = v_9_perm_0, x = var_1130)[name = tensor("transpose_25")]; - tensor inner_9 = matmul(transpose_x = inner_9_transpose_x_0, transpose_y = inner_9_transpose_y_0, x = var_1148, y = v_9)[name = tensor("inner_9")]; - tensor var_1150_transpose_x_0 = const()[name = tensor("op_1150_transpose_x_0"), val = tensor(false)]; - tensor var_1150_transpose_y_0 = const()[name = tensor("op_1150_transpose_y_0"), val = tensor(false)]; - tensor var_1150 = matmul(transpose_x = var_1150_transpose_x_0, transpose_y = var_1150_transpose_y_0, x = q_9, y = prev_kv_9)[name = tensor("op_1150")]; - tensor var_1151 = real_div(x = sqrt_s0_9, y = sqrt_s_t_9)[name = tensor("op_1151")]; - tensor var_1152 = const()[name = tensor("op_1152"), val = tensor([1, 1, 4, 1])]; - tensor var_1153 = reshape(shape = var_1152, x = var_1151)[name = tensor("op_1153")]; - tensor cross_9 = mul(x = var_1150, y = var_1153)[name = tensor("cross_9")]; + tensor v_9 = transpose(perm = v_9_perm_0, x = var_1163)[name = tensor("transpose_25")]; + tensor inner_9 = matmul(transpose_x = inner_9_transpose_x_0, transpose_y = inner_9_transpose_y_0, x = var_1181, y = v_9)[name = tensor("inner_9")]; + tensor var_1183_transpose_x_0 = const()[name = tensor("op_1183_transpose_x_0"), val = tensor(false)]; + tensor var_1183_transpose_y_0 = const()[name = tensor("op_1183_transpose_y_0"), val = tensor(false)]; + tensor var_1183 = matmul(transpose_x = var_1183_transpose_x_0, transpose_y = var_1183_transpose_y_0, x = q_9, y = prev_kv_9)[name = tensor("op_1183")]; + tensor var_1184 = real_div(x = sqrt_s0_9, y = sqrt_s_t_9)[name = tensor("op_1184")]; + tensor var_1185 = const()[name = tensor("op_1185"), val = tensor([1, 1, 4, 1])]; + tensor var_1186 = reshape(shape = var_1185, x = var_1184)[name = tensor("op_1186")]; + tensor cross_9 = mul(x = var_1183, y = var_1186)[name = tensor("cross_9")]; tensor out_25 = add(x = inner_9, y = cross_9)[name = tensor("out_25")]; - tensor var_1156 = const()[name = tensor("op_1156"), val = tensor([1, 1, 4, 1])]; - tensor var_1157 = reshape(shape = var_1156, x = valid_mask)[name = tensor("op_1157")]; - tensor v_masked_1 = mul(x = v_9, y = var_1157)[name = tensor("v_masked_1")]; - tensor var_1159 = mul(x = prev_kv_9, y = sqrt_s0_9)[name = tensor("op_1159")]; - tensor var_1161_transpose_x_1 = const()[name = tensor("op_1161_transpose_x_1"), val = tensor(true)]; - tensor var_1161_transpose_y_1 = const()[name = tensor("op_1161_transpose_y_1"), val = tensor(false)]; - tensor var_1161 = matmul(transpose_x = var_1161_transpose_x_1, transpose_y = var_1161_transpose_y_1, x = k_9, y = v_masked_1)[name = tensor("op_1161")]; - tensor new_kv_unnorm_9 = add(x = var_1159, y = var_1161)[name = tensor("new_kv_unnorm_9")]; - tensor var_1163_keep_dims_0 = const()[name = tensor("op_1163_keep_dims_0"), val = tensor(false)]; - tensor var_1163 = reduce_sum(keep_dims = var_1163_keep_dims_0, x = valid_mask)[name = tensor("op_1163")]; - tensor var_1164 = const()[name = tensor("op_1164"), val = tensor([1])]; - tensor var_1165 = reshape(shape = var_1164, x = var_1163)[name = tensor("op_1165")]; - tensor new_scale_9 = add(x = prev_scale_9, y = var_1165)[name = tensor("new_scale_9")]; + tensor var_1189 = const()[name = tensor("op_1189"), val = tensor([1, 1, 4, 1])]; + tensor var_1190 = reshape(shape = var_1189, x = valid_mask)[name = tensor("op_1190")]; + tensor v_masked_1 = mul(x = v_9, y = var_1190)[name = tensor("v_masked_1")]; + tensor var_1192 = mul(x = prev_kv_9, y = sqrt_s0_9)[name = tensor("op_1192")]; + tensor var_1194_transpose_x_1 = const()[name = tensor("op_1194_transpose_x_1"), val = tensor(true)]; + tensor var_1194_transpose_y_1 = const()[name = tensor("op_1194_transpose_y_1"), val = tensor(false)]; + tensor var_1194 = matmul(transpose_x = var_1194_transpose_x_1, transpose_y = var_1194_transpose_y_1, x = k_9, y = v_masked_1)[name = tensor("op_1194")]; + tensor new_kv_unnorm_9 = add(x = var_1192, y = var_1194)[name = tensor("new_kv_unnorm_9")]; + tensor var_1196_keep_dims_0 = const()[name = tensor("op_1196_keep_dims_0"), val = tensor(false)]; + tensor var_1196 = reduce_sum(keep_dims = var_1196_keep_dims_0, x = valid_mask)[name = tensor("op_1196")]; + tensor var_1197 = const()[name = tensor("op_1197"), val = tensor([1])]; + tensor var_1198 = reshape(shape = var_1197, x = var_1196)[name = tensor("op_1198")]; + tensor new_scale_9 = add(x = prev_scale_9, y = var_1198)[name = tensor("new_scale_9")]; tensor const_21 = const()[name = tensor("const_21"), val = tensor(0x1.fffffep+127)]; - tensor clip_2 = clip(alpha = var_1022, beta = const_21, x = new_scale_9)[name = tensor("clip_2")]; + tensor clip_2 = clip(alpha = var_58, beta = const_21, x = new_scale_9)[name = tensor("clip_2")]; tensor sqrt_new_scale_1 = sqrt(x = clip_2)[name = tensor("sqrt_new_scale_1")]; - tensor var_1169 = real_div(x = new_kv_unnorm_9, y = sqrt_new_scale_1)[name = tensor("op_1169")]; - tensor var_1170_perm_0 = const()[name = tensor("op_1170_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_1202 = real_div(x = new_kv_unnorm_9, y = sqrt_new_scale_1)[name = tensor("op_1202")]; + tensor var_1203_perm_0 = const()[name = tensor("op_1203_perm_0"), val = tensor([0, 2, 1, 3])]; tensor out_27_axes_0 = const()[name = tensor("out_27_axes_0"), val = tensor([-1])]; - tensor var_1170 = transpose(perm = var_1170_perm_0, x = out_25)[name = tensor("transpose_24")]; - tensor out_27 = layer_norm(axes = out_27_axes_0, epsilon = var_1019, x = var_1170)[name = tensor("out_27")]; - tensor var_1174 = const()[name = tensor("op_1174"), val = tensor([6, 4, 256])]; - tensor out_29 = reshape(shape = var_1174, x = out_27)[name = tensor("out_29")]; - tensor var_1176 = silu(x = input_169)[name = tensor("op_1176")]; - tensor input_171 = mul(x = var_1176, y = out_29)[name = tensor("input_171")]; - tensor ret_out_9 = linear(bias = decoder_out_proj_0_bias, weight = decoder_out_proj_0_weight, x = input_171)[name = tensor("linear_42")]; - tensor input_173 = add(x = x_29, y = ret_out_9)[name = tensor("input_173")]; + tensor var_1203 = transpose(perm = var_1203_perm_0, x = out_25)[name = tensor("transpose_24")]; + tensor out_27 = layer_norm(axes = out_27_axes_0, epsilon = var_73, x = var_1203)[name = tensor("out_27")]; + tensor var_1207 = const()[name = tensor("op_1207"), val = tensor([6, 4, 256])]; + tensor out_29 = reshape(shape = var_1207, x = out_27)[name = tensor("out_29")]; + tensor var_1209 = silu(x = input_171)[name = tensor("op_1209")]; + tensor input_173 = mul(x = var_1209, y = out_29)[name = tensor("input_173")]; + tensor ret_out_9 = linear(bias = inner_decoder_out_proj_0_bias, weight = inner_decoder_out_proj_0_weight, x = input_173)[name = tensor("linear_42")]; + tensor input_175 = add(x = x_29, y = ret_out_9)[name = tensor("input_175")]; tensor xt_1_axes_0 = const()[name = tensor("xt_1_axes_0"), val = tensor([-1])]; - tensor xt_1 = layer_norm(axes = xt_1_axes_0, beta = decoder_norm11_0_bias, epsilon = var_1017, gamma = decoder_norm11_0_weight, x = input_173)[name = tensor("xt_1")]; - tensor var_1186 = const()[name = tensor("op_1186"), val = tensor([1, 6, 4, 256])]; - tensor var_1187 = reshape(shape = var_1186, x = xt_1)[name = tensor("op_1187")]; - tensor var_1188_perm_0 = const()[name = tensor("op_1188_perm_0"), val = tensor([0, 2, 1, 3])]; - tensor var_1191 = const()[name = tensor("op_1191"), val = tensor([4, 6, 256])]; - tensor var_1188 = transpose(perm = var_1188_perm_0, x = var_1187)[name = tensor("transpose_23")]; - tensor query_1 = reshape(shape = var_1191, x = var_1188)[name = tensor("query_1")]; + tensor xt_1 = layer_norm(axes = xt_1_axes_0, beta = inner_decoder_norm11_0_bias, epsilon = var_65, gamma = inner_decoder_norm11_0_weight, x = input_175)[name = tensor("xt_1")]; + tensor var_1219 = const()[name = tensor("op_1219"), val = tensor([1, 6, 4, 256])]; + tensor var_1220 = reshape(shape = var_1219, x = xt_1)[name = tensor("op_1220")]; + tensor var_1221_perm_0 = const()[name = tensor("op_1221_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_1224 = const()[name = tensor("op_1224"), val = tensor([4, 6, 256])]; + tensor var_1221 = transpose(perm = var_1221_perm_0, x = var_1220)[name = tensor("transpose_23")]; + tensor query_1 = reshape(shape = var_1224, x = var_1221)[name = tensor("query_1")]; tensor query_3_perm_0 = const()[name = tensor("query_3_perm_0"), val = tensor([1, 0, 2])]; tensor query_3 = transpose(perm = query_3_perm_0, x = query_1)[name = tensor("transpose_22")]; - tensor var_1214 = linear(bias = decoder_self_attn2_0_in_proj_bias, weight = decoder_self_attn2_0_in_proj_weight, x = query_3)[name = tensor("linear_43")]; + tensor var_1247 = linear(bias = inner_decoder_self_attn2_0_in_proj_bias, weight = inner_decoder_self_attn2_0_in_proj_weight, x = query_3)[name = tensor("linear_43")]; tensor concat_1 = const()[name = tensor("concat_1"), val = tensor([6, 4, 3, 256])]; - tensor var_1216 = reshape(shape = concat_1, x = var_1214)[name = tensor("op_1216")]; - tensor var_1217_axes_0 = const()[name = tensor("op_1217_axes_0"), val = tensor([0])]; - tensor var_1217 = expand_dims(axes = var_1217_axes_0, x = var_1216)[name = tensor("op_1217")]; - tensor var_1218_perm_0 = const()[name = tensor("op_1218_perm_0"), val = tensor([-2, 1, 2, 0, 4])]; - tensor var_1219_axes_0 = const()[name = tensor("op_1219_axes_0"), val = tensor([-2])]; - tensor var_1218 = transpose(perm = var_1218_perm_0, x = var_1217)[name = tensor("transpose_21")]; - tensor var_1219 = squeeze(axes = var_1219_axes_0, x = var_1218)[name = tensor("op_1219")]; + tensor var_1249 = reshape(shape = concat_1, x = var_1247)[name = tensor("op_1249")]; + tensor var_1250_axes_0 = const()[name = tensor("op_1250_axes_0"), val = tensor([0])]; + tensor var_1250 = expand_dims(axes = var_1250_axes_0, x = var_1249)[name = tensor("op_1250")]; + tensor var_1251_perm_0 = const()[name = tensor("op_1251_perm_0"), val = tensor([-2, 1, 2, 0, 4])]; + tensor var_1252_axes_0 = const()[name = tensor("op_1252_axes_0"), val = tensor([-2])]; + tensor var_1251 = transpose(perm = var_1251_perm_0, x = var_1250)[name = tensor("transpose_21")]; + tensor var_1252 = squeeze(axes = var_1252_axes_0, x = var_1251)[name = tensor("op_1252")]; tensor q_11_begin_0 = const()[name = tensor("q_11_begin_0"), val = tensor([0, 0, 0, 0])]; tensor q_11_end_0 = const()[name = tensor("q_11_end_0"), val = tensor([1, 6, 4, 256])]; tensor q_11_end_mask_0 = const()[name = tensor("q_11_end_mask_0"), val = tensor([false, true, true, true])]; tensor q_11_squeeze_mask_0 = const()[name = tensor("q_11_squeeze_mask_0"), val = tensor([true, false, false, false])]; - tensor q_11 = slice_by_index(begin = q_11_begin_0, end = q_11_end_0, end_mask = q_11_end_mask_0, squeeze_mask = q_11_squeeze_mask_0, x = var_1219)[name = tensor("q_11")]; + tensor q_11 = slice_by_index(begin = q_11_begin_0, end = q_11_end_0, end_mask = q_11_end_mask_0, squeeze_mask = q_11_squeeze_mask_0, x = var_1252)[name = tensor("q_11")]; tensor k_11_begin_0 = const()[name = tensor("k_11_begin_0"), val = tensor([1, 0, 0, 0])]; tensor k_11_end_0 = const()[name = tensor("k_11_end_0"), val = tensor([2, 6, 4, 256])]; tensor k_11_end_mask_0 = const()[name = tensor("k_11_end_mask_0"), val = tensor([false, true, true, true])]; tensor k_11_squeeze_mask_0 = const()[name = tensor("k_11_squeeze_mask_0"), val = tensor([true, false, false, false])]; - tensor k_11 = slice_by_index(begin = k_11_begin_0, end = k_11_end_0, end_mask = k_11_end_mask_0, squeeze_mask = k_11_squeeze_mask_0, x = var_1219)[name = tensor("k_11")]; + tensor k_11 = slice_by_index(begin = k_11_begin_0, end = k_11_end_0, end_mask = k_11_end_mask_0, squeeze_mask = k_11_squeeze_mask_0, x = var_1252)[name = tensor("k_11")]; tensor v_11_begin_0 = const()[name = tensor("v_11_begin_0"), val = tensor([2, 0, 0, 0])]; tensor v_11_end_0 = const()[name = tensor("v_11_end_0"), val = tensor([3, 6, 4, 256])]; tensor v_11_end_mask_0 = const()[name = tensor("v_11_end_mask_0"), val = tensor([false, true, true, true])]; tensor v_11_squeeze_mask_0 = const()[name = tensor("v_11_squeeze_mask_0"), val = tensor([true, false, false, false])]; - tensor v_11 = slice_by_index(begin = v_11_begin_0, end = v_11_end_0, end_mask = v_11_end_mask_0, squeeze_mask = v_11_squeeze_mask_0, x = var_1219)[name = tensor("v_11")]; - tensor var_1227 = const()[name = tensor("op_1227"), val = tensor([6, 16, 64])]; - tensor var_1228 = reshape(shape = var_1227, x = q_11)[name = tensor("op_1228")]; + tensor v_11 = slice_by_index(begin = v_11_begin_0, end = v_11_end_0, end_mask = v_11_end_mask_0, squeeze_mask = v_11_squeeze_mask_0, x = var_1252)[name = tensor("v_11")]; + tensor var_1260 = const()[name = tensor("op_1260"), val = tensor([6, 16, 64])]; + tensor var_1261 = reshape(shape = var_1260, x = q_11)[name = tensor("op_1261")]; tensor q_13_perm_0 = const()[name = tensor("q_13_perm_0"), val = tensor([1, 0, 2])]; - tensor var_1234 = const()[name = tensor("op_1234"), val = tensor([6, 16, 64])]; - tensor var_1235 = reshape(shape = var_1234, x = k_11)[name = tensor("op_1235")]; + tensor var_1267 = const()[name = tensor("op_1267"), val = tensor([6, 16, 64])]; + tensor var_1268 = reshape(shape = var_1267, x = k_11)[name = tensor("op_1268")]; tensor k_13_perm_0 = const()[name = tensor("k_13_perm_0"), val = tensor([1, 0, 2])]; - tensor var_1241 = const()[name = tensor("op_1241"), val = tensor([6, 16, 64])]; - tensor var_1242 = reshape(shape = var_1241, x = v_11)[name = tensor("op_1242")]; + tensor var_1274 = const()[name = tensor("op_1274"), val = tensor([6, 16, 64])]; + tensor var_1275 = reshape(shape = var_1274, x = v_11)[name = tensor("op_1275")]; tensor v_13_perm_0 = const()[name = tensor("v_13_perm_0"), val = tensor([1, 0, 2])]; - tensor var_1245 = const()[name = tensor("op_1245"), val = tensor([4, 4, 6, 64])]; - tensor q_13 = transpose(perm = q_13_perm_0, x = var_1228)[name = tensor("transpose_20")]; - tensor q_15 = reshape(shape = var_1245, x = q_13)[name = tensor("q_15")]; - tensor var_1247 = const()[name = tensor("op_1247"), val = tensor([4, 4, 6, 64])]; - tensor k_13 = transpose(perm = k_13_perm_0, x = var_1235)[name = tensor("transpose_19")]; - tensor k_15 = reshape(shape = var_1247, x = k_13)[name = tensor("k_15")]; - tensor var_1249 = const()[name = tensor("op_1249"), val = tensor([4, 4, 6, 64])]; - tensor v_13 = transpose(perm = v_13_perm_0, x = var_1242)[name = tensor("transpose_18")]; - tensor v_15 = reshape(shape = var_1249, x = v_13)[name = tensor("v_15")]; + tensor var_1278 = const()[name = tensor("op_1278"), val = tensor([4, 4, 6, 64])]; + tensor q_13 = transpose(perm = q_13_perm_0, x = var_1261)[name = tensor("transpose_20")]; + tensor q_15 = reshape(shape = var_1278, x = q_13)[name = tensor("q_15")]; + tensor var_1280 = const()[name = tensor("op_1280"), val = tensor([4, 4, 6, 64])]; + tensor k_13 = transpose(perm = k_13_perm_0, x = var_1268)[name = tensor("transpose_19")]; + tensor k_15 = reshape(shape = var_1280, x = k_13)[name = tensor("k_15")]; + tensor var_1282 = const()[name = tensor("op_1282"), val = tensor([4, 4, 6, 64])]; + tensor v_13 = transpose(perm = v_13_perm_0, x = var_1275)[name = tensor("transpose_18")]; + tensor v_15 = reshape(shape = var_1282, x = v_13)[name = tensor("v_15")]; tensor mul_1_y_0 = const()[name = tensor("mul_1_y_0"), val = tensor(0x1p-3)]; tensor mul_1 = mul(x = q_15, y = mul_1_y_0)[name = tensor("mul_1")]; tensor matmul_0_transpose_y_0 = const()[name = tensor("matmul_0_transpose_y_0"), val = tensor(true)]; @@ -1119,30 +1134,30 @@ program(1.0) tensor attn_output_1_transpose_x_0 = const()[name = tensor("attn_output_1_transpose_x_0"), val = tensor(false)]; tensor attn_output_1_transpose_y_0 = const()[name = tensor("attn_output_1_transpose_y_0"), val = tensor(false)]; tensor attn_output_1 = matmul(transpose_x = attn_output_1_transpose_x_0, transpose_y = attn_output_1_transpose_y_0, x = softmax_0, y = v_15)[name = tensor("attn_output_1")]; - tensor var_1252 = const()[name = tensor("op_1252"), val = tensor([2, 0, 1, 3])]; - tensor var_1257 = const()[name = tensor("op_1257"), val = tensor([24, 256])]; - tensor var_1253 = transpose(perm = var_1252, x = attn_output_1)[name = tensor("transpose_17")]; - tensor attn_output_3 = reshape(shape = var_1257, x = var_1253)[name = tensor("attn_output_3")]; - tensor attn_output_5 = linear(bias = decoder_self_attn2_0_out_proj_bias, weight = decoder_self_attn2_0_out_proj_weight, x = attn_output_3)[name = tensor("linear_44")]; - tensor var_1261 = const()[name = tensor("op_1261"), val = tensor([6, 4, 256])]; - tensor attn_output_7 = reshape(shape = var_1261, x = attn_output_5)[name = tensor("attn_output_7")]; + tensor var_1285 = const()[name = tensor("op_1285"), val = tensor([2, 0, 1, 3])]; + tensor var_1290 = const()[name = tensor("op_1290"), val = tensor([24, 256])]; + tensor var_1286 = transpose(perm = var_1285, x = attn_output_1)[name = tensor("transpose_17")]; + tensor attn_output_3 = reshape(shape = var_1290, x = var_1286)[name = tensor("attn_output_3")]; + tensor attn_output_5 = linear(bias = inner_decoder_self_attn2_0_out_proj_bias, weight = inner_decoder_self_attn2_0_out_proj_weight, x = attn_output_3)[name = tensor("linear_44")]; + tensor var_1294 = const()[name = tensor("op_1294"), val = tensor([6, 4, 256])]; + tensor attn_output_7 = reshape(shape = var_1294, x = attn_output_5)[name = tensor("attn_output_7")]; tensor sa2_out_1_perm_0 = const()[name = tensor("sa2_out_1_perm_0"), val = tensor([1, 0, 2])]; tensor sa2_out_1 = transpose(perm = sa2_out_1_perm_0, x = attn_output_7)[name = tensor("transpose_16")]; - tensor input_175 = add(x = query_1, y = sa2_out_1)[name = tensor("input_175")]; - tensor input_177_axes_0 = const()[name = tensor("input_177_axes_0"), val = tensor([-1])]; - tensor input_177 = layer_norm(axes = input_177_axes_0, beta = decoder_norm21_0_bias, epsilon = var_1017, gamma = decoder_norm21_0_weight, x = input_175)[name = tensor("input_177")]; - tensor input_179 = linear(bias = decoder_linear1_0_bias, weight = decoder_linear1_0_weight, x = input_177)[name = tensor("linear_45")]; - tensor input_181 = relu(x = input_179)[name = tensor("input_181")]; - tensor ffn_out_1 = linear(bias = decoder_linear2_0_bias, weight = decoder_linear2_0_weight, x = input_181)[name = tensor("linear_46")]; - tensor input_183 = add(x = input_177, y = ffn_out_1)[name = tensor("input_183")]; + tensor input_177 = add(x = query_1, y = sa2_out_1)[name = tensor("input_177")]; + tensor input_179_axes_0 = const()[name = tensor("input_179_axes_0"), val = tensor([-1])]; + tensor input_179 = layer_norm(axes = input_179_axes_0, beta = inner_decoder_norm21_0_bias, epsilon = var_65, gamma = inner_decoder_norm21_0_weight, x = input_177)[name = tensor("input_179")]; + tensor input_181 = linear(bias = inner_decoder_linear1_0_bias, weight = inner_decoder_linear1_0_weight, x = input_179)[name = tensor("linear_45")]; + tensor input_183 = relu(x = input_181)[name = tensor("input_183")]; + tensor ffn_out_1 = linear(bias = inner_decoder_linear2_0_bias, weight = inner_decoder_linear2_0_weight, x = input_183)[name = tensor("linear_46")]; + tensor input_185 = add(x = input_179, y = ffn_out_1)[name = tensor("input_185")]; tensor xt_3_axes_0 = const()[name = tensor("xt_3_axes_0"), val = tensor([-1])]; - tensor xt_3 = layer_norm(axes = xt_3_axes_0, beta = decoder_norm22_0_bias, epsilon = var_1017, gamma = decoder_norm22_0_weight, x = input_183)[name = tensor("xt_3")]; - tensor var_1281 = const()[name = tensor("op_1281"), val = tensor([1, 4, 6, 256])]; - tensor x_31 = reshape(shape = var_1281, x = xt_3)[name = tensor("x_31")]; - tensor var_1283_perm_0 = const()[name = tensor("op_1283_perm_0"), val = tensor([0, 2, 1, 3])]; - tensor var_1287 = const()[name = tensor("op_1287"), val = tensor([6, 4, 256])]; - tensor var_1283 = transpose(perm = var_1283_perm_0, x = x_31)[name = tensor("transpose_15")]; - tensor x = reshape(shape = var_1287, x = var_1283)[name = tensor("x")]; + tensor xt_3 = layer_norm(axes = xt_3_axes_0, beta = inner_decoder_norm22_0_bias, epsilon = var_65, gamma = inner_decoder_norm22_0_weight, x = input_185)[name = tensor("xt_3")]; + tensor var_1314 = const()[name = tensor("op_1314"), val = tensor([1, 4, 6, 256])]; + tensor x_31 = reshape(shape = var_1314, x = xt_3)[name = tensor("x_31")]; + tensor var_1316_perm_0 = const()[name = tensor("op_1316_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_1320 = const()[name = tensor("op_1320"), val = tensor([6, 4, 256])]; + tensor var_1316 = transpose(perm = var_1316_perm_0, x = x_31)[name = tensor("transpose_15")]; + tensor x = reshape(shape = var_1320, x = var_1316)[name = tensor("x")]; tensor prev_kv_begin_0 = const()[name = tensor("prev_kv_begin_0"), val = tensor([1, 0, 0, 0, 0])]; tensor prev_kv_end_0 = const()[name = tensor("prev_kv_end_0"), val = tensor([2, 6, 4, 64, 64])]; tensor prev_kv_end_mask_0 = const()[name = tensor("prev_kv_end_mask_0"), val = tensor([false, true, true, true, true])]; @@ -1153,120 +1168,120 @@ program(1.0) tensor prev_scale_end_mask_0 = const()[name = tensor("prev_scale_end_mask_0"), val = tensor([false, true])]; tensor prev_scale_squeeze_mask_0 = const()[name = tensor("prev_scale_squeeze_mask_0"), val = tensor([true, false])]; tensor prev_scale = slice_by_index(begin = prev_scale_begin_0, end = prev_scale_end_0, end_mask = prev_scale_end_mask_0, squeeze_mask = prev_scale_squeeze_mask_0, x = dec_scale)[name = tensor("prev_scale")]; - tensor var_1295 = linear(bias = decoder_q_proj_1_bias, weight = decoder_q_proj_1_weight, x = x)[name = tensor("linear_47")]; - tensor var_1296 = const()[name = tensor("op_1296"), val = tensor([6, 4, 4, 64])]; - tensor var_1297 = reshape(shape = var_1296, x = var_1295)[name = tensor("op_1297")]; + tensor var_1328 = linear(bias = inner_decoder_q_proj_1_bias, weight = inner_decoder_q_proj_1_weight, x = x)[name = tensor("linear_47")]; + tensor var_1329 = const()[name = tensor("op_1329"), val = tensor([6, 4, 4, 64])]; + tensor var_1330 = reshape(shape = var_1329, x = var_1328)[name = tensor("op_1330")]; tensor q_17_perm_0 = const()[name = tensor("q_17_perm_0"), val = tensor([0, 2, 1, 3])]; - tensor var_1301 = linear(bias = decoder_k_proj_1_bias, weight = decoder_k_proj_1_weight, x = x)[name = tensor("linear_48")]; - tensor var_1302 = const()[name = tensor("op_1302"), val = tensor(0x1p-3)]; - tensor var_1303 = mul(x = var_1301, y = var_1302)[name = tensor("op_1303")]; - tensor var_1304 = const()[name = tensor("op_1304"), val = tensor([6, 4, 4, 64])]; - tensor var_1305 = reshape(shape = var_1304, x = var_1303)[name = tensor("op_1305")]; + tensor var_1334 = linear(bias = inner_decoder_k_proj_1_bias, weight = inner_decoder_k_proj_1_weight, x = x)[name = tensor("linear_48")]; + tensor var_1335 = const()[name = tensor("op_1335"), val = tensor(0x1p-3)]; + tensor var_1336 = mul(x = var_1334, y = var_1335)[name = tensor("op_1336")]; + tensor var_1337 = const()[name = tensor("op_1337"), val = tensor([6, 4, 4, 64])]; + tensor var_1338 = reshape(shape = var_1337, x = var_1336)[name = tensor("op_1338")]; tensor k_17_perm_0 = const()[name = tensor("k_17_perm_0"), val = tensor([0, 2, 1, 3])]; - tensor var_1309 = linear(bias = decoder_v_proj_1_bias, weight = decoder_v_proj_1_weight, x = x)[name = tensor("linear_49")]; - tensor var_1310 = const()[name = tensor("op_1310"), val = tensor([6, 4, 4, 64])]; - tensor var_1311 = reshape(shape = var_1310, x = var_1309)[name = tensor("op_1311")]; + tensor var_1342 = linear(bias = inner_decoder_v_proj_1_bias, weight = inner_decoder_v_proj_1_weight, x = x)[name = tensor("linear_49")]; + tensor var_1343 = const()[name = tensor("op_1343"), val = tensor([6, 4, 4, 64])]; + tensor var_1344 = reshape(shape = var_1343, x = var_1342)[name = tensor("op_1344")]; tensor v_17_perm_0 = const()[name = tensor("v_17_perm_0"), val = tensor([0, 2, 1, 3])]; - tensor input_187 = linear(bias = decoder_g_proj_1_bias, weight = decoder_g_proj_1_weight, x = x)[name = tensor("linear_50")]; + tensor input_189 = linear(bias = inner_decoder_g_proj_1_bias, weight = inner_decoder_g_proj_1_weight, x = x)[name = tensor("linear_50")]; tensor sqrt_s0 = sqrt(x = prev_scale)[name = tensor("sqrt_s0")]; tensor s_t = add(x = prev_scale, y = cumsum_mask_1)[name = tensor("s_t")]; tensor const_32 = const()[name = tensor("const_32"), val = tensor(0x1.fffffep+127)]; - tensor clip_3 = clip(alpha = var_1022, beta = const_32, x = s_t)[name = tensor("clip_3")]; + tensor clip_3 = clip(alpha = var_58, beta = const_32, x = s_t)[name = tensor("clip_3")]; tensor sqrt_s_t = sqrt(x = clip_3)[name = tensor("sqrt_s_t")]; tensor qk_transpose_x_1 = const()[name = tensor("qk_transpose_x_1"), val = tensor(false)]; tensor qk_transpose_y_1 = const()[name = tensor("qk_transpose_y_1"), val = tensor(true)]; - tensor k_17 = transpose(perm = k_17_perm_0, x = var_1305)[name = tensor("transpose_13")]; - tensor q_17 = transpose(perm = q_17_perm_0, x = var_1297)[name = tensor("transpose_14")]; + tensor k_17 = transpose(perm = k_17_perm_0, x = var_1338)[name = tensor("transpose_13")]; + tensor q_17 = transpose(perm = q_17_perm_0, x = var_1330)[name = tensor("transpose_14")]; tensor qk = matmul(transpose_x = qk_transpose_x_1, transpose_y = qk_transpose_y_1, x = q_17, y = k_17)[name = tensor("qk")]; - tensor var_1326 = const()[name = tensor("op_1326"), val = tensor([4, 1])]; - tensor var_1327 = reshape(shape = var_1326, x = sqrt_s_t)[name = tensor("op_1327")]; - tensor M = real_div(x = causal_with_valid_1, y = var_1327)[name = tensor("M")]; - tensor var_1329 = mul(x = qk, y = M)[name = tensor("op_1329")]; - tensor inner_transpose_x_0 = const()[name = tensor("inner_transpose_x_0"), val = tensor(false)]; - tensor inner_transpose_y_0 = const()[name = tensor("inner_transpose_y_0"), val = tensor(false)]; - tensor v_17 = transpose(perm = v_17_perm_0, x = var_1311)[name = tensor("transpose_12")]; - tensor inner = matmul(transpose_x = inner_transpose_x_0, transpose_y = inner_transpose_y_0, x = var_1329, y = v_17)[name = tensor("inner")]; - tensor var_1331_transpose_x_0 = const()[name = tensor("op_1331_transpose_x_0"), val = tensor(false)]; - tensor var_1331_transpose_y_0 = const()[name = tensor("op_1331_transpose_y_0"), val = tensor(false)]; - tensor var_1331 = matmul(transpose_x = var_1331_transpose_x_0, transpose_y = var_1331_transpose_y_0, x = q_17, y = prev_kv)[name = tensor("op_1331")]; - tensor var_1332 = real_div(x = sqrt_s0, y = sqrt_s_t)[name = tensor("op_1332")]; - tensor var_1333 = const()[name = tensor("op_1333"), val = tensor([1, 1, 4, 1])]; - tensor var_1334 = reshape(shape = var_1333, x = var_1332)[name = tensor("op_1334")]; - tensor cross = mul(x = var_1331, y = var_1334)[name = tensor("cross")]; - tensor out_31 = add(x = inner, y = cross)[name = tensor("out_31")]; - tensor v_masked = mul(x = v_17, y = var_1157)[name = tensor("v_masked")]; - tensor var_1340 = mul(x = prev_kv, y = sqrt_s0)[name = tensor("op_1340")]; - tensor var_1342_transpose_x_1 = const()[name = tensor("op_1342_transpose_x_1"), val = tensor(true)]; - tensor var_1342_transpose_y_1 = const()[name = tensor("op_1342_transpose_y_1"), val = tensor(false)]; - tensor var_1342 = matmul(transpose_x = var_1342_transpose_x_1, transpose_y = var_1342_transpose_y_1, x = k_17, y = v_masked)[name = tensor("op_1342")]; - tensor new_kv_unnorm = add(x = var_1340, y = var_1342)[name = tensor("new_kv_unnorm")]; - tensor new_scale = add(x = prev_scale, y = var_1165)[name = tensor("new_scale")]; + tensor var_1359 = const()[name = tensor("op_1359"), val = tensor([4, 1])]; + tensor var_1360 = reshape(shape = var_1359, x = sqrt_s_t)[name = tensor("op_1360")]; + tensor M = real_div(x = causal_with_valid_1, y = var_1360)[name = tensor("M")]; + tensor var_1362 = mul(x = qk, y = M)[name = tensor("op_1362")]; + tensor inner_11_transpose_x_0 = const()[name = tensor("inner_11_transpose_x_0"), val = tensor(false)]; + tensor inner_11_transpose_y_0 = const()[name = tensor("inner_11_transpose_y_0"), val = tensor(false)]; + tensor v_17 = transpose(perm = v_17_perm_0, x = var_1344)[name = tensor("transpose_12")]; + tensor inner_11 = matmul(transpose_x = inner_11_transpose_x_0, transpose_y = inner_11_transpose_y_0, x = var_1362, y = v_17)[name = tensor("inner_11")]; + tensor var_1364_transpose_x_0 = const()[name = tensor("op_1364_transpose_x_0"), val = tensor(false)]; + tensor var_1364_transpose_y_0 = const()[name = tensor("op_1364_transpose_y_0"), val = tensor(false)]; + tensor var_1364 = matmul(transpose_x = var_1364_transpose_x_0, transpose_y = var_1364_transpose_y_0, x = q_17, y = prev_kv)[name = tensor("op_1364")]; + tensor var_1365 = real_div(x = sqrt_s0, y = sqrt_s_t)[name = tensor("op_1365")]; + tensor var_1366 = const()[name = tensor("op_1366"), val = tensor([1, 1, 4, 1])]; + tensor var_1367 = reshape(shape = var_1366, x = var_1365)[name = tensor("op_1367")]; + tensor cross = mul(x = var_1364, y = var_1367)[name = tensor("cross")]; + tensor out_31 = add(x = inner_11, y = cross)[name = tensor("out_31")]; + tensor v_masked = mul(x = v_17, y = var_1190)[name = tensor("v_masked")]; + tensor var_1373 = mul(x = prev_kv, y = sqrt_s0)[name = tensor("op_1373")]; + tensor var_1375_transpose_x_1 = const()[name = tensor("op_1375_transpose_x_1"), val = tensor(true)]; + tensor var_1375_transpose_y_1 = const()[name = tensor("op_1375_transpose_y_1"), val = tensor(false)]; + tensor var_1375 = matmul(transpose_x = var_1375_transpose_x_1, transpose_y = var_1375_transpose_y_1, x = k_17, y = v_masked)[name = tensor("op_1375")]; + tensor new_kv_unnorm = add(x = var_1373, y = var_1375)[name = tensor("new_kv_unnorm")]; + tensor new_scale = add(x = prev_scale, y = var_1198)[name = tensor("new_scale")]; tensor const_33 = const()[name = tensor("const_33"), val = tensor(0x1.fffffep+127)]; - tensor clip_4 = clip(alpha = var_1022, beta = const_33, x = new_scale)[name = tensor("clip_4")]; + tensor clip_4 = clip(alpha = var_58, beta = const_33, x = new_scale)[name = tensor("clip_4")]; tensor sqrt_new_scale = sqrt(x = clip_4)[name = tensor("sqrt_new_scale")]; tensor nkv = real_div(x = new_kv_unnorm, y = sqrt_new_scale)[name = tensor("nkv")]; - tensor var_1351_perm_0 = const()[name = tensor("op_1351_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_1384_perm_0 = const()[name = tensor("op_1384_perm_0"), val = tensor([0, 2, 1, 3])]; tensor out_33_axes_0 = const()[name = tensor("out_33_axes_0"), val = tensor([-1])]; - tensor var_1351 = transpose(perm = var_1351_perm_0, x = out_31)[name = tensor("transpose_11")]; - tensor out_33 = layer_norm(axes = out_33_axes_0, epsilon = var_1019, x = var_1351)[name = tensor("out_33")]; - tensor var_1355 = const()[name = tensor("op_1355"), val = tensor([6, 4, 256])]; - tensor out = reshape(shape = var_1355, x = out_33)[name = tensor("out")]; - tensor var_1357 = silu(x = input_187)[name = tensor("op_1357")]; - tensor input_189 = mul(x = var_1357, y = out)[name = tensor("input_189")]; - tensor ret_out = linear(bias = decoder_out_proj_1_bias, weight = decoder_out_proj_1_weight, x = input_189)[name = tensor("linear_51")]; - tensor input_191 = add(x = x, y = ret_out)[name = tensor("input_191")]; + tensor var_1384 = transpose(perm = var_1384_perm_0, x = out_31)[name = tensor("transpose_11")]; + tensor out_33 = layer_norm(axes = out_33_axes_0, epsilon = var_73, x = var_1384)[name = tensor("out_33")]; + tensor var_1388 = const()[name = tensor("op_1388"), val = tensor([6, 4, 256])]; + tensor out = reshape(shape = var_1388, x = out_33)[name = tensor("out")]; + tensor var_1390 = silu(x = input_189)[name = tensor("op_1390")]; + tensor input_191 = mul(x = var_1390, y = out)[name = tensor("input_191")]; + tensor ret_out = linear(bias = inner_decoder_out_proj_1_bias, weight = inner_decoder_out_proj_1_weight, x = input_191)[name = tensor("linear_51")]; + tensor input_193 = add(x = x, y = ret_out)[name = tensor("input_193")]; tensor xt_5_axes_0 = const()[name = tensor("xt_5_axes_0"), val = tensor([-1])]; - tensor xt_5 = layer_norm(axes = xt_5_axes_0, beta = decoder_norm11_1_bias, epsilon = var_1017, gamma = decoder_norm11_1_weight, x = input_191)[name = tensor("xt_5")]; - tensor var_1367 = const()[name = tensor("op_1367"), val = tensor([1, 6, 4, 256])]; - tensor var_1368 = reshape(shape = var_1367, x = xt_5)[name = tensor("op_1368")]; - tensor var_1369_perm_0 = const()[name = tensor("op_1369_perm_0"), val = tensor([0, 2, 1, 3])]; - tensor var_1372 = const()[name = tensor("op_1372"), val = tensor([4, 6, 256])]; - tensor var_1369 = transpose(perm = var_1369_perm_0, x = var_1368)[name = tensor("transpose_10")]; - tensor query_5 = reshape(shape = var_1372, x = var_1369)[name = tensor("query_5")]; + tensor xt_5 = layer_norm(axes = xt_5_axes_0, beta = inner_decoder_norm11_1_bias, epsilon = var_65, gamma = inner_decoder_norm11_1_weight, x = input_193)[name = tensor("xt_5")]; + tensor var_1400 = const()[name = tensor("op_1400"), val = tensor([1, 6, 4, 256])]; + tensor var_1401 = reshape(shape = var_1400, x = xt_5)[name = tensor("op_1401")]; + tensor var_1402_perm_0 = const()[name = tensor("op_1402_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_1405 = const()[name = tensor("op_1405"), val = tensor([4, 6, 256])]; + tensor var_1402 = transpose(perm = var_1402_perm_0, x = var_1401)[name = tensor("transpose_10")]; + tensor query_5 = reshape(shape = var_1405, x = var_1402)[name = tensor("query_5")]; tensor query_perm_0 = const()[name = tensor("query_perm_0"), val = tensor([1, 0, 2])]; tensor query = transpose(perm = query_perm_0, x = query_5)[name = tensor("transpose_9")]; - tensor var_1395 = linear(bias = decoder_self_attn2_1_in_proj_bias, weight = decoder_self_attn2_1_in_proj_weight, x = query)[name = tensor("linear_52")]; + tensor var_1428 = linear(bias = inner_decoder_self_attn2_1_in_proj_bias, weight = inner_decoder_self_attn2_1_in_proj_weight, x = query)[name = tensor("linear_52")]; tensor concat_2 = const()[name = tensor("concat_2"), val = tensor([6, 4, 3, 256])]; - tensor var_1397 = reshape(shape = concat_2, x = var_1395)[name = tensor("op_1397")]; - tensor var_1398_axes_0 = const()[name = tensor("op_1398_axes_0"), val = tensor([0])]; - tensor var_1398 = expand_dims(axes = var_1398_axes_0, x = var_1397)[name = tensor("op_1398")]; - tensor var_1399_perm_0 = const()[name = tensor("op_1399_perm_0"), val = tensor([-2, 1, 2, 0, 4])]; - tensor var_1400_axes_0 = const()[name = tensor("op_1400_axes_0"), val = tensor([-2])]; - tensor var_1399 = transpose(perm = var_1399_perm_0, x = var_1398)[name = tensor("transpose_8")]; - tensor var_1400 = squeeze(axes = var_1400_axes_0, x = var_1399)[name = tensor("op_1400")]; + tensor var_1430 = reshape(shape = concat_2, x = var_1428)[name = tensor("op_1430")]; + tensor var_1431_axes_0 = const()[name = tensor("op_1431_axes_0"), val = tensor([0])]; + tensor var_1431 = expand_dims(axes = var_1431_axes_0, x = var_1430)[name = tensor("op_1431")]; + tensor var_1432_perm_0 = const()[name = tensor("op_1432_perm_0"), val = tensor([-2, 1, 2, 0, 4])]; + tensor var_1433_axes_0 = const()[name = tensor("op_1433_axes_0"), val = tensor([-2])]; + tensor var_1432 = transpose(perm = var_1432_perm_0, x = var_1431)[name = tensor("transpose_8")]; + tensor var_1433 = squeeze(axes = var_1433_axes_0, x = var_1432)[name = tensor("op_1433")]; tensor q_19_begin_0 = const()[name = tensor("q_19_begin_0"), val = tensor([0, 0, 0, 0])]; tensor q_19_end_0 = const()[name = tensor("q_19_end_0"), val = tensor([1, 6, 4, 256])]; tensor q_19_end_mask_0 = const()[name = tensor("q_19_end_mask_0"), val = tensor([false, true, true, true])]; tensor q_19_squeeze_mask_0 = const()[name = tensor("q_19_squeeze_mask_0"), val = tensor([true, false, false, false])]; - tensor q_19 = slice_by_index(begin = q_19_begin_0, end = q_19_end_0, end_mask = q_19_end_mask_0, squeeze_mask = q_19_squeeze_mask_0, x = var_1400)[name = tensor("q_19")]; + tensor q_19 = slice_by_index(begin = q_19_begin_0, end = q_19_end_0, end_mask = q_19_end_mask_0, squeeze_mask = q_19_squeeze_mask_0, x = var_1433)[name = tensor("q_19")]; tensor k_19_begin_0 = const()[name = tensor("k_19_begin_0"), val = tensor([1, 0, 0, 0])]; tensor k_19_end_0 = const()[name = tensor("k_19_end_0"), val = tensor([2, 6, 4, 256])]; tensor k_19_end_mask_0 = const()[name = tensor("k_19_end_mask_0"), val = tensor([false, true, true, true])]; tensor k_19_squeeze_mask_0 = const()[name = tensor("k_19_squeeze_mask_0"), val = tensor([true, false, false, false])]; - tensor k_19 = slice_by_index(begin = k_19_begin_0, end = k_19_end_0, end_mask = k_19_end_mask_0, squeeze_mask = k_19_squeeze_mask_0, x = var_1400)[name = tensor("k_19")]; + tensor k_19 = slice_by_index(begin = k_19_begin_0, end = k_19_end_0, end_mask = k_19_end_mask_0, squeeze_mask = k_19_squeeze_mask_0, x = var_1433)[name = tensor("k_19")]; tensor v_19_begin_0 = const()[name = tensor("v_19_begin_0"), val = tensor([2, 0, 0, 0])]; tensor v_19_end_0 = const()[name = tensor("v_19_end_0"), val = tensor([3, 6, 4, 256])]; tensor v_19_end_mask_0 = const()[name = tensor("v_19_end_mask_0"), val = tensor([false, true, true, true])]; tensor v_19_squeeze_mask_0 = const()[name = tensor("v_19_squeeze_mask_0"), val = tensor([true, false, false, false])]; - tensor v_19 = slice_by_index(begin = v_19_begin_0, end = v_19_end_0, end_mask = v_19_end_mask_0, squeeze_mask = v_19_squeeze_mask_0, x = var_1400)[name = tensor("v_19")]; - tensor var_1408 = const()[name = tensor("op_1408"), val = tensor([6, 16, 64])]; - tensor var_1409 = reshape(shape = var_1408, x = q_19)[name = tensor("op_1409")]; + tensor v_19 = slice_by_index(begin = v_19_begin_0, end = v_19_end_0, end_mask = v_19_end_mask_0, squeeze_mask = v_19_squeeze_mask_0, x = var_1433)[name = tensor("v_19")]; + tensor var_1441 = const()[name = tensor("op_1441"), val = tensor([6, 16, 64])]; + tensor var_1442 = reshape(shape = var_1441, x = q_19)[name = tensor("op_1442")]; tensor q_21_perm_0 = const()[name = tensor("q_21_perm_0"), val = tensor([1, 0, 2])]; - tensor var_1415 = const()[name = tensor("op_1415"), val = tensor([6, 16, 64])]; - tensor var_1416 = reshape(shape = var_1415, x = k_19)[name = tensor("op_1416")]; + tensor var_1448 = const()[name = tensor("op_1448"), val = tensor([6, 16, 64])]; + tensor var_1449 = reshape(shape = var_1448, x = k_19)[name = tensor("op_1449")]; tensor k_21_perm_0 = const()[name = tensor("k_21_perm_0"), val = tensor([1, 0, 2])]; - tensor var_1422 = const()[name = tensor("op_1422"), val = tensor([6, 16, 64])]; - tensor var_1423 = reshape(shape = var_1422, x = v_19)[name = tensor("op_1423")]; + tensor var_1455 = const()[name = tensor("op_1455"), val = tensor([6, 16, 64])]; + tensor var_1456 = reshape(shape = var_1455, x = v_19)[name = tensor("op_1456")]; tensor v_21_perm_0 = const()[name = tensor("v_21_perm_0"), val = tensor([1, 0, 2])]; - tensor var_1426 = const()[name = tensor("op_1426"), val = tensor([4, 4, 6, 64])]; - tensor q_21 = transpose(perm = q_21_perm_0, x = var_1409)[name = tensor("transpose_7")]; - tensor q = reshape(shape = var_1426, x = q_21)[name = tensor("q")]; - tensor var_1428 = const()[name = tensor("op_1428"), val = tensor([4, 4, 6, 64])]; - tensor k_21 = transpose(perm = k_21_perm_0, x = var_1416)[name = tensor("transpose_6")]; - tensor k = reshape(shape = var_1428, x = k_21)[name = tensor("k")]; - tensor var_1430 = const()[name = tensor("op_1430"), val = tensor([4, 4, 6, 64])]; - tensor v_21 = transpose(perm = v_21_perm_0, x = var_1423)[name = tensor("transpose_5")]; - tensor v = reshape(shape = var_1430, x = v_21)[name = tensor("v")]; + tensor var_1459 = const()[name = tensor("op_1459"), val = tensor([4, 4, 6, 64])]; + tensor q_21 = transpose(perm = q_21_perm_0, x = var_1442)[name = tensor("transpose_7")]; + tensor q = reshape(shape = var_1459, x = q_21)[name = tensor("q")]; + tensor var_1461 = const()[name = tensor("op_1461"), val = tensor([4, 4, 6, 64])]; + tensor k_21 = transpose(perm = k_21_perm_0, x = var_1449)[name = tensor("transpose_6")]; + tensor k = reshape(shape = var_1461, x = k_21)[name = tensor("k")]; + tensor var_1463 = const()[name = tensor("op_1463"), val = tensor([4, 4, 6, 64])]; + tensor v_21 = transpose(perm = v_21_perm_0, x = var_1456)[name = tensor("transpose_5")]; + tensor v = reshape(shape = var_1463, x = v_21)[name = tensor("v")]; tensor mul_3_y_0 = const()[name = tensor("mul_3_y_0"), val = tensor(0x1p-3)]; tensor mul_3 = mul(x = q, y = mul_3_y_0)[name = tensor("mul_3")]; tensor matmul_1_transpose_y_0 = const()[name = tensor("matmul_1_transpose_y_0"), val = tensor(true)]; @@ -1277,36 +1292,36 @@ program(1.0) tensor attn_output_9_transpose_x_0 = const()[name = tensor("attn_output_9_transpose_x_0"), val = tensor(false)]; tensor attn_output_9_transpose_y_0 = const()[name = tensor("attn_output_9_transpose_y_0"), val = tensor(false)]; tensor attn_output_9 = matmul(transpose_x = attn_output_9_transpose_x_0, transpose_y = attn_output_9_transpose_y_0, x = softmax_1, y = v)[name = tensor("attn_output_9")]; - tensor var_1433 = const()[name = tensor("op_1433"), val = tensor([2, 0, 1, 3])]; - tensor var_1438 = const()[name = tensor("op_1438"), val = tensor([24, 256])]; - tensor var_1434 = transpose(perm = var_1433, x = attn_output_9)[name = tensor("transpose_4")]; - tensor attn_output_11 = reshape(shape = var_1438, x = var_1434)[name = tensor("attn_output_11")]; - tensor attn_output_13 = linear(bias = decoder_self_attn2_1_out_proj_bias, weight = decoder_self_attn2_1_out_proj_weight, x = attn_output_11)[name = tensor("linear_53")]; - tensor var_1442 = const()[name = tensor("op_1442"), val = tensor([6, 4, 256])]; - tensor attn_output = reshape(shape = var_1442, x = attn_output_13)[name = tensor("attn_output")]; + tensor var_1466 = const()[name = tensor("op_1466"), val = tensor([2, 0, 1, 3])]; + tensor var_1471 = const()[name = tensor("op_1471"), val = tensor([24, 256])]; + tensor var_1467 = transpose(perm = var_1466, x = attn_output_9)[name = tensor("transpose_4")]; + tensor attn_output_11 = reshape(shape = var_1471, x = var_1467)[name = tensor("attn_output_11")]; + tensor attn_output_13 = linear(bias = inner_decoder_self_attn2_1_out_proj_bias, weight = inner_decoder_self_attn2_1_out_proj_weight, x = attn_output_11)[name = tensor("linear_53")]; + tensor var_1475 = const()[name = tensor("op_1475"), val = tensor([6, 4, 256])]; + tensor attn_output = reshape(shape = var_1475, x = attn_output_13)[name = tensor("attn_output")]; tensor sa2_out_perm_0 = const()[name = tensor("sa2_out_perm_0"), val = tensor([1, 0, 2])]; tensor sa2_out = transpose(perm = sa2_out_perm_0, x = attn_output)[name = tensor("transpose_3")]; - tensor input_193 = add(x = query_5, y = sa2_out)[name = tensor("input_193")]; - tensor input_195_axes_0 = const()[name = tensor("input_195_axes_0"), val = tensor([-1])]; - tensor input_195 = layer_norm(axes = input_195_axes_0, beta = decoder_norm21_1_bias, epsilon = var_1017, gamma = decoder_norm21_1_weight, x = input_193)[name = tensor("input_195")]; - tensor input_197 = linear(bias = decoder_linear1_1_bias, weight = decoder_linear1_1_weight, x = input_195)[name = tensor("linear_54")]; - tensor input_199 = relu(x = input_197)[name = tensor("input_199")]; - tensor ffn_out = linear(bias = decoder_linear2_1_bias, weight = decoder_linear2_1_weight, x = input_199)[name = tensor("linear_55")]; - tensor input_201 = add(x = input_195, y = ffn_out)[name = tensor("input_201")]; + tensor input_195 = add(x = query_5, y = sa2_out)[name = tensor("input_195")]; + tensor input_197_axes_0 = const()[name = tensor("input_197_axes_0"), val = tensor([-1])]; + tensor input_197 = layer_norm(axes = input_197_axes_0, beta = inner_decoder_norm21_1_bias, epsilon = var_65, gamma = inner_decoder_norm21_1_weight, x = input_195)[name = tensor("input_197")]; + tensor input_199 = linear(bias = inner_decoder_linear1_1_bias, weight = inner_decoder_linear1_1_weight, x = input_197)[name = tensor("linear_54")]; + tensor input_201 = relu(x = input_199)[name = tensor("input_201")]; + tensor ffn_out = linear(bias = inner_decoder_linear2_1_bias, weight = inner_decoder_linear2_1_weight, x = input_201)[name = tensor("linear_55")]; + tensor input_203 = add(x = input_197, y = ffn_out)[name = tensor("input_203")]; tensor xt_axes_0 = const()[name = tensor("xt_axes_0"), val = tensor([-1])]; - tensor xt = layer_norm(axes = xt_axes_0, beta = decoder_norm22_1_bias, epsilon = var_1017, gamma = decoder_norm22_1_weight, x = input_201)[name = tensor("xt")]; - tensor var_1462 = const()[name = tensor("op_1462"), val = tensor([1, 4, 6, 256])]; - tensor input = reshape(shape = var_1462, x = xt)[name = tensor("input")]; - tensor var_1464 = const()[name = tensor("op_1464"), val = tensor([-1])]; - tensor var_1465 = reduce_l2_norm(axes = var_1464, keep_dims = var_1020, x = input)[name = tensor("op_1465")]; + tensor xt = layer_norm(axes = xt_axes_0, beta = inner_decoder_norm22_1_bias, epsilon = var_65, gamma = inner_decoder_norm22_1_weight, x = input_203)[name = tensor("xt")]; + tensor var_1495 = const()[name = tensor("op_1495"), val = tensor([1, 4, 6, 256])]; + tensor input = reshape(shape = var_1495, x = xt)[name = tensor("input")]; + tensor var_1497 = const()[name = tensor("op_1497"), val = tensor([-1])]; + tensor var_1498 = reduce_l2_norm(axes = var_1497, keep_dims = var_64, x = input)[name = tensor("op_1498")]; tensor const_42 = const()[name = tensor("const_42"), val = tensor(0x1.fffffep+127)]; - tensor clip_5 = clip(alpha = var_1012, beta = const_42, x = var_1465)[name = tensor("clip_5")]; - tensor var_1467 = real_div(x = input, y = clip_5)[name = tensor("op_1467")]; + tensor clip_5 = clip(alpha = var_78, beta = const_42, x = var_1498)[name = tensor("clip_5")]; + tensor var_1500 = real_div(x = input, y = clip_5)[name = tensor("op_1500")]; tensor concat_6 = const()[name = tensor("concat_6"), val = tensor([4, 1, 256])]; tensor reshape_0 = reshape(shape = concat_6, x = emb)[name = tensor("reshape_0")]; tensor transpose_1_perm_0 = const()[name = tensor("transpose_1_perm_0"), val = tensor([0, 1, 3, 2])]; tensor concat_7 = const()[name = tensor("concat_7"), val = tensor([4, 256, 6])]; - tensor transpose_1 = transpose(perm = transpose_1_perm_0, x = var_1467)[name = tensor("transpose_2")]; + tensor transpose_1 = transpose(perm = transpose_1_perm_0, x = var_1500)[name = tensor("transpose_2")]; tensor reshape_1 = reshape(shape = concat_7, x = transpose_1)[name = tensor("reshape_1")]; tensor matmul_2_transpose_x_0 = const()[name = tensor("matmul_2_transpose_x_0"), val = tensor(false)]; tensor matmul_2_transpose_y_0 = const()[name = tensor("matmul_2_transpose_y_0"), val = tensor(false)]; @@ -1317,10 +1332,10 @@ program(1.0) tensor output_end_0 = const()[name = tensor("output_end_0"), val = tensor([1, 4, 5])]; tensor output_end_mask_0 = const()[name = tensor("output_end_mask_0"), val = tensor([true, true, false])]; tensor output = slice_by_index(begin = output_begin_0, end = output_end_0, end_mask = output_end_mask_0, x = reshape_2)[name = tensor("output")]; - tensor probs = sigmoid(x = output)[name = tensor("op_1471")]; - tensor var_1473_axis_0 = const()[name = tensor("op_1473_axis_0"), val = tensor(0)]; - tensor dec_kv_new = stack(axis = var_1473_axis_0, values = (var_1169, nkv))[name = tensor("op_1473")]; - tensor var_1475_axis_0 = const()[name = tensor("op_1475_axis_0"), val = tensor(0)]; - tensor dec_scale_new = stack(axis = var_1475_axis_0, values = (new_scale_9, new_scale))[name = tensor("op_1475")]; + tensor probs = sigmoid(x = output)[name = tensor("op_1504")]; + tensor var_1506_axis_0 = const()[name = tensor("op_1506_axis_0"), val = tensor(0)]; + tensor dec_kv_new = stack(axis = var_1506_axis_0, values = (var_1202, nkv))[name = tensor("op_1506")]; + tensor var_1508_axis_0 = const()[name = tensor("op_1508_axis_0"), val = tensor(0)]; + tensor dec_scale_new = stack(axis = var_1508_axis_0, values = (new_scale_9, new_scale))[name = tensor("op_1508")]; } -> (probs, enc_kv_new, enc_scale_new, enc_conv_cache_new, cnn_window_new, dec_kv_new, dec_scale_new); } \ No newline at end of file diff --git a/optimized/ami/400ms/ls_eend_ami_400ms.mlpackage/Data/com.apple.CoreML/model.mlmodel b/optimized/ami/400ms/ls_eend_ami_400ms.mlpackage/Data/com.apple.CoreML/model.mlmodel index 11bd6cc8bc201ab8dc743d904c5a906849392429..67687bbb2394cad079080ece84b036a41b37d124 100644 --- a/optimized/ami/400ms/ls_eend_ami_400ms.mlpackage/Data/com.apple.CoreML/model.mlmodel +++ b/optimized/ami/400ms/ls_eend_ami_400ms.mlpackage/Data/com.apple.CoreML/model.mlmodel @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:3bf0e98390dbb2eac96d1c7400ce29fa396132e83c940c9f6fb9aa1d779f97c3 -size 191035 +oid sha256:139c50454d888783e9c845f69f1948c913ba3b6d9e5773f5abae72887cb75a72 +size 197107 diff --git a/optimized/ami/400ms/ls_eend_ami_400ms.mlpackage/Manifest.json b/optimized/ami/400ms/ls_eend_ami_400ms.mlpackage/Manifest.json index 4ea926ed488b5227c9d9a317766f7e6b911916dd..26fc622c438e51bb17223f8c3f9f85d13685006a 100644 --- a/optimized/ami/400ms/ls_eend_ami_400ms.mlpackage/Manifest.json +++ b/optimized/ami/400ms/ls_eend_ami_400ms.mlpackage/Manifest.json @@ -1,18 +1,18 @@ { "fileFormatVersion": "1.0.0", "itemInfoEntries": { - "111C110F-82D7-4E61-A58C-05908C6E61C8": { - "author": "com.apple.CoreML", - "description": "CoreML Model Weights", - "name": "weights", - "path": "com.apple.CoreML/weights" - }, - "A106D900-1931-4022-84A9-F374DF9D3F12": { + "DEA911EA-DD5F-40B3-ACAD-71757B857265": { "author": "com.apple.CoreML", "description": "CoreML Model Specification", "name": "model.mlmodel", "path": "com.apple.CoreML/model.mlmodel" + }, + "EE26D82A-DA8A-4B4C-938F-63934CA9722D": { + "author": "com.apple.CoreML", + "description": "CoreML Model Weights", + "name": "weights", + "path": "com.apple.CoreML/weights" } }, - "rootModelIdentifier": "A106D900-1931-4022-84A9-F374DF9D3F12" + "rootModelIdentifier": "DEA911EA-DD5F-40B3-ACAD-71757B857265" } diff --git a/optimized/ami/500ms/ls_eend_ami_500ms.mlmodelc/analytics/coremldata.bin b/optimized/ami/500ms/ls_eend_ami_500ms.mlmodelc/analytics/coremldata.bin index 38627aa1b674f9fdb4fc74b0266575b8228d9989..329d62e0e526a6645a8be26cf6cdc4659b12e0c9 100644 --- a/optimized/ami/500ms/ls_eend_ami_500ms.mlmodelc/analytics/coremldata.bin +++ b/optimized/ami/500ms/ls_eend_ami_500ms.mlmodelc/analytics/coremldata.bin @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:e15806e285f1582d15c27637594c49104f55b3b7b72abdaad4b32313223bee5b +oid sha256:8c8d6032e92c8c43fe974f203d4a9041453e83932bbc133eaa00605afe3464b4 size 243 diff --git a/optimized/ami/500ms/ls_eend_ami_500ms.mlmodelc/coremldata.bin b/optimized/ami/500ms/ls_eend_ami_500ms.mlmodelc/coremldata.bin index 93ccd031fa1a60d8bc1e9e1900eebe4ed013ded1..a6c28c2ec1c0ed0e854ae17762c90fa173e4266b 100644 --- a/optimized/ami/500ms/ls_eend_ami_500ms.mlmodelc/coremldata.bin +++ b/optimized/ami/500ms/ls_eend_ami_500ms.mlmodelc/coremldata.bin @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:77dd81a1ebcc4c55be5fef0d9f8639561671cb3d7b23bed333b23b53fe9a5ab7 -size 1292 +oid sha256:984dbda533de03f37b496d41400be4fb0cad97233106f868482ef79302526a38 +size 1395 diff --git a/optimized/ami/500ms/ls_eend_ami_500ms.mlmodelc/metadata.json b/optimized/ami/500ms/ls_eend_ami_500ms.mlmodelc/metadata.json index 410cbbe3ea4581edc9b16eeffe96a1435ac493c2..3ec79fce1e284c0951daf5eba6209646633f21bd 100644 --- a/optimized/ami/500ms/ls_eend_ami_500ms.mlmodelc/metadata.json +++ b/optimized/ami/500ms/ls_eend_ami_500ms.mlmodelc/metadata.json @@ -1,7 +1,7 @@ [ { "metadataOutputVersion" : "3.0", - "shortDescription" : "LS-EEND AMI streaming diarizer (pipeline, T=5, max_speakers=4)", + "shortDescription" : "LS-EEND AMI streaming diarizer (pipeline, T=5, max_speakers=4, layout=raw_mel, cu=all)", "outputSchema" : [ { "hasShapeFlexibility" : "0", @@ -81,7 +81,7 @@ "specificationVersion" : 8, "mlProgramOperationTypeHistogram" : { "Ios16.reduceL2Norm" : 2, - "Ios17.reshape" : 68, + "Ios17.reshape" : 69, "Ios16.softmax" : 2, "Ios17.matmul" : 29, "Ios17.transpose" : 57, @@ -89,7 +89,7 @@ "Ios17.expandDims" : 3, "Ios17.add" : 46, "Ios16.sigmoid" : 5, - "Ios17.sliceByIndex" : 72, + "Ios17.sliceByIndex" : 77, "Tile" : 1, "Ios16.reduceSum" : 1, "Ios17.squeeze" : 2, @@ -101,7 +101,7 @@ "Ios16.silu" : 18, "Ios17.realDiv" : 20, "Ios17.linear" : 56, - "Stack" : 5, + "Stack" : 6, "Ios17.concat" : 26, "Ios16.relu" : 2, "Ios16.cumsum" : 1, @@ -128,9 +128,9 @@ "hasShapeFlexibility" : "0", "isOptional" : "0", "dataType" : "Float32", - "formattedType" : "MultiArray (Float32 1 × 5 × 345)", + "formattedType" : "MultiArray (Float32 1 × 55 × 23)", "shortDescription" : "", - "shape" : "[1, 5, 345]", + "shape" : "[1, 55, 23]", "name" : "features", "type" : "MultiArray" }, @@ -206,8 +206,8 @@ } ], "userDefinedMetadata" : { - "com.github.apple.coremltools.conversion_date" : "2026-04-16", - "config" : "{\"model_name\": \"ami\", \"model_label\": \"AMI\", \"variant\": \"pipeline\", \"chunk_size\": 5, \"step_duration_ms\": 500, \"frame_duration_ms\": 100, \"frame_duration_seconds\": 0.1, \"max_speakers\": 4, \"max_nspks\": 6, \"n_units\": 256, \"n_heads\": 4, \"enc_n_layers\": 4, \"dec_n_layers\": 2, \"conv_kernel_size\": 16, \"conv_delay\": 9, \"sample_rate\": 8000, \"win_length\": 200, \"hop_length\": 80, \"n_mels\": 23, \"context_size\": 7, \"subsampling\": 10, \"feat_type\": \"logmel23_cummn\", \"pure_roll\": true}", + "com.github.apple.coremltools.conversion_date" : "2026-04-18", + "config" : "{\"model_name\": \"ami\", \"model_label\": \"AMI\", \"variant\": \"pipeline\", \"chunk_size\": 5, \"step_duration_ms\": 500, \"frame_duration_ms\": 100, \"frame_duration_seconds\": 0.1, \"max_speakers\": 4, \"max_nspks\": 6, \"n_units\": 256, \"n_heads\": 4, \"enc_n_layers\": 4, \"dec_n_layers\": 2, \"conv_kernel_size\": 16, \"conv_delay\": 9, \"sample_rate\": 8000, \"win_length\": 200, \"hop_length\": 80, \"n_mels\": 23, \"context_size\": 7, \"subsampling\": 10, \"feat_type\": \"logmel23_cummn\", \"pure_roll\": true, \"input_layout\": \"raw_mel\", \"compute_units_export\": \"all\", \"raw_mel_length\": 55}", "com.github.apple.coremltools.source" : "torch==2.6.0", "com.github.apple.coremltools.version" : "9.0", "com.github.apple.coremltools.source_dialect" : "TorchScript" diff --git a/optimized/ami/500ms/ls_eend_ami_500ms.mlmodelc/model.mil b/optimized/ami/500ms/ls_eend_ami_500ms.mlmodelc/model.mil index f062932fd3c10f5f9a6925badb848bf8e06e6c01..7f57cbf699017c4f570f6ccdd73a5252e8b8642e 100644 --- a/optimized/ami/500ms/ls_eend_ami_500ms.mlmodelc/model.mil +++ b/optimized/ami/500ms/ls_eend_ami_500ms.mlmodelc/model.mil @@ -1,234 +1,260 @@ program(1.0) [buildInfo = dict, tensor>({{"coremlc-component-MIL", "3510.2.1"}, {"coremlc-version", "3500.32.1"}, {"coremltools-component-torch", "2.6.0"}, {"coremltools-source-dialect", "TorchScript"}, {"coremltools-version", "9.0"}})] { - func main(tensor cnn_window, tensor dec_kv, tensor dec_scale, tensor enc_conv_cache, tensor enc_kv, tensor enc_scale, tensor features, tensor valid_mask) { - tensor encoder_cnn_bias = const()[name = tensor("encoder_cnn_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(64)))]; - tensor encoder_cnn_weight = const()[name = tensor("encoder_cnn_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1152)))]; - tensor encoder__causal_mask = const()[name = tensor("encoder__causal_mask"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4981952)))]; - tensor encoder__t_index = const()[name = tensor("encoder__t_index"), val = tensor([0x1p+0, 0x1p+1, 0x1.8p+1, 0x1p+2, 0x1.4p+2])]; - tensor encoder_input_projection_linear_bias = const()[name = tensor("encoder_input_projection_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4982144)))]; - tensor encoder_input_projection_linear_weight = const()[name = tensor("encoder_input_projection_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4983232)))]; - tensor encoder_pre_layer_norm_bias = const()[name = tensor("encoder_pre_layer_norm_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(5336576)))]; - tensor encoder_pre_layer_norm_weight = const()[name = tensor("encoder_pre_layer_norm_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(5337664)))]; - tensor encoder_ffn1_0_module_sequential_0_bias = const()[name = tensor("encoder_ffn1_0_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(5338752)))]; - tensor encoder_ffn1_0_module_sequential_0_weight = const()[name = tensor("encoder_ffn1_0_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(5339840)))]; - tensor encoder_ffn1_0_module_sequential_1_linear_bias = const()[name = tensor("encoder_ffn1_0_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(5340928)))]; - tensor encoder_ffn1_0_module_sequential_1_linear_weight = const()[name = tensor("encoder_ffn1_0_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(5345088)))]; - tensor encoder_ffn1_0_module_sequential_4_linear_bias = const()[name = tensor("encoder_ffn1_0_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(6393728)))]; - tensor encoder_ffn1_0_module_sequential_4_linear_weight = const()[name = tensor("encoder_ffn1_0_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(6394816)))]; - tensor encoder_ret_lns_0_bias = const()[name = tensor("encoder_ret_lns_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7443456)))]; - tensor encoder_ret_lns_0_weight = const()[name = tensor("encoder_ret_lns_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7444544)))]; - tensor encoder_q_proj_0_bias = const()[name = tensor("encoder_q_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7445632)))]; - tensor encoder_q_proj_0_weight = const()[name = tensor("encoder_q_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7446720)))]; - tensor encoder_k_proj_0_bias = const()[name = tensor("encoder_k_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7708928)))]; - tensor encoder_k_proj_0_weight = const()[name = tensor("encoder_k_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7710016)))]; - tensor encoder_v_proj_0_bias = const()[name = tensor("encoder_v_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7972224)))]; - tensor encoder_v_proj_0_weight = const()[name = tensor("encoder_v_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7973312)))]; - tensor encoder_g_proj_0_bias = const()[name = tensor("encoder_g_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8235520)))]; - tensor encoder_g_proj_0_weight = const()[name = tensor("encoder_g_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8236608)))]; - tensor encoder_out_proj_0_bias = const()[name = tensor("encoder_out_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8498816)))]; - tensor encoder_out_proj_0_weight = const()[name = tensor("encoder_out_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8499904)))]; - tensor encoder_conv_module_0_sequential_0_bias = const()[name = tensor("encoder_conv_module_0_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8762112)))]; - tensor encoder_conv_module_0_sequential_0_weight = const()[name = tensor("encoder_conv_module_0_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8763200)))]; - tensor encoder_conv_module_0_sequential_2_conv_bias = const()[name = tensor("encoder_conv_module_0_sequential_2_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8764288)))]; - tensor encoder_conv_module_0_sequential_2_conv_weight = const()[name = tensor("encoder_conv_module_0_sequential_2_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8766400)))]; - tensor encoder_conv_module_0_sequential_4_conv_weight = const()[name = tensor("encoder_conv_module_0_sequential_4_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9290752)))]; - tensor encoder_conv_module_0_sequential_5_running_var = const()[name = tensor("encoder_conv_module_0_sequential_5_running_var"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9307200)))]; - tensor encoder_conv_module_0_sequential_5_running_mean = const()[name = tensor("encoder_conv_module_0_sequential_5_running_mean"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9308288)))]; - tensor encoder_conv_module_0_sequential_5_bias = const()[name = tensor("encoder_conv_module_0_sequential_5_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9309376)))]; - tensor encoder_conv_module_0_sequential_5_weight = const()[name = tensor("encoder_conv_module_0_sequential_5_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9310464)))]; - tensor encoder_conv_module_0_sequential_7_conv_bias = const()[name = tensor("encoder_conv_module_0_sequential_7_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9311552)))]; - tensor encoder_conv_module_0_sequential_7_conv_weight = const()[name = tensor("encoder_conv_module_0_sequential_7_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9312640)))]; - tensor encoder_ffn2_0_module_sequential_0_bias = const()[name = tensor("encoder_ffn2_0_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9574848)))]; - tensor encoder_ffn2_0_module_sequential_0_weight = const()[name = tensor("encoder_ffn2_0_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9575936)))]; - tensor encoder_ffn2_0_module_sequential_1_linear_bias = const()[name = tensor("encoder_ffn2_0_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9577024)))]; - tensor encoder_ffn2_0_module_sequential_1_linear_weight = const()[name = tensor("encoder_ffn2_0_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9581184)))]; - tensor encoder_ffn2_0_module_sequential_4_linear_bias = const()[name = tensor("encoder_ffn2_0_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(10629824)))]; - tensor encoder_ffn2_0_module_sequential_4_linear_weight = const()[name = tensor("encoder_ffn2_0_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(10630912)))]; - tensor encoder_layer_norm_0_bias = const()[name = tensor("encoder_layer_norm_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11679552)))]; - tensor encoder_layer_norm_0_weight = const()[name = tensor("encoder_layer_norm_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11680640)))]; - tensor encoder_ffn1_1_module_sequential_0_bias = const()[name = tensor("encoder_ffn1_1_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11681728)))]; - tensor encoder_ffn1_1_module_sequential_0_weight = const()[name = tensor("encoder_ffn1_1_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11682816)))]; - tensor encoder_ffn1_1_module_sequential_1_linear_bias = const()[name = tensor("encoder_ffn1_1_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11683904)))]; - tensor encoder_ffn1_1_module_sequential_1_linear_weight = const()[name = tensor("encoder_ffn1_1_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11688064)))]; - tensor encoder_ffn1_1_module_sequential_4_linear_bias = const()[name = tensor("encoder_ffn1_1_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(12736704)))]; - tensor encoder_ffn1_1_module_sequential_4_linear_weight = const()[name = tensor("encoder_ffn1_1_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(12737792)))]; - tensor encoder_ret_lns_1_bias = const()[name = tensor("encoder_ret_lns_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(13786432)))]; - tensor encoder_ret_lns_1_weight = const()[name = tensor("encoder_ret_lns_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(13787520)))]; - tensor encoder_q_proj_1_bias = const()[name = tensor("encoder_q_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(13788608)))]; - tensor encoder_q_proj_1_weight = const()[name = tensor("encoder_q_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(13789696)))]; - tensor encoder_k_proj_1_bias = const()[name = tensor("encoder_k_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14051904)))]; - tensor encoder_k_proj_1_weight = const()[name = tensor("encoder_k_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14052992)))]; - tensor encoder_v_proj_1_bias = const()[name = tensor("encoder_v_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14315200)))]; - tensor encoder_v_proj_1_weight = const()[name = tensor("encoder_v_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14316288)))]; - tensor encoder_g_proj_1_bias = const()[name = tensor("encoder_g_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14578496)))]; - tensor encoder_g_proj_1_weight = const()[name = tensor("encoder_g_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14579584)))]; - tensor encoder_out_proj_1_bias = const()[name = tensor("encoder_out_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14841792)))]; - tensor encoder_out_proj_1_weight = const()[name = tensor("encoder_out_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14842880)))]; - tensor encoder_conv_module_1_sequential_0_bias = const()[name = tensor("encoder_conv_module_1_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15105088)))]; - tensor encoder_conv_module_1_sequential_0_weight = const()[name = tensor("encoder_conv_module_1_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15106176)))]; - tensor encoder_conv_module_1_sequential_2_conv_bias = const()[name = tensor("encoder_conv_module_1_sequential_2_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15107264)))]; - tensor encoder_conv_module_1_sequential_2_conv_weight = const()[name = tensor("encoder_conv_module_1_sequential_2_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15109376)))]; - tensor encoder_conv_module_1_sequential_4_conv_weight = const()[name = tensor("encoder_conv_module_1_sequential_4_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15633728)))]; - tensor encoder_conv_module_1_sequential_5_running_var = const()[name = tensor("encoder_conv_module_1_sequential_5_running_var"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15650176)))]; - tensor encoder_conv_module_1_sequential_5_running_mean = const()[name = tensor("encoder_conv_module_1_sequential_5_running_mean"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15651264)))]; - tensor encoder_conv_module_1_sequential_5_bias = const()[name = tensor("encoder_conv_module_1_sequential_5_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15652352)))]; - tensor encoder_conv_module_1_sequential_5_weight = const()[name = tensor("encoder_conv_module_1_sequential_5_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15653440)))]; - tensor encoder_conv_module_1_sequential_7_conv_bias = const()[name = tensor("encoder_conv_module_1_sequential_7_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15654528)))]; - tensor encoder_conv_module_1_sequential_7_conv_weight = const()[name = tensor("encoder_conv_module_1_sequential_7_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15655616)))]; - tensor encoder_ffn2_1_module_sequential_0_bias = const()[name = tensor("encoder_ffn2_1_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15917824)))]; - tensor encoder_ffn2_1_module_sequential_0_weight = const()[name = tensor("encoder_ffn2_1_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15918912)))]; - tensor encoder_ffn2_1_module_sequential_1_linear_bias = const()[name = tensor("encoder_ffn2_1_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15920000)))]; - tensor encoder_ffn2_1_module_sequential_1_linear_weight = const()[name = tensor("encoder_ffn2_1_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15924160)))]; - tensor encoder_ffn2_1_module_sequential_4_linear_bias = const()[name = tensor("encoder_ffn2_1_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(16972800)))]; - tensor encoder_ffn2_1_module_sequential_4_linear_weight = const()[name = tensor("encoder_ffn2_1_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(16973888)))]; - tensor encoder_layer_norm_1_bias = const()[name = tensor("encoder_layer_norm_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18022528)))]; - tensor encoder_layer_norm_1_weight = const()[name = tensor("encoder_layer_norm_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18023616)))]; - tensor encoder_ffn1_2_module_sequential_0_bias = const()[name = tensor("encoder_ffn1_2_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18024704)))]; - tensor encoder_ffn1_2_module_sequential_0_weight = const()[name = tensor("encoder_ffn1_2_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18025792)))]; - tensor encoder_ffn1_2_module_sequential_1_linear_bias = const()[name = tensor("encoder_ffn1_2_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18026880)))]; - tensor encoder_ffn1_2_module_sequential_1_linear_weight = const()[name = tensor("encoder_ffn1_2_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18031040)))]; - tensor encoder_ffn1_2_module_sequential_4_linear_bias = const()[name = tensor("encoder_ffn1_2_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(19079680)))]; - tensor encoder_ffn1_2_module_sequential_4_linear_weight = const()[name = tensor("encoder_ffn1_2_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(19080768)))]; - tensor encoder_ret_lns_2_bias = const()[name = tensor("encoder_ret_lns_2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20129408)))]; - tensor encoder_ret_lns_2_weight = const()[name = tensor("encoder_ret_lns_2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20130496)))]; - tensor encoder_q_proj_2_bias = const()[name = tensor("encoder_q_proj_2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20131584)))]; - tensor encoder_q_proj_2_weight = const()[name = tensor("encoder_q_proj_2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20132672)))]; - tensor encoder_k_proj_2_bias = const()[name = tensor("encoder_k_proj_2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20394880)))]; - tensor encoder_k_proj_2_weight = const()[name = tensor("encoder_k_proj_2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20395968)))]; - tensor encoder_v_proj_2_bias = const()[name = tensor("encoder_v_proj_2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20658176)))]; - tensor encoder_v_proj_2_weight = const()[name = tensor("encoder_v_proj_2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20659264)))]; - tensor encoder_g_proj_2_bias = const()[name = tensor("encoder_g_proj_2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20921472)))]; - tensor encoder_g_proj_2_weight = const()[name = tensor("encoder_g_proj_2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20922560)))]; - tensor encoder_out_proj_2_bias = const()[name = tensor("encoder_out_proj_2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21184768)))]; - tensor encoder_out_proj_2_weight = const()[name = tensor("encoder_out_proj_2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21185856)))]; - tensor encoder_conv_module_2_sequential_0_bias = const()[name = tensor("encoder_conv_module_2_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21448064)))]; - tensor encoder_conv_module_2_sequential_0_weight = const()[name = tensor("encoder_conv_module_2_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21449152)))]; - tensor encoder_conv_module_2_sequential_2_conv_bias = const()[name = tensor("encoder_conv_module_2_sequential_2_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21450240)))]; - tensor encoder_conv_module_2_sequential_2_conv_weight = const()[name = tensor("encoder_conv_module_2_sequential_2_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21452352)))]; - tensor encoder_conv_module_2_sequential_4_conv_weight = const()[name = tensor("encoder_conv_module_2_sequential_4_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21976704)))]; - tensor encoder_conv_module_2_sequential_5_running_var = const()[name = tensor("encoder_conv_module_2_sequential_5_running_var"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21993152)))]; - tensor encoder_conv_module_2_sequential_5_running_mean = const()[name = tensor("encoder_conv_module_2_sequential_5_running_mean"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21994240)))]; - tensor encoder_conv_module_2_sequential_5_bias = const()[name = tensor("encoder_conv_module_2_sequential_5_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21995328)))]; - tensor encoder_conv_module_2_sequential_5_weight = const()[name = tensor("encoder_conv_module_2_sequential_5_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21996416)))]; - tensor encoder_conv_module_2_sequential_7_conv_bias = const()[name = tensor("encoder_conv_module_2_sequential_7_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21997504)))]; - tensor encoder_conv_module_2_sequential_7_conv_weight = const()[name = tensor("encoder_conv_module_2_sequential_7_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21998592)))]; - tensor encoder_ffn2_2_module_sequential_0_bias = const()[name = tensor("encoder_ffn2_2_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(22260800)))]; - tensor encoder_ffn2_2_module_sequential_0_weight = const()[name = tensor("encoder_ffn2_2_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(22261888)))]; - tensor encoder_ffn2_2_module_sequential_1_linear_bias = const()[name = tensor("encoder_ffn2_2_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(22262976)))]; - tensor encoder_ffn2_2_module_sequential_1_linear_weight = const()[name = tensor("encoder_ffn2_2_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(22267136)))]; - tensor encoder_ffn2_2_module_sequential_4_linear_bias = const()[name = tensor("encoder_ffn2_2_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(23315776)))]; - tensor encoder_ffn2_2_module_sequential_4_linear_weight = const()[name = tensor("encoder_ffn2_2_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(23316864)))]; - tensor encoder_layer_norm_2_bias = const()[name = tensor("encoder_layer_norm_2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(24365504)))]; - tensor encoder_layer_norm_2_weight = const()[name = tensor("encoder_layer_norm_2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(24366592)))]; - tensor encoder_ffn1_3_module_sequential_0_bias = const()[name = tensor("encoder_ffn1_3_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(24367680)))]; - tensor encoder_ffn1_3_module_sequential_0_weight = const()[name = tensor("encoder_ffn1_3_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(24368768)))]; - tensor encoder_ffn1_3_module_sequential_1_linear_bias = const()[name = tensor("encoder_ffn1_3_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(24369856)))]; - tensor encoder_ffn1_3_module_sequential_1_linear_weight = const()[name = tensor("encoder_ffn1_3_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(24374016)))]; - tensor encoder_ffn1_3_module_sequential_4_linear_bias = const()[name = tensor("encoder_ffn1_3_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(25422656)))]; - tensor encoder_ffn1_3_module_sequential_4_linear_weight = const()[name = tensor("encoder_ffn1_3_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(25423744)))]; - tensor encoder_ret_lns_3_bias = const()[name = tensor("encoder_ret_lns_3_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(26472384)))]; - tensor encoder_ret_lns_3_weight = const()[name = tensor("encoder_ret_lns_3_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(26473472)))]; - tensor encoder_q_proj_3_bias = const()[name = tensor("encoder_q_proj_3_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(26474560)))]; - tensor encoder_q_proj_3_weight = const()[name = tensor("encoder_q_proj_3_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(26475648)))]; - tensor encoder_k_proj_3_bias = const()[name = tensor("encoder_k_proj_3_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(26737856)))]; - tensor encoder_k_proj_3_weight = const()[name = tensor("encoder_k_proj_3_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(26738944)))]; - tensor encoder_v_proj_3_bias = const()[name = tensor("encoder_v_proj_3_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27001152)))]; - tensor encoder_v_proj_3_weight = const()[name = tensor("encoder_v_proj_3_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27002240)))]; - tensor encoder_g_proj_3_bias = const()[name = tensor("encoder_g_proj_3_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27264448)))]; - tensor encoder_g_proj_3_weight = const()[name = tensor("encoder_g_proj_3_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27265536)))]; - tensor encoder_out_proj_3_bias = const()[name = tensor("encoder_out_proj_3_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27527744)))]; - tensor encoder_out_proj_3_weight = const()[name = tensor("encoder_out_proj_3_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27528832)))]; - tensor encoder_conv_module_3_sequential_0_bias = const()[name = tensor("encoder_conv_module_3_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27791040)))]; - tensor encoder_conv_module_3_sequential_0_weight = const()[name = tensor("encoder_conv_module_3_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27792128)))]; - tensor encoder_conv_module_3_sequential_2_conv_bias = const()[name = tensor("encoder_conv_module_3_sequential_2_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27793216)))]; - tensor encoder_conv_module_3_sequential_2_conv_weight = const()[name = tensor("encoder_conv_module_3_sequential_2_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27795328)))]; - tensor encoder_conv_module_3_sequential_4_conv_weight = const()[name = tensor("encoder_conv_module_3_sequential_4_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28319680)))]; - tensor encoder_conv_module_3_sequential_5_running_var = const()[name = tensor("encoder_conv_module_3_sequential_5_running_var"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28336128)))]; - tensor encoder_conv_module_3_sequential_5_running_mean = const()[name = tensor("encoder_conv_module_3_sequential_5_running_mean"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28337216)))]; - tensor encoder_conv_module_3_sequential_5_bias = const()[name = tensor("encoder_conv_module_3_sequential_5_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28338304)))]; - tensor encoder_conv_module_3_sequential_5_weight = const()[name = tensor("encoder_conv_module_3_sequential_5_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28339392)))]; - tensor encoder_conv_module_3_sequential_7_conv_bias = const()[name = tensor("encoder_conv_module_3_sequential_7_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28340480)))]; - tensor encoder_conv_module_3_sequential_7_conv_weight = const()[name = tensor("encoder_conv_module_3_sequential_7_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28341568)))]; - tensor encoder_ffn2_3_module_sequential_0_bias = const()[name = tensor("encoder_ffn2_3_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28603776)))]; - tensor encoder_ffn2_3_module_sequential_0_weight = const()[name = tensor("encoder_ffn2_3_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28604864)))]; - tensor encoder_ffn2_3_module_sequential_1_linear_bias = const()[name = tensor("encoder_ffn2_3_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28605952)))]; - tensor encoder_ffn2_3_module_sequential_1_linear_weight = const()[name = tensor("encoder_ffn2_3_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28610112)))]; - tensor encoder_ffn2_3_module_sequential_4_linear_bias = const()[name = tensor("encoder_ffn2_3_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(29658752)))]; - tensor encoder_ffn2_3_module_sequential_4_linear_weight = const()[name = tensor("encoder_ffn2_3_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(29659840)))]; - tensor encoder_layer_norm_3_bias = const()[name = tensor("encoder_layer_norm_3_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(30708480)))]; - tensor encoder_layer_norm_3_weight = const()[name = tensor("encoder_layer_norm_3_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(30709568)))]; - tensor decoder__causal_mask = const()[name = tensor("decoder__causal_mask"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(30710656)))]; - tensor decoder_convert_bias = const()[name = tensor("decoder_convert_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(30710848)))]; - tensor decoder_convert_weight = const()[name = tensor("decoder_convert_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(30711936)))]; - tensor decoder_q_proj_0_bias = const()[name = tensor("decoder_q_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(31236288)))]; - tensor decoder_q_proj_0_weight = const()[name = tensor("decoder_q_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(31237376)))]; - tensor decoder_k_proj_0_bias = const()[name = tensor("decoder_k_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(31499584)))]; - tensor decoder_k_proj_0_weight = const()[name = tensor("decoder_k_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(31500672)))]; - tensor decoder_v_proj_0_bias = const()[name = tensor("decoder_v_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(31762880)))]; - tensor decoder_v_proj_0_weight = const()[name = tensor("decoder_v_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(31763968)))]; - tensor decoder_g_proj_0_bias = const()[name = tensor("decoder_g_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32026176)))]; - tensor decoder_g_proj_0_weight = const()[name = tensor("decoder_g_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32027264)))]; - tensor decoder_out_proj_0_bias = const()[name = tensor("decoder_out_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32289472)))]; - tensor decoder_out_proj_0_weight = const()[name = tensor("decoder_out_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32290560)))]; - tensor decoder_norm11_0_bias = const()[name = tensor("decoder_norm11_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32552768)))]; - tensor decoder_norm11_0_weight = const()[name = tensor("decoder_norm11_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32553856)))]; - tensor decoder_self_attn2_0_out_proj_bias = const()[name = tensor("decoder_self_attn2_0_out_proj_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32554944)))]; - tensor decoder_self_attn2_0_out_proj_weight = const()[name = tensor("decoder_self_attn2_0_out_proj_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32556032)))]; - tensor decoder_self_attn2_0_in_proj_bias = const()[name = tensor("decoder_self_attn2_0_in_proj_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32818240)))]; - tensor decoder_self_attn2_0_in_proj_weight = const()[name = tensor("decoder_self_attn2_0_in_proj_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32821376)))]; - tensor decoder_norm21_0_bias = const()[name = tensor("decoder_norm21_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(33607872)))]; - tensor decoder_norm21_0_weight = const()[name = tensor("decoder_norm21_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(33608960)))]; - tensor decoder_linear1_0_bias = const()[name = tensor("decoder_linear1_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(33610048)))]; - tensor decoder_linear1_0_weight = const()[name = tensor("decoder_linear1_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(33618304)))]; - tensor decoder_linear2_0_bias = const()[name = tensor("decoder_linear2_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(35715520)))]; - tensor decoder_linear2_0_weight = const()[name = tensor("decoder_linear2_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(35716608)))]; - tensor decoder_norm22_0_bias = const()[name = tensor("decoder_norm22_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(37813824)))]; - tensor decoder_norm22_0_weight = const()[name = tensor("decoder_norm22_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(37814912)))]; - tensor decoder_q_proj_1_bias = const()[name = tensor("decoder_q_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(37816000)))]; - tensor decoder_q_proj_1_weight = const()[name = tensor("decoder_q_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(37817088)))]; - tensor decoder_k_proj_1_bias = const()[name = tensor("decoder_k_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38079296)))]; - tensor decoder_k_proj_1_weight = const()[name = tensor("decoder_k_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38080384)))]; - tensor decoder_v_proj_1_bias = const()[name = tensor("decoder_v_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38342592)))]; - tensor decoder_v_proj_1_weight = const()[name = tensor("decoder_v_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38343680)))]; - tensor decoder_g_proj_1_bias = const()[name = tensor("decoder_g_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38605888)))]; - tensor decoder_g_proj_1_weight = const()[name = tensor("decoder_g_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38606976)))]; - tensor decoder_out_proj_1_bias = const()[name = tensor("decoder_out_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38869184)))]; - tensor decoder_out_proj_1_weight = const()[name = tensor("decoder_out_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38870272)))]; - tensor decoder_norm11_1_bias = const()[name = tensor("decoder_norm11_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(39132480)))]; - tensor decoder_norm11_1_weight = const()[name = tensor("decoder_norm11_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(39133568)))]; - tensor decoder_self_attn2_1_out_proj_bias = const()[name = tensor("decoder_self_attn2_1_out_proj_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(39134656)))]; - tensor decoder_self_attn2_1_out_proj_weight = const()[name = tensor("decoder_self_attn2_1_out_proj_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(39135744)))]; - tensor decoder_self_attn2_1_in_proj_bias = const()[name = tensor("decoder_self_attn2_1_in_proj_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(39397952)))]; - tensor decoder_self_attn2_1_in_proj_weight = const()[name = tensor("decoder_self_attn2_1_in_proj_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(39401088)))]; - tensor decoder_norm21_1_bias = const()[name = tensor("decoder_norm21_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(40187584)))]; - tensor decoder_norm21_1_weight = const()[name = tensor("decoder_norm21_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(40188672)))]; - tensor decoder_linear1_1_bias = const()[name = tensor("decoder_linear1_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(40189760)))]; - tensor decoder_linear1_1_weight = const()[name = tensor("decoder_linear1_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(40198016)))]; - tensor decoder_linear2_1_bias = const()[name = tensor("decoder_linear2_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(42295232)))]; - tensor decoder_linear2_1_weight = const()[name = tensor("decoder_linear2_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(42296320)))]; - tensor decoder_norm22_1_bias = const()[name = tensor("decoder_norm22_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(44393536)))]; - tensor decoder_norm22_1_weight = const()[name = tensor("decoder_norm22_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(44394624)))]; - tensor var_11 = const()[name = tensor("op_11"), val = tensor(0x1.5798eep-27)]; - tensor var_18 = const()[name = tensor("op_18"), val = tensor(0x1.0c6f7ap-20)]; - tensor var_21 = const()[name = tensor("op_21"), val = tensor(2)]; - tensor var_24 = const()[name = tensor("op_24"), val = tensor(0)]; - tensor var_27 = const()[name = tensor("op_27"), val = tensor(1)]; - tensor var_29 = const()[name = tensor("op_29"), val = tensor(0x1.4f8b58p-17)]; - tensor var_30 = const()[name = tensor("op_30"), val = tensor(true)]; - tensor input_1 = linear(bias = encoder_input_projection_linear_bias, weight = encoder_input_projection_linear_weight, x = features)[name = tensor("linear_0")]; - tensor input_3_axes_0 = const()[name = tensor("input_3_axes_0"), val = tensor([-1])]; - tensor input_3 = layer_norm(axes = input_3_axes_0, beta = encoder_pre_layer_norm_bias, epsilon = var_29, gamma = encoder_pre_layer_norm_weight, x = input_1)[name = tensor("input_3")]; + func main(tensor cnn_window, tensor dec_kv, tensor dec_scale, tensor enc_conv_cache, tensor enc_kv, tensor enc_scale, tensor features, tensor valid_mask) { + tensor inner_encoder_cnn_bias = const()[name = tensor("inner_encoder_cnn_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(64)))]; + tensor inner_encoder_cnn_weight = const()[name = tensor("inner_encoder_cnn_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1152)))]; + tensor inner_encoder__causal_mask = const()[name = tensor("inner_encoder__causal_mask"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4981952)))]; + tensor inner_encoder__t_index = const()[name = tensor("inner_encoder__t_index"), val = tensor([0x1p+0, 0x1p+1, 0x1.8p+1, 0x1p+2, 0x1.4p+2])]; + tensor inner_encoder_input_projection_linear_bias = const()[name = tensor("inner_encoder_input_projection_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4982144)))]; + tensor inner_encoder_input_projection_linear_weight = const()[name = tensor("inner_encoder_input_projection_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4983232)))]; + tensor inner_encoder_pre_layer_norm_bias = const()[name = tensor("inner_encoder_pre_layer_norm_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(5336576)))]; + tensor inner_encoder_pre_layer_norm_weight = const()[name = tensor("inner_encoder_pre_layer_norm_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(5337664)))]; + tensor inner_encoder_ffn1_0_module_sequential_0_bias = const()[name = tensor("inner_encoder_ffn1_0_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(5338752)))]; + tensor inner_encoder_ffn1_0_module_sequential_0_weight = const()[name = tensor("inner_encoder_ffn1_0_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(5339840)))]; + tensor inner_encoder_ffn1_0_module_sequential_1_linear_bias = const()[name = tensor("inner_encoder_ffn1_0_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(5340928)))]; + tensor inner_encoder_ffn1_0_module_sequential_1_linear_weight = const()[name = tensor("inner_encoder_ffn1_0_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(5345088)))]; + tensor inner_encoder_ffn1_0_module_sequential_4_linear_bias = const()[name = tensor("inner_encoder_ffn1_0_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(6393728)))]; + tensor inner_encoder_ffn1_0_module_sequential_4_linear_weight = const()[name = tensor("inner_encoder_ffn1_0_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(6394816)))]; + tensor inner_encoder_ret_lns_0_bias = const()[name = tensor("inner_encoder_ret_lns_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7443456)))]; + tensor inner_encoder_ret_lns_0_weight = const()[name = tensor("inner_encoder_ret_lns_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7444544)))]; + tensor inner_encoder_q_proj_0_bias = const()[name = tensor("inner_encoder_q_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7445632)))]; + tensor inner_encoder_q_proj_0_weight = const()[name = tensor("inner_encoder_q_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7446720)))]; + tensor inner_encoder_k_proj_0_bias = const()[name = tensor("inner_encoder_k_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7708928)))]; + tensor inner_encoder_k_proj_0_weight = const()[name = tensor("inner_encoder_k_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7710016)))]; + tensor inner_encoder_v_proj_0_bias = const()[name = tensor("inner_encoder_v_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7972224)))]; + tensor inner_encoder_v_proj_0_weight = const()[name = tensor("inner_encoder_v_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7973312)))]; + tensor inner_encoder_g_proj_0_bias = const()[name = tensor("inner_encoder_g_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8235520)))]; + tensor inner_encoder_g_proj_0_weight = const()[name = tensor("inner_encoder_g_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8236608)))]; + tensor inner_encoder_out_proj_0_bias = const()[name = tensor("inner_encoder_out_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8498816)))]; + tensor inner_encoder_out_proj_0_weight = const()[name = tensor("inner_encoder_out_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8499904)))]; + tensor inner_encoder_conv_module_0_sequential_0_bias = const()[name = tensor("inner_encoder_conv_module_0_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8762112)))]; + tensor inner_encoder_conv_module_0_sequential_0_weight = const()[name = tensor("inner_encoder_conv_module_0_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8763200)))]; + tensor inner_encoder_conv_module_0_sequential_2_conv_bias = const()[name = tensor("inner_encoder_conv_module_0_sequential_2_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8764288)))]; + tensor inner_encoder_conv_module_0_sequential_2_conv_weight = const()[name = tensor("inner_encoder_conv_module_0_sequential_2_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8766400)))]; + tensor inner_encoder_conv_module_0_sequential_4_conv_weight = const()[name = tensor("inner_encoder_conv_module_0_sequential_4_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9290752)))]; + tensor inner_encoder_conv_module_0_sequential_5_running_var = const()[name = tensor("inner_encoder_conv_module_0_sequential_5_running_var"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9307200)))]; + tensor inner_encoder_conv_module_0_sequential_5_running_mean = const()[name = tensor("inner_encoder_conv_module_0_sequential_5_running_mean"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9308288)))]; + tensor inner_encoder_conv_module_0_sequential_5_bias = const()[name = tensor("inner_encoder_conv_module_0_sequential_5_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9309376)))]; + tensor inner_encoder_conv_module_0_sequential_5_weight = const()[name = tensor("inner_encoder_conv_module_0_sequential_5_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9310464)))]; + tensor inner_encoder_conv_module_0_sequential_7_conv_bias = const()[name = tensor("inner_encoder_conv_module_0_sequential_7_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9311552)))]; + tensor inner_encoder_conv_module_0_sequential_7_conv_weight = const()[name = tensor("inner_encoder_conv_module_0_sequential_7_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9312640)))]; + tensor inner_encoder_ffn2_0_module_sequential_0_bias = const()[name = tensor("inner_encoder_ffn2_0_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9574848)))]; + tensor inner_encoder_ffn2_0_module_sequential_0_weight = const()[name = tensor("inner_encoder_ffn2_0_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9575936)))]; + tensor inner_encoder_ffn2_0_module_sequential_1_linear_bias = const()[name = tensor("inner_encoder_ffn2_0_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9577024)))]; + tensor inner_encoder_ffn2_0_module_sequential_1_linear_weight = const()[name = tensor("inner_encoder_ffn2_0_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9581184)))]; + tensor inner_encoder_ffn2_0_module_sequential_4_linear_bias = const()[name = tensor("inner_encoder_ffn2_0_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(10629824)))]; + tensor inner_encoder_ffn2_0_module_sequential_4_linear_weight = const()[name = tensor("inner_encoder_ffn2_0_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(10630912)))]; + tensor inner_encoder_layer_norm_0_bias = const()[name = tensor("inner_encoder_layer_norm_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11679552)))]; + tensor inner_encoder_layer_norm_0_weight = const()[name = tensor("inner_encoder_layer_norm_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11680640)))]; + tensor inner_encoder_ffn1_1_module_sequential_0_bias = const()[name = tensor("inner_encoder_ffn1_1_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11681728)))]; + tensor inner_encoder_ffn1_1_module_sequential_0_weight = const()[name = tensor("inner_encoder_ffn1_1_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11682816)))]; + tensor inner_encoder_ffn1_1_module_sequential_1_linear_bias = const()[name = tensor("inner_encoder_ffn1_1_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11683904)))]; + tensor inner_encoder_ffn1_1_module_sequential_1_linear_weight = const()[name = tensor("inner_encoder_ffn1_1_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11688064)))]; + tensor inner_encoder_ffn1_1_module_sequential_4_linear_bias = const()[name = tensor("inner_encoder_ffn1_1_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(12736704)))]; + tensor inner_encoder_ffn1_1_module_sequential_4_linear_weight = const()[name = tensor("inner_encoder_ffn1_1_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(12737792)))]; + tensor inner_encoder_ret_lns_1_bias = const()[name = tensor("inner_encoder_ret_lns_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(13786432)))]; + tensor inner_encoder_ret_lns_1_weight = const()[name = tensor("inner_encoder_ret_lns_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(13787520)))]; + tensor inner_encoder_q_proj_1_bias = const()[name = tensor("inner_encoder_q_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(13788608)))]; + tensor inner_encoder_q_proj_1_weight = const()[name = tensor("inner_encoder_q_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(13789696)))]; + tensor inner_encoder_k_proj_1_bias = const()[name = tensor("inner_encoder_k_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14051904)))]; + tensor inner_encoder_k_proj_1_weight = const()[name = tensor("inner_encoder_k_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14052992)))]; + tensor inner_encoder_v_proj_1_bias = const()[name = tensor("inner_encoder_v_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14315200)))]; + tensor inner_encoder_v_proj_1_weight = const()[name = tensor("inner_encoder_v_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14316288)))]; + tensor inner_encoder_g_proj_1_bias = const()[name = tensor("inner_encoder_g_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14578496)))]; + tensor inner_encoder_g_proj_1_weight = const()[name = tensor("inner_encoder_g_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14579584)))]; + tensor inner_encoder_out_proj_1_bias = const()[name = tensor("inner_encoder_out_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14841792)))]; + tensor inner_encoder_out_proj_1_weight = const()[name = tensor("inner_encoder_out_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14842880)))]; + tensor inner_encoder_conv_module_1_sequential_0_bias = const()[name = tensor("inner_encoder_conv_module_1_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15105088)))]; + tensor inner_encoder_conv_module_1_sequential_0_weight = const()[name = tensor("inner_encoder_conv_module_1_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15106176)))]; + tensor inner_encoder_conv_module_1_sequential_2_conv_bias = const()[name = tensor("inner_encoder_conv_module_1_sequential_2_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15107264)))]; + tensor inner_encoder_conv_module_1_sequential_2_conv_weight = const()[name = tensor("inner_encoder_conv_module_1_sequential_2_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15109376)))]; + tensor inner_encoder_conv_module_1_sequential_4_conv_weight = const()[name = tensor("inner_encoder_conv_module_1_sequential_4_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15633728)))]; + tensor inner_encoder_conv_module_1_sequential_5_running_var = const()[name = tensor("inner_encoder_conv_module_1_sequential_5_running_var"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15650176)))]; + tensor inner_encoder_conv_module_1_sequential_5_running_mean = const()[name = tensor("inner_encoder_conv_module_1_sequential_5_running_mean"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15651264)))]; + tensor inner_encoder_conv_module_1_sequential_5_bias = const()[name = tensor("inner_encoder_conv_module_1_sequential_5_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15652352)))]; + tensor inner_encoder_conv_module_1_sequential_5_weight = const()[name = tensor("inner_encoder_conv_module_1_sequential_5_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15653440)))]; + tensor inner_encoder_conv_module_1_sequential_7_conv_bias = const()[name = tensor("inner_encoder_conv_module_1_sequential_7_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15654528)))]; + tensor inner_encoder_conv_module_1_sequential_7_conv_weight = const()[name = tensor("inner_encoder_conv_module_1_sequential_7_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15655616)))]; + tensor inner_encoder_ffn2_1_module_sequential_0_bias = const()[name = tensor("inner_encoder_ffn2_1_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15917824)))]; + tensor inner_encoder_ffn2_1_module_sequential_0_weight = const()[name = tensor("inner_encoder_ffn2_1_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15918912)))]; + tensor inner_encoder_ffn2_1_module_sequential_1_linear_bias = const()[name = tensor("inner_encoder_ffn2_1_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15920000)))]; + tensor inner_encoder_ffn2_1_module_sequential_1_linear_weight = const()[name = tensor("inner_encoder_ffn2_1_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15924160)))]; + tensor inner_encoder_ffn2_1_module_sequential_4_linear_bias = const()[name = tensor("inner_encoder_ffn2_1_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(16972800)))]; + tensor inner_encoder_ffn2_1_module_sequential_4_linear_weight = const()[name = tensor("inner_encoder_ffn2_1_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(16973888)))]; + tensor inner_encoder_layer_norm_1_bias = const()[name = tensor("inner_encoder_layer_norm_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18022528)))]; + tensor inner_encoder_layer_norm_1_weight = const()[name = tensor("inner_encoder_layer_norm_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18023616)))]; + tensor inner_encoder_ffn1_2_module_sequential_0_bias = const()[name = tensor("inner_encoder_ffn1_2_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18024704)))]; + tensor inner_encoder_ffn1_2_module_sequential_0_weight = const()[name = tensor("inner_encoder_ffn1_2_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18025792)))]; + tensor inner_encoder_ffn1_2_module_sequential_1_linear_bias = const()[name = tensor("inner_encoder_ffn1_2_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18026880)))]; + tensor inner_encoder_ffn1_2_module_sequential_1_linear_weight = const()[name = tensor("inner_encoder_ffn1_2_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18031040)))]; + tensor inner_encoder_ffn1_2_module_sequential_4_linear_bias = const()[name = tensor("inner_encoder_ffn1_2_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(19079680)))]; + tensor inner_encoder_ffn1_2_module_sequential_4_linear_weight = const()[name = tensor("inner_encoder_ffn1_2_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(19080768)))]; + tensor inner_encoder_ret_lns_2_bias = const()[name = tensor("inner_encoder_ret_lns_2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20129408)))]; + tensor inner_encoder_ret_lns_2_weight = const()[name = tensor("inner_encoder_ret_lns_2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20130496)))]; + tensor inner_encoder_q_proj_2_bias = const()[name = tensor("inner_encoder_q_proj_2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20131584)))]; + tensor inner_encoder_q_proj_2_weight = const()[name = tensor("inner_encoder_q_proj_2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20132672)))]; + tensor inner_encoder_k_proj_2_bias = const()[name = tensor("inner_encoder_k_proj_2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20394880)))]; + tensor inner_encoder_k_proj_2_weight = const()[name = tensor("inner_encoder_k_proj_2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20395968)))]; + tensor inner_encoder_v_proj_2_bias = const()[name = tensor("inner_encoder_v_proj_2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20658176)))]; + tensor inner_encoder_v_proj_2_weight = const()[name = tensor("inner_encoder_v_proj_2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20659264)))]; + tensor inner_encoder_g_proj_2_bias = const()[name = tensor("inner_encoder_g_proj_2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20921472)))]; + tensor inner_encoder_g_proj_2_weight = const()[name = tensor("inner_encoder_g_proj_2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20922560)))]; + tensor inner_encoder_out_proj_2_bias = const()[name = tensor("inner_encoder_out_proj_2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21184768)))]; + tensor inner_encoder_out_proj_2_weight = const()[name = tensor("inner_encoder_out_proj_2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21185856)))]; + tensor inner_encoder_conv_module_2_sequential_0_bias = const()[name = tensor("inner_encoder_conv_module_2_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21448064)))]; + tensor inner_encoder_conv_module_2_sequential_0_weight = const()[name = tensor("inner_encoder_conv_module_2_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21449152)))]; + tensor inner_encoder_conv_module_2_sequential_2_conv_bias = const()[name = tensor("inner_encoder_conv_module_2_sequential_2_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21450240)))]; + tensor inner_encoder_conv_module_2_sequential_2_conv_weight = const()[name = tensor("inner_encoder_conv_module_2_sequential_2_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21452352)))]; + tensor inner_encoder_conv_module_2_sequential_4_conv_weight = const()[name = tensor("inner_encoder_conv_module_2_sequential_4_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21976704)))]; + tensor inner_encoder_conv_module_2_sequential_5_running_var = const()[name = tensor("inner_encoder_conv_module_2_sequential_5_running_var"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21993152)))]; + tensor inner_encoder_conv_module_2_sequential_5_running_mean = const()[name = tensor("inner_encoder_conv_module_2_sequential_5_running_mean"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21994240)))]; + tensor inner_encoder_conv_module_2_sequential_5_bias = const()[name = tensor("inner_encoder_conv_module_2_sequential_5_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21995328)))]; + tensor inner_encoder_conv_module_2_sequential_5_weight = const()[name = tensor("inner_encoder_conv_module_2_sequential_5_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21996416)))]; + tensor inner_encoder_conv_module_2_sequential_7_conv_bias = const()[name = tensor("inner_encoder_conv_module_2_sequential_7_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21997504)))]; + tensor inner_encoder_conv_module_2_sequential_7_conv_weight = const()[name = tensor("inner_encoder_conv_module_2_sequential_7_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21998592)))]; + tensor inner_encoder_ffn2_2_module_sequential_0_bias = const()[name = tensor("inner_encoder_ffn2_2_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(22260800)))]; + tensor inner_encoder_ffn2_2_module_sequential_0_weight = const()[name = tensor("inner_encoder_ffn2_2_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(22261888)))]; + tensor inner_encoder_ffn2_2_module_sequential_1_linear_bias = const()[name = tensor("inner_encoder_ffn2_2_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(22262976)))]; + tensor inner_encoder_ffn2_2_module_sequential_1_linear_weight = const()[name = tensor("inner_encoder_ffn2_2_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(22267136)))]; + tensor inner_encoder_ffn2_2_module_sequential_4_linear_bias = const()[name = tensor("inner_encoder_ffn2_2_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(23315776)))]; + tensor inner_encoder_ffn2_2_module_sequential_4_linear_weight = const()[name = tensor("inner_encoder_ffn2_2_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(23316864)))]; + tensor inner_encoder_layer_norm_2_bias = const()[name = tensor("inner_encoder_layer_norm_2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(24365504)))]; + tensor inner_encoder_layer_norm_2_weight = const()[name = tensor("inner_encoder_layer_norm_2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(24366592)))]; + tensor inner_encoder_ffn1_3_module_sequential_0_bias = const()[name = tensor("inner_encoder_ffn1_3_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(24367680)))]; + tensor inner_encoder_ffn1_3_module_sequential_0_weight = const()[name = tensor("inner_encoder_ffn1_3_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(24368768)))]; + tensor inner_encoder_ffn1_3_module_sequential_1_linear_bias = const()[name = tensor("inner_encoder_ffn1_3_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(24369856)))]; + tensor inner_encoder_ffn1_3_module_sequential_1_linear_weight = const()[name = tensor("inner_encoder_ffn1_3_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(24374016)))]; + tensor inner_encoder_ffn1_3_module_sequential_4_linear_bias = const()[name = tensor("inner_encoder_ffn1_3_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(25422656)))]; + tensor inner_encoder_ffn1_3_module_sequential_4_linear_weight = const()[name = tensor("inner_encoder_ffn1_3_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(25423744)))]; + tensor inner_encoder_ret_lns_3_bias = const()[name = tensor("inner_encoder_ret_lns_3_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(26472384)))]; + tensor inner_encoder_ret_lns_3_weight = const()[name = tensor("inner_encoder_ret_lns_3_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(26473472)))]; + tensor inner_encoder_q_proj_3_bias = const()[name = tensor("inner_encoder_q_proj_3_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(26474560)))]; + tensor inner_encoder_q_proj_3_weight = const()[name = tensor("inner_encoder_q_proj_3_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(26475648)))]; + tensor inner_encoder_k_proj_3_bias = const()[name = tensor("inner_encoder_k_proj_3_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(26737856)))]; + tensor inner_encoder_k_proj_3_weight = const()[name = tensor("inner_encoder_k_proj_3_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(26738944)))]; + tensor inner_encoder_v_proj_3_bias = const()[name = tensor("inner_encoder_v_proj_3_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27001152)))]; + tensor inner_encoder_v_proj_3_weight = const()[name = tensor("inner_encoder_v_proj_3_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27002240)))]; + tensor inner_encoder_g_proj_3_bias = const()[name = tensor("inner_encoder_g_proj_3_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27264448)))]; + tensor inner_encoder_g_proj_3_weight = const()[name = tensor("inner_encoder_g_proj_3_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27265536)))]; + tensor inner_encoder_out_proj_3_bias = const()[name = tensor("inner_encoder_out_proj_3_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27527744)))]; + tensor inner_encoder_out_proj_3_weight = const()[name = tensor("inner_encoder_out_proj_3_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27528832)))]; + tensor inner_encoder_conv_module_3_sequential_0_bias = const()[name = tensor("inner_encoder_conv_module_3_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27791040)))]; + tensor inner_encoder_conv_module_3_sequential_0_weight = const()[name = tensor("inner_encoder_conv_module_3_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27792128)))]; + tensor inner_encoder_conv_module_3_sequential_2_conv_bias = const()[name = tensor("inner_encoder_conv_module_3_sequential_2_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27793216)))]; + tensor inner_encoder_conv_module_3_sequential_2_conv_weight = const()[name = tensor("inner_encoder_conv_module_3_sequential_2_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27795328)))]; + tensor inner_encoder_conv_module_3_sequential_4_conv_weight = const()[name = tensor("inner_encoder_conv_module_3_sequential_4_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28319680)))]; + tensor inner_encoder_conv_module_3_sequential_5_running_var = const()[name = tensor("inner_encoder_conv_module_3_sequential_5_running_var"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28336128)))]; + tensor inner_encoder_conv_module_3_sequential_5_running_mean = const()[name = tensor("inner_encoder_conv_module_3_sequential_5_running_mean"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28337216)))]; + tensor inner_encoder_conv_module_3_sequential_5_bias = const()[name = tensor("inner_encoder_conv_module_3_sequential_5_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28338304)))]; + tensor inner_encoder_conv_module_3_sequential_5_weight = const()[name = tensor("inner_encoder_conv_module_3_sequential_5_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28339392)))]; + tensor inner_encoder_conv_module_3_sequential_7_conv_bias = const()[name = tensor("inner_encoder_conv_module_3_sequential_7_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28340480)))]; + tensor inner_encoder_conv_module_3_sequential_7_conv_weight = const()[name = tensor("inner_encoder_conv_module_3_sequential_7_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28341568)))]; + tensor inner_encoder_ffn2_3_module_sequential_0_bias = const()[name = tensor("inner_encoder_ffn2_3_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28603776)))]; + tensor inner_encoder_ffn2_3_module_sequential_0_weight = const()[name = tensor("inner_encoder_ffn2_3_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28604864)))]; + tensor inner_encoder_ffn2_3_module_sequential_1_linear_bias = const()[name = tensor("inner_encoder_ffn2_3_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28605952)))]; + tensor inner_encoder_ffn2_3_module_sequential_1_linear_weight = const()[name = tensor("inner_encoder_ffn2_3_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28610112)))]; + tensor inner_encoder_ffn2_3_module_sequential_4_linear_bias = const()[name = tensor("inner_encoder_ffn2_3_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(29658752)))]; + tensor inner_encoder_ffn2_3_module_sequential_4_linear_weight = const()[name = tensor("inner_encoder_ffn2_3_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(29659840)))]; + tensor inner_encoder_layer_norm_3_bias = const()[name = tensor("inner_encoder_layer_norm_3_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(30708480)))]; + tensor inner_encoder_layer_norm_3_weight = const()[name = tensor("inner_encoder_layer_norm_3_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(30709568)))]; + tensor inner_decoder__causal_mask = const()[name = tensor("inner_decoder__causal_mask"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(30710656)))]; + tensor inner_decoder_convert_bias = const()[name = tensor("inner_decoder_convert_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(30710848)))]; + tensor inner_decoder_convert_weight = const()[name = tensor("inner_decoder_convert_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(30711936)))]; + tensor inner_decoder_q_proj_0_bias = const()[name = tensor("inner_decoder_q_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(31236288)))]; + tensor inner_decoder_q_proj_0_weight = const()[name = tensor("inner_decoder_q_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(31237376)))]; + tensor inner_decoder_k_proj_0_bias = const()[name = tensor("inner_decoder_k_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(31499584)))]; + tensor inner_decoder_k_proj_0_weight = const()[name = tensor("inner_decoder_k_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(31500672)))]; + tensor inner_decoder_v_proj_0_bias = const()[name = tensor("inner_decoder_v_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(31762880)))]; + tensor inner_decoder_v_proj_0_weight = const()[name = tensor("inner_decoder_v_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(31763968)))]; + tensor inner_decoder_g_proj_0_bias = const()[name = tensor("inner_decoder_g_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32026176)))]; + tensor inner_decoder_g_proj_0_weight = const()[name = tensor("inner_decoder_g_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32027264)))]; + tensor inner_decoder_out_proj_0_bias = const()[name = tensor("inner_decoder_out_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32289472)))]; + tensor inner_decoder_out_proj_0_weight = const()[name = tensor("inner_decoder_out_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32290560)))]; + tensor inner_decoder_norm11_0_bias = const()[name = tensor("inner_decoder_norm11_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32552768)))]; + tensor inner_decoder_norm11_0_weight = const()[name = tensor("inner_decoder_norm11_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32553856)))]; + tensor inner_decoder_self_attn2_0_out_proj_bias = const()[name = tensor("inner_decoder_self_attn2_0_out_proj_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32554944)))]; + tensor inner_decoder_self_attn2_0_out_proj_weight = const()[name = tensor("inner_decoder_self_attn2_0_out_proj_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32556032)))]; + tensor inner_decoder_self_attn2_0_in_proj_bias = const()[name = tensor("inner_decoder_self_attn2_0_in_proj_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32818240)))]; + tensor inner_decoder_self_attn2_0_in_proj_weight = const()[name = tensor("inner_decoder_self_attn2_0_in_proj_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32821376)))]; + tensor inner_decoder_norm21_0_bias = const()[name = tensor("inner_decoder_norm21_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(33607872)))]; + tensor inner_decoder_norm21_0_weight = const()[name = tensor("inner_decoder_norm21_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(33608960)))]; + tensor inner_decoder_linear1_0_bias = const()[name = tensor("inner_decoder_linear1_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(33610048)))]; + tensor inner_decoder_linear1_0_weight = const()[name = tensor("inner_decoder_linear1_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(33618304)))]; + tensor inner_decoder_linear2_0_bias = const()[name = tensor("inner_decoder_linear2_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(35715520)))]; + tensor inner_decoder_linear2_0_weight = const()[name = tensor("inner_decoder_linear2_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(35716608)))]; + tensor inner_decoder_norm22_0_bias = const()[name = tensor("inner_decoder_norm22_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(37813824)))]; + tensor inner_decoder_norm22_0_weight = const()[name = tensor("inner_decoder_norm22_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(37814912)))]; + tensor inner_decoder_q_proj_1_bias = const()[name = tensor("inner_decoder_q_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(37816000)))]; + tensor inner_decoder_q_proj_1_weight = const()[name = tensor("inner_decoder_q_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(37817088)))]; + tensor inner_decoder_k_proj_1_bias = const()[name = tensor("inner_decoder_k_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38079296)))]; + tensor inner_decoder_k_proj_1_weight = const()[name = tensor("inner_decoder_k_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38080384)))]; + tensor inner_decoder_v_proj_1_bias = const()[name = tensor("inner_decoder_v_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38342592)))]; + tensor inner_decoder_v_proj_1_weight = const()[name = tensor("inner_decoder_v_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38343680)))]; + tensor inner_decoder_g_proj_1_bias = const()[name = tensor("inner_decoder_g_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38605888)))]; + tensor inner_decoder_g_proj_1_weight = const()[name = tensor("inner_decoder_g_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38606976)))]; + tensor inner_decoder_out_proj_1_bias = const()[name = tensor("inner_decoder_out_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38869184)))]; + tensor inner_decoder_out_proj_1_weight = const()[name = tensor("inner_decoder_out_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38870272)))]; + tensor inner_decoder_norm11_1_bias = const()[name = tensor("inner_decoder_norm11_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(39132480)))]; + tensor inner_decoder_norm11_1_weight = const()[name = tensor("inner_decoder_norm11_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(39133568)))]; + tensor inner_decoder_self_attn2_1_out_proj_bias = const()[name = tensor("inner_decoder_self_attn2_1_out_proj_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(39134656)))]; + tensor inner_decoder_self_attn2_1_out_proj_weight = const()[name = tensor("inner_decoder_self_attn2_1_out_proj_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(39135744)))]; + tensor inner_decoder_self_attn2_1_in_proj_bias = const()[name = tensor("inner_decoder_self_attn2_1_in_proj_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(39397952)))]; + tensor inner_decoder_self_attn2_1_in_proj_weight = const()[name = tensor("inner_decoder_self_attn2_1_in_proj_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(39401088)))]; + tensor inner_decoder_norm21_1_bias = const()[name = tensor("inner_decoder_norm21_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(40187584)))]; + tensor inner_decoder_norm21_1_weight = const()[name = tensor("inner_decoder_norm21_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(40188672)))]; + tensor inner_decoder_linear1_1_bias = const()[name = tensor("inner_decoder_linear1_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(40189760)))]; + tensor inner_decoder_linear1_1_weight = const()[name = tensor("inner_decoder_linear1_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(40198016)))]; + tensor inner_decoder_linear2_1_bias = const()[name = tensor("inner_decoder_linear2_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(42295232)))]; + tensor inner_decoder_linear2_1_weight = const()[name = tensor("inner_decoder_linear2_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(42296320)))]; + tensor inner_decoder_norm22_1_bias = const()[name = tensor("inner_decoder_norm22_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(44393536)))]; + tensor inner_decoder_norm22_1_weight = const()[name = tensor("inner_decoder_norm22_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(44394624)))]; + tensor var_19_begin_0 = const()[name = tensor("op_19_begin_0"), val = tensor([0, 0, 0])]; + tensor var_19_end_0 = const()[name = tensor("op_19_end_0"), val = tensor([1, 15, 23])]; + tensor var_19_end_mask_0 = const()[name = tensor("op_19_end_mask_0"), val = tensor([true, false, true])]; + tensor var_19 = slice_by_index(begin = var_19_begin_0, end = var_19_end_0, end_mask = var_19_end_mask_0, x = features)[name = tensor("op_19")]; + tensor var_29_begin_0 = const()[name = tensor("op_29_begin_0"), val = tensor([0, 10, 0])]; + tensor var_29_end_0 = const()[name = tensor("op_29_end_0"), val = tensor([1, 25, 23])]; + tensor var_29_end_mask_0 = const()[name = tensor("op_29_end_mask_0"), val = tensor([true, false, true])]; + tensor var_29 = slice_by_index(begin = var_29_begin_0, end = var_29_end_0, end_mask = var_29_end_mask_0, x = features)[name = tensor("op_29")]; + tensor var_39_begin_0 = const()[name = tensor("op_39_begin_0"), val = tensor([0, 20, 0])]; + tensor var_39_end_0 = const()[name = tensor("op_39_end_0"), val = tensor([1, 35, 23])]; + tensor var_39_end_mask_0 = const()[name = tensor("op_39_end_mask_0"), val = tensor([true, false, true])]; + tensor var_39 = slice_by_index(begin = var_39_begin_0, end = var_39_end_0, end_mask = var_39_end_mask_0, x = features)[name = tensor("op_39")]; + tensor var_49_begin_0 = const()[name = tensor("op_49_begin_0"), val = tensor([0, 30, 0])]; + tensor var_49_end_0 = const()[name = tensor("op_49_end_0"), val = tensor([1, 45, 23])]; + tensor var_49_end_mask_0 = const()[name = tensor("op_49_end_mask_0"), val = tensor([true, false, true])]; + tensor var_49 = slice_by_index(begin = var_49_begin_0, end = var_49_end_0, end_mask = var_49_end_mask_0, x = features)[name = tensor("op_49")]; + tensor var_59_begin_0 = const()[name = tensor("op_59_begin_0"), val = tensor([0, 40, 0])]; + tensor var_59_end_0 = const()[name = tensor("op_59_end_0"), val = tensor([1, 1, 23])]; + tensor var_59_end_mask_0 = const()[name = tensor("op_59_end_mask_0"), val = tensor([true, true, true])]; + tensor var_59 = slice_by_index(begin = var_59_begin_0, end = var_59_end_0, end_mask = var_59_end_mask_0, x = features)[name = tensor("op_59")]; + tensor stacked_axis_0 = const()[name = tensor("stacked_axis_0"), val = tensor(1)]; + tensor stacked = stack(axis = stacked_axis_0, values = (var_19, var_29, var_39, var_49, var_59))[name = tensor("stacked")]; + tensor var_66 = const()[name = tensor("op_66"), val = tensor([1, 5, 345])]; + tensor input_1 = reshape(shape = var_66, x = stacked)[name = tensor("input_1")]; + tensor var_68 = const()[name = tensor("op_68"), val = tensor(0x1p+0)]; + tensor var_73 = const()[name = tensor("op_73"), val = tensor(true)]; + tensor var_74 = const()[name = tensor("op_74"), val = tensor(0x1.4f8b58p-17)]; + tensor var_77 = const()[name = tensor("op_77"), val = tensor(0)]; + tensor var_79 = const()[name = tensor("op_79"), val = tensor(2)]; + tensor var_80 = const()[name = tensor("op_80"), val = tensor(-1)]; + tensor var_82 = const()[name = tensor("op_82"), val = tensor(0x1.0c6f7ap-20)]; + tensor var_88 = const()[name = tensor("op_88"), val = tensor(0x1.5798eep-27)]; + tensor var_92 = const()[name = tensor("op_92"), val = tensor(1)]; + tensor input_3 = linear(bias = inner_encoder_input_projection_linear_bias, weight = inner_encoder_input_projection_linear_weight, x = input_1)[name = tensor("linear_0")]; tensor input_5_axes_0 = const()[name = tensor("input_5_axes_0"), val = tensor([-1])]; - tensor input_5 = layer_norm(axes = input_5_axes_0, beta = encoder_ffn1_0_module_sequential_0_bias, epsilon = var_29, gamma = encoder_ffn1_0_module_sequential_0_weight, x = input_3)[name = tensor("input_5")]; - tensor inputs_1 = linear(bias = encoder_ffn1_0_module_sequential_1_linear_bias, weight = encoder_ffn1_0_module_sequential_1_linear_weight, x = input_5)[name = tensor("linear_1")]; - tensor input_7 = silu(x = inputs_1)[name = tensor("input_7")]; - tensor input_11 = linear(bias = encoder_ffn1_0_module_sequential_4_linear_bias, weight = encoder_ffn1_0_module_sequential_4_linear_weight, x = input_7)[name = tensor("linear_2")]; - tensor var_148 = const()[name = tensor("op_148"), val = tensor(0x1p-1)]; - tensor var_149 = mul(x = input_11, y = var_148)[name = tensor("op_149")]; - tensor input_13 = add(x = var_149, y = input_3)[name = tensor("input_13")]; + tensor input_5 = layer_norm(axes = input_5_axes_0, beta = inner_encoder_pre_layer_norm_bias, epsilon = var_74, gamma = inner_encoder_pre_layer_norm_weight, x = input_3)[name = tensor("input_5")]; + tensor input_7_axes_0 = const()[name = tensor("input_7_axes_0"), val = tensor([-1])]; + tensor input_7 = layer_norm(axes = input_7_axes_0, beta = inner_encoder_ffn1_0_module_sequential_0_bias, epsilon = var_74, gamma = inner_encoder_ffn1_0_module_sequential_0_weight, x = input_5)[name = tensor("input_7")]; + tensor inputs_1 = linear(bias = inner_encoder_ffn1_0_module_sequential_1_linear_bias, weight = inner_encoder_ffn1_0_module_sequential_1_linear_weight, x = input_7)[name = tensor("linear_1")]; + tensor input_9 = silu(x = inputs_1)[name = tensor("input_9")]; + tensor input_13 = linear(bias = inner_encoder_ffn1_0_module_sequential_4_linear_bias, weight = inner_encoder_ffn1_0_module_sequential_4_linear_weight, x = input_9)[name = tensor("linear_2")]; + tensor var_213 = const()[name = tensor("op_213"), val = tensor(0x1p-1)]; + tensor var_214 = mul(x = input_13, y = var_213)[name = tensor("op_214")]; + tensor input_15 = add(x = var_214, y = input_5)[name = tensor("input_15")]; tensor x_1_axes_0 = const()[name = tensor("x_1_axes_0"), val = tensor([-1])]; - tensor x_1 = layer_norm(axes = x_1_axes_0, beta = encoder_ret_lns_0_bias, epsilon = var_29, gamma = encoder_ret_lns_0_weight, x = input_13)[name = tensor("x_1")]; + tensor x_1 = layer_norm(axes = x_1_axes_0, beta = inner_encoder_ret_lns_0_bias, epsilon = var_74, gamma = inner_encoder_ret_lns_0_weight, x = input_15)[name = tensor("x_1")]; tensor prev_kv_1_begin_0 = const()[name = tensor("prev_kv_1_begin_0"), val = tensor([0, 0, 0, 0, 0])]; tensor prev_kv_1_end_0 = const()[name = tensor("prev_kv_1_end_0"), val = tensor([1, 1, 4, 64, 64])]; tensor prev_kv_1_end_mask_0 = const()[name = tensor("prev_kv_1_end_mask_0"), val = tensor([false, true, true, true, true])]; @@ -239,183 +265,183 @@ program(1.0) tensor prev_scale_1_end_mask_0 = const()[name = tensor("prev_scale_1_end_mask_0"), val = tensor([false, true])]; tensor prev_scale_1_squeeze_mask_0 = const()[name = tensor("prev_scale_1_squeeze_mask_0"), val = tensor([true, false])]; tensor prev_scale_1 = slice_by_index(begin = prev_scale_1_begin_0, end = prev_scale_1_end_0, end_mask = prev_scale_1_end_mask_0, squeeze_mask = prev_scale_1_squeeze_mask_0, x = enc_scale)[name = tensor("prev_scale_1")]; - tensor var_163 = linear(bias = encoder_q_proj_0_bias, weight = encoder_q_proj_0_weight, x = x_1)[name = tensor("linear_3")]; - tensor var_164 = const()[name = tensor("op_164"), val = tensor([1, 5, 4, 64])]; - tensor var_165 = reshape(shape = var_164, x = var_163)[name = tensor("op_165")]; + tensor var_228 = linear(bias = inner_encoder_q_proj_0_bias, weight = inner_encoder_q_proj_0_weight, x = x_1)[name = tensor("linear_3")]; + tensor var_229 = const()[name = tensor("op_229"), val = tensor([1, 5, 4, 64])]; + tensor var_230 = reshape(shape = var_229, x = var_228)[name = tensor("op_230")]; tensor q_1_perm_0 = const()[name = tensor("q_1_perm_0"), val = tensor([0, 2, 1, 3])]; - tensor var_169 = linear(bias = encoder_k_proj_0_bias, weight = encoder_k_proj_0_weight, x = x_1)[name = tensor("linear_4")]; - tensor var_170 = const()[name = tensor("op_170"), val = tensor(0x1p-3)]; - tensor var_171 = mul(x = var_169, y = var_170)[name = tensor("op_171")]; - tensor var_172 = const()[name = tensor("op_172"), val = tensor([1, 5, 4, 64])]; - tensor var_173 = reshape(shape = var_172, x = var_171)[name = tensor("op_173")]; + tensor var_234 = linear(bias = inner_encoder_k_proj_0_bias, weight = inner_encoder_k_proj_0_weight, x = x_1)[name = tensor("linear_4")]; + tensor var_235 = const()[name = tensor("op_235"), val = tensor(0x1p-3)]; + tensor var_236 = mul(x = var_234, y = var_235)[name = tensor("op_236")]; + tensor var_237 = const()[name = tensor("op_237"), val = tensor([1, 5, 4, 64])]; + tensor var_238 = reshape(shape = var_237, x = var_236)[name = tensor("op_238")]; tensor k_1_perm_0 = const()[name = tensor("k_1_perm_0"), val = tensor([0, 2, 1, 3])]; - tensor var_177 = linear(bias = encoder_v_proj_0_bias, weight = encoder_v_proj_0_weight, x = x_1)[name = tensor("linear_5")]; - tensor var_178 = const()[name = tensor("op_178"), val = tensor([1, 5, 4, 64])]; - tensor var_179 = reshape(shape = var_178, x = var_177)[name = tensor("op_179")]; + tensor var_242 = linear(bias = inner_encoder_v_proj_0_bias, weight = inner_encoder_v_proj_0_weight, x = x_1)[name = tensor("linear_5")]; + tensor var_243 = const()[name = tensor("op_243"), val = tensor([1, 5, 4, 64])]; + tensor var_244 = reshape(shape = var_243, x = var_242)[name = tensor("op_244")]; tensor v_1_perm_0 = const()[name = tensor("v_1_perm_0"), val = tensor([0, 2, 1, 3])]; - tensor input_17 = linear(bias = encoder_g_proj_0_bias, weight = encoder_g_proj_0_weight, x = x_1)[name = tensor("linear_6")]; + tensor input_19 = linear(bias = inner_encoder_g_proj_0_bias, weight = inner_encoder_g_proj_0_weight, x = x_1)[name = tensor("linear_6")]; tensor sqrt_s0_1 = sqrt(x = prev_scale_1)[name = tensor("sqrt_s0_1")]; - tensor s_t_1 = add(x = prev_scale_1, y = encoder__t_index)[name = tensor("s_t_1")]; + tensor s_t_1 = add(x = prev_scale_1, y = inner_encoder__t_index)[name = tensor("s_t_1")]; tensor sqrt_s_t_1 = sqrt(x = s_t_1)[name = tensor("sqrt_s_t_1")]; tensor qk_1_transpose_x_1 = const()[name = tensor("qk_1_transpose_x_1"), val = tensor(false)]; tensor qk_1_transpose_y_1 = const()[name = tensor("qk_1_transpose_y_1"), val = tensor(true)]; - tensor k_1 = transpose(perm = k_1_perm_0, x = var_173)[name = tensor("transpose_57")]; - tensor q_1 = transpose(perm = q_1_perm_0, x = var_165)[name = tensor("transpose_58")]; + tensor k_1 = transpose(perm = k_1_perm_0, x = var_238)[name = tensor("transpose_57")]; + tensor q_1 = transpose(perm = q_1_perm_0, x = var_230)[name = tensor("transpose_58")]; tensor qk_1 = matmul(transpose_x = qk_1_transpose_x_1, transpose_y = qk_1_transpose_y_1, x = q_1, y = k_1)[name = tensor("qk_1")]; - tensor var_189 = const()[name = tensor("op_189"), val = tensor([5, 1])]; - tensor var_190 = reshape(shape = var_189, x = sqrt_s_t_1)[name = tensor("op_190")]; - tensor M_1 = real_div(x = encoder__causal_mask, y = var_190)[name = tensor("M_1")]; - tensor var_192 = mul(x = qk_1, y = M_1)[name = tensor("op_192")]; + tensor var_254 = const()[name = tensor("op_254"), val = tensor([5, 1])]; + tensor var_255 = reshape(shape = var_254, x = sqrt_s_t_1)[name = tensor("op_255")]; + tensor M_1 = real_div(x = inner_encoder__causal_mask, y = var_255)[name = tensor("M_1")]; + tensor var_257 = mul(x = qk_1, y = M_1)[name = tensor("op_257")]; tensor inner_1_transpose_x_0 = const()[name = tensor("inner_1_transpose_x_0"), val = tensor(false)]; tensor inner_1_transpose_y_0 = const()[name = tensor("inner_1_transpose_y_0"), val = tensor(false)]; - tensor v_1 = transpose(perm = v_1_perm_0, x = var_179)[name = tensor("transpose_56")]; - tensor inner_1 = matmul(transpose_x = inner_1_transpose_x_0, transpose_y = inner_1_transpose_y_0, x = var_192, y = v_1)[name = tensor("inner_1")]; - tensor var_194_transpose_x_0 = const()[name = tensor("op_194_transpose_x_0"), val = tensor(false)]; - tensor var_194_transpose_y_0 = const()[name = tensor("op_194_transpose_y_0"), val = tensor(false)]; - tensor var_194 = matmul(transpose_x = var_194_transpose_x_0, transpose_y = var_194_transpose_y_0, x = q_1, y = prev_kv_1)[name = tensor("op_194")]; - tensor var_195 = real_div(x = sqrt_s0_1, y = sqrt_s_t_1)[name = tensor("op_195")]; - tensor var_196 = const()[name = tensor("op_196"), val = tensor([1, 1, 5, 1])]; - tensor var_197 = reshape(shape = var_196, x = var_195)[name = tensor("op_197")]; - tensor cross_1 = mul(x = var_194, y = var_197)[name = tensor("cross_1")]; + tensor v_1 = transpose(perm = v_1_perm_0, x = var_244)[name = tensor("transpose_56")]; + tensor inner_1 = matmul(transpose_x = inner_1_transpose_x_0, transpose_y = inner_1_transpose_y_0, x = var_257, y = v_1)[name = tensor("inner_1")]; + tensor var_259_transpose_x_0 = const()[name = tensor("op_259_transpose_x_0"), val = tensor(false)]; + tensor var_259_transpose_y_0 = const()[name = tensor("op_259_transpose_y_0"), val = tensor(false)]; + tensor var_259 = matmul(transpose_x = var_259_transpose_x_0, transpose_y = var_259_transpose_y_0, x = q_1, y = prev_kv_1)[name = tensor("op_259")]; + tensor var_260 = real_div(x = sqrt_s0_1, y = sqrt_s_t_1)[name = tensor("op_260")]; + tensor var_261 = const()[name = tensor("op_261"), val = tensor([1, 1, 5, 1])]; + tensor var_262 = reshape(shape = var_261, x = var_260)[name = tensor("op_262")]; + tensor cross_1 = mul(x = var_259, y = var_262)[name = tensor("cross_1")]; tensor out_1 = add(x = inner_1, y = cross_1)[name = tensor("out_1")]; - tensor var_200 = mul(x = prev_kv_1, y = sqrt_s0_1)[name = tensor("op_200")]; - tensor var_202_transpose_x_1 = const()[name = tensor("op_202_transpose_x_1"), val = tensor(true)]; - tensor var_202_transpose_y_1 = const()[name = tensor("op_202_transpose_y_1"), val = tensor(false)]; - tensor var_202 = matmul(transpose_x = var_202_transpose_x_1, transpose_y = var_202_transpose_y_1, x = k_1, y = v_1)[name = tensor("op_202")]; - tensor new_kv_unnorm_1 = add(x = var_200, y = var_202)[name = tensor("new_kv_unnorm_1")]; - tensor var_204 = const()[name = tensor("op_204"), val = tensor(0x1.4p+2)]; - tensor new_scale_1 = add(x = prev_scale_1, y = var_204)[name = tensor("new_scale_1")]; - tensor var_206 = sqrt(x = new_scale_1)[name = tensor("op_206")]; - tensor var_207 = real_div(x = new_kv_unnorm_1, y = var_206)[name = tensor("op_207")]; - tensor var_208_perm_0 = const()[name = tensor("op_208_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_265 = mul(x = prev_kv_1, y = sqrt_s0_1)[name = tensor("op_265")]; + tensor var_267_transpose_x_1 = const()[name = tensor("op_267_transpose_x_1"), val = tensor(true)]; + tensor var_267_transpose_y_1 = const()[name = tensor("op_267_transpose_y_1"), val = tensor(false)]; + tensor var_267 = matmul(transpose_x = var_267_transpose_x_1, transpose_y = var_267_transpose_y_1, x = k_1, y = v_1)[name = tensor("op_267")]; + tensor new_kv_unnorm_1 = add(x = var_265, y = var_267)[name = tensor("new_kv_unnorm_1")]; + tensor var_269 = const()[name = tensor("op_269"), val = tensor(0x1.4p+2)]; + tensor new_scale_1 = add(x = prev_scale_1, y = var_269)[name = tensor("new_scale_1")]; + tensor var_271 = sqrt(x = new_scale_1)[name = tensor("op_271")]; + tensor var_272 = real_div(x = new_kv_unnorm_1, y = var_271)[name = tensor("op_272")]; + tensor var_273_perm_0 = const()[name = tensor("op_273_perm_0"), val = tensor([0, 2, 1, 3])]; tensor out_3_axes_0 = const()[name = tensor("out_3_axes_0"), val = tensor([-1])]; - tensor var_208 = transpose(perm = var_208_perm_0, x = out_1)[name = tensor("transpose_55")]; - tensor out_3 = layer_norm(axes = out_3_axes_0, epsilon = var_18, x = var_208)[name = tensor("out_3")]; - tensor var_212 = const()[name = tensor("op_212"), val = tensor([1, 5, 256])]; - tensor out_5 = reshape(shape = var_212, x = out_3)[name = tensor("out_5")]; - tensor var_214 = silu(x = input_17)[name = tensor("op_214")]; - tensor input_19 = mul(x = var_214, y = out_5)[name = tensor("input_19")]; - tensor ret_out_1 = linear(bias = encoder_out_proj_0_bias, weight = encoder_out_proj_0_weight, x = input_19)[name = tensor("linear_7")]; - tensor x_3 = add(x = input_13, y = ret_out_1)[name = tensor("x_3")]; + tensor var_273 = transpose(perm = var_273_perm_0, x = out_1)[name = tensor("transpose_55")]; + tensor out_3 = layer_norm(axes = out_3_axes_0, epsilon = var_82, x = var_273)[name = tensor("out_3")]; + tensor var_277 = const()[name = tensor("op_277"), val = tensor([1, 5, 256])]; + tensor out_5 = reshape(shape = var_277, x = out_3)[name = tensor("out_5")]; + tensor var_279 = silu(x = input_19)[name = tensor("op_279")]; + tensor input_21 = mul(x = var_279, y = out_5)[name = tensor("input_21")]; + tensor ret_out_1 = linear(bias = inner_encoder_out_proj_0_bias, weight = inner_encoder_out_proj_0_weight, x = input_21)[name = tensor("linear_7")]; + tensor x_3 = add(x = input_15, y = ret_out_1)[name = tensor("x_3")]; tensor window_1_begin_0 = const()[name = tensor("window_1_begin_0"), val = tensor([0, 0, 0, 0])]; tensor window_1_end_0 = const()[name = tensor("window_1_end_0"), val = tensor([1, 1, 16, 256])]; tensor window_1_end_mask_0 = const()[name = tensor("window_1_end_mask_0"), val = tensor([false, true, true, true])]; tensor window_1_squeeze_mask_0 = const()[name = tensor("window_1_squeeze_mask_0"), val = tensor([true, false, false, false])]; tensor window_1 = slice_by_index(begin = window_1_begin_0, end = window_1_end_0, end_mask = window_1_end_mask_0, squeeze_mask = window_1_squeeze_mask_0, x = enc_conv_cache)[name = tensor("window_1")]; - tensor var_222_begin_0 = const()[name = tensor("op_222_begin_0"), val = tensor([0, 0, 0])]; - tensor var_222_end_0 = const()[name = tensor("op_222_end_0"), val = tensor([1, 1, 256])]; - tensor var_222_end_mask_0 = const()[name = tensor("op_222_end_mask_0"), val = tensor([true, false, true])]; - tensor var_222 = slice_by_index(begin = var_222_begin_0, end = var_222_end_0, end_mask = var_222_end_mask_0, x = x_3)[name = tensor("op_222")]; - tensor var_225_begin_0 = const()[name = tensor("op_225_begin_0"), val = tensor([0, 1, 0])]; - tensor var_225_end_0 = const()[name = tensor("op_225_end_0"), val = tensor([1, 16, 256])]; - tensor var_225_end_mask_0 = const()[name = tensor("op_225_end_mask_0"), val = tensor([true, true, true])]; - tensor var_225 = slice_by_index(begin = var_225_begin_0, end = var_225_end_0, end_mask = var_225_end_mask_0, x = window_1)[name = tensor("op_225")]; + tensor var_287_begin_0 = const()[name = tensor("op_287_begin_0"), val = tensor([0, 0, 0])]; + tensor var_287_end_0 = const()[name = tensor("op_287_end_0"), val = tensor([1, 1, 256])]; + tensor var_287_end_mask_0 = const()[name = tensor("op_287_end_mask_0"), val = tensor([true, false, true])]; + tensor var_287 = slice_by_index(begin = var_287_begin_0, end = var_287_end_0, end_mask = var_287_end_mask_0, x = x_3)[name = tensor("op_287")]; + tensor var_290_begin_0 = const()[name = tensor("op_290_begin_0"), val = tensor([0, 1, 0])]; + tensor var_290_end_0 = const()[name = tensor("op_290_end_0"), val = tensor([1, 16, 256])]; + tensor var_290_end_mask_0 = const()[name = tensor("op_290_end_mask_0"), val = tensor([true, true, true])]; + tensor var_290 = slice_by_index(begin = var_290_begin_0, end = var_290_end_0, end_mask = var_290_end_mask_0, x = window_1)[name = tensor("op_290")]; tensor window_3_interleave_0 = const()[name = tensor("window_3_interleave_0"), val = tensor(false)]; - tensor window_3 = concat(axis = var_27, interleave = window_3_interleave_0, values = (var_225, var_222))[name = tensor("window_3")]; - tensor var_230_begin_0 = const()[name = tensor("op_230_begin_0"), val = tensor([0, 1, 0])]; - tensor var_230_end_0 = const()[name = tensor("op_230_end_0"), val = tensor([1, 2, 256])]; - tensor var_230_end_mask_0 = const()[name = tensor("op_230_end_mask_0"), val = tensor([true, false, true])]; - tensor var_230 = slice_by_index(begin = var_230_begin_0, end = var_230_end_0, end_mask = var_230_end_mask_0, x = x_3)[name = tensor("op_230")]; - tensor var_233_begin_0 = const()[name = tensor("op_233_begin_0"), val = tensor([0, 1, 0])]; - tensor var_233_end_0 = const()[name = tensor("op_233_end_0"), val = tensor([1, 16, 256])]; - tensor var_233_end_mask_0 = const()[name = tensor("op_233_end_mask_0"), val = tensor([true, true, true])]; - tensor var_233 = slice_by_index(begin = var_233_begin_0, end = var_233_end_0, end_mask = var_233_end_mask_0, x = window_3)[name = tensor("op_233")]; + tensor window_3 = concat(axis = var_92, interleave = window_3_interleave_0, values = (var_290, var_287))[name = tensor("window_3")]; + tensor var_295_begin_0 = const()[name = tensor("op_295_begin_0"), val = tensor([0, 1, 0])]; + tensor var_295_end_0 = const()[name = tensor("op_295_end_0"), val = tensor([1, 2, 256])]; + tensor var_295_end_mask_0 = const()[name = tensor("op_295_end_mask_0"), val = tensor([true, false, true])]; + tensor var_295 = slice_by_index(begin = var_295_begin_0, end = var_295_end_0, end_mask = var_295_end_mask_0, x = x_3)[name = tensor("op_295")]; + tensor var_298_begin_0 = const()[name = tensor("op_298_begin_0"), val = tensor([0, 1, 0])]; + tensor var_298_end_0 = const()[name = tensor("op_298_end_0"), val = tensor([1, 16, 256])]; + tensor var_298_end_mask_0 = const()[name = tensor("op_298_end_mask_0"), val = tensor([true, true, true])]; + tensor var_298 = slice_by_index(begin = var_298_begin_0, end = var_298_end_0, end_mask = var_298_end_mask_0, x = window_3)[name = tensor("op_298")]; tensor window_5_interleave_0 = const()[name = tensor("window_5_interleave_0"), val = tensor(false)]; - tensor window_5 = concat(axis = var_27, interleave = window_5_interleave_0, values = (var_233, var_230))[name = tensor("window_5")]; - tensor var_238_begin_0 = const()[name = tensor("op_238_begin_0"), val = tensor([0, 2, 0])]; - tensor var_238_end_0 = const()[name = tensor("op_238_end_0"), val = tensor([1, 3, 256])]; - tensor var_238_end_mask_0 = const()[name = tensor("op_238_end_mask_0"), val = tensor([true, false, true])]; - tensor var_238 = slice_by_index(begin = var_238_begin_0, end = var_238_end_0, end_mask = var_238_end_mask_0, x = x_3)[name = tensor("op_238")]; - tensor var_241_begin_0 = const()[name = tensor("op_241_begin_0"), val = tensor([0, 1, 0])]; - tensor var_241_end_0 = const()[name = tensor("op_241_end_0"), val = tensor([1, 16, 256])]; - tensor var_241_end_mask_0 = const()[name = tensor("op_241_end_mask_0"), val = tensor([true, true, true])]; - tensor var_241 = slice_by_index(begin = var_241_begin_0, end = var_241_end_0, end_mask = var_241_end_mask_0, x = window_5)[name = tensor("op_241")]; + tensor window_5 = concat(axis = var_92, interleave = window_5_interleave_0, values = (var_298, var_295))[name = tensor("window_5")]; + tensor var_303_begin_0 = const()[name = tensor("op_303_begin_0"), val = tensor([0, 2, 0])]; + tensor var_303_end_0 = const()[name = tensor("op_303_end_0"), val = tensor([1, 3, 256])]; + tensor var_303_end_mask_0 = const()[name = tensor("op_303_end_mask_0"), val = tensor([true, false, true])]; + tensor var_303 = slice_by_index(begin = var_303_begin_0, end = var_303_end_0, end_mask = var_303_end_mask_0, x = x_3)[name = tensor("op_303")]; + tensor var_306_begin_0 = const()[name = tensor("op_306_begin_0"), val = tensor([0, 1, 0])]; + tensor var_306_end_0 = const()[name = tensor("op_306_end_0"), val = tensor([1, 16, 256])]; + tensor var_306_end_mask_0 = const()[name = tensor("op_306_end_mask_0"), val = tensor([true, true, true])]; + tensor var_306 = slice_by_index(begin = var_306_begin_0, end = var_306_end_0, end_mask = var_306_end_mask_0, x = window_5)[name = tensor("op_306")]; tensor window_7_interleave_0 = const()[name = tensor("window_7_interleave_0"), val = tensor(false)]; - tensor window_7 = concat(axis = var_27, interleave = window_7_interleave_0, values = (var_241, var_238))[name = tensor("window_7")]; - tensor var_246_begin_0 = const()[name = tensor("op_246_begin_0"), val = tensor([0, 3, 0])]; - tensor var_246_end_0 = const()[name = tensor("op_246_end_0"), val = tensor([1, 4, 256])]; - tensor var_246_end_mask_0 = const()[name = tensor("op_246_end_mask_0"), val = tensor([true, false, true])]; - tensor var_246 = slice_by_index(begin = var_246_begin_0, end = var_246_end_0, end_mask = var_246_end_mask_0, x = x_3)[name = tensor("op_246")]; - tensor var_249_begin_0 = const()[name = tensor("op_249_begin_0"), val = tensor([0, 1, 0])]; - tensor var_249_end_0 = const()[name = tensor("op_249_end_0"), val = tensor([1, 16, 256])]; - tensor var_249_end_mask_0 = const()[name = tensor("op_249_end_mask_0"), val = tensor([true, true, true])]; - tensor var_249 = slice_by_index(begin = var_249_begin_0, end = var_249_end_0, end_mask = var_249_end_mask_0, x = window_7)[name = tensor("op_249")]; + tensor window_7 = concat(axis = var_92, interleave = window_7_interleave_0, values = (var_306, var_303))[name = tensor("window_7")]; + tensor var_311_begin_0 = const()[name = tensor("op_311_begin_0"), val = tensor([0, 3, 0])]; + tensor var_311_end_0 = const()[name = tensor("op_311_end_0"), val = tensor([1, 4, 256])]; + tensor var_311_end_mask_0 = const()[name = tensor("op_311_end_mask_0"), val = tensor([true, false, true])]; + tensor var_311 = slice_by_index(begin = var_311_begin_0, end = var_311_end_0, end_mask = var_311_end_mask_0, x = x_3)[name = tensor("op_311")]; + tensor var_314_begin_0 = const()[name = tensor("op_314_begin_0"), val = tensor([0, 1, 0])]; + tensor var_314_end_0 = const()[name = tensor("op_314_end_0"), val = tensor([1, 16, 256])]; + tensor var_314_end_mask_0 = const()[name = tensor("op_314_end_mask_0"), val = tensor([true, true, true])]; + tensor var_314 = slice_by_index(begin = var_314_begin_0, end = var_314_end_0, end_mask = var_314_end_mask_0, x = window_7)[name = tensor("op_314")]; tensor window_9_interleave_0 = const()[name = tensor("window_9_interleave_0"), val = tensor(false)]; - tensor window_9 = concat(axis = var_27, interleave = window_9_interleave_0, values = (var_249, var_246))[name = tensor("window_9")]; - tensor var_254_begin_0 = const()[name = tensor("op_254_begin_0"), val = tensor([0, 4, 0])]; - tensor var_254_end_0 = const()[name = tensor("op_254_end_0"), val = tensor([1, 1, 256])]; - tensor var_254_end_mask_0 = const()[name = tensor("op_254_end_mask_0"), val = tensor([true, true, true])]; - tensor var_254 = slice_by_index(begin = var_254_begin_0, end = var_254_end_0, end_mask = var_254_end_mask_0, x = x_3)[name = tensor("op_254")]; - tensor var_257_begin_0 = const()[name = tensor("op_257_begin_0"), val = tensor([0, 1, 0])]; - tensor var_257_end_0 = const()[name = tensor("op_257_end_0"), val = tensor([1, 16, 256])]; - tensor var_257_end_mask_0 = const()[name = tensor("op_257_end_mask_0"), val = tensor([true, true, true])]; - tensor var_257 = slice_by_index(begin = var_257_begin_0, end = var_257_end_0, end_mask = var_257_end_mask_0, x = window_9)[name = tensor("op_257")]; + tensor window_9 = concat(axis = var_92, interleave = window_9_interleave_0, values = (var_314, var_311))[name = tensor("window_9")]; + tensor var_319_begin_0 = const()[name = tensor("op_319_begin_0"), val = tensor([0, 4, 0])]; + tensor var_319_end_0 = const()[name = tensor("op_319_end_0"), val = tensor([1, 1, 256])]; + tensor var_319_end_mask_0 = const()[name = tensor("op_319_end_mask_0"), val = tensor([true, true, true])]; + tensor var_319 = slice_by_index(begin = var_319_begin_0, end = var_319_end_0, end_mask = var_319_end_mask_0, x = x_3)[name = tensor("op_319")]; + tensor var_322_begin_0 = const()[name = tensor("op_322_begin_0"), val = tensor([0, 1, 0])]; + tensor var_322_end_0 = const()[name = tensor("op_322_end_0"), val = tensor([1, 16, 256])]; + tensor var_322_end_mask_0 = const()[name = tensor("op_322_end_mask_0"), val = tensor([true, true, true])]; + tensor var_322 = slice_by_index(begin = var_322_begin_0, end = var_322_end_0, end_mask = var_322_end_mask_0, x = window_9)[name = tensor("op_322")]; tensor window_11_interleave_0 = const()[name = tensor("window_11_interleave_0"), val = tensor(false)]; - tensor window_11 = concat(axis = var_27, interleave = window_11_interleave_0, values = (var_257, var_254))[name = tensor("window_11")]; - tensor input_21_interleave_0 = const()[name = tensor("input_21_interleave_0"), val = tensor(false)]; - tensor input_21 = concat(axis = var_24, interleave = input_21_interleave_0, values = (window_3, window_5, window_7, window_9, window_11))[name = tensor("input_21")]; + tensor window_11 = concat(axis = var_92, interleave = window_11_interleave_0, values = (var_322, var_319))[name = tensor("window_11")]; + tensor input_23_interleave_0 = const()[name = tensor("input_23_interleave_0"), val = tensor(false)]; + tensor input_23 = concat(axis = var_77, interleave = input_23_interleave_0, values = (window_3, window_5, window_7, window_9, window_11))[name = tensor("input_23")]; tensor x_5_axes_0 = const()[name = tensor("x_5_axes_0"), val = tensor([-1])]; - tensor x_5 = layer_norm(axes = x_5_axes_0, beta = encoder_conv_module_0_sequential_0_bias, epsilon = var_29, gamma = encoder_conv_module_0_sequential_0_weight, x = input_21)[name = tensor("x_5")]; - tensor input_23_perm_0 = const()[name = tensor("input_23_perm_0"), val = tensor([0, 2, 1])]; + tensor x_5 = layer_norm(axes = x_5_axes_0, beta = inner_encoder_conv_module_0_sequential_0_bias, epsilon = var_74, gamma = inner_encoder_conv_module_0_sequential_0_weight, x = input_23)[name = tensor("x_5")]; + tensor input_25_perm_0 = const()[name = tensor("input_25_perm_0"), val = tensor([0, 2, 1])]; tensor inputs_3_pad_type_0 = const()[name = tensor("inputs_3_pad_type_0"), val = tensor("valid")]; tensor inputs_3_strides_0 = const()[name = tensor("inputs_3_strides_0"), val = tensor([1])]; tensor inputs_3_pad_0 = const()[name = tensor("inputs_3_pad_0"), val = tensor([0, 0])]; tensor inputs_3_dilations_0 = const()[name = tensor("inputs_3_dilations_0"), val = tensor([1])]; tensor inputs_3_groups_0 = const()[name = tensor("inputs_3_groups_0"), val = tensor(1)]; - tensor input_23 = transpose(perm = input_23_perm_0, x = x_5)[name = tensor("transpose_54")]; - tensor inputs_3 = conv(bias = encoder_conv_module_0_sequential_2_conv_bias, dilations = inputs_3_dilations_0, groups = inputs_3_groups_0, pad = inputs_3_pad_0, pad_type = inputs_3_pad_type_0, strides = inputs_3_strides_0, weight = encoder_conv_module_0_sequential_2_conv_weight, x = input_23)[name = tensor("inputs_3")]; - tensor var_282_split_sizes_0 = const()[name = tensor("op_282_split_sizes_0"), val = tensor([256, 256])]; - tensor var_282_axis_0 = const()[name = tensor("op_282_axis_0"), val = tensor(1)]; - tensor var_282_0, tensor var_282_1 = split(axis = var_282_axis_0, split_sizes = var_282_split_sizes_0, x = inputs_3)[name = tensor("op_282")]; - tensor var_284 = sigmoid(x = var_282_1)[name = tensor("op_284")]; - tensor inputs_5 = mul(x = var_282_0, y = var_284)[name = tensor("inputs_5")]; + tensor input_25 = transpose(perm = input_25_perm_0, x = x_5)[name = tensor("transpose_54")]; + tensor inputs_3 = conv(bias = inner_encoder_conv_module_0_sequential_2_conv_bias, dilations = inputs_3_dilations_0, groups = inputs_3_groups_0, pad = inputs_3_pad_0, pad_type = inputs_3_pad_type_0, strides = inputs_3_strides_0, weight = inner_encoder_conv_module_0_sequential_2_conv_weight, x = input_25)[name = tensor("inputs_3")]; + tensor var_347_split_sizes_0 = const()[name = tensor("op_347_split_sizes_0"), val = tensor([256, 256])]; + tensor var_347_axis_0 = const()[name = tensor("op_347_axis_0"), val = tensor(1)]; + tensor var_347_0, tensor var_347_1 = split(axis = var_347_axis_0, split_sizes = var_347_split_sizes_0, x = inputs_3)[name = tensor("op_347")]; + tensor var_349 = sigmoid(x = var_347_1)[name = tensor("op_349")]; + tensor inputs_5 = mul(x = var_347_0, y = var_349)[name = tensor("inputs_5")]; tensor outputs_aug_1_pad_type_0 = const()[name = tensor("outputs_aug_1_pad_type_0"), val = tensor("custom")]; tensor outputs_aug_1_pad_0 = const()[name = tensor("outputs_aug_1_pad_0"), val = tensor([15, 15])]; tensor outputs_aug_1_groups_0 = const()[name = tensor("outputs_aug_1_groups_0"), val = tensor(256)]; tensor outputs_aug_1_strides_0 = const()[name = tensor("outputs_aug_1_strides_0"), val = tensor([1])]; tensor outputs_aug_1_dilations_0 = const()[name = tensor("outputs_aug_1_dilations_0"), val = tensor([1])]; - tensor outputs_aug_1 = conv(dilations = outputs_aug_1_dilations_0, groups = outputs_aug_1_groups_0, pad = outputs_aug_1_pad_0, pad_type = outputs_aug_1_pad_type_0, strides = outputs_aug_1_strides_0, weight = encoder_conv_module_0_sequential_4_conv_weight, x = inputs_5)[name = tensor("outputs_aug_1")]; - tensor input_25_begin_0 = const()[name = tensor("input_25_begin_0"), val = tensor([0, 0, 0])]; - tensor input_25_end_0 = const()[name = tensor("input_25_end_0"), val = tensor([5, 256, 16])]; - tensor input_25_end_mask_0 = const()[name = tensor("input_25_end_mask_0"), val = tensor([true, true, false])]; - tensor input_25 = slice_by_index(begin = input_25_begin_0, end = input_25_end_0, end_mask = input_25_end_mask_0, x = outputs_aug_1)[name = tensor("input_25")]; - tensor inputs_7 = batch_norm(beta = encoder_conv_module_0_sequential_5_bias, epsilon = var_29, gamma = encoder_conv_module_0_sequential_5_weight, mean = encoder_conv_module_0_sequential_5_running_mean, variance = encoder_conv_module_0_sequential_5_running_var, x = input_25)[name = tensor("inputs_7")]; - tensor input_27 = silu(x = inputs_7)[name = tensor("input_27")]; - tensor input_29_pad_type_0 = const()[name = tensor("input_29_pad_type_0"), val = tensor("valid")]; - tensor input_29_strides_0 = const()[name = tensor("input_29_strides_0"), val = tensor([1])]; - tensor input_29_pad_0 = const()[name = tensor("input_29_pad_0"), val = tensor([0, 0])]; - tensor input_29_dilations_0 = const()[name = tensor("input_29_dilations_0"), val = tensor([1])]; - tensor input_29_groups_0 = const()[name = tensor("input_29_groups_0"), val = tensor(1)]; - tensor input_29 = conv(bias = encoder_conv_module_0_sequential_7_conv_bias, dilations = input_29_dilations_0, groups = input_29_groups_0, pad = input_29_pad_0, pad_type = input_29_pad_type_0, strides = input_29_strides_0, weight = encoder_conv_module_0_sequential_7_conv_weight, x = input_27)[name = tensor("input_29")]; + tensor outputs_aug_1 = conv(dilations = outputs_aug_1_dilations_0, groups = outputs_aug_1_groups_0, pad = outputs_aug_1_pad_0, pad_type = outputs_aug_1_pad_type_0, strides = outputs_aug_1_strides_0, weight = inner_encoder_conv_module_0_sequential_4_conv_weight, x = inputs_5)[name = tensor("outputs_aug_1")]; + tensor input_27_begin_0 = const()[name = tensor("input_27_begin_0"), val = tensor([0, 0, 0])]; + tensor input_27_end_0 = const()[name = tensor("input_27_end_0"), val = tensor([5, 256, 16])]; + tensor input_27_end_mask_0 = const()[name = tensor("input_27_end_mask_0"), val = tensor([true, true, false])]; + tensor input_27 = slice_by_index(begin = input_27_begin_0, end = input_27_end_0, end_mask = input_27_end_mask_0, x = outputs_aug_1)[name = tensor("input_27")]; + tensor inputs_7 = batch_norm(beta = inner_encoder_conv_module_0_sequential_5_bias, epsilon = var_74, gamma = inner_encoder_conv_module_0_sequential_5_weight, mean = inner_encoder_conv_module_0_sequential_5_running_mean, variance = inner_encoder_conv_module_0_sequential_5_running_var, x = input_27)[name = tensor("inputs_7")]; + tensor input_29 = silu(x = inputs_7)[name = tensor("input_29")]; + tensor input_31_pad_type_0 = const()[name = tensor("input_31_pad_type_0"), val = tensor("valid")]; + tensor input_31_strides_0 = const()[name = tensor("input_31_strides_0"), val = tensor([1])]; + tensor input_31_pad_0 = const()[name = tensor("input_31_pad_0"), val = tensor([0, 0])]; + tensor input_31_dilations_0 = const()[name = tensor("input_31_dilations_0"), val = tensor([1])]; + tensor input_31_groups_0 = const()[name = tensor("input_31_groups_0"), val = tensor(1)]; + tensor input_31 = conv(bias = inner_encoder_conv_module_0_sequential_7_conv_bias, dilations = input_31_dilations_0, groups = input_31_groups_0, pad = input_31_pad_0, pad_type = input_31_pad_type_0, strides = input_31_strides_0, weight = inner_encoder_conv_module_0_sequential_7_conv_weight, x = input_29)[name = tensor("input_31")]; tensor conv_out_1_perm_0 = const()[name = tensor("conv_out_1_perm_0"), val = tensor([0, 2, 1])]; - tensor var_315_begin_0 = const()[name = tensor("op_315_begin_0"), val = tensor([0, -1, 0])]; - tensor var_315_end_0 = const()[name = tensor("op_315_end_0"), val = tensor([5, 16, 256])]; - tensor var_315_end_mask_0 = const()[name = tensor("op_315_end_mask_0"), val = tensor([true, true, true])]; - tensor conv_out_1 = transpose(perm = conv_out_1_perm_0, x = input_29)[name = tensor("transpose_53")]; - tensor var_315 = slice_by_index(begin = var_315_begin_0, end = var_315_end_0, end_mask = var_315_end_mask_0, x = conv_out_1)[name = tensor("op_315")]; - tensor var_317_perm_0 = const()[name = tensor("op_317_perm_0"), val = tensor([1, 0, 2])]; - tensor var_317 = transpose(perm = var_317_perm_0, x = var_315)[name = tensor("transpose_52")]; - tensor input_31 = add(x = x_3, y = var_317)[name = tensor("input_31")]; - tensor input_33_axes_0 = const()[name = tensor("input_33_axes_0"), val = tensor([-1])]; - tensor input_33 = layer_norm(axes = input_33_axes_0, beta = encoder_ffn2_0_module_sequential_0_bias, epsilon = var_29, gamma = encoder_ffn2_0_module_sequential_0_weight, x = input_31)[name = tensor("input_33")]; - tensor inputs_9 = linear(bias = encoder_ffn2_0_module_sequential_1_linear_bias, weight = encoder_ffn2_0_module_sequential_1_linear_weight, x = input_33)[name = tensor("linear_8")]; - tensor input_35 = silu(x = inputs_9)[name = tensor("input_35")]; - tensor input_39 = linear(bias = encoder_ffn2_0_module_sequential_4_linear_bias, weight = encoder_ffn2_0_module_sequential_4_linear_weight, x = input_35)[name = tensor("linear_9")]; - tensor var_340 = const()[name = tensor("op_340"), val = tensor(0x1p-1)]; - tensor var_341 = mul(x = input_39, y = var_340)[name = tensor("op_341")]; - tensor input_41 = add(x = var_341, y = input_31)[name = tensor("input_41")]; - tensor input_43_axes_0 = const()[name = tensor("input_43_axes_0"), val = tensor([-1])]; - tensor input_43 = layer_norm(axes = input_43_axes_0, beta = encoder_layer_norm_0_bias, epsilon = var_29, gamma = encoder_layer_norm_0_weight, x = input_41)[name = tensor("input_43")]; + tensor var_380_begin_0 = const()[name = tensor("op_380_begin_0"), val = tensor([0, -1, 0])]; + tensor var_380_end_0 = const()[name = tensor("op_380_end_0"), val = tensor([5, 16, 256])]; + tensor var_380_end_mask_0 = const()[name = tensor("op_380_end_mask_0"), val = tensor([true, true, true])]; + tensor conv_out_1 = transpose(perm = conv_out_1_perm_0, x = input_31)[name = tensor("transpose_53")]; + tensor var_380 = slice_by_index(begin = var_380_begin_0, end = var_380_end_0, end_mask = var_380_end_mask_0, x = conv_out_1)[name = tensor("op_380")]; + tensor var_382_perm_0 = const()[name = tensor("op_382_perm_0"), val = tensor([1, 0, 2])]; + tensor var_382 = transpose(perm = var_382_perm_0, x = var_380)[name = tensor("transpose_52")]; + tensor input_33 = add(x = x_3, y = var_382)[name = tensor("input_33")]; + tensor input_35_axes_0 = const()[name = tensor("input_35_axes_0"), val = tensor([-1])]; + tensor input_35 = layer_norm(axes = input_35_axes_0, beta = inner_encoder_ffn2_0_module_sequential_0_bias, epsilon = var_74, gamma = inner_encoder_ffn2_0_module_sequential_0_weight, x = input_33)[name = tensor("input_35")]; + tensor inputs_9 = linear(bias = inner_encoder_ffn2_0_module_sequential_1_linear_bias, weight = inner_encoder_ffn2_0_module_sequential_1_linear_weight, x = input_35)[name = tensor("linear_8")]; + tensor input_37 = silu(x = inputs_9)[name = tensor("input_37")]; + tensor input_41 = linear(bias = inner_encoder_ffn2_0_module_sequential_4_linear_bias, weight = inner_encoder_ffn2_0_module_sequential_4_linear_weight, x = input_37)[name = tensor("linear_9")]; + tensor var_405 = const()[name = tensor("op_405"), val = tensor(0x1p-1)]; + tensor var_406 = mul(x = input_41, y = var_405)[name = tensor("op_406")]; + tensor input_43 = add(x = var_406, y = input_33)[name = tensor("input_43")]; tensor input_45_axes_0 = const()[name = tensor("input_45_axes_0"), val = tensor([-1])]; - tensor input_45 = layer_norm(axes = input_45_axes_0, beta = encoder_ffn1_1_module_sequential_0_bias, epsilon = var_29, gamma = encoder_ffn1_1_module_sequential_0_weight, x = input_43)[name = tensor("input_45")]; - tensor inputs_11 = linear(bias = encoder_ffn1_1_module_sequential_1_linear_bias, weight = encoder_ffn1_1_module_sequential_1_linear_weight, x = input_45)[name = tensor("linear_10")]; - tensor input_47 = silu(x = inputs_11)[name = tensor("input_47")]; - tensor input_51 = linear(bias = encoder_ffn1_1_module_sequential_4_linear_bias, weight = encoder_ffn1_1_module_sequential_4_linear_weight, x = input_47)[name = tensor("linear_11")]; - tensor var_370 = const()[name = tensor("op_370"), val = tensor(0x1p-1)]; - tensor var_371 = mul(x = input_51, y = var_370)[name = tensor("op_371")]; - tensor input_53 = add(x = var_371, y = input_43)[name = tensor("input_53")]; + tensor input_45 = layer_norm(axes = input_45_axes_0, beta = inner_encoder_layer_norm_0_bias, epsilon = var_74, gamma = inner_encoder_layer_norm_0_weight, x = input_43)[name = tensor("input_45")]; + tensor input_47_axes_0 = const()[name = tensor("input_47_axes_0"), val = tensor([-1])]; + tensor input_47 = layer_norm(axes = input_47_axes_0, beta = inner_encoder_ffn1_1_module_sequential_0_bias, epsilon = var_74, gamma = inner_encoder_ffn1_1_module_sequential_0_weight, x = input_45)[name = tensor("input_47")]; + tensor inputs_11 = linear(bias = inner_encoder_ffn1_1_module_sequential_1_linear_bias, weight = inner_encoder_ffn1_1_module_sequential_1_linear_weight, x = input_47)[name = tensor("linear_10")]; + tensor input_49 = silu(x = inputs_11)[name = tensor("input_49")]; + tensor input_53 = linear(bias = inner_encoder_ffn1_1_module_sequential_4_linear_bias, weight = inner_encoder_ffn1_1_module_sequential_4_linear_weight, x = input_49)[name = tensor("linear_11")]; + tensor var_435 = const()[name = tensor("op_435"), val = tensor(0x1p-1)]; + tensor var_436 = mul(x = input_53, y = var_435)[name = tensor("op_436")]; + tensor input_55 = add(x = var_436, y = input_45)[name = tensor("input_55")]; tensor x_7_axes_0 = const()[name = tensor("x_7_axes_0"), val = tensor([-1])]; - tensor x_7 = layer_norm(axes = x_7_axes_0, beta = encoder_ret_lns_1_bias, epsilon = var_29, gamma = encoder_ret_lns_1_weight, x = input_53)[name = tensor("x_7")]; + tensor x_7 = layer_norm(axes = x_7_axes_0, beta = inner_encoder_ret_lns_1_bias, epsilon = var_74, gamma = inner_encoder_ret_lns_1_weight, x = input_55)[name = tensor("x_7")]; tensor prev_kv_3_begin_0 = const()[name = tensor("prev_kv_3_begin_0"), val = tensor([1, 0, 0, 0, 0])]; tensor prev_kv_3_end_0 = const()[name = tensor("prev_kv_3_end_0"), val = tensor([2, 1, 4, 64, 64])]; tensor prev_kv_3_end_mask_0 = const()[name = tensor("prev_kv_3_end_mask_0"), val = tensor([false, true, true, true, true])]; @@ -426,183 +452,183 @@ program(1.0) tensor prev_scale_3_end_mask_0 = const()[name = tensor("prev_scale_3_end_mask_0"), val = tensor([false, true])]; tensor prev_scale_3_squeeze_mask_0 = const()[name = tensor("prev_scale_3_squeeze_mask_0"), val = tensor([true, false])]; tensor prev_scale_3 = slice_by_index(begin = prev_scale_3_begin_0, end = prev_scale_3_end_0, end_mask = prev_scale_3_end_mask_0, squeeze_mask = prev_scale_3_squeeze_mask_0, x = enc_scale)[name = tensor("prev_scale_3")]; - tensor var_385 = linear(bias = encoder_q_proj_1_bias, weight = encoder_q_proj_1_weight, x = x_7)[name = tensor("linear_12")]; - tensor var_386 = const()[name = tensor("op_386"), val = tensor([1, 5, 4, 64])]; - tensor var_387 = reshape(shape = var_386, x = var_385)[name = tensor("op_387")]; + tensor var_450 = linear(bias = inner_encoder_q_proj_1_bias, weight = inner_encoder_q_proj_1_weight, x = x_7)[name = tensor("linear_12")]; + tensor var_451 = const()[name = tensor("op_451"), val = tensor([1, 5, 4, 64])]; + tensor var_452 = reshape(shape = var_451, x = var_450)[name = tensor("op_452")]; tensor q_3_perm_0 = const()[name = tensor("q_3_perm_0"), val = tensor([0, 2, 1, 3])]; - tensor var_391 = linear(bias = encoder_k_proj_1_bias, weight = encoder_k_proj_1_weight, x = x_7)[name = tensor("linear_13")]; - tensor var_392 = const()[name = tensor("op_392"), val = tensor(0x1p-3)]; - tensor var_393 = mul(x = var_391, y = var_392)[name = tensor("op_393")]; - tensor var_394 = const()[name = tensor("op_394"), val = tensor([1, 5, 4, 64])]; - tensor var_395 = reshape(shape = var_394, x = var_393)[name = tensor("op_395")]; + tensor var_456 = linear(bias = inner_encoder_k_proj_1_bias, weight = inner_encoder_k_proj_1_weight, x = x_7)[name = tensor("linear_13")]; + tensor var_457 = const()[name = tensor("op_457"), val = tensor(0x1p-3)]; + tensor var_458 = mul(x = var_456, y = var_457)[name = tensor("op_458")]; + tensor var_459 = const()[name = tensor("op_459"), val = tensor([1, 5, 4, 64])]; + tensor var_460 = reshape(shape = var_459, x = var_458)[name = tensor("op_460")]; tensor k_3_perm_0 = const()[name = tensor("k_3_perm_0"), val = tensor([0, 2, 1, 3])]; - tensor var_399 = linear(bias = encoder_v_proj_1_bias, weight = encoder_v_proj_1_weight, x = x_7)[name = tensor("linear_14")]; - tensor var_400 = const()[name = tensor("op_400"), val = tensor([1, 5, 4, 64])]; - tensor var_401 = reshape(shape = var_400, x = var_399)[name = tensor("op_401")]; + tensor var_464 = linear(bias = inner_encoder_v_proj_1_bias, weight = inner_encoder_v_proj_1_weight, x = x_7)[name = tensor("linear_14")]; + tensor var_465 = const()[name = tensor("op_465"), val = tensor([1, 5, 4, 64])]; + tensor var_466 = reshape(shape = var_465, x = var_464)[name = tensor("op_466")]; tensor v_3_perm_0 = const()[name = tensor("v_3_perm_0"), val = tensor([0, 2, 1, 3])]; - tensor input_57 = linear(bias = encoder_g_proj_1_bias, weight = encoder_g_proj_1_weight, x = x_7)[name = tensor("linear_15")]; + tensor input_59 = linear(bias = inner_encoder_g_proj_1_bias, weight = inner_encoder_g_proj_1_weight, x = x_7)[name = tensor("linear_15")]; tensor sqrt_s0_3 = sqrt(x = prev_scale_3)[name = tensor("sqrt_s0_3")]; - tensor s_t_3 = add(x = prev_scale_3, y = encoder__t_index)[name = tensor("s_t_3")]; + tensor s_t_3 = add(x = prev_scale_3, y = inner_encoder__t_index)[name = tensor("s_t_3")]; tensor sqrt_s_t_3 = sqrt(x = s_t_3)[name = tensor("sqrt_s_t_3")]; tensor qk_3_transpose_x_1 = const()[name = tensor("qk_3_transpose_x_1"), val = tensor(false)]; tensor qk_3_transpose_y_1 = const()[name = tensor("qk_3_transpose_y_1"), val = tensor(true)]; - tensor k_3 = transpose(perm = k_3_perm_0, x = var_395)[name = tensor("transpose_50")]; - tensor q_3 = transpose(perm = q_3_perm_0, x = var_387)[name = tensor("transpose_51")]; + tensor k_3 = transpose(perm = k_3_perm_0, x = var_460)[name = tensor("transpose_50")]; + tensor q_3 = transpose(perm = q_3_perm_0, x = var_452)[name = tensor("transpose_51")]; tensor qk_3 = matmul(transpose_x = qk_3_transpose_x_1, transpose_y = qk_3_transpose_y_1, x = q_3, y = k_3)[name = tensor("qk_3")]; - tensor var_411 = const()[name = tensor("op_411"), val = tensor([5, 1])]; - tensor var_412 = reshape(shape = var_411, x = sqrt_s_t_3)[name = tensor("op_412")]; - tensor M_3 = real_div(x = encoder__causal_mask, y = var_412)[name = tensor("M_3")]; - tensor var_414 = mul(x = qk_3, y = M_3)[name = tensor("op_414")]; + tensor var_476 = const()[name = tensor("op_476"), val = tensor([5, 1])]; + tensor var_477 = reshape(shape = var_476, x = sqrt_s_t_3)[name = tensor("op_477")]; + tensor M_3 = real_div(x = inner_encoder__causal_mask, y = var_477)[name = tensor("M_3")]; + tensor var_479 = mul(x = qk_3, y = M_3)[name = tensor("op_479")]; tensor inner_3_transpose_x_0 = const()[name = tensor("inner_3_transpose_x_0"), val = tensor(false)]; tensor inner_3_transpose_y_0 = const()[name = tensor("inner_3_transpose_y_0"), val = tensor(false)]; - tensor v_3 = transpose(perm = v_3_perm_0, x = var_401)[name = tensor("transpose_49")]; - tensor inner_3 = matmul(transpose_x = inner_3_transpose_x_0, transpose_y = inner_3_transpose_y_0, x = var_414, y = v_3)[name = tensor("inner_3")]; - tensor var_416_transpose_x_0 = const()[name = tensor("op_416_transpose_x_0"), val = tensor(false)]; - tensor var_416_transpose_y_0 = const()[name = tensor("op_416_transpose_y_0"), val = tensor(false)]; - tensor var_416 = matmul(transpose_x = var_416_transpose_x_0, transpose_y = var_416_transpose_y_0, x = q_3, y = prev_kv_3)[name = tensor("op_416")]; - tensor var_417 = real_div(x = sqrt_s0_3, y = sqrt_s_t_3)[name = tensor("op_417")]; - tensor var_418 = const()[name = tensor("op_418"), val = tensor([1, 1, 5, 1])]; - tensor var_419 = reshape(shape = var_418, x = var_417)[name = tensor("op_419")]; - tensor cross_3 = mul(x = var_416, y = var_419)[name = tensor("cross_3")]; + tensor v_3 = transpose(perm = v_3_perm_0, x = var_466)[name = tensor("transpose_49")]; + tensor inner_3 = matmul(transpose_x = inner_3_transpose_x_0, transpose_y = inner_3_transpose_y_0, x = var_479, y = v_3)[name = tensor("inner_3")]; + tensor var_481_transpose_x_0 = const()[name = tensor("op_481_transpose_x_0"), val = tensor(false)]; + tensor var_481_transpose_y_0 = const()[name = tensor("op_481_transpose_y_0"), val = tensor(false)]; + tensor var_481 = matmul(transpose_x = var_481_transpose_x_0, transpose_y = var_481_transpose_y_0, x = q_3, y = prev_kv_3)[name = tensor("op_481")]; + tensor var_482 = real_div(x = sqrt_s0_3, y = sqrt_s_t_3)[name = tensor("op_482")]; + tensor var_483 = const()[name = tensor("op_483"), val = tensor([1, 1, 5, 1])]; + tensor var_484 = reshape(shape = var_483, x = var_482)[name = tensor("op_484")]; + tensor cross_3 = mul(x = var_481, y = var_484)[name = tensor("cross_3")]; tensor out_7 = add(x = inner_3, y = cross_3)[name = tensor("out_7")]; - tensor var_422 = mul(x = prev_kv_3, y = sqrt_s0_3)[name = tensor("op_422")]; - tensor var_424_transpose_x_1 = const()[name = tensor("op_424_transpose_x_1"), val = tensor(true)]; - tensor var_424_transpose_y_1 = const()[name = tensor("op_424_transpose_y_1"), val = tensor(false)]; - tensor var_424 = matmul(transpose_x = var_424_transpose_x_1, transpose_y = var_424_transpose_y_1, x = k_3, y = v_3)[name = tensor("op_424")]; - tensor new_kv_unnorm_3 = add(x = var_422, y = var_424)[name = tensor("new_kv_unnorm_3")]; - tensor var_426 = const()[name = tensor("op_426"), val = tensor(0x1.4p+2)]; - tensor new_scale_3 = add(x = prev_scale_3, y = var_426)[name = tensor("new_scale_3")]; - tensor var_428 = sqrt(x = new_scale_3)[name = tensor("op_428")]; - tensor var_429 = real_div(x = new_kv_unnorm_3, y = var_428)[name = tensor("op_429")]; - tensor var_430_perm_0 = const()[name = tensor("op_430_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_487 = mul(x = prev_kv_3, y = sqrt_s0_3)[name = tensor("op_487")]; + tensor var_489_transpose_x_1 = const()[name = tensor("op_489_transpose_x_1"), val = tensor(true)]; + tensor var_489_transpose_y_1 = const()[name = tensor("op_489_transpose_y_1"), val = tensor(false)]; + tensor var_489 = matmul(transpose_x = var_489_transpose_x_1, transpose_y = var_489_transpose_y_1, x = k_3, y = v_3)[name = tensor("op_489")]; + tensor new_kv_unnorm_3 = add(x = var_487, y = var_489)[name = tensor("new_kv_unnorm_3")]; + tensor var_491 = const()[name = tensor("op_491"), val = tensor(0x1.4p+2)]; + tensor new_scale_3 = add(x = prev_scale_3, y = var_491)[name = tensor("new_scale_3")]; + tensor var_493 = sqrt(x = new_scale_3)[name = tensor("op_493")]; + tensor var_494 = real_div(x = new_kv_unnorm_3, y = var_493)[name = tensor("op_494")]; + tensor var_495_perm_0 = const()[name = tensor("op_495_perm_0"), val = tensor([0, 2, 1, 3])]; tensor out_9_axes_0 = const()[name = tensor("out_9_axes_0"), val = tensor([-1])]; - tensor var_430 = transpose(perm = var_430_perm_0, x = out_7)[name = tensor("transpose_48")]; - tensor out_9 = layer_norm(axes = out_9_axes_0, epsilon = var_18, x = var_430)[name = tensor("out_9")]; - tensor var_434 = const()[name = tensor("op_434"), val = tensor([1, 5, 256])]; - tensor out_11 = reshape(shape = var_434, x = out_9)[name = tensor("out_11")]; - tensor var_436 = silu(x = input_57)[name = tensor("op_436")]; - tensor input_59 = mul(x = var_436, y = out_11)[name = tensor("input_59")]; - tensor ret_out_3 = linear(bias = encoder_out_proj_1_bias, weight = encoder_out_proj_1_weight, x = input_59)[name = tensor("linear_16")]; - tensor x_9 = add(x = input_53, y = ret_out_3)[name = tensor("x_9")]; + tensor var_495 = transpose(perm = var_495_perm_0, x = out_7)[name = tensor("transpose_48")]; + tensor out_9 = layer_norm(axes = out_9_axes_0, epsilon = var_82, x = var_495)[name = tensor("out_9")]; + tensor var_499 = const()[name = tensor("op_499"), val = tensor([1, 5, 256])]; + tensor out_11 = reshape(shape = var_499, x = out_9)[name = tensor("out_11")]; + tensor var_501 = silu(x = input_59)[name = tensor("op_501")]; + tensor input_61 = mul(x = var_501, y = out_11)[name = tensor("input_61")]; + tensor ret_out_3 = linear(bias = inner_encoder_out_proj_1_bias, weight = inner_encoder_out_proj_1_weight, x = input_61)[name = tensor("linear_16")]; + tensor x_9 = add(x = input_55, y = ret_out_3)[name = tensor("x_9")]; tensor window_13_begin_0 = const()[name = tensor("window_13_begin_0"), val = tensor([1, 0, 0, 0])]; tensor window_13_end_0 = const()[name = tensor("window_13_end_0"), val = tensor([2, 1, 16, 256])]; tensor window_13_end_mask_0 = const()[name = tensor("window_13_end_mask_0"), val = tensor([false, true, true, true])]; tensor window_13_squeeze_mask_0 = const()[name = tensor("window_13_squeeze_mask_0"), val = tensor([true, false, false, false])]; tensor window_13 = slice_by_index(begin = window_13_begin_0, end = window_13_end_0, end_mask = window_13_end_mask_0, squeeze_mask = window_13_squeeze_mask_0, x = enc_conv_cache)[name = tensor("window_13")]; - tensor var_444_begin_0 = const()[name = tensor("op_444_begin_0"), val = tensor([0, 0, 0])]; - tensor var_444_end_0 = const()[name = tensor("op_444_end_0"), val = tensor([1, 1, 256])]; - tensor var_444_end_mask_0 = const()[name = tensor("op_444_end_mask_0"), val = tensor([true, false, true])]; - tensor var_444 = slice_by_index(begin = var_444_begin_0, end = var_444_end_0, end_mask = var_444_end_mask_0, x = x_9)[name = tensor("op_444")]; - tensor var_447_begin_0 = const()[name = tensor("op_447_begin_0"), val = tensor([0, 1, 0])]; - tensor var_447_end_0 = const()[name = tensor("op_447_end_0"), val = tensor([1, 16, 256])]; - tensor var_447_end_mask_0 = const()[name = tensor("op_447_end_mask_0"), val = tensor([true, true, true])]; - tensor var_447 = slice_by_index(begin = var_447_begin_0, end = var_447_end_0, end_mask = var_447_end_mask_0, x = window_13)[name = tensor("op_447")]; + tensor var_509_begin_0 = const()[name = tensor("op_509_begin_0"), val = tensor([0, 0, 0])]; + tensor var_509_end_0 = const()[name = tensor("op_509_end_0"), val = tensor([1, 1, 256])]; + tensor var_509_end_mask_0 = const()[name = tensor("op_509_end_mask_0"), val = tensor([true, false, true])]; + tensor var_509 = slice_by_index(begin = var_509_begin_0, end = var_509_end_0, end_mask = var_509_end_mask_0, x = x_9)[name = tensor("op_509")]; + tensor var_512_begin_0 = const()[name = tensor("op_512_begin_0"), val = tensor([0, 1, 0])]; + tensor var_512_end_0 = const()[name = tensor("op_512_end_0"), val = tensor([1, 16, 256])]; + tensor var_512_end_mask_0 = const()[name = tensor("op_512_end_mask_0"), val = tensor([true, true, true])]; + tensor var_512 = slice_by_index(begin = var_512_begin_0, end = var_512_end_0, end_mask = var_512_end_mask_0, x = window_13)[name = tensor("op_512")]; tensor window_15_interleave_0 = const()[name = tensor("window_15_interleave_0"), val = tensor(false)]; - tensor window_15 = concat(axis = var_27, interleave = window_15_interleave_0, values = (var_447, var_444))[name = tensor("window_15")]; - tensor var_452_begin_0 = const()[name = tensor("op_452_begin_0"), val = tensor([0, 1, 0])]; - tensor var_452_end_0 = const()[name = tensor("op_452_end_0"), val = tensor([1, 2, 256])]; - tensor var_452_end_mask_0 = const()[name = tensor("op_452_end_mask_0"), val = tensor([true, false, true])]; - tensor var_452 = slice_by_index(begin = var_452_begin_0, end = var_452_end_0, end_mask = var_452_end_mask_0, x = x_9)[name = tensor("op_452")]; - tensor var_455_begin_0 = const()[name = tensor("op_455_begin_0"), val = tensor([0, 1, 0])]; - tensor var_455_end_0 = const()[name = tensor("op_455_end_0"), val = tensor([1, 16, 256])]; - tensor var_455_end_mask_0 = const()[name = tensor("op_455_end_mask_0"), val = tensor([true, true, true])]; - tensor var_455 = slice_by_index(begin = var_455_begin_0, end = var_455_end_0, end_mask = var_455_end_mask_0, x = window_15)[name = tensor("op_455")]; + tensor window_15 = concat(axis = var_92, interleave = window_15_interleave_0, values = (var_512, var_509))[name = tensor("window_15")]; + tensor var_517_begin_0 = const()[name = tensor("op_517_begin_0"), val = tensor([0, 1, 0])]; + tensor var_517_end_0 = const()[name = tensor("op_517_end_0"), val = tensor([1, 2, 256])]; + tensor var_517_end_mask_0 = const()[name = tensor("op_517_end_mask_0"), val = tensor([true, false, true])]; + tensor var_517 = slice_by_index(begin = var_517_begin_0, end = var_517_end_0, end_mask = var_517_end_mask_0, x = x_9)[name = tensor("op_517")]; + tensor var_520_begin_0 = const()[name = tensor("op_520_begin_0"), val = tensor([0, 1, 0])]; + tensor var_520_end_0 = const()[name = tensor("op_520_end_0"), val = tensor([1, 16, 256])]; + tensor var_520_end_mask_0 = const()[name = tensor("op_520_end_mask_0"), val = tensor([true, true, true])]; + tensor var_520 = slice_by_index(begin = var_520_begin_0, end = var_520_end_0, end_mask = var_520_end_mask_0, x = window_15)[name = tensor("op_520")]; tensor window_17_interleave_0 = const()[name = tensor("window_17_interleave_0"), val = tensor(false)]; - tensor window_17 = concat(axis = var_27, interleave = window_17_interleave_0, values = (var_455, var_452))[name = tensor("window_17")]; - tensor var_460_begin_0 = const()[name = tensor("op_460_begin_0"), val = tensor([0, 2, 0])]; - tensor var_460_end_0 = const()[name = tensor("op_460_end_0"), val = tensor([1, 3, 256])]; - tensor var_460_end_mask_0 = const()[name = tensor("op_460_end_mask_0"), val = tensor([true, false, true])]; - tensor var_460 = slice_by_index(begin = var_460_begin_0, end = var_460_end_0, end_mask = var_460_end_mask_0, x = x_9)[name = tensor("op_460")]; - tensor var_463_begin_0 = const()[name = tensor("op_463_begin_0"), val = tensor([0, 1, 0])]; - tensor var_463_end_0 = const()[name = tensor("op_463_end_0"), val = tensor([1, 16, 256])]; - tensor var_463_end_mask_0 = const()[name = tensor("op_463_end_mask_0"), val = tensor([true, true, true])]; - tensor var_463 = slice_by_index(begin = var_463_begin_0, end = var_463_end_0, end_mask = var_463_end_mask_0, x = window_17)[name = tensor("op_463")]; + tensor window_17 = concat(axis = var_92, interleave = window_17_interleave_0, values = (var_520, var_517))[name = tensor("window_17")]; + tensor var_525_begin_0 = const()[name = tensor("op_525_begin_0"), val = tensor([0, 2, 0])]; + tensor var_525_end_0 = const()[name = tensor("op_525_end_0"), val = tensor([1, 3, 256])]; + tensor var_525_end_mask_0 = const()[name = tensor("op_525_end_mask_0"), val = tensor([true, false, true])]; + tensor var_525 = slice_by_index(begin = var_525_begin_0, end = var_525_end_0, end_mask = var_525_end_mask_0, x = x_9)[name = tensor("op_525")]; + tensor var_528_begin_0 = const()[name = tensor("op_528_begin_0"), val = tensor([0, 1, 0])]; + tensor var_528_end_0 = const()[name = tensor("op_528_end_0"), val = tensor([1, 16, 256])]; + tensor var_528_end_mask_0 = const()[name = tensor("op_528_end_mask_0"), val = tensor([true, true, true])]; + tensor var_528 = slice_by_index(begin = var_528_begin_0, end = var_528_end_0, end_mask = var_528_end_mask_0, x = window_17)[name = tensor("op_528")]; tensor window_19_interleave_0 = const()[name = tensor("window_19_interleave_0"), val = tensor(false)]; - tensor window_19 = concat(axis = var_27, interleave = window_19_interleave_0, values = (var_463, var_460))[name = tensor("window_19")]; - tensor var_468_begin_0 = const()[name = tensor("op_468_begin_0"), val = tensor([0, 3, 0])]; - tensor var_468_end_0 = const()[name = tensor("op_468_end_0"), val = tensor([1, 4, 256])]; - tensor var_468_end_mask_0 = const()[name = tensor("op_468_end_mask_0"), val = tensor([true, false, true])]; - tensor var_468 = slice_by_index(begin = var_468_begin_0, end = var_468_end_0, end_mask = var_468_end_mask_0, x = x_9)[name = tensor("op_468")]; - tensor var_471_begin_0 = const()[name = tensor("op_471_begin_0"), val = tensor([0, 1, 0])]; - tensor var_471_end_0 = const()[name = tensor("op_471_end_0"), val = tensor([1, 16, 256])]; - tensor var_471_end_mask_0 = const()[name = tensor("op_471_end_mask_0"), val = tensor([true, true, true])]; - tensor var_471 = slice_by_index(begin = var_471_begin_0, end = var_471_end_0, end_mask = var_471_end_mask_0, x = window_19)[name = tensor("op_471")]; + tensor window_19 = concat(axis = var_92, interleave = window_19_interleave_0, values = (var_528, var_525))[name = tensor("window_19")]; + tensor var_533_begin_0 = const()[name = tensor("op_533_begin_0"), val = tensor([0, 3, 0])]; + tensor var_533_end_0 = const()[name = tensor("op_533_end_0"), val = tensor([1, 4, 256])]; + tensor var_533_end_mask_0 = const()[name = tensor("op_533_end_mask_0"), val = tensor([true, false, true])]; + tensor var_533 = slice_by_index(begin = var_533_begin_0, end = var_533_end_0, end_mask = var_533_end_mask_0, x = x_9)[name = tensor("op_533")]; + tensor var_536_begin_0 = const()[name = tensor("op_536_begin_0"), val = tensor([0, 1, 0])]; + tensor var_536_end_0 = const()[name = tensor("op_536_end_0"), val = tensor([1, 16, 256])]; + tensor var_536_end_mask_0 = const()[name = tensor("op_536_end_mask_0"), val = tensor([true, true, true])]; + tensor var_536 = slice_by_index(begin = var_536_begin_0, end = var_536_end_0, end_mask = var_536_end_mask_0, x = window_19)[name = tensor("op_536")]; tensor window_21_interleave_0 = const()[name = tensor("window_21_interleave_0"), val = tensor(false)]; - tensor window_21 = concat(axis = var_27, interleave = window_21_interleave_0, values = (var_471, var_468))[name = tensor("window_21")]; - tensor var_476_begin_0 = const()[name = tensor("op_476_begin_0"), val = tensor([0, 4, 0])]; - tensor var_476_end_0 = const()[name = tensor("op_476_end_0"), val = tensor([1, 1, 256])]; - tensor var_476_end_mask_0 = const()[name = tensor("op_476_end_mask_0"), val = tensor([true, true, true])]; - tensor var_476 = slice_by_index(begin = var_476_begin_0, end = var_476_end_0, end_mask = var_476_end_mask_0, x = x_9)[name = tensor("op_476")]; - tensor var_479_begin_0 = const()[name = tensor("op_479_begin_0"), val = tensor([0, 1, 0])]; - tensor var_479_end_0 = const()[name = tensor("op_479_end_0"), val = tensor([1, 16, 256])]; - tensor var_479_end_mask_0 = const()[name = tensor("op_479_end_mask_0"), val = tensor([true, true, true])]; - tensor var_479 = slice_by_index(begin = var_479_begin_0, end = var_479_end_0, end_mask = var_479_end_mask_0, x = window_21)[name = tensor("op_479")]; + tensor window_21 = concat(axis = var_92, interleave = window_21_interleave_0, values = (var_536, var_533))[name = tensor("window_21")]; + tensor var_541_begin_0 = const()[name = tensor("op_541_begin_0"), val = tensor([0, 4, 0])]; + tensor var_541_end_0 = const()[name = tensor("op_541_end_0"), val = tensor([1, 1, 256])]; + tensor var_541_end_mask_0 = const()[name = tensor("op_541_end_mask_0"), val = tensor([true, true, true])]; + tensor var_541 = slice_by_index(begin = var_541_begin_0, end = var_541_end_0, end_mask = var_541_end_mask_0, x = x_9)[name = tensor("op_541")]; + tensor var_544_begin_0 = const()[name = tensor("op_544_begin_0"), val = tensor([0, 1, 0])]; + tensor var_544_end_0 = const()[name = tensor("op_544_end_0"), val = tensor([1, 16, 256])]; + tensor var_544_end_mask_0 = const()[name = tensor("op_544_end_mask_0"), val = tensor([true, true, true])]; + tensor var_544 = slice_by_index(begin = var_544_begin_0, end = var_544_end_0, end_mask = var_544_end_mask_0, x = window_21)[name = tensor("op_544")]; tensor window_23_interleave_0 = const()[name = tensor("window_23_interleave_0"), val = tensor(false)]; - tensor window_23 = concat(axis = var_27, interleave = window_23_interleave_0, values = (var_479, var_476))[name = tensor("window_23")]; - tensor input_61_interleave_0 = const()[name = tensor("input_61_interleave_0"), val = tensor(false)]; - tensor input_61 = concat(axis = var_24, interleave = input_61_interleave_0, values = (window_15, window_17, window_19, window_21, window_23))[name = tensor("input_61")]; + tensor window_23 = concat(axis = var_92, interleave = window_23_interleave_0, values = (var_544, var_541))[name = tensor("window_23")]; + tensor input_63_interleave_0 = const()[name = tensor("input_63_interleave_0"), val = tensor(false)]; + tensor input_63 = concat(axis = var_77, interleave = input_63_interleave_0, values = (window_15, window_17, window_19, window_21, window_23))[name = tensor("input_63")]; tensor x_11_axes_0 = const()[name = tensor("x_11_axes_0"), val = tensor([-1])]; - tensor x_11 = layer_norm(axes = x_11_axes_0, beta = encoder_conv_module_1_sequential_0_bias, epsilon = var_29, gamma = encoder_conv_module_1_sequential_0_weight, x = input_61)[name = tensor("x_11")]; - tensor input_63_perm_0 = const()[name = tensor("input_63_perm_0"), val = tensor([0, 2, 1])]; + tensor x_11 = layer_norm(axes = x_11_axes_0, beta = inner_encoder_conv_module_1_sequential_0_bias, epsilon = var_74, gamma = inner_encoder_conv_module_1_sequential_0_weight, x = input_63)[name = tensor("x_11")]; + tensor input_65_perm_0 = const()[name = tensor("input_65_perm_0"), val = tensor([0, 2, 1])]; tensor inputs_13_pad_type_0 = const()[name = tensor("inputs_13_pad_type_0"), val = tensor("valid")]; tensor inputs_13_strides_0 = const()[name = tensor("inputs_13_strides_0"), val = tensor([1])]; tensor inputs_13_pad_0 = const()[name = tensor("inputs_13_pad_0"), val = tensor([0, 0])]; tensor inputs_13_dilations_0 = const()[name = tensor("inputs_13_dilations_0"), val = tensor([1])]; tensor inputs_13_groups_0 = const()[name = tensor("inputs_13_groups_0"), val = tensor(1)]; - tensor input_63 = transpose(perm = input_63_perm_0, x = x_11)[name = tensor("transpose_47")]; - tensor inputs_13 = conv(bias = encoder_conv_module_1_sequential_2_conv_bias, dilations = inputs_13_dilations_0, groups = inputs_13_groups_0, pad = inputs_13_pad_0, pad_type = inputs_13_pad_type_0, strides = inputs_13_strides_0, weight = encoder_conv_module_1_sequential_2_conv_weight, x = input_63)[name = tensor("inputs_13")]; - tensor var_504_split_sizes_0 = const()[name = tensor("op_504_split_sizes_0"), val = tensor([256, 256])]; - tensor var_504_axis_0 = const()[name = tensor("op_504_axis_0"), val = tensor(1)]; - tensor var_504_0, tensor var_504_1 = split(axis = var_504_axis_0, split_sizes = var_504_split_sizes_0, x = inputs_13)[name = tensor("op_504")]; - tensor var_506 = sigmoid(x = var_504_1)[name = tensor("op_506")]; - tensor inputs_15 = mul(x = var_504_0, y = var_506)[name = tensor("inputs_15")]; + tensor input_65 = transpose(perm = input_65_perm_0, x = x_11)[name = tensor("transpose_47")]; + tensor inputs_13 = conv(bias = inner_encoder_conv_module_1_sequential_2_conv_bias, dilations = inputs_13_dilations_0, groups = inputs_13_groups_0, pad = inputs_13_pad_0, pad_type = inputs_13_pad_type_0, strides = inputs_13_strides_0, weight = inner_encoder_conv_module_1_sequential_2_conv_weight, x = input_65)[name = tensor("inputs_13")]; + tensor var_569_split_sizes_0 = const()[name = tensor("op_569_split_sizes_0"), val = tensor([256, 256])]; + tensor var_569_axis_0 = const()[name = tensor("op_569_axis_0"), val = tensor(1)]; + tensor var_569_0, tensor var_569_1 = split(axis = var_569_axis_0, split_sizes = var_569_split_sizes_0, x = inputs_13)[name = tensor("op_569")]; + tensor var_571 = sigmoid(x = var_569_1)[name = tensor("op_571")]; + tensor inputs_15 = mul(x = var_569_0, y = var_571)[name = tensor("inputs_15")]; tensor outputs_aug_3_pad_type_0 = const()[name = tensor("outputs_aug_3_pad_type_0"), val = tensor("custom")]; tensor outputs_aug_3_pad_0 = const()[name = tensor("outputs_aug_3_pad_0"), val = tensor([15, 15])]; tensor outputs_aug_3_groups_0 = const()[name = tensor("outputs_aug_3_groups_0"), val = tensor(256)]; tensor outputs_aug_3_strides_0 = const()[name = tensor("outputs_aug_3_strides_0"), val = tensor([1])]; tensor outputs_aug_3_dilations_0 = const()[name = tensor("outputs_aug_3_dilations_0"), val = tensor([1])]; - tensor outputs_aug_3 = conv(dilations = outputs_aug_3_dilations_0, groups = outputs_aug_3_groups_0, pad = outputs_aug_3_pad_0, pad_type = outputs_aug_3_pad_type_0, strides = outputs_aug_3_strides_0, weight = encoder_conv_module_1_sequential_4_conv_weight, x = inputs_15)[name = tensor("outputs_aug_3")]; - tensor input_65_begin_0 = const()[name = tensor("input_65_begin_0"), val = tensor([0, 0, 0])]; - tensor input_65_end_0 = const()[name = tensor("input_65_end_0"), val = tensor([5, 256, 16])]; - tensor input_65_end_mask_0 = const()[name = tensor("input_65_end_mask_0"), val = tensor([true, true, false])]; - tensor input_65 = slice_by_index(begin = input_65_begin_0, end = input_65_end_0, end_mask = input_65_end_mask_0, x = outputs_aug_3)[name = tensor("input_65")]; - tensor inputs_17 = batch_norm(beta = encoder_conv_module_1_sequential_5_bias, epsilon = var_29, gamma = encoder_conv_module_1_sequential_5_weight, mean = encoder_conv_module_1_sequential_5_running_mean, variance = encoder_conv_module_1_sequential_5_running_var, x = input_65)[name = tensor("inputs_17")]; - tensor input_67 = silu(x = inputs_17)[name = tensor("input_67")]; - tensor input_69_pad_type_0 = const()[name = tensor("input_69_pad_type_0"), val = tensor("valid")]; - tensor input_69_strides_0 = const()[name = tensor("input_69_strides_0"), val = tensor([1])]; - tensor input_69_pad_0 = const()[name = tensor("input_69_pad_0"), val = tensor([0, 0])]; - tensor input_69_dilations_0 = const()[name = tensor("input_69_dilations_0"), val = tensor([1])]; - tensor input_69_groups_0 = const()[name = tensor("input_69_groups_0"), val = tensor(1)]; - tensor input_69 = conv(bias = encoder_conv_module_1_sequential_7_conv_bias, dilations = input_69_dilations_0, groups = input_69_groups_0, pad = input_69_pad_0, pad_type = input_69_pad_type_0, strides = input_69_strides_0, weight = encoder_conv_module_1_sequential_7_conv_weight, x = input_67)[name = tensor("input_69")]; + tensor outputs_aug_3 = conv(dilations = outputs_aug_3_dilations_0, groups = outputs_aug_3_groups_0, pad = outputs_aug_3_pad_0, pad_type = outputs_aug_3_pad_type_0, strides = outputs_aug_3_strides_0, weight = inner_encoder_conv_module_1_sequential_4_conv_weight, x = inputs_15)[name = tensor("outputs_aug_3")]; + tensor input_67_begin_0 = const()[name = tensor("input_67_begin_0"), val = tensor([0, 0, 0])]; + tensor input_67_end_0 = const()[name = tensor("input_67_end_0"), val = tensor([5, 256, 16])]; + tensor input_67_end_mask_0 = const()[name = tensor("input_67_end_mask_0"), val = tensor([true, true, false])]; + tensor input_67 = slice_by_index(begin = input_67_begin_0, end = input_67_end_0, end_mask = input_67_end_mask_0, x = outputs_aug_3)[name = tensor("input_67")]; + tensor inputs_17 = batch_norm(beta = inner_encoder_conv_module_1_sequential_5_bias, epsilon = var_74, gamma = inner_encoder_conv_module_1_sequential_5_weight, mean = inner_encoder_conv_module_1_sequential_5_running_mean, variance = inner_encoder_conv_module_1_sequential_5_running_var, x = input_67)[name = tensor("inputs_17")]; + tensor input_69 = silu(x = inputs_17)[name = tensor("input_69")]; + tensor input_71_pad_type_0 = const()[name = tensor("input_71_pad_type_0"), val = tensor("valid")]; + tensor input_71_strides_0 = const()[name = tensor("input_71_strides_0"), val = tensor([1])]; + tensor input_71_pad_0 = const()[name = tensor("input_71_pad_0"), val = tensor([0, 0])]; + tensor input_71_dilations_0 = const()[name = tensor("input_71_dilations_0"), val = tensor([1])]; + tensor input_71_groups_0 = const()[name = tensor("input_71_groups_0"), val = tensor(1)]; + tensor input_71 = conv(bias = inner_encoder_conv_module_1_sequential_7_conv_bias, dilations = input_71_dilations_0, groups = input_71_groups_0, pad = input_71_pad_0, pad_type = input_71_pad_type_0, strides = input_71_strides_0, weight = inner_encoder_conv_module_1_sequential_7_conv_weight, x = input_69)[name = tensor("input_71")]; tensor conv_out_3_perm_0 = const()[name = tensor("conv_out_3_perm_0"), val = tensor([0, 2, 1])]; - tensor var_537_begin_0 = const()[name = tensor("op_537_begin_0"), val = tensor([0, -1, 0])]; - tensor var_537_end_0 = const()[name = tensor("op_537_end_0"), val = tensor([5, 16, 256])]; - tensor var_537_end_mask_0 = const()[name = tensor("op_537_end_mask_0"), val = tensor([true, true, true])]; - tensor conv_out_3 = transpose(perm = conv_out_3_perm_0, x = input_69)[name = tensor("transpose_46")]; - tensor var_537 = slice_by_index(begin = var_537_begin_0, end = var_537_end_0, end_mask = var_537_end_mask_0, x = conv_out_3)[name = tensor("op_537")]; - tensor var_539_perm_0 = const()[name = tensor("op_539_perm_0"), val = tensor([1, 0, 2])]; - tensor var_539 = transpose(perm = var_539_perm_0, x = var_537)[name = tensor("transpose_45")]; - tensor input_71 = add(x = x_9, y = var_539)[name = tensor("input_71")]; - tensor input_73_axes_0 = const()[name = tensor("input_73_axes_0"), val = tensor([-1])]; - tensor input_73 = layer_norm(axes = input_73_axes_0, beta = encoder_ffn2_1_module_sequential_0_bias, epsilon = var_29, gamma = encoder_ffn2_1_module_sequential_0_weight, x = input_71)[name = tensor("input_73")]; - tensor inputs_19 = linear(bias = encoder_ffn2_1_module_sequential_1_linear_bias, weight = encoder_ffn2_1_module_sequential_1_linear_weight, x = input_73)[name = tensor("linear_17")]; - tensor input_75 = silu(x = inputs_19)[name = tensor("input_75")]; - tensor input_79 = linear(bias = encoder_ffn2_1_module_sequential_4_linear_bias, weight = encoder_ffn2_1_module_sequential_4_linear_weight, x = input_75)[name = tensor("linear_18")]; - tensor var_562 = const()[name = tensor("op_562"), val = tensor(0x1p-1)]; - tensor var_563 = mul(x = input_79, y = var_562)[name = tensor("op_563")]; - tensor input_81 = add(x = var_563, y = input_71)[name = tensor("input_81")]; - tensor input_83_axes_0 = const()[name = tensor("input_83_axes_0"), val = tensor([-1])]; - tensor input_83 = layer_norm(axes = input_83_axes_0, beta = encoder_layer_norm_1_bias, epsilon = var_29, gamma = encoder_layer_norm_1_weight, x = input_81)[name = tensor("input_83")]; + tensor var_602_begin_0 = const()[name = tensor("op_602_begin_0"), val = tensor([0, -1, 0])]; + tensor var_602_end_0 = const()[name = tensor("op_602_end_0"), val = tensor([5, 16, 256])]; + tensor var_602_end_mask_0 = const()[name = tensor("op_602_end_mask_0"), val = tensor([true, true, true])]; + tensor conv_out_3 = transpose(perm = conv_out_3_perm_0, x = input_71)[name = tensor("transpose_46")]; + tensor var_602 = slice_by_index(begin = var_602_begin_0, end = var_602_end_0, end_mask = var_602_end_mask_0, x = conv_out_3)[name = tensor("op_602")]; + tensor var_604_perm_0 = const()[name = tensor("op_604_perm_0"), val = tensor([1, 0, 2])]; + tensor var_604 = transpose(perm = var_604_perm_0, x = var_602)[name = tensor("transpose_45")]; + tensor input_73 = add(x = x_9, y = var_604)[name = tensor("input_73")]; + tensor input_75_axes_0 = const()[name = tensor("input_75_axes_0"), val = tensor([-1])]; + tensor input_75 = layer_norm(axes = input_75_axes_0, beta = inner_encoder_ffn2_1_module_sequential_0_bias, epsilon = var_74, gamma = inner_encoder_ffn2_1_module_sequential_0_weight, x = input_73)[name = tensor("input_75")]; + tensor inputs_19 = linear(bias = inner_encoder_ffn2_1_module_sequential_1_linear_bias, weight = inner_encoder_ffn2_1_module_sequential_1_linear_weight, x = input_75)[name = tensor("linear_17")]; + tensor input_77 = silu(x = inputs_19)[name = tensor("input_77")]; + tensor input_81 = linear(bias = inner_encoder_ffn2_1_module_sequential_4_linear_bias, weight = inner_encoder_ffn2_1_module_sequential_4_linear_weight, x = input_77)[name = tensor("linear_18")]; + tensor var_627 = const()[name = tensor("op_627"), val = tensor(0x1p-1)]; + tensor var_628 = mul(x = input_81, y = var_627)[name = tensor("op_628")]; + tensor input_83 = add(x = var_628, y = input_73)[name = tensor("input_83")]; tensor input_85_axes_0 = const()[name = tensor("input_85_axes_0"), val = tensor([-1])]; - tensor input_85 = layer_norm(axes = input_85_axes_0, beta = encoder_ffn1_2_module_sequential_0_bias, epsilon = var_29, gamma = encoder_ffn1_2_module_sequential_0_weight, x = input_83)[name = tensor("input_85")]; - tensor inputs_21 = linear(bias = encoder_ffn1_2_module_sequential_1_linear_bias, weight = encoder_ffn1_2_module_sequential_1_linear_weight, x = input_85)[name = tensor("linear_19")]; - tensor input_87 = silu(x = inputs_21)[name = tensor("input_87")]; - tensor input_91 = linear(bias = encoder_ffn1_2_module_sequential_4_linear_bias, weight = encoder_ffn1_2_module_sequential_4_linear_weight, x = input_87)[name = tensor("linear_20")]; - tensor var_592 = const()[name = tensor("op_592"), val = tensor(0x1p-1)]; - tensor var_593 = mul(x = input_91, y = var_592)[name = tensor("op_593")]; - tensor input_93 = add(x = var_593, y = input_83)[name = tensor("input_93")]; + tensor input_85 = layer_norm(axes = input_85_axes_0, beta = inner_encoder_layer_norm_1_bias, epsilon = var_74, gamma = inner_encoder_layer_norm_1_weight, x = input_83)[name = tensor("input_85")]; + tensor input_87_axes_0 = const()[name = tensor("input_87_axes_0"), val = tensor([-1])]; + tensor input_87 = layer_norm(axes = input_87_axes_0, beta = inner_encoder_ffn1_2_module_sequential_0_bias, epsilon = var_74, gamma = inner_encoder_ffn1_2_module_sequential_0_weight, x = input_85)[name = tensor("input_87")]; + tensor inputs_21 = linear(bias = inner_encoder_ffn1_2_module_sequential_1_linear_bias, weight = inner_encoder_ffn1_2_module_sequential_1_linear_weight, x = input_87)[name = tensor("linear_19")]; + tensor input_89 = silu(x = inputs_21)[name = tensor("input_89")]; + tensor input_93 = linear(bias = inner_encoder_ffn1_2_module_sequential_4_linear_bias, weight = inner_encoder_ffn1_2_module_sequential_4_linear_weight, x = input_89)[name = tensor("linear_20")]; + tensor var_657 = const()[name = tensor("op_657"), val = tensor(0x1p-1)]; + tensor var_658 = mul(x = input_93, y = var_657)[name = tensor("op_658")]; + tensor input_95 = add(x = var_658, y = input_85)[name = tensor("input_95")]; tensor x_13_axes_0 = const()[name = tensor("x_13_axes_0"), val = tensor([-1])]; - tensor x_13 = layer_norm(axes = x_13_axes_0, beta = encoder_ret_lns_2_bias, epsilon = var_29, gamma = encoder_ret_lns_2_weight, x = input_93)[name = tensor("x_13")]; + tensor x_13 = layer_norm(axes = x_13_axes_0, beta = inner_encoder_ret_lns_2_bias, epsilon = var_74, gamma = inner_encoder_ret_lns_2_weight, x = input_95)[name = tensor("x_13")]; tensor prev_kv_5_begin_0 = const()[name = tensor("prev_kv_5_begin_0"), val = tensor([2, 0, 0, 0, 0])]; tensor prev_kv_5_end_0 = const()[name = tensor("prev_kv_5_end_0"), val = tensor([3, 1, 4, 64, 64])]; tensor prev_kv_5_end_mask_0 = const()[name = tensor("prev_kv_5_end_mask_0"), val = tensor([false, true, true, true, true])]; @@ -613,183 +639,183 @@ program(1.0) tensor prev_scale_5_end_mask_0 = const()[name = tensor("prev_scale_5_end_mask_0"), val = tensor([false, true])]; tensor prev_scale_5_squeeze_mask_0 = const()[name = tensor("prev_scale_5_squeeze_mask_0"), val = tensor([true, false])]; tensor prev_scale_5 = slice_by_index(begin = prev_scale_5_begin_0, end = prev_scale_5_end_0, end_mask = prev_scale_5_end_mask_0, squeeze_mask = prev_scale_5_squeeze_mask_0, x = enc_scale)[name = tensor("prev_scale_5")]; - tensor var_607 = linear(bias = encoder_q_proj_2_bias, weight = encoder_q_proj_2_weight, x = x_13)[name = tensor("linear_21")]; - tensor var_608 = const()[name = tensor("op_608"), val = tensor([1, 5, 4, 64])]; - tensor var_609 = reshape(shape = var_608, x = var_607)[name = tensor("op_609")]; + tensor var_672 = linear(bias = inner_encoder_q_proj_2_bias, weight = inner_encoder_q_proj_2_weight, x = x_13)[name = tensor("linear_21")]; + tensor var_673 = const()[name = tensor("op_673"), val = tensor([1, 5, 4, 64])]; + tensor var_674 = reshape(shape = var_673, x = var_672)[name = tensor("op_674")]; tensor q_5_perm_0 = const()[name = tensor("q_5_perm_0"), val = tensor([0, 2, 1, 3])]; - tensor var_613 = linear(bias = encoder_k_proj_2_bias, weight = encoder_k_proj_2_weight, x = x_13)[name = tensor("linear_22")]; - tensor var_614 = const()[name = tensor("op_614"), val = tensor(0x1p-3)]; - tensor var_615 = mul(x = var_613, y = var_614)[name = tensor("op_615")]; - tensor var_616 = const()[name = tensor("op_616"), val = tensor([1, 5, 4, 64])]; - tensor var_617 = reshape(shape = var_616, x = var_615)[name = tensor("op_617")]; + tensor var_678 = linear(bias = inner_encoder_k_proj_2_bias, weight = inner_encoder_k_proj_2_weight, x = x_13)[name = tensor("linear_22")]; + tensor var_679 = const()[name = tensor("op_679"), val = tensor(0x1p-3)]; + tensor var_680 = mul(x = var_678, y = var_679)[name = tensor("op_680")]; + tensor var_681 = const()[name = tensor("op_681"), val = tensor([1, 5, 4, 64])]; + tensor var_682 = reshape(shape = var_681, x = var_680)[name = tensor("op_682")]; tensor k_5_perm_0 = const()[name = tensor("k_5_perm_0"), val = tensor([0, 2, 1, 3])]; - tensor var_621 = linear(bias = encoder_v_proj_2_bias, weight = encoder_v_proj_2_weight, x = x_13)[name = tensor("linear_23")]; - tensor var_622 = const()[name = tensor("op_622"), val = tensor([1, 5, 4, 64])]; - tensor var_623 = reshape(shape = var_622, x = var_621)[name = tensor("op_623")]; + tensor var_686 = linear(bias = inner_encoder_v_proj_2_bias, weight = inner_encoder_v_proj_2_weight, x = x_13)[name = tensor("linear_23")]; + tensor var_687 = const()[name = tensor("op_687"), val = tensor([1, 5, 4, 64])]; + tensor var_688 = reshape(shape = var_687, x = var_686)[name = tensor("op_688")]; tensor v_5_perm_0 = const()[name = tensor("v_5_perm_0"), val = tensor([0, 2, 1, 3])]; - tensor input_97 = linear(bias = encoder_g_proj_2_bias, weight = encoder_g_proj_2_weight, x = x_13)[name = tensor("linear_24")]; + tensor input_99 = linear(bias = inner_encoder_g_proj_2_bias, weight = inner_encoder_g_proj_2_weight, x = x_13)[name = tensor("linear_24")]; tensor sqrt_s0_5 = sqrt(x = prev_scale_5)[name = tensor("sqrt_s0_5")]; - tensor s_t_5 = add(x = prev_scale_5, y = encoder__t_index)[name = tensor("s_t_5")]; + tensor s_t_5 = add(x = prev_scale_5, y = inner_encoder__t_index)[name = tensor("s_t_5")]; tensor sqrt_s_t_5 = sqrt(x = s_t_5)[name = tensor("sqrt_s_t_5")]; tensor qk_5_transpose_x_1 = const()[name = tensor("qk_5_transpose_x_1"), val = tensor(false)]; tensor qk_5_transpose_y_1 = const()[name = tensor("qk_5_transpose_y_1"), val = tensor(true)]; - tensor k_5 = transpose(perm = k_5_perm_0, x = var_617)[name = tensor("transpose_43")]; - tensor q_5 = transpose(perm = q_5_perm_0, x = var_609)[name = tensor("transpose_44")]; + tensor k_5 = transpose(perm = k_5_perm_0, x = var_682)[name = tensor("transpose_43")]; + tensor q_5 = transpose(perm = q_5_perm_0, x = var_674)[name = tensor("transpose_44")]; tensor qk_5 = matmul(transpose_x = qk_5_transpose_x_1, transpose_y = qk_5_transpose_y_1, x = q_5, y = k_5)[name = tensor("qk_5")]; - tensor var_633 = const()[name = tensor("op_633"), val = tensor([5, 1])]; - tensor var_634 = reshape(shape = var_633, x = sqrt_s_t_5)[name = tensor("op_634")]; - tensor M_5 = real_div(x = encoder__causal_mask, y = var_634)[name = tensor("M_5")]; - tensor var_636 = mul(x = qk_5, y = M_5)[name = tensor("op_636")]; + tensor var_698 = const()[name = tensor("op_698"), val = tensor([5, 1])]; + tensor var_699 = reshape(shape = var_698, x = sqrt_s_t_5)[name = tensor("op_699")]; + tensor M_5 = real_div(x = inner_encoder__causal_mask, y = var_699)[name = tensor("M_5")]; + tensor var_701 = mul(x = qk_5, y = M_5)[name = tensor("op_701")]; tensor inner_5_transpose_x_0 = const()[name = tensor("inner_5_transpose_x_0"), val = tensor(false)]; tensor inner_5_transpose_y_0 = const()[name = tensor("inner_5_transpose_y_0"), val = tensor(false)]; - tensor v_5 = transpose(perm = v_5_perm_0, x = var_623)[name = tensor("transpose_42")]; - tensor inner_5 = matmul(transpose_x = inner_5_transpose_x_0, transpose_y = inner_5_transpose_y_0, x = var_636, y = v_5)[name = tensor("inner_5")]; - tensor var_638_transpose_x_0 = const()[name = tensor("op_638_transpose_x_0"), val = tensor(false)]; - tensor var_638_transpose_y_0 = const()[name = tensor("op_638_transpose_y_0"), val = tensor(false)]; - tensor var_638 = matmul(transpose_x = var_638_transpose_x_0, transpose_y = var_638_transpose_y_0, x = q_5, y = prev_kv_5)[name = tensor("op_638")]; - tensor var_639 = real_div(x = sqrt_s0_5, y = sqrt_s_t_5)[name = tensor("op_639")]; - tensor var_640 = const()[name = tensor("op_640"), val = tensor([1, 1, 5, 1])]; - tensor var_641 = reshape(shape = var_640, x = var_639)[name = tensor("op_641")]; - tensor cross_5 = mul(x = var_638, y = var_641)[name = tensor("cross_5")]; + tensor v_5 = transpose(perm = v_5_perm_0, x = var_688)[name = tensor("transpose_42")]; + tensor inner_5 = matmul(transpose_x = inner_5_transpose_x_0, transpose_y = inner_5_transpose_y_0, x = var_701, y = v_5)[name = tensor("inner_5")]; + tensor var_703_transpose_x_0 = const()[name = tensor("op_703_transpose_x_0"), val = tensor(false)]; + tensor var_703_transpose_y_0 = const()[name = tensor("op_703_transpose_y_0"), val = tensor(false)]; + tensor var_703 = matmul(transpose_x = var_703_transpose_x_0, transpose_y = var_703_transpose_y_0, x = q_5, y = prev_kv_5)[name = tensor("op_703")]; + tensor var_704 = real_div(x = sqrt_s0_5, y = sqrt_s_t_5)[name = tensor("op_704")]; + tensor var_705 = const()[name = tensor("op_705"), val = tensor([1, 1, 5, 1])]; + tensor var_706 = reshape(shape = var_705, x = var_704)[name = tensor("op_706")]; + tensor cross_5 = mul(x = var_703, y = var_706)[name = tensor("cross_5")]; tensor out_13 = add(x = inner_5, y = cross_5)[name = tensor("out_13")]; - tensor var_644 = mul(x = prev_kv_5, y = sqrt_s0_5)[name = tensor("op_644")]; - tensor var_646_transpose_x_1 = const()[name = tensor("op_646_transpose_x_1"), val = tensor(true)]; - tensor var_646_transpose_y_1 = const()[name = tensor("op_646_transpose_y_1"), val = tensor(false)]; - tensor var_646 = matmul(transpose_x = var_646_transpose_x_1, transpose_y = var_646_transpose_y_1, x = k_5, y = v_5)[name = tensor("op_646")]; - tensor new_kv_unnorm_5 = add(x = var_644, y = var_646)[name = tensor("new_kv_unnorm_5")]; - tensor var_648 = const()[name = tensor("op_648"), val = tensor(0x1.4p+2)]; - tensor new_scale_5 = add(x = prev_scale_5, y = var_648)[name = tensor("new_scale_5")]; - tensor var_650 = sqrt(x = new_scale_5)[name = tensor("op_650")]; - tensor var_651 = real_div(x = new_kv_unnorm_5, y = var_650)[name = tensor("op_651")]; - tensor var_652_perm_0 = const()[name = tensor("op_652_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_709 = mul(x = prev_kv_5, y = sqrt_s0_5)[name = tensor("op_709")]; + tensor var_711_transpose_x_1 = const()[name = tensor("op_711_transpose_x_1"), val = tensor(true)]; + tensor var_711_transpose_y_1 = const()[name = tensor("op_711_transpose_y_1"), val = tensor(false)]; + tensor var_711 = matmul(transpose_x = var_711_transpose_x_1, transpose_y = var_711_transpose_y_1, x = k_5, y = v_5)[name = tensor("op_711")]; + tensor new_kv_unnorm_5 = add(x = var_709, y = var_711)[name = tensor("new_kv_unnorm_5")]; + tensor var_713 = const()[name = tensor("op_713"), val = tensor(0x1.4p+2)]; + tensor new_scale_5 = add(x = prev_scale_5, y = var_713)[name = tensor("new_scale_5")]; + tensor var_715 = sqrt(x = new_scale_5)[name = tensor("op_715")]; + tensor var_716 = real_div(x = new_kv_unnorm_5, y = var_715)[name = tensor("op_716")]; + tensor var_717_perm_0 = const()[name = tensor("op_717_perm_0"), val = tensor([0, 2, 1, 3])]; tensor out_15_axes_0 = const()[name = tensor("out_15_axes_0"), val = tensor([-1])]; - tensor var_652 = transpose(perm = var_652_perm_0, x = out_13)[name = tensor("transpose_41")]; - tensor out_15 = layer_norm(axes = out_15_axes_0, epsilon = var_18, x = var_652)[name = tensor("out_15")]; - tensor var_656 = const()[name = tensor("op_656"), val = tensor([1, 5, 256])]; - tensor out_17 = reshape(shape = var_656, x = out_15)[name = tensor("out_17")]; - tensor var_658 = silu(x = input_97)[name = tensor("op_658")]; - tensor input_99 = mul(x = var_658, y = out_17)[name = tensor("input_99")]; - tensor ret_out_5 = linear(bias = encoder_out_proj_2_bias, weight = encoder_out_proj_2_weight, x = input_99)[name = tensor("linear_25")]; - tensor x_15 = add(x = input_93, y = ret_out_5)[name = tensor("x_15")]; + tensor var_717 = transpose(perm = var_717_perm_0, x = out_13)[name = tensor("transpose_41")]; + tensor out_15 = layer_norm(axes = out_15_axes_0, epsilon = var_82, x = var_717)[name = tensor("out_15")]; + tensor var_721 = const()[name = tensor("op_721"), val = tensor([1, 5, 256])]; + tensor out_17 = reshape(shape = var_721, x = out_15)[name = tensor("out_17")]; + tensor var_723 = silu(x = input_99)[name = tensor("op_723")]; + tensor input_101 = mul(x = var_723, y = out_17)[name = tensor("input_101")]; + tensor ret_out_5 = linear(bias = inner_encoder_out_proj_2_bias, weight = inner_encoder_out_proj_2_weight, x = input_101)[name = tensor("linear_25")]; + tensor x_15 = add(x = input_95, y = ret_out_5)[name = tensor("x_15")]; tensor window_25_begin_0 = const()[name = tensor("window_25_begin_0"), val = tensor([2, 0, 0, 0])]; tensor window_25_end_0 = const()[name = tensor("window_25_end_0"), val = tensor([3, 1, 16, 256])]; tensor window_25_end_mask_0 = const()[name = tensor("window_25_end_mask_0"), val = tensor([false, true, true, true])]; tensor window_25_squeeze_mask_0 = const()[name = tensor("window_25_squeeze_mask_0"), val = tensor([true, false, false, false])]; tensor window_25 = slice_by_index(begin = window_25_begin_0, end = window_25_end_0, end_mask = window_25_end_mask_0, squeeze_mask = window_25_squeeze_mask_0, x = enc_conv_cache)[name = tensor("window_25")]; - tensor var_666_begin_0 = const()[name = tensor("op_666_begin_0"), val = tensor([0, 0, 0])]; - tensor var_666_end_0 = const()[name = tensor("op_666_end_0"), val = tensor([1, 1, 256])]; - tensor var_666_end_mask_0 = const()[name = tensor("op_666_end_mask_0"), val = tensor([true, false, true])]; - tensor var_666 = slice_by_index(begin = var_666_begin_0, end = var_666_end_0, end_mask = var_666_end_mask_0, x = x_15)[name = tensor("op_666")]; - tensor var_669_begin_0 = const()[name = tensor("op_669_begin_0"), val = tensor([0, 1, 0])]; - tensor var_669_end_0 = const()[name = tensor("op_669_end_0"), val = tensor([1, 16, 256])]; - tensor var_669_end_mask_0 = const()[name = tensor("op_669_end_mask_0"), val = tensor([true, true, true])]; - tensor var_669 = slice_by_index(begin = var_669_begin_0, end = var_669_end_0, end_mask = var_669_end_mask_0, x = window_25)[name = tensor("op_669")]; + tensor var_731_begin_0 = const()[name = tensor("op_731_begin_0"), val = tensor([0, 0, 0])]; + tensor var_731_end_0 = const()[name = tensor("op_731_end_0"), val = tensor([1, 1, 256])]; + tensor var_731_end_mask_0 = const()[name = tensor("op_731_end_mask_0"), val = tensor([true, false, true])]; + tensor var_731 = slice_by_index(begin = var_731_begin_0, end = var_731_end_0, end_mask = var_731_end_mask_0, x = x_15)[name = tensor("op_731")]; + tensor var_734_begin_0 = const()[name = tensor("op_734_begin_0"), val = tensor([0, 1, 0])]; + tensor var_734_end_0 = const()[name = tensor("op_734_end_0"), val = tensor([1, 16, 256])]; + tensor var_734_end_mask_0 = const()[name = tensor("op_734_end_mask_0"), val = tensor([true, true, true])]; + tensor var_734 = slice_by_index(begin = var_734_begin_0, end = var_734_end_0, end_mask = var_734_end_mask_0, x = window_25)[name = tensor("op_734")]; tensor window_27_interleave_0 = const()[name = tensor("window_27_interleave_0"), val = tensor(false)]; - tensor window_27 = concat(axis = var_27, interleave = window_27_interleave_0, values = (var_669, var_666))[name = tensor("window_27")]; - tensor var_674_begin_0 = const()[name = tensor("op_674_begin_0"), val = tensor([0, 1, 0])]; - tensor var_674_end_0 = const()[name = tensor("op_674_end_0"), val = tensor([1, 2, 256])]; - tensor var_674_end_mask_0 = const()[name = tensor("op_674_end_mask_0"), val = tensor([true, false, true])]; - tensor var_674 = slice_by_index(begin = var_674_begin_0, end = var_674_end_0, end_mask = var_674_end_mask_0, x = x_15)[name = tensor("op_674")]; - tensor var_677_begin_0 = const()[name = tensor("op_677_begin_0"), val = tensor([0, 1, 0])]; - tensor var_677_end_0 = const()[name = tensor("op_677_end_0"), val = tensor([1, 16, 256])]; - tensor var_677_end_mask_0 = const()[name = tensor("op_677_end_mask_0"), val = tensor([true, true, true])]; - tensor var_677 = slice_by_index(begin = var_677_begin_0, end = var_677_end_0, end_mask = var_677_end_mask_0, x = window_27)[name = tensor("op_677")]; + tensor window_27 = concat(axis = var_92, interleave = window_27_interleave_0, values = (var_734, var_731))[name = tensor("window_27")]; + tensor var_739_begin_0 = const()[name = tensor("op_739_begin_0"), val = tensor([0, 1, 0])]; + tensor var_739_end_0 = const()[name = tensor("op_739_end_0"), val = tensor([1, 2, 256])]; + tensor var_739_end_mask_0 = const()[name = tensor("op_739_end_mask_0"), val = tensor([true, false, true])]; + tensor var_739 = slice_by_index(begin = var_739_begin_0, end = var_739_end_0, end_mask = var_739_end_mask_0, x = x_15)[name = tensor("op_739")]; + tensor var_742_begin_0 = const()[name = tensor("op_742_begin_0"), val = tensor([0, 1, 0])]; + tensor var_742_end_0 = const()[name = tensor("op_742_end_0"), val = tensor([1, 16, 256])]; + tensor var_742_end_mask_0 = const()[name = tensor("op_742_end_mask_0"), val = tensor([true, true, true])]; + tensor var_742 = slice_by_index(begin = var_742_begin_0, end = var_742_end_0, end_mask = var_742_end_mask_0, x = window_27)[name = tensor("op_742")]; tensor window_29_interleave_0 = const()[name = tensor("window_29_interleave_0"), val = tensor(false)]; - tensor window_29 = concat(axis = var_27, interleave = window_29_interleave_0, values = (var_677, var_674))[name = tensor("window_29")]; - tensor var_682_begin_0 = const()[name = tensor("op_682_begin_0"), val = tensor([0, 2, 0])]; - tensor var_682_end_0 = const()[name = tensor("op_682_end_0"), val = tensor([1, 3, 256])]; - tensor var_682_end_mask_0 = const()[name = tensor("op_682_end_mask_0"), val = tensor([true, false, true])]; - tensor var_682 = slice_by_index(begin = var_682_begin_0, end = var_682_end_0, end_mask = var_682_end_mask_0, x = x_15)[name = tensor("op_682")]; - tensor var_685_begin_0 = const()[name = tensor("op_685_begin_0"), val = tensor([0, 1, 0])]; - tensor var_685_end_0 = const()[name = tensor("op_685_end_0"), val = tensor([1, 16, 256])]; - tensor var_685_end_mask_0 = const()[name = tensor("op_685_end_mask_0"), val = tensor([true, true, true])]; - tensor var_685 = slice_by_index(begin = var_685_begin_0, end = var_685_end_0, end_mask = var_685_end_mask_0, x = window_29)[name = tensor("op_685")]; + tensor window_29 = concat(axis = var_92, interleave = window_29_interleave_0, values = (var_742, var_739))[name = tensor("window_29")]; + tensor var_747_begin_0 = const()[name = tensor("op_747_begin_0"), val = tensor([0, 2, 0])]; + tensor var_747_end_0 = const()[name = tensor("op_747_end_0"), val = tensor([1, 3, 256])]; + tensor var_747_end_mask_0 = const()[name = tensor("op_747_end_mask_0"), val = tensor([true, false, true])]; + tensor var_747 = slice_by_index(begin = var_747_begin_0, end = var_747_end_0, end_mask = var_747_end_mask_0, x = x_15)[name = tensor("op_747")]; + tensor var_750_begin_0 = const()[name = tensor("op_750_begin_0"), val = tensor([0, 1, 0])]; + tensor var_750_end_0 = const()[name = tensor("op_750_end_0"), val = tensor([1, 16, 256])]; + tensor var_750_end_mask_0 = const()[name = tensor("op_750_end_mask_0"), val = tensor([true, true, true])]; + tensor var_750 = slice_by_index(begin = var_750_begin_0, end = var_750_end_0, end_mask = var_750_end_mask_0, x = window_29)[name = tensor("op_750")]; tensor window_31_interleave_0 = const()[name = tensor("window_31_interleave_0"), val = tensor(false)]; - tensor window_31 = concat(axis = var_27, interleave = window_31_interleave_0, values = (var_685, var_682))[name = tensor("window_31")]; - tensor var_690_begin_0 = const()[name = tensor("op_690_begin_0"), val = tensor([0, 3, 0])]; - tensor var_690_end_0 = const()[name = tensor("op_690_end_0"), val = tensor([1, 4, 256])]; - tensor var_690_end_mask_0 = const()[name = tensor("op_690_end_mask_0"), val = tensor([true, false, true])]; - tensor var_690 = slice_by_index(begin = var_690_begin_0, end = var_690_end_0, end_mask = var_690_end_mask_0, x = x_15)[name = tensor("op_690")]; - tensor var_693_begin_0 = const()[name = tensor("op_693_begin_0"), val = tensor([0, 1, 0])]; - tensor var_693_end_0 = const()[name = tensor("op_693_end_0"), val = tensor([1, 16, 256])]; - tensor var_693_end_mask_0 = const()[name = tensor("op_693_end_mask_0"), val = tensor([true, true, true])]; - tensor var_693 = slice_by_index(begin = var_693_begin_0, end = var_693_end_0, end_mask = var_693_end_mask_0, x = window_31)[name = tensor("op_693")]; + tensor window_31 = concat(axis = var_92, interleave = window_31_interleave_0, values = (var_750, var_747))[name = tensor("window_31")]; + tensor var_755_begin_0 = const()[name = tensor("op_755_begin_0"), val = tensor([0, 3, 0])]; + tensor var_755_end_0 = const()[name = tensor("op_755_end_0"), val = tensor([1, 4, 256])]; + tensor var_755_end_mask_0 = const()[name = tensor("op_755_end_mask_0"), val = tensor([true, false, true])]; + tensor var_755 = slice_by_index(begin = var_755_begin_0, end = var_755_end_0, end_mask = var_755_end_mask_0, x = x_15)[name = tensor("op_755")]; + tensor var_758_begin_0 = const()[name = tensor("op_758_begin_0"), val = tensor([0, 1, 0])]; + tensor var_758_end_0 = const()[name = tensor("op_758_end_0"), val = tensor([1, 16, 256])]; + tensor var_758_end_mask_0 = const()[name = tensor("op_758_end_mask_0"), val = tensor([true, true, true])]; + tensor var_758 = slice_by_index(begin = var_758_begin_0, end = var_758_end_0, end_mask = var_758_end_mask_0, x = window_31)[name = tensor("op_758")]; tensor window_33_interleave_0 = const()[name = tensor("window_33_interleave_0"), val = tensor(false)]; - tensor window_33 = concat(axis = var_27, interleave = window_33_interleave_0, values = (var_693, var_690))[name = tensor("window_33")]; - tensor var_698_begin_0 = const()[name = tensor("op_698_begin_0"), val = tensor([0, 4, 0])]; - tensor var_698_end_0 = const()[name = tensor("op_698_end_0"), val = tensor([1, 1, 256])]; - tensor var_698_end_mask_0 = const()[name = tensor("op_698_end_mask_0"), val = tensor([true, true, true])]; - tensor var_698 = slice_by_index(begin = var_698_begin_0, end = var_698_end_0, end_mask = var_698_end_mask_0, x = x_15)[name = tensor("op_698")]; - tensor var_701_begin_0 = const()[name = tensor("op_701_begin_0"), val = tensor([0, 1, 0])]; - tensor var_701_end_0 = const()[name = tensor("op_701_end_0"), val = tensor([1, 16, 256])]; - tensor var_701_end_mask_0 = const()[name = tensor("op_701_end_mask_0"), val = tensor([true, true, true])]; - tensor var_701 = slice_by_index(begin = var_701_begin_0, end = var_701_end_0, end_mask = var_701_end_mask_0, x = window_33)[name = tensor("op_701")]; + tensor window_33 = concat(axis = var_92, interleave = window_33_interleave_0, values = (var_758, var_755))[name = tensor("window_33")]; + tensor var_763_begin_0 = const()[name = tensor("op_763_begin_0"), val = tensor([0, 4, 0])]; + tensor var_763_end_0 = const()[name = tensor("op_763_end_0"), val = tensor([1, 1, 256])]; + tensor var_763_end_mask_0 = const()[name = tensor("op_763_end_mask_0"), val = tensor([true, true, true])]; + tensor var_763 = slice_by_index(begin = var_763_begin_0, end = var_763_end_0, end_mask = var_763_end_mask_0, x = x_15)[name = tensor("op_763")]; + tensor var_766_begin_0 = const()[name = tensor("op_766_begin_0"), val = tensor([0, 1, 0])]; + tensor var_766_end_0 = const()[name = tensor("op_766_end_0"), val = tensor([1, 16, 256])]; + tensor var_766_end_mask_0 = const()[name = tensor("op_766_end_mask_0"), val = tensor([true, true, true])]; + tensor var_766 = slice_by_index(begin = var_766_begin_0, end = var_766_end_0, end_mask = var_766_end_mask_0, x = window_33)[name = tensor("op_766")]; tensor window_35_interleave_0 = const()[name = tensor("window_35_interleave_0"), val = tensor(false)]; - tensor window_35 = concat(axis = var_27, interleave = window_35_interleave_0, values = (var_701, var_698))[name = tensor("window_35")]; - tensor input_101_interleave_0 = const()[name = tensor("input_101_interleave_0"), val = tensor(false)]; - tensor input_101 = concat(axis = var_24, interleave = input_101_interleave_0, values = (window_27, window_29, window_31, window_33, window_35))[name = tensor("input_101")]; + tensor window_35 = concat(axis = var_92, interleave = window_35_interleave_0, values = (var_766, var_763))[name = tensor("window_35")]; + tensor input_103_interleave_0 = const()[name = tensor("input_103_interleave_0"), val = tensor(false)]; + tensor input_103 = concat(axis = var_77, interleave = input_103_interleave_0, values = (window_27, window_29, window_31, window_33, window_35))[name = tensor("input_103")]; tensor x_17_axes_0 = const()[name = tensor("x_17_axes_0"), val = tensor([-1])]; - tensor x_17 = layer_norm(axes = x_17_axes_0, beta = encoder_conv_module_2_sequential_0_bias, epsilon = var_29, gamma = encoder_conv_module_2_sequential_0_weight, x = input_101)[name = tensor("x_17")]; - tensor input_103_perm_0 = const()[name = tensor("input_103_perm_0"), val = tensor([0, 2, 1])]; + tensor x_17 = layer_norm(axes = x_17_axes_0, beta = inner_encoder_conv_module_2_sequential_0_bias, epsilon = var_74, gamma = inner_encoder_conv_module_2_sequential_0_weight, x = input_103)[name = tensor("x_17")]; + tensor input_105_perm_0 = const()[name = tensor("input_105_perm_0"), val = tensor([0, 2, 1])]; tensor inputs_23_pad_type_0 = const()[name = tensor("inputs_23_pad_type_0"), val = tensor("valid")]; tensor inputs_23_strides_0 = const()[name = tensor("inputs_23_strides_0"), val = tensor([1])]; tensor inputs_23_pad_0 = const()[name = tensor("inputs_23_pad_0"), val = tensor([0, 0])]; tensor inputs_23_dilations_0 = const()[name = tensor("inputs_23_dilations_0"), val = tensor([1])]; tensor inputs_23_groups_0 = const()[name = tensor("inputs_23_groups_0"), val = tensor(1)]; - tensor input_103 = transpose(perm = input_103_perm_0, x = x_17)[name = tensor("transpose_40")]; - tensor inputs_23 = conv(bias = encoder_conv_module_2_sequential_2_conv_bias, dilations = inputs_23_dilations_0, groups = inputs_23_groups_0, pad = inputs_23_pad_0, pad_type = inputs_23_pad_type_0, strides = inputs_23_strides_0, weight = encoder_conv_module_2_sequential_2_conv_weight, x = input_103)[name = tensor("inputs_23")]; - tensor var_726_split_sizes_0 = const()[name = tensor("op_726_split_sizes_0"), val = tensor([256, 256])]; - tensor var_726_axis_0 = const()[name = tensor("op_726_axis_0"), val = tensor(1)]; - tensor var_726_0, tensor var_726_1 = split(axis = var_726_axis_0, split_sizes = var_726_split_sizes_0, x = inputs_23)[name = tensor("op_726")]; - tensor var_728 = sigmoid(x = var_726_1)[name = tensor("op_728")]; - tensor inputs_25 = mul(x = var_726_0, y = var_728)[name = tensor("inputs_25")]; + tensor input_105 = transpose(perm = input_105_perm_0, x = x_17)[name = tensor("transpose_40")]; + tensor inputs_23 = conv(bias = inner_encoder_conv_module_2_sequential_2_conv_bias, dilations = inputs_23_dilations_0, groups = inputs_23_groups_0, pad = inputs_23_pad_0, pad_type = inputs_23_pad_type_0, strides = inputs_23_strides_0, weight = inner_encoder_conv_module_2_sequential_2_conv_weight, x = input_105)[name = tensor("inputs_23")]; + tensor var_791_split_sizes_0 = const()[name = tensor("op_791_split_sizes_0"), val = tensor([256, 256])]; + tensor var_791_axis_0 = const()[name = tensor("op_791_axis_0"), val = tensor(1)]; + tensor var_791_0, tensor var_791_1 = split(axis = var_791_axis_0, split_sizes = var_791_split_sizes_0, x = inputs_23)[name = tensor("op_791")]; + tensor var_793 = sigmoid(x = var_791_1)[name = tensor("op_793")]; + tensor inputs_25 = mul(x = var_791_0, y = var_793)[name = tensor("inputs_25")]; tensor outputs_aug_5_pad_type_0 = const()[name = tensor("outputs_aug_5_pad_type_0"), val = tensor("custom")]; tensor outputs_aug_5_pad_0 = const()[name = tensor("outputs_aug_5_pad_0"), val = tensor([15, 15])]; tensor outputs_aug_5_groups_0 = const()[name = tensor("outputs_aug_5_groups_0"), val = tensor(256)]; tensor outputs_aug_5_strides_0 = const()[name = tensor("outputs_aug_5_strides_0"), val = tensor([1])]; tensor outputs_aug_5_dilations_0 = const()[name = tensor("outputs_aug_5_dilations_0"), val = tensor([1])]; - tensor outputs_aug_5 = conv(dilations = outputs_aug_5_dilations_0, groups = outputs_aug_5_groups_0, pad = outputs_aug_5_pad_0, pad_type = outputs_aug_5_pad_type_0, strides = outputs_aug_5_strides_0, weight = encoder_conv_module_2_sequential_4_conv_weight, x = inputs_25)[name = tensor("outputs_aug_5")]; - tensor input_105_begin_0 = const()[name = tensor("input_105_begin_0"), val = tensor([0, 0, 0])]; - tensor input_105_end_0 = const()[name = tensor("input_105_end_0"), val = tensor([5, 256, 16])]; - tensor input_105_end_mask_0 = const()[name = tensor("input_105_end_mask_0"), val = tensor([true, true, false])]; - tensor input_105 = slice_by_index(begin = input_105_begin_0, end = input_105_end_0, end_mask = input_105_end_mask_0, x = outputs_aug_5)[name = tensor("input_105")]; - tensor inputs_27 = batch_norm(beta = encoder_conv_module_2_sequential_5_bias, epsilon = var_29, gamma = encoder_conv_module_2_sequential_5_weight, mean = encoder_conv_module_2_sequential_5_running_mean, variance = encoder_conv_module_2_sequential_5_running_var, x = input_105)[name = tensor("inputs_27")]; - tensor input_107 = silu(x = inputs_27)[name = tensor("input_107")]; - tensor input_109_pad_type_0 = const()[name = tensor("input_109_pad_type_0"), val = tensor("valid")]; - tensor input_109_strides_0 = const()[name = tensor("input_109_strides_0"), val = tensor([1])]; - tensor input_109_pad_0 = const()[name = tensor("input_109_pad_0"), val = tensor([0, 0])]; - tensor input_109_dilations_0 = const()[name = tensor("input_109_dilations_0"), val = tensor([1])]; - tensor input_109_groups_0 = const()[name = tensor("input_109_groups_0"), val = tensor(1)]; - tensor input_109 = conv(bias = encoder_conv_module_2_sequential_7_conv_bias, dilations = input_109_dilations_0, groups = input_109_groups_0, pad = input_109_pad_0, pad_type = input_109_pad_type_0, strides = input_109_strides_0, weight = encoder_conv_module_2_sequential_7_conv_weight, x = input_107)[name = tensor("input_109")]; + tensor outputs_aug_5 = conv(dilations = outputs_aug_5_dilations_0, groups = outputs_aug_5_groups_0, pad = outputs_aug_5_pad_0, pad_type = outputs_aug_5_pad_type_0, strides = outputs_aug_5_strides_0, weight = inner_encoder_conv_module_2_sequential_4_conv_weight, x = inputs_25)[name = tensor("outputs_aug_5")]; + tensor input_107_begin_0 = const()[name = tensor("input_107_begin_0"), val = tensor([0, 0, 0])]; + tensor input_107_end_0 = const()[name = tensor("input_107_end_0"), val = tensor([5, 256, 16])]; + tensor input_107_end_mask_0 = const()[name = tensor("input_107_end_mask_0"), val = tensor([true, true, false])]; + tensor input_107 = slice_by_index(begin = input_107_begin_0, end = input_107_end_0, end_mask = input_107_end_mask_0, x = outputs_aug_5)[name = tensor("input_107")]; + tensor inputs_27 = batch_norm(beta = inner_encoder_conv_module_2_sequential_5_bias, epsilon = var_74, gamma = inner_encoder_conv_module_2_sequential_5_weight, mean = inner_encoder_conv_module_2_sequential_5_running_mean, variance = inner_encoder_conv_module_2_sequential_5_running_var, x = input_107)[name = tensor("inputs_27")]; + tensor input_109 = silu(x = inputs_27)[name = tensor("input_109")]; + tensor input_111_pad_type_0 = const()[name = tensor("input_111_pad_type_0"), val = tensor("valid")]; + tensor input_111_strides_0 = const()[name = tensor("input_111_strides_0"), val = tensor([1])]; + tensor input_111_pad_0 = const()[name = tensor("input_111_pad_0"), val = tensor([0, 0])]; + tensor input_111_dilations_0 = const()[name = tensor("input_111_dilations_0"), val = tensor([1])]; + tensor input_111_groups_0 = const()[name = tensor("input_111_groups_0"), val = tensor(1)]; + tensor input_111 = conv(bias = inner_encoder_conv_module_2_sequential_7_conv_bias, dilations = input_111_dilations_0, groups = input_111_groups_0, pad = input_111_pad_0, pad_type = input_111_pad_type_0, strides = input_111_strides_0, weight = inner_encoder_conv_module_2_sequential_7_conv_weight, x = input_109)[name = tensor("input_111")]; tensor conv_out_5_perm_0 = const()[name = tensor("conv_out_5_perm_0"), val = tensor([0, 2, 1])]; - tensor var_759_begin_0 = const()[name = tensor("op_759_begin_0"), val = tensor([0, -1, 0])]; - tensor var_759_end_0 = const()[name = tensor("op_759_end_0"), val = tensor([5, 16, 256])]; - tensor var_759_end_mask_0 = const()[name = tensor("op_759_end_mask_0"), val = tensor([true, true, true])]; - tensor conv_out_5 = transpose(perm = conv_out_5_perm_0, x = input_109)[name = tensor("transpose_39")]; - tensor var_759 = slice_by_index(begin = var_759_begin_0, end = var_759_end_0, end_mask = var_759_end_mask_0, x = conv_out_5)[name = tensor("op_759")]; - tensor var_761_perm_0 = const()[name = tensor("op_761_perm_0"), val = tensor([1, 0, 2])]; - tensor var_761 = transpose(perm = var_761_perm_0, x = var_759)[name = tensor("transpose_38")]; - tensor input_111 = add(x = x_15, y = var_761)[name = tensor("input_111")]; - tensor input_113_axes_0 = const()[name = tensor("input_113_axes_0"), val = tensor([-1])]; - tensor input_113 = layer_norm(axes = input_113_axes_0, beta = encoder_ffn2_2_module_sequential_0_bias, epsilon = var_29, gamma = encoder_ffn2_2_module_sequential_0_weight, x = input_111)[name = tensor("input_113")]; - tensor inputs_29 = linear(bias = encoder_ffn2_2_module_sequential_1_linear_bias, weight = encoder_ffn2_2_module_sequential_1_linear_weight, x = input_113)[name = tensor("linear_26")]; - tensor input_115 = silu(x = inputs_29)[name = tensor("input_115")]; - tensor input_119 = linear(bias = encoder_ffn2_2_module_sequential_4_linear_bias, weight = encoder_ffn2_2_module_sequential_4_linear_weight, x = input_115)[name = tensor("linear_27")]; - tensor var_784 = const()[name = tensor("op_784"), val = tensor(0x1p-1)]; - tensor var_785 = mul(x = input_119, y = var_784)[name = tensor("op_785")]; - tensor input_121 = add(x = var_785, y = input_111)[name = tensor("input_121")]; - tensor input_123_axes_0 = const()[name = tensor("input_123_axes_0"), val = tensor([-1])]; - tensor input_123 = layer_norm(axes = input_123_axes_0, beta = encoder_layer_norm_2_bias, epsilon = var_29, gamma = encoder_layer_norm_2_weight, x = input_121)[name = tensor("input_123")]; + tensor var_824_begin_0 = const()[name = tensor("op_824_begin_0"), val = tensor([0, -1, 0])]; + tensor var_824_end_0 = const()[name = tensor("op_824_end_0"), val = tensor([5, 16, 256])]; + tensor var_824_end_mask_0 = const()[name = tensor("op_824_end_mask_0"), val = tensor([true, true, true])]; + tensor conv_out_5 = transpose(perm = conv_out_5_perm_0, x = input_111)[name = tensor("transpose_39")]; + tensor var_824 = slice_by_index(begin = var_824_begin_0, end = var_824_end_0, end_mask = var_824_end_mask_0, x = conv_out_5)[name = tensor("op_824")]; + tensor var_826_perm_0 = const()[name = tensor("op_826_perm_0"), val = tensor([1, 0, 2])]; + tensor var_826 = transpose(perm = var_826_perm_0, x = var_824)[name = tensor("transpose_38")]; + tensor input_113 = add(x = x_15, y = var_826)[name = tensor("input_113")]; + tensor input_115_axes_0 = const()[name = tensor("input_115_axes_0"), val = tensor([-1])]; + tensor input_115 = layer_norm(axes = input_115_axes_0, beta = inner_encoder_ffn2_2_module_sequential_0_bias, epsilon = var_74, gamma = inner_encoder_ffn2_2_module_sequential_0_weight, x = input_113)[name = tensor("input_115")]; + tensor inputs_29 = linear(bias = inner_encoder_ffn2_2_module_sequential_1_linear_bias, weight = inner_encoder_ffn2_2_module_sequential_1_linear_weight, x = input_115)[name = tensor("linear_26")]; + tensor input_117 = silu(x = inputs_29)[name = tensor("input_117")]; + tensor input_121 = linear(bias = inner_encoder_ffn2_2_module_sequential_4_linear_bias, weight = inner_encoder_ffn2_2_module_sequential_4_linear_weight, x = input_117)[name = tensor("linear_27")]; + tensor var_849 = const()[name = tensor("op_849"), val = tensor(0x1p-1)]; + tensor var_850 = mul(x = input_121, y = var_849)[name = tensor("op_850")]; + tensor input_123 = add(x = var_850, y = input_113)[name = tensor("input_123")]; tensor input_125_axes_0 = const()[name = tensor("input_125_axes_0"), val = tensor([-1])]; - tensor input_125 = layer_norm(axes = input_125_axes_0, beta = encoder_ffn1_3_module_sequential_0_bias, epsilon = var_29, gamma = encoder_ffn1_3_module_sequential_0_weight, x = input_123)[name = tensor("input_125")]; - tensor inputs_31 = linear(bias = encoder_ffn1_3_module_sequential_1_linear_bias, weight = encoder_ffn1_3_module_sequential_1_linear_weight, x = input_125)[name = tensor("linear_28")]; - tensor input_127 = silu(x = inputs_31)[name = tensor("input_127")]; - tensor input_131 = linear(bias = encoder_ffn1_3_module_sequential_4_linear_bias, weight = encoder_ffn1_3_module_sequential_4_linear_weight, x = input_127)[name = tensor("linear_29")]; - tensor var_814 = const()[name = tensor("op_814"), val = tensor(0x1p-1)]; - tensor var_815 = mul(x = input_131, y = var_814)[name = tensor("op_815")]; - tensor input_133 = add(x = var_815, y = input_123)[name = tensor("input_133")]; + tensor input_125 = layer_norm(axes = input_125_axes_0, beta = inner_encoder_layer_norm_2_bias, epsilon = var_74, gamma = inner_encoder_layer_norm_2_weight, x = input_123)[name = tensor("input_125")]; + tensor input_127_axes_0 = const()[name = tensor("input_127_axes_0"), val = tensor([-1])]; + tensor input_127 = layer_norm(axes = input_127_axes_0, beta = inner_encoder_ffn1_3_module_sequential_0_bias, epsilon = var_74, gamma = inner_encoder_ffn1_3_module_sequential_0_weight, x = input_125)[name = tensor("input_127")]; + tensor inputs_31 = linear(bias = inner_encoder_ffn1_3_module_sequential_1_linear_bias, weight = inner_encoder_ffn1_3_module_sequential_1_linear_weight, x = input_127)[name = tensor("linear_28")]; + tensor input_129 = silu(x = inputs_31)[name = tensor("input_129")]; + tensor input_133 = linear(bias = inner_encoder_ffn1_3_module_sequential_4_linear_bias, weight = inner_encoder_ffn1_3_module_sequential_4_linear_weight, x = input_129)[name = tensor("linear_29")]; + tensor var_879 = const()[name = tensor("op_879"), val = tensor(0x1p-1)]; + tensor var_880 = mul(x = input_133, y = var_879)[name = tensor("op_880")]; + tensor input_135 = add(x = var_880, y = input_125)[name = tensor("input_135")]; tensor x_19_axes_0 = const()[name = tensor("x_19_axes_0"), val = tensor([-1])]; - tensor x_19 = layer_norm(axes = x_19_axes_0, beta = encoder_ret_lns_3_bias, epsilon = var_29, gamma = encoder_ret_lns_3_weight, x = input_133)[name = tensor("x_19")]; + tensor x_19 = layer_norm(axes = x_19_axes_0, beta = inner_encoder_ret_lns_3_bias, epsilon = var_74, gamma = inner_encoder_ret_lns_3_weight, x = input_135)[name = tensor("x_19")]; tensor prev_kv_7_begin_0 = const()[name = tensor("prev_kv_7_begin_0"), val = tensor([3, 0, 0, 0, 0])]; tensor prev_kv_7_end_0 = const()[name = tensor("prev_kv_7_end_0"), val = tensor([4, 1, 4, 64, 64])]; tensor prev_kv_7_end_mask_0 = const()[name = tensor("prev_kv_7_end_mask_0"), val = tensor([false, true, true, true, true])]; @@ -800,219 +826,212 @@ program(1.0) tensor prev_scale_7_end_mask_0 = const()[name = tensor("prev_scale_7_end_mask_0"), val = tensor([false, true])]; tensor prev_scale_7_squeeze_mask_0 = const()[name = tensor("prev_scale_7_squeeze_mask_0"), val = tensor([true, false])]; tensor prev_scale_7 = slice_by_index(begin = prev_scale_7_begin_0, end = prev_scale_7_end_0, end_mask = prev_scale_7_end_mask_0, squeeze_mask = prev_scale_7_squeeze_mask_0, x = enc_scale)[name = tensor("prev_scale_7")]; - tensor var_829 = linear(bias = encoder_q_proj_3_bias, weight = encoder_q_proj_3_weight, x = x_19)[name = tensor("linear_30")]; - tensor var_830 = const()[name = tensor("op_830"), val = tensor([1, 5, 4, 64])]; - tensor var_831 = reshape(shape = var_830, x = var_829)[name = tensor("op_831")]; + tensor var_894 = linear(bias = inner_encoder_q_proj_3_bias, weight = inner_encoder_q_proj_3_weight, x = x_19)[name = tensor("linear_30")]; + tensor var_895 = const()[name = tensor("op_895"), val = tensor([1, 5, 4, 64])]; + tensor var_896 = reshape(shape = var_895, x = var_894)[name = tensor("op_896")]; tensor q_7_perm_0 = const()[name = tensor("q_7_perm_0"), val = tensor([0, 2, 1, 3])]; - tensor var_835 = linear(bias = encoder_k_proj_3_bias, weight = encoder_k_proj_3_weight, x = x_19)[name = tensor("linear_31")]; - tensor var_836 = const()[name = tensor("op_836"), val = tensor(0x1p-3)]; - tensor var_837 = mul(x = var_835, y = var_836)[name = tensor("op_837")]; - tensor var_838 = const()[name = tensor("op_838"), val = tensor([1, 5, 4, 64])]; - tensor var_839 = reshape(shape = var_838, x = var_837)[name = tensor("op_839")]; + tensor var_900 = linear(bias = inner_encoder_k_proj_3_bias, weight = inner_encoder_k_proj_3_weight, x = x_19)[name = tensor("linear_31")]; + tensor var_901 = const()[name = tensor("op_901"), val = tensor(0x1p-3)]; + tensor var_902 = mul(x = var_900, y = var_901)[name = tensor("op_902")]; + tensor var_903 = const()[name = tensor("op_903"), val = tensor([1, 5, 4, 64])]; + tensor var_904 = reshape(shape = var_903, x = var_902)[name = tensor("op_904")]; tensor k_7_perm_0 = const()[name = tensor("k_7_perm_0"), val = tensor([0, 2, 1, 3])]; - tensor var_843 = linear(bias = encoder_v_proj_3_bias, weight = encoder_v_proj_3_weight, x = x_19)[name = tensor("linear_32")]; - tensor var_844 = const()[name = tensor("op_844"), val = tensor([1, 5, 4, 64])]; - tensor var_845 = reshape(shape = var_844, x = var_843)[name = tensor("op_845")]; + tensor var_908 = linear(bias = inner_encoder_v_proj_3_bias, weight = inner_encoder_v_proj_3_weight, x = x_19)[name = tensor("linear_32")]; + tensor var_909 = const()[name = tensor("op_909"), val = tensor([1, 5, 4, 64])]; + tensor var_910 = reshape(shape = var_909, x = var_908)[name = tensor("op_910")]; tensor v_7_perm_0 = const()[name = tensor("v_7_perm_0"), val = tensor([0, 2, 1, 3])]; - tensor input_137 = linear(bias = encoder_g_proj_3_bias, weight = encoder_g_proj_3_weight, x = x_19)[name = tensor("linear_33")]; + tensor input_139 = linear(bias = inner_encoder_g_proj_3_bias, weight = inner_encoder_g_proj_3_weight, x = x_19)[name = tensor("linear_33")]; tensor sqrt_s0_7 = sqrt(x = prev_scale_7)[name = tensor("sqrt_s0_7")]; - tensor s_t_7 = add(x = prev_scale_7, y = encoder__t_index)[name = tensor("s_t_7")]; + tensor s_t_7 = add(x = prev_scale_7, y = inner_encoder__t_index)[name = tensor("s_t_7")]; tensor sqrt_s_t_7 = sqrt(x = s_t_7)[name = tensor("sqrt_s_t_7")]; tensor qk_7_transpose_x_1 = const()[name = tensor("qk_7_transpose_x_1"), val = tensor(false)]; tensor qk_7_transpose_y_1 = const()[name = tensor("qk_7_transpose_y_1"), val = tensor(true)]; - tensor k_7 = transpose(perm = k_7_perm_0, x = var_839)[name = tensor("transpose_36")]; - tensor q_7 = transpose(perm = q_7_perm_0, x = var_831)[name = tensor("transpose_37")]; + tensor k_7 = transpose(perm = k_7_perm_0, x = var_904)[name = tensor("transpose_36")]; + tensor q_7 = transpose(perm = q_7_perm_0, x = var_896)[name = tensor("transpose_37")]; tensor qk_7 = matmul(transpose_x = qk_7_transpose_x_1, transpose_y = qk_7_transpose_y_1, x = q_7, y = k_7)[name = tensor("qk_7")]; - tensor var_855 = const()[name = tensor("op_855"), val = tensor([5, 1])]; - tensor var_856 = reshape(shape = var_855, x = sqrt_s_t_7)[name = tensor("op_856")]; - tensor M_7 = real_div(x = encoder__causal_mask, y = var_856)[name = tensor("M_7")]; - tensor var_858 = mul(x = qk_7, y = M_7)[name = tensor("op_858")]; + tensor var_920 = const()[name = tensor("op_920"), val = tensor([5, 1])]; + tensor var_921 = reshape(shape = var_920, x = sqrt_s_t_7)[name = tensor("op_921")]; + tensor M_7 = real_div(x = inner_encoder__causal_mask, y = var_921)[name = tensor("M_7")]; + tensor var_923 = mul(x = qk_7, y = M_7)[name = tensor("op_923")]; tensor inner_7_transpose_x_0 = const()[name = tensor("inner_7_transpose_x_0"), val = tensor(false)]; tensor inner_7_transpose_y_0 = const()[name = tensor("inner_7_transpose_y_0"), val = tensor(false)]; - tensor v_7 = transpose(perm = v_7_perm_0, x = var_845)[name = tensor("transpose_35")]; - tensor inner_7 = matmul(transpose_x = inner_7_transpose_x_0, transpose_y = inner_7_transpose_y_0, x = var_858, y = v_7)[name = tensor("inner_7")]; - tensor var_860_transpose_x_0 = const()[name = tensor("op_860_transpose_x_0"), val = tensor(false)]; - tensor var_860_transpose_y_0 = const()[name = tensor("op_860_transpose_y_0"), val = tensor(false)]; - tensor var_860 = matmul(transpose_x = var_860_transpose_x_0, transpose_y = var_860_transpose_y_0, x = q_7, y = prev_kv_7)[name = tensor("op_860")]; - tensor var_861 = real_div(x = sqrt_s0_7, y = sqrt_s_t_7)[name = tensor("op_861")]; - tensor var_862 = const()[name = tensor("op_862"), val = tensor([1, 1, 5, 1])]; - tensor var_863 = reshape(shape = var_862, x = var_861)[name = tensor("op_863")]; - tensor cross_7 = mul(x = var_860, y = var_863)[name = tensor("cross_7")]; + tensor v_7 = transpose(perm = v_7_perm_0, x = var_910)[name = tensor("transpose_35")]; + tensor inner_7 = matmul(transpose_x = inner_7_transpose_x_0, transpose_y = inner_7_transpose_y_0, x = var_923, y = v_7)[name = tensor("inner_7")]; + tensor var_925_transpose_x_0 = const()[name = tensor("op_925_transpose_x_0"), val = tensor(false)]; + tensor var_925_transpose_y_0 = const()[name = tensor("op_925_transpose_y_0"), val = tensor(false)]; + tensor var_925 = matmul(transpose_x = var_925_transpose_x_0, transpose_y = var_925_transpose_y_0, x = q_7, y = prev_kv_7)[name = tensor("op_925")]; + tensor var_926 = real_div(x = sqrt_s0_7, y = sqrt_s_t_7)[name = tensor("op_926")]; + tensor var_927 = const()[name = tensor("op_927"), val = tensor([1, 1, 5, 1])]; + tensor var_928 = reshape(shape = var_927, x = var_926)[name = tensor("op_928")]; + tensor cross_7 = mul(x = var_925, y = var_928)[name = tensor("cross_7")]; tensor out_19 = add(x = inner_7, y = cross_7)[name = tensor("out_19")]; - tensor var_866 = mul(x = prev_kv_7, y = sqrt_s0_7)[name = tensor("op_866")]; - tensor var_868_transpose_x_1 = const()[name = tensor("op_868_transpose_x_1"), val = tensor(true)]; - tensor var_868_transpose_y_1 = const()[name = tensor("op_868_transpose_y_1"), val = tensor(false)]; - tensor var_868 = matmul(transpose_x = var_868_transpose_x_1, transpose_y = var_868_transpose_y_1, x = k_7, y = v_7)[name = tensor("op_868")]; - tensor new_kv_unnorm_7 = add(x = var_866, y = var_868)[name = tensor("new_kv_unnorm_7")]; - tensor var_870 = const()[name = tensor("op_870"), val = tensor(0x1.4p+2)]; - tensor new_scale_7 = add(x = prev_scale_7, y = var_870)[name = tensor("new_scale_7")]; - tensor var_872 = sqrt(x = new_scale_7)[name = tensor("op_872")]; - tensor nkv_1 = real_div(x = new_kv_unnorm_7, y = var_872)[name = tensor("nkv_1")]; - tensor var_874_perm_0 = const()[name = tensor("op_874_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_931 = mul(x = prev_kv_7, y = sqrt_s0_7)[name = tensor("op_931")]; + tensor var_933_transpose_x_1 = const()[name = tensor("op_933_transpose_x_1"), val = tensor(true)]; + tensor var_933_transpose_y_1 = const()[name = tensor("op_933_transpose_y_1"), val = tensor(false)]; + tensor var_933 = matmul(transpose_x = var_933_transpose_x_1, transpose_y = var_933_transpose_y_1, x = k_7, y = v_7)[name = tensor("op_933")]; + tensor new_kv_unnorm_7 = add(x = var_931, y = var_933)[name = tensor("new_kv_unnorm_7")]; + tensor var_935 = const()[name = tensor("op_935"), val = tensor(0x1.4p+2)]; + tensor new_scale_7 = add(x = prev_scale_7, y = var_935)[name = tensor("new_scale_7")]; + tensor var_937 = sqrt(x = new_scale_7)[name = tensor("op_937")]; + tensor nkv_1 = real_div(x = new_kv_unnorm_7, y = var_937)[name = tensor("nkv_1")]; + tensor var_939_perm_0 = const()[name = tensor("op_939_perm_0"), val = tensor([0, 2, 1, 3])]; tensor out_21_axes_0 = const()[name = tensor("out_21_axes_0"), val = tensor([-1])]; - tensor var_874 = transpose(perm = var_874_perm_0, x = out_19)[name = tensor("transpose_34")]; - tensor out_21 = layer_norm(axes = out_21_axes_0, epsilon = var_18, x = var_874)[name = tensor("out_21")]; - tensor var_878 = const()[name = tensor("op_878"), val = tensor([1, 5, 256])]; - tensor out_23 = reshape(shape = var_878, x = out_21)[name = tensor("out_23")]; - tensor var_880 = silu(x = input_137)[name = tensor("op_880")]; - tensor input_139 = mul(x = var_880, y = out_23)[name = tensor("input_139")]; - tensor ret_out_7 = linear(bias = encoder_out_proj_3_bias, weight = encoder_out_proj_3_weight, x = input_139)[name = tensor("linear_34")]; - tensor x_21 = add(x = input_133, y = ret_out_7)[name = tensor("x_21")]; + tensor var_939 = transpose(perm = var_939_perm_0, x = out_19)[name = tensor("transpose_34")]; + tensor out_21 = layer_norm(axes = out_21_axes_0, epsilon = var_82, x = var_939)[name = tensor("out_21")]; + tensor var_943 = const()[name = tensor("op_943"), val = tensor([1, 5, 256])]; + tensor out_23 = reshape(shape = var_943, x = out_21)[name = tensor("out_23")]; + tensor var_945 = silu(x = input_139)[name = tensor("op_945")]; + tensor input_141 = mul(x = var_945, y = out_23)[name = tensor("input_141")]; + tensor ret_out_7 = linear(bias = inner_encoder_out_proj_3_bias, weight = inner_encoder_out_proj_3_weight, x = input_141)[name = tensor("linear_34")]; + tensor x_21 = add(x = input_135, y = ret_out_7)[name = tensor("x_21")]; tensor window_37_begin_0 = const()[name = tensor("window_37_begin_0"), val = tensor([3, 0, 0, 0])]; tensor window_37_end_0 = const()[name = tensor("window_37_end_0"), val = tensor([4, 1, 16, 256])]; tensor window_37_end_mask_0 = const()[name = tensor("window_37_end_mask_0"), val = tensor([false, true, true, true])]; tensor window_37_squeeze_mask_0 = const()[name = tensor("window_37_squeeze_mask_0"), val = tensor([true, false, false, false])]; tensor window_37 = slice_by_index(begin = window_37_begin_0, end = window_37_end_0, end_mask = window_37_end_mask_0, squeeze_mask = window_37_squeeze_mask_0, x = enc_conv_cache)[name = tensor("window_37")]; - tensor var_888_begin_0 = const()[name = tensor("op_888_begin_0"), val = tensor([0, 0, 0])]; - tensor var_888_end_0 = const()[name = tensor("op_888_end_0"), val = tensor([1, 1, 256])]; - tensor var_888_end_mask_0 = const()[name = tensor("op_888_end_mask_0"), val = tensor([true, false, true])]; - tensor var_888 = slice_by_index(begin = var_888_begin_0, end = var_888_end_0, end_mask = var_888_end_mask_0, x = x_21)[name = tensor("op_888")]; - tensor var_891_begin_0 = const()[name = tensor("op_891_begin_0"), val = tensor([0, 1, 0])]; - tensor var_891_end_0 = const()[name = tensor("op_891_end_0"), val = tensor([1, 16, 256])]; - tensor var_891_end_mask_0 = const()[name = tensor("op_891_end_mask_0"), val = tensor([true, true, true])]; - tensor var_891 = slice_by_index(begin = var_891_begin_0, end = var_891_end_0, end_mask = var_891_end_mask_0, x = window_37)[name = tensor("op_891")]; + tensor var_953_begin_0 = const()[name = tensor("op_953_begin_0"), val = tensor([0, 0, 0])]; + tensor var_953_end_0 = const()[name = tensor("op_953_end_0"), val = tensor([1, 1, 256])]; + tensor var_953_end_mask_0 = const()[name = tensor("op_953_end_mask_0"), val = tensor([true, false, true])]; + tensor var_953 = slice_by_index(begin = var_953_begin_0, end = var_953_end_0, end_mask = var_953_end_mask_0, x = x_21)[name = tensor("op_953")]; + tensor var_956_begin_0 = const()[name = tensor("op_956_begin_0"), val = tensor([0, 1, 0])]; + tensor var_956_end_0 = const()[name = tensor("op_956_end_0"), val = tensor([1, 16, 256])]; + tensor var_956_end_mask_0 = const()[name = tensor("op_956_end_mask_0"), val = tensor([true, true, true])]; + tensor var_956 = slice_by_index(begin = var_956_begin_0, end = var_956_end_0, end_mask = var_956_end_mask_0, x = window_37)[name = tensor("op_956")]; tensor window_39_interleave_0 = const()[name = tensor("window_39_interleave_0"), val = tensor(false)]; - tensor window_39 = concat(axis = var_27, interleave = window_39_interleave_0, values = (var_891, var_888))[name = tensor("window_39")]; - tensor var_896_begin_0 = const()[name = tensor("op_896_begin_0"), val = tensor([0, 1, 0])]; - tensor var_896_end_0 = const()[name = tensor("op_896_end_0"), val = tensor([1, 2, 256])]; - tensor var_896_end_mask_0 = const()[name = tensor("op_896_end_mask_0"), val = tensor([true, false, true])]; - tensor var_896 = slice_by_index(begin = var_896_begin_0, end = var_896_end_0, end_mask = var_896_end_mask_0, x = x_21)[name = tensor("op_896")]; - tensor var_899_begin_0 = const()[name = tensor("op_899_begin_0"), val = tensor([0, 1, 0])]; - tensor var_899_end_0 = const()[name = tensor("op_899_end_0"), val = tensor([1, 16, 256])]; - tensor var_899_end_mask_0 = const()[name = tensor("op_899_end_mask_0"), val = tensor([true, true, true])]; - tensor var_899 = slice_by_index(begin = var_899_begin_0, end = var_899_end_0, end_mask = var_899_end_mask_0, x = window_39)[name = tensor("op_899")]; + tensor window_39 = concat(axis = var_92, interleave = window_39_interleave_0, values = (var_956, var_953))[name = tensor("window_39")]; + tensor var_961_begin_0 = const()[name = tensor("op_961_begin_0"), val = tensor([0, 1, 0])]; + tensor var_961_end_0 = const()[name = tensor("op_961_end_0"), val = tensor([1, 2, 256])]; + tensor var_961_end_mask_0 = const()[name = tensor("op_961_end_mask_0"), val = tensor([true, false, true])]; + tensor var_961 = slice_by_index(begin = var_961_begin_0, end = var_961_end_0, end_mask = var_961_end_mask_0, x = x_21)[name = tensor("op_961")]; + tensor var_964_begin_0 = const()[name = tensor("op_964_begin_0"), val = tensor([0, 1, 0])]; + tensor var_964_end_0 = const()[name = tensor("op_964_end_0"), val = tensor([1, 16, 256])]; + tensor var_964_end_mask_0 = const()[name = tensor("op_964_end_mask_0"), val = tensor([true, true, true])]; + tensor var_964 = slice_by_index(begin = var_964_begin_0, end = var_964_end_0, end_mask = var_964_end_mask_0, x = window_39)[name = tensor("op_964")]; tensor window_41_interleave_0 = const()[name = tensor("window_41_interleave_0"), val = tensor(false)]; - tensor window_41 = concat(axis = var_27, interleave = window_41_interleave_0, values = (var_899, var_896))[name = tensor("window_41")]; - tensor var_904_begin_0 = const()[name = tensor("op_904_begin_0"), val = tensor([0, 2, 0])]; - tensor var_904_end_0 = const()[name = tensor("op_904_end_0"), val = tensor([1, 3, 256])]; - tensor var_904_end_mask_0 = const()[name = tensor("op_904_end_mask_0"), val = tensor([true, false, true])]; - tensor var_904 = slice_by_index(begin = var_904_begin_0, end = var_904_end_0, end_mask = var_904_end_mask_0, x = x_21)[name = tensor("op_904")]; - tensor var_907_begin_0 = const()[name = tensor("op_907_begin_0"), val = tensor([0, 1, 0])]; - tensor var_907_end_0 = const()[name = tensor("op_907_end_0"), val = tensor([1, 16, 256])]; - tensor var_907_end_mask_0 = const()[name = tensor("op_907_end_mask_0"), val = tensor([true, true, true])]; - tensor var_907 = slice_by_index(begin = var_907_begin_0, end = var_907_end_0, end_mask = var_907_end_mask_0, x = window_41)[name = tensor("op_907")]; + tensor window_41 = concat(axis = var_92, interleave = window_41_interleave_0, values = (var_964, var_961))[name = tensor("window_41")]; + tensor var_969_begin_0 = const()[name = tensor("op_969_begin_0"), val = tensor([0, 2, 0])]; + tensor var_969_end_0 = const()[name = tensor("op_969_end_0"), val = tensor([1, 3, 256])]; + tensor var_969_end_mask_0 = const()[name = tensor("op_969_end_mask_0"), val = tensor([true, false, true])]; + tensor var_969 = slice_by_index(begin = var_969_begin_0, end = var_969_end_0, end_mask = var_969_end_mask_0, x = x_21)[name = tensor("op_969")]; + tensor var_972_begin_0 = const()[name = tensor("op_972_begin_0"), val = tensor([0, 1, 0])]; + tensor var_972_end_0 = const()[name = tensor("op_972_end_0"), val = tensor([1, 16, 256])]; + tensor var_972_end_mask_0 = const()[name = tensor("op_972_end_mask_0"), val = tensor([true, true, true])]; + tensor var_972 = slice_by_index(begin = var_972_begin_0, end = var_972_end_0, end_mask = var_972_end_mask_0, x = window_41)[name = tensor("op_972")]; tensor window_43_interleave_0 = const()[name = tensor("window_43_interleave_0"), val = tensor(false)]; - tensor window_43 = concat(axis = var_27, interleave = window_43_interleave_0, values = (var_907, var_904))[name = tensor("window_43")]; - tensor var_912_begin_0 = const()[name = tensor("op_912_begin_0"), val = tensor([0, 3, 0])]; - tensor var_912_end_0 = const()[name = tensor("op_912_end_0"), val = tensor([1, 4, 256])]; - tensor var_912_end_mask_0 = const()[name = tensor("op_912_end_mask_0"), val = tensor([true, false, true])]; - tensor var_912 = slice_by_index(begin = var_912_begin_0, end = var_912_end_0, end_mask = var_912_end_mask_0, x = x_21)[name = tensor("op_912")]; - tensor var_915_begin_0 = const()[name = tensor("op_915_begin_0"), val = tensor([0, 1, 0])]; - tensor var_915_end_0 = const()[name = tensor("op_915_end_0"), val = tensor([1, 16, 256])]; - tensor var_915_end_mask_0 = const()[name = tensor("op_915_end_mask_0"), val = tensor([true, true, true])]; - tensor var_915 = slice_by_index(begin = var_915_begin_0, end = var_915_end_0, end_mask = var_915_end_mask_0, x = window_43)[name = tensor("op_915")]; + tensor window_43 = concat(axis = var_92, interleave = window_43_interleave_0, values = (var_972, var_969))[name = tensor("window_43")]; + tensor var_977_begin_0 = const()[name = tensor("op_977_begin_0"), val = tensor([0, 3, 0])]; + tensor var_977_end_0 = const()[name = tensor("op_977_end_0"), val = tensor([1, 4, 256])]; + tensor var_977_end_mask_0 = const()[name = tensor("op_977_end_mask_0"), val = tensor([true, false, true])]; + tensor var_977 = slice_by_index(begin = var_977_begin_0, end = var_977_end_0, end_mask = var_977_end_mask_0, x = x_21)[name = tensor("op_977")]; + tensor var_980_begin_0 = const()[name = tensor("op_980_begin_0"), val = tensor([0, 1, 0])]; + tensor var_980_end_0 = const()[name = tensor("op_980_end_0"), val = tensor([1, 16, 256])]; + tensor var_980_end_mask_0 = const()[name = tensor("op_980_end_mask_0"), val = tensor([true, true, true])]; + tensor var_980 = slice_by_index(begin = var_980_begin_0, end = var_980_end_0, end_mask = var_980_end_mask_0, x = window_43)[name = tensor("op_980")]; tensor window_45_interleave_0 = const()[name = tensor("window_45_interleave_0"), val = tensor(false)]; - tensor window_45 = concat(axis = var_27, interleave = window_45_interleave_0, values = (var_915, var_912))[name = tensor("window_45")]; - tensor var_920_begin_0 = const()[name = tensor("op_920_begin_0"), val = tensor([0, 4, 0])]; - tensor var_920_end_0 = const()[name = tensor("op_920_end_0"), val = tensor([1, 1, 256])]; - tensor var_920_end_mask_0 = const()[name = tensor("op_920_end_mask_0"), val = tensor([true, true, true])]; - tensor var_920 = slice_by_index(begin = var_920_begin_0, end = var_920_end_0, end_mask = var_920_end_mask_0, x = x_21)[name = tensor("op_920")]; - tensor var_923_begin_0 = const()[name = tensor("op_923_begin_0"), val = tensor([0, 1, 0])]; - tensor var_923_end_0 = const()[name = tensor("op_923_end_0"), val = tensor([1, 16, 256])]; - tensor var_923_end_mask_0 = const()[name = tensor("op_923_end_mask_0"), val = tensor([true, true, true])]; - tensor var_923 = slice_by_index(begin = var_923_begin_0, end = var_923_end_0, end_mask = var_923_end_mask_0, x = window_45)[name = tensor("op_923")]; + tensor window_45 = concat(axis = var_92, interleave = window_45_interleave_0, values = (var_980, var_977))[name = tensor("window_45")]; + tensor var_985_begin_0 = const()[name = tensor("op_985_begin_0"), val = tensor([0, 4, 0])]; + tensor var_985_end_0 = const()[name = tensor("op_985_end_0"), val = tensor([1, 1, 256])]; + tensor var_985_end_mask_0 = const()[name = tensor("op_985_end_mask_0"), val = tensor([true, true, true])]; + tensor var_985 = slice_by_index(begin = var_985_begin_0, end = var_985_end_0, end_mask = var_985_end_mask_0, x = x_21)[name = tensor("op_985")]; + tensor var_988_begin_0 = const()[name = tensor("op_988_begin_0"), val = tensor([0, 1, 0])]; + tensor var_988_end_0 = const()[name = tensor("op_988_end_0"), val = tensor([1, 16, 256])]; + tensor var_988_end_mask_0 = const()[name = tensor("op_988_end_mask_0"), val = tensor([true, true, true])]; + tensor var_988 = slice_by_index(begin = var_988_begin_0, end = var_988_end_0, end_mask = var_988_end_mask_0, x = window_45)[name = tensor("op_988")]; tensor window_interleave_0 = const()[name = tensor("window_interleave_0"), val = tensor(false)]; - tensor window = concat(axis = var_27, interleave = window_interleave_0, values = (var_923, var_920))[name = tensor("window")]; - tensor input_141_interleave_0 = const()[name = tensor("input_141_interleave_0"), val = tensor(false)]; - tensor input_141 = concat(axis = var_24, interleave = input_141_interleave_0, values = (window_39, window_41, window_43, window_45, window))[name = tensor("input_141")]; + tensor window = concat(axis = var_92, interleave = window_interleave_0, values = (var_988, var_985))[name = tensor("window")]; + tensor input_143_interleave_0 = const()[name = tensor("input_143_interleave_0"), val = tensor(false)]; + tensor input_143 = concat(axis = var_77, interleave = input_143_interleave_0, values = (window_39, window_41, window_43, window_45, window))[name = tensor("input_143")]; tensor x_23_axes_0 = const()[name = tensor("x_23_axes_0"), val = tensor([-1])]; - tensor x_23 = layer_norm(axes = x_23_axes_0, beta = encoder_conv_module_3_sequential_0_bias, epsilon = var_29, gamma = encoder_conv_module_3_sequential_0_weight, x = input_141)[name = tensor("x_23")]; - tensor input_143_perm_0 = const()[name = tensor("input_143_perm_0"), val = tensor([0, 2, 1])]; + tensor x_23 = layer_norm(axes = x_23_axes_0, beta = inner_encoder_conv_module_3_sequential_0_bias, epsilon = var_74, gamma = inner_encoder_conv_module_3_sequential_0_weight, x = input_143)[name = tensor("x_23")]; + tensor input_145_perm_0 = const()[name = tensor("input_145_perm_0"), val = tensor([0, 2, 1])]; tensor inputs_33_pad_type_0 = const()[name = tensor("inputs_33_pad_type_0"), val = tensor("valid")]; tensor inputs_33_strides_0 = const()[name = tensor("inputs_33_strides_0"), val = tensor([1])]; tensor inputs_33_pad_0 = const()[name = tensor("inputs_33_pad_0"), val = tensor([0, 0])]; tensor inputs_33_dilations_0 = const()[name = tensor("inputs_33_dilations_0"), val = tensor([1])]; tensor inputs_33_groups_0 = const()[name = tensor("inputs_33_groups_0"), val = tensor(1)]; - tensor input_143 = transpose(perm = input_143_perm_0, x = x_23)[name = tensor("transpose_33")]; - tensor inputs_33 = conv(bias = encoder_conv_module_3_sequential_2_conv_bias, dilations = inputs_33_dilations_0, groups = inputs_33_groups_0, pad = inputs_33_pad_0, pad_type = inputs_33_pad_type_0, strides = inputs_33_strides_0, weight = encoder_conv_module_3_sequential_2_conv_weight, x = input_143)[name = tensor("inputs_33")]; - tensor var_948_split_sizes_0 = const()[name = tensor("op_948_split_sizes_0"), val = tensor([256, 256])]; - tensor var_948_axis_0 = const()[name = tensor("op_948_axis_0"), val = tensor(1)]; - tensor var_948_0, tensor var_948_1 = split(axis = var_948_axis_0, split_sizes = var_948_split_sizes_0, x = inputs_33)[name = tensor("op_948")]; - tensor var_950 = sigmoid(x = var_948_1)[name = tensor("op_950")]; - tensor inputs_35 = mul(x = var_948_0, y = var_950)[name = tensor("inputs_35")]; + tensor input_145 = transpose(perm = input_145_perm_0, x = x_23)[name = tensor("transpose_33")]; + tensor inputs_33 = conv(bias = inner_encoder_conv_module_3_sequential_2_conv_bias, dilations = inputs_33_dilations_0, groups = inputs_33_groups_0, pad = inputs_33_pad_0, pad_type = inputs_33_pad_type_0, strides = inputs_33_strides_0, weight = inner_encoder_conv_module_3_sequential_2_conv_weight, x = input_145)[name = tensor("inputs_33")]; + tensor var_1013_split_sizes_0 = const()[name = tensor("op_1013_split_sizes_0"), val = tensor([256, 256])]; + tensor var_1013_axis_0 = const()[name = tensor("op_1013_axis_0"), val = tensor(1)]; + tensor var_1013_0, tensor var_1013_1 = split(axis = var_1013_axis_0, split_sizes = var_1013_split_sizes_0, x = inputs_33)[name = tensor("op_1013")]; + tensor var_1015 = sigmoid(x = var_1013_1)[name = tensor("op_1015")]; + tensor inputs_35 = mul(x = var_1013_0, y = var_1015)[name = tensor("inputs_35")]; tensor outputs_aug_pad_type_0 = const()[name = tensor("outputs_aug_pad_type_0"), val = tensor("custom")]; tensor outputs_aug_pad_0 = const()[name = tensor("outputs_aug_pad_0"), val = tensor([15, 15])]; tensor outputs_aug_groups_0 = const()[name = tensor("outputs_aug_groups_0"), val = tensor(256)]; tensor outputs_aug_strides_0 = const()[name = tensor("outputs_aug_strides_0"), val = tensor([1])]; tensor outputs_aug_dilations_0 = const()[name = tensor("outputs_aug_dilations_0"), val = tensor([1])]; - tensor outputs_aug = conv(dilations = outputs_aug_dilations_0, groups = outputs_aug_groups_0, pad = outputs_aug_pad_0, pad_type = outputs_aug_pad_type_0, strides = outputs_aug_strides_0, weight = encoder_conv_module_3_sequential_4_conv_weight, x = inputs_35)[name = tensor("outputs_aug")]; - tensor input_145_begin_0 = const()[name = tensor("input_145_begin_0"), val = tensor([0, 0, 0])]; - tensor input_145_end_0 = const()[name = tensor("input_145_end_0"), val = tensor([5, 256, 16])]; - tensor input_145_end_mask_0 = const()[name = tensor("input_145_end_mask_0"), val = tensor([true, true, false])]; - tensor input_145 = slice_by_index(begin = input_145_begin_0, end = input_145_end_0, end_mask = input_145_end_mask_0, x = outputs_aug)[name = tensor("input_145")]; - tensor inputs_37 = batch_norm(beta = encoder_conv_module_3_sequential_5_bias, epsilon = var_29, gamma = encoder_conv_module_3_sequential_5_weight, mean = encoder_conv_module_3_sequential_5_running_mean, variance = encoder_conv_module_3_sequential_5_running_var, x = input_145)[name = tensor("inputs_37")]; - tensor input_147 = silu(x = inputs_37)[name = tensor("input_147")]; - tensor input_149_pad_type_0 = const()[name = tensor("input_149_pad_type_0"), val = tensor("valid")]; - tensor input_149_strides_0 = const()[name = tensor("input_149_strides_0"), val = tensor([1])]; - tensor input_149_pad_0 = const()[name = tensor("input_149_pad_0"), val = tensor([0, 0])]; - tensor input_149_dilations_0 = const()[name = tensor("input_149_dilations_0"), val = tensor([1])]; - tensor input_149_groups_0 = const()[name = tensor("input_149_groups_0"), val = tensor(1)]; - tensor input_149 = conv(bias = encoder_conv_module_3_sequential_7_conv_bias, dilations = input_149_dilations_0, groups = input_149_groups_0, pad = input_149_pad_0, pad_type = input_149_pad_type_0, strides = input_149_strides_0, weight = encoder_conv_module_3_sequential_7_conv_weight, x = input_147)[name = tensor("input_149")]; + tensor outputs_aug = conv(dilations = outputs_aug_dilations_0, groups = outputs_aug_groups_0, pad = outputs_aug_pad_0, pad_type = outputs_aug_pad_type_0, strides = outputs_aug_strides_0, weight = inner_encoder_conv_module_3_sequential_4_conv_weight, x = inputs_35)[name = tensor("outputs_aug")]; + tensor input_147_begin_0 = const()[name = tensor("input_147_begin_0"), val = tensor([0, 0, 0])]; + tensor input_147_end_0 = const()[name = tensor("input_147_end_0"), val = tensor([5, 256, 16])]; + tensor input_147_end_mask_0 = const()[name = tensor("input_147_end_mask_0"), val = tensor([true, true, false])]; + tensor input_147 = slice_by_index(begin = input_147_begin_0, end = input_147_end_0, end_mask = input_147_end_mask_0, x = outputs_aug)[name = tensor("input_147")]; + tensor inputs_37 = batch_norm(beta = inner_encoder_conv_module_3_sequential_5_bias, epsilon = var_74, gamma = inner_encoder_conv_module_3_sequential_5_weight, mean = inner_encoder_conv_module_3_sequential_5_running_mean, variance = inner_encoder_conv_module_3_sequential_5_running_var, x = input_147)[name = tensor("inputs_37")]; + tensor input_149 = silu(x = inputs_37)[name = tensor("input_149")]; + tensor input_151_pad_type_0 = const()[name = tensor("input_151_pad_type_0"), val = tensor("valid")]; + tensor input_151_strides_0 = const()[name = tensor("input_151_strides_0"), val = tensor([1])]; + tensor input_151_pad_0 = const()[name = tensor("input_151_pad_0"), val = tensor([0, 0])]; + tensor input_151_dilations_0 = const()[name = tensor("input_151_dilations_0"), val = tensor([1])]; + tensor input_151_groups_0 = const()[name = tensor("input_151_groups_0"), val = tensor(1)]; + tensor input_151 = conv(bias = inner_encoder_conv_module_3_sequential_7_conv_bias, dilations = input_151_dilations_0, groups = input_151_groups_0, pad = input_151_pad_0, pad_type = input_151_pad_type_0, strides = input_151_strides_0, weight = inner_encoder_conv_module_3_sequential_7_conv_weight, x = input_149)[name = tensor("input_151")]; tensor conv_out_7_perm_0 = const()[name = tensor("conv_out_7_perm_0"), val = tensor([0, 2, 1])]; - tensor var_981_begin_0 = const()[name = tensor("op_981_begin_0"), val = tensor([0, -1, 0])]; - tensor var_981_end_0 = const()[name = tensor("op_981_end_0"), val = tensor([5, 16, 256])]; - tensor var_981_end_mask_0 = const()[name = tensor("op_981_end_mask_0"), val = tensor([true, true, true])]; - tensor conv_out_7 = transpose(perm = conv_out_7_perm_0, x = input_149)[name = tensor("transpose_32")]; - tensor var_981 = slice_by_index(begin = var_981_begin_0, end = var_981_end_0, end_mask = var_981_end_mask_0, x = conv_out_7)[name = tensor("op_981")]; - tensor var_983_perm_0 = const()[name = tensor("op_983_perm_0"), val = tensor([1, 0, 2])]; - tensor var_983 = transpose(perm = var_983_perm_0, x = var_981)[name = tensor("transpose_31")]; - tensor input_151 = add(x = x_21, y = var_983)[name = tensor("input_151")]; - tensor input_153_axes_0 = const()[name = tensor("input_153_axes_0"), val = tensor([-1])]; - tensor input_153 = layer_norm(axes = input_153_axes_0, beta = encoder_ffn2_3_module_sequential_0_bias, epsilon = var_29, gamma = encoder_ffn2_3_module_sequential_0_weight, x = input_151)[name = tensor("input_153")]; - tensor inputs = linear(bias = encoder_ffn2_3_module_sequential_1_linear_bias, weight = encoder_ffn2_3_module_sequential_1_linear_weight, x = input_153)[name = tensor("linear_35")]; - tensor input_155 = silu(x = inputs)[name = tensor("input_155")]; - tensor input_159 = linear(bias = encoder_ffn2_3_module_sequential_4_linear_bias, weight = encoder_ffn2_3_module_sequential_4_linear_weight, x = input_155)[name = tensor("linear_36")]; - tensor var_1006 = const()[name = tensor("op_1006"), val = tensor(0x1p-1)]; - tensor var_1007 = mul(x = input_159, y = var_1006)[name = tensor("op_1007")]; - tensor input_161 = add(x = var_1007, y = input_151)[name = tensor("input_161")]; + tensor var_1046_begin_0 = const()[name = tensor("op_1046_begin_0"), val = tensor([0, -1, 0])]; + tensor var_1046_end_0 = const()[name = tensor("op_1046_end_0"), val = tensor([5, 16, 256])]; + tensor var_1046_end_mask_0 = const()[name = tensor("op_1046_end_mask_0"), val = tensor([true, true, true])]; + tensor conv_out_7 = transpose(perm = conv_out_7_perm_0, x = input_151)[name = tensor("transpose_32")]; + tensor var_1046 = slice_by_index(begin = var_1046_begin_0, end = var_1046_end_0, end_mask = var_1046_end_mask_0, x = conv_out_7)[name = tensor("op_1046")]; + tensor var_1048_perm_0 = const()[name = tensor("op_1048_perm_0"), val = tensor([1, 0, 2])]; + tensor var_1048 = transpose(perm = var_1048_perm_0, x = var_1046)[name = tensor("transpose_31")]; + tensor input_153 = add(x = x_21, y = var_1048)[name = tensor("input_153")]; + tensor input_155_axes_0 = const()[name = tensor("input_155_axes_0"), val = tensor([-1])]; + tensor input_155 = layer_norm(axes = input_155_axes_0, beta = inner_encoder_ffn2_3_module_sequential_0_bias, epsilon = var_74, gamma = inner_encoder_ffn2_3_module_sequential_0_weight, x = input_153)[name = tensor("input_155")]; + tensor inputs = linear(bias = inner_encoder_ffn2_3_module_sequential_1_linear_bias, weight = inner_encoder_ffn2_3_module_sequential_1_linear_weight, x = input_155)[name = tensor("linear_35")]; + tensor input_157 = silu(x = inputs)[name = tensor("input_157")]; + tensor input_161 = linear(bias = inner_encoder_ffn2_3_module_sequential_4_linear_bias, weight = inner_encoder_ffn2_3_module_sequential_4_linear_weight, x = input_157)[name = tensor("linear_36")]; + tensor var_1071 = const()[name = tensor("op_1071"), val = tensor(0x1p-1)]; + tensor var_1072 = mul(x = input_161, y = var_1071)[name = tensor("op_1072")]; + tensor input_163 = add(x = var_1072, y = input_153)[name = tensor("input_163")]; tensor x_25_axes_0 = const()[name = tensor("x_25_axes_0"), val = tensor([-1])]; - tensor x_25 = layer_norm(axes = x_25_axes_0, beta = encoder_layer_norm_3_bias, epsilon = var_29, gamma = encoder_layer_norm_3_weight, x = input_161)[name = tensor("x_25")]; + tensor x_25 = layer_norm(axes = x_25_axes_0, beta = inner_encoder_layer_norm_3_bias, epsilon = var_74, gamma = inner_encoder_layer_norm_3_weight, x = input_163)[name = tensor("x_25")]; tensor x_ct_perm_0 = const()[name = tensor("x_ct_perm_0"), val = tensor([0, 2, 1])]; tensor cat_interleave_0 = const()[name = tensor("cat_interleave_0"), val = tensor(false)]; tensor x_ct = transpose(perm = x_ct_perm_0, x = x_25)[name = tensor("transpose_30")]; - tensor cat = concat(axis = var_21, interleave = cat_interleave_0, values = (cnn_window, x_ct))[name = tensor("cat")]; + tensor cat = concat(axis = var_79, interleave = cat_interleave_0, values = (cnn_window, x_ct))[name = tensor("cat")]; tensor conv_out_pad_type_0 = const()[name = tensor("conv_out_pad_type_0"), val = tensor("valid")]; tensor conv_out_strides_0 = const()[name = tensor("conv_out_strides_0"), val = tensor([1])]; tensor conv_out_pad_0 = const()[name = tensor("conv_out_pad_0"), val = tensor([0, 0])]; tensor conv_out_dilations_0 = const()[name = tensor("conv_out_dilations_0"), val = tensor([1])]; tensor conv_out_groups_0 = const()[name = tensor("conv_out_groups_0"), val = tensor(1)]; - tensor conv_out = conv(bias = encoder_cnn_bias, dilations = conv_out_dilations_0, groups = conv_out_groups_0, pad = conv_out_pad_0, pad_type = conv_out_pad_type_0, strides = conv_out_strides_0, weight = encoder_cnn_weight, x = cat)[name = tensor("conv_out")]; - tensor var_1025_begin_0 = const()[name = tensor("op_1025_begin_0"), val = tensor([0, 0, 5])]; - tensor var_1025_end_0 = const()[name = tensor("op_1025_end_0"), val = tensor([1, 256, 23])]; - tensor var_1025_end_mask_0 = const()[name = tensor("op_1025_end_mask_0"), val = tensor([true, true, true])]; - tensor cnn_window_new = slice_by_index(begin = var_1025_begin_0, end = var_1025_end_0, end_mask = var_1025_end_mask_0, x = cat)[name = tensor("op_1025")]; - tensor input_163_perm_0 = const()[name = tensor("input_163_perm_0"), val = tensor([0, 2, 1])]; - tensor var_1027 = const()[name = tensor("op_1027"), val = tensor([-1])]; - tensor input_163 = transpose(perm = input_163_perm_0, x = conv_out)[name = tensor("transpose_29")]; - tensor var_1028 = reduce_l2_norm(axes = var_1027, keep_dims = var_30, x = input_163)[name = tensor("op_1028")]; + tensor conv_out = conv(bias = inner_encoder_cnn_bias, dilations = conv_out_dilations_0, groups = conv_out_groups_0, pad = conv_out_pad_0, pad_type = conv_out_pad_type_0, strides = conv_out_strides_0, weight = inner_encoder_cnn_weight, x = cat)[name = tensor("conv_out")]; + tensor var_1090_begin_0 = const()[name = tensor("op_1090_begin_0"), val = tensor([0, 0, 5])]; + tensor var_1090_end_0 = const()[name = tensor("op_1090_end_0"), val = tensor([1, 256, 23])]; + tensor var_1090_end_mask_0 = const()[name = tensor("op_1090_end_mask_0"), val = tensor([true, true, true])]; + tensor cnn_window_new = slice_by_index(begin = var_1090_begin_0, end = var_1090_end_0, end_mask = var_1090_end_mask_0, x = cat)[name = tensor("op_1090")]; + tensor input_165_perm_0 = const()[name = tensor("input_165_perm_0"), val = tensor([0, 2, 1])]; + tensor var_1092 = const()[name = tensor("op_1092"), val = tensor([-1])]; + tensor input_165 = transpose(perm = input_165_perm_0, x = conv_out)[name = tensor("transpose_29")]; + tensor var_1093 = reduce_l2_norm(axes = var_1092, keep_dims = var_73, x = input_165)[name = tensor("op_1093")]; tensor const_12 = const()[name = tensor("const_12"), val = tensor(0x1.fffffep+127)]; - tensor clip_0 = clip(alpha = var_11, beta = const_12, x = var_1028)[name = tensor("clip_0")]; - tensor emb = real_div(x = input_163, y = clip_0)[name = tensor("emb")]; - tensor var_1032_axis_0 = const()[name = tensor("op_1032_axis_0"), val = tensor(0)]; - tensor enc_kv_new = stack(axis = var_1032_axis_0, values = (var_207, var_429, var_651, nkv_1))[name = tensor("op_1032")]; - tensor var_1034_axis_0 = const()[name = tensor("op_1034_axis_0"), val = tensor(0)]; - tensor enc_scale_new = stack(axis = var_1034_axis_0, values = (new_scale_1, new_scale_3, new_scale_5, new_scale_7))[name = tensor("op_1034")]; - tensor var_1036_axis_0 = const()[name = tensor("op_1036_axis_0"), val = tensor(0)]; - tensor enc_conv_cache_new = stack(axis = var_1036_axis_0, values = (window_11, window_23, window_35, window))[name = tensor("op_1036")]; - tensor var_1045 = const()[name = tensor("op_1045"), val = tensor(0x1.5798eep-27)]; - tensor var_1050 = const()[name = tensor("op_1050"), val = tensor(0x1.4f8b58p-17)]; - tensor var_1052 = const()[name = tensor("op_1052"), val = tensor(0x1.0c6f7ap-20)]; - tensor var_1053 = const()[name = tensor("op_1053"), val = tensor(true)]; - tensor var_1055 = const()[name = tensor("op_1055"), val = tensor(0x1p+0)]; - tensor var_1059 = const()[name = tensor("op_1059"), val = tensor(-1)]; - tensor var_1065 = const()[name = tensor("op_1065"), val = tensor(0)]; + tensor clip_0 = clip(alpha = var_88, beta = const_12, x = var_1093)[name = tensor("clip_0")]; + tensor emb = real_div(x = input_165, y = clip_0)[name = tensor("emb")]; + tensor var_1097_axis_0 = const()[name = tensor("op_1097_axis_0"), val = tensor(0)]; + tensor enc_kv_new = stack(axis = var_1097_axis_0, values = (var_272, var_494, var_716, nkv_1))[name = tensor("op_1097")]; + tensor var_1099_axis_0 = const()[name = tensor("op_1099_axis_0"), val = tensor(0)]; + tensor enc_scale_new = stack(axis = var_1099_axis_0, values = (new_scale_1, new_scale_3, new_scale_5, new_scale_7))[name = tensor("op_1099")]; + tensor var_1101_axis_0 = const()[name = tensor("op_1101_axis_0"), val = tensor(0)]; + tensor enc_conv_cache_new = stack(axis = var_1101_axis_0, values = (window_11, window_23, window_35, window))[name = tensor("op_1101")]; tensor pos = const()[name = tensor("pos"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(44395712)))]; - tensor var_1127_axes_0 = const()[name = tensor("op_1127_axes_0"), val = tensor([2])]; - tensor var_1127 = expand_dims(axes = var_1127_axes_0, x = emb)[name = tensor("op_1127")]; + tensor var_1169_axes_0 = const()[name = tensor("op_1169_axes_0"), val = tensor([2])]; + tensor var_1169 = expand_dims(axes = var_1169_axes_0, x = emb)[name = tensor("op_1169")]; tensor emb_exp_reps_0 = const()[name = tensor("emb_exp_reps_0"), val = tensor([1, 1, 6, 1])]; - tensor emb_exp = tile(reps = emb_exp_reps_0, x = var_1127)[name = tensor("emb_exp")]; - tensor input_165_interleave_0 = const()[name = tensor("input_165_interleave_0"), val = tensor(false)]; - tensor input_165 = concat(axis = var_1059, interleave = input_165_interleave_0, values = (emb_exp, pos))[name = tensor("input_165")]; - tensor x_27 = linear(bias = decoder_convert_bias, weight = decoder_convert_weight, x = input_165)[name = tensor("linear_37")]; - tensor var_1135_perm_0 = const()[name = tensor("op_1135_perm_0"), val = tensor([0, 2, 1, 3])]; - tensor var_1139 = const()[name = tensor("op_1139"), val = tensor([6, 5, 256])]; - tensor var_1135 = transpose(perm = var_1135_perm_0, x = x_27)[name = tensor("transpose_28")]; - tensor x_29 = reshape(shape = var_1139, x = var_1135)[name = tensor("x_29")]; + tensor emb_exp = tile(reps = emb_exp_reps_0, x = var_1169)[name = tensor("emb_exp")]; + tensor input_167_interleave_0 = const()[name = tensor("input_167_interleave_0"), val = tensor(false)]; + tensor input_167 = concat(axis = var_80, interleave = input_167_interleave_0, values = (emb_exp, pos))[name = tensor("input_167")]; + tensor x_27 = linear(bias = inner_decoder_convert_bias, weight = inner_decoder_convert_weight, x = input_167)[name = tensor("linear_37")]; + tensor var_1177_perm_0 = const()[name = tensor("op_1177_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_1181 = const()[name = tensor("op_1181"), val = tensor([6, 5, 256])]; + tensor var_1177 = transpose(perm = var_1177_perm_0, x = x_27)[name = tensor("transpose_28")]; + tensor x_29 = reshape(shape = var_1181, x = var_1177)[name = tensor("x_29")]; tensor prev_kv_9_begin_0 = const()[name = tensor("prev_kv_9_begin_0"), val = tensor([0, 0, 0, 0, 0])]; tensor prev_kv_9_end_0 = const()[name = tensor("prev_kv_9_end_0"), val = tensor([1, 6, 4, 64, 64])]; tensor prev_kv_9_end_mask_0 = const()[name = tensor("prev_kv_9_end_mask_0"), val = tensor([false, true, true, true, true])]; @@ -1023,132 +1042,132 @@ program(1.0) tensor prev_scale_9_end_mask_0 = const()[name = tensor("prev_scale_9_end_mask_0"), val = tensor([false, true])]; tensor prev_scale_9_squeeze_mask_0 = const()[name = tensor("prev_scale_9_squeeze_mask_0"), val = tensor([true, false])]; tensor prev_scale_9 = slice_by_index(begin = prev_scale_9_begin_0, end = prev_scale_9_end_0, end_mask = prev_scale_9_end_mask_0, squeeze_mask = prev_scale_9_squeeze_mask_0, x = dec_scale)[name = tensor("prev_scale_9")]; - tensor var_1147 = linear(bias = decoder_q_proj_0_bias, weight = decoder_q_proj_0_weight, x = x_29)[name = tensor("linear_38")]; - tensor var_1148 = const()[name = tensor("op_1148"), val = tensor([6, 5, 4, 64])]; - tensor var_1149 = reshape(shape = var_1148, x = var_1147)[name = tensor("op_1149")]; + tensor var_1189 = linear(bias = inner_decoder_q_proj_0_bias, weight = inner_decoder_q_proj_0_weight, x = x_29)[name = tensor("linear_38")]; + tensor var_1190 = const()[name = tensor("op_1190"), val = tensor([6, 5, 4, 64])]; + tensor var_1191 = reshape(shape = var_1190, x = var_1189)[name = tensor("op_1191")]; tensor q_9_perm_0 = const()[name = tensor("q_9_perm_0"), val = tensor([0, 2, 1, 3])]; - tensor var_1153 = linear(bias = decoder_k_proj_0_bias, weight = decoder_k_proj_0_weight, x = x_29)[name = tensor("linear_39")]; - tensor var_1154 = const()[name = tensor("op_1154"), val = tensor(0x1p-3)]; - tensor var_1155 = mul(x = var_1153, y = var_1154)[name = tensor("op_1155")]; - tensor var_1156 = const()[name = tensor("op_1156"), val = tensor([6, 5, 4, 64])]; - tensor var_1157 = reshape(shape = var_1156, x = var_1155)[name = tensor("op_1157")]; + tensor var_1195 = linear(bias = inner_decoder_k_proj_0_bias, weight = inner_decoder_k_proj_0_weight, x = x_29)[name = tensor("linear_39")]; + tensor var_1196 = const()[name = tensor("op_1196"), val = tensor(0x1p-3)]; + tensor var_1197 = mul(x = var_1195, y = var_1196)[name = tensor("op_1197")]; + tensor var_1198 = const()[name = tensor("op_1198"), val = tensor([6, 5, 4, 64])]; + tensor var_1199 = reshape(shape = var_1198, x = var_1197)[name = tensor("op_1199")]; tensor k_9_perm_0 = const()[name = tensor("k_9_perm_0"), val = tensor([0, 2, 1, 3])]; - tensor var_1161 = linear(bias = decoder_v_proj_0_bias, weight = decoder_v_proj_0_weight, x = x_29)[name = tensor("linear_40")]; - tensor var_1162 = const()[name = tensor("op_1162"), val = tensor([6, 5, 4, 64])]; - tensor var_1163 = reshape(shape = var_1162, x = var_1161)[name = tensor("op_1163")]; + tensor var_1203 = linear(bias = inner_decoder_v_proj_0_bias, weight = inner_decoder_v_proj_0_weight, x = x_29)[name = tensor("linear_40")]; + tensor var_1204 = const()[name = tensor("op_1204"), val = tensor([6, 5, 4, 64])]; + tensor var_1205 = reshape(shape = var_1204, x = var_1203)[name = tensor("op_1205")]; tensor v_9_perm_0 = const()[name = tensor("v_9_perm_0"), val = tensor([0, 2, 1, 3])]; - tensor input_169 = linear(bias = decoder_g_proj_0_bias, weight = decoder_g_proj_0_weight, x = x_29)[name = tensor("linear_41")]; + tensor input_171 = linear(bias = inner_decoder_g_proj_0_bias, weight = inner_decoder_g_proj_0_weight, x = x_29)[name = tensor("linear_41")]; tensor cumsum_mask_1_exclusive_0 = const()[name = tensor("cumsum_mask_1_exclusive_0"), val = tensor(false)]; tensor cumsum_mask_1_reverse_0 = const()[name = tensor("cumsum_mask_1_reverse_0"), val = tensor(false)]; - tensor cumsum_mask_1 = cumsum(axis = var_1065, exclusive = cumsum_mask_1_exclusive_0, reverse = cumsum_mask_1_reverse_0, x = valid_mask)[name = tensor("cumsum_mask_1")]; + tensor cumsum_mask_1 = cumsum(axis = var_77, exclusive = cumsum_mask_1_exclusive_0, reverse = cumsum_mask_1_reverse_0, x = valid_mask)[name = tensor("cumsum_mask_1")]; tensor sqrt_s0_9 = sqrt(x = prev_scale_9)[name = tensor("sqrt_s0_9")]; tensor s_t_9 = add(x = prev_scale_9, y = cumsum_mask_1)[name = tensor("s_t_9")]; tensor const_20 = const()[name = tensor("const_20"), val = tensor(0x1.fffffep+127)]; - tensor clip_1 = clip(alpha = var_1055, beta = const_20, x = s_t_9)[name = tensor("clip_1")]; + tensor clip_1 = clip(alpha = var_68, beta = const_20, x = s_t_9)[name = tensor("clip_1")]; tensor sqrt_s_t_9 = sqrt(x = clip_1)[name = tensor("sqrt_s_t_9")]; tensor qk_9_transpose_x_1 = const()[name = tensor("qk_9_transpose_x_1"), val = tensor(false)]; tensor qk_9_transpose_y_1 = const()[name = tensor("qk_9_transpose_y_1"), val = tensor(true)]; - tensor k_9 = transpose(perm = k_9_perm_0, x = var_1157)[name = tensor("transpose_26")]; - tensor q_9 = transpose(perm = q_9_perm_0, x = var_1149)[name = tensor("transpose_27")]; + tensor k_9 = transpose(perm = k_9_perm_0, x = var_1199)[name = tensor("transpose_26")]; + tensor q_9 = transpose(perm = q_9_perm_0, x = var_1191)[name = tensor("transpose_27")]; tensor qk_9 = matmul(transpose_x = qk_9_transpose_x_1, transpose_y = qk_9_transpose_y_1, x = q_9, y = k_9)[name = tensor("qk_9")]; - tensor var_1175 = const()[name = tensor("op_1175"), val = tensor([1, 5])]; - tensor var_1176 = reshape(shape = var_1175, x = valid_mask)[name = tensor("op_1176")]; - tensor causal_with_valid_1 = mul(x = decoder__causal_mask, y = var_1176)[name = tensor("causal_with_valid_1")]; - tensor var_1178 = const()[name = tensor("op_1178"), val = tensor([5, 1])]; - tensor var_1179 = reshape(shape = var_1178, x = sqrt_s_t_9)[name = tensor("op_1179")]; - tensor M_9 = real_div(x = causal_with_valid_1, y = var_1179)[name = tensor("M_9")]; - tensor var_1181 = mul(x = qk_9, y = M_9)[name = tensor("op_1181")]; + tensor var_1217 = const()[name = tensor("op_1217"), val = tensor([1, 5])]; + tensor var_1218 = reshape(shape = var_1217, x = valid_mask)[name = tensor("op_1218")]; + tensor causal_with_valid_1 = mul(x = inner_decoder__causal_mask, y = var_1218)[name = tensor("causal_with_valid_1")]; + tensor var_1220 = const()[name = tensor("op_1220"), val = tensor([5, 1])]; + tensor var_1221 = reshape(shape = var_1220, x = sqrt_s_t_9)[name = tensor("op_1221")]; + tensor M_9 = real_div(x = causal_with_valid_1, y = var_1221)[name = tensor("M_9")]; + tensor var_1223 = mul(x = qk_9, y = M_9)[name = tensor("op_1223")]; tensor inner_9_transpose_x_0 = const()[name = tensor("inner_9_transpose_x_0"), val = tensor(false)]; tensor inner_9_transpose_y_0 = const()[name = tensor("inner_9_transpose_y_0"), val = tensor(false)]; - tensor v_9 = transpose(perm = v_9_perm_0, x = var_1163)[name = tensor("transpose_25")]; - tensor inner_9 = matmul(transpose_x = inner_9_transpose_x_0, transpose_y = inner_9_transpose_y_0, x = var_1181, y = v_9)[name = tensor("inner_9")]; - tensor var_1183_transpose_x_0 = const()[name = tensor("op_1183_transpose_x_0"), val = tensor(false)]; - tensor var_1183_transpose_y_0 = const()[name = tensor("op_1183_transpose_y_0"), val = tensor(false)]; - tensor var_1183 = matmul(transpose_x = var_1183_transpose_x_0, transpose_y = var_1183_transpose_y_0, x = q_9, y = prev_kv_9)[name = tensor("op_1183")]; - tensor var_1184 = real_div(x = sqrt_s0_9, y = sqrt_s_t_9)[name = tensor("op_1184")]; - tensor var_1185 = const()[name = tensor("op_1185"), val = tensor([1, 1, 5, 1])]; - tensor var_1186 = reshape(shape = var_1185, x = var_1184)[name = tensor("op_1186")]; - tensor cross_9 = mul(x = var_1183, y = var_1186)[name = tensor("cross_9")]; + tensor v_9 = transpose(perm = v_9_perm_0, x = var_1205)[name = tensor("transpose_25")]; + tensor inner_9 = matmul(transpose_x = inner_9_transpose_x_0, transpose_y = inner_9_transpose_y_0, x = var_1223, y = v_9)[name = tensor("inner_9")]; + tensor var_1225_transpose_x_0 = const()[name = tensor("op_1225_transpose_x_0"), val = tensor(false)]; + tensor var_1225_transpose_y_0 = const()[name = tensor("op_1225_transpose_y_0"), val = tensor(false)]; + tensor var_1225 = matmul(transpose_x = var_1225_transpose_x_0, transpose_y = var_1225_transpose_y_0, x = q_9, y = prev_kv_9)[name = tensor("op_1225")]; + tensor var_1226 = real_div(x = sqrt_s0_9, y = sqrt_s_t_9)[name = tensor("op_1226")]; + tensor var_1227 = const()[name = tensor("op_1227"), val = tensor([1, 1, 5, 1])]; + tensor var_1228 = reshape(shape = var_1227, x = var_1226)[name = tensor("op_1228")]; + tensor cross_9 = mul(x = var_1225, y = var_1228)[name = tensor("cross_9")]; tensor out_25 = add(x = inner_9, y = cross_9)[name = tensor("out_25")]; - tensor var_1189 = const()[name = tensor("op_1189"), val = tensor([1, 1, 5, 1])]; - tensor var_1190 = reshape(shape = var_1189, x = valid_mask)[name = tensor("op_1190")]; - tensor v_masked_1 = mul(x = v_9, y = var_1190)[name = tensor("v_masked_1")]; - tensor var_1192 = mul(x = prev_kv_9, y = sqrt_s0_9)[name = tensor("op_1192")]; - tensor var_1194_transpose_x_1 = const()[name = tensor("op_1194_transpose_x_1"), val = tensor(true)]; - tensor var_1194_transpose_y_1 = const()[name = tensor("op_1194_transpose_y_1"), val = tensor(false)]; - tensor var_1194 = matmul(transpose_x = var_1194_transpose_x_1, transpose_y = var_1194_transpose_y_1, x = k_9, y = v_masked_1)[name = tensor("op_1194")]; - tensor new_kv_unnorm_9 = add(x = var_1192, y = var_1194)[name = tensor("new_kv_unnorm_9")]; - tensor var_1196_keep_dims_0 = const()[name = tensor("op_1196_keep_dims_0"), val = tensor(false)]; - tensor var_1196 = reduce_sum(keep_dims = var_1196_keep_dims_0, x = valid_mask)[name = tensor("op_1196")]; - tensor var_1197 = const()[name = tensor("op_1197"), val = tensor([1])]; - tensor var_1198 = reshape(shape = var_1197, x = var_1196)[name = tensor("op_1198")]; - tensor new_scale_9 = add(x = prev_scale_9, y = var_1198)[name = tensor("new_scale_9")]; + tensor var_1231 = const()[name = tensor("op_1231"), val = tensor([1, 1, 5, 1])]; + tensor var_1232 = reshape(shape = var_1231, x = valid_mask)[name = tensor("op_1232")]; + tensor v_masked_1 = mul(x = v_9, y = var_1232)[name = tensor("v_masked_1")]; + tensor var_1234 = mul(x = prev_kv_9, y = sqrt_s0_9)[name = tensor("op_1234")]; + tensor var_1236_transpose_x_1 = const()[name = tensor("op_1236_transpose_x_1"), val = tensor(true)]; + tensor var_1236_transpose_y_1 = const()[name = tensor("op_1236_transpose_y_1"), val = tensor(false)]; + tensor var_1236 = matmul(transpose_x = var_1236_transpose_x_1, transpose_y = var_1236_transpose_y_1, x = k_9, y = v_masked_1)[name = tensor("op_1236")]; + tensor new_kv_unnorm_9 = add(x = var_1234, y = var_1236)[name = tensor("new_kv_unnorm_9")]; + tensor var_1238_keep_dims_0 = const()[name = tensor("op_1238_keep_dims_0"), val = tensor(false)]; + tensor var_1238 = reduce_sum(keep_dims = var_1238_keep_dims_0, x = valid_mask)[name = tensor("op_1238")]; + tensor var_1239 = const()[name = tensor("op_1239"), val = tensor([1])]; + tensor var_1240 = reshape(shape = var_1239, x = var_1238)[name = tensor("op_1240")]; + tensor new_scale_9 = add(x = prev_scale_9, y = var_1240)[name = tensor("new_scale_9")]; tensor const_21 = const()[name = tensor("const_21"), val = tensor(0x1.fffffep+127)]; - tensor clip_2 = clip(alpha = var_1055, beta = const_21, x = new_scale_9)[name = tensor("clip_2")]; + tensor clip_2 = clip(alpha = var_68, beta = const_21, x = new_scale_9)[name = tensor("clip_2")]; tensor sqrt_new_scale_1 = sqrt(x = clip_2)[name = tensor("sqrt_new_scale_1")]; - tensor var_1202 = real_div(x = new_kv_unnorm_9, y = sqrt_new_scale_1)[name = tensor("op_1202")]; - tensor var_1203_perm_0 = const()[name = tensor("op_1203_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_1244 = real_div(x = new_kv_unnorm_9, y = sqrt_new_scale_1)[name = tensor("op_1244")]; + tensor var_1245_perm_0 = const()[name = tensor("op_1245_perm_0"), val = tensor([0, 2, 1, 3])]; tensor out_27_axes_0 = const()[name = tensor("out_27_axes_0"), val = tensor([-1])]; - tensor var_1203 = transpose(perm = var_1203_perm_0, x = out_25)[name = tensor("transpose_24")]; - tensor out_27 = layer_norm(axes = out_27_axes_0, epsilon = var_1052, x = var_1203)[name = tensor("out_27")]; - tensor var_1207 = const()[name = tensor("op_1207"), val = tensor([6, 5, 256])]; - tensor out_29 = reshape(shape = var_1207, x = out_27)[name = tensor("out_29")]; - tensor var_1209 = silu(x = input_169)[name = tensor("op_1209")]; - tensor input_171 = mul(x = var_1209, y = out_29)[name = tensor("input_171")]; - tensor ret_out_9 = linear(bias = decoder_out_proj_0_bias, weight = decoder_out_proj_0_weight, x = input_171)[name = tensor("linear_42")]; - tensor input_173 = add(x = x_29, y = ret_out_9)[name = tensor("input_173")]; + tensor var_1245 = transpose(perm = var_1245_perm_0, x = out_25)[name = tensor("transpose_24")]; + tensor out_27 = layer_norm(axes = out_27_axes_0, epsilon = var_82, x = var_1245)[name = tensor("out_27")]; + tensor var_1249 = const()[name = tensor("op_1249"), val = tensor([6, 5, 256])]; + tensor out_29 = reshape(shape = var_1249, x = out_27)[name = tensor("out_29")]; + tensor var_1251 = silu(x = input_171)[name = tensor("op_1251")]; + tensor input_173 = mul(x = var_1251, y = out_29)[name = tensor("input_173")]; + tensor ret_out_9 = linear(bias = inner_decoder_out_proj_0_bias, weight = inner_decoder_out_proj_0_weight, x = input_173)[name = tensor("linear_42")]; + tensor input_175 = add(x = x_29, y = ret_out_9)[name = tensor("input_175")]; tensor xt_1_axes_0 = const()[name = tensor("xt_1_axes_0"), val = tensor([-1])]; - tensor xt_1 = layer_norm(axes = xt_1_axes_0, beta = decoder_norm11_0_bias, epsilon = var_1050, gamma = decoder_norm11_0_weight, x = input_173)[name = tensor("xt_1")]; - tensor var_1219 = const()[name = tensor("op_1219"), val = tensor([1, 6, 5, 256])]; - tensor var_1220 = reshape(shape = var_1219, x = xt_1)[name = tensor("op_1220")]; - tensor var_1221_perm_0 = const()[name = tensor("op_1221_perm_0"), val = tensor([0, 2, 1, 3])]; - tensor var_1224 = const()[name = tensor("op_1224"), val = tensor([5, 6, 256])]; - tensor var_1221 = transpose(perm = var_1221_perm_0, x = var_1220)[name = tensor("transpose_23")]; - tensor query_1 = reshape(shape = var_1224, x = var_1221)[name = tensor("query_1")]; + tensor xt_1 = layer_norm(axes = xt_1_axes_0, beta = inner_decoder_norm11_0_bias, epsilon = var_74, gamma = inner_decoder_norm11_0_weight, x = input_175)[name = tensor("xt_1")]; + tensor var_1261 = const()[name = tensor("op_1261"), val = tensor([1, 6, 5, 256])]; + tensor var_1262 = reshape(shape = var_1261, x = xt_1)[name = tensor("op_1262")]; + tensor var_1263_perm_0 = const()[name = tensor("op_1263_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_1266 = const()[name = tensor("op_1266"), val = tensor([5, 6, 256])]; + tensor var_1263 = transpose(perm = var_1263_perm_0, x = var_1262)[name = tensor("transpose_23")]; + tensor query_1 = reshape(shape = var_1266, x = var_1263)[name = tensor("query_1")]; tensor query_3_perm_0 = const()[name = tensor("query_3_perm_0"), val = tensor([1, 0, 2])]; tensor query_3 = transpose(perm = query_3_perm_0, x = query_1)[name = tensor("transpose_22")]; - tensor var_1247 = linear(bias = decoder_self_attn2_0_in_proj_bias, weight = decoder_self_attn2_0_in_proj_weight, x = query_3)[name = tensor("linear_43")]; + tensor var_1289 = linear(bias = inner_decoder_self_attn2_0_in_proj_bias, weight = inner_decoder_self_attn2_0_in_proj_weight, x = query_3)[name = tensor("linear_43")]; tensor concat_1 = const()[name = tensor("concat_1"), val = tensor([6, 5, 3, 256])]; - tensor var_1249 = reshape(shape = concat_1, x = var_1247)[name = tensor("op_1249")]; - tensor var_1250_axes_0 = const()[name = tensor("op_1250_axes_0"), val = tensor([0])]; - tensor var_1250 = expand_dims(axes = var_1250_axes_0, x = var_1249)[name = tensor("op_1250")]; - tensor var_1251_perm_0 = const()[name = tensor("op_1251_perm_0"), val = tensor([-2, 1, 2, 0, 4])]; - tensor var_1252_axes_0 = const()[name = tensor("op_1252_axes_0"), val = tensor([-2])]; - tensor var_1251 = transpose(perm = var_1251_perm_0, x = var_1250)[name = tensor("transpose_21")]; - tensor var_1252 = squeeze(axes = var_1252_axes_0, x = var_1251)[name = tensor("op_1252")]; + tensor var_1291 = reshape(shape = concat_1, x = var_1289)[name = tensor("op_1291")]; + tensor var_1292_axes_0 = const()[name = tensor("op_1292_axes_0"), val = tensor([0])]; + tensor var_1292 = expand_dims(axes = var_1292_axes_0, x = var_1291)[name = tensor("op_1292")]; + tensor var_1293_perm_0 = const()[name = tensor("op_1293_perm_0"), val = tensor([-2, 1, 2, 0, 4])]; + tensor var_1294_axes_0 = const()[name = tensor("op_1294_axes_0"), val = tensor([-2])]; + tensor var_1293 = transpose(perm = var_1293_perm_0, x = var_1292)[name = tensor("transpose_21")]; + tensor var_1294 = squeeze(axes = var_1294_axes_0, x = var_1293)[name = tensor("op_1294")]; tensor q_11_begin_0 = const()[name = tensor("q_11_begin_0"), val = tensor([0, 0, 0, 0])]; tensor q_11_end_0 = const()[name = tensor("q_11_end_0"), val = tensor([1, 6, 5, 256])]; tensor q_11_end_mask_0 = const()[name = tensor("q_11_end_mask_0"), val = tensor([false, true, true, true])]; tensor q_11_squeeze_mask_0 = const()[name = tensor("q_11_squeeze_mask_0"), val = tensor([true, false, false, false])]; - tensor q_11 = slice_by_index(begin = q_11_begin_0, end = q_11_end_0, end_mask = q_11_end_mask_0, squeeze_mask = q_11_squeeze_mask_0, x = var_1252)[name = tensor("q_11")]; + tensor q_11 = slice_by_index(begin = q_11_begin_0, end = q_11_end_0, end_mask = q_11_end_mask_0, squeeze_mask = q_11_squeeze_mask_0, x = var_1294)[name = tensor("q_11")]; tensor k_11_begin_0 = const()[name = tensor("k_11_begin_0"), val = tensor([1, 0, 0, 0])]; tensor k_11_end_0 = const()[name = tensor("k_11_end_0"), val = tensor([2, 6, 5, 256])]; tensor k_11_end_mask_0 = const()[name = tensor("k_11_end_mask_0"), val = tensor([false, true, true, true])]; tensor k_11_squeeze_mask_0 = const()[name = tensor("k_11_squeeze_mask_0"), val = tensor([true, false, false, false])]; - tensor k_11 = slice_by_index(begin = k_11_begin_0, end = k_11_end_0, end_mask = k_11_end_mask_0, squeeze_mask = k_11_squeeze_mask_0, x = var_1252)[name = tensor("k_11")]; + tensor k_11 = slice_by_index(begin = k_11_begin_0, end = k_11_end_0, end_mask = k_11_end_mask_0, squeeze_mask = k_11_squeeze_mask_0, x = var_1294)[name = tensor("k_11")]; tensor v_11_begin_0 = const()[name = tensor("v_11_begin_0"), val = tensor([2, 0, 0, 0])]; tensor v_11_end_0 = const()[name = tensor("v_11_end_0"), val = tensor([3, 6, 5, 256])]; tensor v_11_end_mask_0 = const()[name = tensor("v_11_end_mask_0"), val = tensor([false, true, true, true])]; tensor v_11_squeeze_mask_0 = const()[name = tensor("v_11_squeeze_mask_0"), val = tensor([true, false, false, false])]; - tensor v_11 = slice_by_index(begin = v_11_begin_0, end = v_11_end_0, end_mask = v_11_end_mask_0, squeeze_mask = v_11_squeeze_mask_0, x = var_1252)[name = tensor("v_11")]; - tensor var_1260 = const()[name = tensor("op_1260"), val = tensor([6, 20, 64])]; - tensor var_1261 = reshape(shape = var_1260, x = q_11)[name = tensor("op_1261")]; + tensor v_11 = slice_by_index(begin = v_11_begin_0, end = v_11_end_0, end_mask = v_11_end_mask_0, squeeze_mask = v_11_squeeze_mask_0, x = var_1294)[name = tensor("v_11")]; + tensor var_1302 = const()[name = tensor("op_1302"), val = tensor([6, 20, 64])]; + tensor var_1303 = reshape(shape = var_1302, x = q_11)[name = tensor("op_1303")]; tensor q_13_perm_0 = const()[name = tensor("q_13_perm_0"), val = tensor([1, 0, 2])]; - tensor var_1267 = const()[name = tensor("op_1267"), val = tensor([6, 20, 64])]; - tensor var_1268 = reshape(shape = var_1267, x = k_11)[name = tensor("op_1268")]; + tensor var_1309 = const()[name = tensor("op_1309"), val = tensor([6, 20, 64])]; + tensor var_1310 = reshape(shape = var_1309, x = k_11)[name = tensor("op_1310")]; tensor k_13_perm_0 = const()[name = tensor("k_13_perm_0"), val = tensor([1, 0, 2])]; - tensor var_1274 = const()[name = tensor("op_1274"), val = tensor([6, 20, 64])]; - tensor var_1275 = reshape(shape = var_1274, x = v_11)[name = tensor("op_1275")]; + tensor var_1316 = const()[name = tensor("op_1316"), val = tensor([6, 20, 64])]; + tensor var_1317 = reshape(shape = var_1316, x = v_11)[name = tensor("op_1317")]; tensor v_13_perm_0 = const()[name = tensor("v_13_perm_0"), val = tensor([1, 0, 2])]; - tensor var_1278 = const()[name = tensor("op_1278"), val = tensor([5, 4, 6, 64])]; - tensor q_13 = transpose(perm = q_13_perm_0, x = var_1261)[name = tensor("transpose_20")]; - tensor q_15 = reshape(shape = var_1278, x = q_13)[name = tensor("q_15")]; - tensor var_1280 = const()[name = tensor("op_1280"), val = tensor([5, 4, 6, 64])]; - tensor k_13 = transpose(perm = k_13_perm_0, x = var_1268)[name = tensor("transpose_19")]; - tensor k_15 = reshape(shape = var_1280, x = k_13)[name = tensor("k_15")]; - tensor var_1282 = const()[name = tensor("op_1282"), val = tensor([5, 4, 6, 64])]; - tensor v_13 = transpose(perm = v_13_perm_0, x = var_1275)[name = tensor("transpose_18")]; - tensor v_15 = reshape(shape = var_1282, x = v_13)[name = tensor("v_15")]; + tensor var_1320 = const()[name = tensor("op_1320"), val = tensor([5, 4, 6, 64])]; + tensor q_13 = transpose(perm = q_13_perm_0, x = var_1303)[name = tensor("transpose_20")]; + tensor q_15 = reshape(shape = var_1320, x = q_13)[name = tensor("q_15")]; + tensor var_1322 = const()[name = tensor("op_1322"), val = tensor([5, 4, 6, 64])]; + tensor k_13 = transpose(perm = k_13_perm_0, x = var_1310)[name = tensor("transpose_19")]; + tensor k_15 = reshape(shape = var_1322, x = k_13)[name = tensor("k_15")]; + tensor var_1324 = const()[name = tensor("op_1324"), val = tensor([5, 4, 6, 64])]; + tensor v_13 = transpose(perm = v_13_perm_0, x = var_1317)[name = tensor("transpose_18")]; + tensor v_15 = reshape(shape = var_1324, x = v_13)[name = tensor("v_15")]; tensor mul_1_y_0 = const()[name = tensor("mul_1_y_0"), val = tensor(0x1p-3)]; tensor mul_1 = mul(x = q_15, y = mul_1_y_0)[name = tensor("mul_1")]; tensor matmul_0_transpose_y_0 = const()[name = tensor("matmul_0_transpose_y_0"), val = tensor(true)]; @@ -1159,30 +1178,30 @@ program(1.0) tensor attn_output_1_transpose_x_0 = const()[name = tensor("attn_output_1_transpose_x_0"), val = tensor(false)]; tensor attn_output_1_transpose_y_0 = const()[name = tensor("attn_output_1_transpose_y_0"), val = tensor(false)]; tensor attn_output_1 = matmul(transpose_x = attn_output_1_transpose_x_0, transpose_y = attn_output_1_transpose_y_0, x = softmax_0, y = v_15)[name = tensor("attn_output_1")]; - tensor var_1285 = const()[name = tensor("op_1285"), val = tensor([2, 0, 1, 3])]; - tensor var_1290 = const()[name = tensor("op_1290"), val = tensor([30, 256])]; - tensor var_1286 = transpose(perm = var_1285, x = attn_output_1)[name = tensor("transpose_17")]; - tensor attn_output_3 = reshape(shape = var_1290, x = var_1286)[name = tensor("attn_output_3")]; - tensor attn_output_5 = linear(bias = decoder_self_attn2_0_out_proj_bias, weight = decoder_self_attn2_0_out_proj_weight, x = attn_output_3)[name = tensor("linear_44")]; - tensor var_1294 = const()[name = tensor("op_1294"), val = tensor([6, 5, 256])]; - tensor attn_output_7 = reshape(shape = var_1294, x = attn_output_5)[name = tensor("attn_output_7")]; + tensor var_1327 = const()[name = tensor("op_1327"), val = tensor([2, 0, 1, 3])]; + tensor var_1332 = const()[name = tensor("op_1332"), val = tensor([30, 256])]; + tensor var_1328 = transpose(perm = var_1327, x = attn_output_1)[name = tensor("transpose_17")]; + tensor attn_output_3 = reshape(shape = var_1332, x = var_1328)[name = tensor("attn_output_3")]; + tensor attn_output_5 = linear(bias = inner_decoder_self_attn2_0_out_proj_bias, weight = inner_decoder_self_attn2_0_out_proj_weight, x = attn_output_3)[name = tensor("linear_44")]; + tensor var_1336 = const()[name = tensor("op_1336"), val = tensor([6, 5, 256])]; + tensor attn_output_7 = reshape(shape = var_1336, x = attn_output_5)[name = tensor("attn_output_7")]; tensor sa2_out_1_perm_0 = const()[name = tensor("sa2_out_1_perm_0"), val = tensor([1, 0, 2])]; tensor sa2_out_1 = transpose(perm = sa2_out_1_perm_0, x = attn_output_7)[name = tensor("transpose_16")]; - tensor input_175 = add(x = query_1, y = sa2_out_1)[name = tensor("input_175")]; - tensor input_177_axes_0 = const()[name = tensor("input_177_axes_0"), val = tensor([-1])]; - tensor input_177 = layer_norm(axes = input_177_axes_0, beta = decoder_norm21_0_bias, epsilon = var_1050, gamma = decoder_norm21_0_weight, x = input_175)[name = tensor("input_177")]; - tensor input_179 = linear(bias = decoder_linear1_0_bias, weight = decoder_linear1_0_weight, x = input_177)[name = tensor("linear_45")]; - tensor input_181 = relu(x = input_179)[name = tensor("input_181")]; - tensor ffn_out_1 = linear(bias = decoder_linear2_0_bias, weight = decoder_linear2_0_weight, x = input_181)[name = tensor("linear_46")]; - tensor input_183 = add(x = input_177, y = ffn_out_1)[name = tensor("input_183")]; + tensor input_177 = add(x = query_1, y = sa2_out_1)[name = tensor("input_177")]; + tensor input_179_axes_0 = const()[name = tensor("input_179_axes_0"), val = tensor([-1])]; + tensor input_179 = layer_norm(axes = input_179_axes_0, beta = inner_decoder_norm21_0_bias, epsilon = var_74, gamma = inner_decoder_norm21_0_weight, x = input_177)[name = tensor("input_179")]; + tensor input_181 = linear(bias = inner_decoder_linear1_0_bias, weight = inner_decoder_linear1_0_weight, x = input_179)[name = tensor("linear_45")]; + tensor input_183 = relu(x = input_181)[name = tensor("input_183")]; + tensor ffn_out_1 = linear(bias = inner_decoder_linear2_0_bias, weight = inner_decoder_linear2_0_weight, x = input_183)[name = tensor("linear_46")]; + tensor input_185 = add(x = input_179, y = ffn_out_1)[name = tensor("input_185")]; tensor xt_3_axes_0 = const()[name = tensor("xt_3_axes_0"), val = tensor([-1])]; - tensor xt_3 = layer_norm(axes = xt_3_axes_0, beta = decoder_norm22_0_bias, epsilon = var_1050, gamma = decoder_norm22_0_weight, x = input_183)[name = tensor("xt_3")]; - tensor var_1314 = const()[name = tensor("op_1314"), val = tensor([1, 5, 6, 256])]; - tensor x_31 = reshape(shape = var_1314, x = xt_3)[name = tensor("x_31")]; - tensor var_1316_perm_0 = const()[name = tensor("op_1316_perm_0"), val = tensor([0, 2, 1, 3])]; - tensor var_1320 = const()[name = tensor("op_1320"), val = tensor([6, 5, 256])]; - tensor var_1316 = transpose(perm = var_1316_perm_0, x = x_31)[name = tensor("transpose_15")]; - tensor x = reshape(shape = var_1320, x = var_1316)[name = tensor("x")]; + tensor xt_3 = layer_norm(axes = xt_3_axes_0, beta = inner_decoder_norm22_0_bias, epsilon = var_74, gamma = inner_decoder_norm22_0_weight, x = input_185)[name = tensor("xt_3")]; + tensor var_1356 = const()[name = tensor("op_1356"), val = tensor([1, 5, 6, 256])]; + tensor x_31 = reshape(shape = var_1356, x = xt_3)[name = tensor("x_31")]; + tensor var_1358_perm_0 = const()[name = tensor("op_1358_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_1362 = const()[name = tensor("op_1362"), val = tensor([6, 5, 256])]; + tensor var_1358 = transpose(perm = var_1358_perm_0, x = x_31)[name = tensor("transpose_15")]; + tensor x = reshape(shape = var_1362, x = var_1358)[name = tensor("x")]; tensor prev_kv_begin_0 = const()[name = tensor("prev_kv_begin_0"), val = tensor([1, 0, 0, 0, 0])]; tensor prev_kv_end_0 = const()[name = tensor("prev_kv_end_0"), val = tensor([2, 6, 4, 64, 64])]; tensor prev_kv_end_mask_0 = const()[name = tensor("prev_kv_end_mask_0"), val = tensor([false, true, true, true, true])]; @@ -1193,120 +1212,120 @@ program(1.0) tensor prev_scale_end_mask_0 = const()[name = tensor("prev_scale_end_mask_0"), val = tensor([false, true])]; tensor prev_scale_squeeze_mask_0 = const()[name = tensor("prev_scale_squeeze_mask_0"), val = tensor([true, false])]; tensor prev_scale = slice_by_index(begin = prev_scale_begin_0, end = prev_scale_end_0, end_mask = prev_scale_end_mask_0, squeeze_mask = prev_scale_squeeze_mask_0, x = dec_scale)[name = tensor("prev_scale")]; - tensor var_1328 = linear(bias = decoder_q_proj_1_bias, weight = decoder_q_proj_1_weight, x = x)[name = tensor("linear_47")]; - tensor var_1329 = const()[name = tensor("op_1329"), val = tensor([6, 5, 4, 64])]; - tensor var_1330 = reshape(shape = var_1329, x = var_1328)[name = tensor("op_1330")]; + tensor var_1370 = linear(bias = inner_decoder_q_proj_1_bias, weight = inner_decoder_q_proj_1_weight, x = x)[name = tensor("linear_47")]; + tensor var_1371 = const()[name = tensor("op_1371"), val = tensor([6, 5, 4, 64])]; + tensor var_1372 = reshape(shape = var_1371, x = var_1370)[name = tensor("op_1372")]; tensor q_17_perm_0 = const()[name = tensor("q_17_perm_0"), val = tensor([0, 2, 1, 3])]; - tensor var_1334 = linear(bias = decoder_k_proj_1_bias, weight = decoder_k_proj_1_weight, x = x)[name = tensor("linear_48")]; - tensor var_1335 = const()[name = tensor("op_1335"), val = tensor(0x1p-3)]; - tensor var_1336 = mul(x = var_1334, y = var_1335)[name = tensor("op_1336")]; - tensor var_1337 = const()[name = tensor("op_1337"), val = tensor([6, 5, 4, 64])]; - tensor var_1338 = reshape(shape = var_1337, x = var_1336)[name = tensor("op_1338")]; + tensor var_1376 = linear(bias = inner_decoder_k_proj_1_bias, weight = inner_decoder_k_proj_1_weight, x = x)[name = tensor("linear_48")]; + tensor var_1377 = const()[name = tensor("op_1377"), val = tensor(0x1p-3)]; + tensor var_1378 = mul(x = var_1376, y = var_1377)[name = tensor("op_1378")]; + tensor var_1379 = const()[name = tensor("op_1379"), val = tensor([6, 5, 4, 64])]; + tensor var_1380 = reshape(shape = var_1379, x = var_1378)[name = tensor("op_1380")]; tensor k_17_perm_0 = const()[name = tensor("k_17_perm_0"), val = tensor([0, 2, 1, 3])]; - tensor var_1342 = linear(bias = decoder_v_proj_1_bias, weight = decoder_v_proj_1_weight, x = x)[name = tensor("linear_49")]; - tensor var_1343 = const()[name = tensor("op_1343"), val = tensor([6, 5, 4, 64])]; - tensor var_1344 = reshape(shape = var_1343, x = var_1342)[name = tensor("op_1344")]; + tensor var_1384 = linear(bias = inner_decoder_v_proj_1_bias, weight = inner_decoder_v_proj_1_weight, x = x)[name = tensor("linear_49")]; + tensor var_1385 = const()[name = tensor("op_1385"), val = tensor([6, 5, 4, 64])]; + tensor var_1386 = reshape(shape = var_1385, x = var_1384)[name = tensor("op_1386")]; tensor v_17_perm_0 = const()[name = tensor("v_17_perm_0"), val = tensor([0, 2, 1, 3])]; - tensor input_187 = linear(bias = decoder_g_proj_1_bias, weight = decoder_g_proj_1_weight, x = x)[name = tensor("linear_50")]; + tensor input_189 = linear(bias = inner_decoder_g_proj_1_bias, weight = inner_decoder_g_proj_1_weight, x = x)[name = tensor("linear_50")]; tensor sqrt_s0 = sqrt(x = prev_scale)[name = tensor("sqrt_s0")]; tensor s_t = add(x = prev_scale, y = cumsum_mask_1)[name = tensor("s_t")]; tensor const_32 = const()[name = tensor("const_32"), val = tensor(0x1.fffffep+127)]; - tensor clip_3 = clip(alpha = var_1055, beta = const_32, x = s_t)[name = tensor("clip_3")]; + tensor clip_3 = clip(alpha = var_68, beta = const_32, x = s_t)[name = tensor("clip_3")]; tensor sqrt_s_t = sqrt(x = clip_3)[name = tensor("sqrt_s_t")]; tensor qk_transpose_x_1 = const()[name = tensor("qk_transpose_x_1"), val = tensor(false)]; tensor qk_transpose_y_1 = const()[name = tensor("qk_transpose_y_1"), val = tensor(true)]; - tensor k_17 = transpose(perm = k_17_perm_0, x = var_1338)[name = tensor("transpose_13")]; - tensor q_17 = transpose(perm = q_17_perm_0, x = var_1330)[name = tensor("transpose_14")]; + tensor k_17 = transpose(perm = k_17_perm_0, x = var_1380)[name = tensor("transpose_13")]; + tensor q_17 = transpose(perm = q_17_perm_0, x = var_1372)[name = tensor("transpose_14")]; tensor qk = matmul(transpose_x = qk_transpose_x_1, transpose_y = qk_transpose_y_1, x = q_17, y = k_17)[name = tensor("qk")]; - tensor var_1359 = const()[name = tensor("op_1359"), val = tensor([5, 1])]; - tensor var_1360 = reshape(shape = var_1359, x = sqrt_s_t)[name = tensor("op_1360")]; - tensor M = real_div(x = causal_with_valid_1, y = var_1360)[name = tensor("M")]; - tensor var_1362 = mul(x = qk, y = M)[name = tensor("op_1362")]; - tensor inner_transpose_x_0 = const()[name = tensor("inner_transpose_x_0"), val = tensor(false)]; - tensor inner_transpose_y_0 = const()[name = tensor("inner_transpose_y_0"), val = tensor(false)]; - tensor v_17 = transpose(perm = v_17_perm_0, x = var_1344)[name = tensor("transpose_12")]; - tensor inner = matmul(transpose_x = inner_transpose_x_0, transpose_y = inner_transpose_y_0, x = var_1362, y = v_17)[name = tensor("inner")]; - tensor var_1364_transpose_x_0 = const()[name = tensor("op_1364_transpose_x_0"), val = tensor(false)]; - tensor var_1364_transpose_y_0 = const()[name = tensor("op_1364_transpose_y_0"), val = tensor(false)]; - tensor var_1364 = matmul(transpose_x = var_1364_transpose_x_0, transpose_y = var_1364_transpose_y_0, x = q_17, y = prev_kv)[name = tensor("op_1364")]; - tensor var_1365 = real_div(x = sqrt_s0, y = sqrt_s_t)[name = tensor("op_1365")]; - tensor var_1366 = const()[name = tensor("op_1366"), val = tensor([1, 1, 5, 1])]; - tensor var_1367 = reshape(shape = var_1366, x = var_1365)[name = tensor("op_1367")]; - tensor cross = mul(x = var_1364, y = var_1367)[name = tensor("cross")]; - tensor out_31 = add(x = inner, y = cross)[name = tensor("out_31")]; - tensor v_masked = mul(x = v_17, y = var_1190)[name = tensor("v_masked")]; - tensor var_1373 = mul(x = prev_kv, y = sqrt_s0)[name = tensor("op_1373")]; - tensor var_1375_transpose_x_1 = const()[name = tensor("op_1375_transpose_x_1"), val = tensor(true)]; - tensor var_1375_transpose_y_1 = const()[name = tensor("op_1375_transpose_y_1"), val = tensor(false)]; - tensor var_1375 = matmul(transpose_x = var_1375_transpose_x_1, transpose_y = var_1375_transpose_y_1, x = k_17, y = v_masked)[name = tensor("op_1375")]; - tensor new_kv_unnorm = add(x = var_1373, y = var_1375)[name = tensor("new_kv_unnorm")]; - tensor new_scale = add(x = prev_scale, y = var_1198)[name = tensor("new_scale")]; + tensor var_1401 = const()[name = tensor("op_1401"), val = tensor([5, 1])]; + tensor var_1402 = reshape(shape = var_1401, x = sqrt_s_t)[name = tensor("op_1402")]; + tensor M = real_div(x = causal_with_valid_1, y = var_1402)[name = tensor("M")]; + tensor var_1404 = mul(x = qk, y = M)[name = tensor("op_1404")]; + tensor inner_11_transpose_x_0 = const()[name = tensor("inner_11_transpose_x_0"), val = tensor(false)]; + tensor inner_11_transpose_y_0 = const()[name = tensor("inner_11_transpose_y_0"), val = tensor(false)]; + tensor v_17 = transpose(perm = v_17_perm_0, x = var_1386)[name = tensor("transpose_12")]; + tensor inner_11 = matmul(transpose_x = inner_11_transpose_x_0, transpose_y = inner_11_transpose_y_0, x = var_1404, y = v_17)[name = tensor("inner_11")]; + tensor var_1406_transpose_x_0 = const()[name = tensor("op_1406_transpose_x_0"), val = tensor(false)]; + tensor var_1406_transpose_y_0 = const()[name = tensor("op_1406_transpose_y_0"), val = tensor(false)]; + tensor var_1406 = matmul(transpose_x = var_1406_transpose_x_0, transpose_y = var_1406_transpose_y_0, x = q_17, y = prev_kv)[name = tensor("op_1406")]; + tensor var_1407 = real_div(x = sqrt_s0, y = sqrt_s_t)[name = tensor("op_1407")]; + tensor var_1408 = const()[name = tensor("op_1408"), val = tensor([1, 1, 5, 1])]; + tensor var_1409 = reshape(shape = var_1408, x = var_1407)[name = tensor("op_1409")]; + tensor cross = mul(x = var_1406, y = var_1409)[name = tensor("cross")]; + tensor out_31 = add(x = inner_11, y = cross)[name = tensor("out_31")]; + tensor v_masked = mul(x = v_17, y = var_1232)[name = tensor("v_masked")]; + tensor var_1415 = mul(x = prev_kv, y = sqrt_s0)[name = tensor("op_1415")]; + tensor var_1417_transpose_x_1 = const()[name = tensor("op_1417_transpose_x_1"), val = tensor(true)]; + tensor var_1417_transpose_y_1 = const()[name = tensor("op_1417_transpose_y_1"), val = tensor(false)]; + tensor var_1417 = matmul(transpose_x = var_1417_transpose_x_1, transpose_y = var_1417_transpose_y_1, x = k_17, y = v_masked)[name = tensor("op_1417")]; + tensor new_kv_unnorm = add(x = var_1415, y = var_1417)[name = tensor("new_kv_unnorm")]; + tensor new_scale = add(x = prev_scale, y = var_1240)[name = tensor("new_scale")]; tensor const_33 = const()[name = tensor("const_33"), val = tensor(0x1.fffffep+127)]; - tensor clip_4 = clip(alpha = var_1055, beta = const_33, x = new_scale)[name = tensor("clip_4")]; + tensor clip_4 = clip(alpha = var_68, beta = const_33, x = new_scale)[name = tensor("clip_4")]; tensor sqrt_new_scale = sqrt(x = clip_4)[name = tensor("sqrt_new_scale")]; tensor nkv = real_div(x = new_kv_unnorm, y = sqrt_new_scale)[name = tensor("nkv")]; - tensor var_1384_perm_0 = const()[name = tensor("op_1384_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_1426_perm_0 = const()[name = tensor("op_1426_perm_0"), val = tensor([0, 2, 1, 3])]; tensor out_33_axes_0 = const()[name = tensor("out_33_axes_0"), val = tensor([-1])]; - tensor var_1384 = transpose(perm = var_1384_perm_0, x = out_31)[name = tensor("transpose_11")]; - tensor out_33 = layer_norm(axes = out_33_axes_0, epsilon = var_1052, x = var_1384)[name = tensor("out_33")]; - tensor var_1388 = const()[name = tensor("op_1388"), val = tensor([6, 5, 256])]; - tensor out = reshape(shape = var_1388, x = out_33)[name = tensor("out")]; - tensor var_1390 = silu(x = input_187)[name = tensor("op_1390")]; - tensor input_189 = mul(x = var_1390, y = out)[name = tensor("input_189")]; - tensor ret_out = linear(bias = decoder_out_proj_1_bias, weight = decoder_out_proj_1_weight, x = input_189)[name = tensor("linear_51")]; - tensor input_191 = add(x = x, y = ret_out)[name = tensor("input_191")]; + tensor var_1426 = transpose(perm = var_1426_perm_0, x = out_31)[name = tensor("transpose_11")]; + tensor out_33 = layer_norm(axes = out_33_axes_0, epsilon = var_82, x = var_1426)[name = tensor("out_33")]; + tensor var_1430 = const()[name = tensor("op_1430"), val = tensor([6, 5, 256])]; + tensor out = reshape(shape = var_1430, x = out_33)[name = tensor("out")]; + tensor var_1432 = silu(x = input_189)[name = tensor("op_1432")]; + tensor input_191 = mul(x = var_1432, y = out)[name = tensor("input_191")]; + tensor ret_out = linear(bias = inner_decoder_out_proj_1_bias, weight = inner_decoder_out_proj_1_weight, x = input_191)[name = tensor("linear_51")]; + tensor input_193 = add(x = x, y = ret_out)[name = tensor("input_193")]; tensor xt_5_axes_0 = const()[name = tensor("xt_5_axes_0"), val = tensor([-1])]; - tensor xt_5 = layer_norm(axes = xt_5_axes_0, beta = decoder_norm11_1_bias, epsilon = var_1050, gamma = decoder_norm11_1_weight, x = input_191)[name = tensor("xt_5")]; - tensor var_1400 = const()[name = tensor("op_1400"), val = tensor([1, 6, 5, 256])]; - tensor var_1401 = reshape(shape = var_1400, x = xt_5)[name = tensor("op_1401")]; - tensor var_1402_perm_0 = const()[name = tensor("op_1402_perm_0"), val = tensor([0, 2, 1, 3])]; - tensor var_1405 = const()[name = tensor("op_1405"), val = tensor([5, 6, 256])]; - tensor var_1402 = transpose(perm = var_1402_perm_0, x = var_1401)[name = tensor("transpose_10")]; - tensor query_5 = reshape(shape = var_1405, x = var_1402)[name = tensor("query_5")]; + tensor xt_5 = layer_norm(axes = xt_5_axes_0, beta = inner_decoder_norm11_1_bias, epsilon = var_74, gamma = inner_decoder_norm11_1_weight, x = input_193)[name = tensor("xt_5")]; + tensor var_1442 = const()[name = tensor("op_1442"), val = tensor([1, 6, 5, 256])]; + tensor var_1443 = reshape(shape = var_1442, x = xt_5)[name = tensor("op_1443")]; + tensor var_1444_perm_0 = const()[name = tensor("op_1444_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_1447 = const()[name = tensor("op_1447"), val = tensor([5, 6, 256])]; + tensor var_1444 = transpose(perm = var_1444_perm_0, x = var_1443)[name = tensor("transpose_10")]; + tensor query_5 = reshape(shape = var_1447, x = var_1444)[name = tensor("query_5")]; tensor query_perm_0 = const()[name = tensor("query_perm_0"), val = tensor([1, 0, 2])]; tensor query = transpose(perm = query_perm_0, x = query_5)[name = tensor("transpose_9")]; - tensor var_1428 = linear(bias = decoder_self_attn2_1_in_proj_bias, weight = decoder_self_attn2_1_in_proj_weight, x = query)[name = tensor("linear_52")]; + tensor var_1470 = linear(bias = inner_decoder_self_attn2_1_in_proj_bias, weight = inner_decoder_self_attn2_1_in_proj_weight, x = query)[name = tensor("linear_52")]; tensor concat_2 = const()[name = tensor("concat_2"), val = tensor([6, 5, 3, 256])]; - tensor var_1430 = reshape(shape = concat_2, x = var_1428)[name = tensor("op_1430")]; - tensor var_1431_axes_0 = const()[name = tensor("op_1431_axes_0"), val = tensor([0])]; - tensor var_1431 = expand_dims(axes = var_1431_axes_0, x = var_1430)[name = tensor("op_1431")]; - tensor var_1432_perm_0 = const()[name = tensor("op_1432_perm_0"), val = tensor([-2, 1, 2, 0, 4])]; - tensor var_1433_axes_0 = const()[name = tensor("op_1433_axes_0"), val = tensor([-2])]; - tensor var_1432 = transpose(perm = var_1432_perm_0, x = var_1431)[name = tensor("transpose_8")]; - tensor var_1433 = squeeze(axes = var_1433_axes_0, x = var_1432)[name = tensor("op_1433")]; + tensor var_1472 = reshape(shape = concat_2, x = var_1470)[name = tensor("op_1472")]; + tensor var_1473_axes_0 = const()[name = tensor("op_1473_axes_0"), val = tensor([0])]; + tensor var_1473 = expand_dims(axes = var_1473_axes_0, x = var_1472)[name = tensor("op_1473")]; + tensor var_1474_perm_0 = const()[name = tensor("op_1474_perm_0"), val = tensor([-2, 1, 2, 0, 4])]; + tensor var_1475_axes_0 = const()[name = tensor("op_1475_axes_0"), val = tensor([-2])]; + tensor var_1474 = transpose(perm = var_1474_perm_0, x = var_1473)[name = tensor("transpose_8")]; + tensor var_1475 = squeeze(axes = var_1475_axes_0, x = var_1474)[name = tensor("op_1475")]; tensor q_19_begin_0 = const()[name = tensor("q_19_begin_0"), val = tensor([0, 0, 0, 0])]; tensor q_19_end_0 = const()[name = tensor("q_19_end_0"), val = tensor([1, 6, 5, 256])]; tensor q_19_end_mask_0 = const()[name = tensor("q_19_end_mask_0"), val = tensor([false, true, true, true])]; tensor q_19_squeeze_mask_0 = const()[name = tensor("q_19_squeeze_mask_0"), val = tensor([true, false, false, false])]; - tensor q_19 = slice_by_index(begin = q_19_begin_0, end = q_19_end_0, end_mask = q_19_end_mask_0, squeeze_mask = q_19_squeeze_mask_0, x = var_1433)[name = tensor("q_19")]; + tensor q_19 = slice_by_index(begin = q_19_begin_0, end = q_19_end_0, end_mask = q_19_end_mask_0, squeeze_mask = q_19_squeeze_mask_0, x = var_1475)[name = tensor("q_19")]; tensor k_19_begin_0 = const()[name = tensor("k_19_begin_0"), val = tensor([1, 0, 0, 0])]; tensor k_19_end_0 = const()[name = tensor("k_19_end_0"), val = tensor([2, 6, 5, 256])]; tensor k_19_end_mask_0 = const()[name = tensor("k_19_end_mask_0"), val = tensor([false, true, true, true])]; tensor k_19_squeeze_mask_0 = const()[name = tensor("k_19_squeeze_mask_0"), val = tensor([true, false, false, false])]; - tensor k_19 = slice_by_index(begin = k_19_begin_0, end = k_19_end_0, end_mask = k_19_end_mask_0, squeeze_mask = k_19_squeeze_mask_0, x = var_1433)[name = tensor("k_19")]; + tensor k_19 = slice_by_index(begin = k_19_begin_0, end = k_19_end_0, end_mask = k_19_end_mask_0, squeeze_mask = k_19_squeeze_mask_0, x = var_1475)[name = tensor("k_19")]; tensor v_19_begin_0 = const()[name = tensor("v_19_begin_0"), val = tensor([2, 0, 0, 0])]; tensor v_19_end_0 = const()[name = tensor("v_19_end_0"), val = tensor([3, 6, 5, 256])]; tensor v_19_end_mask_0 = const()[name = tensor("v_19_end_mask_0"), val = tensor([false, true, true, true])]; tensor v_19_squeeze_mask_0 = const()[name = tensor("v_19_squeeze_mask_0"), val = tensor([true, false, false, false])]; - tensor v_19 = slice_by_index(begin = v_19_begin_0, end = v_19_end_0, end_mask = v_19_end_mask_0, squeeze_mask = v_19_squeeze_mask_0, x = var_1433)[name = tensor("v_19")]; - tensor var_1441 = const()[name = tensor("op_1441"), val = tensor([6, 20, 64])]; - tensor var_1442 = reshape(shape = var_1441, x = q_19)[name = tensor("op_1442")]; + tensor v_19 = slice_by_index(begin = v_19_begin_0, end = v_19_end_0, end_mask = v_19_end_mask_0, squeeze_mask = v_19_squeeze_mask_0, x = var_1475)[name = tensor("v_19")]; + tensor var_1483 = const()[name = tensor("op_1483"), val = tensor([6, 20, 64])]; + tensor var_1484 = reshape(shape = var_1483, x = q_19)[name = tensor("op_1484")]; tensor q_21_perm_0 = const()[name = tensor("q_21_perm_0"), val = tensor([1, 0, 2])]; - tensor var_1448 = const()[name = tensor("op_1448"), val = tensor([6, 20, 64])]; - tensor var_1449 = reshape(shape = var_1448, x = k_19)[name = tensor("op_1449")]; + tensor var_1490 = const()[name = tensor("op_1490"), val = tensor([6, 20, 64])]; + tensor var_1491 = reshape(shape = var_1490, x = k_19)[name = tensor("op_1491")]; tensor k_21_perm_0 = const()[name = tensor("k_21_perm_0"), val = tensor([1, 0, 2])]; - tensor var_1455 = const()[name = tensor("op_1455"), val = tensor([6, 20, 64])]; - tensor var_1456 = reshape(shape = var_1455, x = v_19)[name = tensor("op_1456")]; + tensor var_1497 = const()[name = tensor("op_1497"), val = tensor([6, 20, 64])]; + tensor var_1498 = reshape(shape = var_1497, x = v_19)[name = tensor("op_1498")]; tensor v_21_perm_0 = const()[name = tensor("v_21_perm_0"), val = tensor([1, 0, 2])]; - tensor var_1459 = const()[name = tensor("op_1459"), val = tensor([5, 4, 6, 64])]; - tensor q_21 = transpose(perm = q_21_perm_0, x = var_1442)[name = tensor("transpose_7")]; - tensor q = reshape(shape = var_1459, x = q_21)[name = tensor("q")]; - tensor var_1461 = const()[name = tensor("op_1461"), val = tensor([5, 4, 6, 64])]; - tensor k_21 = transpose(perm = k_21_perm_0, x = var_1449)[name = tensor("transpose_6")]; - tensor k = reshape(shape = var_1461, x = k_21)[name = tensor("k")]; - tensor var_1463 = const()[name = tensor("op_1463"), val = tensor([5, 4, 6, 64])]; - tensor v_21 = transpose(perm = v_21_perm_0, x = var_1456)[name = tensor("transpose_5")]; - tensor v = reshape(shape = var_1463, x = v_21)[name = tensor("v")]; + tensor var_1501 = const()[name = tensor("op_1501"), val = tensor([5, 4, 6, 64])]; + tensor q_21 = transpose(perm = q_21_perm_0, x = var_1484)[name = tensor("transpose_7")]; + tensor q = reshape(shape = var_1501, x = q_21)[name = tensor("q")]; + tensor var_1503 = const()[name = tensor("op_1503"), val = tensor([5, 4, 6, 64])]; + tensor k_21 = transpose(perm = k_21_perm_0, x = var_1491)[name = tensor("transpose_6")]; + tensor k = reshape(shape = var_1503, x = k_21)[name = tensor("k")]; + tensor var_1505 = const()[name = tensor("op_1505"), val = tensor([5, 4, 6, 64])]; + tensor v_21 = transpose(perm = v_21_perm_0, x = var_1498)[name = tensor("transpose_5")]; + tensor v = reshape(shape = var_1505, x = v_21)[name = tensor("v")]; tensor mul_3_y_0 = const()[name = tensor("mul_3_y_0"), val = tensor(0x1p-3)]; tensor mul_3 = mul(x = q, y = mul_3_y_0)[name = tensor("mul_3")]; tensor matmul_1_transpose_y_0 = const()[name = tensor("matmul_1_transpose_y_0"), val = tensor(true)]; @@ -1317,36 +1336,36 @@ program(1.0) tensor attn_output_9_transpose_x_0 = const()[name = tensor("attn_output_9_transpose_x_0"), val = tensor(false)]; tensor attn_output_9_transpose_y_0 = const()[name = tensor("attn_output_9_transpose_y_0"), val = tensor(false)]; tensor attn_output_9 = matmul(transpose_x = attn_output_9_transpose_x_0, transpose_y = attn_output_9_transpose_y_0, x = softmax_1, y = v)[name = tensor("attn_output_9")]; - tensor var_1466 = const()[name = tensor("op_1466"), val = tensor([2, 0, 1, 3])]; - tensor var_1471 = const()[name = tensor("op_1471"), val = tensor([30, 256])]; - tensor var_1467 = transpose(perm = var_1466, x = attn_output_9)[name = tensor("transpose_4")]; - tensor attn_output_11 = reshape(shape = var_1471, x = var_1467)[name = tensor("attn_output_11")]; - tensor attn_output_13 = linear(bias = decoder_self_attn2_1_out_proj_bias, weight = decoder_self_attn2_1_out_proj_weight, x = attn_output_11)[name = tensor("linear_53")]; - tensor var_1475 = const()[name = tensor("op_1475"), val = tensor([6, 5, 256])]; - tensor attn_output = reshape(shape = var_1475, x = attn_output_13)[name = tensor("attn_output")]; + tensor var_1508 = const()[name = tensor("op_1508"), val = tensor([2, 0, 1, 3])]; + tensor var_1513 = const()[name = tensor("op_1513"), val = tensor([30, 256])]; + tensor var_1509 = transpose(perm = var_1508, x = attn_output_9)[name = tensor("transpose_4")]; + tensor attn_output_11 = reshape(shape = var_1513, x = var_1509)[name = tensor("attn_output_11")]; + tensor attn_output_13 = linear(bias = inner_decoder_self_attn2_1_out_proj_bias, weight = inner_decoder_self_attn2_1_out_proj_weight, x = attn_output_11)[name = tensor("linear_53")]; + tensor var_1517 = const()[name = tensor("op_1517"), val = tensor([6, 5, 256])]; + tensor attn_output = reshape(shape = var_1517, x = attn_output_13)[name = tensor("attn_output")]; tensor sa2_out_perm_0 = const()[name = tensor("sa2_out_perm_0"), val = tensor([1, 0, 2])]; tensor sa2_out = transpose(perm = sa2_out_perm_0, x = attn_output)[name = tensor("transpose_3")]; - tensor input_193 = add(x = query_5, y = sa2_out)[name = tensor("input_193")]; - tensor input_195_axes_0 = const()[name = tensor("input_195_axes_0"), val = tensor([-1])]; - tensor input_195 = layer_norm(axes = input_195_axes_0, beta = decoder_norm21_1_bias, epsilon = var_1050, gamma = decoder_norm21_1_weight, x = input_193)[name = tensor("input_195")]; - tensor input_197 = linear(bias = decoder_linear1_1_bias, weight = decoder_linear1_1_weight, x = input_195)[name = tensor("linear_54")]; - tensor input_199 = relu(x = input_197)[name = tensor("input_199")]; - tensor ffn_out = linear(bias = decoder_linear2_1_bias, weight = decoder_linear2_1_weight, x = input_199)[name = tensor("linear_55")]; - tensor input_201 = add(x = input_195, y = ffn_out)[name = tensor("input_201")]; + tensor input_195 = add(x = query_5, y = sa2_out)[name = tensor("input_195")]; + tensor input_197_axes_0 = const()[name = tensor("input_197_axes_0"), val = tensor([-1])]; + tensor input_197 = layer_norm(axes = input_197_axes_0, beta = inner_decoder_norm21_1_bias, epsilon = var_74, gamma = inner_decoder_norm21_1_weight, x = input_195)[name = tensor("input_197")]; + tensor input_199 = linear(bias = inner_decoder_linear1_1_bias, weight = inner_decoder_linear1_1_weight, x = input_197)[name = tensor("linear_54")]; + tensor input_201 = relu(x = input_199)[name = tensor("input_201")]; + tensor ffn_out = linear(bias = inner_decoder_linear2_1_bias, weight = inner_decoder_linear2_1_weight, x = input_201)[name = tensor("linear_55")]; + tensor input_203 = add(x = input_197, y = ffn_out)[name = tensor("input_203")]; tensor xt_axes_0 = const()[name = tensor("xt_axes_0"), val = tensor([-1])]; - tensor xt = layer_norm(axes = xt_axes_0, beta = decoder_norm22_1_bias, epsilon = var_1050, gamma = decoder_norm22_1_weight, x = input_201)[name = tensor("xt")]; - tensor var_1495 = const()[name = tensor("op_1495"), val = tensor([1, 5, 6, 256])]; - tensor input = reshape(shape = var_1495, x = xt)[name = tensor("input")]; - tensor var_1497 = const()[name = tensor("op_1497"), val = tensor([-1])]; - tensor var_1498 = reduce_l2_norm(axes = var_1497, keep_dims = var_1053, x = input)[name = tensor("op_1498")]; + tensor xt = layer_norm(axes = xt_axes_0, beta = inner_decoder_norm22_1_bias, epsilon = var_74, gamma = inner_decoder_norm22_1_weight, x = input_203)[name = tensor("xt")]; + tensor var_1537 = const()[name = tensor("op_1537"), val = tensor([1, 5, 6, 256])]; + tensor input = reshape(shape = var_1537, x = xt)[name = tensor("input")]; + tensor var_1539 = const()[name = tensor("op_1539"), val = tensor([-1])]; + tensor var_1540 = reduce_l2_norm(axes = var_1539, keep_dims = var_73, x = input)[name = tensor("op_1540")]; tensor const_42 = const()[name = tensor("const_42"), val = tensor(0x1.fffffep+127)]; - tensor clip_5 = clip(alpha = var_1045, beta = const_42, x = var_1498)[name = tensor("clip_5")]; - tensor var_1500 = real_div(x = input, y = clip_5)[name = tensor("op_1500")]; + tensor clip_5 = clip(alpha = var_88, beta = const_42, x = var_1540)[name = tensor("clip_5")]; + tensor var_1542 = real_div(x = input, y = clip_5)[name = tensor("op_1542")]; tensor concat_6 = const()[name = tensor("concat_6"), val = tensor([5, 1, 256])]; tensor reshape_0 = reshape(shape = concat_6, x = emb)[name = tensor("reshape_0")]; tensor transpose_1_perm_0 = const()[name = tensor("transpose_1_perm_0"), val = tensor([0, 1, 3, 2])]; tensor concat_7 = const()[name = tensor("concat_7"), val = tensor([5, 256, 6])]; - tensor transpose_1 = transpose(perm = transpose_1_perm_0, x = var_1500)[name = tensor("transpose_2")]; + tensor transpose_1 = transpose(perm = transpose_1_perm_0, x = var_1542)[name = tensor("transpose_2")]; tensor reshape_1 = reshape(shape = concat_7, x = transpose_1)[name = tensor("reshape_1")]; tensor matmul_2_transpose_x_0 = const()[name = tensor("matmul_2_transpose_x_0"), val = tensor(false)]; tensor matmul_2_transpose_y_0 = const()[name = tensor("matmul_2_transpose_y_0"), val = tensor(false)]; @@ -1357,10 +1376,10 @@ program(1.0) tensor output_end_0 = const()[name = tensor("output_end_0"), val = tensor([1, 5, 5])]; tensor output_end_mask_0 = const()[name = tensor("output_end_mask_0"), val = tensor([true, true, false])]; tensor output = slice_by_index(begin = output_begin_0, end = output_end_0, end_mask = output_end_mask_0, x = reshape_2)[name = tensor("output")]; - tensor probs = sigmoid(x = output)[name = tensor("op_1504")]; - tensor var_1506_axis_0 = const()[name = tensor("op_1506_axis_0"), val = tensor(0)]; - tensor dec_kv_new = stack(axis = var_1506_axis_0, values = (var_1202, nkv))[name = tensor("op_1506")]; - tensor var_1508_axis_0 = const()[name = tensor("op_1508_axis_0"), val = tensor(0)]; - tensor dec_scale_new = stack(axis = var_1508_axis_0, values = (new_scale_9, new_scale))[name = tensor("op_1508")]; + tensor probs = sigmoid(x = output)[name = tensor("op_1546")]; + tensor var_1548_axis_0 = const()[name = tensor("op_1548_axis_0"), val = tensor(0)]; + tensor dec_kv_new = stack(axis = var_1548_axis_0, values = (var_1244, nkv))[name = tensor("op_1548")]; + tensor var_1550_axis_0 = const()[name = tensor("op_1550_axis_0"), val = tensor(0)]; + tensor dec_scale_new = stack(axis = var_1550_axis_0, values = (new_scale_9, new_scale))[name = tensor("op_1550")]; } -> (probs, enc_kv_new, enc_scale_new, enc_conv_cache_new, cnn_window_new, dec_kv_new, dec_scale_new); } \ No newline at end of file diff --git a/optimized/ami/500ms/ls_eend_ami_500ms.mlpackage/Data/com.apple.CoreML/model.mlmodel b/optimized/ami/500ms/ls_eend_ami_500ms.mlpackage/Data/com.apple.CoreML/model.mlmodel index c2635f870434e5d00bdf7febc221c1084a938f1b..b439a634d1a038a58cc98042708019768152db58 100644 --- a/optimized/ami/500ms/ls_eend_ami_500ms.mlpackage/Data/com.apple.CoreML/model.mlmodel +++ b/optimized/ami/500ms/ls_eend_ami_500ms.mlpackage/Data/com.apple.CoreML/model.mlmodel @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:bb1fc846d5fe66b99cc3d94f2f27dc499fd4bcd04080e23e0297dd02f0f61f7b -size 196611 +oid sha256:449581250eeb3f6eda599089dca398f8bd8fb4e4433e96a35c5ccfdf9a84e6c7 +size 203211 diff --git a/optimized/ami/500ms/ls_eend_ami_500ms.mlpackage/Manifest.json b/optimized/ami/500ms/ls_eend_ami_500ms.mlpackage/Manifest.json index 719daafccc33381f00ca4380bee97b50ce8c7019..3aae5688c1b26b6cb8cdfd89ba0210bc978433ea 100644 --- a/optimized/ami/500ms/ls_eend_ami_500ms.mlpackage/Manifest.json +++ b/optimized/ami/500ms/ls_eend_ami_500ms.mlpackage/Manifest.json @@ -1,18 +1,18 @@ { "fileFormatVersion": "1.0.0", "itemInfoEntries": { - "BFFA710C-2BC9-41A2-80C8-E692BA4158D2": { + "42F69C96-0A26-4074-BA5C-B0CA1B7F580E": { "author": "com.apple.CoreML", "description": "CoreML Model Weights", "name": "weights", "path": "com.apple.CoreML/weights" }, - "F337566C-1A74-4DED-8712-B263AAD1B3CF": { + "516AD8CE-7461-455E-837D-56DC0B7EE116": { "author": "com.apple.CoreML", "description": "CoreML Model Specification", "name": "model.mlmodel", "path": "com.apple.CoreML/model.mlmodel" } }, - "rootModelIdentifier": "F337566C-1A74-4DED-8712-B263AAD1B3CF" + "rootModelIdentifier": "516AD8CE-7461-455E-837D-56DC0B7EE116" } diff --git a/optimized/ch/100ms/ls_eend_ch_100ms.mlmodelc/analytics/coremldata.bin b/optimized/ch/100ms/ls_eend_ch_100ms.mlmodelc/analytics/coremldata.bin index d14d2473e765f039603cf751a34e8895b6d8f152..31fee0a94bede8702423f9026567ab93f857b5a1 100644 --- a/optimized/ch/100ms/ls_eend_ch_100ms.mlmodelc/analytics/coremldata.bin +++ b/optimized/ch/100ms/ls_eend_ch_100ms.mlmodelc/analytics/coremldata.bin @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:162c9b06e4452aa666026e5cf6d51ebd396faac77145f39864e8e1fee2ec569e +oid sha256:75d3cc235f66ab174d6aa9cba0d97746fe4f69ba198a3ade13dc3cd450a16a2a size 243 diff --git a/optimized/ch/100ms/ls_eend_ch_100ms.mlmodelc/coremldata.bin b/optimized/ch/100ms/ls_eend_ch_100ms.mlmodelc/coremldata.bin index 4f3888dde2e2edab7506a40b902b161ff6a88031..8ac0ea92ab52934795d546e4b456121c116daaa2 100644 --- a/optimized/ch/100ms/ls_eend_ch_100ms.mlmodelc/coremldata.bin +++ b/optimized/ch/100ms/ls_eend_ch_100ms.mlmodelc/coremldata.bin @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:6e2c9fc79e93736f5e7e4017f5282df7b689fac77791ab360645dd0732ca2af9 -size 1301 +oid sha256:e097920817284b8134fcc6a8906e568bdf28e5a55bd78818f9e0485c71019561 +size 1404 diff --git a/optimized/ch/100ms/ls_eend_ch_100ms.mlmodelc/metadata.json b/optimized/ch/100ms/ls_eend_ch_100ms.mlmodelc/metadata.json index c0c60029c9245e143c373a94558005a27729a798..7354bbdc1dd030e949e3666df3cfdc7562a14d38 100644 --- a/optimized/ch/100ms/ls_eend_ch_100ms.mlmodelc/metadata.json +++ b/optimized/ch/100ms/ls_eend_ch_100ms.mlmodelc/metadata.json @@ -1,7 +1,7 @@ [ { "metadataOutputVersion" : "3.0", - "shortDescription" : "LS-EEND CALLHOME streaming diarizer (pipeline, T=1, max_speakers=7)", + "shortDescription" : "LS-EEND CALLHOME streaming diarizer (pipeline, T=1, max_speakers=7, layout=raw_mel, cu=all)", "outputSchema" : [ { "hasShapeFlexibility" : "0", @@ -81,12 +81,12 @@ "specificationVersion" : 8, "mlProgramOperationTypeHistogram" : { "Ios16.reduceL2Norm" : 2, - "Ios17.reshape" : 66, + "Ios17.reshape" : 67, "Ios16.softmax" : 2, "Ios17.matmul" : 29, "Ios17.transpose" : 57, "Split" : 4, - "Ios17.expandDims" : 3, + "Ios17.expandDims" : 4, "Ios17.add" : 46, "Ios16.sigmoid" : 5, "Ios17.sliceByIndex" : 36, @@ -128,9 +128,9 @@ "hasShapeFlexibility" : "0", "isOptional" : "0", "dataType" : "Float32", - "formattedType" : "MultiArray (Float32 1 × 1 × 345)", + "formattedType" : "MultiArray (Float32 1 × 15 × 23)", "shortDescription" : "", - "shape" : "[1, 1, 345]", + "shape" : "[1, 15, 23]", "name" : "features", "type" : "MultiArray" }, @@ -206,8 +206,8 @@ } ], "userDefinedMetadata" : { - "com.github.apple.coremltools.conversion_date" : "2026-04-16", - "config" : "{\"model_name\": \"ch\", \"model_label\": \"CALLHOME\", \"variant\": \"pipeline\", \"chunk_size\": 1, \"step_duration_ms\": 100, \"frame_duration_ms\": 100, \"frame_duration_seconds\": 0.1, \"max_speakers\": 7, \"max_nspks\": 9, \"n_units\": 256, \"n_heads\": 4, \"enc_n_layers\": 4, \"dec_n_layers\": 2, \"conv_kernel_size\": 16, \"conv_delay\": 9, \"sample_rate\": 8000, \"win_length\": 200, \"hop_length\": 80, \"n_mels\": 23, \"context_size\": 7, \"subsampling\": 10, \"feat_type\": \"logmel23_cummn\", \"pure_roll\": true}", + "com.github.apple.coremltools.conversion_date" : "2026-04-18", + "config" : "{\"model_name\": \"ch\", \"model_label\": \"CALLHOME\", \"variant\": \"pipeline\", \"chunk_size\": 1, \"step_duration_ms\": 100, \"frame_duration_ms\": 100, \"frame_duration_seconds\": 0.1, \"max_speakers\": 7, \"max_nspks\": 9, \"n_units\": 256, \"n_heads\": 4, \"enc_n_layers\": 4, \"dec_n_layers\": 2, \"conv_kernel_size\": 16, \"conv_delay\": 9, \"sample_rate\": 8000, \"win_length\": 200, \"hop_length\": 80, \"n_mels\": 23, \"context_size\": 7, \"subsampling\": 10, \"feat_type\": \"logmel23_cummn\", \"pure_roll\": true, \"input_layout\": \"raw_mel\", \"compute_units_export\": \"all\", \"raw_mel_length\": 15}", "com.github.apple.coremltools.source" : "torch==2.6.0", "com.github.apple.coremltools.version" : "9.0", "com.github.apple.coremltools.source_dialect" : "TorchScript" diff --git a/optimized/ch/100ms/ls_eend_ch_100ms.mlmodelc/model.mil b/optimized/ch/100ms/ls_eend_ch_100ms.mlmodelc/model.mil index c6564bd8478dcda29a396e2ab8dcb7eda1e4a19b..dfa2adbe0753d8a6dffeec554a5c9c51bd856833 100644 --- a/optimized/ch/100ms/ls_eend_ch_100ms.mlmodelc/model.mil +++ b/optimized/ch/100ms/ls_eend_ch_100ms.mlmodelc/model.mil @@ -1,233 +1,239 @@ program(1.0) [buildInfo = dict, tensor>({{"coremlc-component-MIL", "3510.2.1"}, {"coremlc-version", "3500.32.1"}, {"coremltools-component-torch", "2.6.0"}, {"coremltools-source-dialect", "TorchScript"}, {"coremltools-version", "9.0"}})] { - func main(tensor cnn_window, tensor dec_kv, tensor dec_scale, tensor enc_conv_cache, tensor enc_kv, tensor enc_scale, tensor features, tensor valid_mask) { - tensor encoder_cnn_bias = const()[name = tensor("encoder_cnn_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(64)))]; - tensor encoder_cnn_weight = const()[name = tensor("encoder_cnn_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1152)))]; - tensor encoder__causal_mask = const()[name = tensor("encoder__causal_mask"), val = tensor([[0x1p+0]])]; - tensor encoder__t_index = const()[name = tensor("encoder__t_index"), val = tensor([0x1p+0])]; - tensor encoder_input_projection_linear_bias = const()[name = tensor("encoder_input_projection_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4981952)))]; - tensor encoder_input_projection_linear_weight = const()[name = tensor("encoder_input_projection_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4983040)))]; - tensor encoder_pre_layer_norm_bias = const()[name = tensor("encoder_pre_layer_norm_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(5336384)))]; - tensor encoder_pre_layer_norm_weight = const()[name = tensor("encoder_pre_layer_norm_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(5337472)))]; - tensor encoder_ffn1_0_module_sequential_0_bias = const()[name = tensor("encoder_ffn1_0_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(5338560)))]; - tensor encoder_ffn1_0_module_sequential_0_weight = const()[name = tensor("encoder_ffn1_0_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(5339648)))]; - tensor encoder_ffn1_0_module_sequential_1_linear_bias = const()[name = tensor("encoder_ffn1_0_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(5340736)))]; - tensor encoder_ffn1_0_module_sequential_1_linear_weight = const()[name = tensor("encoder_ffn1_0_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(5344896)))]; - tensor encoder_ffn1_0_module_sequential_4_linear_bias = const()[name = tensor("encoder_ffn1_0_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(6393536)))]; - tensor encoder_ffn1_0_module_sequential_4_linear_weight = const()[name = tensor("encoder_ffn1_0_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(6394624)))]; - tensor encoder_ret_lns_0_bias = const()[name = tensor("encoder_ret_lns_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7443264)))]; - tensor encoder_ret_lns_0_weight = const()[name = tensor("encoder_ret_lns_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7444352)))]; - tensor encoder_q_proj_0_bias = const()[name = tensor("encoder_q_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7445440)))]; - tensor encoder_q_proj_0_weight = const()[name = tensor("encoder_q_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7446528)))]; - tensor encoder_k_proj_0_bias = const()[name = tensor("encoder_k_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7708736)))]; - tensor encoder_k_proj_0_weight = const()[name = tensor("encoder_k_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7709824)))]; - tensor encoder_v_proj_0_bias = const()[name = tensor("encoder_v_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7972032)))]; - tensor encoder_v_proj_0_weight = const()[name = tensor("encoder_v_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7973120)))]; - tensor encoder_g_proj_0_bias = const()[name = tensor("encoder_g_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8235328)))]; - tensor encoder_g_proj_0_weight = const()[name = tensor("encoder_g_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8236416)))]; - tensor encoder_out_proj_0_bias = const()[name = tensor("encoder_out_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8498624)))]; - tensor encoder_out_proj_0_weight = const()[name = tensor("encoder_out_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8499712)))]; - tensor encoder_conv_module_0_sequential_0_bias = const()[name = tensor("encoder_conv_module_0_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8761920)))]; - tensor encoder_conv_module_0_sequential_0_weight = const()[name = tensor("encoder_conv_module_0_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8763008)))]; - tensor encoder_conv_module_0_sequential_2_conv_bias = const()[name = tensor("encoder_conv_module_0_sequential_2_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8764096)))]; - tensor encoder_conv_module_0_sequential_2_conv_weight = const()[name = tensor("encoder_conv_module_0_sequential_2_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8766208)))]; - tensor encoder_conv_module_0_sequential_4_conv_weight = const()[name = tensor("encoder_conv_module_0_sequential_4_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9290560)))]; - tensor encoder_conv_module_0_sequential_5_running_var = const()[name = tensor("encoder_conv_module_0_sequential_5_running_var"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9307008)))]; - tensor encoder_conv_module_0_sequential_5_running_mean = const()[name = tensor("encoder_conv_module_0_sequential_5_running_mean"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9308096)))]; - tensor encoder_conv_module_0_sequential_5_bias = const()[name = tensor("encoder_conv_module_0_sequential_5_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9309184)))]; - tensor encoder_conv_module_0_sequential_5_weight = const()[name = tensor("encoder_conv_module_0_sequential_5_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9310272)))]; - tensor encoder_conv_module_0_sequential_7_conv_bias = const()[name = tensor("encoder_conv_module_0_sequential_7_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9311360)))]; - tensor encoder_conv_module_0_sequential_7_conv_weight = const()[name = tensor("encoder_conv_module_0_sequential_7_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9312448)))]; - tensor encoder_ffn2_0_module_sequential_0_bias = const()[name = tensor("encoder_ffn2_0_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9574656)))]; - tensor encoder_ffn2_0_module_sequential_0_weight = const()[name = tensor("encoder_ffn2_0_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9575744)))]; - tensor encoder_ffn2_0_module_sequential_1_linear_bias = const()[name = tensor("encoder_ffn2_0_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9576832)))]; - tensor encoder_ffn2_0_module_sequential_1_linear_weight = const()[name = tensor("encoder_ffn2_0_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9580992)))]; - tensor encoder_ffn2_0_module_sequential_4_linear_bias = const()[name = tensor("encoder_ffn2_0_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(10629632)))]; - tensor encoder_ffn2_0_module_sequential_4_linear_weight = const()[name = tensor("encoder_ffn2_0_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(10630720)))]; - tensor encoder_layer_norm_0_bias = const()[name = tensor("encoder_layer_norm_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11679360)))]; - tensor encoder_layer_norm_0_weight = const()[name = tensor("encoder_layer_norm_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11680448)))]; - tensor encoder_ffn1_1_module_sequential_0_bias = const()[name = tensor("encoder_ffn1_1_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11681536)))]; - tensor encoder_ffn1_1_module_sequential_0_weight = const()[name = tensor("encoder_ffn1_1_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11682624)))]; - tensor encoder_ffn1_1_module_sequential_1_linear_bias = const()[name = tensor("encoder_ffn1_1_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11683712)))]; - tensor encoder_ffn1_1_module_sequential_1_linear_weight = const()[name = tensor("encoder_ffn1_1_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11687872)))]; - tensor encoder_ffn1_1_module_sequential_4_linear_bias = const()[name = tensor("encoder_ffn1_1_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(12736512)))]; - tensor encoder_ffn1_1_module_sequential_4_linear_weight = const()[name = tensor("encoder_ffn1_1_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(12737600)))]; - tensor encoder_ret_lns_1_bias = const()[name = tensor("encoder_ret_lns_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(13786240)))]; - tensor encoder_ret_lns_1_weight = const()[name = tensor("encoder_ret_lns_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(13787328)))]; - tensor encoder_q_proj_1_bias = const()[name = tensor("encoder_q_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(13788416)))]; - tensor encoder_q_proj_1_weight = const()[name = tensor("encoder_q_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(13789504)))]; - tensor encoder_k_proj_1_bias = const()[name = tensor("encoder_k_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14051712)))]; - tensor encoder_k_proj_1_weight = const()[name = tensor("encoder_k_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14052800)))]; - tensor encoder_v_proj_1_bias = const()[name = tensor("encoder_v_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14315008)))]; - tensor encoder_v_proj_1_weight = const()[name = tensor("encoder_v_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14316096)))]; - tensor encoder_g_proj_1_bias = const()[name = tensor("encoder_g_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14578304)))]; - tensor encoder_g_proj_1_weight = const()[name = tensor("encoder_g_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14579392)))]; - tensor encoder_out_proj_1_bias = const()[name = tensor("encoder_out_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14841600)))]; - tensor encoder_out_proj_1_weight = const()[name = tensor("encoder_out_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14842688)))]; - tensor encoder_conv_module_1_sequential_0_bias = const()[name = tensor("encoder_conv_module_1_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15104896)))]; - tensor encoder_conv_module_1_sequential_0_weight = const()[name = tensor("encoder_conv_module_1_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15105984)))]; - tensor encoder_conv_module_1_sequential_2_conv_bias = const()[name = tensor("encoder_conv_module_1_sequential_2_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15107072)))]; - tensor encoder_conv_module_1_sequential_2_conv_weight = const()[name = tensor("encoder_conv_module_1_sequential_2_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15109184)))]; - tensor encoder_conv_module_1_sequential_4_conv_weight = const()[name = tensor("encoder_conv_module_1_sequential_4_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15633536)))]; - tensor encoder_conv_module_1_sequential_5_running_var = const()[name = tensor("encoder_conv_module_1_sequential_5_running_var"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15649984)))]; - tensor encoder_conv_module_1_sequential_5_running_mean = const()[name = tensor("encoder_conv_module_1_sequential_5_running_mean"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15651072)))]; - tensor encoder_conv_module_1_sequential_5_bias = const()[name = tensor("encoder_conv_module_1_sequential_5_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15652160)))]; - tensor encoder_conv_module_1_sequential_5_weight = const()[name = tensor("encoder_conv_module_1_sequential_5_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15653248)))]; - tensor encoder_conv_module_1_sequential_7_conv_bias = const()[name = tensor("encoder_conv_module_1_sequential_7_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15654336)))]; - tensor encoder_conv_module_1_sequential_7_conv_weight = const()[name = tensor("encoder_conv_module_1_sequential_7_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15655424)))]; - tensor encoder_ffn2_1_module_sequential_0_bias = const()[name = tensor("encoder_ffn2_1_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15917632)))]; - tensor encoder_ffn2_1_module_sequential_0_weight = const()[name = tensor("encoder_ffn2_1_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15918720)))]; - tensor encoder_ffn2_1_module_sequential_1_linear_bias = const()[name = tensor("encoder_ffn2_1_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15919808)))]; - tensor encoder_ffn2_1_module_sequential_1_linear_weight = const()[name = tensor("encoder_ffn2_1_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15923968)))]; - tensor encoder_ffn2_1_module_sequential_4_linear_bias = const()[name = tensor("encoder_ffn2_1_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(16972608)))]; - tensor encoder_ffn2_1_module_sequential_4_linear_weight = const()[name = tensor("encoder_ffn2_1_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(16973696)))]; - tensor encoder_layer_norm_1_bias = const()[name = tensor("encoder_layer_norm_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18022336)))]; - tensor encoder_layer_norm_1_weight = const()[name = tensor("encoder_layer_norm_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18023424)))]; - tensor encoder_ffn1_2_module_sequential_0_bias = const()[name = tensor("encoder_ffn1_2_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18024512)))]; - tensor encoder_ffn1_2_module_sequential_0_weight = const()[name = tensor("encoder_ffn1_2_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18025600)))]; - tensor encoder_ffn1_2_module_sequential_1_linear_bias = const()[name = tensor("encoder_ffn1_2_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18026688)))]; - tensor encoder_ffn1_2_module_sequential_1_linear_weight = const()[name = tensor("encoder_ffn1_2_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18030848)))]; - tensor encoder_ffn1_2_module_sequential_4_linear_bias = const()[name = tensor("encoder_ffn1_2_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(19079488)))]; - tensor encoder_ffn1_2_module_sequential_4_linear_weight = const()[name = tensor("encoder_ffn1_2_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(19080576)))]; - tensor encoder_ret_lns_2_bias = const()[name = tensor("encoder_ret_lns_2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20129216)))]; - tensor encoder_ret_lns_2_weight = const()[name = tensor("encoder_ret_lns_2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20130304)))]; - tensor encoder_q_proj_2_bias = const()[name = tensor("encoder_q_proj_2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20131392)))]; - tensor encoder_q_proj_2_weight = const()[name = tensor("encoder_q_proj_2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20132480)))]; - tensor encoder_k_proj_2_bias = const()[name = tensor("encoder_k_proj_2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20394688)))]; - tensor encoder_k_proj_2_weight = const()[name = tensor("encoder_k_proj_2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20395776)))]; - tensor encoder_v_proj_2_bias = const()[name = tensor("encoder_v_proj_2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20657984)))]; - tensor encoder_v_proj_2_weight = const()[name = tensor("encoder_v_proj_2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20659072)))]; - tensor encoder_g_proj_2_bias = const()[name = tensor("encoder_g_proj_2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20921280)))]; - tensor encoder_g_proj_2_weight = const()[name = tensor("encoder_g_proj_2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20922368)))]; - tensor encoder_out_proj_2_bias = const()[name = tensor("encoder_out_proj_2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21184576)))]; - tensor encoder_out_proj_2_weight = const()[name = tensor("encoder_out_proj_2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21185664)))]; - tensor encoder_conv_module_2_sequential_0_bias = const()[name = tensor("encoder_conv_module_2_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21447872)))]; - tensor encoder_conv_module_2_sequential_0_weight = const()[name = tensor("encoder_conv_module_2_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21448960)))]; - tensor encoder_conv_module_2_sequential_2_conv_bias = const()[name = tensor("encoder_conv_module_2_sequential_2_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21450048)))]; - tensor encoder_conv_module_2_sequential_2_conv_weight = const()[name = tensor("encoder_conv_module_2_sequential_2_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21452160)))]; - tensor encoder_conv_module_2_sequential_4_conv_weight = const()[name = tensor("encoder_conv_module_2_sequential_4_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21976512)))]; - tensor encoder_conv_module_2_sequential_5_running_var = const()[name = tensor("encoder_conv_module_2_sequential_5_running_var"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21992960)))]; - tensor encoder_conv_module_2_sequential_5_running_mean = const()[name = tensor("encoder_conv_module_2_sequential_5_running_mean"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21994048)))]; - tensor encoder_conv_module_2_sequential_5_bias = const()[name = tensor("encoder_conv_module_2_sequential_5_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21995136)))]; - tensor encoder_conv_module_2_sequential_5_weight = const()[name = tensor("encoder_conv_module_2_sequential_5_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21996224)))]; - tensor encoder_conv_module_2_sequential_7_conv_bias = const()[name = tensor("encoder_conv_module_2_sequential_7_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21997312)))]; - tensor encoder_conv_module_2_sequential_7_conv_weight = const()[name = tensor("encoder_conv_module_2_sequential_7_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21998400)))]; - tensor encoder_ffn2_2_module_sequential_0_bias = const()[name = tensor("encoder_ffn2_2_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(22260608)))]; - tensor encoder_ffn2_2_module_sequential_0_weight = const()[name = tensor("encoder_ffn2_2_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(22261696)))]; - tensor encoder_ffn2_2_module_sequential_1_linear_bias = const()[name = tensor("encoder_ffn2_2_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(22262784)))]; - tensor encoder_ffn2_2_module_sequential_1_linear_weight = const()[name = tensor("encoder_ffn2_2_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(22266944)))]; - tensor encoder_ffn2_2_module_sequential_4_linear_bias = const()[name = tensor("encoder_ffn2_2_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(23315584)))]; - tensor encoder_ffn2_2_module_sequential_4_linear_weight = const()[name = tensor("encoder_ffn2_2_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(23316672)))]; - tensor encoder_layer_norm_2_bias = const()[name = tensor("encoder_layer_norm_2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(24365312)))]; - tensor encoder_layer_norm_2_weight = const()[name = tensor("encoder_layer_norm_2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(24366400)))]; - tensor encoder_ffn1_3_module_sequential_0_bias = const()[name = tensor("encoder_ffn1_3_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(24367488)))]; - tensor encoder_ffn1_3_module_sequential_0_weight = const()[name = tensor("encoder_ffn1_3_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(24368576)))]; - tensor encoder_ffn1_3_module_sequential_1_linear_bias = const()[name = tensor("encoder_ffn1_3_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(24369664)))]; - tensor encoder_ffn1_3_module_sequential_1_linear_weight = const()[name = tensor("encoder_ffn1_3_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(24373824)))]; - tensor encoder_ffn1_3_module_sequential_4_linear_bias = const()[name = tensor("encoder_ffn1_3_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(25422464)))]; - tensor encoder_ffn1_3_module_sequential_4_linear_weight = const()[name = tensor("encoder_ffn1_3_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(25423552)))]; - tensor encoder_ret_lns_3_bias = const()[name = tensor("encoder_ret_lns_3_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(26472192)))]; - tensor encoder_ret_lns_3_weight = const()[name = tensor("encoder_ret_lns_3_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(26473280)))]; - tensor encoder_q_proj_3_bias = const()[name = tensor("encoder_q_proj_3_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(26474368)))]; - tensor encoder_q_proj_3_weight = const()[name = tensor("encoder_q_proj_3_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(26475456)))]; - tensor encoder_k_proj_3_bias = const()[name = tensor("encoder_k_proj_3_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(26737664)))]; - tensor encoder_k_proj_3_weight = const()[name = tensor("encoder_k_proj_3_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(26738752)))]; - tensor encoder_v_proj_3_bias = const()[name = tensor("encoder_v_proj_3_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27000960)))]; - tensor encoder_v_proj_3_weight = const()[name = tensor("encoder_v_proj_3_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27002048)))]; - tensor encoder_g_proj_3_bias = const()[name = tensor("encoder_g_proj_3_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27264256)))]; - tensor encoder_g_proj_3_weight = const()[name = tensor("encoder_g_proj_3_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27265344)))]; - tensor encoder_out_proj_3_bias = const()[name = tensor("encoder_out_proj_3_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27527552)))]; - tensor encoder_out_proj_3_weight = const()[name = tensor("encoder_out_proj_3_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27528640)))]; - tensor encoder_conv_module_3_sequential_0_bias = const()[name = tensor("encoder_conv_module_3_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27790848)))]; - tensor encoder_conv_module_3_sequential_0_weight = const()[name = tensor("encoder_conv_module_3_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27791936)))]; - tensor encoder_conv_module_3_sequential_2_conv_bias = const()[name = tensor("encoder_conv_module_3_sequential_2_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27793024)))]; - tensor encoder_conv_module_3_sequential_2_conv_weight = const()[name = tensor("encoder_conv_module_3_sequential_2_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27795136)))]; - tensor encoder_conv_module_3_sequential_4_conv_weight = const()[name = tensor("encoder_conv_module_3_sequential_4_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28319488)))]; - tensor encoder_conv_module_3_sequential_5_running_var = const()[name = tensor("encoder_conv_module_3_sequential_5_running_var"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28335936)))]; - tensor encoder_conv_module_3_sequential_5_running_mean = const()[name = tensor("encoder_conv_module_3_sequential_5_running_mean"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28337024)))]; - tensor encoder_conv_module_3_sequential_5_bias = const()[name = tensor("encoder_conv_module_3_sequential_5_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28338112)))]; - tensor encoder_conv_module_3_sequential_5_weight = const()[name = tensor("encoder_conv_module_3_sequential_5_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28339200)))]; - tensor encoder_conv_module_3_sequential_7_conv_bias = const()[name = tensor("encoder_conv_module_3_sequential_7_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28340288)))]; - tensor encoder_conv_module_3_sequential_7_conv_weight = const()[name = tensor("encoder_conv_module_3_sequential_7_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28341376)))]; - tensor encoder_ffn2_3_module_sequential_0_bias = const()[name = tensor("encoder_ffn2_3_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28603584)))]; - tensor encoder_ffn2_3_module_sequential_0_weight = const()[name = tensor("encoder_ffn2_3_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28604672)))]; - tensor encoder_ffn2_3_module_sequential_1_linear_bias = const()[name = tensor("encoder_ffn2_3_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28605760)))]; - tensor encoder_ffn2_3_module_sequential_1_linear_weight = const()[name = tensor("encoder_ffn2_3_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28609920)))]; - tensor encoder_ffn2_3_module_sequential_4_linear_bias = const()[name = tensor("encoder_ffn2_3_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(29658560)))]; - tensor encoder_ffn2_3_module_sequential_4_linear_weight = const()[name = tensor("encoder_ffn2_3_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(29659648)))]; - tensor encoder_layer_norm_3_bias = const()[name = tensor("encoder_layer_norm_3_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(30708288)))]; - tensor encoder_layer_norm_3_weight = const()[name = tensor("encoder_layer_norm_3_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(30709376)))]; - tensor decoder_convert_bias = const()[name = tensor("decoder_convert_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(30710464)))]; - tensor decoder_convert_weight = const()[name = tensor("decoder_convert_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(30711552)))]; - tensor decoder_q_proj_0_bias = const()[name = tensor("decoder_q_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(31235904)))]; - tensor decoder_q_proj_0_weight = const()[name = tensor("decoder_q_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(31236992)))]; - tensor decoder_k_proj_0_bias = const()[name = tensor("decoder_k_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(31499200)))]; - tensor decoder_k_proj_0_weight = const()[name = tensor("decoder_k_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(31500288)))]; - tensor decoder_v_proj_0_bias = const()[name = tensor("decoder_v_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(31762496)))]; - tensor decoder_v_proj_0_weight = const()[name = tensor("decoder_v_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(31763584)))]; - tensor decoder_g_proj_0_bias = const()[name = tensor("decoder_g_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32025792)))]; - tensor decoder_g_proj_0_weight = const()[name = tensor("decoder_g_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32026880)))]; - tensor decoder_out_proj_0_bias = const()[name = tensor("decoder_out_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32289088)))]; - tensor decoder_out_proj_0_weight = const()[name = tensor("decoder_out_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32290176)))]; - tensor decoder_norm11_0_bias = const()[name = tensor("decoder_norm11_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32552384)))]; - tensor decoder_norm11_0_weight = const()[name = tensor("decoder_norm11_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32553472)))]; - tensor decoder_self_attn2_0_out_proj_bias = const()[name = tensor("decoder_self_attn2_0_out_proj_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32554560)))]; - tensor decoder_self_attn2_0_out_proj_weight = const()[name = tensor("decoder_self_attn2_0_out_proj_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32555648)))]; - tensor decoder_self_attn2_0_in_proj_bias = const()[name = tensor("decoder_self_attn2_0_in_proj_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32817856)))]; - tensor decoder_self_attn2_0_in_proj_weight = const()[name = tensor("decoder_self_attn2_0_in_proj_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32820992)))]; - tensor decoder_norm21_0_bias = const()[name = tensor("decoder_norm21_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(33607488)))]; - tensor decoder_norm21_0_weight = const()[name = tensor("decoder_norm21_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(33608576)))]; - tensor decoder_linear1_0_bias = const()[name = tensor("decoder_linear1_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(33609664)))]; - tensor decoder_linear1_0_weight = const()[name = tensor("decoder_linear1_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(33617920)))]; - tensor decoder_linear2_0_bias = const()[name = tensor("decoder_linear2_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(35715136)))]; - tensor decoder_linear2_0_weight = const()[name = tensor("decoder_linear2_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(35716224)))]; - tensor decoder_norm22_0_bias = const()[name = tensor("decoder_norm22_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(37813440)))]; - tensor decoder_norm22_0_weight = const()[name = tensor("decoder_norm22_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(37814528)))]; - tensor decoder_q_proj_1_bias = const()[name = tensor("decoder_q_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(37815616)))]; - tensor decoder_q_proj_1_weight = const()[name = tensor("decoder_q_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(37816704)))]; - tensor decoder_k_proj_1_bias = const()[name = tensor("decoder_k_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38078912)))]; - tensor decoder_k_proj_1_weight = const()[name = tensor("decoder_k_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38080000)))]; - tensor decoder_v_proj_1_bias = const()[name = tensor("decoder_v_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38342208)))]; - tensor decoder_v_proj_1_weight = const()[name = tensor("decoder_v_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38343296)))]; - tensor decoder_g_proj_1_bias = const()[name = tensor("decoder_g_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38605504)))]; - tensor decoder_g_proj_1_weight = const()[name = tensor("decoder_g_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38606592)))]; - tensor decoder_out_proj_1_bias = const()[name = tensor("decoder_out_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38868800)))]; - tensor decoder_out_proj_1_weight = const()[name = tensor("decoder_out_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38869888)))]; - tensor decoder_norm11_1_bias = const()[name = tensor("decoder_norm11_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(39132096)))]; - tensor decoder_norm11_1_weight = const()[name = tensor("decoder_norm11_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(39133184)))]; - tensor decoder_self_attn2_1_out_proj_bias = const()[name = tensor("decoder_self_attn2_1_out_proj_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(39134272)))]; - tensor decoder_self_attn2_1_out_proj_weight = const()[name = tensor("decoder_self_attn2_1_out_proj_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(39135360)))]; - tensor decoder_self_attn2_1_in_proj_bias = const()[name = tensor("decoder_self_attn2_1_in_proj_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(39397568)))]; - tensor decoder_self_attn2_1_in_proj_weight = const()[name = tensor("decoder_self_attn2_1_in_proj_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(39400704)))]; - tensor decoder_norm21_1_bias = const()[name = tensor("decoder_norm21_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(40187200)))]; - tensor decoder_norm21_1_weight = const()[name = tensor("decoder_norm21_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(40188288)))]; - tensor decoder_linear1_1_bias = const()[name = tensor("decoder_linear1_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(40189376)))]; - tensor decoder_linear1_1_weight = const()[name = tensor("decoder_linear1_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(40197632)))]; - tensor decoder_linear2_1_bias = const()[name = tensor("decoder_linear2_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(42294848)))]; - tensor decoder_linear2_1_weight = const()[name = tensor("decoder_linear2_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(42295936)))]; - tensor decoder_norm22_1_bias = const()[name = tensor("decoder_norm22_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(44393152)))]; - tensor decoder_norm22_1_weight = const()[name = tensor("decoder_norm22_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(44394240)))]; - tensor var_11 = const()[name = tensor("op_11"), val = tensor(0x1.5798eep-27)]; - tensor var_17 = const()[name = tensor("op_17"), val = tensor(0x1.0c6f7ap-20)]; - tensor var_20 = const()[name = tensor("op_20"), val = tensor(2)]; - tensor var_23 = const()[name = tensor("op_23"), val = tensor(0)]; - tensor var_26 = const()[name = tensor("op_26"), val = tensor(1)]; - tensor var_28 = const()[name = tensor("op_28"), val = tensor(0x1.4f8b58p-17)]; - tensor var_29 = const()[name = tensor("op_29"), val = tensor(true)]; - tensor input_1 = linear(bias = encoder_input_projection_linear_bias, weight = encoder_input_projection_linear_weight, x = features)[name = tensor("linear_0")]; - tensor input_3_axes_0 = const()[name = tensor("input_3_axes_0"), val = tensor([-1])]; - tensor input_3 = layer_norm(axes = input_3_axes_0, beta = encoder_pre_layer_norm_bias, epsilon = var_28, gamma = encoder_pre_layer_norm_weight, x = input_1)[name = tensor("input_3")]; + func main(tensor cnn_window, tensor dec_kv, tensor dec_scale, tensor enc_conv_cache, tensor enc_kv, tensor enc_scale, tensor features, tensor valid_mask) { + tensor inner_encoder_cnn_bias = const()[name = tensor("inner_encoder_cnn_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(64)))]; + tensor inner_encoder_cnn_weight = const()[name = tensor("inner_encoder_cnn_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1152)))]; + tensor inner_encoder__causal_mask = const()[name = tensor("inner_encoder__causal_mask"), val = tensor([[0x1p+0]])]; + tensor inner_encoder__t_index = const()[name = tensor("inner_encoder__t_index"), val = tensor([0x1p+0])]; + tensor inner_encoder_input_projection_linear_bias = const()[name = tensor("inner_encoder_input_projection_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4981952)))]; + tensor inner_encoder_input_projection_linear_weight = const()[name = tensor("inner_encoder_input_projection_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4983040)))]; + tensor inner_encoder_pre_layer_norm_bias = const()[name = tensor("inner_encoder_pre_layer_norm_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(5336384)))]; + tensor inner_encoder_pre_layer_norm_weight = const()[name = tensor("inner_encoder_pre_layer_norm_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(5337472)))]; + tensor inner_encoder_ffn1_0_module_sequential_0_bias = const()[name = tensor("inner_encoder_ffn1_0_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(5338560)))]; + tensor inner_encoder_ffn1_0_module_sequential_0_weight = const()[name = tensor("inner_encoder_ffn1_0_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(5339648)))]; + tensor inner_encoder_ffn1_0_module_sequential_1_linear_bias = const()[name = tensor("inner_encoder_ffn1_0_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(5340736)))]; + tensor inner_encoder_ffn1_0_module_sequential_1_linear_weight = const()[name = tensor("inner_encoder_ffn1_0_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(5344896)))]; + tensor inner_encoder_ffn1_0_module_sequential_4_linear_bias = const()[name = tensor("inner_encoder_ffn1_0_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(6393536)))]; + tensor inner_encoder_ffn1_0_module_sequential_4_linear_weight = const()[name = tensor("inner_encoder_ffn1_0_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(6394624)))]; + tensor inner_encoder_ret_lns_0_bias = const()[name = tensor("inner_encoder_ret_lns_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7443264)))]; + tensor inner_encoder_ret_lns_0_weight = const()[name = tensor("inner_encoder_ret_lns_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7444352)))]; + tensor inner_encoder_q_proj_0_bias = const()[name = tensor("inner_encoder_q_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7445440)))]; + tensor inner_encoder_q_proj_0_weight = const()[name = tensor("inner_encoder_q_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7446528)))]; + tensor inner_encoder_k_proj_0_bias = const()[name = tensor("inner_encoder_k_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7708736)))]; + tensor inner_encoder_k_proj_0_weight = const()[name = tensor("inner_encoder_k_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7709824)))]; + tensor inner_encoder_v_proj_0_bias = const()[name = tensor("inner_encoder_v_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7972032)))]; + tensor inner_encoder_v_proj_0_weight = const()[name = tensor("inner_encoder_v_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7973120)))]; + tensor inner_encoder_g_proj_0_bias = const()[name = tensor("inner_encoder_g_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8235328)))]; + tensor inner_encoder_g_proj_0_weight = const()[name = tensor("inner_encoder_g_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8236416)))]; + tensor inner_encoder_out_proj_0_bias = const()[name = tensor("inner_encoder_out_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8498624)))]; + tensor inner_encoder_out_proj_0_weight = const()[name = tensor("inner_encoder_out_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8499712)))]; + tensor inner_encoder_conv_module_0_sequential_0_bias = const()[name = tensor("inner_encoder_conv_module_0_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8761920)))]; + tensor inner_encoder_conv_module_0_sequential_0_weight = const()[name = tensor("inner_encoder_conv_module_0_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8763008)))]; + tensor inner_encoder_conv_module_0_sequential_2_conv_bias = const()[name = tensor("inner_encoder_conv_module_0_sequential_2_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8764096)))]; + tensor inner_encoder_conv_module_0_sequential_2_conv_weight = const()[name = tensor("inner_encoder_conv_module_0_sequential_2_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8766208)))]; + tensor inner_encoder_conv_module_0_sequential_4_conv_weight = const()[name = tensor("inner_encoder_conv_module_0_sequential_4_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9290560)))]; + tensor inner_encoder_conv_module_0_sequential_5_running_var = const()[name = tensor("inner_encoder_conv_module_0_sequential_5_running_var"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9307008)))]; + tensor inner_encoder_conv_module_0_sequential_5_running_mean = const()[name = tensor("inner_encoder_conv_module_0_sequential_5_running_mean"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9308096)))]; + tensor inner_encoder_conv_module_0_sequential_5_bias = const()[name = tensor("inner_encoder_conv_module_0_sequential_5_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9309184)))]; + tensor inner_encoder_conv_module_0_sequential_5_weight = const()[name = tensor("inner_encoder_conv_module_0_sequential_5_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9310272)))]; + tensor inner_encoder_conv_module_0_sequential_7_conv_bias = const()[name = tensor("inner_encoder_conv_module_0_sequential_7_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9311360)))]; + tensor inner_encoder_conv_module_0_sequential_7_conv_weight = const()[name = tensor("inner_encoder_conv_module_0_sequential_7_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9312448)))]; + tensor inner_encoder_ffn2_0_module_sequential_0_bias = const()[name = tensor("inner_encoder_ffn2_0_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9574656)))]; + tensor inner_encoder_ffn2_0_module_sequential_0_weight = const()[name = tensor("inner_encoder_ffn2_0_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9575744)))]; + tensor inner_encoder_ffn2_0_module_sequential_1_linear_bias = const()[name = tensor("inner_encoder_ffn2_0_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9576832)))]; + tensor inner_encoder_ffn2_0_module_sequential_1_linear_weight = const()[name = tensor("inner_encoder_ffn2_0_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9580992)))]; + tensor inner_encoder_ffn2_0_module_sequential_4_linear_bias = const()[name = tensor("inner_encoder_ffn2_0_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(10629632)))]; + tensor inner_encoder_ffn2_0_module_sequential_4_linear_weight = const()[name = tensor("inner_encoder_ffn2_0_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(10630720)))]; + tensor inner_encoder_layer_norm_0_bias = const()[name = tensor("inner_encoder_layer_norm_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11679360)))]; + tensor inner_encoder_layer_norm_0_weight = const()[name = tensor("inner_encoder_layer_norm_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11680448)))]; + tensor inner_encoder_ffn1_1_module_sequential_0_bias = const()[name = tensor("inner_encoder_ffn1_1_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11681536)))]; + tensor inner_encoder_ffn1_1_module_sequential_0_weight = const()[name = tensor("inner_encoder_ffn1_1_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11682624)))]; + tensor inner_encoder_ffn1_1_module_sequential_1_linear_bias = const()[name = tensor("inner_encoder_ffn1_1_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11683712)))]; + tensor inner_encoder_ffn1_1_module_sequential_1_linear_weight = const()[name = tensor("inner_encoder_ffn1_1_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11687872)))]; + tensor inner_encoder_ffn1_1_module_sequential_4_linear_bias = const()[name = tensor("inner_encoder_ffn1_1_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(12736512)))]; + tensor inner_encoder_ffn1_1_module_sequential_4_linear_weight = const()[name = tensor("inner_encoder_ffn1_1_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(12737600)))]; + tensor inner_encoder_ret_lns_1_bias = const()[name = tensor("inner_encoder_ret_lns_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(13786240)))]; + tensor inner_encoder_ret_lns_1_weight = const()[name = tensor("inner_encoder_ret_lns_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(13787328)))]; + tensor inner_encoder_q_proj_1_bias = const()[name = tensor("inner_encoder_q_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(13788416)))]; + tensor inner_encoder_q_proj_1_weight = const()[name = tensor("inner_encoder_q_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(13789504)))]; + tensor inner_encoder_k_proj_1_bias = const()[name = tensor("inner_encoder_k_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14051712)))]; + tensor inner_encoder_k_proj_1_weight = const()[name = tensor("inner_encoder_k_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14052800)))]; + tensor inner_encoder_v_proj_1_bias = const()[name = tensor("inner_encoder_v_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14315008)))]; + tensor inner_encoder_v_proj_1_weight = const()[name = tensor("inner_encoder_v_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14316096)))]; + tensor inner_encoder_g_proj_1_bias = const()[name = tensor("inner_encoder_g_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14578304)))]; + tensor inner_encoder_g_proj_1_weight = const()[name = tensor("inner_encoder_g_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14579392)))]; + tensor inner_encoder_out_proj_1_bias = const()[name = tensor("inner_encoder_out_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14841600)))]; + tensor inner_encoder_out_proj_1_weight = const()[name = tensor("inner_encoder_out_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14842688)))]; + tensor inner_encoder_conv_module_1_sequential_0_bias = const()[name = tensor("inner_encoder_conv_module_1_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15104896)))]; + tensor inner_encoder_conv_module_1_sequential_0_weight = const()[name = tensor("inner_encoder_conv_module_1_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15105984)))]; + tensor inner_encoder_conv_module_1_sequential_2_conv_bias = const()[name = tensor("inner_encoder_conv_module_1_sequential_2_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15107072)))]; + tensor inner_encoder_conv_module_1_sequential_2_conv_weight = const()[name = tensor("inner_encoder_conv_module_1_sequential_2_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15109184)))]; + tensor inner_encoder_conv_module_1_sequential_4_conv_weight = const()[name = tensor("inner_encoder_conv_module_1_sequential_4_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15633536)))]; + tensor inner_encoder_conv_module_1_sequential_5_running_var = const()[name = tensor("inner_encoder_conv_module_1_sequential_5_running_var"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15649984)))]; + tensor inner_encoder_conv_module_1_sequential_5_running_mean = const()[name = tensor("inner_encoder_conv_module_1_sequential_5_running_mean"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15651072)))]; + tensor inner_encoder_conv_module_1_sequential_5_bias = const()[name = tensor("inner_encoder_conv_module_1_sequential_5_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15652160)))]; + tensor inner_encoder_conv_module_1_sequential_5_weight = const()[name = tensor("inner_encoder_conv_module_1_sequential_5_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15653248)))]; + tensor inner_encoder_conv_module_1_sequential_7_conv_bias = const()[name = tensor("inner_encoder_conv_module_1_sequential_7_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15654336)))]; + tensor inner_encoder_conv_module_1_sequential_7_conv_weight = const()[name = tensor("inner_encoder_conv_module_1_sequential_7_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15655424)))]; + tensor inner_encoder_ffn2_1_module_sequential_0_bias = const()[name = tensor("inner_encoder_ffn2_1_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15917632)))]; + tensor inner_encoder_ffn2_1_module_sequential_0_weight = const()[name = tensor("inner_encoder_ffn2_1_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15918720)))]; + tensor inner_encoder_ffn2_1_module_sequential_1_linear_bias = const()[name = tensor("inner_encoder_ffn2_1_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15919808)))]; + tensor inner_encoder_ffn2_1_module_sequential_1_linear_weight = const()[name = tensor("inner_encoder_ffn2_1_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15923968)))]; + tensor inner_encoder_ffn2_1_module_sequential_4_linear_bias = const()[name = tensor("inner_encoder_ffn2_1_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(16972608)))]; + tensor inner_encoder_ffn2_1_module_sequential_4_linear_weight = const()[name = tensor("inner_encoder_ffn2_1_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(16973696)))]; + tensor inner_encoder_layer_norm_1_bias = const()[name = tensor("inner_encoder_layer_norm_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18022336)))]; + tensor inner_encoder_layer_norm_1_weight = const()[name = tensor("inner_encoder_layer_norm_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18023424)))]; + tensor inner_encoder_ffn1_2_module_sequential_0_bias = const()[name = tensor("inner_encoder_ffn1_2_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18024512)))]; + tensor inner_encoder_ffn1_2_module_sequential_0_weight = const()[name = tensor("inner_encoder_ffn1_2_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18025600)))]; + tensor inner_encoder_ffn1_2_module_sequential_1_linear_bias = const()[name = tensor("inner_encoder_ffn1_2_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18026688)))]; + tensor inner_encoder_ffn1_2_module_sequential_1_linear_weight = const()[name = tensor("inner_encoder_ffn1_2_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18030848)))]; + tensor inner_encoder_ffn1_2_module_sequential_4_linear_bias = const()[name = tensor("inner_encoder_ffn1_2_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(19079488)))]; + tensor inner_encoder_ffn1_2_module_sequential_4_linear_weight = const()[name = tensor("inner_encoder_ffn1_2_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(19080576)))]; + tensor inner_encoder_ret_lns_2_bias = const()[name = tensor("inner_encoder_ret_lns_2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20129216)))]; + tensor inner_encoder_ret_lns_2_weight = const()[name = tensor("inner_encoder_ret_lns_2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20130304)))]; + tensor inner_encoder_q_proj_2_bias = const()[name = tensor("inner_encoder_q_proj_2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20131392)))]; + tensor inner_encoder_q_proj_2_weight = const()[name = tensor("inner_encoder_q_proj_2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20132480)))]; + tensor inner_encoder_k_proj_2_bias = const()[name = tensor("inner_encoder_k_proj_2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20394688)))]; + tensor inner_encoder_k_proj_2_weight = const()[name = tensor("inner_encoder_k_proj_2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20395776)))]; + tensor inner_encoder_v_proj_2_bias = const()[name = tensor("inner_encoder_v_proj_2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20657984)))]; + tensor inner_encoder_v_proj_2_weight = const()[name = tensor("inner_encoder_v_proj_2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20659072)))]; + tensor inner_encoder_g_proj_2_bias = const()[name = tensor("inner_encoder_g_proj_2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20921280)))]; + tensor inner_encoder_g_proj_2_weight = const()[name = tensor("inner_encoder_g_proj_2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20922368)))]; + tensor inner_encoder_out_proj_2_bias = const()[name = tensor("inner_encoder_out_proj_2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21184576)))]; + tensor inner_encoder_out_proj_2_weight = const()[name = tensor("inner_encoder_out_proj_2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21185664)))]; + tensor inner_encoder_conv_module_2_sequential_0_bias = const()[name = tensor("inner_encoder_conv_module_2_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21447872)))]; + tensor inner_encoder_conv_module_2_sequential_0_weight = const()[name = tensor("inner_encoder_conv_module_2_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21448960)))]; + tensor inner_encoder_conv_module_2_sequential_2_conv_bias = const()[name = tensor("inner_encoder_conv_module_2_sequential_2_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21450048)))]; + tensor inner_encoder_conv_module_2_sequential_2_conv_weight = const()[name = tensor("inner_encoder_conv_module_2_sequential_2_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21452160)))]; + tensor inner_encoder_conv_module_2_sequential_4_conv_weight = const()[name = tensor("inner_encoder_conv_module_2_sequential_4_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21976512)))]; + tensor inner_encoder_conv_module_2_sequential_5_running_var = const()[name = tensor("inner_encoder_conv_module_2_sequential_5_running_var"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21992960)))]; + tensor inner_encoder_conv_module_2_sequential_5_running_mean = const()[name = tensor("inner_encoder_conv_module_2_sequential_5_running_mean"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21994048)))]; + tensor inner_encoder_conv_module_2_sequential_5_bias = const()[name = tensor("inner_encoder_conv_module_2_sequential_5_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21995136)))]; + tensor inner_encoder_conv_module_2_sequential_5_weight = const()[name = tensor("inner_encoder_conv_module_2_sequential_5_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21996224)))]; + tensor inner_encoder_conv_module_2_sequential_7_conv_bias = const()[name = tensor("inner_encoder_conv_module_2_sequential_7_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21997312)))]; + tensor inner_encoder_conv_module_2_sequential_7_conv_weight = const()[name = tensor("inner_encoder_conv_module_2_sequential_7_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21998400)))]; + tensor inner_encoder_ffn2_2_module_sequential_0_bias = const()[name = tensor("inner_encoder_ffn2_2_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(22260608)))]; + tensor inner_encoder_ffn2_2_module_sequential_0_weight = const()[name = tensor("inner_encoder_ffn2_2_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(22261696)))]; + tensor inner_encoder_ffn2_2_module_sequential_1_linear_bias = const()[name = tensor("inner_encoder_ffn2_2_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(22262784)))]; + tensor inner_encoder_ffn2_2_module_sequential_1_linear_weight = const()[name = tensor("inner_encoder_ffn2_2_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(22266944)))]; + tensor inner_encoder_ffn2_2_module_sequential_4_linear_bias = const()[name = tensor("inner_encoder_ffn2_2_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(23315584)))]; + tensor inner_encoder_ffn2_2_module_sequential_4_linear_weight = const()[name = tensor("inner_encoder_ffn2_2_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(23316672)))]; + tensor inner_encoder_layer_norm_2_bias = const()[name = tensor("inner_encoder_layer_norm_2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(24365312)))]; + tensor inner_encoder_layer_norm_2_weight = const()[name = tensor("inner_encoder_layer_norm_2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(24366400)))]; + tensor inner_encoder_ffn1_3_module_sequential_0_bias = const()[name = tensor("inner_encoder_ffn1_3_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(24367488)))]; + tensor inner_encoder_ffn1_3_module_sequential_0_weight = const()[name = tensor("inner_encoder_ffn1_3_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(24368576)))]; + tensor inner_encoder_ffn1_3_module_sequential_1_linear_bias = const()[name = tensor("inner_encoder_ffn1_3_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(24369664)))]; + tensor inner_encoder_ffn1_3_module_sequential_1_linear_weight = const()[name = tensor("inner_encoder_ffn1_3_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(24373824)))]; + tensor inner_encoder_ffn1_3_module_sequential_4_linear_bias = const()[name = tensor("inner_encoder_ffn1_3_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(25422464)))]; + tensor inner_encoder_ffn1_3_module_sequential_4_linear_weight = const()[name = tensor("inner_encoder_ffn1_3_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(25423552)))]; + tensor inner_encoder_ret_lns_3_bias = const()[name = tensor("inner_encoder_ret_lns_3_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(26472192)))]; + tensor inner_encoder_ret_lns_3_weight = const()[name = tensor("inner_encoder_ret_lns_3_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(26473280)))]; + tensor inner_encoder_q_proj_3_bias = const()[name = tensor("inner_encoder_q_proj_3_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(26474368)))]; + tensor inner_encoder_q_proj_3_weight = const()[name = tensor("inner_encoder_q_proj_3_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(26475456)))]; + tensor inner_encoder_k_proj_3_bias = const()[name = tensor("inner_encoder_k_proj_3_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(26737664)))]; + tensor inner_encoder_k_proj_3_weight = const()[name = tensor("inner_encoder_k_proj_3_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(26738752)))]; + tensor inner_encoder_v_proj_3_bias = const()[name = tensor("inner_encoder_v_proj_3_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27000960)))]; + tensor inner_encoder_v_proj_3_weight = const()[name = tensor("inner_encoder_v_proj_3_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27002048)))]; + tensor inner_encoder_g_proj_3_bias = const()[name = tensor("inner_encoder_g_proj_3_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27264256)))]; + tensor inner_encoder_g_proj_3_weight = const()[name = tensor("inner_encoder_g_proj_3_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27265344)))]; + tensor inner_encoder_out_proj_3_bias = const()[name = tensor("inner_encoder_out_proj_3_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27527552)))]; + tensor inner_encoder_out_proj_3_weight = const()[name = tensor("inner_encoder_out_proj_3_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27528640)))]; + tensor inner_encoder_conv_module_3_sequential_0_bias = const()[name = tensor("inner_encoder_conv_module_3_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27790848)))]; + tensor inner_encoder_conv_module_3_sequential_0_weight = const()[name = tensor("inner_encoder_conv_module_3_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27791936)))]; + tensor inner_encoder_conv_module_3_sequential_2_conv_bias = const()[name = tensor("inner_encoder_conv_module_3_sequential_2_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27793024)))]; + tensor inner_encoder_conv_module_3_sequential_2_conv_weight = const()[name = tensor("inner_encoder_conv_module_3_sequential_2_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27795136)))]; + tensor inner_encoder_conv_module_3_sequential_4_conv_weight = const()[name = tensor("inner_encoder_conv_module_3_sequential_4_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28319488)))]; + tensor inner_encoder_conv_module_3_sequential_5_running_var = const()[name = tensor("inner_encoder_conv_module_3_sequential_5_running_var"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28335936)))]; + tensor inner_encoder_conv_module_3_sequential_5_running_mean = const()[name = tensor("inner_encoder_conv_module_3_sequential_5_running_mean"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28337024)))]; + tensor inner_encoder_conv_module_3_sequential_5_bias = const()[name = tensor("inner_encoder_conv_module_3_sequential_5_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28338112)))]; + tensor inner_encoder_conv_module_3_sequential_5_weight = const()[name = tensor("inner_encoder_conv_module_3_sequential_5_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28339200)))]; + tensor inner_encoder_conv_module_3_sequential_7_conv_bias = const()[name = tensor("inner_encoder_conv_module_3_sequential_7_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28340288)))]; + tensor inner_encoder_conv_module_3_sequential_7_conv_weight = const()[name = tensor("inner_encoder_conv_module_3_sequential_7_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28341376)))]; + tensor inner_encoder_ffn2_3_module_sequential_0_bias = const()[name = tensor("inner_encoder_ffn2_3_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28603584)))]; + tensor inner_encoder_ffn2_3_module_sequential_0_weight = const()[name = tensor("inner_encoder_ffn2_3_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28604672)))]; + tensor inner_encoder_ffn2_3_module_sequential_1_linear_bias = const()[name = tensor("inner_encoder_ffn2_3_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28605760)))]; + tensor inner_encoder_ffn2_3_module_sequential_1_linear_weight = const()[name = tensor("inner_encoder_ffn2_3_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28609920)))]; + tensor inner_encoder_ffn2_3_module_sequential_4_linear_bias = const()[name = tensor("inner_encoder_ffn2_3_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(29658560)))]; + tensor inner_encoder_ffn2_3_module_sequential_4_linear_weight = const()[name = tensor("inner_encoder_ffn2_3_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(29659648)))]; + tensor inner_encoder_layer_norm_3_bias = const()[name = tensor("inner_encoder_layer_norm_3_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(30708288)))]; + tensor inner_encoder_layer_norm_3_weight = const()[name = tensor("inner_encoder_layer_norm_3_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(30709376)))]; + tensor inner_decoder_convert_bias = const()[name = tensor("inner_decoder_convert_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(30710464)))]; + tensor inner_decoder_convert_weight = const()[name = tensor("inner_decoder_convert_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(30711552)))]; + tensor inner_decoder_q_proj_0_bias = const()[name = tensor("inner_decoder_q_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(31235904)))]; + tensor inner_decoder_q_proj_0_weight = const()[name = tensor("inner_decoder_q_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(31236992)))]; + tensor inner_decoder_k_proj_0_bias = const()[name = tensor("inner_decoder_k_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(31499200)))]; + tensor inner_decoder_k_proj_0_weight = const()[name = tensor("inner_decoder_k_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(31500288)))]; + tensor inner_decoder_v_proj_0_bias = const()[name = tensor("inner_decoder_v_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(31762496)))]; + tensor inner_decoder_v_proj_0_weight = const()[name = tensor("inner_decoder_v_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(31763584)))]; + tensor inner_decoder_g_proj_0_bias = const()[name = tensor("inner_decoder_g_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32025792)))]; + tensor inner_decoder_g_proj_0_weight = const()[name = tensor("inner_decoder_g_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32026880)))]; + tensor inner_decoder_out_proj_0_bias = const()[name = tensor("inner_decoder_out_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32289088)))]; + tensor inner_decoder_out_proj_0_weight = const()[name = tensor("inner_decoder_out_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32290176)))]; + tensor inner_decoder_norm11_0_bias = const()[name = tensor("inner_decoder_norm11_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32552384)))]; + tensor inner_decoder_norm11_0_weight = const()[name = tensor("inner_decoder_norm11_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32553472)))]; + tensor inner_decoder_self_attn2_0_out_proj_bias = const()[name = tensor("inner_decoder_self_attn2_0_out_proj_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32554560)))]; + tensor inner_decoder_self_attn2_0_out_proj_weight = const()[name = tensor("inner_decoder_self_attn2_0_out_proj_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32555648)))]; + tensor inner_decoder_self_attn2_0_in_proj_bias = const()[name = tensor("inner_decoder_self_attn2_0_in_proj_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32817856)))]; + tensor inner_decoder_self_attn2_0_in_proj_weight = const()[name = tensor("inner_decoder_self_attn2_0_in_proj_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32820992)))]; + tensor inner_decoder_norm21_0_bias = const()[name = tensor("inner_decoder_norm21_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(33607488)))]; + tensor inner_decoder_norm21_0_weight = const()[name = tensor("inner_decoder_norm21_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(33608576)))]; + tensor inner_decoder_linear1_0_bias = const()[name = tensor("inner_decoder_linear1_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(33609664)))]; + tensor inner_decoder_linear1_0_weight = const()[name = tensor("inner_decoder_linear1_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(33617920)))]; + tensor inner_decoder_linear2_0_bias = const()[name = tensor("inner_decoder_linear2_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(35715136)))]; + tensor inner_decoder_linear2_0_weight = const()[name = tensor("inner_decoder_linear2_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(35716224)))]; + tensor inner_decoder_norm22_0_bias = const()[name = tensor("inner_decoder_norm22_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(37813440)))]; + tensor inner_decoder_norm22_0_weight = const()[name = tensor("inner_decoder_norm22_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(37814528)))]; + tensor inner_decoder_q_proj_1_bias = const()[name = tensor("inner_decoder_q_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(37815616)))]; + tensor inner_decoder_q_proj_1_weight = const()[name = tensor("inner_decoder_q_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(37816704)))]; + tensor inner_decoder_k_proj_1_bias = const()[name = tensor("inner_decoder_k_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38078912)))]; + tensor inner_decoder_k_proj_1_weight = const()[name = tensor("inner_decoder_k_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38080000)))]; + tensor inner_decoder_v_proj_1_bias = const()[name = tensor("inner_decoder_v_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38342208)))]; + tensor inner_decoder_v_proj_1_weight = const()[name = tensor("inner_decoder_v_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38343296)))]; + tensor inner_decoder_g_proj_1_bias = const()[name = tensor("inner_decoder_g_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38605504)))]; + tensor inner_decoder_g_proj_1_weight = const()[name = tensor("inner_decoder_g_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38606592)))]; + tensor inner_decoder_out_proj_1_bias = const()[name = tensor("inner_decoder_out_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38868800)))]; + tensor inner_decoder_out_proj_1_weight = const()[name = tensor("inner_decoder_out_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38869888)))]; + tensor inner_decoder_norm11_1_bias = const()[name = tensor("inner_decoder_norm11_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(39132096)))]; + tensor inner_decoder_norm11_1_weight = const()[name = tensor("inner_decoder_norm11_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(39133184)))]; + tensor inner_decoder_self_attn2_1_out_proj_bias = const()[name = tensor("inner_decoder_self_attn2_1_out_proj_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(39134272)))]; + tensor inner_decoder_self_attn2_1_out_proj_weight = const()[name = tensor("inner_decoder_self_attn2_1_out_proj_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(39135360)))]; + tensor inner_decoder_self_attn2_1_in_proj_bias = const()[name = tensor("inner_decoder_self_attn2_1_in_proj_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(39397568)))]; + tensor inner_decoder_self_attn2_1_in_proj_weight = const()[name = tensor("inner_decoder_self_attn2_1_in_proj_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(39400704)))]; + tensor inner_decoder_norm21_1_bias = const()[name = tensor("inner_decoder_norm21_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(40187200)))]; + tensor inner_decoder_norm21_1_weight = const()[name = tensor("inner_decoder_norm21_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(40188288)))]; + tensor inner_decoder_linear1_1_bias = const()[name = tensor("inner_decoder_linear1_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(40189376)))]; + tensor inner_decoder_linear1_1_weight = const()[name = tensor("inner_decoder_linear1_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(40197632)))]; + tensor inner_decoder_linear2_1_bias = const()[name = tensor("inner_decoder_linear2_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(42294848)))]; + tensor inner_decoder_linear2_1_weight = const()[name = tensor("inner_decoder_linear2_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(42295936)))]; + tensor inner_decoder_norm22_1_bias = const()[name = tensor("inner_decoder_norm22_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(44393152)))]; + tensor inner_decoder_norm22_1_weight = const()[name = tensor("inner_decoder_norm22_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(44394240)))]; + tensor stacked_axes_0 = const()[name = tensor("stacked_axes_0"), val = tensor([1])]; + tensor stacked = expand_dims(axes = stacked_axes_0, x = features)[name = tensor("stacked")]; + tensor var_26 = const()[name = tensor("op_26"), val = tensor([1, 1, 345])]; + tensor input_1 = reshape(shape = var_26, x = stacked)[name = tensor("input_1")]; + tensor var_29 = const()[name = tensor("op_29"), val = tensor(0x1p+0)]; + tensor var_35 = const()[name = tensor("op_35"), val = tensor(true)]; + tensor var_36 = const()[name = tensor("op_36"), val = tensor(0x1.4f8b58p-17)]; + tensor var_39 = const()[name = tensor("op_39"), val = tensor(0)]; + tensor var_41 = const()[name = tensor("op_41"), val = tensor(2)]; + tensor var_42 = const()[name = tensor("op_42"), val = tensor(-1)]; + tensor var_44 = const()[name = tensor("op_44"), val = tensor(0x1.0c6f7ap-20)]; + tensor var_49 = const()[name = tensor("op_49"), val = tensor(0x1.5798eep-27)]; + tensor var_52 = const()[name = tensor("op_52"), val = tensor(1)]; + tensor input_3 = linear(bias = inner_encoder_input_projection_linear_bias, weight = inner_encoder_input_projection_linear_weight, x = input_1)[name = tensor("linear_0")]; tensor input_5_axes_0 = const()[name = tensor("input_5_axes_0"), val = tensor([-1])]; - tensor input_5 = layer_norm(axes = input_5_axes_0, beta = encoder_ffn1_0_module_sequential_0_bias, epsilon = var_28, gamma = encoder_ffn1_0_module_sequential_0_weight, x = input_3)[name = tensor("input_5")]; - tensor inputs_1 = linear(bias = encoder_ffn1_0_module_sequential_1_linear_bias, weight = encoder_ffn1_0_module_sequential_1_linear_weight, x = input_5)[name = tensor("linear_1")]; - tensor input_7 = silu(x = inputs_1)[name = tensor("input_7")]; - tensor input_11 = linear(bias = encoder_ffn1_0_module_sequential_4_linear_bias, weight = encoder_ffn1_0_module_sequential_4_linear_weight, x = input_7)[name = tensor("linear_2")]; - tensor var_147 = const()[name = tensor("op_147"), val = tensor(0x1p-1)]; - tensor var_148 = mul(x = input_11, y = var_147)[name = tensor("op_148")]; - tensor input_13 = add(x = var_148, y = input_3)[name = tensor("input_13")]; + tensor input_5 = layer_norm(axes = input_5_axes_0, beta = inner_encoder_pre_layer_norm_bias, epsilon = var_36, gamma = inner_encoder_pre_layer_norm_weight, x = input_3)[name = tensor("input_5")]; + tensor input_7_axes_0 = const()[name = tensor("input_7_axes_0"), val = tensor([-1])]; + tensor input_7 = layer_norm(axes = input_7_axes_0, beta = inner_encoder_ffn1_0_module_sequential_0_bias, epsilon = var_36, gamma = inner_encoder_ffn1_0_module_sequential_0_weight, x = input_5)[name = tensor("input_7")]; + tensor inputs_1 = linear(bias = inner_encoder_ffn1_0_module_sequential_1_linear_bias, weight = inner_encoder_ffn1_0_module_sequential_1_linear_weight, x = input_7)[name = tensor("linear_1")]; + tensor input_9 = silu(x = inputs_1)[name = tensor("input_9")]; + tensor input_13 = linear(bias = inner_encoder_ffn1_0_module_sequential_4_linear_bias, weight = inner_encoder_ffn1_0_module_sequential_4_linear_weight, x = input_9)[name = tensor("linear_2")]; + tensor var_173 = const()[name = tensor("op_173"), val = tensor(0x1p-1)]; + tensor var_174 = mul(x = input_13, y = var_173)[name = tensor("op_174")]; + tensor input_15 = add(x = var_174, y = input_5)[name = tensor("input_15")]; tensor x_1_axes_0 = const()[name = tensor("x_1_axes_0"), val = tensor([-1])]; - tensor x_1 = layer_norm(axes = x_1_axes_0, beta = encoder_ret_lns_0_bias, epsilon = var_28, gamma = encoder_ret_lns_0_weight, x = input_13)[name = tensor("x_1")]; + tensor x_1 = layer_norm(axes = x_1_axes_0, beta = inner_encoder_ret_lns_0_bias, epsilon = var_36, gamma = inner_encoder_ret_lns_0_weight, x = input_15)[name = tensor("x_1")]; tensor prev_kv_1_begin_0 = const()[name = tensor("prev_kv_1_begin_0"), val = tensor([0, 0, 0, 0, 0])]; tensor prev_kv_1_end_0 = const()[name = tensor("prev_kv_1_end_0"), val = tensor([1, 1, 4, 64, 64])]; tensor prev_kv_1_end_mask_0 = const()[name = tensor("prev_kv_1_end_mask_0"), val = tensor([false, true, true, true, true])]; @@ -238,139 +244,139 @@ program(1.0) tensor prev_scale_1_end_mask_0 = const()[name = tensor("prev_scale_1_end_mask_0"), val = tensor([false, true])]; tensor prev_scale_1_squeeze_mask_0 = const()[name = tensor("prev_scale_1_squeeze_mask_0"), val = tensor([true, false])]; tensor prev_scale_1 = slice_by_index(begin = prev_scale_1_begin_0, end = prev_scale_1_end_0, end_mask = prev_scale_1_end_mask_0, squeeze_mask = prev_scale_1_squeeze_mask_0, x = enc_scale)[name = tensor("prev_scale_1")]; - tensor var_162 = linear(bias = encoder_q_proj_0_bias, weight = encoder_q_proj_0_weight, x = x_1)[name = tensor("linear_3")]; - tensor var_163 = const()[name = tensor("op_163"), val = tensor([1, 1, 4, 64])]; - tensor var_164 = reshape(shape = var_163, x = var_162)[name = tensor("op_164")]; + tensor var_188 = linear(bias = inner_encoder_q_proj_0_bias, weight = inner_encoder_q_proj_0_weight, x = x_1)[name = tensor("linear_3")]; + tensor var_189 = const()[name = tensor("op_189"), val = tensor([1, 1, 4, 64])]; + tensor var_190 = reshape(shape = var_189, x = var_188)[name = tensor("op_190")]; tensor q_1_perm_0 = const()[name = tensor("q_1_perm_0"), val = tensor([0, 2, 1, 3])]; - tensor var_168 = linear(bias = encoder_k_proj_0_bias, weight = encoder_k_proj_0_weight, x = x_1)[name = tensor("linear_4")]; - tensor var_169 = const()[name = tensor("op_169"), val = tensor(0x1p-3)]; - tensor var_170 = mul(x = var_168, y = var_169)[name = tensor("op_170")]; - tensor var_171 = const()[name = tensor("op_171"), val = tensor([1, 1, 4, 64])]; - tensor var_172 = reshape(shape = var_171, x = var_170)[name = tensor("op_172")]; + tensor var_194 = linear(bias = inner_encoder_k_proj_0_bias, weight = inner_encoder_k_proj_0_weight, x = x_1)[name = tensor("linear_4")]; + tensor var_195 = const()[name = tensor("op_195"), val = tensor(0x1p-3)]; + tensor var_196 = mul(x = var_194, y = var_195)[name = tensor("op_196")]; + tensor var_197 = const()[name = tensor("op_197"), val = tensor([1, 1, 4, 64])]; + tensor var_198 = reshape(shape = var_197, x = var_196)[name = tensor("op_198")]; tensor k_1_perm_0 = const()[name = tensor("k_1_perm_0"), val = tensor([0, 2, 1, 3])]; - tensor var_176 = linear(bias = encoder_v_proj_0_bias, weight = encoder_v_proj_0_weight, x = x_1)[name = tensor("linear_5")]; - tensor var_177 = const()[name = tensor("op_177"), val = tensor([1, 1, 4, 64])]; - tensor var_178 = reshape(shape = var_177, x = var_176)[name = tensor("op_178")]; + tensor var_202 = linear(bias = inner_encoder_v_proj_0_bias, weight = inner_encoder_v_proj_0_weight, x = x_1)[name = tensor("linear_5")]; + tensor var_203 = const()[name = tensor("op_203"), val = tensor([1, 1, 4, 64])]; + tensor var_204 = reshape(shape = var_203, x = var_202)[name = tensor("op_204")]; tensor v_1_perm_0 = const()[name = tensor("v_1_perm_0"), val = tensor([0, 2, 1, 3])]; - tensor input_17 = linear(bias = encoder_g_proj_0_bias, weight = encoder_g_proj_0_weight, x = x_1)[name = tensor("linear_6")]; + tensor input_19 = linear(bias = inner_encoder_g_proj_0_bias, weight = inner_encoder_g_proj_0_weight, x = x_1)[name = tensor("linear_6")]; tensor sqrt_s0_1 = sqrt(x = prev_scale_1)[name = tensor("sqrt_s0_1")]; - tensor s_t_1 = add(x = prev_scale_1, y = encoder__t_index)[name = tensor("s_t_1")]; + tensor s_t_1 = add(x = prev_scale_1, y = inner_encoder__t_index)[name = tensor("s_t_1")]; tensor sqrt_s_t_1 = sqrt(x = s_t_1)[name = tensor("sqrt_s_t_1")]; tensor qk_1_transpose_x_1 = const()[name = tensor("qk_1_transpose_x_1"), val = tensor(false)]; tensor qk_1_transpose_y_1 = const()[name = tensor("qk_1_transpose_y_1"), val = tensor(true)]; - tensor k_1 = transpose(perm = k_1_perm_0, x = var_172)[name = tensor("transpose_57")]; - tensor q_1 = transpose(perm = q_1_perm_0, x = var_164)[name = tensor("transpose_58")]; + tensor k_1 = transpose(perm = k_1_perm_0, x = var_198)[name = tensor("transpose_57")]; + tensor q_1 = transpose(perm = q_1_perm_0, x = var_190)[name = tensor("transpose_58")]; tensor qk_1 = matmul(transpose_x = qk_1_transpose_x_1, transpose_y = qk_1_transpose_y_1, x = q_1, y = k_1)[name = tensor("qk_1")]; - tensor var_188 = const()[name = tensor("op_188"), val = tensor([1, 1])]; - tensor var_189 = reshape(shape = var_188, x = sqrt_s_t_1)[name = tensor("op_189")]; - tensor M_1 = real_div(x = encoder__causal_mask, y = var_189)[name = tensor("M_1")]; - tensor var_191 = mul(x = qk_1, y = M_1)[name = tensor("op_191")]; + tensor var_214 = const()[name = tensor("op_214"), val = tensor([1, 1])]; + tensor var_215 = reshape(shape = var_214, x = sqrt_s_t_1)[name = tensor("op_215")]; + tensor M_1 = real_div(x = inner_encoder__causal_mask, y = var_215)[name = tensor("M_1")]; + tensor var_217 = mul(x = qk_1, y = M_1)[name = tensor("op_217")]; tensor inner_1_transpose_x_0 = const()[name = tensor("inner_1_transpose_x_0"), val = tensor(false)]; tensor inner_1_transpose_y_0 = const()[name = tensor("inner_1_transpose_y_0"), val = tensor(false)]; - tensor v_1 = transpose(perm = v_1_perm_0, x = var_178)[name = tensor("transpose_56")]; - tensor inner_1 = matmul(transpose_x = inner_1_transpose_x_0, transpose_y = inner_1_transpose_y_0, x = var_191, y = v_1)[name = tensor("inner_1")]; - tensor var_193_transpose_x_0 = const()[name = tensor("op_193_transpose_x_0"), val = tensor(false)]; - tensor var_193_transpose_y_0 = const()[name = tensor("op_193_transpose_y_0"), val = tensor(false)]; - tensor var_193 = matmul(transpose_x = var_193_transpose_x_0, transpose_y = var_193_transpose_y_0, x = q_1, y = prev_kv_1)[name = tensor("op_193")]; - tensor var_194 = real_div(x = sqrt_s0_1, y = sqrt_s_t_1)[name = tensor("op_194")]; - tensor var_195 = const()[name = tensor("op_195"), val = tensor([1, 1, 1, 1])]; - tensor var_196 = reshape(shape = var_195, x = var_194)[name = tensor("op_196")]; - tensor cross_1 = mul(x = var_193, y = var_196)[name = tensor("cross_1")]; + tensor v_1 = transpose(perm = v_1_perm_0, x = var_204)[name = tensor("transpose_56")]; + tensor inner_1 = matmul(transpose_x = inner_1_transpose_x_0, transpose_y = inner_1_transpose_y_0, x = var_217, y = v_1)[name = tensor("inner_1")]; + tensor var_219_transpose_x_0 = const()[name = tensor("op_219_transpose_x_0"), val = tensor(false)]; + tensor var_219_transpose_y_0 = const()[name = tensor("op_219_transpose_y_0"), val = tensor(false)]; + tensor var_219 = matmul(transpose_x = var_219_transpose_x_0, transpose_y = var_219_transpose_y_0, x = q_1, y = prev_kv_1)[name = tensor("op_219")]; + tensor var_220 = real_div(x = sqrt_s0_1, y = sqrt_s_t_1)[name = tensor("op_220")]; + tensor var_221 = const()[name = tensor("op_221"), val = tensor([1, 1, 1, 1])]; + tensor var_222 = reshape(shape = var_221, x = var_220)[name = tensor("op_222")]; + tensor cross_1 = mul(x = var_219, y = var_222)[name = tensor("cross_1")]; tensor out_1 = add(x = inner_1, y = cross_1)[name = tensor("out_1")]; - tensor var_199 = mul(x = prev_kv_1, y = sqrt_s0_1)[name = tensor("op_199")]; - tensor var_201_transpose_x_1 = const()[name = tensor("op_201_transpose_x_1"), val = tensor(true)]; - tensor var_201_transpose_y_1 = const()[name = tensor("op_201_transpose_y_1"), val = tensor(false)]; - tensor var_201 = matmul(transpose_x = var_201_transpose_x_1, transpose_y = var_201_transpose_y_1, x = k_1, y = v_1)[name = tensor("op_201")]; - tensor new_kv_unnorm_1 = add(x = var_199, y = var_201)[name = tensor("new_kv_unnorm_1")]; - tensor var_203 = const()[name = tensor("op_203"), val = tensor(0x1p+0)]; - tensor new_scale_1 = add(x = prev_scale_1, y = var_203)[name = tensor("new_scale_1")]; - tensor var_205 = sqrt(x = new_scale_1)[name = tensor("op_205")]; - tensor var_206 = real_div(x = new_kv_unnorm_1, y = var_205)[name = tensor("op_206")]; - tensor var_207_perm_0 = const()[name = tensor("op_207_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_225 = mul(x = prev_kv_1, y = sqrt_s0_1)[name = tensor("op_225")]; + tensor var_227_transpose_x_1 = const()[name = tensor("op_227_transpose_x_1"), val = tensor(true)]; + tensor var_227_transpose_y_1 = const()[name = tensor("op_227_transpose_y_1"), val = tensor(false)]; + tensor var_227 = matmul(transpose_x = var_227_transpose_x_1, transpose_y = var_227_transpose_y_1, x = k_1, y = v_1)[name = tensor("op_227")]; + tensor new_kv_unnorm_1 = add(x = var_225, y = var_227)[name = tensor("new_kv_unnorm_1")]; + tensor var_229 = const()[name = tensor("op_229"), val = tensor(0x1p+0)]; + tensor new_scale_1 = add(x = prev_scale_1, y = var_229)[name = tensor("new_scale_1")]; + tensor var_231 = sqrt(x = new_scale_1)[name = tensor("op_231")]; + tensor var_232 = real_div(x = new_kv_unnorm_1, y = var_231)[name = tensor("op_232")]; + tensor var_233_perm_0 = const()[name = tensor("op_233_perm_0"), val = tensor([0, 2, 1, 3])]; tensor out_3_axes_0 = const()[name = tensor("out_3_axes_0"), val = tensor([-1])]; - tensor var_207 = transpose(perm = var_207_perm_0, x = out_1)[name = tensor("transpose_55")]; - tensor out_3 = layer_norm(axes = out_3_axes_0, epsilon = var_17, x = var_207)[name = tensor("out_3")]; - tensor var_211 = const()[name = tensor("op_211"), val = tensor([1, 1, 256])]; - tensor out_5 = reshape(shape = var_211, x = out_3)[name = tensor("out_5")]; - tensor var_213 = silu(x = input_17)[name = tensor("op_213")]; - tensor input_19 = mul(x = var_213, y = out_5)[name = tensor("input_19")]; - tensor ret_out_1 = linear(bias = encoder_out_proj_0_bias, weight = encoder_out_proj_0_weight, x = input_19)[name = tensor("linear_7")]; - tensor x_3 = add(x = input_13, y = ret_out_1)[name = tensor("x_3")]; + tensor var_233 = transpose(perm = var_233_perm_0, x = out_1)[name = tensor("transpose_55")]; + tensor out_3 = layer_norm(axes = out_3_axes_0, epsilon = var_44, x = var_233)[name = tensor("out_3")]; + tensor var_237 = const()[name = tensor("op_237"), val = tensor([1, 1, 256])]; + tensor out_5 = reshape(shape = var_237, x = out_3)[name = tensor("out_5")]; + tensor var_239 = silu(x = input_19)[name = tensor("op_239")]; + tensor input_21 = mul(x = var_239, y = out_5)[name = tensor("input_21")]; + tensor ret_out_1 = linear(bias = inner_encoder_out_proj_0_bias, weight = inner_encoder_out_proj_0_weight, x = input_21)[name = tensor("linear_7")]; + tensor x_3 = add(x = input_15, y = ret_out_1)[name = tensor("x_3")]; tensor window_1_begin_0 = const()[name = tensor("window_1_begin_0"), val = tensor([0, 0, 0, 0])]; tensor window_1_end_0 = const()[name = tensor("window_1_end_0"), val = tensor([1, 1, 16, 256])]; tensor window_1_end_mask_0 = const()[name = tensor("window_1_end_mask_0"), val = tensor([false, true, true, true])]; tensor window_1_squeeze_mask_0 = const()[name = tensor("window_1_squeeze_mask_0"), val = tensor([true, false, false, false])]; tensor window_1 = slice_by_index(begin = window_1_begin_0, end = window_1_end_0, end_mask = window_1_end_mask_0, squeeze_mask = window_1_squeeze_mask_0, x = enc_conv_cache)[name = tensor("window_1")]; - tensor var_224_begin_0 = const()[name = tensor("op_224_begin_0"), val = tensor([0, 1, 0])]; - tensor var_224_end_0 = const()[name = tensor("op_224_end_0"), val = tensor([1, 16, 256])]; - tensor var_224_end_mask_0 = const()[name = tensor("op_224_end_mask_0"), val = tensor([true, true, true])]; - tensor var_224 = slice_by_index(begin = var_224_begin_0, end = var_224_end_0, end_mask = var_224_end_mask_0, x = window_1)[name = tensor("op_224")]; + tensor var_250_begin_0 = const()[name = tensor("op_250_begin_0"), val = tensor([0, 1, 0])]; + tensor var_250_end_0 = const()[name = tensor("op_250_end_0"), val = tensor([1, 16, 256])]; + tensor var_250_end_mask_0 = const()[name = tensor("op_250_end_mask_0"), val = tensor([true, true, true])]; + tensor var_250 = slice_by_index(begin = var_250_begin_0, end = var_250_end_0, end_mask = var_250_end_mask_0, x = window_1)[name = tensor("op_250")]; tensor window_3_interleave_0 = const()[name = tensor("window_3_interleave_0"), val = tensor(false)]; - tensor window_3 = concat(axis = var_26, interleave = window_3_interleave_0, values = (var_224, x_3))[name = tensor("window_3")]; - tensor input_21_interleave_0 = const()[name = tensor("input_21_interleave_0"), val = tensor(false)]; - tensor input_21 = concat(axis = var_23, interleave = input_21_interleave_0, values = window_3)[name = tensor("input_21")]; + tensor window_3 = concat(axis = var_52, interleave = window_3_interleave_0, values = (var_250, x_3))[name = tensor("window_3")]; + tensor input_23_interleave_0 = const()[name = tensor("input_23_interleave_0"), val = tensor(false)]; + tensor input_23 = concat(axis = var_39, interleave = input_23_interleave_0, values = window_3)[name = tensor("input_23")]; tensor x_5_axes_0 = const()[name = tensor("x_5_axes_0"), val = tensor([-1])]; - tensor x_5 = layer_norm(axes = x_5_axes_0, beta = encoder_conv_module_0_sequential_0_bias, epsilon = var_28, gamma = encoder_conv_module_0_sequential_0_weight, x = input_21)[name = tensor("x_5")]; - tensor input_23_perm_0 = const()[name = tensor("input_23_perm_0"), val = tensor([0, 2, 1])]; + tensor x_5 = layer_norm(axes = x_5_axes_0, beta = inner_encoder_conv_module_0_sequential_0_bias, epsilon = var_36, gamma = inner_encoder_conv_module_0_sequential_0_weight, x = input_23)[name = tensor("x_5")]; + tensor input_25_perm_0 = const()[name = tensor("input_25_perm_0"), val = tensor([0, 2, 1])]; tensor inputs_3_pad_type_0 = const()[name = tensor("inputs_3_pad_type_0"), val = tensor("valid")]; tensor inputs_3_strides_0 = const()[name = tensor("inputs_3_strides_0"), val = tensor([1])]; tensor inputs_3_pad_0 = const()[name = tensor("inputs_3_pad_0"), val = tensor([0, 0])]; tensor inputs_3_dilations_0 = const()[name = tensor("inputs_3_dilations_0"), val = tensor([1])]; tensor inputs_3_groups_0 = const()[name = tensor("inputs_3_groups_0"), val = tensor(1)]; - tensor input_23 = transpose(perm = input_23_perm_0, x = x_5)[name = tensor("transpose_54")]; - tensor inputs_3 = conv(bias = encoder_conv_module_0_sequential_2_conv_bias, dilations = inputs_3_dilations_0, groups = inputs_3_groups_0, pad = inputs_3_pad_0, pad_type = inputs_3_pad_type_0, strides = inputs_3_strides_0, weight = encoder_conv_module_0_sequential_2_conv_weight, x = input_23)[name = tensor("inputs_3")]; - tensor var_249_split_sizes_0 = const()[name = tensor("op_249_split_sizes_0"), val = tensor([256, 256])]; - tensor var_249_axis_0 = const()[name = tensor("op_249_axis_0"), val = tensor(1)]; - tensor var_249_0, tensor var_249_1 = split(axis = var_249_axis_0, split_sizes = var_249_split_sizes_0, x = inputs_3)[name = tensor("op_249")]; - tensor var_251 = sigmoid(x = var_249_1)[name = tensor("op_251")]; - tensor inputs_5 = mul(x = var_249_0, y = var_251)[name = tensor("inputs_5")]; + tensor input_25 = transpose(perm = input_25_perm_0, x = x_5)[name = tensor("transpose_54")]; + tensor inputs_3 = conv(bias = inner_encoder_conv_module_0_sequential_2_conv_bias, dilations = inputs_3_dilations_0, groups = inputs_3_groups_0, pad = inputs_3_pad_0, pad_type = inputs_3_pad_type_0, strides = inputs_3_strides_0, weight = inner_encoder_conv_module_0_sequential_2_conv_weight, x = input_25)[name = tensor("inputs_3")]; + tensor var_275_split_sizes_0 = const()[name = tensor("op_275_split_sizes_0"), val = tensor([256, 256])]; + tensor var_275_axis_0 = const()[name = tensor("op_275_axis_0"), val = tensor(1)]; + tensor var_275_0, tensor var_275_1 = split(axis = var_275_axis_0, split_sizes = var_275_split_sizes_0, x = inputs_3)[name = tensor("op_275")]; + tensor var_277 = sigmoid(x = var_275_1)[name = tensor("op_277")]; + tensor inputs_5 = mul(x = var_275_0, y = var_277)[name = tensor("inputs_5")]; tensor outputs_aug_1_pad_type_0 = const()[name = tensor("outputs_aug_1_pad_type_0"), val = tensor("custom")]; tensor outputs_aug_1_pad_0 = const()[name = tensor("outputs_aug_1_pad_0"), val = tensor([15, 15])]; tensor outputs_aug_1_groups_0 = const()[name = tensor("outputs_aug_1_groups_0"), val = tensor(256)]; tensor outputs_aug_1_strides_0 = const()[name = tensor("outputs_aug_1_strides_0"), val = tensor([1])]; tensor outputs_aug_1_dilations_0 = const()[name = tensor("outputs_aug_1_dilations_0"), val = tensor([1])]; - tensor outputs_aug_1 = conv(dilations = outputs_aug_1_dilations_0, groups = outputs_aug_1_groups_0, pad = outputs_aug_1_pad_0, pad_type = outputs_aug_1_pad_type_0, strides = outputs_aug_1_strides_0, weight = encoder_conv_module_0_sequential_4_conv_weight, x = inputs_5)[name = tensor("outputs_aug_1")]; - tensor input_25_begin_0 = const()[name = tensor("input_25_begin_0"), val = tensor([0, 0, 0])]; - tensor input_25_end_0 = const()[name = tensor("input_25_end_0"), val = tensor([1, 256, 16])]; - tensor input_25_end_mask_0 = const()[name = tensor("input_25_end_mask_0"), val = tensor([true, true, false])]; - tensor input_25 = slice_by_index(begin = input_25_begin_0, end = input_25_end_0, end_mask = input_25_end_mask_0, x = outputs_aug_1)[name = tensor("input_25")]; - tensor inputs_7 = batch_norm(beta = encoder_conv_module_0_sequential_5_bias, epsilon = var_28, gamma = encoder_conv_module_0_sequential_5_weight, mean = encoder_conv_module_0_sequential_5_running_mean, variance = encoder_conv_module_0_sequential_5_running_var, x = input_25)[name = tensor("inputs_7")]; - tensor input_27 = silu(x = inputs_7)[name = tensor("input_27")]; - tensor input_29_pad_type_0 = const()[name = tensor("input_29_pad_type_0"), val = tensor("valid")]; - tensor input_29_strides_0 = const()[name = tensor("input_29_strides_0"), val = tensor([1])]; - tensor input_29_pad_0 = const()[name = tensor("input_29_pad_0"), val = tensor([0, 0])]; - tensor input_29_dilations_0 = const()[name = tensor("input_29_dilations_0"), val = tensor([1])]; - tensor input_29_groups_0 = const()[name = tensor("input_29_groups_0"), val = tensor(1)]; - tensor input_29 = conv(bias = encoder_conv_module_0_sequential_7_conv_bias, dilations = input_29_dilations_0, groups = input_29_groups_0, pad = input_29_pad_0, pad_type = input_29_pad_type_0, strides = input_29_strides_0, weight = encoder_conv_module_0_sequential_7_conv_weight, x = input_27)[name = tensor("input_29")]; + tensor outputs_aug_1 = conv(dilations = outputs_aug_1_dilations_0, groups = outputs_aug_1_groups_0, pad = outputs_aug_1_pad_0, pad_type = outputs_aug_1_pad_type_0, strides = outputs_aug_1_strides_0, weight = inner_encoder_conv_module_0_sequential_4_conv_weight, x = inputs_5)[name = tensor("outputs_aug_1")]; + tensor input_27_begin_0 = const()[name = tensor("input_27_begin_0"), val = tensor([0, 0, 0])]; + tensor input_27_end_0 = const()[name = tensor("input_27_end_0"), val = tensor([1, 256, 16])]; + tensor input_27_end_mask_0 = const()[name = tensor("input_27_end_mask_0"), val = tensor([true, true, false])]; + tensor input_27 = slice_by_index(begin = input_27_begin_0, end = input_27_end_0, end_mask = input_27_end_mask_0, x = outputs_aug_1)[name = tensor("input_27")]; + tensor inputs_7 = batch_norm(beta = inner_encoder_conv_module_0_sequential_5_bias, epsilon = var_36, gamma = inner_encoder_conv_module_0_sequential_5_weight, mean = inner_encoder_conv_module_0_sequential_5_running_mean, variance = inner_encoder_conv_module_0_sequential_5_running_var, x = input_27)[name = tensor("inputs_7")]; + tensor input_29 = silu(x = inputs_7)[name = tensor("input_29")]; + tensor input_31_pad_type_0 = const()[name = tensor("input_31_pad_type_0"), val = tensor("valid")]; + tensor input_31_strides_0 = const()[name = tensor("input_31_strides_0"), val = tensor([1])]; + tensor input_31_pad_0 = const()[name = tensor("input_31_pad_0"), val = tensor([0, 0])]; + tensor input_31_dilations_0 = const()[name = tensor("input_31_dilations_0"), val = tensor([1])]; + tensor input_31_groups_0 = const()[name = tensor("input_31_groups_0"), val = tensor(1)]; + tensor input_31 = conv(bias = inner_encoder_conv_module_0_sequential_7_conv_bias, dilations = input_31_dilations_0, groups = input_31_groups_0, pad = input_31_pad_0, pad_type = input_31_pad_type_0, strides = input_31_strides_0, weight = inner_encoder_conv_module_0_sequential_7_conv_weight, x = input_29)[name = tensor("input_31")]; tensor conv_out_1_perm_0 = const()[name = tensor("conv_out_1_perm_0"), val = tensor([0, 2, 1])]; - tensor var_282_begin_0 = const()[name = tensor("op_282_begin_0"), val = tensor([0, -1, 0])]; - tensor var_282_end_0 = const()[name = tensor("op_282_end_0"), val = tensor([1, 16, 256])]; - tensor var_282_end_mask_0 = const()[name = tensor("op_282_end_mask_0"), val = tensor([true, true, true])]; - tensor conv_out_1 = transpose(perm = conv_out_1_perm_0, x = input_29)[name = tensor("transpose_53")]; - tensor var_282 = slice_by_index(begin = var_282_begin_0, end = var_282_end_0, end_mask = var_282_end_mask_0, x = conv_out_1)[name = tensor("op_282")]; - tensor var_284_perm_0 = const()[name = tensor("op_284_perm_0"), val = tensor([1, 0, 2])]; - tensor var_284 = transpose(perm = var_284_perm_0, x = var_282)[name = tensor("transpose_52")]; - tensor input_31 = add(x = x_3, y = var_284)[name = tensor("input_31")]; - tensor input_33_axes_0 = const()[name = tensor("input_33_axes_0"), val = tensor([-1])]; - tensor input_33 = layer_norm(axes = input_33_axes_0, beta = encoder_ffn2_0_module_sequential_0_bias, epsilon = var_28, gamma = encoder_ffn2_0_module_sequential_0_weight, x = input_31)[name = tensor("input_33")]; - tensor inputs_9 = linear(bias = encoder_ffn2_0_module_sequential_1_linear_bias, weight = encoder_ffn2_0_module_sequential_1_linear_weight, x = input_33)[name = tensor("linear_8")]; - tensor input_35 = silu(x = inputs_9)[name = tensor("input_35")]; - tensor input_39 = linear(bias = encoder_ffn2_0_module_sequential_4_linear_bias, weight = encoder_ffn2_0_module_sequential_4_linear_weight, x = input_35)[name = tensor("linear_9")]; - tensor var_307 = const()[name = tensor("op_307"), val = tensor(0x1p-1)]; - tensor var_308 = mul(x = input_39, y = var_307)[name = tensor("op_308")]; - tensor input_41 = add(x = var_308, y = input_31)[name = tensor("input_41")]; - tensor input_43_axes_0 = const()[name = tensor("input_43_axes_0"), val = tensor([-1])]; - tensor input_43 = layer_norm(axes = input_43_axes_0, beta = encoder_layer_norm_0_bias, epsilon = var_28, gamma = encoder_layer_norm_0_weight, x = input_41)[name = tensor("input_43")]; + tensor var_308_begin_0 = const()[name = tensor("op_308_begin_0"), val = tensor([0, -1, 0])]; + tensor var_308_end_0 = const()[name = tensor("op_308_end_0"), val = tensor([1, 16, 256])]; + tensor var_308_end_mask_0 = const()[name = tensor("op_308_end_mask_0"), val = tensor([true, true, true])]; + tensor conv_out_1 = transpose(perm = conv_out_1_perm_0, x = input_31)[name = tensor("transpose_53")]; + tensor var_308 = slice_by_index(begin = var_308_begin_0, end = var_308_end_0, end_mask = var_308_end_mask_0, x = conv_out_1)[name = tensor("op_308")]; + tensor var_310_perm_0 = const()[name = tensor("op_310_perm_0"), val = tensor([1, 0, 2])]; + tensor var_310 = transpose(perm = var_310_perm_0, x = var_308)[name = tensor("transpose_52")]; + tensor input_33 = add(x = x_3, y = var_310)[name = tensor("input_33")]; + tensor input_35_axes_0 = const()[name = tensor("input_35_axes_0"), val = tensor([-1])]; + tensor input_35 = layer_norm(axes = input_35_axes_0, beta = inner_encoder_ffn2_0_module_sequential_0_bias, epsilon = var_36, gamma = inner_encoder_ffn2_0_module_sequential_0_weight, x = input_33)[name = tensor("input_35")]; + tensor inputs_9 = linear(bias = inner_encoder_ffn2_0_module_sequential_1_linear_bias, weight = inner_encoder_ffn2_0_module_sequential_1_linear_weight, x = input_35)[name = tensor("linear_8")]; + tensor input_37 = silu(x = inputs_9)[name = tensor("input_37")]; + tensor input_41 = linear(bias = inner_encoder_ffn2_0_module_sequential_4_linear_bias, weight = inner_encoder_ffn2_0_module_sequential_4_linear_weight, x = input_37)[name = tensor("linear_9")]; + tensor var_333 = const()[name = tensor("op_333"), val = tensor(0x1p-1)]; + tensor var_334 = mul(x = input_41, y = var_333)[name = tensor("op_334")]; + tensor input_43 = add(x = var_334, y = input_33)[name = tensor("input_43")]; tensor input_45_axes_0 = const()[name = tensor("input_45_axes_0"), val = tensor([-1])]; - tensor input_45 = layer_norm(axes = input_45_axes_0, beta = encoder_ffn1_1_module_sequential_0_bias, epsilon = var_28, gamma = encoder_ffn1_1_module_sequential_0_weight, x = input_43)[name = tensor("input_45")]; - tensor inputs_11 = linear(bias = encoder_ffn1_1_module_sequential_1_linear_bias, weight = encoder_ffn1_1_module_sequential_1_linear_weight, x = input_45)[name = tensor("linear_10")]; - tensor input_47 = silu(x = inputs_11)[name = tensor("input_47")]; - tensor input_51 = linear(bias = encoder_ffn1_1_module_sequential_4_linear_bias, weight = encoder_ffn1_1_module_sequential_4_linear_weight, x = input_47)[name = tensor("linear_11")]; - tensor var_337 = const()[name = tensor("op_337"), val = tensor(0x1p-1)]; - tensor var_338 = mul(x = input_51, y = var_337)[name = tensor("op_338")]; - tensor input_53 = add(x = var_338, y = input_43)[name = tensor("input_53")]; + tensor input_45 = layer_norm(axes = input_45_axes_0, beta = inner_encoder_layer_norm_0_bias, epsilon = var_36, gamma = inner_encoder_layer_norm_0_weight, x = input_43)[name = tensor("input_45")]; + tensor input_47_axes_0 = const()[name = tensor("input_47_axes_0"), val = tensor([-1])]; + tensor input_47 = layer_norm(axes = input_47_axes_0, beta = inner_encoder_ffn1_1_module_sequential_0_bias, epsilon = var_36, gamma = inner_encoder_ffn1_1_module_sequential_0_weight, x = input_45)[name = tensor("input_47")]; + tensor inputs_11 = linear(bias = inner_encoder_ffn1_1_module_sequential_1_linear_bias, weight = inner_encoder_ffn1_1_module_sequential_1_linear_weight, x = input_47)[name = tensor("linear_10")]; + tensor input_49 = silu(x = inputs_11)[name = tensor("input_49")]; + tensor input_53 = linear(bias = inner_encoder_ffn1_1_module_sequential_4_linear_bias, weight = inner_encoder_ffn1_1_module_sequential_4_linear_weight, x = input_49)[name = tensor("linear_11")]; + tensor var_363 = const()[name = tensor("op_363"), val = tensor(0x1p-1)]; + tensor var_364 = mul(x = input_53, y = var_363)[name = tensor("op_364")]; + tensor input_55 = add(x = var_364, y = input_45)[name = tensor("input_55")]; tensor x_7_axes_0 = const()[name = tensor("x_7_axes_0"), val = tensor([-1])]; - tensor x_7 = layer_norm(axes = x_7_axes_0, beta = encoder_ret_lns_1_bias, epsilon = var_28, gamma = encoder_ret_lns_1_weight, x = input_53)[name = tensor("x_7")]; + tensor x_7 = layer_norm(axes = x_7_axes_0, beta = inner_encoder_ret_lns_1_bias, epsilon = var_36, gamma = inner_encoder_ret_lns_1_weight, x = input_55)[name = tensor("x_7")]; tensor prev_kv_3_begin_0 = const()[name = tensor("prev_kv_3_begin_0"), val = tensor([1, 0, 0, 0, 0])]; tensor prev_kv_3_end_0 = const()[name = tensor("prev_kv_3_end_0"), val = tensor([2, 1, 4, 64, 64])]; tensor prev_kv_3_end_mask_0 = const()[name = tensor("prev_kv_3_end_mask_0"), val = tensor([false, true, true, true, true])]; @@ -381,139 +387,139 @@ program(1.0) tensor prev_scale_3_end_mask_0 = const()[name = tensor("prev_scale_3_end_mask_0"), val = tensor([false, true])]; tensor prev_scale_3_squeeze_mask_0 = const()[name = tensor("prev_scale_3_squeeze_mask_0"), val = tensor([true, false])]; tensor prev_scale_3 = slice_by_index(begin = prev_scale_3_begin_0, end = prev_scale_3_end_0, end_mask = prev_scale_3_end_mask_0, squeeze_mask = prev_scale_3_squeeze_mask_0, x = enc_scale)[name = tensor("prev_scale_3")]; - tensor var_352 = linear(bias = encoder_q_proj_1_bias, weight = encoder_q_proj_1_weight, x = x_7)[name = tensor("linear_12")]; - tensor var_353 = const()[name = tensor("op_353"), val = tensor([1, 1, 4, 64])]; - tensor var_354 = reshape(shape = var_353, x = var_352)[name = tensor("op_354")]; + tensor var_378 = linear(bias = inner_encoder_q_proj_1_bias, weight = inner_encoder_q_proj_1_weight, x = x_7)[name = tensor("linear_12")]; + tensor var_379 = const()[name = tensor("op_379"), val = tensor([1, 1, 4, 64])]; + tensor var_380 = reshape(shape = var_379, x = var_378)[name = tensor("op_380")]; tensor q_3_perm_0 = const()[name = tensor("q_3_perm_0"), val = tensor([0, 2, 1, 3])]; - tensor var_358 = linear(bias = encoder_k_proj_1_bias, weight = encoder_k_proj_1_weight, x = x_7)[name = tensor("linear_13")]; - tensor var_359 = const()[name = tensor("op_359"), val = tensor(0x1p-3)]; - tensor var_360 = mul(x = var_358, y = var_359)[name = tensor("op_360")]; - tensor var_361 = const()[name = tensor("op_361"), val = tensor([1, 1, 4, 64])]; - tensor var_362 = reshape(shape = var_361, x = var_360)[name = tensor("op_362")]; + tensor var_384 = linear(bias = inner_encoder_k_proj_1_bias, weight = inner_encoder_k_proj_1_weight, x = x_7)[name = tensor("linear_13")]; + tensor var_385 = const()[name = tensor("op_385"), val = tensor(0x1p-3)]; + tensor var_386 = mul(x = var_384, y = var_385)[name = tensor("op_386")]; + tensor var_387 = const()[name = tensor("op_387"), val = tensor([1, 1, 4, 64])]; + tensor var_388 = reshape(shape = var_387, x = var_386)[name = tensor("op_388")]; tensor k_3_perm_0 = const()[name = tensor("k_3_perm_0"), val = tensor([0, 2, 1, 3])]; - tensor var_366 = linear(bias = encoder_v_proj_1_bias, weight = encoder_v_proj_1_weight, x = x_7)[name = tensor("linear_14")]; - tensor var_367 = const()[name = tensor("op_367"), val = tensor([1, 1, 4, 64])]; - tensor var_368 = reshape(shape = var_367, x = var_366)[name = tensor("op_368")]; + tensor var_392 = linear(bias = inner_encoder_v_proj_1_bias, weight = inner_encoder_v_proj_1_weight, x = x_7)[name = tensor("linear_14")]; + tensor var_393 = const()[name = tensor("op_393"), val = tensor([1, 1, 4, 64])]; + tensor var_394 = reshape(shape = var_393, x = var_392)[name = tensor("op_394")]; tensor v_3_perm_0 = const()[name = tensor("v_3_perm_0"), val = tensor([0, 2, 1, 3])]; - tensor input_57 = linear(bias = encoder_g_proj_1_bias, weight = encoder_g_proj_1_weight, x = x_7)[name = tensor("linear_15")]; + tensor input_59 = linear(bias = inner_encoder_g_proj_1_bias, weight = inner_encoder_g_proj_1_weight, x = x_7)[name = tensor("linear_15")]; tensor sqrt_s0_3 = sqrt(x = prev_scale_3)[name = tensor("sqrt_s0_3")]; - tensor s_t_3 = add(x = prev_scale_3, y = encoder__t_index)[name = tensor("s_t_3")]; + tensor s_t_3 = add(x = prev_scale_3, y = inner_encoder__t_index)[name = tensor("s_t_3")]; tensor sqrt_s_t_3 = sqrt(x = s_t_3)[name = tensor("sqrt_s_t_3")]; tensor qk_3_transpose_x_1 = const()[name = tensor("qk_3_transpose_x_1"), val = tensor(false)]; tensor qk_3_transpose_y_1 = const()[name = tensor("qk_3_transpose_y_1"), val = tensor(true)]; - tensor k_3 = transpose(perm = k_3_perm_0, x = var_362)[name = tensor("transpose_50")]; - tensor q_3 = transpose(perm = q_3_perm_0, x = var_354)[name = tensor("transpose_51")]; + tensor k_3 = transpose(perm = k_3_perm_0, x = var_388)[name = tensor("transpose_50")]; + tensor q_3 = transpose(perm = q_3_perm_0, x = var_380)[name = tensor("transpose_51")]; tensor qk_3 = matmul(transpose_x = qk_3_transpose_x_1, transpose_y = qk_3_transpose_y_1, x = q_3, y = k_3)[name = tensor("qk_3")]; - tensor var_378 = const()[name = tensor("op_378"), val = tensor([1, 1])]; - tensor var_379 = reshape(shape = var_378, x = sqrt_s_t_3)[name = tensor("op_379")]; - tensor M_3 = real_div(x = encoder__causal_mask, y = var_379)[name = tensor("M_3")]; - tensor var_381 = mul(x = qk_3, y = M_3)[name = tensor("op_381")]; + tensor var_404 = const()[name = tensor("op_404"), val = tensor([1, 1])]; + tensor var_405 = reshape(shape = var_404, x = sqrt_s_t_3)[name = tensor("op_405")]; + tensor M_3 = real_div(x = inner_encoder__causal_mask, y = var_405)[name = tensor("M_3")]; + tensor var_407 = mul(x = qk_3, y = M_3)[name = tensor("op_407")]; tensor inner_3_transpose_x_0 = const()[name = tensor("inner_3_transpose_x_0"), val = tensor(false)]; tensor inner_3_transpose_y_0 = const()[name = tensor("inner_3_transpose_y_0"), val = tensor(false)]; - tensor v_3 = transpose(perm = v_3_perm_0, x = var_368)[name = tensor("transpose_49")]; - tensor inner_3 = matmul(transpose_x = inner_3_transpose_x_0, transpose_y = inner_3_transpose_y_0, x = var_381, y = v_3)[name = tensor("inner_3")]; - tensor var_383_transpose_x_0 = const()[name = tensor("op_383_transpose_x_0"), val = tensor(false)]; - tensor var_383_transpose_y_0 = const()[name = tensor("op_383_transpose_y_0"), val = tensor(false)]; - tensor var_383 = matmul(transpose_x = var_383_transpose_x_0, transpose_y = var_383_transpose_y_0, x = q_3, y = prev_kv_3)[name = tensor("op_383")]; - tensor var_384 = real_div(x = sqrt_s0_3, y = sqrt_s_t_3)[name = tensor("op_384")]; - tensor var_385 = const()[name = tensor("op_385"), val = tensor([1, 1, 1, 1])]; - tensor var_386 = reshape(shape = var_385, x = var_384)[name = tensor("op_386")]; - tensor cross_3 = mul(x = var_383, y = var_386)[name = tensor("cross_3")]; + tensor v_3 = transpose(perm = v_3_perm_0, x = var_394)[name = tensor("transpose_49")]; + tensor inner_3 = matmul(transpose_x = inner_3_transpose_x_0, transpose_y = inner_3_transpose_y_0, x = var_407, y = v_3)[name = tensor("inner_3")]; + tensor var_409_transpose_x_0 = const()[name = tensor("op_409_transpose_x_0"), val = tensor(false)]; + tensor var_409_transpose_y_0 = const()[name = tensor("op_409_transpose_y_0"), val = tensor(false)]; + tensor var_409 = matmul(transpose_x = var_409_transpose_x_0, transpose_y = var_409_transpose_y_0, x = q_3, y = prev_kv_3)[name = tensor("op_409")]; + tensor var_410 = real_div(x = sqrt_s0_3, y = sqrt_s_t_3)[name = tensor("op_410")]; + tensor var_411 = const()[name = tensor("op_411"), val = tensor([1, 1, 1, 1])]; + tensor var_412 = reshape(shape = var_411, x = var_410)[name = tensor("op_412")]; + tensor cross_3 = mul(x = var_409, y = var_412)[name = tensor("cross_3")]; tensor out_7 = add(x = inner_3, y = cross_3)[name = tensor("out_7")]; - tensor var_389 = mul(x = prev_kv_3, y = sqrt_s0_3)[name = tensor("op_389")]; - tensor var_391_transpose_x_1 = const()[name = tensor("op_391_transpose_x_1"), val = tensor(true)]; - tensor var_391_transpose_y_1 = const()[name = tensor("op_391_transpose_y_1"), val = tensor(false)]; - tensor var_391 = matmul(transpose_x = var_391_transpose_x_1, transpose_y = var_391_transpose_y_1, x = k_3, y = v_3)[name = tensor("op_391")]; - tensor new_kv_unnorm_3 = add(x = var_389, y = var_391)[name = tensor("new_kv_unnorm_3")]; - tensor var_393 = const()[name = tensor("op_393"), val = tensor(0x1p+0)]; - tensor new_scale_3 = add(x = prev_scale_3, y = var_393)[name = tensor("new_scale_3")]; - tensor var_395 = sqrt(x = new_scale_3)[name = tensor("op_395")]; - tensor var_396 = real_div(x = new_kv_unnorm_3, y = var_395)[name = tensor("op_396")]; - tensor var_397_perm_0 = const()[name = tensor("op_397_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_415 = mul(x = prev_kv_3, y = sqrt_s0_3)[name = tensor("op_415")]; + tensor var_417_transpose_x_1 = const()[name = tensor("op_417_transpose_x_1"), val = tensor(true)]; + tensor var_417_transpose_y_1 = const()[name = tensor("op_417_transpose_y_1"), val = tensor(false)]; + tensor var_417 = matmul(transpose_x = var_417_transpose_x_1, transpose_y = var_417_transpose_y_1, x = k_3, y = v_3)[name = tensor("op_417")]; + tensor new_kv_unnorm_3 = add(x = var_415, y = var_417)[name = tensor("new_kv_unnorm_3")]; + tensor var_419 = const()[name = tensor("op_419"), val = tensor(0x1p+0)]; + tensor new_scale_3 = add(x = prev_scale_3, y = var_419)[name = tensor("new_scale_3")]; + tensor var_421 = sqrt(x = new_scale_3)[name = tensor("op_421")]; + tensor var_422 = real_div(x = new_kv_unnorm_3, y = var_421)[name = tensor("op_422")]; + tensor var_423_perm_0 = const()[name = tensor("op_423_perm_0"), val = tensor([0, 2, 1, 3])]; tensor out_9_axes_0 = const()[name = tensor("out_9_axes_0"), val = tensor([-1])]; - tensor var_397 = transpose(perm = var_397_perm_0, x = out_7)[name = tensor("transpose_48")]; - tensor out_9 = layer_norm(axes = out_9_axes_0, epsilon = var_17, x = var_397)[name = tensor("out_9")]; - tensor var_401 = const()[name = tensor("op_401"), val = tensor([1, 1, 256])]; - tensor out_11 = reshape(shape = var_401, x = out_9)[name = tensor("out_11")]; - tensor var_403 = silu(x = input_57)[name = tensor("op_403")]; - tensor input_59 = mul(x = var_403, y = out_11)[name = tensor("input_59")]; - tensor ret_out_3 = linear(bias = encoder_out_proj_1_bias, weight = encoder_out_proj_1_weight, x = input_59)[name = tensor("linear_16")]; - tensor x_9 = add(x = input_53, y = ret_out_3)[name = tensor("x_9")]; + tensor var_423 = transpose(perm = var_423_perm_0, x = out_7)[name = tensor("transpose_48")]; + tensor out_9 = layer_norm(axes = out_9_axes_0, epsilon = var_44, x = var_423)[name = tensor("out_9")]; + tensor var_427 = const()[name = tensor("op_427"), val = tensor([1, 1, 256])]; + tensor out_11 = reshape(shape = var_427, x = out_9)[name = tensor("out_11")]; + tensor var_429 = silu(x = input_59)[name = tensor("op_429")]; + tensor input_61 = mul(x = var_429, y = out_11)[name = tensor("input_61")]; + tensor ret_out_3 = linear(bias = inner_encoder_out_proj_1_bias, weight = inner_encoder_out_proj_1_weight, x = input_61)[name = tensor("linear_16")]; + tensor x_9 = add(x = input_55, y = ret_out_3)[name = tensor("x_9")]; tensor window_5_begin_0 = const()[name = tensor("window_5_begin_0"), val = tensor([1, 0, 0, 0])]; tensor window_5_end_0 = const()[name = tensor("window_5_end_0"), val = tensor([2, 1, 16, 256])]; tensor window_5_end_mask_0 = const()[name = tensor("window_5_end_mask_0"), val = tensor([false, true, true, true])]; tensor window_5_squeeze_mask_0 = const()[name = tensor("window_5_squeeze_mask_0"), val = tensor([true, false, false, false])]; tensor window_5 = slice_by_index(begin = window_5_begin_0, end = window_5_end_0, end_mask = window_5_end_mask_0, squeeze_mask = window_5_squeeze_mask_0, x = enc_conv_cache)[name = tensor("window_5")]; - tensor var_414_begin_0 = const()[name = tensor("op_414_begin_0"), val = tensor([0, 1, 0])]; - tensor var_414_end_0 = const()[name = tensor("op_414_end_0"), val = tensor([1, 16, 256])]; - tensor var_414_end_mask_0 = const()[name = tensor("op_414_end_mask_0"), val = tensor([true, true, true])]; - tensor var_414 = slice_by_index(begin = var_414_begin_0, end = var_414_end_0, end_mask = var_414_end_mask_0, x = window_5)[name = tensor("op_414")]; + tensor var_440_begin_0 = const()[name = tensor("op_440_begin_0"), val = tensor([0, 1, 0])]; + tensor var_440_end_0 = const()[name = tensor("op_440_end_0"), val = tensor([1, 16, 256])]; + tensor var_440_end_mask_0 = const()[name = tensor("op_440_end_mask_0"), val = tensor([true, true, true])]; + tensor var_440 = slice_by_index(begin = var_440_begin_0, end = var_440_end_0, end_mask = var_440_end_mask_0, x = window_5)[name = tensor("op_440")]; tensor window_7_interleave_0 = const()[name = tensor("window_7_interleave_0"), val = tensor(false)]; - tensor window_7 = concat(axis = var_26, interleave = window_7_interleave_0, values = (var_414, x_9))[name = tensor("window_7")]; - tensor input_61_interleave_0 = const()[name = tensor("input_61_interleave_0"), val = tensor(false)]; - tensor input_61 = concat(axis = var_23, interleave = input_61_interleave_0, values = window_7)[name = tensor("input_61")]; + tensor window_7 = concat(axis = var_52, interleave = window_7_interleave_0, values = (var_440, x_9))[name = tensor("window_7")]; + tensor input_63_interleave_0 = const()[name = tensor("input_63_interleave_0"), val = tensor(false)]; + tensor input_63 = concat(axis = var_39, interleave = input_63_interleave_0, values = window_7)[name = tensor("input_63")]; tensor x_11_axes_0 = const()[name = tensor("x_11_axes_0"), val = tensor([-1])]; - tensor x_11 = layer_norm(axes = x_11_axes_0, beta = encoder_conv_module_1_sequential_0_bias, epsilon = var_28, gamma = encoder_conv_module_1_sequential_0_weight, x = input_61)[name = tensor("x_11")]; - tensor input_63_perm_0 = const()[name = tensor("input_63_perm_0"), val = tensor([0, 2, 1])]; + tensor x_11 = layer_norm(axes = x_11_axes_0, beta = inner_encoder_conv_module_1_sequential_0_bias, epsilon = var_36, gamma = inner_encoder_conv_module_1_sequential_0_weight, x = input_63)[name = tensor("x_11")]; + tensor input_65_perm_0 = const()[name = tensor("input_65_perm_0"), val = tensor([0, 2, 1])]; tensor inputs_13_pad_type_0 = const()[name = tensor("inputs_13_pad_type_0"), val = tensor("valid")]; tensor inputs_13_strides_0 = const()[name = tensor("inputs_13_strides_0"), val = tensor([1])]; tensor inputs_13_pad_0 = const()[name = tensor("inputs_13_pad_0"), val = tensor([0, 0])]; tensor inputs_13_dilations_0 = const()[name = tensor("inputs_13_dilations_0"), val = tensor([1])]; tensor inputs_13_groups_0 = const()[name = tensor("inputs_13_groups_0"), val = tensor(1)]; - tensor input_63 = transpose(perm = input_63_perm_0, x = x_11)[name = tensor("transpose_47")]; - tensor inputs_13 = conv(bias = encoder_conv_module_1_sequential_2_conv_bias, dilations = inputs_13_dilations_0, groups = inputs_13_groups_0, pad = inputs_13_pad_0, pad_type = inputs_13_pad_type_0, strides = inputs_13_strides_0, weight = encoder_conv_module_1_sequential_2_conv_weight, x = input_63)[name = tensor("inputs_13")]; - tensor var_439_split_sizes_0 = const()[name = tensor("op_439_split_sizes_0"), val = tensor([256, 256])]; - tensor var_439_axis_0 = const()[name = tensor("op_439_axis_0"), val = tensor(1)]; - tensor var_439_0, tensor var_439_1 = split(axis = var_439_axis_0, split_sizes = var_439_split_sizes_0, x = inputs_13)[name = tensor("op_439")]; - tensor var_441 = sigmoid(x = var_439_1)[name = tensor("op_441")]; - tensor inputs_15 = mul(x = var_439_0, y = var_441)[name = tensor("inputs_15")]; + tensor input_65 = transpose(perm = input_65_perm_0, x = x_11)[name = tensor("transpose_47")]; + tensor inputs_13 = conv(bias = inner_encoder_conv_module_1_sequential_2_conv_bias, dilations = inputs_13_dilations_0, groups = inputs_13_groups_0, pad = inputs_13_pad_0, pad_type = inputs_13_pad_type_0, strides = inputs_13_strides_0, weight = inner_encoder_conv_module_1_sequential_2_conv_weight, x = input_65)[name = tensor("inputs_13")]; + tensor var_465_split_sizes_0 = const()[name = tensor("op_465_split_sizes_0"), val = tensor([256, 256])]; + tensor var_465_axis_0 = const()[name = tensor("op_465_axis_0"), val = tensor(1)]; + tensor var_465_0, tensor var_465_1 = split(axis = var_465_axis_0, split_sizes = var_465_split_sizes_0, x = inputs_13)[name = tensor("op_465")]; + tensor var_467 = sigmoid(x = var_465_1)[name = tensor("op_467")]; + tensor inputs_15 = mul(x = var_465_0, y = var_467)[name = tensor("inputs_15")]; tensor outputs_aug_3_pad_type_0 = const()[name = tensor("outputs_aug_3_pad_type_0"), val = tensor("custom")]; tensor outputs_aug_3_pad_0 = const()[name = tensor("outputs_aug_3_pad_0"), val = tensor([15, 15])]; tensor outputs_aug_3_groups_0 = const()[name = tensor("outputs_aug_3_groups_0"), val = tensor(256)]; tensor outputs_aug_3_strides_0 = const()[name = tensor("outputs_aug_3_strides_0"), val = tensor([1])]; tensor outputs_aug_3_dilations_0 = const()[name = tensor("outputs_aug_3_dilations_0"), val = tensor([1])]; - tensor outputs_aug_3 = conv(dilations = outputs_aug_3_dilations_0, groups = outputs_aug_3_groups_0, pad = outputs_aug_3_pad_0, pad_type = outputs_aug_3_pad_type_0, strides = outputs_aug_3_strides_0, weight = encoder_conv_module_1_sequential_4_conv_weight, x = inputs_15)[name = tensor("outputs_aug_3")]; - tensor input_65_begin_0 = const()[name = tensor("input_65_begin_0"), val = tensor([0, 0, 0])]; - tensor input_65_end_0 = const()[name = tensor("input_65_end_0"), val = tensor([1, 256, 16])]; - tensor input_65_end_mask_0 = const()[name = tensor("input_65_end_mask_0"), val = tensor([true, true, false])]; - tensor input_65 = slice_by_index(begin = input_65_begin_0, end = input_65_end_0, end_mask = input_65_end_mask_0, x = outputs_aug_3)[name = tensor("input_65")]; - tensor inputs_17 = batch_norm(beta = encoder_conv_module_1_sequential_5_bias, epsilon = var_28, gamma = encoder_conv_module_1_sequential_5_weight, mean = encoder_conv_module_1_sequential_5_running_mean, variance = encoder_conv_module_1_sequential_5_running_var, x = input_65)[name = tensor("inputs_17")]; - tensor input_67 = silu(x = inputs_17)[name = tensor("input_67")]; - tensor input_69_pad_type_0 = const()[name = tensor("input_69_pad_type_0"), val = tensor("valid")]; - tensor input_69_strides_0 = const()[name = tensor("input_69_strides_0"), val = tensor([1])]; - tensor input_69_pad_0 = const()[name = tensor("input_69_pad_0"), val = tensor([0, 0])]; - tensor input_69_dilations_0 = const()[name = tensor("input_69_dilations_0"), val = tensor([1])]; - tensor input_69_groups_0 = const()[name = tensor("input_69_groups_0"), val = tensor(1)]; - tensor input_69 = conv(bias = encoder_conv_module_1_sequential_7_conv_bias, dilations = input_69_dilations_0, groups = input_69_groups_0, pad = input_69_pad_0, pad_type = input_69_pad_type_0, strides = input_69_strides_0, weight = encoder_conv_module_1_sequential_7_conv_weight, x = input_67)[name = tensor("input_69")]; + tensor outputs_aug_3 = conv(dilations = outputs_aug_3_dilations_0, groups = outputs_aug_3_groups_0, pad = outputs_aug_3_pad_0, pad_type = outputs_aug_3_pad_type_0, strides = outputs_aug_3_strides_0, weight = inner_encoder_conv_module_1_sequential_4_conv_weight, x = inputs_15)[name = tensor("outputs_aug_3")]; + tensor input_67_begin_0 = const()[name = tensor("input_67_begin_0"), val = tensor([0, 0, 0])]; + tensor input_67_end_0 = const()[name = tensor("input_67_end_0"), val = tensor([1, 256, 16])]; + tensor input_67_end_mask_0 = const()[name = tensor("input_67_end_mask_0"), val = tensor([true, true, false])]; + tensor input_67 = slice_by_index(begin = input_67_begin_0, end = input_67_end_0, end_mask = input_67_end_mask_0, x = outputs_aug_3)[name = tensor("input_67")]; + tensor inputs_17 = batch_norm(beta = inner_encoder_conv_module_1_sequential_5_bias, epsilon = var_36, gamma = inner_encoder_conv_module_1_sequential_5_weight, mean = inner_encoder_conv_module_1_sequential_5_running_mean, variance = inner_encoder_conv_module_1_sequential_5_running_var, x = input_67)[name = tensor("inputs_17")]; + tensor input_69 = silu(x = inputs_17)[name = tensor("input_69")]; + tensor input_71_pad_type_0 = const()[name = tensor("input_71_pad_type_0"), val = tensor("valid")]; + tensor input_71_strides_0 = const()[name = tensor("input_71_strides_0"), val = tensor([1])]; + tensor input_71_pad_0 = const()[name = tensor("input_71_pad_0"), val = tensor([0, 0])]; + tensor input_71_dilations_0 = const()[name = tensor("input_71_dilations_0"), val = tensor([1])]; + tensor input_71_groups_0 = const()[name = tensor("input_71_groups_0"), val = tensor(1)]; + tensor input_71 = conv(bias = inner_encoder_conv_module_1_sequential_7_conv_bias, dilations = input_71_dilations_0, groups = input_71_groups_0, pad = input_71_pad_0, pad_type = input_71_pad_type_0, strides = input_71_strides_0, weight = inner_encoder_conv_module_1_sequential_7_conv_weight, x = input_69)[name = tensor("input_71")]; tensor conv_out_3_perm_0 = const()[name = tensor("conv_out_3_perm_0"), val = tensor([0, 2, 1])]; - tensor var_472_begin_0 = const()[name = tensor("op_472_begin_0"), val = tensor([0, -1, 0])]; - tensor var_472_end_0 = const()[name = tensor("op_472_end_0"), val = tensor([1, 16, 256])]; - tensor var_472_end_mask_0 = const()[name = tensor("op_472_end_mask_0"), val = tensor([true, true, true])]; - tensor conv_out_3 = transpose(perm = conv_out_3_perm_0, x = input_69)[name = tensor("transpose_46")]; - tensor var_472 = slice_by_index(begin = var_472_begin_0, end = var_472_end_0, end_mask = var_472_end_mask_0, x = conv_out_3)[name = tensor("op_472")]; - tensor var_474_perm_0 = const()[name = tensor("op_474_perm_0"), val = tensor([1, 0, 2])]; - tensor var_474 = transpose(perm = var_474_perm_0, x = var_472)[name = tensor("transpose_45")]; - tensor input_71 = add(x = x_9, y = var_474)[name = tensor("input_71")]; - tensor input_73_axes_0 = const()[name = tensor("input_73_axes_0"), val = tensor([-1])]; - tensor input_73 = layer_norm(axes = input_73_axes_0, beta = encoder_ffn2_1_module_sequential_0_bias, epsilon = var_28, gamma = encoder_ffn2_1_module_sequential_0_weight, x = input_71)[name = tensor("input_73")]; - tensor inputs_19 = linear(bias = encoder_ffn2_1_module_sequential_1_linear_bias, weight = encoder_ffn2_1_module_sequential_1_linear_weight, x = input_73)[name = tensor("linear_17")]; - tensor input_75 = silu(x = inputs_19)[name = tensor("input_75")]; - tensor input_79 = linear(bias = encoder_ffn2_1_module_sequential_4_linear_bias, weight = encoder_ffn2_1_module_sequential_4_linear_weight, x = input_75)[name = tensor("linear_18")]; - tensor var_497 = const()[name = tensor("op_497"), val = tensor(0x1p-1)]; - tensor var_498 = mul(x = input_79, y = var_497)[name = tensor("op_498")]; - tensor input_81 = add(x = var_498, y = input_71)[name = tensor("input_81")]; - tensor input_83_axes_0 = const()[name = tensor("input_83_axes_0"), val = tensor([-1])]; - tensor input_83 = layer_norm(axes = input_83_axes_0, beta = encoder_layer_norm_1_bias, epsilon = var_28, gamma = encoder_layer_norm_1_weight, x = input_81)[name = tensor("input_83")]; + tensor var_498_begin_0 = const()[name = tensor("op_498_begin_0"), val = tensor([0, -1, 0])]; + tensor var_498_end_0 = const()[name = tensor("op_498_end_0"), val = tensor([1, 16, 256])]; + tensor var_498_end_mask_0 = const()[name = tensor("op_498_end_mask_0"), val = tensor([true, true, true])]; + tensor conv_out_3 = transpose(perm = conv_out_3_perm_0, x = input_71)[name = tensor("transpose_46")]; + tensor var_498 = slice_by_index(begin = var_498_begin_0, end = var_498_end_0, end_mask = var_498_end_mask_0, x = conv_out_3)[name = tensor("op_498")]; + tensor var_500_perm_0 = const()[name = tensor("op_500_perm_0"), val = tensor([1, 0, 2])]; + tensor var_500 = transpose(perm = var_500_perm_0, x = var_498)[name = tensor("transpose_45")]; + tensor input_73 = add(x = x_9, y = var_500)[name = tensor("input_73")]; + tensor input_75_axes_0 = const()[name = tensor("input_75_axes_0"), val = tensor([-1])]; + tensor input_75 = layer_norm(axes = input_75_axes_0, beta = inner_encoder_ffn2_1_module_sequential_0_bias, epsilon = var_36, gamma = inner_encoder_ffn2_1_module_sequential_0_weight, x = input_73)[name = tensor("input_75")]; + tensor inputs_19 = linear(bias = inner_encoder_ffn2_1_module_sequential_1_linear_bias, weight = inner_encoder_ffn2_1_module_sequential_1_linear_weight, x = input_75)[name = tensor("linear_17")]; + tensor input_77 = silu(x = inputs_19)[name = tensor("input_77")]; + tensor input_81 = linear(bias = inner_encoder_ffn2_1_module_sequential_4_linear_bias, weight = inner_encoder_ffn2_1_module_sequential_4_linear_weight, x = input_77)[name = tensor("linear_18")]; + tensor var_523 = const()[name = tensor("op_523"), val = tensor(0x1p-1)]; + tensor var_524 = mul(x = input_81, y = var_523)[name = tensor("op_524")]; + tensor input_83 = add(x = var_524, y = input_73)[name = tensor("input_83")]; tensor input_85_axes_0 = const()[name = tensor("input_85_axes_0"), val = tensor([-1])]; - tensor input_85 = layer_norm(axes = input_85_axes_0, beta = encoder_ffn1_2_module_sequential_0_bias, epsilon = var_28, gamma = encoder_ffn1_2_module_sequential_0_weight, x = input_83)[name = tensor("input_85")]; - tensor inputs_21 = linear(bias = encoder_ffn1_2_module_sequential_1_linear_bias, weight = encoder_ffn1_2_module_sequential_1_linear_weight, x = input_85)[name = tensor("linear_19")]; - tensor input_87 = silu(x = inputs_21)[name = tensor("input_87")]; - tensor input_91 = linear(bias = encoder_ffn1_2_module_sequential_4_linear_bias, weight = encoder_ffn1_2_module_sequential_4_linear_weight, x = input_87)[name = tensor("linear_20")]; - tensor var_527 = const()[name = tensor("op_527"), val = tensor(0x1p-1)]; - tensor var_528 = mul(x = input_91, y = var_527)[name = tensor("op_528")]; - tensor input_93 = add(x = var_528, y = input_83)[name = tensor("input_93")]; + tensor input_85 = layer_norm(axes = input_85_axes_0, beta = inner_encoder_layer_norm_1_bias, epsilon = var_36, gamma = inner_encoder_layer_norm_1_weight, x = input_83)[name = tensor("input_85")]; + tensor input_87_axes_0 = const()[name = tensor("input_87_axes_0"), val = tensor([-1])]; + tensor input_87 = layer_norm(axes = input_87_axes_0, beta = inner_encoder_ffn1_2_module_sequential_0_bias, epsilon = var_36, gamma = inner_encoder_ffn1_2_module_sequential_0_weight, x = input_85)[name = tensor("input_87")]; + tensor inputs_21 = linear(bias = inner_encoder_ffn1_2_module_sequential_1_linear_bias, weight = inner_encoder_ffn1_2_module_sequential_1_linear_weight, x = input_87)[name = tensor("linear_19")]; + tensor input_89 = silu(x = inputs_21)[name = tensor("input_89")]; + tensor input_93 = linear(bias = inner_encoder_ffn1_2_module_sequential_4_linear_bias, weight = inner_encoder_ffn1_2_module_sequential_4_linear_weight, x = input_89)[name = tensor("linear_20")]; + tensor var_553 = const()[name = tensor("op_553"), val = tensor(0x1p-1)]; + tensor var_554 = mul(x = input_93, y = var_553)[name = tensor("op_554")]; + tensor input_95 = add(x = var_554, y = input_85)[name = tensor("input_95")]; tensor x_13_axes_0 = const()[name = tensor("x_13_axes_0"), val = tensor([-1])]; - tensor x_13 = layer_norm(axes = x_13_axes_0, beta = encoder_ret_lns_2_bias, epsilon = var_28, gamma = encoder_ret_lns_2_weight, x = input_93)[name = tensor("x_13")]; + tensor x_13 = layer_norm(axes = x_13_axes_0, beta = inner_encoder_ret_lns_2_bias, epsilon = var_36, gamma = inner_encoder_ret_lns_2_weight, x = input_95)[name = tensor("x_13")]; tensor prev_kv_5_begin_0 = const()[name = tensor("prev_kv_5_begin_0"), val = tensor([2, 0, 0, 0, 0])]; tensor prev_kv_5_end_0 = const()[name = tensor("prev_kv_5_end_0"), val = tensor([3, 1, 4, 64, 64])]; tensor prev_kv_5_end_mask_0 = const()[name = tensor("prev_kv_5_end_mask_0"), val = tensor([false, true, true, true, true])]; @@ -524,139 +530,139 @@ program(1.0) tensor prev_scale_5_end_mask_0 = const()[name = tensor("prev_scale_5_end_mask_0"), val = tensor([false, true])]; tensor prev_scale_5_squeeze_mask_0 = const()[name = tensor("prev_scale_5_squeeze_mask_0"), val = tensor([true, false])]; tensor prev_scale_5 = slice_by_index(begin = prev_scale_5_begin_0, end = prev_scale_5_end_0, end_mask = prev_scale_5_end_mask_0, squeeze_mask = prev_scale_5_squeeze_mask_0, x = enc_scale)[name = tensor("prev_scale_5")]; - tensor var_542 = linear(bias = encoder_q_proj_2_bias, weight = encoder_q_proj_2_weight, x = x_13)[name = tensor("linear_21")]; - tensor var_543 = const()[name = tensor("op_543"), val = tensor([1, 1, 4, 64])]; - tensor var_544 = reshape(shape = var_543, x = var_542)[name = tensor("op_544")]; + tensor var_568 = linear(bias = inner_encoder_q_proj_2_bias, weight = inner_encoder_q_proj_2_weight, x = x_13)[name = tensor("linear_21")]; + tensor var_569 = const()[name = tensor("op_569"), val = tensor([1, 1, 4, 64])]; + tensor var_570 = reshape(shape = var_569, x = var_568)[name = tensor("op_570")]; tensor q_5_perm_0 = const()[name = tensor("q_5_perm_0"), val = tensor([0, 2, 1, 3])]; - tensor var_548 = linear(bias = encoder_k_proj_2_bias, weight = encoder_k_proj_2_weight, x = x_13)[name = tensor("linear_22")]; - tensor var_549 = const()[name = tensor("op_549"), val = tensor(0x1p-3)]; - tensor var_550 = mul(x = var_548, y = var_549)[name = tensor("op_550")]; - tensor var_551 = const()[name = tensor("op_551"), val = tensor([1, 1, 4, 64])]; - tensor var_552 = reshape(shape = var_551, x = var_550)[name = tensor("op_552")]; + tensor var_574 = linear(bias = inner_encoder_k_proj_2_bias, weight = inner_encoder_k_proj_2_weight, x = x_13)[name = tensor("linear_22")]; + tensor var_575 = const()[name = tensor("op_575"), val = tensor(0x1p-3)]; + tensor var_576 = mul(x = var_574, y = var_575)[name = tensor("op_576")]; + tensor var_577 = const()[name = tensor("op_577"), val = tensor([1, 1, 4, 64])]; + tensor var_578 = reshape(shape = var_577, x = var_576)[name = tensor("op_578")]; tensor k_5_perm_0 = const()[name = tensor("k_5_perm_0"), val = tensor([0, 2, 1, 3])]; - tensor var_556 = linear(bias = encoder_v_proj_2_bias, weight = encoder_v_proj_2_weight, x = x_13)[name = tensor("linear_23")]; - tensor var_557 = const()[name = tensor("op_557"), val = tensor([1, 1, 4, 64])]; - tensor var_558 = reshape(shape = var_557, x = var_556)[name = tensor("op_558")]; + tensor var_582 = linear(bias = inner_encoder_v_proj_2_bias, weight = inner_encoder_v_proj_2_weight, x = x_13)[name = tensor("linear_23")]; + tensor var_583 = const()[name = tensor("op_583"), val = tensor([1, 1, 4, 64])]; + tensor var_584 = reshape(shape = var_583, x = var_582)[name = tensor("op_584")]; tensor v_5_perm_0 = const()[name = tensor("v_5_perm_0"), val = tensor([0, 2, 1, 3])]; - tensor input_97 = linear(bias = encoder_g_proj_2_bias, weight = encoder_g_proj_2_weight, x = x_13)[name = tensor("linear_24")]; + tensor input_99 = linear(bias = inner_encoder_g_proj_2_bias, weight = inner_encoder_g_proj_2_weight, x = x_13)[name = tensor("linear_24")]; tensor sqrt_s0_5 = sqrt(x = prev_scale_5)[name = tensor("sqrt_s0_5")]; - tensor s_t_5 = add(x = prev_scale_5, y = encoder__t_index)[name = tensor("s_t_5")]; + tensor s_t_5 = add(x = prev_scale_5, y = inner_encoder__t_index)[name = tensor("s_t_5")]; tensor sqrt_s_t_5 = sqrt(x = s_t_5)[name = tensor("sqrt_s_t_5")]; tensor qk_5_transpose_x_1 = const()[name = tensor("qk_5_transpose_x_1"), val = tensor(false)]; tensor qk_5_transpose_y_1 = const()[name = tensor("qk_5_transpose_y_1"), val = tensor(true)]; - tensor k_5 = transpose(perm = k_5_perm_0, x = var_552)[name = tensor("transpose_43")]; - tensor q_5 = transpose(perm = q_5_perm_0, x = var_544)[name = tensor("transpose_44")]; + tensor k_5 = transpose(perm = k_5_perm_0, x = var_578)[name = tensor("transpose_43")]; + tensor q_5 = transpose(perm = q_5_perm_0, x = var_570)[name = tensor("transpose_44")]; tensor qk_5 = matmul(transpose_x = qk_5_transpose_x_1, transpose_y = qk_5_transpose_y_1, x = q_5, y = k_5)[name = tensor("qk_5")]; - tensor var_568 = const()[name = tensor("op_568"), val = tensor([1, 1])]; - tensor var_569 = reshape(shape = var_568, x = sqrt_s_t_5)[name = tensor("op_569")]; - tensor M_5 = real_div(x = encoder__causal_mask, y = var_569)[name = tensor("M_5")]; - tensor var_571 = mul(x = qk_5, y = M_5)[name = tensor("op_571")]; + tensor var_594 = const()[name = tensor("op_594"), val = tensor([1, 1])]; + tensor var_595 = reshape(shape = var_594, x = sqrt_s_t_5)[name = tensor("op_595")]; + tensor M_5 = real_div(x = inner_encoder__causal_mask, y = var_595)[name = tensor("M_5")]; + tensor var_597 = mul(x = qk_5, y = M_5)[name = tensor("op_597")]; tensor inner_5_transpose_x_0 = const()[name = tensor("inner_5_transpose_x_0"), val = tensor(false)]; tensor inner_5_transpose_y_0 = const()[name = tensor("inner_5_transpose_y_0"), val = tensor(false)]; - tensor v_5 = transpose(perm = v_5_perm_0, x = var_558)[name = tensor("transpose_42")]; - tensor inner_5 = matmul(transpose_x = inner_5_transpose_x_0, transpose_y = inner_5_transpose_y_0, x = var_571, y = v_5)[name = tensor("inner_5")]; - tensor var_573_transpose_x_0 = const()[name = tensor("op_573_transpose_x_0"), val = tensor(false)]; - tensor var_573_transpose_y_0 = const()[name = tensor("op_573_transpose_y_0"), val = tensor(false)]; - tensor var_573 = matmul(transpose_x = var_573_transpose_x_0, transpose_y = var_573_transpose_y_0, x = q_5, y = prev_kv_5)[name = tensor("op_573")]; - tensor var_574 = real_div(x = sqrt_s0_5, y = sqrt_s_t_5)[name = tensor("op_574")]; - tensor var_575 = const()[name = tensor("op_575"), val = tensor([1, 1, 1, 1])]; - tensor var_576 = reshape(shape = var_575, x = var_574)[name = tensor("op_576")]; - tensor cross_5 = mul(x = var_573, y = var_576)[name = tensor("cross_5")]; + tensor v_5 = transpose(perm = v_5_perm_0, x = var_584)[name = tensor("transpose_42")]; + tensor inner_5 = matmul(transpose_x = inner_5_transpose_x_0, transpose_y = inner_5_transpose_y_0, x = var_597, y = v_5)[name = tensor("inner_5")]; + tensor var_599_transpose_x_0 = const()[name = tensor("op_599_transpose_x_0"), val = tensor(false)]; + tensor var_599_transpose_y_0 = const()[name = tensor("op_599_transpose_y_0"), val = tensor(false)]; + tensor var_599 = matmul(transpose_x = var_599_transpose_x_0, transpose_y = var_599_transpose_y_0, x = q_5, y = prev_kv_5)[name = tensor("op_599")]; + tensor var_600 = real_div(x = sqrt_s0_5, y = sqrt_s_t_5)[name = tensor("op_600")]; + tensor var_601 = const()[name = tensor("op_601"), val = tensor([1, 1, 1, 1])]; + tensor var_602 = reshape(shape = var_601, x = var_600)[name = tensor("op_602")]; + tensor cross_5 = mul(x = var_599, y = var_602)[name = tensor("cross_5")]; tensor out_13 = add(x = inner_5, y = cross_5)[name = tensor("out_13")]; - tensor var_579 = mul(x = prev_kv_5, y = sqrt_s0_5)[name = tensor("op_579")]; - tensor var_581_transpose_x_1 = const()[name = tensor("op_581_transpose_x_1"), val = tensor(true)]; - tensor var_581_transpose_y_1 = const()[name = tensor("op_581_transpose_y_1"), val = tensor(false)]; - tensor var_581 = matmul(transpose_x = var_581_transpose_x_1, transpose_y = var_581_transpose_y_1, x = k_5, y = v_5)[name = tensor("op_581")]; - tensor new_kv_unnorm_5 = add(x = var_579, y = var_581)[name = tensor("new_kv_unnorm_5")]; - tensor var_583 = const()[name = tensor("op_583"), val = tensor(0x1p+0)]; - tensor new_scale_5 = add(x = prev_scale_5, y = var_583)[name = tensor("new_scale_5")]; - tensor var_585 = sqrt(x = new_scale_5)[name = tensor("op_585")]; - tensor var_586 = real_div(x = new_kv_unnorm_5, y = var_585)[name = tensor("op_586")]; - tensor var_587_perm_0 = const()[name = tensor("op_587_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_605 = mul(x = prev_kv_5, y = sqrt_s0_5)[name = tensor("op_605")]; + tensor var_607_transpose_x_1 = const()[name = tensor("op_607_transpose_x_1"), val = tensor(true)]; + tensor var_607_transpose_y_1 = const()[name = tensor("op_607_transpose_y_1"), val = tensor(false)]; + tensor var_607 = matmul(transpose_x = var_607_transpose_x_1, transpose_y = var_607_transpose_y_1, x = k_5, y = v_5)[name = tensor("op_607")]; + tensor new_kv_unnorm_5 = add(x = var_605, y = var_607)[name = tensor("new_kv_unnorm_5")]; + tensor var_609 = const()[name = tensor("op_609"), val = tensor(0x1p+0)]; + tensor new_scale_5 = add(x = prev_scale_5, y = var_609)[name = tensor("new_scale_5")]; + tensor var_611 = sqrt(x = new_scale_5)[name = tensor("op_611")]; + tensor var_612 = real_div(x = new_kv_unnorm_5, y = var_611)[name = tensor("op_612")]; + tensor var_613_perm_0 = const()[name = tensor("op_613_perm_0"), val = tensor([0, 2, 1, 3])]; tensor out_15_axes_0 = const()[name = tensor("out_15_axes_0"), val = tensor([-1])]; - tensor var_587 = transpose(perm = var_587_perm_0, x = out_13)[name = tensor("transpose_41")]; - tensor out_15 = layer_norm(axes = out_15_axes_0, epsilon = var_17, x = var_587)[name = tensor("out_15")]; - tensor var_591 = const()[name = tensor("op_591"), val = tensor([1, 1, 256])]; - tensor out_17 = reshape(shape = var_591, x = out_15)[name = tensor("out_17")]; - tensor var_593 = silu(x = input_97)[name = tensor("op_593")]; - tensor input_99 = mul(x = var_593, y = out_17)[name = tensor("input_99")]; - tensor ret_out_5 = linear(bias = encoder_out_proj_2_bias, weight = encoder_out_proj_2_weight, x = input_99)[name = tensor("linear_25")]; - tensor x_15 = add(x = input_93, y = ret_out_5)[name = tensor("x_15")]; + tensor var_613 = transpose(perm = var_613_perm_0, x = out_13)[name = tensor("transpose_41")]; + tensor out_15 = layer_norm(axes = out_15_axes_0, epsilon = var_44, x = var_613)[name = tensor("out_15")]; + tensor var_617 = const()[name = tensor("op_617"), val = tensor([1, 1, 256])]; + tensor out_17 = reshape(shape = var_617, x = out_15)[name = tensor("out_17")]; + tensor var_619 = silu(x = input_99)[name = tensor("op_619")]; + tensor input_101 = mul(x = var_619, y = out_17)[name = tensor("input_101")]; + tensor ret_out_5 = linear(bias = inner_encoder_out_proj_2_bias, weight = inner_encoder_out_proj_2_weight, x = input_101)[name = tensor("linear_25")]; + tensor x_15 = add(x = input_95, y = ret_out_5)[name = tensor("x_15")]; tensor window_9_begin_0 = const()[name = tensor("window_9_begin_0"), val = tensor([2, 0, 0, 0])]; tensor window_9_end_0 = const()[name = tensor("window_9_end_0"), val = tensor([3, 1, 16, 256])]; tensor window_9_end_mask_0 = const()[name = tensor("window_9_end_mask_0"), val = tensor([false, true, true, true])]; tensor window_9_squeeze_mask_0 = const()[name = tensor("window_9_squeeze_mask_0"), val = tensor([true, false, false, false])]; tensor window_9 = slice_by_index(begin = window_9_begin_0, end = window_9_end_0, end_mask = window_9_end_mask_0, squeeze_mask = window_9_squeeze_mask_0, x = enc_conv_cache)[name = tensor("window_9")]; - tensor var_604_begin_0 = const()[name = tensor("op_604_begin_0"), val = tensor([0, 1, 0])]; - tensor var_604_end_0 = const()[name = tensor("op_604_end_0"), val = tensor([1, 16, 256])]; - tensor var_604_end_mask_0 = const()[name = tensor("op_604_end_mask_0"), val = tensor([true, true, true])]; - tensor var_604 = slice_by_index(begin = var_604_begin_0, end = var_604_end_0, end_mask = var_604_end_mask_0, x = window_9)[name = tensor("op_604")]; + tensor var_630_begin_0 = const()[name = tensor("op_630_begin_0"), val = tensor([0, 1, 0])]; + tensor var_630_end_0 = const()[name = tensor("op_630_end_0"), val = tensor([1, 16, 256])]; + tensor var_630_end_mask_0 = const()[name = tensor("op_630_end_mask_0"), val = tensor([true, true, true])]; + tensor var_630 = slice_by_index(begin = var_630_begin_0, end = var_630_end_0, end_mask = var_630_end_mask_0, x = window_9)[name = tensor("op_630")]; tensor window_11_interleave_0 = const()[name = tensor("window_11_interleave_0"), val = tensor(false)]; - tensor window_11 = concat(axis = var_26, interleave = window_11_interleave_0, values = (var_604, x_15))[name = tensor("window_11")]; - tensor input_101_interleave_0 = const()[name = tensor("input_101_interleave_0"), val = tensor(false)]; - tensor input_101 = concat(axis = var_23, interleave = input_101_interleave_0, values = window_11)[name = tensor("input_101")]; + tensor window_11 = concat(axis = var_52, interleave = window_11_interleave_0, values = (var_630, x_15))[name = tensor("window_11")]; + tensor input_103_interleave_0 = const()[name = tensor("input_103_interleave_0"), val = tensor(false)]; + tensor input_103 = concat(axis = var_39, interleave = input_103_interleave_0, values = window_11)[name = tensor("input_103")]; tensor x_17_axes_0 = const()[name = tensor("x_17_axes_0"), val = tensor([-1])]; - tensor x_17 = layer_norm(axes = x_17_axes_0, beta = encoder_conv_module_2_sequential_0_bias, epsilon = var_28, gamma = encoder_conv_module_2_sequential_0_weight, x = input_101)[name = tensor("x_17")]; - tensor input_103_perm_0 = const()[name = tensor("input_103_perm_0"), val = tensor([0, 2, 1])]; + tensor x_17 = layer_norm(axes = x_17_axes_0, beta = inner_encoder_conv_module_2_sequential_0_bias, epsilon = var_36, gamma = inner_encoder_conv_module_2_sequential_0_weight, x = input_103)[name = tensor("x_17")]; + tensor input_105_perm_0 = const()[name = tensor("input_105_perm_0"), val = tensor([0, 2, 1])]; tensor inputs_23_pad_type_0 = const()[name = tensor("inputs_23_pad_type_0"), val = tensor("valid")]; tensor inputs_23_strides_0 = const()[name = tensor("inputs_23_strides_0"), val = tensor([1])]; tensor inputs_23_pad_0 = const()[name = tensor("inputs_23_pad_0"), val = tensor([0, 0])]; tensor inputs_23_dilations_0 = const()[name = tensor("inputs_23_dilations_0"), val = tensor([1])]; tensor inputs_23_groups_0 = const()[name = tensor("inputs_23_groups_0"), val = tensor(1)]; - tensor input_103 = transpose(perm = input_103_perm_0, x = x_17)[name = tensor("transpose_40")]; - tensor inputs_23 = conv(bias = encoder_conv_module_2_sequential_2_conv_bias, dilations = inputs_23_dilations_0, groups = inputs_23_groups_0, pad = inputs_23_pad_0, pad_type = inputs_23_pad_type_0, strides = inputs_23_strides_0, weight = encoder_conv_module_2_sequential_2_conv_weight, x = input_103)[name = tensor("inputs_23")]; - tensor var_629_split_sizes_0 = const()[name = tensor("op_629_split_sizes_0"), val = tensor([256, 256])]; - tensor var_629_axis_0 = const()[name = tensor("op_629_axis_0"), val = tensor(1)]; - tensor var_629_0, tensor var_629_1 = split(axis = var_629_axis_0, split_sizes = var_629_split_sizes_0, x = inputs_23)[name = tensor("op_629")]; - tensor var_631 = sigmoid(x = var_629_1)[name = tensor("op_631")]; - tensor inputs_25 = mul(x = var_629_0, y = var_631)[name = tensor("inputs_25")]; + tensor input_105 = transpose(perm = input_105_perm_0, x = x_17)[name = tensor("transpose_40")]; + tensor inputs_23 = conv(bias = inner_encoder_conv_module_2_sequential_2_conv_bias, dilations = inputs_23_dilations_0, groups = inputs_23_groups_0, pad = inputs_23_pad_0, pad_type = inputs_23_pad_type_0, strides = inputs_23_strides_0, weight = inner_encoder_conv_module_2_sequential_2_conv_weight, x = input_105)[name = tensor("inputs_23")]; + tensor var_655_split_sizes_0 = const()[name = tensor("op_655_split_sizes_0"), val = tensor([256, 256])]; + tensor var_655_axis_0 = const()[name = tensor("op_655_axis_0"), val = tensor(1)]; + tensor var_655_0, tensor var_655_1 = split(axis = var_655_axis_0, split_sizes = var_655_split_sizes_0, x = inputs_23)[name = tensor("op_655")]; + tensor var_657 = sigmoid(x = var_655_1)[name = tensor("op_657")]; + tensor inputs_25 = mul(x = var_655_0, y = var_657)[name = tensor("inputs_25")]; tensor outputs_aug_5_pad_type_0 = const()[name = tensor("outputs_aug_5_pad_type_0"), val = tensor("custom")]; tensor outputs_aug_5_pad_0 = const()[name = tensor("outputs_aug_5_pad_0"), val = tensor([15, 15])]; tensor outputs_aug_5_groups_0 = const()[name = tensor("outputs_aug_5_groups_0"), val = tensor(256)]; tensor outputs_aug_5_strides_0 = const()[name = tensor("outputs_aug_5_strides_0"), val = tensor([1])]; tensor outputs_aug_5_dilations_0 = const()[name = tensor("outputs_aug_5_dilations_0"), val = tensor([1])]; - tensor outputs_aug_5 = conv(dilations = outputs_aug_5_dilations_0, groups = outputs_aug_5_groups_0, pad = outputs_aug_5_pad_0, pad_type = outputs_aug_5_pad_type_0, strides = outputs_aug_5_strides_0, weight = encoder_conv_module_2_sequential_4_conv_weight, x = inputs_25)[name = tensor("outputs_aug_5")]; - tensor input_105_begin_0 = const()[name = tensor("input_105_begin_0"), val = tensor([0, 0, 0])]; - tensor input_105_end_0 = const()[name = tensor("input_105_end_0"), val = tensor([1, 256, 16])]; - tensor input_105_end_mask_0 = const()[name = tensor("input_105_end_mask_0"), val = tensor([true, true, false])]; - tensor input_105 = slice_by_index(begin = input_105_begin_0, end = input_105_end_0, end_mask = input_105_end_mask_0, x = outputs_aug_5)[name = tensor("input_105")]; - tensor inputs_27 = batch_norm(beta = encoder_conv_module_2_sequential_5_bias, epsilon = var_28, gamma = encoder_conv_module_2_sequential_5_weight, mean = encoder_conv_module_2_sequential_5_running_mean, variance = encoder_conv_module_2_sequential_5_running_var, x = input_105)[name = tensor("inputs_27")]; - tensor input_107 = silu(x = inputs_27)[name = tensor("input_107")]; - tensor input_109_pad_type_0 = const()[name = tensor("input_109_pad_type_0"), val = tensor("valid")]; - tensor input_109_strides_0 = const()[name = tensor("input_109_strides_0"), val = tensor([1])]; - tensor input_109_pad_0 = const()[name = tensor("input_109_pad_0"), val = tensor([0, 0])]; - tensor input_109_dilations_0 = const()[name = tensor("input_109_dilations_0"), val = tensor([1])]; - tensor input_109_groups_0 = const()[name = tensor("input_109_groups_0"), val = tensor(1)]; - tensor input_109 = conv(bias = encoder_conv_module_2_sequential_7_conv_bias, dilations = input_109_dilations_0, groups = input_109_groups_0, pad = input_109_pad_0, pad_type = input_109_pad_type_0, strides = input_109_strides_0, weight = encoder_conv_module_2_sequential_7_conv_weight, x = input_107)[name = tensor("input_109")]; + tensor outputs_aug_5 = conv(dilations = outputs_aug_5_dilations_0, groups = outputs_aug_5_groups_0, pad = outputs_aug_5_pad_0, pad_type = outputs_aug_5_pad_type_0, strides = outputs_aug_5_strides_0, weight = inner_encoder_conv_module_2_sequential_4_conv_weight, x = inputs_25)[name = tensor("outputs_aug_5")]; + tensor input_107_begin_0 = const()[name = tensor("input_107_begin_0"), val = tensor([0, 0, 0])]; + tensor input_107_end_0 = const()[name = tensor("input_107_end_0"), val = tensor([1, 256, 16])]; + tensor input_107_end_mask_0 = const()[name = tensor("input_107_end_mask_0"), val = tensor([true, true, false])]; + tensor input_107 = slice_by_index(begin = input_107_begin_0, end = input_107_end_0, end_mask = input_107_end_mask_0, x = outputs_aug_5)[name = tensor("input_107")]; + tensor inputs_27 = batch_norm(beta = inner_encoder_conv_module_2_sequential_5_bias, epsilon = var_36, gamma = inner_encoder_conv_module_2_sequential_5_weight, mean = inner_encoder_conv_module_2_sequential_5_running_mean, variance = inner_encoder_conv_module_2_sequential_5_running_var, x = input_107)[name = tensor("inputs_27")]; + tensor input_109 = silu(x = inputs_27)[name = tensor("input_109")]; + tensor input_111_pad_type_0 = const()[name = tensor("input_111_pad_type_0"), val = tensor("valid")]; + tensor input_111_strides_0 = const()[name = tensor("input_111_strides_0"), val = tensor([1])]; + tensor input_111_pad_0 = const()[name = tensor("input_111_pad_0"), val = tensor([0, 0])]; + tensor input_111_dilations_0 = const()[name = tensor("input_111_dilations_0"), val = tensor([1])]; + tensor input_111_groups_0 = const()[name = tensor("input_111_groups_0"), val = tensor(1)]; + tensor input_111 = conv(bias = inner_encoder_conv_module_2_sequential_7_conv_bias, dilations = input_111_dilations_0, groups = input_111_groups_0, pad = input_111_pad_0, pad_type = input_111_pad_type_0, strides = input_111_strides_0, weight = inner_encoder_conv_module_2_sequential_7_conv_weight, x = input_109)[name = tensor("input_111")]; tensor conv_out_5_perm_0 = const()[name = tensor("conv_out_5_perm_0"), val = tensor([0, 2, 1])]; - tensor var_662_begin_0 = const()[name = tensor("op_662_begin_0"), val = tensor([0, -1, 0])]; - tensor var_662_end_0 = const()[name = tensor("op_662_end_0"), val = tensor([1, 16, 256])]; - tensor var_662_end_mask_0 = const()[name = tensor("op_662_end_mask_0"), val = tensor([true, true, true])]; - tensor conv_out_5 = transpose(perm = conv_out_5_perm_0, x = input_109)[name = tensor("transpose_39")]; - tensor var_662 = slice_by_index(begin = var_662_begin_0, end = var_662_end_0, end_mask = var_662_end_mask_0, x = conv_out_5)[name = tensor("op_662")]; - tensor var_664_perm_0 = const()[name = tensor("op_664_perm_0"), val = tensor([1, 0, 2])]; - tensor var_664 = transpose(perm = var_664_perm_0, x = var_662)[name = tensor("transpose_38")]; - tensor input_111 = add(x = x_15, y = var_664)[name = tensor("input_111")]; - tensor input_113_axes_0 = const()[name = tensor("input_113_axes_0"), val = tensor([-1])]; - tensor input_113 = layer_norm(axes = input_113_axes_0, beta = encoder_ffn2_2_module_sequential_0_bias, epsilon = var_28, gamma = encoder_ffn2_2_module_sequential_0_weight, x = input_111)[name = tensor("input_113")]; - tensor inputs_29 = linear(bias = encoder_ffn2_2_module_sequential_1_linear_bias, weight = encoder_ffn2_2_module_sequential_1_linear_weight, x = input_113)[name = tensor("linear_26")]; - tensor input_115 = silu(x = inputs_29)[name = tensor("input_115")]; - tensor input_119 = linear(bias = encoder_ffn2_2_module_sequential_4_linear_bias, weight = encoder_ffn2_2_module_sequential_4_linear_weight, x = input_115)[name = tensor("linear_27")]; - tensor var_687 = const()[name = tensor("op_687"), val = tensor(0x1p-1)]; - tensor var_688 = mul(x = input_119, y = var_687)[name = tensor("op_688")]; - tensor input_121 = add(x = var_688, y = input_111)[name = tensor("input_121")]; - tensor input_123_axes_0 = const()[name = tensor("input_123_axes_0"), val = tensor([-1])]; - tensor input_123 = layer_norm(axes = input_123_axes_0, beta = encoder_layer_norm_2_bias, epsilon = var_28, gamma = encoder_layer_norm_2_weight, x = input_121)[name = tensor("input_123")]; + tensor var_688_begin_0 = const()[name = tensor("op_688_begin_0"), val = tensor([0, -1, 0])]; + tensor var_688_end_0 = const()[name = tensor("op_688_end_0"), val = tensor([1, 16, 256])]; + tensor var_688_end_mask_0 = const()[name = tensor("op_688_end_mask_0"), val = tensor([true, true, true])]; + tensor conv_out_5 = transpose(perm = conv_out_5_perm_0, x = input_111)[name = tensor("transpose_39")]; + tensor var_688 = slice_by_index(begin = var_688_begin_0, end = var_688_end_0, end_mask = var_688_end_mask_0, x = conv_out_5)[name = tensor("op_688")]; + tensor var_690_perm_0 = const()[name = tensor("op_690_perm_0"), val = tensor([1, 0, 2])]; + tensor var_690 = transpose(perm = var_690_perm_0, x = var_688)[name = tensor("transpose_38")]; + tensor input_113 = add(x = x_15, y = var_690)[name = tensor("input_113")]; + tensor input_115_axes_0 = const()[name = tensor("input_115_axes_0"), val = tensor([-1])]; + tensor input_115 = layer_norm(axes = input_115_axes_0, beta = inner_encoder_ffn2_2_module_sequential_0_bias, epsilon = var_36, gamma = inner_encoder_ffn2_2_module_sequential_0_weight, x = input_113)[name = tensor("input_115")]; + tensor inputs_29 = linear(bias = inner_encoder_ffn2_2_module_sequential_1_linear_bias, weight = inner_encoder_ffn2_2_module_sequential_1_linear_weight, x = input_115)[name = tensor("linear_26")]; + tensor input_117 = silu(x = inputs_29)[name = tensor("input_117")]; + tensor input_121 = linear(bias = inner_encoder_ffn2_2_module_sequential_4_linear_bias, weight = inner_encoder_ffn2_2_module_sequential_4_linear_weight, x = input_117)[name = tensor("linear_27")]; + tensor var_713 = const()[name = tensor("op_713"), val = tensor(0x1p-1)]; + tensor var_714 = mul(x = input_121, y = var_713)[name = tensor("op_714")]; + tensor input_123 = add(x = var_714, y = input_113)[name = tensor("input_123")]; tensor input_125_axes_0 = const()[name = tensor("input_125_axes_0"), val = tensor([-1])]; - tensor input_125 = layer_norm(axes = input_125_axes_0, beta = encoder_ffn1_3_module_sequential_0_bias, epsilon = var_28, gamma = encoder_ffn1_3_module_sequential_0_weight, x = input_123)[name = tensor("input_125")]; - tensor inputs_31 = linear(bias = encoder_ffn1_3_module_sequential_1_linear_bias, weight = encoder_ffn1_3_module_sequential_1_linear_weight, x = input_125)[name = tensor("linear_28")]; - tensor input_127 = silu(x = inputs_31)[name = tensor("input_127")]; - tensor input_131 = linear(bias = encoder_ffn1_3_module_sequential_4_linear_bias, weight = encoder_ffn1_3_module_sequential_4_linear_weight, x = input_127)[name = tensor("linear_29")]; - tensor var_717 = const()[name = tensor("op_717"), val = tensor(0x1p-1)]; - tensor var_718 = mul(x = input_131, y = var_717)[name = tensor("op_718")]; - tensor input_133 = add(x = var_718, y = input_123)[name = tensor("input_133")]; + tensor input_125 = layer_norm(axes = input_125_axes_0, beta = inner_encoder_layer_norm_2_bias, epsilon = var_36, gamma = inner_encoder_layer_norm_2_weight, x = input_123)[name = tensor("input_125")]; + tensor input_127_axes_0 = const()[name = tensor("input_127_axes_0"), val = tensor([-1])]; + tensor input_127 = layer_norm(axes = input_127_axes_0, beta = inner_encoder_ffn1_3_module_sequential_0_bias, epsilon = var_36, gamma = inner_encoder_ffn1_3_module_sequential_0_weight, x = input_125)[name = tensor("input_127")]; + tensor inputs_31 = linear(bias = inner_encoder_ffn1_3_module_sequential_1_linear_bias, weight = inner_encoder_ffn1_3_module_sequential_1_linear_weight, x = input_127)[name = tensor("linear_28")]; + tensor input_129 = silu(x = inputs_31)[name = tensor("input_129")]; + tensor input_133 = linear(bias = inner_encoder_ffn1_3_module_sequential_4_linear_bias, weight = inner_encoder_ffn1_3_module_sequential_4_linear_weight, x = input_129)[name = tensor("linear_29")]; + tensor var_743 = const()[name = tensor("op_743"), val = tensor(0x1p-1)]; + tensor var_744 = mul(x = input_133, y = var_743)[name = tensor("op_744")]; + tensor input_135 = add(x = var_744, y = input_125)[name = tensor("input_135")]; tensor x_19_axes_0 = const()[name = tensor("x_19_axes_0"), val = tensor([-1])]; - tensor x_19 = layer_norm(axes = x_19_axes_0, beta = encoder_ret_lns_3_bias, epsilon = var_28, gamma = encoder_ret_lns_3_weight, x = input_133)[name = tensor("x_19")]; + tensor x_19 = layer_norm(axes = x_19_axes_0, beta = inner_encoder_ret_lns_3_bias, epsilon = var_36, gamma = inner_encoder_ret_lns_3_weight, x = input_135)[name = tensor("x_19")]; tensor prev_kv_7_begin_0 = const()[name = tensor("prev_kv_7_begin_0"), val = tensor([3, 0, 0, 0, 0])]; tensor prev_kv_7_end_0 = const()[name = tensor("prev_kv_7_end_0"), val = tensor([4, 1, 4, 64, 64])]; tensor prev_kv_7_end_mask_0 = const()[name = tensor("prev_kv_7_end_mask_0"), val = tensor([false, true, true, true, true])]; @@ -667,175 +673,168 @@ program(1.0) tensor prev_scale_7_end_mask_0 = const()[name = tensor("prev_scale_7_end_mask_0"), val = tensor([false, true])]; tensor prev_scale_7_squeeze_mask_0 = const()[name = tensor("prev_scale_7_squeeze_mask_0"), val = tensor([true, false])]; tensor prev_scale_7 = slice_by_index(begin = prev_scale_7_begin_0, end = prev_scale_7_end_0, end_mask = prev_scale_7_end_mask_0, squeeze_mask = prev_scale_7_squeeze_mask_0, x = enc_scale)[name = tensor("prev_scale_7")]; - tensor var_732 = linear(bias = encoder_q_proj_3_bias, weight = encoder_q_proj_3_weight, x = x_19)[name = tensor("linear_30")]; - tensor var_733 = const()[name = tensor("op_733"), val = tensor([1, 1, 4, 64])]; - tensor var_734 = reshape(shape = var_733, x = var_732)[name = tensor("op_734")]; + tensor var_758 = linear(bias = inner_encoder_q_proj_3_bias, weight = inner_encoder_q_proj_3_weight, x = x_19)[name = tensor("linear_30")]; + tensor var_759 = const()[name = tensor("op_759"), val = tensor([1, 1, 4, 64])]; + tensor var_760 = reshape(shape = var_759, x = var_758)[name = tensor("op_760")]; tensor q_7_perm_0 = const()[name = tensor("q_7_perm_0"), val = tensor([0, 2, 1, 3])]; - tensor var_738 = linear(bias = encoder_k_proj_3_bias, weight = encoder_k_proj_3_weight, x = x_19)[name = tensor("linear_31")]; - tensor var_739 = const()[name = tensor("op_739"), val = tensor(0x1p-3)]; - tensor var_740 = mul(x = var_738, y = var_739)[name = tensor("op_740")]; - tensor var_741 = const()[name = tensor("op_741"), val = tensor([1, 1, 4, 64])]; - tensor var_742 = reshape(shape = var_741, x = var_740)[name = tensor("op_742")]; + tensor var_764 = linear(bias = inner_encoder_k_proj_3_bias, weight = inner_encoder_k_proj_3_weight, x = x_19)[name = tensor("linear_31")]; + tensor var_765 = const()[name = tensor("op_765"), val = tensor(0x1p-3)]; + tensor var_766 = mul(x = var_764, y = var_765)[name = tensor("op_766")]; + tensor var_767 = const()[name = tensor("op_767"), val = tensor([1, 1, 4, 64])]; + tensor var_768 = reshape(shape = var_767, x = var_766)[name = tensor("op_768")]; tensor k_7_perm_0 = const()[name = tensor("k_7_perm_0"), val = tensor([0, 2, 1, 3])]; - tensor var_746 = linear(bias = encoder_v_proj_3_bias, weight = encoder_v_proj_3_weight, x = x_19)[name = tensor("linear_32")]; - tensor var_747 = const()[name = tensor("op_747"), val = tensor([1, 1, 4, 64])]; - tensor var_748 = reshape(shape = var_747, x = var_746)[name = tensor("op_748")]; + tensor var_772 = linear(bias = inner_encoder_v_proj_3_bias, weight = inner_encoder_v_proj_3_weight, x = x_19)[name = tensor("linear_32")]; + tensor var_773 = const()[name = tensor("op_773"), val = tensor([1, 1, 4, 64])]; + tensor var_774 = reshape(shape = var_773, x = var_772)[name = tensor("op_774")]; tensor v_7_perm_0 = const()[name = tensor("v_7_perm_0"), val = tensor([0, 2, 1, 3])]; - tensor input_137 = linear(bias = encoder_g_proj_3_bias, weight = encoder_g_proj_3_weight, x = x_19)[name = tensor("linear_33")]; + tensor input_139 = linear(bias = inner_encoder_g_proj_3_bias, weight = inner_encoder_g_proj_3_weight, x = x_19)[name = tensor("linear_33")]; tensor sqrt_s0_7 = sqrt(x = prev_scale_7)[name = tensor("sqrt_s0_7")]; - tensor s_t_7 = add(x = prev_scale_7, y = encoder__t_index)[name = tensor("s_t_7")]; + tensor s_t_7 = add(x = prev_scale_7, y = inner_encoder__t_index)[name = tensor("s_t_7")]; tensor sqrt_s_t_7 = sqrt(x = s_t_7)[name = tensor("sqrt_s_t_7")]; tensor qk_7_transpose_x_1 = const()[name = tensor("qk_7_transpose_x_1"), val = tensor(false)]; tensor qk_7_transpose_y_1 = const()[name = tensor("qk_7_transpose_y_1"), val = tensor(true)]; - tensor k_7 = transpose(perm = k_7_perm_0, x = var_742)[name = tensor("transpose_36")]; - tensor q_7 = transpose(perm = q_7_perm_0, x = var_734)[name = tensor("transpose_37")]; + tensor k_7 = transpose(perm = k_7_perm_0, x = var_768)[name = tensor("transpose_36")]; + tensor q_7 = transpose(perm = q_7_perm_0, x = var_760)[name = tensor("transpose_37")]; tensor qk_7 = matmul(transpose_x = qk_7_transpose_x_1, transpose_y = qk_7_transpose_y_1, x = q_7, y = k_7)[name = tensor("qk_7")]; - tensor var_758 = const()[name = tensor("op_758"), val = tensor([1, 1])]; - tensor var_759 = reshape(shape = var_758, x = sqrt_s_t_7)[name = tensor("op_759")]; - tensor M_7 = real_div(x = encoder__causal_mask, y = var_759)[name = tensor("M_7")]; - tensor var_761 = mul(x = qk_7, y = M_7)[name = tensor("op_761")]; + tensor var_784 = const()[name = tensor("op_784"), val = tensor([1, 1])]; + tensor var_785 = reshape(shape = var_784, x = sqrt_s_t_7)[name = tensor("op_785")]; + tensor M_7 = real_div(x = inner_encoder__causal_mask, y = var_785)[name = tensor("M_7")]; + tensor var_787 = mul(x = qk_7, y = M_7)[name = tensor("op_787")]; tensor inner_7_transpose_x_0 = const()[name = tensor("inner_7_transpose_x_0"), val = tensor(false)]; tensor inner_7_transpose_y_0 = const()[name = tensor("inner_7_transpose_y_0"), val = tensor(false)]; - tensor v_7 = transpose(perm = v_7_perm_0, x = var_748)[name = tensor("transpose_35")]; - tensor inner_7 = matmul(transpose_x = inner_7_transpose_x_0, transpose_y = inner_7_transpose_y_0, x = var_761, y = v_7)[name = tensor("inner_7")]; - tensor var_763_transpose_x_0 = const()[name = tensor("op_763_transpose_x_0"), val = tensor(false)]; - tensor var_763_transpose_y_0 = const()[name = tensor("op_763_transpose_y_0"), val = tensor(false)]; - tensor var_763 = matmul(transpose_x = var_763_transpose_x_0, transpose_y = var_763_transpose_y_0, x = q_7, y = prev_kv_7)[name = tensor("op_763")]; - tensor var_764 = real_div(x = sqrt_s0_7, y = sqrt_s_t_7)[name = tensor("op_764")]; - tensor var_765 = const()[name = tensor("op_765"), val = tensor([1, 1, 1, 1])]; - tensor var_766 = reshape(shape = var_765, x = var_764)[name = tensor("op_766")]; - tensor cross_7 = mul(x = var_763, y = var_766)[name = tensor("cross_7")]; + tensor v_7 = transpose(perm = v_7_perm_0, x = var_774)[name = tensor("transpose_35")]; + tensor inner_7 = matmul(transpose_x = inner_7_transpose_x_0, transpose_y = inner_7_transpose_y_0, x = var_787, y = v_7)[name = tensor("inner_7")]; + tensor var_789_transpose_x_0 = const()[name = tensor("op_789_transpose_x_0"), val = tensor(false)]; + tensor var_789_transpose_y_0 = const()[name = tensor("op_789_transpose_y_0"), val = tensor(false)]; + tensor var_789 = matmul(transpose_x = var_789_transpose_x_0, transpose_y = var_789_transpose_y_0, x = q_7, y = prev_kv_7)[name = tensor("op_789")]; + tensor var_790 = real_div(x = sqrt_s0_7, y = sqrt_s_t_7)[name = tensor("op_790")]; + tensor var_791 = const()[name = tensor("op_791"), val = tensor([1, 1, 1, 1])]; + tensor var_792 = reshape(shape = var_791, x = var_790)[name = tensor("op_792")]; + tensor cross_7 = mul(x = var_789, y = var_792)[name = tensor("cross_7")]; tensor out_19 = add(x = inner_7, y = cross_7)[name = tensor("out_19")]; - tensor var_769 = mul(x = prev_kv_7, y = sqrt_s0_7)[name = tensor("op_769")]; - tensor var_771_transpose_x_1 = const()[name = tensor("op_771_transpose_x_1"), val = tensor(true)]; - tensor var_771_transpose_y_1 = const()[name = tensor("op_771_transpose_y_1"), val = tensor(false)]; - tensor var_771 = matmul(transpose_x = var_771_transpose_x_1, transpose_y = var_771_transpose_y_1, x = k_7, y = v_7)[name = tensor("op_771")]; - tensor new_kv_unnorm_7 = add(x = var_769, y = var_771)[name = tensor("new_kv_unnorm_7")]; - tensor var_773 = const()[name = tensor("op_773"), val = tensor(0x1p+0)]; - tensor new_scale_7 = add(x = prev_scale_7, y = var_773)[name = tensor("new_scale_7")]; - tensor var_775 = sqrt(x = new_scale_7)[name = tensor("op_775")]; - tensor nkv_1 = real_div(x = new_kv_unnorm_7, y = var_775)[name = tensor("nkv_1")]; - tensor var_777_perm_0 = const()[name = tensor("op_777_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_795 = mul(x = prev_kv_7, y = sqrt_s0_7)[name = tensor("op_795")]; + tensor var_797_transpose_x_1 = const()[name = tensor("op_797_transpose_x_1"), val = tensor(true)]; + tensor var_797_transpose_y_1 = const()[name = tensor("op_797_transpose_y_1"), val = tensor(false)]; + tensor var_797 = matmul(transpose_x = var_797_transpose_x_1, transpose_y = var_797_transpose_y_1, x = k_7, y = v_7)[name = tensor("op_797")]; + tensor new_kv_unnorm_7 = add(x = var_795, y = var_797)[name = tensor("new_kv_unnorm_7")]; + tensor var_799 = const()[name = tensor("op_799"), val = tensor(0x1p+0)]; + tensor new_scale_7 = add(x = prev_scale_7, y = var_799)[name = tensor("new_scale_7")]; + tensor var_801 = sqrt(x = new_scale_7)[name = tensor("op_801")]; + tensor nkv_1 = real_div(x = new_kv_unnorm_7, y = var_801)[name = tensor("nkv_1")]; + tensor var_803_perm_0 = const()[name = tensor("op_803_perm_0"), val = tensor([0, 2, 1, 3])]; tensor out_21_axes_0 = const()[name = tensor("out_21_axes_0"), val = tensor([-1])]; - tensor var_777 = transpose(perm = var_777_perm_0, x = out_19)[name = tensor("transpose_34")]; - tensor out_21 = layer_norm(axes = out_21_axes_0, epsilon = var_17, x = var_777)[name = tensor("out_21")]; - tensor var_781 = const()[name = tensor("op_781"), val = tensor([1, 1, 256])]; - tensor out_23 = reshape(shape = var_781, x = out_21)[name = tensor("out_23")]; - tensor var_783 = silu(x = input_137)[name = tensor("op_783")]; - tensor input_139 = mul(x = var_783, y = out_23)[name = tensor("input_139")]; - tensor ret_out_7 = linear(bias = encoder_out_proj_3_bias, weight = encoder_out_proj_3_weight, x = input_139)[name = tensor("linear_34")]; - tensor x_21 = add(x = input_133, y = ret_out_7)[name = tensor("x_21")]; + tensor var_803 = transpose(perm = var_803_perm_0, x = out_19)[name = tensor("transpose_34")]; + tensor out_21 = layer_norm(axes = out_21_axes_0, epsilon = var_44, x = var_803)[name = tensor("out_21")]; + tensor var_807 = const()[name = tensor("op_807"), val = tensor([1, 1, 256])]; + tensor out_23 = reshape(shape = var_807, x = out_21)[name = tensor("out_23")]; + tensor var_809 = silu(x = input_139)[name = tensor("op_809")]; + tensor input_141 = mul(x = var_809, y = out_23)[name = tensor("input_141")]; + tensor ret_out_7 = linear(bias = inner_encoder_out_proj_3_bias, weight = inner_encoder_out_proj_3_weight, x = input_141)[name = tensor("linear_34")]; + tensor x_21 = add(x = input_135, y = ret_out_7)[name = tensor("x_21")]; tensor window_13_begin_0 = const()[name = tensor("window_13_begin_0"), val = tensor([3, 0, 0, 0])]; tensor window_13_end_0 = const()[name = tensor("window_13_end_0"), val = tensor([4, 1, 16, 256])]; tensor window_13_end_mask_0 = const()[name = tensor("window_13_end_mask_0"), val = tensor([false, true, true, true])]; tensor window_13_squeeze_mask_0 = const()[name = tensor("window_13_squeeze_mask_0"), val = tensor([true, false, false, false])]; tensor window_13 = slice_by_index(begin = window_13_begin_0, end = window_13_end_0, end_mask = window_13_end_mask_0, squeeze_mask = window_13_squeeze_mask_0, x = enc_conv_cache)[name = tensor("window_13")]; - tensor var_794_begin_0 = const()[name = tensor("op_794_begin_0"), val = tensor([0, 1, 0])]; - tensor var_794_end_0 = const()[name = tensor("op_794_end_0"), val = tensor([1, 16, 256])]; - tensor var_794_end_mask_0 = const()[name = tensor("op_794_end_mask_0"), val = tensor([true, true, true])]; - tensor var_794 = slice_by_index(begin = var_794_begin_0, end = var_794_end_0, end_mask = var_794_end_mask_0, x = window_13)[name = tensor("op_794")]; + tensor var_820_begin_0 = const()[name = tensor("op_820_begin_0"), val = tensor([0, 1, 0])]; + tensor var_820_end_0 = const()[name = tensor("op_820_end_0"), val = tensor([1, 16, 256])]; + tensor var_820_end_mask_0 = const()[name = tensor("op_820_end_mask_0"), val = tensor([true, true, true])]; + tensor var_820 = slice_by_index(begin = var_820_begin_0, end = var_820_end_0, end_mask = var_820_end_mask_0, x = window_13)[name = tensor("op_820")]; tensor window_interleave_0 = const()[name = tensor("window_interleave_0"), val = tensor(false)]; - tensor window = concat(axis = var_26, interleave = window_interleave_0, values = (var_794, x_21))[name = tensor("window")]; - tensor input_141_interleave_0 = const()[name = tensor("input_141_interleave_0"), val = tensor(false)]; - tensor input_141 = concat(axis = var_23, interleave = input_141_interleave_0, values = window)[name = tensor("input_141")]; + tensor window = concat(axis = var_52, interleave = window_interleave_0, values = (var_820, x_21))[name = tensor("window")]; + tensor input_143_interleave_0 = const()[name = tensor("input_143_interleave_0"), val = tensor(false)]; + tensor input_143 = concat(axis = var_39, interleave = input_143_interleave_0, values = window)[name = tensor("input_143")]; tensor x_23_axes_0 = const()[name = tensor("x_23_axes_0"), val = tensor([-1])]; - tensor x_23 = layer_norm(axes = x_23_axes_0, beta = encoder_conv_module_3_sequential_0_bias, epsilon = var_28, gamma = encoder_conv_module_3_sequential_0_weight, x = input_141)[name = tensor("x_23")]; - tensor input_143_perm_0 = const()[name = tensor("input_143_perm_0"), val = tensor([0, 2, 1])]; + tensor x_23 = layer_norm(axes = x_23_axes_0, beta = inner_encoder_conv_module_3_sequential_0_bias, epsilon = var_36, gamma = inner_encoder_conv_module_3_sequential_0_weight, x = input_143)[name = tensor("x_23")]; + tensor input_145_perm_0 = const()[name = tensor("input_145_perm_0"), val = tensor([0, 2, 1])]; tensor inputs_33_pad_type_0 = const()[name = tensor("inputs_33_pad_type_0"), val = tensor("valid")]; tensor inputs_33_strides_0 = const()[name = tensor("inputs_33_strides_0"), val = tensor([1])]; tensor inputs_33_pad_0 = const()[name = tensor("inputs_33_pad_0"), val = tensor([0, 0])]; tensor inputs_33_dilations_0 = const()[name = tensor("inputs_33_dilations_0"), val = tensor([1])]; tensor inputs_33_groups_0 = const()[name = tensor("inputs_33_groups_0"), val = tensor(1)]; - tensor input_143 = transpose(perm = input_143_perm_0, x = x_23)[name = tensor("transpose_33")]; - tensor inputs_33 = conv(bias = encoder_conv_module_3_sequential_2_conv_bias, dilations = inputs_33_dilations_0, groups = inputs_33_groups_0, pad = inputs_33_pad_0, pad_type = inputs_33_pad_type_0, strides = inputs_33_strides_0, weight = encoder_conv_module_3_sequential_2_conv_weight, x = input_143)[name = tensor("inputs_33")]; - tensor var_819_split_sizes_0 = const()[name = tensor("op_819_split_sizes_0"), val = tensor([256, 256])]; - tensor var_819_axis_0 = const()[name = tensor("op_819_axis_0"), val = tensor(1)]; - tensor var_819_0, tensor var_819_1 = split(axis = var_819_axis_0, split_sizes = var_819_split_sizes_0, x = inputs_33)[name = tensor("op_819")]; - tensor var_821 = sigmoid(x = var_819_1)[name = tensor("op_821")]; - tensor inputs_35 = mul(x = var_819_0, y = var_821)[name = tensor("inputs_35")]; + tensor input_145 = transpose(perm = input_145_perm_0, x = x_23)[name = tensor("transpose_33")]; + tensor inputs_33 = conv(bias = inner_encoder_conv_module_3_sequential_2_conv_bias, dilations = inputs_33_dilations_0, groups = inputs_33_groups_0, pad = inputs_33_pad_0, pad_type = inputs_33_pad_type_0, strides = inputs_33_strides_0, weight = inner_encoder_conv_module_3_sequential_2_conv_weight, x = input_145)[name = tensor("inputs_33")]; + tensor var_845_split_sizes_0 = const()[name = tensor("op_845_split_sizes_0"), val = tensor([256, 256])]; + tensor var_845_axis_0 = const()[name = tensor("op_845_axis_0"), val = tensor(1)]; + tensor var_845_0, tensor var_845_1 = split(axis = var_845_axis_0, split_sizes = var_845_split_sizes_0, x = inputs_33)[name = tensor("op_845")]; + tensor var_847 = sigmoid(x = var_845_1)[name = tensor("op_847")]; + tensor inputs_35 = mul(x = var_845_0, y = var_847)[name = tensor("inputs_35")]; tensor outputs_aug_pad_type_0 = const()[name = tensor("outputs_aug_pad_type_0"), val = tensor("custom")]; tensor outputs_aug_pad_0 = const()[name = tensor("outputs_aug_pad_0"), val = tensor([15, 15])]; tensor outputs_aug_groups_0 = const()[name = tensor("outputs_aug_groups_0"), val = tensor(256)]; tensor outputs_aug_strides_0 = const()[name = tensor("outputs_aug_strides_0"), val = tensor([1])]; tensor outputs_aug_dilations_0 = const()[name = tensor("outputs_aug_dilations_0"), val = tensor([1])]; - tensor outputs_aug = conv(dilations = outputs_aug_dilations_0, groups = outputs_aug_groups_0, pad = outputs_aug_pad_0, pad_type = outputs_aug_pad_type_0, strides = outputs_aug_strides_0, weight = encoder_conv_module_3_sequential_4_conv_weight, x = inputs_35)[name = tensor("outputs_aug")]; - tensor input_145_begin_0 = const()[name = tensor("input_145_begin_0"), val = tensor([0, 0, 0])]; - tensor input_145_end_0 = const()[name = tensor("input_145_end_0"), val = tensor([1, 256, 16])]; - tensor input_145_end_mask_0 = const()[name = tensor("input_145_end_mask_0"), val = tensor([true, true, false])]; - tensor input_145 = slice_by_index(begin = input_145_begin_0, end = input_145_end_0, end_mask = input_145_end_mask_0, x = outputs_aug)[name = tensor("input_145")]; - tensor inputs_37 = batch_norm(beta = encoder_conv_module_3_sequential_5_bias, epsilon = var_28, gamma = encoder_conv_module_3_sequential_5_weight, mean = encoder_conv_module_3_sequential_5_running_mean, variance = encoder_conv_module_3_sequential_5_running_var, x = input_145)[name = tensor("inputs_37")]; - tensor input_147 = silu(x = inputs_37)[name = tensor("input_147")]; - tensor input_149_pad_type_0 = const()[name = tensor("input_149_pad_type_0"), val = tensor("valid")]; - tensor input_149_strides_0 = const()[name = tensor("input_149_strides_0"), val = tensor([1])]; - tensor input_149_pad_0 = const()[name = tensor("input_149_pad_0"), val = tensor([0, 0])]; - tensor input_149_dilations_0 = const()[name = tensor("input_149_dilations_0"), val = tensor([1])]; - tensor input_149_groups_0 = const()[name = tensor("input_149_groups_0"), val = tensor(1)]; - tensor input_149 = conv(bias = encoder_conv_module_3_sequential_7_conv_bias, dilations = input_149_dilations_0, groups = input_149_groups_0, pad = input_149_pad_0, pad_type = input_149_pad_type_0, strides = input_149_strides_0, weight = encoder_conv_module_3_sequential_7_conv_weight, x = input_147)[name = tensor("input_149")]; + tensor outputs_aug = conv(dilations = outputs_aug_dilations_0, groups = outputs_aug_groups_0, pad = outputs_aug_pad_0, pad_type = outputs_aug_pad_type_0, strides = outputs_aug_strides_0, weight = inner_encoder_conv_module_3_sequential_4_conv_weight, x = inputs_35)[name = tensor("outputs_aug")]; + tensor input_147_begin_0 = const()[name = tensor("input_147_begin_0"), val = tensor([0, 0, 0])]; + tensor input_147_end_0 = const()[name = tensor("input_147_end_0"), val = tensor([1, 256, 16])]; + tensor input_147_end_mask_0 = const()[name = tensor("input_147_end_mask_0"), val = tensor([true, true, false])]; + tensor input_147 = slice_by_index(begin = input_147_begin_0, end = input_147_end_0, end_mask = input_147_end_mask_0, x = outputs_aug)[name = tensor("input_147")]; + tensor inputs_37 = batch_norm(beta = inner_encoder_conv_module_3_sequential_5_bias, epsilon = var_36, gamma = inner_encoder_conv_module_3_sequential_5_weight, mean = inner_encoder_conv_module_3_sequential_5_running_mean, variance = inner_encoder_conv_module_3_sequential_5_running_var, x = input_147)[name = tensor("inputs_37")]; + tensor input_149 = silu(x = inputs_37)[name = tensor("input_149")]; + tensor input_151_pad_type_0 = const()[name = tensor("input_151_pad_type_0"), val = tensor("valid")]; + tensor input_151_strides_0 = const()[name = tensor("input_151_strides_0"), val = tensor([1])]; + tensor input_151_pad_0 = const()[name = tensor("input_151_pad_0"), val = tensor([0, 0])]; + tensor input_151_dilations_0 = const()[name = tensor("input_151_dilations_0"), val = tensor([1])]; + tensor input_151_groups_0 = const()[name = tensor("input_151_groups_0"), val = tensor(1)]; + tensor input_151 = conv(bias = inner_encoder_conv_module_3_sequential_7_conv_bias, dilations = input_151_dilations_0, groups = input_151_groups_0, pad = input_151_pad_0, pad_type = input_151_pad_type_0, strides = input_151_strides_0, weight = inner_encoder_conv_module_3_sequential_7_conv_weight, x = input_149)[name = tensor("input_151")]; tensor conv_out_7_perm_0 = const()[name = tensor("conv_out_7_perm_0"), val = tensor([0, 2, 1])]; - tensor var_852_begin_0 = const()[name = tensor("op_852_begin_0"), val = tensor([0, -1, 0])]; - tensor var_852_end_0 = const()[name = tensor("op_852_end_0"), val = tensor([1, 16, 256])]; - tensor var_852_end_mask_0 = const()[name = tensor("op_852_end_mask_0"), val = tensor([true, true, true])]; - tensor conv_out_7 = transpose(perm = conv_out_7_perm_0, x = input_149)[name = tensor("transpose_32")]; - tensor var_852 = slice_by_index(begin = var_852_begin_0, end = var_852_end_0, end_mask = var_852_end_mask_0, x = conv_out_7)[name = tensor("op_852")]; - tensor var_854_perm_0 = const()[name = tensor("op_854_perm_0"), val = tensor([1, 0, 2])]; - tensor var_854 = transpose(perm = var_854_perm_0, x = var_852)[name = tensor("transpose_31")]; - tensor input_151 = add(x = x_21, y = var_854)[name = tensor("input_151")]; - tensor input_153_axes_0 = const()[name = tensor("input_153_axes_0"), val = tensor([-1])]; - tensor input_153 = layer_norm(axes = input_153_axes_0, beta = encoder_ffn2_3_module_sequential_0_bias, epsilon = var_28, gamma = encoder_ffn2_3_module_sequential_0_weight, x = input_151)[name = tensor("input_153")]; - tensor inputs = linear(bias = encoder_ffn2_3_module_sequential_1_linear_bias, weight = encoder_ffn2_3_module_sequential_1_linear_weight, x = input_153)[name = tensor("linear_35")]; - tensor input_155 = silu(x = inputs)[name = tensor("input_155")]; - tensor input_159 = linear(bias = encoder_ffn2_3_module_sequential_4_linear_bias, weight = encoder_ffn2_3_module_sequential_4_linear_weight, x = input_155)[name = tensor("linear_36")]; - tensor var_877 = const()[name = tensor("op_877"), val = tensor(0x1p-1)]; - tensor var_878 = mul(x = input_159, y = var_877)[name = tensor("op_878")]; - tensor input_161 = add(x = var_878, y = input_151)[name = tensor("input_161")]; + tensor var_878_begin_0 = const()[name = tensor("op_878_begin_0"), val = tensor([0, -1, 0])]; + tensor var_878_end_0 = const()[name = tensor("op_878_end_0"), val = tensor([1, 16, 256])]; + tensor var_878_end_mask_0 = const()[name = tensor("op_878_end_mask_0"), val = tensor([true, true, true])]; + tensor conv_out_7 = transpose(perm = conv_out_7_perm_0, x = input_151)[name = tensor("transpose_32")]; + tensor var_878 = slice_by_index(begin = var_878_begin_0, end = var_878_end_0, end_mask = var_878_end_mask_0, x = conv_out_7)[name = tensor("op_878")]; + tensor var_880_perm_0 = const()[name = tensor("op_880_perm_0"), val = tensor([1, 0, 2])]; + tensor var_880 = transpose(perm = var_880_perm_0, x = var_878)[name = tensor("transpose_31")]; + tensor input_153 = add(x = x_21, y = var_880)[name = tensor("input_153")]; + tensor input_155_axes_0 = const()[name = tensor("input_155_axes_0"), val = tensor([-1])]; + tensor input_155 = layer_norm(axes = input_155_axes_0, beta = inner_encoder_ffn2_3_module_sequential_0_bias, epsilon = var_36, gamma = inner_encoder_ffn2_3_module_sequential_0_weight, x = input_153)[name = tensor("input_155")]; + tensor inputs = linear(bias = inner_encoder_ffn2_3_module_sequential_1_linear_bias, weight = inner_encoder_ffn2_3_module_sequential_1_linear_weight, x = input_155)[name = tensor("linear_35")]; + tensor input_157 = silu(x = inputs)[name = tensor("input_157")]; + tensor input_161 = linear(bias = inner_encoder_ffn2_3_module_sequential_4_linear_bias, weight = inner_encoder_ffn2_3_module_sequential_4_linear_weight, x = input_157)[name = tensor("linear_36")]; + tensor var_903 = const()[name = tensor("op_903"), val = tensor(0x1p-1)]; + tensor var_904 = mul(x = input_161, y = var_903)[name = tensor("op_904")]; + tensor input_163 = add(x = var_904, y = input_153)[name = tensor("input_163")]; tensor x_25_axes_0 = const()[name = tensor("x_25_axes_0"), val = tensor([-1])]; - tensor x_25 = layer_norm(axes = x_25_axes_0, beta = encoder_layer_norm_3_bias, epsilon = var_28, gamma = encoder_layer_norm_3_weight, x = input_161)[name = tensor("x_25")]; + tensor x_25 = layer_norm(axes = x_25_axes_0, beta = inner_encoder_layer_norm_3_bias, epsilon = var_36, gamma = inner_encoder_layer_norm_3_weight, x = input_163)[name = tensor("x_25")]; tensor x_ct_perm_0 = const()[name = tensor("x_ct_perm_0"), val = tensor([0, 2, 1])]; tensor cat_interleave_0 = const()[name = tensor("cat_interleave_0"), val = tensor(false)]; tensor x_ct = transpose(perm = x_ct_perm_0, x = x_25)[name = tensor("transpose_30")]; - tensor cat = concat(axis = var_20, interleave = cat_interleave_0, values = (cnn_window, x_ct))[name = tensor("cat")]; + tensor cat = concat(axis = var_41, interleave = cat_interleave_0, values = (cnn_window, x_ct))[name = tensor("cat")]; tensor conv_out_pad_type_0 = const()[name = tensor("conv_out_pad_type_0"), val = tensor("valid")]; tensor conv_out_strides_0 = const()[name = tensor("conv_out_strides_0"), val = tensor([1])]; tensor conv_out_pad_0 = const()[name = tensor("conv_out_pad_0"), val = tensor([0, 0])]; tensor conv_out_dilations_0 = const()[name = tensor("conv_out_dilations_0"), val = tensor([1])]; tensor conv_out_groups_0 = const()[name = tensor("conv_out_groups_0"), val = tensor(1)]; - tensor conv_out = conv(bias = encoder_cnn_bias, dilations = conv_out_dilations_0, groups = conv_out_groups_0, pad = conv_out_pad_0, pad_type = conv_out_pad_type_0, strides = conv_out_strides_0, weight = encoder_cnn_weight, x = cat)[name = tensor("conv_out")]; - tensor var_896_begin_0 = const()[name = tensor("op_896_begin_0"), val = tensor([0, 0, 1])]; - tensor var_896_end_0 = const()[name = tensor("op_896_end_0"), val = tensor([1, 256, 19])]; - tensor var_896_end_mask_0 = const()[name = tensor("op_896_end_mask_0"), val = tensor([true, true, true])]; - tensor cnn_window_new = slice_by_index(begin = var_896_begin_0, end = var_896_end_0, end_mask = var_896_end_mask_0, x = cat)[name = tensor("op_896")]; - tensor input_163_perm_0 = const()[name = tensor("input_163_perm_0"), val = tensor([0, 2, 1])]; - tensor var_898 = const()[name = tensor("op_898"), val = tensor([-1])]; - tensor input_163 = transpose(perm = input_163_perm_0, x = conv_out)[name = tensor("transpose_29")]; - tensor var_899 = reduce_l2_norm(axes = var_898, keep_dims = var_29, x = input_163)[name = tensor("op_899")]; + tensor conv_out = conv(bias = inner_encoder_cnn_bias, dilations = conv_out_dilations_0, groups = conv_out_groups_0, pad = conv_out_pad_0, pad_type = conv_out_pad_type_0, strides = conv_out_strides_0, weight = inner_encoder_cnn_weight, x = cat)[name = tensor("conv_out")]; + tensor var_922_begin_0 = const()[name = tensor("op_922_begin_0"), val = tensor([0, 0, 1])]; + tensor var_922_end_0 = const()[name = tensor("op_922_end_0"), val = tensor([1, 256, 19])]; + tensor var_922_end_mask_0 = const()[name = tensor("op_922_end_mask_0"), val = tensor([true, true, true])]; + tensor cnn_window_new = slice_by_index(begin = var_922_begin_0, end = var_922_end_0, end_mask = var_922_end_mask_0, x = cat)[name = tensor("op_922")]; + tensor input_165_perm_0 = const()[name = tensor("input_165_perm_0"), val = tensor([0, 2, 1])]; + tensor var_924 = const()[name = tensor("op_924"), val = tensor([-1])]; + tensor input_165 = transpose(perm = input_165_perm_0, x = conv_out)[name = tensor("transpose_29")]; + tensor var_925 = reduce_l2_norm(axes = var_924, keep_dims = var_35, x = input_165)[name = tensor("op_925")]; tensor const_12 = const()[name = tensor("const_12"), val = tensor(0x1.fffffep+127)]; - tensor clip_0 = clip(alpha = var_11, beta = const_12, x = var_899)[name = tensor("clip_0")]; - tensor emb = real_div(x = input_163, y = clip_0)[name = tensor("emb")]; - tensor var_903_axis_0 = const()[name = tensor("op_903_axis_0"), val = tensor(0)]; - tensor enc_kv_new = stack(axis = var_903_axis_0, values = (var_206, var_396, var_586, nkv_1))[name = tensor("op_903")]; - tensor var_905_axis_0 = const()[name = tensor("op_905_axis_0"), val = tensor(0)]; - tensor enc_scale_new = stack(axis = var_905_axis_0, values = (new_scale_1, new_scale_3, new_scale_5, new_scale_7))[name = tensor("op_905")]; - tensor var_907_axis_0 = const()[name = tensor("op_907_axis_0"), val = tensor(0)]; - tensor enc_conv_cache_new = stack(axis = var_907_axis_0, values = (window_3, window_7, window_11, window))[name = tensor("op_907")]; - tensor var_916 = const()[name = tensor("op_916"), val = tensor(0x1.5798eep-27)]; - tensor var_921 = const()[name = tensor("op_921"), val = tensor(0x1.4f8b58p-17)]; - tensor var_923 = const()[name = tensor("op_923"), val = tensor(0x1.0c6f7ap-20)]; - tensor var_924 = const()[name = tensor("op_924"), val = tensor(true)]; - tensor var_926 = const()[name = tensor("op_926"), val = tensor(0x1p+0)]; - tensor var_930 = const()[name = tensor("op_930"), val = tensor(-1)]; - tensor var_936 = const()[name = tensor("op_936"), val = tensor(0)]; - tensor var_993 = const()[name = tensor("op_993"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(44395328)))]; - tensor var_998_axes_0 = const()[name = tensor("op_998_axes_0"), val = tensor([2])]; - tensor var_998 = expand_dims(axes = var_998_axes_0, x = emb)[name = tensor("op_998")]; + tensor clip_0 = clip(alpha = var_49, beta = const_12, x = var_925)[name = tensor("clip_0")]; + tensor emb = real_div(x = input_165, y = clip_0)[name = tensor("emb")]; + tensor var_929_axis_0 = const()[name = tensor("op_929_axis_0"), val = tensor(0)]; + tensor enc_kv_new = stack(axis = var_929_axis_0, values = (var_232, var_422, var_612, nkv_1))[name = tensor("op_929")]; + tensor var_931_axis_0 = const()[name = tensor("op_931_axis_0"), val = tensor(0)]; + tensor enc_scale_new = stack(axis = var_931_axis_0, values = (new_scale_1, new_scale_3, new_scale_5, new_scale_7))[name = tensor("op_931")]; + tensor var_933_axis_0 = const()[name = tensor("op_933_axis_0"), val = tensor(0)]; + tensor enc_conv_cache_new = stack(axis = var_933_axis_0, values = (window_3, window_7, window_11, window))[name = tensor("op_933")]; + tensor var_996 = const()[name = tensor("op_996"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(44395328)))]; + tensor var_1001_axes_0 = const()[name = tensor("op_1001_axes_0"), val = tensor([2])]; + tensor var_1001 = expand_dims(axes = var_1001_axes_0, x = emb)[name = tensor("op_1001")]; tensor emb_exp_reps_0 = const()[name = tensor("emb_exp_reps_0"), val = tensor([1, 1, 9, 1])]; - tensor emb_exp = tile(reps = emb_exp_reps_0, x = var_998)[name = tensor("emb_exp")]; - tensor input_165_interleave_0 = const()[name = tensor("input_165_interleave_0"), val = tensor(false)]; - tensor input_165 = concat(axis = var_930, interleave = input_165_interleave_0, values = (emb_exp, var_993))[name = tensor("input_165")]; - tensor x_27 = linear(bias = decoder_convert_bias, weight = decoder_convert_weight, x = input_165)[name = tensor("linear_37")]; - tensor var_1006_perm_0 = const()[name = tensor("op_1006_perm_0"), val = tensor([0, 2, 1, 3])]; - tensor var_1010 = const()[name = tensor("op_1010"), val = tensor([9, 1, 256])]; - tensor var_1006 = transpose(perm = var_1006_perm_0, x = x_27)[name = tensor("transpose_28")]; - tensor x_29 = reshape(shape = var_1010, x = var_1006)[name = tensor("x_29")]; + tensor emb_exp = tile(reps = emb_exp_reps_0, x = var_1001)[name = tensor("emb_exp")]; + tensor input_167_interleave_0 = const()[name = tensor("input_167_interleave_0"), val = tensor(false)]; + tensor input_167 = concat(axis = var_42, interleave = input_167_interleave_0, values = (emb_exp, var_996))[name = tensor("input_167")]; + tensor x_27 = linear(bias = inner_decoder_convert_bias, weight = inner_decoder_convert_weight, x = input_167)[name = tensor("linear_37")]; + tensor var_1009_perm_0 = const()[name = tensor("op_1009_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_1013 = const()[name = tensor("op_1013"), val = tensor([9, 1, 256])]; + tensor var_1009 = transpose(perm = var_1009_perm_0, x = x_27)[name = tensor("transpose_28")]; + tensor x_29 = reshape(shape = var_1013, x = var_1009)[name = tensor("x_29")]; tensor prev_kv_9_begin_0 = const()[name = tensor("prev_kv_9_begin_0"), val = tensor([0, 0, 0, 0, 0])]; tensor prev_kv_9_end_0 = const()[name = tensor("prev_kv_9_end_0"), val = tensor([1, 9, 4, 64, 64])]; tensor prev_kv_9_end_mask_0 = const()[name = tensor("prev_kv_9_end_mask_0"), val = tensor([false, true, true, true, true])]; @@ -846,131 +845,131 @@ program(1.0) tensor prev_scale_9_end_mask_0 = const()[name = tensor("prev_scale_9_end_mask_0"), val = tensor([false, true])]; tensor prev_scale_9_squeeze_mask_0 = const()[name = tensor("prev_scale_9_squeeze_mask_0"), val = tensor([true, false])]; tensor prev_scale_9 = slice_by_index(begin = prev_scale_9_begin_0, end = prev_scale_9_end_0, end_mask = prev_scale_9_end_mask_0, squeeze_mask = prev_scale_9_squeeze_mask_0, x = dec_scale)[name = tensor("prev_scale_9")]; - tensor var_1018 = linear(bias = decoder_q_proj_0_bias, weight = decoder_q_proj_0_weight, x = x_29)[name = tensor("linear_38")]; - tensor var_1019 = const()[name = tensor("op_1019"), val = tensor([9, 1, 4, 64])]; - tensor var_1020 = reshape(shape = var_1019, x = var_1018)[name = tensor("op_1020")]; + tensor var_1021 = linear(bias = inner_decoder_q_proj_0_bias, weight = inner_decoder_q_proj_0_weight, x = x_29)[name = tensor("linear_38")]; + tensor var_1022 = const()[name = tensor("op_1022"), val = tensor([9, 1, 4, 64])]; + tensor var_1023 = reshape(shape = var_1022, x = var_1021)[name = tensor("op_1023")]; tensor q_9_perm_0 = const()[name = tensor("q_9_perm_0"), val = tensor([0, 2, 1, 3])]; - tensor var_1024 = linear(bias = decoder_k_proj_0_bias, weight = decoder_k_proj_0_weight, x = x_29)[name = tensor("linear_39")]; - tensor var_1025 = const()[name = tensor("op_1025"), val = tensor(0x1p-3)]; - tensor var_1026 = mul(x = var_1024, y = var_1025)[name = tensor("op_1026")]; - tensor var_1027 = const()[name = tensor("op_1027"), val = tensor([9, 1, 4, 64])]; - tensor var_1028 = reshape(shape = var_1027, x = var_1026)[name = tensor("op_1028")]; + tensor var_1027 = linear(bias = inner_decoder_k_proj_0_bias, weight = inner_decoder_k_proj_0_weight, x = x_29)[name = tensor("linear_39")]; + tensor var_1028 = const()[name = tensor("op_1028"), val = tensor(0x1p-3)]; + tensor var_1029 = mul(x = var_1027, y = var_1028)[name = tensor("op_1029")]; + tensor var_1030 = const()[name = tensor("op_1030"), val = tensor([9, 1, 4, 64])]; + tensor var_1031 = reshape(shape = var_1030, x = var_1029)[name = tensor("op_1031")]; tensor k_9_perm_0 = const()[name = tensor("k_9_perm_0"), val = tensor([0, 2, 1, 3])]; - tensor var_1032 = linear(bias = decoder_v_proj_0_bias, weight = decoder_v_proj_0_weight, x = x_29)[name = tensor("linear_40")]; - tensor var_1033 = const()[name = tensor("op_1033"), val = tensor([9, 1, 4, 64])]; - tensor var_1034 = reshape(shape = var_1033, x = var_1032)[name = tensor("op_1034")]; + tensor var_1035 = linear(bias = inner_decoder_v_proj_0_bias, weight = inner_decoder_v_proj_0_weight, x = x_29)[name = tensor("linear_40")]; + tensor var_1036 = const()[name = tensor("op_1036"), val = tensor([9, 1, 4, 64])]; + tensor var_1037 = reshape(shape = var_1036, x = var_1035)[name = tensor("op_1037")]; tensor v_9_perm_0 = const()[name = tensor("v_9_perm_0"), val = tensor([0, 2, 1, 3])]; - tensor input_169 = linear(bias = decoder_g_proj_0_bias, weight = decoder_g_proj_0_weight, x = x_29)[name = tensor("linear_41")]; + tensor input_171 = linear(bias = inner_decoder_g_proj_0_bias, weight = inner_decoder_g_proj_0_weight, x = x_29)[name = tensor("linear_41")]; tensor cumsum_mask_1_exclusive_0 = const()[name = tensor("cumsum_mask_1_exclusive_0"), val = tensor(false)]; tensor cumsum_mask_1_reverse_0 = const()[name = tensor("cumsum_mask_1_reverse_0"), val = tensor(false)]; - tensor cumsum_mask_1 = cumsum(axis = var_936, exclusive = cumsum_mask_1_exclusive_0, reverse = cumsum_mask_1_reverse_0, x = valid_mask)[name = tensor("cumsum_mask_1")]; + tensor cumsum_mask_1 = cumsum(axis = var_39, exclusive = cumsum_mask_1_exclusive_0, reverse = cumsum_mask_1_reverse_0, x = valid_mask)[name = tensor("cumsum_mask_1")]; tensor sqrt_s0_9 = sqrt(x = prev_scale_9)[name = tensor("sqrt_s0_9")]; tensor s_t_9 = add(x = prev_scale_9, y = cumsum_mask_1)[name = tensor("s_t_9")]; tensor const_20 = const()[name = tensor("const_20"), val = tensor(0x1.fffffep+127)]; - tensor clip_1 = clip(alpha = var_926, beta = const_20, x = s_t_9)[name = tensor("clip_1")]; + tensor clip_1 = clip(alpha = var_29, beta = const_20, x = s_t_9)[name = tensor("clip_1")]; tensor sqrt_s_t_9 = sqrt(x = clip_1)[name = tensor("sqrt_s_t_9")]; tensor qk_9_transpose_x_1 = const()[name = tensor("qk_9_transpose_x_1"), val = tensor(false)]; tensor qk_9_transpose_y_1 = const()[name = tensor("qk_9_transpose_y_1"), val = tensor(true)]; - tensor k_9 = transpose(perm = k_9_perm_0, x = var_1028)[name = tensor("transpose_26")]; - tensor q_9 = transpose(perm = q_9_perm_0, x = var_1020)[name = tensor("transpose_27")]; + tensor k_9 = transpose(perm = k_9_perm_0, x = var_1031)[name = tensor("transpose_26")]; + tensor q_9 = transpose(perm = q_9_perm_0, x = var_1023)[name = tensor("transpose_27")]; tensor qk_9 = matmul(transpose_x = qk_9_transpose_x_1, transpose_y = qk_9_transpose_y_1, x = q_9, y = k_9)[name = tensor("qk_9")]; - tensor var_1046 = const()[name = tensor("op_1046"), val = tensor([1, 1])]; - tensor var_1047 = reshape(shape = var_1046, x = valid_mask)[name = tensor("op_1047")]; tensor var_1049 = const()[name = tensor("op_1049"), val = tensor([1, 1])]; - tensor var_1050 = reshape(shape = var_1049, x = sqrt_s_t_9)[name = tensor("op_1050")]; - tensor M_9 = real_div(x = var_1047, y = var_1050)[name = tensor("M_9")]; - tensor var_1052 = mul(x = qk_9, y = M_9)[name = tensor("op_1052")]; + tensor var_1050 = reshape(shape = var_1049, x = valid_mask)[name = tensor("op_1050")]; + tensor var_1052 = const()[name = tensor("op_1052"), val = tensor([1, 1])]; + tensor var_1053 = reshape(shape = var_1052, x = sqrt_s_t_9)[name = tensor("op_1053")]; + tensor M_9 = real_div(x = var_1050, y = var_1053)[name = tensor("M_9")]; + tensor var_1055 = mul(x = qk_9, y = M_9)[name = tensor("op_1055")]; tensor inner_9_transpose_x_0 = const()[name = tensor("inner_9_transpose_x_0"), val = tensor(false)]; tensor inner_9_transpose_y_0 = const()[name = tensor("inner_9_transpose_y_0"), val = tensor(false)]; - tensor v_9 = transpose(perm = v_9_perm_0, x = var_1034)[name = tensor("transpose_25")]; - tensor inner_9 = matmul(transpose_x = inner_9_transpose_x_0, transpose_y = inner_9_transpose_y_0, x = var_1052, y = v_9)[name = tensor("inner_9")]; - tensor var_1054_transpose_x_0 = const()[name = tensor("op_1054_transpose_x_0"), val = tensor(false)]; - tensor var_1054_transpose_y_0 = const()[name = tensor("op_1054_transpose_y_0"), val = tensor(false)]; - tensor var_1054 = matmul(transpose_x = var_1054_transpose_x_0, transpose_y = var_1054_transpose_y_0, x = q_9, y = prev_kv_9)[name = tensor("op_1054")]; - tensor var_1055 = real_div(x = sqrt_s0_9, y = sqrt_s_t_9)[name = tensor("op_1055")]; - tensor var_1056 = const()[name = tensor("op_1056"), val = tensor([1, 1, 1, 1])]; - tensor var_1057 = reshape(shape = var_1056, x = var_1055)[name = tensor("op_1057")]; - tensor cross_9 = mul(x = var_1054, y = var_1057)[name = tensor("cross_9")]; + tensor v_9 = transpose(perm = v_9_perm_0, x = var_1037)[name = tensor("transpose_25")]; + tensor inner_9 = matmul(transpose_x = inner_9_transpose_x_0, transpose_y = inner_9_transpose_y_0, x = var_1055, y = v_9)[name = tensor("inner_9")]; + tensor var_1057_transpose_x_0 = const()[name = tensor("op_1057_transpose_x_0"), val = tensor(false)]; + tensor var_1057_transpose_y_0 = const()[name = tensor("op_1057_transpose_y_0"), val = tensor(false)]; + tensor var_1057 = matmul(transpose_x = var_1057_transpose_x_0, transpose_y = var_1057_transpose_y_0, x = q_9, y = prev_kv_9)[name = tensor("op_1057")]; + tensor var_1058 = real_div(x = sqrt_s0_9, y = sqrt_s_t_9)[name = tensor("op_1058")]; + tensor var_1059 = const()[name = tensor("op_1059"), val = tensor([1, 1, 1, 1])]; + tensor var_1060 = reshape(shape = var_1059, x = var_1058)[name = tensor("op_1060")]; + tensor cross_9 = mul(x = var_1057, y = var_1060)[name = tensor("cross_9")]; tensor out_25 = add(x = inner_9, y = cross_9)[name = tensor("out_25")]; - tensor var_1060 = const()[name = tensor("op_1060"), val = tensor([1, 1, 1, 1])]; - tensor var_1061 = reshape(shape = var_1060, x = valid_mask)[name = tensor("op_1061")]; - tensor v_masked_1 = mul(x = v_9, y = var_1061)[name = tensor("v_masked_1")]; - tensor var_1063 = mul(x = prev_kv_9, y = sqrt_s0_9)[name = tensor("op_1063")]; - tensor var_1065_transpose_x_1 = const()[name = tensor("op_1065_transpose_x_1"), val = tensor(true)]; - tensor var_1065_transpose_y_1 = const()[name = tensor("op_1065_transpose_y_1"), val = tensor(false)]; - tensor var_1065 = matmul(transpose_x = var_1065_transpose_x_1, transpose_y = var_1065_transpose_y_1, x = k_9, y = v_masked_1)[name = tensor("op_1065")]; - tensor new_kv_unnorm_9 = add(x = var_1063, y = var_1065)[name = tensor("new_kv_unnorm_9")]; - tensor var_1067_keep_dims_0 = const()[name = tensor("op_1067_keep_dims_0"), val = tensor(false)]; - tensor var_1067 = reduce_sum(keep_dims = var_1067_keep_dims_0, x = valid_mask)[name = tensor("op_1067")]; - tensor var_1068 = const()[name = tensor("op_1068"), val = tensor([1])]; - tensor var_1069 = reshape(shape = var_1068, x = var_1067)[name = tensor("op_1069")]; - tensor new_scale_9 = add(x = prev_scale_9, y = var_1069)[name = tensor("new_scale_9")]; + tensor var_1063 = const()[name = tensor("op_1063"), val = tensor([1, 1, 1, 1])]; + tensor var_1064 = reshape(shape = var_1063, x = valid_mask)[name = tensor("op_1064")]; + tensor v_masked_1 = mul(x = v_9, y = var_1064)[name = tensor("v_masked_1")]; + tensor var_1066 = mul(x = prev_kv_9, y = sqrt_s0_9)[name = tensor("op_1066")]; + tensor var_1068_transpose_x_1 = const()[name = tensor("op_1068_transpose_x_1"), val = tensor(true)]; + tensor var_1068_transpose_y_1 = const()[name = tensor("op_1068_transpose_y_1"), val = tensor(false)]; + tensor var_1068 = matmul(transpose_x = var_1068_transpose_x_1, transpose_y = var_1068_transpose_y_1, x = k_9, y = v_masked_1)[name = tensor("op_1068")]; + tensor new_kv_unnorm_9 = add(x = var_1066, y = var_1068)[name = tensor("new_kv_unnorm_9")]; + tensor var_1070_keep_dims_0 = const()[name = tensor("op_1070_keep_dims_0"), val = tensor(false)]; + tensor var_1070 = reduce_sum(keep_dims = var_1070_keep_dims_0, x = valid_mask)[name = tensor("op_1070")]; + tensor var_1071 = const()[name = tensor("op_1071"), val = tensor([1])]; + tensor var_1072 = reshape(shape = var_1071, x = var_1070)[name = tensor("op_1072")]; + tensor new_scale_9 = add(x = prev_scale_9, y = var_1072)[name = tensor("new_scale_9")]; tensor const_21 = const()[name = tensor("const_21"), val = tensor(0x1.fffffep+127)]; - tensor clip_2 = clip(alpha = var_926, beta = const_21, x = new_scale_9)[name = tensor("clip_2")]; + tensor clip_2 = clip(alpha = var_29, beta = const_21, x = new_scale_9)[name = tensor("clip_2")]; tensor sqrt_new_scale_1 = sqrt(x = clip_2)[name = tensor("sqrt_new_scale_1")]; - tensor var_1073 = real_div(x = new_kv_unnorm_9, y = sqrt_new_scale_1)[name = tensor("op_1073")]; - tensor var_1074_perm_0 = const()[name = tensor("op_1074_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_1076 = real_div(x = new_kv_unnorm_9, y = sqrt_new_scale_1)[name = tensor("op_1076")]; + tensor var_1077_perm_0 = const()[name = tensor("op_1077_perm_0"), val = tensor([0, 2, 1, 3])]; tensor out_27_axes_0 = const()[name = tensor("out_27_axes_0"), val = tensor([-1])]; - tensor var_1074 = transpose(perm = var_1074_perm_0, x = out_25)[name = tensor("transpose_24")]; - tensor out_27 = layer_norm(axes = out_27_axes_0, epsilon = var_923, x = var_1074)[name = tensor("out_27")]; - tensor var_1078 = const()[name = tensor("op_1078"), val = tensor([9, 1, 256])]; - tensor out_29 = reshape(shape = var_1078, x = out_27)[name = tensor("out_29")]; - tensor var_1080 = silu(x = input_169)[name = tensor("op_1080")]; - tensor input_171 = mul(x = var_1080, y = out_29)[name = tensor("input_171")]; - tensor ret_out_9 = linear(bias = decoder_out_proj_0_bias, weight = decoder_out_proj_0_weight, x = input_171)[name = tensor("linear_42")]; - tensor input_173 = add(x = x_29, y = ret_out_9)[name = tensor("input_173")]; + tensor var_1077 = transpose(perm = var_1077_perm_0, x = out_25)[name = tensor("transpose_24")]; + tensor out_27 = layer_norm(axes = out_27_axes_0, epsilon = var_44, x = var_1077)[name = tensor("out_27")]; + tensor var_1081 = const()[name = tensor("op_1081"), val = tensor([9, 1, 256])]; + tensor out_29 = reshape(shape = var_1081, x = out_27)[name = tensor("out_29")]; + tensor var_1083 = silu(x = input_171)[name = tensor("op_1083")]; + tensor input_173 = mul(x = var_1083, y = out_29)[name = tensor("input_173")]; + tensor ret_out_9 = linear(bias = inner_decoder_out_proj_0_bias, weight = inner_decoder_out_proj_0_weight, x = input_173)[name = tensor("linear_42")]; + tensor input_175 = add(x = x_29, y = ret_out_9)[name = tensor("input_175")]; tensor xt_1_axes_0 = const()[name = tensor("xt_1_axes_0"), val = tensor([-1])]; - tensor xt_1 = layer_norm(axes = xt_1_axes_0, beta = decoder_norm11_0_bias, epsilon = var_921, gamma = decoder_norm11_0_weight, x = input_173)[name = tensor("xt_1")]; - tensor var_1090 = const()[name = tensor("op_1090"), val = tensor([1, 9, 1, 256])]; - tensor var_1091 = reshape(shape = var_1090, x = xt_1)[name = tensor("op_1091")]; - tensor var_1092_perm_0 = const()[name = tensor("op_1092_perm_0"), val = tensor([0, 2, 1, 3])]; - tensor var_1095 = const()[name = tensor("op_1095"), val = tensor([1, 9, 256])]; - tensor var_1092 = transpose(perm = var_1092_perm_0, x = var_1091)[name = tensor("transpose_23")]; - tensor query_1 = reshape(shape = var_1095, x = var_1092)[name = tensor("query_1")]; + tensor xt_1 = layer_norm(axes = xt_1_axes_0, beta = inner_decoder_norm11_0_bias, epsilon = var_36, gamma = inner_decoder_norm11_0_weight, x = input_175)[name = tensor("xt_1")]; + tensor var_1093 = const()[name = tensor("op_1093"), val = tensor([1, 9, 1, 256])]; + tensor var_1094 = reshape(shape = var_1093, x = xt_1)[name = tensor("op_1094")]; + tensor var_1095_perm_0 = const()[name = tensor("op_1095_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_1098 = const()[name = tensor("op_1098"), val = tensor([1, 9, 256])]; + tensor var_1095 = transpose(perm = var_1095_perm_0, x = var_1094)[name = tensor("transpose_23")]; + tensor query_1 = reshape(shape = var_1098, x = var_1095)[name = tensor("query_1")]; tensor query_3_perm_0 = const()[name = tensor("query_3_perm_0"), val = tensor([1, 0, 2])]; tensor query_3 = transpose(perm = query_3_perm_0, x = query_1)[name = tensor("transpose_22")]; - tensor var_1118 = linear(bias = decoder_self_attn2_0_in_proj_bias, weight = decoder_self_attn2_0_in_proj_weight, x = query_3)[name = tensor("linear_43")]; + tensor var_1121 = linear(bias = inner_decoder_self_attn2_0_in_proj_bias, weight = inner_decoder_self_attn2_0_in_proj_weight, x = query_3)[name = tensor("linear_43")]; tensor concat_1 = const()[name = tensor("concat_1"), val = tensor([9, 1, 3, 256])]; - tensor var_1120 = reshape(shape = concat_1, x = var_1118)[name = tensor("op_1120")]; - tensor var_1121_axes_0 = const()[name = tensor("op_1121_axes_0"), val = tensor([0])]; - tensor var_1121 = expand_dims(axes = var_1121_axes_0, x = var_1120)[name = tensor("op_1121")]; - tensor var_1122_perm_0 = const()[name = tensor("op_1122_perm_0"), val = tensor([-2, 1, 2, 0, 4])]; - tensor var_1123_axes_0 = const()[name = tensor("op_1123_axes_0"), val = tensor([-2])]; - tensor var_1122 = transpose(perm = var_1122_perm_0, x = var_1121)[name = tensor("transpose_21")]; - tensor var_1123 = squeeze(axes = var_1123_axes_0, x = var_1122)[name = tensor("op_1123")]; + tensor var_1123 = reshape(shape = concat_1, x = var_1121)[name = tensor("op_1123")]; + tensor var_1124_axes_0 = const()[name = tensor("op_1124_axes_0"), val = tensor([0])]; + tensor var_1124 = expand_dims(axes = var_1124_axes_0, x = var_1123)[name = tensor("op_1124")]; + tensor var_1125_perm_0 = const()[name = tensor("op_1125_perm_0"), val = tensor([-2, 1, 2, 0, 4])]; + tensor var_1126_axes_0 = const()[name = tensor("op_1126_axes_0"), val = tensor([-2])]; + tensor var_1125 = transpose(perm = var_1125_perm_0, x = var_1124)[name = tensor("transpose_21")]; + tensor var_1126 = squeeze(axes = var_1126_axes_0, x = var_1125)[name = tensor("op_1126")]; tensor q_11_begin_0 = const()[name = tensor("q_11_begin_0"), val = tensor([0, 0, 0, 0])]; tensor q_11_end_0 = const()[name = tensor("q_11_end_0"), val = tensor([1, 9, 1, 256])]; tensor q_11_end_mask_0 = const()[name = tensor("q_11_end_mask_0"), val = tensor([false, true, true, true])]; tensor q_11_squeeze_mask_0 = const()[name = tensor("q_11_squeeze_mask_0"), val = tensor([true, false, false, false])]; - tensor q_11 = slice_by_index(begin = q_11_begin_0, end = q_11_end_0, end_mask = q_11_end_mask_0, squeeze_mask = q_11_squeeze_mask_0, x = var_1123)[name = tensor("q_11")]; + tensor q_11 = slice_by_index(begin = q_11_begin_0, end = q_11_end_0, end_mask = q_11_end_mask_0, squeeze_mask = q_11_squeeze_mask_0, x = var_1126)[name = tensor("q_11")]; tensor k_11_begin_0 = const()[name = tensor("k_11_begin_0"), val = tensor([1, 0, 0, 0])]; tensor k_11_end_0 = const()[name = tensor("k_11_end_0"), val = tensor([2, 9, 1, 256])]; tensor k_11_end_mask_0 = const()[name = tensor("k_11_end_mask_0"), val = tensor([false, true, true, true])]; tensor k_11_squeeze_mask_0 = const()[name = tensor("k_11_squeeze_mask_0"), val = tensor([true, false, false, false])]; - tensor k_11 = slice_by_index(begin = k_11_begin_0, end = k_11_end_0, end_mask = k_11_end_mask_0, squeeze_mask = k_11_squeeze_mask_0, x = var_1123)[name = tensor("k_11")]; + tensor k_11 = slice_by_index(begin = k_11_begin_0, end = k_11_end_0, end_mask = k_11_end_mask_0, squeeze_mask = k_11_squeeze_mask_0, x = var_1126)[name = tensor("k_11")]; tensor v_11_begin_0 = const()[name = tensor("v_11_begin_0"), val = tensor([2, 0, 0, 0])]; tensor v_11_end_0 = const()[name = tensor("v_11_end_0"), val = tensor([3, 9, 1, 256])]; tensor v_11_end_mask_0 = const()[name = tensor("v_11_end_mask_0"), val = tensor([false, true, true, true])]; tensor v_11_squeeze_mask_0 = const()[name = tensor("v_11_squeeze_mask_0"), val = tensor([true, false, false, false])]; - tensor v_11 = slice_by_index(begin = v_11_begin_0, end = v_11_end_0, end_mask = v_11_end_mask_0, squeeze_mask = v_11_squeeze_mask_0, x = var_1123)[name = tensor("v_11")]; - tensor var_1131 = const()[name = tensor("op_1131"), val = tensor([9, 4, 64])]; - tensor var_1132 = reshape(shape = var_1131, x = q_11)[name = tensor("op_1132")]; + tensor v_11 = slice_by_index(begin = v_11_begin_0, end = v_11_end_0, end_mask = v_11_end_mask_0, squeeze_mask = v_11_squeeze_mask_0, x = var_1126)[name = tensor("v_11")]; + tensor var_1134 = const()[name = tensor("op_1134"), val = tensor([9, 4, 64])]; + tensor var_1135 = reshape(shape = var_1134, x = q_11)[name = tensor("op_1135")]; tensor q_13_perm_0 = const()[name = tensor("q_13_perm_0"), val = tensor([1, 0, 2])]; - tensor var_1138 = const()[name = tensor("op_1138"), val = tensor([9, 4, 64])]; - tensor var_1139 = reshape(shape = var_1138, x = k_11)[name = tensor("op_1139")]; + tensor var_1141 = const()[name = tensor("op_1141"), val = tensor([9, 4, 64])]; + tensor var_1142 = reshape(shape = var_1141, x = k_11)[name = tensor("op_1142")]; tensor k_13_perm_0 = const()[name = tensor("k_13_perm_0"), val = tensor([1, 0, 2])]; - tensor var_1145 = const()[name = tensor("op_1145"), val = tensor([9, 4, 64])]; - tensor var_1146 = reshape(shape = var_1145, x = v_11)[name = tensor("op_1146")]; + tensor var_1148 = const()[name = tensor("op_1148"), val = tensor([9, 4, 64])]; + tensor var_1149 = reshape(shape = var_1148, x = v_11)[name = tensor("op_1149")]; tensor v_13_perm_0 = const()[name = tensor("v_13_perm_0"), val = tensor([1, 0, 2])]; - tensor var_1149 = const()[name = tensor("op_1149"), val = tensor([1, 4, 9, 64])]; - tensor q_13 = transpose(perm = q_13_perm_0, x = var_1132)[name = tensor("transpose_20")]; - tensor q_15 = reshape(shape = var_1149, x = q_13)[name = tensor("q_15")]; - tensor var_1151 = const()[name = tensor("op_1151"), val = tensor([1, 4, 9, 64])]; - tensor k_13 = transpose(perm = k_13_perm_0, x = var_1139)[name = tensor("transpose_19")]; - tensor k_15 = reshape(shape = var_1151, x = k_13)[name = tensor("k_15")]; - tensor var_1153 = const()[name = tensor("op_1153"), val = tensor([1, 4, 9, 64])]; - tensor v_13 = transpose(perm = v_13_perm_0, x = var_1146)[name = tensor("transpose_18")]; - tensor v_15 = reshape(shape = var_1153, x = v_13)[name = tensor("v_15")]; + tensor var_1152 = const()[name = tensor("op_1152"), val = tensor([1, 4, 9, 64])]; + tensor q_13 = transpose(perm = q_13_perm_0, x = var_1135)[name = tensor("transpose_20")]; + tensor q_15 = reshape(shape = var_1152, x = q_13)[name = tensor("q_15")]; + tensor var_1154 = const()[name = tensor("op_1154"), val = tensor([1, 4, 9, 64])]; + tensor k_13 = transpose(perm = k_13_perm_0, x = var_1142)[name = tensor("transpose_19")]; + tensor k_15 = reshape(shape = var_1154, x = k_13)[name = tensor("k_15")]; + tensor var_1156 = const()[name = tensor("op_1156"), val = tensor([1, 4, 9, 64])]; + tensor v_13 = transpose(perm = v_13_perm_0, x = var_1149)[name = tensor("transpose_18")]; + tensor v_15 = reshape(shape = var_1156, x = v_13)[name = tensor("v_15")]; tensor mul_1_y_0 = const()[name = tensor("mul_1_y_0"), val = tensor(0x1p-3)]; tensor mul_1 = mul(x = q_15, y = mul_1_y_0)[name = tensor("mul_1")]; tensor matmul_0_transpose_y_0 = const()[name = tensor("matmul_0_transpose_y_0"), val = tensor(true)]; @@ -981,30 +980,30 @@ program(1.0) tensor attn_output_1_transpose_x_0 = const()[name = tensor("attn_output_1_transpose_x_0"), val = tensor(false)]; tensor attn_output_1_transpose_y_0 = const()[name = tensor("attn_output_1_transpose_y_0"), val = tensor(false)]; tensor attn_output_1 = matmul(transpose_x = attn_output_1_transpose_x_0, transpose_y = attn_output_1_transpose_y_0, x = softmax_0, y = v_15)[name = tensor("attn_output_1")]; - tensor var_1156 = const()[name = tensor("op_1156"), val = tensor([2, 0, 1, 3])]; - tensor var_1161 = const()[name = tensor("op_1161"), val = tensor([9, 256])]; - tensor var_1157 = transpose(perm = var_1156, x = attn_output_1)[name = tensor("transpose_17")]; - tensor attn_output_3 = reshape(shape = var_1161, x = var_1157)[name = tensor("attn_output_3")]; - tensor attn_output_5 = linear(bias = decoder_self_attn2_0_out_proj_bias, weight = decoder_self_attn2_0_out_proj_weight, x = attn_output_3)[name = tensor("linear_44")]; - tensor var_1165 = const()[name = tensor("op_1165"), val = tensor([9, 1, 256])]; - tensor attn_output_7 = reshape(shape = var_1165, x = attn_output_5)[name = tensor("attn_output_7")]; + tensor var_1159 = const()[name = tensor("op_1159"), val = tensor([2, 0, 1, 3])]; + tensor var_1164 = const()[name = tensor("op_1164"), val = tensor([9, 256])]; + tensor var_1160 = transpose(perm = var_1159, x = attn_output_1)[name = tensor("transpose_17")]; + tensor attn_output_3 = reshape(shape = var_1164, x = var_1160)[name = tensor("attn_output_3")]; + tensor attn_output_5 = linear(bias = inner_decoder_self_attn2_0_out_proj_bias, weight = inner_decoder_self_attn2_0_out_proj_weight, x = attn_output_3)[name = tensor("linear_44")]; + tensor var_1168 = const()[name = tensor("op_1168"), val = tensor([9, 1, 256])]; + tensor attn_output_7 = reshape(shape = var_1168, x = attn_output_5)[name = tensor("attn_output_7")]; tensor sa2_out_1_perm_0 = const()[name = tensor("sa2_out_1_perm_0"), val = tensor([1, 0, 2])]; tensor sa2_out_1 = transpose(perm = sa2_out_1_perm_0, x = attn_output_7)[name = tensor("transpose_16")]; - tensor input_175 = add(x = query_1, y = sa2_out_1)[name = tensor("input_175")]; - tensor input_177_axes_0 = const()[name = tensor("input_177_axes_0"), val = tensor([-1])]; - tensor input_177 = layer_norm(axes = input_177_axes_0, beta = decoder_norm21_0_bias, epsilon = var_921, gamma = decoder_norm21_0_weight, x = input_175)[name = tensor("input_177")]; - tensor input_179 = linear(bias = decoder_linear1_0_bias, weight = decoder_linear1_0_weight, x = input_177)[name = tensor("linear_45")]; - tensor input_181 = relu(x = input_179)[name = tensor("input_181")]; - tensor ffn_out_1 = linear(bias = decoder_linear2_0_bias, weight = decoder_linear2_0_weight, x = input_181)[name = tensor("linear_46")]; - tensor input_183 = add(x = input_177, y = ffn_out_1)[name = tensor("input_183")]; + tensor input_177 = add(x = query_1, y = sa2_out_1)[name = tensor("input_177")]; + tensor input_179_axes_0 = const()[name = tensor("input_179_axes_0"), val = tensor([-1])]; + tensor input_179 = layer_norm(axes = input_179_axes_0, beta = inner_decoder_norm21_0_bias, epsilon = var_36, gamma = inner_decoder_norm21_0_weight, x = input_177)[name = tensor("input_179")]; + tensor input_181 = linear(bias = inner_decoder_linear1_0_bias, weight = inner_decoder_linear1_0_weight, x = input_179)[name = tensor("linear_45")]; + tensor input_183 = relu(x = input_181)[name = tensor("input_183")]; + tensor ffn_out_1 = linear(bias = inner_decoder_linear2_0_bias, weight = inner_decoder_linear2_0_weight, x = input_183)[name = tensor("linear_46")]; + tensor input_185 = add(x = input_179, y = ffn_out_1)[name = tensor("input_185")]; tensor xt_3_axes_0 = const()[name = tensor("xt_3_axes_0"), val = tensor([-1])]; - tensor xt_3 = layer_norm(axes = xt_3_axes_0, beta = decoder_norm22_0_bias, epsilon = var_921, gamma = decoder_norm22_0_weight, x = input_183)[name = tensor("xt_3")]; - tensor var_1185 = const()[name = tensor("op_1185"), val = tensor([1, 1, 9, 256])]; - tensor x_31 = reshape(shape = var_1185, x = xt_3)[name = tensor("x_31")]; - tensor var_1187_perm_0 = const()[name = tensor("op_1187_perm_0"), val = tensor([0, 2, 1, 3])]; - tensor var_1191 = const()[name = tensor("op_1191"), val = tensor([9, 1, 256])]; - tensor var_1187 = transpose(perm = var_1187_perm_0, x = x_31)[name = tensor("transpose_15")]; - tensor x = reshape(shape = var_1191, x = var_1187)[name = tensor("x")]; + tensor xt_3 = layer_norm(axes = xt_3_axes_0, beta = inner_decoder_norm22_0_bias, epsilon = var_36, gamma = inner_decoder_norm22_0_weight, x = input_185)[name = tensor("xt_3")]; + tensor var_1188 = const()[name = tensor("op_1188"), val = tensor([1, 1, 9, 256])]; + tensor x_31 = reshape(shape = var_1188, x = xt_3)[name = tensor("x_31")]; + tensor var_1190_perm_0 = const()[name = tensor("op_1190_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_1194 = const()[name = tensor("op_1194"), val = tensor([9, 1, 256])]; + tensor var_1190 = transpose(perm = var_1190_perm_0, x = x_31)[name = tensor("transpose_15")]; + tensor x = reshape(shape = var_1194, x = var_1190)[name = tensor("x")]; tensor prev_kv_begin_0 = const()[name = tensor("prev_kv_begin_0"), val = tensor([1, 0, 0, 0, 0])]; tensor prev_kv_end_0 = const()[name = tensor("prev_kv_end_0"), val = tensor([2, 9, 4, 64, 64])]; tensor prev_kv_end_mask_0 = const()[name = tensor("prev_kv_end_mask_0"), val = tensor([false, true, true, true, true])]; @@ -1015,120 +1014,120 @@ program(1.0) tensor prev_scale_end_mask_0 = const()[name = tensor("prev_scale_end_mask_0"), val = tensor([false, true])]; tensor prev_scale_squeeze_mask_0 = const()[name = tensor("prev_scale_squeeze_mask_0"), val = tensor([true, false])]; tensor prev_scale = slice_by_index(begin = prev_scale_begin_0, end = prev_scale_end_0, end_mask = prev_scale_end_mask_0, squeeze_mask = prev_scale_squeeze_mask_0, x = dec_scale)[name = tensor("prev_scale")]; - tensor var_1199 = linear(bias = decoder_q_proj_1_bias, weight = decoder_q_proj_1_weight, x = x)[name = tensor("linear_47")]; - tensor var_1200 = const()[name = tensor("op_1200"), val = tensor([9, 1, 4, 64])]; - tensor var_1201 = reshape(shape = var_1200, x = var_1199)[name = tensor("op_1201")]; + tensor var_1202 = linear(bias = inner_decoder_q_proj_1_bias, weight = inner_decoder_q_proj_1_weight, x = x)[name = tensor("linear_47")]; + tensor var_1203 = const()[name = tensor("op_1203"), val = tensor([9, 1, 4, 64])]; + tensor var_1204 = reshape(shape = var_1203, x = var_1202)[name = tensor("op_1204")]; tensor q_17_perm_0 = const()[name = tensor("q_17_perm_0"), val = tensor([0, 2, 1, 3])]; - tensor var_1205 = linear(bias = decoder_k_proj_1_bias, weight = decoder_k_proj_1_weight, x = x)[name = tensor("linear_48")]; - tensor var_1206 = const()[name = tensor("op_1206"), val = tensor(0x1p-3)]; - tensor var_1207 = mul(x = var_1205, y = var_1206)[name = tensor("op_1207")]; - tensor var_1208 = const()[name = tensor("op_1208"), val = tensor([9, 1, 4, 64])]; - tensor var_1209 = reshape(shape = var_1208, x = var_1207)[name = tensor("op_1209")]; + tensor var_1208 = linear(bias = inner_decoder_k_proj_1_bias, weight = inner_decoder_k_proj_1_weight, x = x)[name = tensor("linear_48")]; + tensor var_1209 = const()[name = tensor("op_1209"), val = tensor(0x1p-3)]; + tensor var_1210 = mul(x = var_1208, y = var_1209)[name = tensor("op_1210")]; + tensor var_1211 = const()[name = tensor("op_1211"), val = tensor([9, 1, 4, 64])]; + tensor var_1212 = reshape(shape = var_1211, x = var_1210)[name = tensor("op_1212")]; tensor k_17_perm_0 = const()[name = tensor("k_17_perm_0"), val = tensor([0, 2, 1, 3])]; - tensor var_1213 = linear(bias = decoder_v_proj_1_bias, weight = decoder_v_proj_1_weight, x = x)[name = tensor("linear_49")]; - tensor var_1214 = const()[name = tensor("op_1214"), val = tensor([9, 1, 4, 64])]; - tensor var_1215 = reshape(shape = var_1214, x = var_1213)[name = tensor("op_1215")]; + tensor var_1216 = linear(bias = inner_decoder_v_proj_1_bias, weight = inner_decoder_v_proj_1_weight, x = x)[name = tensor("linear_49")]; + tensor var_1217 = const()[name = tensor("op_1217"), val = tensor([9, 1, 4, 64])]; + tensor var_1218 = reshape(shape = var_1217, x = var_1216)[name = tensor("op_1218")]; tensor v_17_perm_0 = const()[name = tensor("v_17_perm_0"), val = tensor([0, 2, 1, 3])]; - tensor input_187 = linear(bias = decoder_g_proj_1_bias, weight = decoder_g_proj_1_weight, x = x)[name = tensor("linear_50")]; + tensor input_189 = linear(bias = inner_decoder_g_proj_1_bias, weight = inner_decoder_g_proj_1_weight, x = x)[name = tensor("linear_50")]; tensor sqrt_s0 = sqrt(x = prev_scale)[name = tensor("sqrt_s0")]; tensor s_t = add(x = prev_scale, y = cumsum_mask_1)[name = tensor("s_t")]; tensor const_32 = const()[name = tensor("const_32"), val = tensor(0x1.fffffep+127)]; - tensor clip_3 = clip(alpha = var_926, beta = const_32, x = s_t)[name = tensor("clip_3")]; + tensor clip_3 = clip(alpha = var_29, beta = const_32, x = s_t)[name = tensor("clip_3")]; tensor sqrt_s_t = sqrt(x = clip_3)[name = tensor("sqrt_s_t")]; tensor qk_transpose_x_1 = const()[name = tensor("qk_transpose_x_1"), val = tensor(false)]; tensor qk_transpose_y_1 = const()[name = tensor("qk_transpose_y_1"), val = tensor(true)]; - tensor k_17 = transpose(perm = k_17_perm_0, x = var_1209)[name = tensor("transpose_13")]; - tensor q_17 = transpose(perm = q_17_perm_0, x = var_1201)[name = tensor("transpose_14")]; + tensor k_17 = transpose(perm = k_17_perm_0, x = var_1212)[name = tensor("transpose_13")]; + tensor q_17 = transpose(perm = q_17_perm_0, x = var_1204)[name = tensor("transpose_14")]; tensor qk = matmul(transpose_x = qk_transpose_x_1, transpose_y = qk_transpose_y_1, x = q_17, y = k_17)[name = tensor("qk")]; - tensor var_1230 = const()[name = tensor("op_1230"), val = tensor([1, 1])]; - tensor var_1231 = reshape(shape = var_1230, x = sqrt_s_t)[name = tensor("op_1231")]; - tensor M = real_div(x = var_1047, y = var_1231)[name = tensor("M")]; - tensor var_1233 = mul(x = qk, y = M)[name = tensor("op_1233")]; - tensor inner_transpose_x_0 = const()[name = tensor("inner_transpose_x_0"), val = tensor(false)]; - tensor inner_transpose_y_0 = const()[name = tensor("inner_transpose_y_0"), val = tensor(false)]; - tensor v_17 = transpose(perm = v_17_perm_0, x = var_1215)[name = tensor("transpose_12")]; - tensor inner = matmul(transpose_x = inner_transpose_x_0, transpose_y = inner_transpose_y_0, x = var_1233, y = v_17)[name = tensor("inner")]; - tensor var_1235_transpose_x_0 = const()[name = tensor("op_1235_transpose_x_0"), val = tensor(false)]; - tensor var_1235_transpose_y_0 = const()[name = tensor("op_1235_transpose_y_0"), val = tensor(false)]; - tensor var_1235 = matmul(transpose_x = var_1235_transpose_x_0, transpose_y = var_1235_transpose_y_0, x = q_17, y = prev_kv)[name = tensor("op_1235")]; - tensor var_1236 = real_div(x = sqrt_s0, y = sqrt_s_t)[name = tensor("op_1236")]; - tensor var_1237 = const()[name = tensor("op_1237"), val = tensor([1, 1, 1, 1])]; - tensor var_1238 = reshape(shape = var_1237, x = var_1236)[name = tensor("op_1238")]; - tensor cross = mul(x = var_1235, y = var_1238)[name = tensor("cross")]; - tensor out_31 = add(x = inner, y = cross)[name = tensor("out_31")]; - tensor v_masked = mul(x = v_17, y = var_1061)[name = tensor("v_masked")]; - tensor var_1244 = mul(x = prev_kv, y = sqrt_s0)[name = tensor("op_1244")]; - tensor var_1246_transpose_x_1 = const()[name = tensor("op_1246_transpose_x_1"), val = tensor(true)]; - tensor var_1246_transpose_y_1 = const()[name = tensor("op_1246_transpose_y_1"), val = tensor(false)]; - tensor var_1246 = matmul(transpose_x = var_1246_transpose_x_1, transpose_y = var_1246_transpose_y_1, x = k_17, y = v_masked)[name = tensor("op_1246")]; - tensor new_kv_unnorm = add(x = var_1244, y = var_1246)[name = tensor("new_kv_unnorm")]; - tensor new_scale = add(x = prev_scale, y = var_1069)[name = tensor("new_scale")]; + tensor var_1233 = const()[name = tensor("op_1233"), val = tensor([1, 1])]; + tensor var_1234 = reshape(shape = var_1233, x = sqrt_s_t)[name = tensor("op_1234")]; + tensor M = real_div(x = var_1050, y = var_1234)[name = tensor("M")]; + tensor var_1236 = mul(x = qk, y = M)[name = tensor("op_1236")]; + tensor inner_11_transpose_x_0 = const()[name = tensor("inner_11_transpose_x_0"), val = tensor(false)]; + tensor inner_11_transpose_y_0 = const()[name = tensor("inner_11_transpose_y_0"), val = tensor(false)]; + tensor v_17 = transpose(perm = v_17_perm_0, x = var_1218)[name = tensor("transpose_12")]; + tensor inner_11 = matmul(transpose_x = inner_11_transpose_x_0, transpose_y = inner_11_transpose_y_0, x = var_1236, y = v_17)[name = tensor("inner_11")]; + tensor var_1238_transpose_x_0 = const()[name = tensor("op_1238_transpose_x_0"), val = tensor(false)]; + tensor var_1238_transpose_y_0 = const()[name = tensor("op_1238_transpose_y_0"), val = tensor(false)]; + tensor var_1238 = matmul(transpose_x = var_1238_transpose_x_0, transpose_y = var_1238_transpose_y_0, x = q_17, y = prev_kv)[name = tensor("op_1238")]; + tensor var_1239 = real_div(x = sqrt_s0, y = sqrt_s_t)[name = tensor("op_1239")]; + tensor var_1240 = const()[name = tensor("op_1240"), val = tensor([1, 1, 1, 1])]; + tensor var_1241 = reshape(shape = var_1240, x = var_1239)[name = tensor("op_1241")]; + tensor cross = mul(x = var_1238, y = var_1241)[name = tensor("cross")]; + tensor out_31 = add(x = inner_11, y = cross)[name = tensor("out_31")]; + tensor v_masked = mul(x = v_17, y = var_1064)[name = tensor("v_masked")]; + tensor var_1247 = mul(x = prev_kv, y = sqrt_s0)[name = tensor("op_1247")]; + tensor var_1249_transpose_x_1 = const()[name = tensor("op_1249_transpose_x_1"), val = tensor(true)]; + tensor var_1249_transpose_y_1 = const()[name = tensor("op_1249_transpose_y_1"), val = tensor(false)]; + tensor var_1249 = matmul(transpose_x = var_1249_transpose_x_1, transpose_y = var_1249_transpose_y_1, x = k_17, y = v_masked)[name = tensor("op_1249")]; + tensor new_kv_unnorm = add(x = var_1247, y = var_1249)[name = tensor("new_kv_unnorm")]; + tensor new_scale = add(x = prev_scale, y = var_1072)[name = tensor("new_scale")]; tensor const_33 = const()[name = tensor("const_33"), val = tensor(0x1.fffffep+127)]; - tensor clip_4 = clip(alpha = var_926, beta = const_33, x = new_scale)[name = tensor("clip_4")]; + tensor clip_4 = clip(alpha = var_29, beta = const_33, x = new_scale)[name = tensor("clip_4")]; tensor sqrt_new_scale = sqrt(x = clip_4)[name = tensor("sqrt_new_scale")]; tensor nkv = real_div(x = new_kv_unnorm, y = sqrt_new_scale)[name = tensor("nkv")]; - tensor var_1255_perm_0 = const()[name = tensor("op_1255_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_1258_perm_0 = const()[name = tensor("op_1258_perm_0"), val = tensor([0, 2, 1, 3])]; tensor out_33_axes_0 = const()[name = tensor("out_33_axes_0"), val = tensor([-1])]; - tensor var_1255 = transpose(perm = var_1255_perm_0, x = out_31)[name = tensor("transpose_11")]; - tensor out_33 = layer_norm(axes = out_33_axes_0, epsilon = var_923, x = var_1255)[name = tensor("out_33")]; - tensor var_1259 = const()[name = tensor("op_1259"), val = tensor([9, 1, 256])]; - tensor out = reshape(shape = var_1259, x = out_33)[name = tensor("out")]; - tensor var_1261 = silu(x = input_187)[name = tensor("op_1261")]; - tensor input_189 = mul(x = var_1261, y = out)[name = tensor("input_189")]; - tensor ret_out = linear(bias = decoder_out_proj_1_bias, weight = decoder_out_proj_1_weight, x = input_189)[name = tensor("linear_51")]; - tensor input_191 = add(x = x, y = ret_out)[name = tensor("input_191")]; + tensor var_1258 = transpose(perm = var_1258_perm_0, x = out_31)[name = tensor("transpose_11")]; + tensor out_33 = layer_norm(axes = out_33_axes_0, epsilon = var_44, x = var_1258)[name = tensor("out_33")]; + tensor var_1262 = const()[name = tensor("op_1262"), val = tensor([9, 1, 256])]; + tensor out = reshape(shape = var_1262, x = out_33)[name = tensor("out")]; + tensor var_1264 = silu(x = input_189)[name = tensor("op_1264")]; + tensor input_191 = mul(x = var_1264, y = out)[name = tensor("input_191")]; + tensor ret_out = linear(bias = inner_decoder_out_proj_1_bias, weight = inner_decoder_out_proj_1_weight, x = input_191)[name = tensor("linear_51")]; + tensor input_193 = add(x = x, y = ret_out)[name = tensor("input_193")]; tensor xt_5_axes_0 = const()[name = tensor("xt_5_axes_0"), val = tensor([-1])]; - tensor xt_5 = layer_norm(axes = xt_5_axes_0, beta = decoder_norm11_1_bias, epsilon = var_921, gamma = decoder_norm11_1_weight, x = input_191)[name = tensor("xt_5")]; - tensor var_1271 = const()[name = tensor("op_1271"), val = tensor([1, 9, 1, 256])]; - tensor var_1272 = reshape(shape = var_1271, x = xt_5)[name = tensor("op_1272")]; - tensor var_1273_perm_0 = const()[name = tensor("op_1273_perm_0"), val = tensor([0, 2, 1, 3])]; - tensor var_1276 = const()[name = tensor("op_1276"), val = tensor([1, 9, 256])]; - tensor var_1273 = transpose(perm = var_1273_perm_0, x = var_1272)[name = tensor("transpose_10")]; - tensor query_5 = reshape(shape = var_1276, x = var_1273)[name = tensor("query_5")]; + tensor xt_5 = layer_norm(axes = xt_5_axes_0, beta = inner_decoder_norm11_1_bias, epsilon = var_36, gamma = inner_decoder_norm11_1_weight, x = input_193)[name = tensor("xt_5")]; + tensor var_1274 = const()[name = tensor("op_1274"), val = tensor([1, 9, 1, 256])]; + tensor var_1275 = reshape(shape = var_1274, x = xt_5)[name = tensor("op_1275")]; + tensor var_1276_perm_0 = const()[name = tensor("op_1276_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_1279 = const()[name = tensor("op_1279"), val = tensor([1, 9, 256])]; + tensor var_1276 = transpose(perm = var_1276_perm_0, x = var_1275)[name = tensor("transpose_10")]; + tensor query_5 = reshape(shape = var_1279, x = var_1276)[name = tensor("query_5")]; tensor query_perm_0 = const()[name = tensor("query_perm_0"), val = tensor([1, 0, 2])]; tensor query = transpose(perm = query_perm_0, x = query_5)[name = tensor("transpose_9")]; - tensor var_1299 = linear(bias = decoder_self_attn2_1_in_proj_bias, weight = decoder_self_attn2_1_in_proj_weight, x = query)[name = tensor("linear_52")]; + tensor var_1302 = linear(bias = inner_decoder_self_attn2_1_in_proj_bias, weight = inner_decoder_self_attn2_1_in_proj_weight, x = query)[name = tensor("linear_52")]; tensor concat_2 = const()[name = tensor("concat_2"), val = tensor([9, 1, 3, 256])]; - tensor var_1301 = reshape(shape = concat_2, x = var_1299)[name = tensor("op_1301")]; - tensor var_1302_axes_0 = const()[name = tensor("op_1302_axes_0"), val = tensor([0])]; - tensor var_1302 = expand_dims(axes = var_1302_axes_0, x = var_1301)[name = tensor("op_1302")]; - tensor var_1303_perm_0 = const()[name = tensor("op_1303_perm_0"), val = tensor([-2, 1, 2, 0, 4])]; - tensor var_1304_axes_0 = const()[name = tensor("op_1304_axes_0"), val = tensor([-2])]; - tensor var_1303 = transpose(perm = var_1303_perm_0, x = var_1302)[name = tensor("transpose_8")]; - tensor var_1304 = squeeze(axes = var_1304_axes_0, x = var_1303)[name = tensor("op_1304")]; + tensor var_1304 = reshape(shape = concat_2, x = var_1302)[name = tensor("op_1304")]; + tensor var_1305_axes_0 = const()[name = tensor("op_1305_axes_0"), val = tensor([0])]; + tensor var_1305 = expand_dims(axes = var_1305_axes_0, x = var_1304)[name = tensor("op_1305")]; + tensor var_1306_perm_0 = const()[name = tensor("op_1306_perm_0"), val = tensor([-2, 1, 2, 0, 4])]; + tensor var_1307_axes_0 = const()[name = tensor("op_1307_axes_0"), val = tensor([-2])]; + tensor var_1306 = transpose(perm = var_1306_perm_0, x = var_1305)[name = tensor("transpose_8")]; + tensor var_1307 = squeeze(axes = var_1307_axes_0, x = var_1306)[name = tensor("op_1307")]; tensor q_19_begin_0 = const()[name = tensor("q_19_begin_0"), val = tensor([0, 0, 0, 0])]; tensor q_19_end_0 = const()[name = tensor("q_19_end_0"), val = tensor([1, 9, 1, 256])]; tensor q_19_end_mask_0 = const()[name = tensor("q_19_end_mask_0"), val = tensor([false, true, true, true])]; tensor q_19_squeeze_mask_0 = const()[name = tensor("q_19_squeeze_mask_0"), val = tensor([true, false, false, false])]; - tensor q_19 = slice_by_index(begin = q_19_begin_0, end = q_19_end_0, end_mask = q_19_end_mask_0, squeeze_mask = q_19_squeeze_mask_0, x = var_1304)[name = tensor("q_19")]; + tensor q_19 = slice_by_index(begin = q_19_begin_0, end = q_19_end_0, end_mask = q_19_end_mask_0, squeeze_mask = q_19_squeeze_mask_0, x = var_1307)[name = tensor("q_19")]; tensor k_19_begin_0 = const()[name = tensor("k_19_begin_0"), val = tensor([1, 0, 0, 0])]; tensor k_19_end_0 = const()[name = tensor("k_19_end_0"), val = tensor([2, 9, 1, 256])]; tensor k_19_end_mask_0 = const()[name = tensor("k_19_end_mask_0"), val = tensor([false, true, true, true])]; tensor k_19_squeeze_mask_0 = const()[name = tensor("k_19_squeeze_mask_0"), val = tensor([true, false, false, false])]; - tensor k_19 = slice_by_index(begin = k_19_begin_0, end = k_19_end_0, end_mask = k_19_end_mask_0, squeeze_mask = k_19_squeeze_mask_0, x = var_1304)[name = tensor("k_19")]; + tensor k_19 = slice_by_index(begin = k_19_begin_0, end = k_19_end_0, end_mask = k_19_end_mask_0, squeeze_mask = k_19_squeeze_mask_0, x = var_1307)[name = tensor("k_19")]; tensor v_19_begin_0 = const()[name = tensor("v_19_begin_0"), val = tensor([2, 0, 0, 0])]; tensor v_19_end_0 = const()[name = tensor("v_19_end_0"), val = tensor([3, 9, 1, 256])]; tensor v_19_end_mask_0 = const()[name = tensor("v_19_end_mask_0"), val = tensor([false, true, true, true])]; tensor v_19_squeeze_mask_0 = const()[name = tensor("v_19_squeeze_mask_0"), val = tensor([true, false, false, false])]; - tensor v_19 = slice_by_index(begin = v_19_begin_0, end = v_19_end_0, end_mask = v_19_end_mask_0, squeeze_mask = v_19_squeeze_mask_0, x = var_1304)[name = tensor("v_19")]; - tensor var_1312 = const()[name = tensor("op_1312"), val = tensor([9, 4, 64])]; - tensor var_1313 = reshape(shape = var_1312, x = q_19)[name = tensor("op_1313")]; + tensor v_19 = slice_by_index(begin = v_19_begin_0, end = v_19_end_0, end_mask = v_19_end_mask_0, squeeze_mask = v_19_squeeze_mask_0, x = var_1307)[name = tensor("v_19")]; + tensor var_1315 = const()[name = tensor("op_1315"), val = tensor([9, 4, 64])]; + tensor var_1316 = reshape(shape = var_1315, x = q_19)[name = tensor("op_1316")]; tensor q_21_perm_0 = const()[name = tensor("q_21_perm_0"), val = tensor([1, 0, 2])]; - tensor var_1319 = const()[name = tensor("op_1319"), val = tensor([9, 4, 64])]; - tensor var_1320 = reshape(shape = var_1319, x = k_19)[name = tensor("op_1320")]; + tensor var_1322 = const()[name = tensor("op_1322"), val = tensor([9, 4, 64])]; + tensor var_1323 = reshape(shape = var_1322, x = k_19)[name = tensor("op_1323")]; tensor k_21_perm_0 = const()[name = tensor("k_21_perm_0"), val = tensor([1, 0, 2])]; - tensor var_1326 = const()[name = tensor("op_1326"), val = tensor([9, 4, 64])]; - tensor var_1327 = reshape(shape = var_1326, x = v_19)[name = tensor("op_1327")]; + tensor var_1329 = const()[name = tensor("op_1329"), val = tensor([9, 4, 64])]; + tensor var_1330 = reshape(shape = var_1329, x = v_19)[name = tensor("op_1330")]; tensor v_21_perm_0 = const()[name = tensor("v_21_perm_0"), val = tensor([1, 0, 2])]; - tensor var_1330 = const()[name = tensor("op_1330"), val = tensor([1, 4, 9, 64])]; - tensor q_21 = transpose(perm = q_21_perm_0, x = var_1313)[name = tensor("transpose_7")]; - tensor q = reshape(shape = var_1330, x = q_21)[name = tensor("q")]; - tensor var_1332 = const()[name = tensor("op_1332"), val = tensor([1, 4, 9, 64])]; - tensor k_21 = transpose(perm = k_21_perm_0, x = var_1320)[name = tensor("transpose_6")]; - tensor k = reshape(shape = var_1332, x = k_21)[name = tensor("k")]; - tensor var_1334 = const()[name = tensor("op_1334"), val = tensor([1, 4, 9, 64])]; - tensor v_21 = transpose(perm = v_21_perm_0, x = var_1327)[name = tensor("transpose_5")]; - tensor v = reshape(shape = var_1334, x = v_21)[name = tensor("v")]; + tensor var_1333 = const()[name = tensor("op_1333"), val = tensor([1, 4, 9, 64])]; + tensor q_21 = transpose(perm = q_21_perm_0, x = var_1316)[name = tensor("transpose_7")]; + tensor q = reshape(shape = var_1333, x = q_21)[name = tensor("q")]; + tensor var_1335 = const()[name = tensor("op_1335"), val = tensor([1, 4, 9, 64])]; + tensor k_21 = transpose(perm = k_21_perm_0, x = var_1323)[name = tensor("transpose_6")]; + tensor k = reshape(shape = var_1335, x = k_21)[name = tensor("k")]; + tensor var_1337 = const()[name = tensor("op_1337"), val = tensor([1, 4, 9, 64])]; + tensor v_21 = transpose(perm = v_21_perm_0, x = var_1330)[name = tensor("transpose_5")]; + tensor v = reshape(shape = var_1337, x = v_21)[name = tensor("v")]; tensor mul_3_y_0 = const()[name = tensor("mul_3_y_0"), val = tensor(0x1p-3)]; tensor mul_3 = mul(x = q, y = mul_3_y_0)[name = tensor("mul_3")]; tensor matmul_1_transpose_y_0 = const()[name = tensor("matmul_1_transpose_y_0"), val = tensor(true)]; @@ -1139,34 +1138,34 @@ program(1.0) tensor attn_output_9_transpose_x_0 = const()[name = tensor("attn_output_9_transpose_x_0"), val = tensor(false)]; tensor attn_output_9_transpose_y_0 = const()[name = tensor("attn_output_9_transpose_y_0"), val = tensor(false)]; tensor attn_output_9 = matmul(transpose_x = attn_output_9_transpose_x_0, transpose_y = attn_output_9_transpose_y_0, x = softmax_1, y = v)[name = tensor("attn_output_9")]; - tensor var_1337 = const()[name = tensor("op_1337"), val = tensor([2, 0, 1, 3])]; - tensor var_1342 = const()[name = tensor("op_1342"), val = tensor([9, 256])]; - tensor var_1338 = transpose(perm = var_1337, x = attn_output_9)[name = tensor("transpose_4")]; - tensor attn_output_11 = reshape(shape = var_1342, x = var_1338)[name = tensor("attn_output_11")]; - tensor attn_output_13 = linear(bias = decoder_self_attn2_1_out_proj_bias, weight = decoder_self_attn2_1_out_proj_weight, x = attn_output_11)[name = tensor("linear_53")]; - tensor var_1346 = const()[name = tensor("op_1346"), val = tensor([9, 1, 256])]; - tensor attn_output = reshape(shape = var_1346, x = attn_output_13)[name = tensor("attn_output")]; + tensor var_1340 = const()[name = tensor("op_1340"), val = tensor([2, 0, 1, 3])]; + tensor var_1345 = const()[name = tensor("op_1345"), val = tensor([9, 256])]; + tensor var_1341 = transpose(perm = var_1340, x = attn_output_9)[name = tensor("transpose_4")]; + tensor attn_output_11 = reshape(shape = var_1345, x = var_1341)[name = tensor("attn_output_11")]; + tensor attn_output_13 = linear(bias = inner_decoder_self_attn2_1_out_proj_bias, weight = inner_decoder_self_attn2_1_out_proj_weight, x = attn_output_11)[name = tensor("linear_53")]; + tensor var_1349 = const()[name = tensor("op_1349"), val = tensor([9, 1, 256])]; + tensor attn_output = reshape(shape = var_1349, x = attn_output_13)[name = tensor("attn_output")]; tensor sa2_out_perm_0 = const()[name = tensor("sa2_out_perm_0"), val = tensor([1, 0, 2])]; tensor sa2_out = transpose(perm = sa2_out_perm_0, x = attn_output)[name = tensor("transpose_3")]; - tensor input_193 = add(x = query_5, y = sa2_out)[name = tensor("input_193")]; - tensor input_195_axes_0 = const()[name = tensor("input_195_axes_0"), val = tensor([-1])]; - tensor input_195 = layer_norm(axes = input_195_axes_0, beta = decoder_norm21_1_bias, epsilon = var_921, gamma = decoder_norm21_1_weight, x = input_193)[name = tensor("input_195")]; - tensor input_197 = linear(bias = decoder_linear1_1_bias, weight = decoder_linear1_1_weight, x = input_195)[name = tensor("linear_54")]; - tensor input_199 = relu(x = input_197)[name = tensor("input_199")]; - tensor ffn_out = linear(bias = decoder_linear2_1_bias, weight = decoder_linear2_1_weight, x = input_199)[name = tensor("linear_55")]; - tensor input_201 = add(x = input_195, y = ffn_out)[name = tensor("input_201")]; + tensor input_195 = add(x = query_5, y = sa2_out)[name = tensor("input_195")]; + tensor input_197_axes_0 = const()[name = tensor("input_197_axes_0"), val = tensor([-1])]; + tensor input_197 = layer_norm(axes = input_197_axes_0, beta = inner_decoder_norm21_1_bias, epsilon = var_36, gamma = inner_decoder_norm21_1_weight, x = input_195)[name = tensor("input_197")]; + tensor input_199 = linear(bias = inner_decoder_linear1_1_bias, weight = inner_decoder_linear1_1_weight, x = input_197)[name = tensor("linear_54")]; + tensor input_201 = relu(x = input_199)[name = tensor("input_201")]; + tensor ffn_out = linear(bias = inner_decoder_linear2_1_bias, weight = inner_decoder_linear2_1_weight, x = input_201)[name = tensor("linear_55")]; + tensor input_203 = add(x = input_197, y = ffn_out)[name = tensor("input_203")]; tensor xt_axes_0 = const()[name = tensor("xt_axes_0"), val = tensor([-1])]; - tensor xt = layer_norm(axes = xt_axes_0, beta = decoder_norm22_1_bias, epsilon = var_921, gamma = decoder_norm22_1_weight, x = input_201)[name = tensor("xt")]; - tensor var_1366 = const()[name = tensor("op_1366"), val = tensor([1, 1, 9, 256])]; - tensor input = reshape(shape = var_1366, x = xt)[name = tensor("input")]; - tensor var_1368 = const()[name = tensor("op_1368"), val = tensor([-1])]; - tensor var_1369 = reduce_l2_norm(axes = var_1368, keep_dims = var_924, x = input)[name = tensor("op_1369")]; + tensor xt = layer_norm(axes = xt_axes_0, beta = inner_decoder_norm22_1_bias, epsilon = var_36, gamma = inner_decoder_norm22_1_weight, x = input_203)[name = tensor("xt")]; + tensor var_1369 = const()[name = tensor("op_1369"), val = tensor([1, 1, 9, 256])]; + tensor input = reshape(shape = var_1369, x = xt)[name = tensor("input")]; + tensor var_1371 = const()[name = tensor("op_1371"), val = tensor([-1])]; + tensor var_1372 = reduce_l2_norm(axes = var_1371, keep_dims = var_35, x = input)[name = tensor("op_1372")]; tensor const_42 = const()[name = tensor("const_42"), val = tensor(0x1.fffffep+127)]; - tensor clip_5 = clip(alpha = var_916, beta = const_42, x = var_1369)[name = tensor("clip_5")]; - tensor var_1371 = real_div(x = input, y = clip_5)[name = tensor("op_1371")]; + tensor clip_5 = clip(alpha = var_49, beta = const_42, x = var_1372)[name = tensor("clip_5")]; + tensor var_1374 = real_div(x = input, y = clip_5)[name = tensor("op_1374")]; tensor transpose_1_perm_0 = const()[name = tensor("transpose_1_perm_0"), val = tensor([0, 1, 3, 2])]; tensor concat_7 = const()[name = tensor("concat_7"), val = tensor([1, 256, 9])]; - tensor transpose_1 = transpose(perm = transpose_1_perm_0, x = var_1371)[name = tensor("transpose_2")]; + tensor transpose_1 = transpose(perm = transpose_1_perm_0, x = var_1374)[name = tensor("transpose_2")]; tensor reshape_1 = reshape(shape = concat_7, x = transpose_1)[name = tensor("reshape_1")]; tensor matmul_2_transpose_x_0 = const()[name = tensor("matmul_2_transpose_x_0"), val = tensor(false)]; tensor matmul_2_transpose_y_0 = const()[name = tensor("matmul_2_transpose_y_0"), val = tensor(false)]; @@ -1175,10 +1174,10 @@ program(1.0) tensor output_end_0 = const()[name = tensor("output_end_0"), val = tensor([1, 1, 8])]; tensor output_end_mask_0 = const()[name = tensor("output_end_mask_0"), val = tensor([true, true, false])]; tensor output = slice_by_index(begin = output_begin_0, end = output_end_0, end_mask = output_end_mask_0, x = matmul_2)[name = tensor("output")]; - tensor probs = sigmoid(x = output)[name = tensor("op_1375")]; - tensor var_1377_axis_0 = const()[name = tensor("op_1377_axis_0"), val = tensor(0)]; - tensor dec_kv_new = stack(axis = var_1377_axis_0, values = (var_1073, nkv))[name = tensor("op_1377")]; - tensor var_1379_axis_0 = const()[name = tensor("op_1379_axis_0"), val = tensor(0)]; - tensor dec_scale_new = stack(axis = var_1379_axis_0, values = (new_scale_9, new_scale))[name = tensor("op_1379")]; + tensor probs = sigmoid(x = output)[name = tensor("op_1378")]; + tensor var_1380_axis_0 = const()[name = tensor("op_1380_axis_0"), val = tensor(0)]; + tensor dec_kv_new = stack(axis = var_1380_axis_0, values = (var_1076, nkv))[name = tensor("op_1380")]; + tensor var_1382_axis_0 = const()[name = tensor("op_1382_axis_0"), val = tensor(0)]; + tensor dec_scale_new = stack(axis = var_1382_axis_0, values = (new_scale_9, new_scale))[name = tensor("op_1382")]; } -> (probs, enc_kv_new, enc_scale_new, enc_conv_cache_new, cnn_window_new, dec_kv_new, dec_scale_new); } \ No newline at end of file diff --git a/optimized/ch/100ms/ls_eend_ch_100ms.mlpackage/Data/com.apple.CoreML/model.mlmodel b/optimized/ch/100ms/ls_eend_ch_100ms.mlpackage/Data/com.apple.CoreML/model.mlmodel index 42c372e11d6d85d4e11b133904269e202b57e146..f9e0a2c3d6a9d275fe99a5b6e10ee2a63f955685 100644 --- a/optimized/ch/100ms/ls_eend_ch_100ms.mlpackage/Data/com.apple.CoreML/model.mlmodel +++ b/optimized/ch/100ms/ls_eend_ch_100ms.mlpackage/Data/com.apple.CoreML/model.mlmodel @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:a15eadb3397bff9c66d9266e9a495a92e7913dd25b985447b5efd0ebd212d182 -size 171357 +oid sha256:d27af30cb86aac3d4db11e7ed026c089dec4ccef4d0c7e12f0e7d59bf0b86092 +size 175275 diff --git a/optimized/ch/100ms/ls_eend_ch_100ms.mlpackage/Manifest.json b/optimized/ch/100ms/ls_eend_ch_100ms.mlpackage/Manifest.json index 6438d33c88f42770e5903b4cf8b6b7fc7ffd893a..9fb9ade7c21b566c958ea9718a94e56ccf55686e 100644 --- a/optimized/ch/100ms/ls_eend_ch_100ms.mlpackage/Manifest.json +++ b/optimized/ch/100ms/ls_eend_ch_100ms.mlpackage/Manifest.json @@ -1,18 +1,18 @@ { "fileFormatVersion": "1.0.0", "itemInfoEntries": { - "15BB7841-F157-4C26-9962-7E006B1F650B": { + "00C8998E-3FCC-440D-86B1-4A910C90D3C0": { "author": "com.apple.CoreML", "description": "CoreML Model Specification", "name": "model.mlmodel", "path": "com.apple.CoreML/model.mlmodel" }, - "69383227-3EDF-4375-8C74-435BD4941612": { + "470AB283-D36B-4390-A0B0-CABAEA84E932": { "author": "com.apple.CoreML", "description": "CoreML Model Weights", "name": "weights", "path": "com.apple.CoreML/weights" } }, - "rootModelIdentifier": "15BB7841-F157-4C26-9962-7E006B1F650B" + "rootModelIdentifier": "00C8998E-3FCC-440D-86B1-4A910C90D3C0" } diff --git a/optimized/ch/200ms/ls_eend_ch_200ms.mlmodelc/analytics/coremldata.bin b/optimized/ch/200ms/ls_eend_ch_200ms.mlmodelc/analytics/coremldata.bin index a37607ed47f4304af3eb57d2df22a9ad4ac88e3e..8f1abe174990c18e0bd327ab363aacda857d642b 100644 --- a/optimized/ch/200ms/ls_eend_ch_200ms.mlmodelc/analytics/coremldata.bin +++ b/optimized/ch/200ms/ls_eend_ch_200ms.mlmodelc/analytics/coremldata.bin @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:e72fc8970f10d313ad089b47fe4fcc1e390363db50c413f17b9fb0d756a8cea1 +oid sha256:b1c408f22c25e1aa2254bb0ea9d086db795029b10e465e8855e0ddb009393d89 size 243 diff --git a/optimized/ch/200ms/ls_eend_ch_200ms.mlmodelc/coremldata.bin b/optimized/ch/200ms/ls_eend_ch_200ms.mlmodelc/coremldata.bin index fd51deca07943f75691c3895fec309e63919fa02..ce1acf08093f90402bc6f90e2d7d5ed0391056e7 100644 --- a/optimized/ch/200ms/ls_eend_ch_200ms.mlmodelc/coremldata.bin +++ b/optimized/ch/200ms/ls_eend_ch_200ms.mlmodelc/coremldata.bin @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:230d025ab4e17869d9301adf650df4fa921472c5d5582a9f365e7d70f3f13585 -size 1301 +oid sha256:25182b8ef75f90ca7ceb64114d45ea7a19e8500cb4ab6b7ce7250f8ad694d7ea +size 1404 diff --git a/optimized/ch/200ms/ls_eend_ch_200ms.mlmodelc/metadata.json b/optimized/ch/200ms/ls_eend_ch_200ms.mlmodelc/metadata.json index ffa39d702bedf6118ed46a822ffe5aee24cfcb30..61203cd26f46e815e985ea745a6b63d352292e3e 100644 --- a/optimized/ch/200ms/ls_eend_ch_200ms.mlmodelc/metadata.json +++ b/optimized/ch/200ms/ls_eend_ch_200ms.mlmodelc/metadata.json @@ -1,7 +1,7 @@ [ { "metadataOutputVersion" : "3.0", - "shortDescription" : "LS-EEND CALLHOME streaming diarizer (pipeline, T=2, max_speakers=7)", + "shortDescription" : "LS-EEND CALLHOME streaming diarizer (pipeline, T=2, max_speakers=7, layout=raw_mel, cu=all)", "outputSchema" : [ { "hasShapeFlexibility" : "0", @@ -81,7 +81,7 @@ "specificationVersion" : 8, "mlProgramOperationTypeHistogram" : { "Ios16.reduceL2Norm" : 2, - "Ios17.reshape" : 68, + "Ios17.reshape" : 69, "Ios16.softmax" : 2, "Ios17.matmul" : 29, "Ios17.transpose" : 57, @@ -89,7 +89,7 @@ "Ios17.expandDims" : 3, "Ios17.add" : 46, "Ios16.sigmoid" : 5, - "Ios17.sliceByIndex" : 48, + "Ios17.sliceByIndex" : 50, "Tile" : 1, "Ios16.reduceSum" : 1, "Ios17.squeeze" : 2, @@ -101,7 +101,7 @@ "Ios16.silu" : 18, "Ios17.realDiv" : 20, "Ios17.linear" : 56, - "Stack" : 5, + "Stack" : 6, "Ios17.concat" : 14, "Ios16.relu" : 2, "Ios16.cumsum" : 1, @@ -128,9 +128,9 @@ "hasShapeFlexibility" : "0", "isOptional" : "0", "dataType" : "Float32", - "formattedType" : "MultiArray (Float32 1 × 2 × 345)", + "formattedType" : "MultiArray (Float32 1 × 25 × 23)", "shortDescription" : "", - "shape" : "[1, 2, 345]", + "shape" : "[1, 25, 23]", "name" : "features", "type" : "MultiArray" }, @@ -206,8 +206,8 @@ } ], "userDefinedMetadata" : { - "com.github.apple.coremltools.conversion_date" : "2026-04-16", - "config" : "{\"model_name\": \"ch\", \"model_label\": \"CALLHOME\", \"variant\": \"pipeline\", \"chunk_size\": 2, \"step_duration_ms\": 200, \"frame_duration_ms\": 100, \"frame_duration_seconds\": 0.1, \"max_speakers\": 7, \"max_nspks\": 9, \"n_units\": 256, \"n_heads\": 4, \"enc_n_layers\": 4, \"dec_n_layers\": 2, \"conv_kernel_size\": 16, \"conv_delay\": 9, \"sample_rate\": 8000, \"win_length\": 200, \"hop_length\": 80, \"n_mels\": 23, \"context_size\": 7, \"subsampling\": 10, \"feat_type\": \"logmel23_cummn\", \"pure_roll\": true}", + "com.github.apple.coremltools.conversion_date" : "2026-04-18", + "config" : "{\"model_name\": \"ch\", \"model_label\": \"CALLHOME\", \"variant\": \"pipeline\", \"chunk_size\": 2, \"step_duration_ms\": 200, \"frame_duration_ms\": 100, \"frame_duration_seconds\": 0.1, \"max_speakers\": 7, \"max_nspks\": 9, \"n_units\": 256, \"n_heads\": 4, \"enc_n_layers\": 4, \"dec_n_layers\": 2, \"conv_kernel_size\": 16, \"conv_delay\": 9, \"sample_rate\": 8000, \"win_length\": 200, \"hop_length\": 80, \"n_mels\": 23, \"context_size\": 7, \"subsampling\": 10, \"feat_type\": \"logmel23_cummn\", \"pure_roll\": true, \"input_layout\": \"raw_mel\", \"compute_units_export\": \"all\", \"raw_mel_length\": 25}", "com.github.apple.coremltools.source" : "torch==2.6.0", "com.github.apple.coremltools.version" : "9.0", "com.github.apple.coremltools.source_dialect" : "TorchScript" diff --git a/optimized/ch/200ms/ls_eend_ch_200ms.mlmodelc/model.mil b/optimized/ch/200ms/ls_eend_ch_200ms.mlmodelc/model.mil index 801b456ed38b11b675323d063b50f9d105936618..3e74603b0651724dff587b0ac9da39a164d753ed 100644 --- a/optimized/ch/200ms/ls_eend_ch_200ms.mlmodelc/model.mil +++ b/optimized/ch/200ms/ls_eend_ch_200ms.mlmodelc/model.mil @@ -1,234 +1,248 @@ program(1.0) [buildInfo = dict, tensor>({{"coremlc-component-MIL", "3510.2.1"}, {"coremlc-version", "3500.32.1"}, {"coremltools-component-torch", "2.6.0"}, {"coremltools-source-dialect", "TorchScript"}, {"coremltools-version", "9.0"}})] { - func main(tensor cnn_window, tensor dec_kv, tensor dec_scale, tensor enc_conv_cache, tensor enc_kv, tensor enc_scale, tensor features, tensor valid_mask) { - tensor encoder_cnn_bias = const()[name = tensor("encoder_cnn_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(64)))]; - tensor encoder_cnn_weight = const()[name = tensor("encoder_cnn_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1152)))]; - tensor encoder__causal_mask = const()[name = tensor("encoder__causal_mask"), val = tensor([[0x1p+0, 0x0p+0], [0x1p+0, 0x1p+0]])]; - tensor encoder__t_index = const()[name = tensor("encoder__t_index"), val = tensor([0x1p+0, 0x1p+1])]; - tensor encoder_input_projection_linear_bias = const()[name = tensor("encoder_input_projection_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4981952)))]; - tensor encoder_input_projection_linear_weight = const()[name = tensor("encoder_input_projection_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4983040)))]; - tensor encoder_pre_layer_norm_bias = const()[name = tensor("encoder_pre_layer_norm_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(5336384)))]; - tensor encoder_pre_layer_norm_weight = const()[name = tensor("encoder_pre_layer_norm_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(5337472)))]; - tensor encoder_ffn1_0_module_sequential_0_bias = const()[name = tensor("encoder_ffn1_0_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(5338560)))]; - tensor encoder_ffn1_0_module_sequential_0_weight = const()[name = tensor("encoder_ffn1_0_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(5339648)))]; - tensor encoder_ffn1_0_module_sequential_1_linear_bias = const()[name = tensor("encoder_ffn1_0_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(5340736)))]; - tensor encoder_ffn1_0_module_sequential_1_linear_weight = const()[name = tensor("encoder_ffn1_0_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(5344896)))]; - tensor encoder_ffn1_0_module_sequential_4_linear_bias = const()[name = tensor("encoder_ffn1_0_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(6393536)))]; - tensor encoder_ffn1_0_module_sequential_4_linear_weight = const()[name = tensor("encoder_ffn1_0_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(6394624)))]; - tensor encoder_ret_lns_0_bias = const()[name = tensor("encoder_ret_lns_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7443264)))]; - tensor encoder_ret_lns_0_weight = const()[name = tensor("encoder_ret_lns_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7444352)))]; - tensor encoder_q_proj_0_bias = const()[name = tensor("encoder_q_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7445440)))]; - tensor encoder_q_proj_0_weight = const()[name = tensor("encoder_q_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7446528)))]; - tensor encoder_k_proj_0_bias = const()[name = tensor("encoder_k_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7708736)))]; - tensor encoder_k_proj_0_weight = const()[name = tensor("encoder_k_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7709824)))]; - tensor encoder_v_proj_0_bias = const()[name = tensor("encoder_v_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7972032)))]; - tensor encoder_v_proj_0_weight = const()[name = tensor("encoder_v_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7973120)))]; - tensor encoder_g_proj_0_bias = const()[name = tensor("encoder_g_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8235328)))]; - tensor encoder_g_proj_0_weight = const()[name = tensor("encoder_g_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8236416)))]; - tensor encoder_out_proj_0_bias = const()[name = tensor("encoder_out_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8498624)))]; - tensor encoder_out_proj_0_weight = const()[name = tensor("encoder_out_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8499712)))]; - tensor encoder_conv_module_0_sequential_0_bias = const()[name = tensor("encoder_conv_module_0_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8761920)))]; - tensor encoder_conv_module_0_sequential_0_weight = const()[name = tensor("encoder_conv_module_0_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8763008)))]; - tensor encoder_conv_module_0_sequential_2_conv_bias = const()[name = tensor("encoder_conv_module_0_sequential_2_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8764096)))]; - tensor encoder_conv_module_0_sequential_2_conv_weight = const()[name = tensor("encoder_conv_module_0_sequential_2_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8766208)))]; - tensor encoder_conv_module_0_sequential_4_conv_weight = const()[name = tensor("encoder_conv_module_0_sequential_4_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9290560)))]; - tensor encoder_conv_module_0_sequential_5_running_var = const()[name = tensor("encoder_conv_module_0_sequential_5_running_var"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9307008)))]; - tensor encoder_conv_module_0_sequential_5_running_mean = const()[name = tensor("encoder_conv_module_0_sequential_5_running_mean"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9308096)))]; - tensor encoder_conv_module_0_sequential_5_bias = const()[name = tensor("encoder_conv_module_0_sequential_5_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9309184)))]; - tensor encoder_conv_module_0_sequential_5_weight = const()[name = tensor("encoder_conv_module_0_sequential_5_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9310272)))]; - tensor encoder_conv_module_0_sequential_7_conv_bias = const()[name = tensor("encoder_conv_module_0_sequential_7_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9311360)))]; - tensor encoder_conv_module_0_sequential_7_conv_weight = const()[name = tensor("encoder_conv_module_0_sequential_7_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9312448)))]; - tensor encoder_ffn2_0_module_sequential_0_bias = const()[name = tensor("encoder_ffn2_0_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9574656)))]; - tensor encoder_ffn2_0_module_sequential_0_weight = const()[name = tensor("encoder_ffn2_0_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9575744)))]; - tensor encoder_ffn2_0_module_sequential_1_linear_bias = const()[name = tensor("encoder_ffn2_0_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9576832)))]; - tensor encoder_ffn2_0_module_sequential_1_linear_weight = const()[name = tensor("encoder_ffn2_0_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9580992)))]; - tensor encoder_ffn2_0_module_sequential_4_linear_bias = const()[name = tensor("encoder_ffn2_0_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(10629632)))]; - tensor encoder_ffn2_0_module_sequential_4_linear_weight = const()[name = tensor("encoder_ffn2_0_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(10630720)))]; - tensor encoder_layer_norm_0_bias = const()[name = tensor("encoder_layer_norm_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11679360)))]; - tensor encoder_layer_norm_0_weight = const()[name = tensor("encoder_layer_norm_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11680448)))]; - tensor encoder_ffn1_1_module_sequential_0_bias = const()[name = tensor("encoder_ffn1_1_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11681536)))]; - tensor encoder_ffn1_1_module_sequential_0_weight = const()[name = tensor("encoder_ffn1_1_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11682624)))]; - tensor encoder_ffn1_1_module_sequential_1_linear_bias = const()[name = tensor("encoder_ffn1_1_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11683712)))]; - tensor encoder_ffn1_1_module_sequential_1_linear_weight = const()[name = tensor("encoder_ffn1_1_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11687872)))]; - tensor encoder_ffn1_1_module_sequential_4_linear_bias = const()[name = tensor("encoder_ffn1_1_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(12736512)))]; - tensor encoder_ffn1_1_module_sequential_4_linear_weight = const()[name = tensor("encoder_ffn1_1_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(12737600)))]; - tensor encoder_ret_lns_1_bias = const()[name = tensor("encoder_ret_lns_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(13786240)))]; - tensor encoder_ret_lns_1_weight = const()[name = tensor("encoder_ret_lns_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(13787328)))]; - tensor encoder_q_proj_1_bias = const()[name = tensor("encoder_q_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(13788416)))]; - tensor encoder_q_proj_1_weight = const()[name = tensor("encoder_q_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(13789504)))]; - tensor encoder_k_proj_1_bias = const()[name = tensor("encoder_k_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14051712)))]; - tensor encoder_k_proj_1_weight = const()[name = tensor("encoder_k_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14052800)))]; - tensor encoder_v_proj_1_bias = const()[name = tensor("encoder_v_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14315008)))]; - tensor encoder_v_proj_1_weight = const()[name = tensor("encoder_v_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14316096)))]; - tensor encoder_g_proj_1_bias = const()[name = tensor("encoder_g_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14578304)))]; - tensor encoder_g_proj_1_weight = const()[name = tensor("encoder_g_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14579392)))]; - tensor encoder_out_proj_1_bias = const()[name = tensor("encoder_out_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14841600)))]; - tensor encoder_out_proj_1_weight = const()[name = tensor("encoder_out_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14842688)))]; - tensor encoder_conv_module_1_sequential_0_bias = const()[name = tensor("encoder_conv_module_1_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15104896)))]; - tensor encoder_conv_module_1_sequential_0_weight = const()[name = tensor("encoder_conv_module_1_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15105984)))]; - tensor encoder_conv_module_1_sequential_2_conv_bias = const()[name = tensor("encoder_conv_module_1_sequential_2_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15107072)))]; - tensor encoder_conv_module_1_sequential_2_conv_weight = const()[name = tensor("encoder_conv_module_1_sequential_2_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15109184)))]; - tensor encoder_conv_module_1_sequential_4_conv_weight = const()[name = tensor("encoder_conv_module_1_sequential_4_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15633536)))]; - tensor encoder_conv_module_1_sequential_5_running_var = const()[name = tensor("encoder_conv_module_1_sequential_5_running_var"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15649984)))]; - tensor encoder_conv_module_1_sequential_5_running_mean = const()[name = tensor("encoder_conv_module_1_sequential_5_running_mean"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15651072)))]; - tensor encoder_conv_module_1_sequential_5_bias = const()[name = tensor("encoder_conv_module_1_sequential_5_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15652160)))]; - tensor encoder_conv_module_1_sequential_5_weight = const()[name = tensor("encoder_conv_module_1_sequential_5_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15653248)))]; - tensor encoder_conv_module_1_sequential_7_conv_bias = const()[name = tensor("encoder_conv_module_1_sequential_7_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15654336)))]; - tensor encoder_conv_module_1_sequential_7_conv_weight = const()[name = tensor("encoder_conv_module_1_sequential_7_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15655424)))]; - tensor encoder_ffn2_1_module_sequential_0_bias = const()[name = tensor("encoder_ffn2_1_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15917632)))]; - tensor encoder_ffn2_1_module_sequential_0_weight = const()[name = tensor("encoder_ffn2_1_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15918720)))]; - tensor encoder_ffn2_1_module_sequential_1_linear_bias = const()[name = tensor("encoder_ffn2_1_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15919808)))]; - tensor encoder_ffn2_1_module_sequential_1_linear_weight = const()[name = tensor("encoder_ffn2_1_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15923968)))]; - tensor encoder_ffn2_1_module_sequential_4_linear_bias = const()[name = tensor("encoder_ffn2_1_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(16972608)))]; - tensor encoder_ffn2_1_module_sequential_4_linear_weight = const()[name = tensor("encoder_ffn2_1_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(16973696)))]; - tensor encoder_layer_norm_1_bias = const()[name = tensor("encoder_layer_norm_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18022336)))]; - tensor encoder_layer_norm_1_weight = const()[name = tensor("encoder_layer_norm_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18023424)))]; - tensor encoder_ffn1_2_module_sequential_0_bias = const()[name = tensor("encoder_ffn1_2_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18024512)))]; - tensor encoder_ffn1_2_module_sequential_0_weight = const()[name = tensor("encoder_ffn1_2_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18025600)))]; - tensor encoder_ffn1_2_module_sequential_1_linear_bias = const()[name = tensor("encoder_ffn1_2_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18026688)))]; - tensor encoder_ffn1_2_module_sequential_1_linear_weight = const()[name = tensor("encoder_ffn1_2_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18030848)))]; - tensor encoder_ffn1_2_module_sequential_4_linear_bias = const()[name = tensor("encoder_ffn1_2_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(19079488)))]; - tensor encoder_ffn1_2_module_sequential_4_linear_weight = const()[name = tensor("encoder_ffn1_2_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(19080576)))]; - tensor encoder_ret_lns_2_bias = const()[name = tensor("encoder_ret_lns_2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20129216)))]; - tensor encoder_ret_lns_2_weight = const()[name = tensor("encoder_ret_lns_2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20130304)))]; - tensor encoder_q_proj_2_bias = const()[name = tensor("encoder_q_proj_2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20131392)))]; - tensor encoder_q_proj_2_weight = const()[name = tensor("encoder_q_proj_2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20132480)))]; - tensor encoder_k_proj_2_bias = const()[name = tensor("encoder_k_proj_2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20394688)))]; - tensor encoder_k_proj_2_weight = const()[name = tensor("encoder_k_proj_2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20395776)))]; - tensor encoder_v_proj_2_bias = const()[name = tensor("encoder_v_proj_2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20657984)))]; - tensor encoder_v_proj_2_weight = const()[name = tensor("encoder_v_proj_2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20659072)))]; - tensor encoder_g_proj_2_bias = const()[name = tensor("encoder_g_proj_2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20921280)))]; - tensor encoder_g_proj_2_weight = const()[name = tensor("encoder_g_proj_2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20922368)))]; - tensor encoder_out_proj_2_bias = const()[name = tensor("encoder_out_proj_2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21184576)))]; - tensor encoder_out_proj_2_weight = const()[name = tensor("encoder_out_proj_2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21185664)))]; - tensor encoder_conv_module_2_sequential_0_bias = const()[name = tensor("encoder_conv_module_2_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21447872)))]; - tensor encoder_conv_module_2_sequential_0_weight = const()[name = tensor("encoder_conv_module_2_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21448960)))]; - tensor encoder_conv_module_2_sequential_2_conv_bias = const()[name = tensor("encoder_conv_module_2_sequential_2_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21450048)))]; - tensor encoder_conv_module_2_sequential_2_conv_weight = const()[name = tensor("encoder_conv_module_2_sequential_2_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21452160)))]; - tensor encoder_conv_module_2_sequential_4_conv_weight = const()[name = tensor("encoder_conv_module_2_sequential_4_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21976512)))]; - tensor encoder_conv_module_2_sequential_5_running_var = const()[name = tensor("encoder_conv_module_2_sequential_5_running_var"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21992960)))]; - tensor encoder_conv_module_2_sequential_5_running_mean = const()[name = tensor("encoder_conv_module_2_sequential_5_running_mean"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21994048)))]; - tensor encoder_conv_module_2_sequential_5_bias = const()[name = tensor("encoder_conv_module_2_sequential_5_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21995136)))]; - tensor encoder_conv_module_2_sequential_5_weight = const()[name = tensor("encoder_conv_module_2_sequential_5_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21996224)))]; - tensor encoder_conv_module_2_sequential_7_conv_bias = const()[name = tensor("encoder_conv_module_2_sequential_7_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21997312)))]; - tensor encoder_conv_module_2_sequential_7_conv_weight = const()[name = tensor("encoder_conv_module_2_sequential_7_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21998400)))]; - tensor encoder_ffn2_2_module_sequential_0_bias = const()[name = tensor("encoder_ffn2_2_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(22260608)))]; - tensor encoder_ffn2_2_module_sequential_0_weight = const()[name = tensor("encoder_ffn2_2_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(22261696)))]; - tensor encoder_ffn2_2_module_sequential_1_linear_bias = const()[name = tensor("encoder_ffn2_2_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(22262784)))]; - tensor encoder_ffn2_2_module_sequential_1_linear_weight = const()[name = tensor("encoder_ffn2_2_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(22266944)))]; - tensor encoder_ffn2_2_module_sequential_4_linear_bias = const()[name = tensor("encoder_ffn2_2_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(23315584)))]; - tensor encoder_ffn2_2_module_sequential_4_linear_weight = const()[name = tensor("encoder_ffn2_2_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(23316672)))]; - tensor encoder_layer_norm_2_bias = const()[name = tensor("encoder_layer_norm_2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(24365312)))]; - tensor encoder_layer_norm_2_weight = const()[name = tensor("encoder_layer_norm_2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(24366400)))]; - tensor encoder_ffn1_3_module_sequential_0_bias = const()[name = tensor("encoder_ffn1_3_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(24367488)))]; - tensor encoder_ffn1_3_module_sequential_0_weight = const()[name = tensor("encoder_ffn1_3_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(24368576)))]; - tensor encoder_ffn1_3_module_sequential_1_linear_bias = const()[name = tensor("encoder_ffn1_3_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(24369664)))]; - tensor encoder_ffn1_3_module_sequential_1_linear_weight = const()[name = tensor("encoder_ffn1_3_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(24373824)))]; - tensor encoder_ffn1_3_module_sequential_4_linear_bias = const()[name = tensor("encoder_ffn1_3_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(25422464)))]; - tensor encoder_ffn1_3_module_sequential_4_linear_weight = const()[name = tensor("encoder_ffn1_3_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(25423552)))]; - tensor encoder_ret_lns_3_bias = const()[name = tensor("encoder_ret_lns_3_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(26472192)))]; - tensor encoder_ret_lns_3_weight = const()[name = tensor("encoder_ret_lns_3_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(26473280)))]; - tensor encoder_q_proj_3_bias = const()[name = tensor("encoder_q_proj_3_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(26474368)))]; - tensor encoder_q_proj_3_weight = const()[name = tensor("encoder_q_proj_3_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(26475456)))]; - tensor encoder_k_proj_3_bias = const()[name = tensor("encoder_k_proj_3_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(26737664)))]; - tensor encoder_k_proj_3_weight = const()[name = tensor("encoder_k_proj_3_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(26738752)))]; - tensor encoder_v_proj_3_bias = const()[name = tensor("encoder_v_proj_3_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27000960)))]; - tensor encoder_v_proj_3_weight = const()[name = tensor("encoder_v_proj_3_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27002048)))]; - tensor encoder_g_proj_3_bias = const()[name = tensor("encoder_g_proj_3_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27264256)))]; - tensor encoder_g_proj_3_weight = const()[name = tensor("encoder_g_proj_3_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27265344)))]; - tensor encoder_out_proj_3_bias = const()[name = tensor("encoder_out_proj_3_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27527552)))]; - tensor encoder_out_proj_3_weight = const()[name = tensor("encoder_out_proj_3_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27528640)))]; - tensor encoder_conv_module_3_sequential_0_bias = const()[name = tensor("encoder_conv_module_3_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27790848)))]; - tensor encoder_conv_module_3_sequential_0_weight = const()[name = tensor("encoder_conv_module_3_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27791936)))]; - tensor encoder_conv_module_3_sequential_2_conv_bias = const()[name = tensor("encoder_conv_module_3_sequential_2_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27793024)))]; - tensor encoder_conv_module_3_sequential_2_conv_weight = const()[name = tensor("encoder_conv_module_3_sequential_2_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27795136)))]; - tensor encoder_conv_module_3_sequential_4_conv_weight = const()[name = tensor("encoder_conv_module_3_sequential_4_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28319488)))]; - tensor encoder_conv_module_3_sequential_5_running_var = const()[name = tensor("encoder_conv_module_3_sequential_5_running_var"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28335936)))]; - tensor encoder_conv_module_3_sequential_5_running_mean = const()[name = tensor("encoder_conv_module_3_sequential_5_running_mean"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28337024)))]; - tensor encoder_conv_module_3_sequential_5_bias = const()[name = tensor("encoder_conv_module_3_sequential_5_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28338112)))]; - tensor encoder_conv_module_3_sequential_5_weight = const()[name = tensor("encoder_conv_module_3_sequential_5_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28339200)))]; - tensor encoder_conv_module_3_sequential_7_conv_bias = const()[name = tensor("encoder_conv_module_3_sequential_7_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28340288)))]; - tensor encoder_conv_module_3_sequential_7_conv_weight = const()[name = tensor("encoder_conv_module_3_sequential_7_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28341376)))]; - tensor encoder_ffn2_3_module_sequential_0_bias = const()[name = tensor("encoder_ffn2_3_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28603584)))]; - tensor encoder_ffn2_3_module_sequential_0_weight = const()[name = tensor("encoder_ffn2_3_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28604672)))]; - tensor encoder_ffn2_3_module_sequential_1_linear_bias = const()[name = tensor("encoder_ffn2_3_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28605760)))]; - tensor encoder_ffn2_3_module_sequential_1_linear_weight = const()[name = tensor("encoder_ffn2_3_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28609920)))]; - tensor encoder_ffn2_3_module_sequential_4_linear_bias = const()[name = tensor("encoder_ffn2_3_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(29658560)))]; - tensor encoder_ffn2_3_module_sequential_4_linear_weight = const()[name = tensor("encoder_ffn2_3_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(29659648)))]; - tensor encoder_layer_norm_3_bias = const()[name = tensor("encoder_layer_norm_3_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(30708288)))]; - tensor encoder_layer_norm_3_weight = const()[name = tensor("encoder_layer_norm_3_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(30709376)))]; - tensor decoder__causal_mask = const()[name = tensor("decoder__causal_mask"), val = tensor([[0x1p+0, 0x0p+0], [0x1p+0, 0x1p+0]])]; - tensor decoder_convert_bias = const()[name = tensor("decoder_convert_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(30710464)))]; - tensor decoder_convert_weight = const()[name = tensor("decoder_convert_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(30711552)))]; - tensor decoder_q_proj_0_bias = const()[name = tensor("decoder_q_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(31235904)))]; - tensor decoder_q_proj_0_weight = const()[name = tensor("decoder_q_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(31236992)))]; - tensor decoder_k_proj_0_bias = const()[name = tensor("decoder_k_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(31499200)))]; - tensor decoder_k_proj_0_weight = const()[name = tensor("decoder_k_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(31500288)))]; - tensor decoder_v_proj_0_bias = const()[name = tensor("decoder_v_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(31762496)))]; - tensor decoder_v_proj_0_weight = const()[name = tensor("decoder_v_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(31763584)))]; - tensor decoder_g_proj_0_bias = const()[name = tensor("decoder_g_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32025792)))]; - tensor decoder_g_proj_0_weight = const()[name = tensor("decoder_g_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32026880)))]; - tensor decoder_out_proj_0_bias = const()[name = tensor("decoder_out_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32289088)))]; - tensor decoder_out_proj_0_weight = const()[name = tensor("decoder_out_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32290176)))]; - tensor decoder_norm11_0_bias = const()[name = tensor("decoder_norm11_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32552384)))]; - tensor decoder_norm11_0_weight = const()[name = tensor("decoder_norm11_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32553472)))]; - tensor decoder_self_attn2_0_out_proj_bias = const()[name = tensor("decoder_self_attn2_0_out_proj_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32554560)))]; - tensor decoder_self_attn2_0_out_proj_weight = const()[name = tensor("decoder_self_attn2_0_out_proj_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32555648)))]; - tensor decoder_self_attn2_0_in_proj_bias = const()[name = tensor("decoder_self_attn2_0_in_proj_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32817856)))]; - tensor decoder_self_attn2_0_in_proj_weight = const()[name = tensor("decoder_self_attn2_0_in_proj_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32820992)))]; - tensor decoder_norm21_0_bias = const()[name = tensor("decoder_norm21_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(33607488)))]; - tensor decoder_norm21_0_weight = const()[name = tensor("decoder_norm21_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(33608576)))]; - tensor decoder_linear1_0_bias = const()[name = tensor("decoder_linear1_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(33609664)))]; - tensor decoder_linear1_0_weight = const()[name = tensor("decoder_linear1_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(33617920)))]; - tensor decoder_linear2_0_bias = const()[name = tensor("decoder_linear2_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(35715136)))]; - tensor decoder_linear2_0_weight = const()[name = tensor("decoder_linear2_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(35716224)))]; - tensor decoder_norm22_0_bias = const()[name = tensor("decoder_norm22_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(37813440)))]; - tensor decoder_norm22_0_weight = const()[name = tensor("decoder_norm22_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(37814528)))]; - tensor decoder_q_proj_1_bias = const()[name = tensor("decoder_q_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(37815616)))]; - tensor decoder_q_proj_1_weight = const()[name = tensor("decoder_q_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(37816704)))]; - tensor decoder_k_proj_1_bias = const()[name = tensor("decoder_k_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38078912)))]; - tensor decoder_k_proj_1_weight = const()[name = tensor("decoder_k_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38080000)))]; - tensor decoder_v_proj_1_bias = const()[name = tensor("decoder_v_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38342208)))]; - tensor decoder_v_proj_1_weight = const()[name = tensor("decoder_v_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38343296)))]; - tensor decoder_g_proj_1_bias = const()[name = tensor("decoder_g_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38605504)))]; - tensor decoder_g_proj_1_weight = const()[name = tensor("decoder_g_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38606592)))]; - tensor decoder_out_proj_1_bias = const()[name = tensor("decoder_out_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38868800)))]; - tensor decoder_out_proj_1_weight = const()[name = tensor("decoder_out_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38869888)))]; - tensor decoder_norm11_1_bias = const()[name = tensor("decoder_norm11_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(39132096)))]; - tensor decoder_norm11_1_weight = const()[name = tensor("decoder_norm11_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(39133184)))]; - tensor decoder_self_attn2_1_out_proj_bias = const()[name = tensor("decoder_self_attn2_1_out_proj_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(39134272)))]; - tensor decoder_self_attn2_1_out_proj_weight = const()[name = tensor("decoder_self_attn2_1_out_proj_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(39135360)))]; - tensor decoder_self_attn2_1_in_proj_bias = const()[name = tensor("decoder_self_attn2_1_in_proj_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(39397568)))]; - tensor decoder_self_attn2_1_in_proj_weight = const()[name = tensor("decoder_self_attn2_1_in_proj_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(39400704)))]; - tensor decoder_norm21_1_bias = const()[name = tensor("decoder_norm21_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(40187200)))]; - tensor decoder_norm21_1_weight = const()[name = tensor("decoder_norm21_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(40188288)))]; - tensor decoder_linear1_1_bias = const()[name = tensor("decoder_linear1_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(40189376)))]; - tensor decoder_linear1_1_weight = const()[name = tensor("decoder_linear1_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(40197632)))]; - tensor decoder_linear2_1_bias = const()[name = tensor("decoder_linear2_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(42294848)))]; - tensor decoder_linear2_1_weight = const()[name = tensor("decoder_linear2_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(42295936)))]; - tensor decoder_norm22_1_bias = const()[name = tensor("decoder_norm22_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(44393152)))]; - tensor decoder_norm22_1_weight = const()[name = tensor("decoder_norm22_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(44394240)))]; - tensor var_11 = const()[name = tensor("op_11"), val = tensor(0x1.5798eep-27)]; - tensor var_17 = const()[name = tensor("op_17"), val = tensor(0x1.0c6f7ap-20)]; - tensor var_20 = const()[name = tensor("op_20"), val = tensor(2)]; - tensor var_23 = const()[name = tensor("op_23"), val = tensor(0)]; - tensor var_26 = const()[name = tensor("op_26"), val = tensor(1)]; - tensor var_28 = const()[name = tensor("op_28"), val = tensor(0x1.4f8b58p-17)]; - tensor var_29 = const()[name = tensor("op_29"), val = tensor(true)]; - tensor input_1 = linear(bias = encoder_input_projection_linear_bias, weight = encoder_input_projection_linear_weight, x = features)[name = tensor("linear_0")]; - tensor input_3_axes_0 = const()[name = tensor("input_3_axes_0"), val = tensor([-1])]; - tensor input_3 = layer_norm(axes = input_3_axes_0, beta = encoder_pre_layer_norm_bias, epsilon = var_28, gamma = encoder_pre_layer_norm_weight, x = input_1)[name = tensor("input_3")]; + func main(tensor cnn_window, tensor dec_kv, tensor dec_scale, tensor enc_conv_cache, tensor enc_kv, tensor enc_scale, tensor features, tensor valid_mask) { + tensor inner_encoder_cnn_bias = const()[name = tensor("inner_encoder_cnn_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(64)))]; + tensor inner_encoder_cnn_weight = const()[name = tensor("inner_encoder_cnn_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1152)))]; + tensor inner_encoder__causal_mask = const()[name = tensor("inner_encoder__causal_mask"), val = tensor([[0x1p+0, 0x0p+0], [0x1p+0, 0x1p+0]])]; + tensor inner_encoder__t_index = const()[name = tensor("inner_encoder__t_index"), val = tensor([0x1p+0, 0x1p+1])]; + tensor inner_encoder_input_projection_linear_bias = const()[name = tensor("inner_encoder_input_projection_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4981952)))]; + tensor inner_encoder_input_projection_linear_weight = const()[name = tensor("inner_encoder_input_projection_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4983040)))]; + tensor inner_encoder_pre_layer_norm_bias = const()[name = tensor("inner_encoder_pre_layer_norm_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(5336384)))]; + tensor inner_encoder_pre_layer_norm_weight = const()[name = tensor("inner_encoder_pre_layer_norm_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(5337472)))]; + tensor inner_encoder_ffn1_0_module_sequential_0_bias = const()[name = tensor("inner_encoder_ffn1_0_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(5338560)))]; + tensor inner_encoder_ffn1_0_module_sequential_0_weight = const()[name = tensor("inner_encoder_ffn1_0_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(5339648)))]; + tensor inner_encoder_ffn1_0_module_sequential_1_linear_bias = const()[name = tensor("inner_encoder_ffn1_0_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(5340736)))]; + tensor inner_encoder_ffn1_0_module_sequential_1_linear_weight = const()[name = tensor("inner_encoder_ffn1_0_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(5344896)))]; + tensor inner_encoder_ffn1_0_module_sequential_4_linear_bias = const()[name = tensor("inner_encoder_ffn1_0_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(6393536)))]; + tensor inner_encoder_ffn1_0_module_sequential_4_linear_weight = const()[name = tensor("inner_encoder_ffn1_0_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(6394624)))]; + tensor inner_encoder_ret_lns_0_bias = const()[name = tensor("inner_encoder_ret_lns_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7443264)))]; + tensor inner_encoder_ret_lns_0_weight = const()[name = tensor("inner_encoder_ret_lns_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7444352)))]; + tensor inner_encoder_q_proj_0_bias = const()[name = tensor("inner_encoder_q_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7445440)))]; + tensor inner_encoder_q_proj_0_weight = const()[name = tensor("inner_encoder_q_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7446528)))]; + tensor inner_encoder_k_proj_0_bias = const()[name = tensor("inner_encoder_k_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7708736)))]; + tensor inner_encoder_k_proj_0_weight = const()[name = tensor("inner_encoder_k_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7709824)))]; + tensor inner_encoder_v_proj_0_bias = const()[name = tensor("inner_encoder_v_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7972032)))]; + tensor inner_encoder_v_proj_0_weight = const()[name = tensor("inner_encoder_v_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7973120)))]; + tensor inner_encoder_g_proj_0_bias = const()[name = tensor("inner_encoder_g_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8235328)))]; + tensor inner_encoder_g_proj_0_weight = const()[name = tensor("inner_encoder_g_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8236416)))]; + tensor inner_encoder_out_proj_0_bias = const()[name = tensor("inner_encoder_out_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8498624)))]; + tensor inner_encoder_out_proj_0_weight = const()[name = tensor("inner_encoder_out_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8499712)))]; + tensor inner_encoder_conv_module_0_sequential_0_bias = const()[name = tensor("inner_encoder_conv_module_0_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8761920)))]; + tensor inner_encoder_conv_module_0_sequential_0_weight = const()[name = tensor("inner_encoder_conv_module_0_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8763008)))]; + tensor inner_encoder_conv_module_0_sequential_2_conv_bias = const()[name = tensor("inner_encoder_conv_module_0_sequential_2_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8764096)))]; + tensor inner_encoder_conv_module_0_sequential_2_conv_weight = const()[name = tensor("inner_encoder_conv_module_0_sequential_2_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8766208)))]; + tensor inner_encoder_conv_module_0_sequential_4_conv_weight = const()[name = tensor("inner_encoder_conv_module_0_sequential_4_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9290560)))]; + tensor inner_encoder_conv_module_0_sequential_5_running_var = const()[name = tensor("inner_encoder_conv_module_0_sequential_5_running_var"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9307008)))]; + tensor inner_encoder_conv_module_0_sequential_5_running_mean = const()[name = tensor("inner_encoder_conv_module_0_sequential_5_running_mean"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9308096)))]; + tensor inner_encoder_conv_module_0_sequential_5_bias = const()[name = tensor("inner_encoder_conv_module_0_sequential_5_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9309184)))]; + tensor inner_encoder_conv_module_0_sequential_5_weight = const()[name = tensor("inner_encoder_conv_module_0_sequential_5_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9310272)))]; + tensor inner_encoder_conv_module_0_sequential_7_conv_bias = const()[name = tensor("inner_encoder_conv_module_0_sequential_7_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9311360)))]; + tensor inner_encoder_conv_module_0_sequential_7_conv_weight = const()[name = tensor("inner_encoder_conv_module_0_sequential_7_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9312448)))]; + tensor inner_encoder_ffn2_0_module_sequential_0_bias = const()[name = tensor("inner_encoder_ffn2_0_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9574656)))]; + tensor inner_encoder_ffn2_0_module_sequential_0_weight = const()[name = tensor("inner_encoder_ffn2_0_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9575744)))]; + tensor inner_encoder_ffn2_0_module_sequential_1_linear_bias = const()[name = tensor("inner_encoder_ffn2_0_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9576832)))]; + tensor inner_encoder_ffn2_0_module_sequential_1_linear_weight = const()[name = tensor("inner_encoder_ffn2_0_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9580992)))]; + tensor inner_encoder_ffn2_0_module_sequential_4_linear_bias = const()[name = tensor("inner_encoder_ffn2_0_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(10629632)))]; + tensor inner_encoder_ffn2_0_module_sequential_4_linear_weight = const()[name = tensor("inner_encoder_ffn2_0_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(10630720)))]; + tensor inner_encoder_layer_norm_0_bias = const()[name = tensor("inner_encoder_layer_norm_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11679360)))]; + tensor inner_encoder_layer_norm_0_weight = const()[name = tensor("inner_encoder_layer_norm_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11680448)))]; + tensor inner_encoder_ffn1_1_module_sequential_0_bias = const()[name = tensor("inner_encoder_ffn1_1_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11681536)))]; + tensor inner_encoder_ffn1_1_module_sequential_0_weight = const()[name = tensor("inner_encoder_ffn1_1_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11682624)))]; + tensor inner_encoder_ffn1_1_module_sequential_1_linear_bias = const()[name = tensor("inner_encoder_ffn1_1_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11683712)))]; + tensor inner_encoder_ffn1_1_module_sequential_1_linear_weight = const()[name = tensor("inner_encoder_ffn1_1_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11687872)))]; + tensor inner_encoder_ffn1_1_module_sequential_4_linear_bias = const()[name = tensor("inner_encoder_ffn1_1_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(12736512)))]; + tensor inner_encoder_ffn1_1_module_sequential_4_linear_weight = const()[name = tensor("inner_encoder_ffn1_1_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(12737600)))]; + tensor inner_encoder_ret_lns_1_bias = const()[name = tensor("inner_encoder_ret_lns_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(13786240)))]; + tensor inner_encoder_ret_lns_1_weight = const()[name = tensor("inner_encoder_ret_lns_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(13787328)))]; + tensor inner_encoder_q_proj_1_bias = const()[name = tensor("inner_encoder_q_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(13788416)))]; + tensor inner_encoder_q_proj_1_weight = const()[name = tensor("inner_encoder_q_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(13789504)))]; + tensor inner_encoder_k_proj_1_bias = const()[name = tensor("inner_encoder_k_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14051712)))]; + tensor inner_encoder_k_proj_1_weight = const()[name = tensor("inner_encoder_k_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14052800)))]; + tensor inner_encoder_v_proj_1_bias = const()[name = tensor("inner_encoder_v_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14315008)))]; + tensor inner_encoder_v_proj_1_weight = const()[name = tensor("inner_encoder_v_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14316096)))]; + tensor inner_encoder_g_proj_1_bias = const()[name = tensor("inner_encoder_g_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14578304)))]; + tensor inner_encoder_g_proj_1_weight = const()[name = tensor("inner_encoder_g_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14579392)))]; + tensor inner_encoder_out_proj_1_bias = const()[name = tensor("inner_encoder_out_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14841600)))]; + tensor inner_encoder_out_proj_1_weight = const()[name = tensor("inner_encoder_out_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14842688)))]; + tensor inner_encoder_conv_module_1_sequential_0_bias = const()[name = tensor("inner_encoder_conv_module_1_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15104896)))]; + tensor inner_encoder_conv_module_1_sequential_0_weight = const()[name = tensor("inner_encoder_conv_module_1_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15105984)))]; + tensor inner_encoder_conv_module_1_sequential_2_conv_bias = const()[name = tensor("inner_encoder_conv_module_1_sequential_2_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15107072)))]; + tensor inner_encoder_conv_module_1_sequential_2_conv_weight = const()[name = tensor("inner_encoder_conv_module_1_sequential_2_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15109184)))]; + tensor inner_encoder_conv_module_1_sequential_4_conv_weight = const()[name = tensor("inner_encoder_conv_module_1_sequential_4_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15633536)))]; + tensor inner_encoder_conv_module_1_sequential_5_running_var = const()[name = tensor("inner_encoder_conv_module_1_sequential_5_running_var"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15649984)))]; + tensor inner_encoder_conv_module_1_sequential_5_running_mean = const()[name = tensor("inner_encoder_conv_module_1_sequential_5_running_mean"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15651072)))]; + tensor inner_encoder_conv_module_1_sequential_5_bias = const()[name = tensor("inner_encoder_conv_module_1_sequential_5_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15652160)))]; + tensor inner_encoder_conv_module_1_sequential_5_weight = const()[name = tensor("inner_encoder_conv_module_1_sequential_5_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15653248)))]; + tensor inner_encoder_conv_module_1_sequential_7_conv_bias = const()[name = tensor("inner_encoder_conv_module_1_sequential_7_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15654336)))]; + tensor inner_encoder_conv_module_1_sequential_7_conv_weight = const()[name = tensor("inner_encoder_conv_module_1_sequential_7_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15655424)))]; + tensor inner_encoder_ffn2_1_module_sequential_0_bias = const()[name = tensor("inner_encoder_ffn2_1_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15917632)))]; + tensor inner_encoder_ffn2_1_module_sequential_0_weight = const()[name = tensor("inner_encoder_ffn2_1_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15918720)))]; + tensor inner_encoder_ffn2_1_module_sequential_1_linear_bias = const()[name = tensor("inner_encoder_ffn2_1_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15919808)))]; + tensor inner_encoder_ffn2_1_module_sequential_1_linear_weight = const()[name = tensor("inner_encoder_ffn2_1_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15923968)))]; + tensor inner_encoder_ffn2_1_module_sequential_4_linear_bias = const()[name = tensor("inner_encoder_ffn2_1_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(16972608)))]; + tensor inner_encoder_ffn2_1_module_sequential_4_linear_weight = const()[name = tensor("inner_encoder_ffn2_1_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(16973696)))]; + tensor inner_encoder_layer_norm_1_bias = const()[name = tensor("inner_encoder_layer_norm_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18022336)))]; + tensor inner_encoder_layer_norm_1_weight = const()[name = tensor("inner_encoder_layer_norm_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18023424)))]; + tensor inner_encoder_ffn1_2_module_sequential_0_bias = const()[name = tensor("inner_encoder_ffn1_2_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18024512)))]; + tensor inner_encoder_ffn1_2_module_sequential_0_weight = const()[name = tensor("inner_encoder_ffn1_2_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18025600)))]; + tensor inner_encoder_ffn1_2_module_sequential_1_linear_bias = const()[name = tensor("inner_encoder_ffn1_2_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18026688)))]; + tensor inner_encoder_ffn1_2_module_sequential_1_linear_weight = const()[name = tensor("inner_encoder_ffn1_2_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18030848)))]; + tensor inner_encoder_ffn1_2_module_sequential_4_linear_bias = const()[name = tensor("inner_encoder_ffn1_2_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(19079488)))]; + tensor inner_encoder_ffn1_2_module_sequential_4_linear_weight = const()[name = tensor("inner_encoder_ffn1_2_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(19080576)))]; + tensor inner_encoder_ret_lns_2_bias = const()[name = tensor("inner_encoder_ret_lns_2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20129216)))]; + tensor inner_encoder_ret_lns_2_weight = const()[name = tensor("inner_encoder_ret_lns_2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20130304)))]; + tensor inner_encoder_q_proj_2_bias = const()[name = tensor("inner_encoder_q_proj_2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20131392)))]; + tensor inner_encoder_q_proj_2_weight = const()[name = tensor("inner_encoder_q_proj_2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20132480)))]; + tensor inner_encoder_k_proj_2_bias = const()[name = tensor("inner_encoder_k_proj_2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20394688)))]; + tensor inner_encoder_k_proj_2_weight = const()[name = tensor("inner_encoder_k_proj_2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20395776)))]; + tensor inner_encoder_v_proj_2_bias = const()[name = tensor("inner_encoder_v_proj_2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20657984)))]; + tensor inner_encoder_v_proj_2_weight = const()[name = tensor("inner_encoder_v_proj_2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20659072)))]; + tensor inner_encoder_g_proj_2_bias = const()[name = tensor("inner_encoder_g_proj_2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20921280)))]; + tensor inner_encoder_g_proj_2_weight = const()[name = tensor("inner_encoder_g_proj_2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20922368)))]; + tensor inner_encoder_out_proj_2_bias = const()[name = tensor("inner_encoder_out_proj_2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21184576)))]; + tensor inner_encoder_out_proj_2_weight = const()[name = tensor("inner_encoder_out_proj_2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21185664)))]; + tensor inner_encoder_conv_module_2_sequential_0_bias = const()[name = tensor("inner_encoder_conv_module_2_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21447872)))]; + tensor inner_encoder_conv_module_2_sequential_0_weight = const()[name = tensor("inner_encoder_conv_module_2_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21448960)))]; + tensor inner_encoder_conv_module_2_sequential_2_conv_bias = const()[name = tensor("inner_encoder_conv_module_2_sequential_2_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21450048)))]; + tensor inner_encoder_conv_module_2_sequential_2_conv_weight = const()[name = tensor("inner_encoder_conv_module_2_sequential_2_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21452160)))]; + tensor inner_encoder_conv_module_2_sequential_4_conv_weight = const()[name = tensor("inner_encoder_conv_module_2_sequential_4_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21976512)))]; + tensor inner_encoder_conv_module_2_sequential_5_running_var = const()[name = tensor("inner_encoder_conv_module_2_sequential_5_running_var"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21992960)))]; + tensor inner_encoder_conv_module_2_sequential_5_running_mean = const()[name = tensor("inner_encoder_conv_module_2_sequential_5_running_mean"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21994048)))]; + tensor inner_encoder_conv_module_2_sequential_5_bias = const()[name = tensor("inner_encoder_conv_module_2_sequential_5_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21995136)))]; + tensor inner_encoder_conv_module_2_sequential_5_weight = const()[name = tensor("inner_encoder_conv_module_2_sequential_5_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21996224)))]; + tensor inner_encoder_conv_module_2_sequential_7_conv_bias = const()[name = tensor("inner_encoder_conv_module_2_sequential_7_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21997312)))]; + tensor inner_encoder_conv_module_2_sequential_7_conv_weight = const()[name = tensor("inner_encoder_conv_module_2_sequential_7_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21998400)))]; + tensor inner_encoder_ffn2_2_module_sequential_0_bias = const()[name = tensor("inner_encoder_ffn2_2_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(22260608)))]; + tensor inner_encoder_ffn2_2_module_sequential_0_weight = const()[name = tensor("inner_encoder_ffn2_2_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(22261696)))]; + tensor inner_encoder_ffn2_2_module_sequential_1_linear_bias = const()[name = tensor("inner_encoder_ffn2_2_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(22262784)))]; + tensor inner_encoder_ffn2_2_module_sequential_1_linear_weight = const()[name = tensor("inner_encoder_ffn2_2_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(22266944)))]; + tensor inner_encoder_ffn2_2_module_sequential_4_linear_bias = const()[name = tensor("inner_encoder_ffn2_2_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(23315584)))]; + tensor inner_encoder_ffn2_2_module_sequential_4_linear_weight = const()[name = tensor("inner_encoder_ffn2_2_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(23316672)))]; + tensor inner_encoder_layer_norm_2_bias = const()[name = tensor("inner_encoder_layer_norm_2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(24365312)))]; + tensor inner_encoder_layer_norm_2_weight = const()[name = tensor("inner_encoder_layer_norm_2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(24366400)))]; + tensor inner_encoder_ffn1_3_module_sequential_0_bias = const()[name = tensor("inner_encoder_ffn1_3_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(24367488)))]; + tensor inner_encoder_ffn1_3_module_sequential_0_weight = const()[name = tensor("inner_encoder_ffn1_3_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(24368576)))]; + tensor inner_encoder_ffn1_3_module_sequential_1_linear_bias = const()[name = tensor("inner_encoder_ffn1_3_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(24369664)))]; + tensor inner_encoder_ffn1_3_module_sequential_1_linear_weight = const()[name = tensor("inner_encoder_ffn1_3_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(24373824)))]; + tensor inner_encoder_ffn1_3_module_sequential_4_linear_bias = const()[name = tensor("inner_encoder_ffn1_3_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(25422464)))]; + tensor inner_encoder_ffn1_3_module_sequential_4_linear_weight = const()[name = tensor("inner_encoder_ffn1_3_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(25423552)))]; + tensor inner_encoder_ret_lns_3_bias = const()[name = tensor("inner_encoder_ret_lns_3_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(26472192)))]; + tensor inner_encoder_ret_lns_3_weight = const()[name = tensor("inner_encoder_ret_lns_3_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(26473280)))]; + tensor inner_encoder_q_proj_3_bias = const()[name = tensor("inner_encoder_q_proj_3_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(26474368)))]; + tensor inner_encoder_q_proj_3_weight = const()[name = tensor("inner_encoder_q_proj_3_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(26475456)))]; + tensor inner_encoder_k_proj_3_bias = const()[name = tensor("inner_encoder_k_proj_3_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(26737664)))]; + tensor inner_encoder_k_proj_3_weight = const()[name = tensor("inner_encoder_k_proj_3_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(26738752)))]; + tensor inner_encoder_v_proj_3_bias = const()[name = tensor("inner_encoder_v_proj_3_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27000960)))]; + tensor inner_encoder_v_proj_3_weight = const()[name = tensor("inner_encoder_v_proj_3_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27002048)))]; + tensor inner_encoder_g_proj_3_bias = const()[name = tensor("inner_encoder_g_proj_3_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27264256)))]; + tensor inner_encoder_g_proj_3_weight = const()[name = tensor("inner_encoder_g_proj_3_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27265344)))]; + tensor inner_encoder_out_proj_3_bias = const()[name = tensor("inner_encoder_out_proj_3_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27527552)))]; + tensor inner_encoder_out_proj_3_weight = const()[name = tensor("inner_encoder_out_proj_3_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27528640)))]; + tensor inner_encoder_conv_module_3_sequential_0_bias = const()[name = tensor("inner_encoder_conv_module_3_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27790848)))]; + tensor inner_encoder_conv_module_3_sequential_0_weight = const()[name = tensor("inner_encoder_conv_module_3_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27791936)))]; + tensor inner_encoder_conv_module_3_sequential_2_conv_bias = const()[name = tensor("inner_encoder_conv_module_3_sequential_2_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27793024)))]; + tensor inner_encoder_conv_module_3_sequential_2_conv_weight = const()[name = tensor("inner_encoder_conv_module_3_sequential_2_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27795136)))]; + tensor inner_encoder_conv_module_3_sequential_4_conv_weight = const()[name = tensor("inner_encoder_conv_module_3_sequential_4_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28319488)))]; + tensor inner_encoder_conv_module_3_sequential_5_running_var = const()[name = tensor("inner_encoder_conv_module_3_sequential_5_running_var"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28335936)))]; + tensor inner_encoder_conv_module_3_sequential_5_running_mean = const()[name = tensor("inner_encoder_conv_module_3_sequential_5_running_mean"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28337024)))]; + tensor inner_encoder_conv_module_3_sequential_5_bias = const()[name = tensor("inner_encoder_conv_module_3_sequential_5_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28338112)))]; + tensor inner_encoder_conv_module_3_sequential_5_weight = const()[name = tensor("inner_encoder_conv_module_3_sequential_5_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28339200)))]; + tensor inner_encoder_conv_module_3_sequential_7_conv_bias = const()[name = tensor("inner_encoder_conv_module_3_sequential_7_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28340288)))]; + tensor inner_encoder_conv_module_3_sequential_7_conv_weight = const()[name = tensor("inner_encoder_conv_module_3_sequential_7_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28341376)))]; + tensor inner_encoder_ffn2_3_module_sequential_0_bias = const()[name = tensor("inner_encoder_ffn2_3_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28603584)))]; + tensor inner_encoder_ffn2_3_module_sequential_0_weight = const()[name = tensor("inner_encoder_ffn2_3_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28604672)))]; + tensor inner_encoder_ffn2_3_module_sequential_1_linear_bias = const()[name = tensor("inner_encoder_ffn2_3_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28605760)))]; + tensor inner_encoder_ffn2_3_module_sequential_1_linear_weight = const()[name = tensor("inner_encoder_ffn2_3_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28609920)))]; + tensor inner_encoder_ffn2_3_module_sequential_4_linear_bias = const()[name = tensor("inner_encoder_ffn2_3_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(29658560)))]; + tensor inner_encoder_ffn2_3_module_sequential_4_linear_weight = const()[name = tensor("inner_encoder_ffn2_3_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(29659648)))]; + tensor inner_encoder_layer_norm_3_bias = const()[name = tensor("inner_encoder_layer_norm_3_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(30708288)))]; + tensor inner_encoder_layer_norm_3_weight = const()[name = tensor("inner_encoder_layer_norm_3_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(30709376)))]; + tensor inner_decoder__causal_mask = const()[name = tensor("inner_decoder__causal_mask"), val = tensor([[0x1p+0, 0x0p+0], [0x1p+0, 0x1p+0]])]; + tensor inner_decoder_convert_bias = const()[name = tensor("inner_decoder_convert_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(30710464)))]; + tensor inner_decoder_convert_weight = const()[name = tensor("inner_decoder_convert_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(30711552)))]; + tensor inner_decoder_q_proj_0_bias = const()[name = tensor("inner_decoder_q_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(31235904)))]; + tensor inner_decoder_q_proj_0_weight = const()[name = tensor("inner_decoder_q_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(31236992)))]; + tensor inner_decoder_k_proj_0_bias = const()[name = tensor("inner_decoder_k_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(31499200)))]; + tensor inner_decoder_k_proj_0_weight = const()[name = tensor("inner_decoder_k_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(31500288)))]; + tensor inner_decoder_v_proj_0_bias = const()[name = tensor("inner_decoder_v_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(31762496)))]; + tensor inner_decoder_v_proj_0_weight = const()[name = tensor("inner_decoder_v_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(31763584)))]; + tensor inner_decoder_g_proj_0_bias = const()[name = tensor("inner_decoder_g_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32025792)))]; + tensor inner_decoder_g_proj_0_weight = const()[name = tensor("inner_decoder_g_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32026880)))]; + tensor inner_decoder_out_proj_0_bias = const()[name = tensor("inner_decoder_out_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32289088)))]; + tensor inner_decoder_out_proj_0_weight = const()[name = tensor("inner_decoder_out_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32290176)))]; + tensor inner_decoder_norm11_0_bias = const()[name = tensor("inner_decoder_norm11_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32552384)))]; + tensor inner_decoder_norm11_0_weight = const()[name = tensor("inner_decoder_norm11_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32553472)))]; + tensor inner_decoder_self_attn2_0_out_proj_bias = const()[name = tensor("inner_decoder_self_attn2_0_out_proj_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32554560)))]; + tensor inner_decoder_self_attn2_0_out_proj_weight = const()[name = tensor("inner_decoder_self_attn2_0_out_proj_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32555648)))]; + tensor inner_decoder_self_attn2_0_in_proj_bias = const()[name = tensor("inner_decoder_self_attn2_0_in_proj_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32817856)))]; + tensor inner_decoder_self_attn2_0_in_proj_weight = const()[name = tensor("inner_decoder_self_attn2_0_in_proj_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32820992)))]; + tensor inner_decoder_norm21_0_bias = const()[name = tensor("inner_decoder_norm21_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(33607488)))]; + tensor inner_decoder_norm21_0_weight = const()[name = tensor("inner_decoder_norm21_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(33608576)))]; + tensor inner_decoder_linear1_0_bias = const()[name = tensor("inner_decoder_linear1_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(33609664)))]; + tensor inner_decoder_linear1_0_weight = const()[name = tensor("inner_decoder_linear1_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(33617920)))]; + tensor inner_decoder_linear2_0_bias = const()[name = tensor("inner_decoder_linear2_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(35715136)))]; + tensor inner_decoder_linear2_0_weight = const()[name = tensor("inner_decoder_linear2_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(35716224)))]; + tensor inner_decoder_norm22_0_bias = const()[name = tensor("inner_decoder_norm22_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(37813440)))]; + tensor inner_decoder_norm22_0_weight = const()[name = tensor("inner_decoder_norm22_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(37814528)))]; + tensor inner_decoder_q_proj_1_bias = const()[name = tensor("inner_decoder_q_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(37815616)))]; + tensor inner_decoder_q_proj_1_weight = const()[name = tensor("inner_decoder_q_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(37816704)))]; + tensor inner_decoder_k_proj_1_bias = const()[name = tensor("inner_decoder_k_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38078912)))]; + tensor inner_decoder_k_proj_1_weight = const()[name = tensor("inner_decoder_k_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38080000)))]; + tensor inner_decoder_v_proj_1_bias = const()[name = tensor("inner_decoder_v_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38342208)))]; + tensor inner_decoder_v_proj_1_weight = const()[name = tensor("inner_decoder_v_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38343296)))]; + tensor inner_decoder_g_proj_1_bias = const()[name = tensor("inner_decoder_g_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38605504)))]; + tensor inner_decoder_g_proj_1_weight = const()[name = tensor("inner_decoder_g_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38606592)))]; + tensor inner_decoder_out_proj_1_bias = const()[name = tensor("inner_decoder_out_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38868800)))]; + tensor inner_decoder_out_proj_1_weight = const()[name = tensor("inner_decoder_out_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38869888)))]; + tensor inner_decoder_norm11_1_bias = const()[name = tensor("inner_decoder_norm11_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(39132096)))]; + tensor inner_decoder_norm11_1_weight = const()[name = tensor("inner_decoder_norm11_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(39133184)))]; + tensor inner_decoder_self_attn2_1_out_proj_bias = const()[name = tensor("inner_decoder_self_attn2_1_out_proj_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(39134272)))]; + tensor inner_decoder_self_attn2_1_out_proj_weight = const()[name = tensor("inner_decoder_self_attn2_1_out_proj_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(39135360)))]; + tensor inner_decoder_self_attn2_1_in_proj_bias = const()[name = tensor("inner_decoder_self_attn2_1_in_proj_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(39397568)))]; + tensor inner_decoder_self_attn2_1_in_proj_weight = const()[name = tensor("inner_decoder_self_attn2_1_in_proj_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(39400704)))]; + tensor inner_decoder_norm21_1_bias = const()[name = tensor("inner_decoder_norm21_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(40187200)))]; + tensor inner_decoder_norm21_1_weight = const()[name = tensor("inner_decoder_norm21_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(40188288)))]; + tensor inner_decoder_linear1_1_bias = const()[name = tensor("inner_decoder_linear1_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(40189376)))]; + tensor inner_decoder_linear1_1_weight = const()[name = tensor("inner_decoder_linear1_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(40197632)))]; + tensor inner_decoder_linear2_1_bias = const()[name = tensor("inner_decoder_linear2_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(42294848)))]; + tensor inner_decoder_linear2_1_weight = const()[name = tensor("inner_decoder_linear2_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(42295936)))]; + tensor inner_decoder_norm22_1_bias = const()[name = tensor("inner_decoder_norm22_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(44393152)))]; + tensor inner_decoder_norm22_1_weight = const()[name = tensor("inner_decoder_norm22_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(44394240)))]; + tensor var_19_begin_0 = const()[name = tensor("op_19_begin_0"), val = tensor([0, 0, 0])]; + tensor var_19_end_0 = const()[name = tensor("op_19_end_0"), val = tensor([1, 15, 23])]; + tensor var_19_end_mask_0 = const()[name = tensor("op_19_end_mask_0"), val = tensor([true, false, true])]; + tensor var_19 = slice_by_index(begin = var_19_begin_0, end = var_19_end_0, end_mask = var_19_end_mask_0, x = features)[name = tensor("op_19")]; + tensor var_29_begin_0 = const()[name = tensor("op_29_begin_0"), val = tensor([0, 10, 0])]; + tensor var_29_end_0 = const()[name = tensor("op_29_end_0"), val = tensor([1, 1, 23])]; + tensor var_29_end_mask_0 = const()[name = tensor("op_29_end_mask_0"), val = tensor([true, true, true])]; + tensor var_29 = slice_by_index(begin = var_29_begin_0, end = var_29_end_0, end_mask = var_29_end_mask_0, x = features)[name = tensor("op_29")]; + tensor stacked_axis_0 = const()[name = tensor("stacked_axis_0"), val = tensor(1)]; + tensor stacked = stack(axis = stacked_axis_0, values = (var_19, var_29))[name = tensor("stacked")]; + tensor var_36 = const()[name = tensor("op_36"), val = tensor([1, 2, 345])]; + tensor input_1 = reshape(shape = var_36, x = stacked)[name = tensor("input_1")]; + tensor var_39 = const()[name = tensor("op_39"), val = tensor(0x1p+0)]; + tensor var_45 = const()[name = tensor("op_45"), val = tensor(true)]; + tensor var_46 = const()[name = tensor("op_46"), val = tensor(0x1.4f8b58p-17)]; + tensor var_49 = const()[name = tensor("op_49"), val = tensor(0)]; + tensor var_51 = const()[name = tensor("op_51"), val = tensor(2)]; + tensor var_52 = const()[name = tensor("op_52"), val = tensor(-1)]; + tensor var_54 = const()[name = tensor("op_54"), val = tensor(0x1.0c6f7ap-20)]; + tensor var_59 = const()[name = tensor("op_59"), val = tensor(0x1.5798eep-27)]; + tensor var_62 = const()[name = tensor("op_62"), val = tensor(1)]; + tensor input_3 = linear(bias = inner_encoder_input_projection_linear_bias, weight = inner_encoder_input_projection_linear_weight, x = input_1)[name = tensor("linear_0")]; tensor input_5_axes_0 = const()[name = tensor("input_5_axes_0"), val = tensor([-1])]; - tensor input_5 = layer_norm(axes = input_5_axes_0, beta = encoder_ffn1_0_module_sequential_0_bias, epsilon = var_28, gamma = encoder_ffn1_0_module_sequential_0_weight, x = input_3)[name = tensor("input_5")]; - tensor inputs_1 = linear(bias = encoder_ffn1_0_module_sequential_1_linear_bias, weight = encoder_ffn1_0_module_sequential_1_linear_weight, x = input_5)[name = tensor("linear_1")]; - tensor input_7 = silu(x = inputs_1)[name = tensor("input_7")]; - tensor input_11 = linear(bias = encoder_ffn1_0_module_sequential_4_linear_bias, weight = encoder_ffn1_0_module_sequential_4_linear_weight, x = input_7)[name = tensor("linear_2")]; - tensor var_147 = const()[name = tensor("op_147"), val = tensor(0x1p-1)]; - tensor var_148 = mul(x = input_11, y = var_147)[name = tensor("op_148")]; - tensor input_13 = add(x = var_148, y = input_3)[name = tensor("input_13")]; + tensor input_5 = layer_norm(axes = input_5_axes_0, beta = inner_encoder_pre_layer_norm_bias, epsilon = var_46, gamma = inner_encoder_pre_layer_norm_weight, x = input_3)[name = tensor("input_5")]; + tensor input_7_axes_0 = const()[name = tensor("input_7_axes_0"), val = tensor([-1])]; + tensor input_7 = layer_norm(axes = input_7_axes_0, beta = inner_encoder_ffn1_0_module_sequential_0_bias, epsilon = var_46, gamma = inner_encoder_ffn1_0_module_sequential_0_weight, x = input_5)[name = tensor("input_7")]; + tensor inputs_1 = linear(bias = inner_encoder_ffn1_0_module_sequential_1_linear_bias, weight = inner_encoder_ffn1_0_module_sequential_1_linear_weight, x = input_7)[name = tensor("linear_1")]; + tensor input_9 = silu(x = inputs_1)[name = tensor("input_9")]; + tensor input_13 = linear(bias = inner_encoder_ffn1_0_module_sequential_4_linear_bias, weight = inner_encoder_ffn1_0_module_sequential_4_linear_weight, x = input_9)[name = tensor("linear_2")]; + tensor var_183 = const()[name = tensor("op_183"), val = tensor(0x1p-1)]; + tensor var_184 = mul(x = input_13, y = var_183)[name = tensor("op_184")]; + tensor input_15 = add(x = var_184, y = input_5)[name = tensor("input_15")]; tensor x_1_axes_0 = const()[name = tensor("x_1_axes_0"), val = tensor([-1])]; - tensor x_1 = layer_norm(axes = x_1_axes_0, beta = encoder_ret_lns_0_bias, epsilon = var_28, gamma = encoder_ret_lns_0_weight, x = input_13)[name = tensor("x_1")]; + tensor x_1 = layer_norm(axes = x_1_axes_0, beta = inner_encoder_ret_lns_0_bias, epsilon = var_46, gamma = inner_encoder_ret_lns_0_weight, x = input_15)[name = tensor("x_1")]; tensor prev_kv_1_begin_0 = const()[name = tensor("prev_kv_1_begin_0"), val = tensor([0, 0, 0, 0, 0])]; tensor prev_kv_1_end_0 = const()[name = tensor("prev_kv_1_end_0"), val = tensor([1, 1, 4, 64, 64])]; tensor prev_kv_1_end_mask_0 = const()[name = tensor("prev_kv_1_end_mask_0"), val = tensor([false, true, true, true, true])]; @@ -239,153 +253,153 @@ program(1.0) tensor prev_scale_1_end_mask_0 = const()[name = tensor("prev_scale_1_end_mask_0"), val = tensor([false, true])]; tensor prev_scale_1_squeeze_mask_0 = const()[name = tensor("prev_scale_1_squeeze_mask_0"), val = tensor([true, false])]; tensor prev_scale_1 = slice_by_index(begin = prev_scale_1_begin_0, end = prev_scale_1_end_0, end_mask = prev_scale_1_end_mask_0, squeeze_mask = prev_scale_1_squeeze_mask_0, x = enc_scale)[name = tensor("prev_scale_1")]; - tensor var_162 = linear(bias = encoder_q_proj_0_bias, weight = encoder_q_proj_0_weight, x = x_1)[name = tensor("linear_3")]; - tensor var_163 = const()[name = tensor("op_163"), val = tensor([1, 2, 4, 64])]; - tensor var_164 = reshape(shape = var_163, x = var_162)[name = tensor("op_164")]; + tensor var_198 = linear(bias = inner_encoder_q_proj_0_bias, weight = inner_encoder_q_proj_0_weight, x = x_1)[name = tensor("linear_3")]; + tensor var_199 = const()[name = tensor("op_199"), val = tensor([1, 2, 4, 64])]; + tensor var_200 = reshape(shape = var_199, x = var_198)[name = tensor("op_200")]; tensor q_1_perm_0 = const()[name = tensor("q_1_perm_0"), val = tensor([0, 2, 1, 3])]; - tensor var_168 = linear(bias = encoder_k_proj_0_bias, weight = encoder_k_proj_0_weight, x = x_1)[name = tensor("linear_4")]; - tensor var_169 = const()[name = tensor("op_169"), val = tensor(0x1p-3)]; - tensor var_170 = mul(x = var_168, y = var_169)[name = tensor("op_170")]; - tensor var_171 = const()[name = tensor("op_171"), val = tensor([1, 2, 4, 64])]; - tensor var_172 = reshape(shape = var_171, x = var_170)[name = tensor("op_172")]; + tensor var_204 = linear(bias = inner_encoder_k_proj_0_bias, weight = inner_encoder_k_proj_0_weight, x = x_1)[name = tensor("linear_4")]; + tensor var_205 = const()[name = tensor("op_205"), val = tensor(0x1p-3)]; + tensor var_206 = mul(x = var_204, y = var_205)[name = tensor("op_206")]; + tensor var_207 = const()[name = tensor("op_207"), val = tensor([1, 2, 4, 64])]; + tensor var_208 = reshape(shape = var_207, x = var_206)[name = tensor("op_208")]; tensor k_1_perm_0 = const()[name = tensor("k_1_perm_0"), val = tensor([0, 2, 1, 3])]; - tensor var_176 = linear(bias = encoder_v_proj_0_bias, weight = encoder_v_proj_0_weight, x = x_1)[name = tensor("linear_5")]; - tensor var_177 = const()[name = tensor("op_177"), val = tensor([1, 2, 4, 64])]; - tensor var_178 = reshape(shape = var_177, x = var_176)[name = tensor("op_178")]; + tensor var_212 = linear(bias = inner_encoder_v_proj_0_bias, weight = inner_encoder_v_proj_0_weight, x = x_1)[name = tensor("linear_5")]; + tensor var_213 = const()[name = tensor("op_213"), val = tensor([1, 2, 4, 64])]; + tensor var_214 = reshape(shape = var_213, x = var_212)[name = tensor("op_214")]; tensor v_1_perm_0 = const()[name = tensor("v_1_perm_0"), val = tensor([0, 2, 1, 3])]; - tensor input_17 = linear(bias = encoder_g_proj_0_bias, weight = encoder_g_proj_0_weight, x = x_1)[name = tensor("linear_6")]; + tensor input_19 = linear(bias = inner_encoder_g_proj_0_bias, weight = inner_encoder_g_proj_0_weight, x = x_1)[name = tensor("linear_6")]; tensor sqrt_s0_1 = sqrt(x = prev_scale_1)[name = tensor("sqrt_s0_1")]; - tensor s_t_1 = add(x = prev_scale_1, y = encoder__t_index)[name = tensor("s_t_1")]; + tensor s_t_1 = add(x = prev_scale_1, y = inner_encoder__t_index)[name = tensor("s_t_1")]; tensor sqrt_s_t_1 = sqrt(x = s_t_1)[name = tensor("sqrt_s_t_1")]; tensor qk_1_transpose_x_1 = const()[name = tensor("qk_1_transpose_x_1"), val = tensor(false)]; tensor qk_1_transpose_y_1 = const()[name = tensor("qk_1_transpose_y_1"), val = tensor(true)]; - tensor k_1 = transpose(perm = k_1_perm_0, x = var_172)[name = tensor("transpose_57")]; - tensor q_1 = transpose(perm = q_1_perm_0, x = var_164)[name = tensor("transpose_58")]; + tensor k_1 = transpose(perm = k_1_perm_0, x = var_208)[name = tensor("transpose_57")]; + tensor q_1 = transpose(perm = q_1_perm_0, x = var_200)[name = tensor("transpose_58")]; tensor qk_1 = matmul(transpose_x = qk_1_transpose_x_1, transpose_y = qk_1_transpose_y_1, x = q_1, y = k_1)[name = tensor("qk_1")]; - tensor var_188 = const()[name = tensor("op_188"), val = tensor([2, 1])]; - tensor var_189 = reshape(shape = var_188, x = sqrt_s_t_1)[name = tensor("op_189")]; - tensor M_1 = real_div(x = encoder__causal_mask, y = var_189)[name = tensor("M_1")]; - tensor var_191 = mul(x = qk_1, y = M_1)[name = tensor("op_191")]; + tensor var_224 = const()[name = tensor("op_224"), val = tensor([2, 1])]; + tensor var_225 = reshape(shape = var_224, x = sqrt_s_t_1)[name = tensor("op_225")]; + tensor M_1 = real_div(x = inner_encoder__causal_mask, y = var_225)[name = tensor("M_1")]; + tensor var_227 = mul(x = qk_1, y = M_1)[name = tensor("op_227")]; tensor inner_1_transpose_x_0 = const()[name = tensor("inner_1_transpose_x_0"), val = tensor(false)]; tensor inner_1_transpose_y_0 = const()[name = tensor("inner_1_transpose_y_0"), val = tensor(false)]; - tensor v_1 = transpose(perm = v_1_perm_0, x = var_178)[name = tensor("transpose_56")]; - tensor inner_1 = matmul(transpose_x = inner_1_transpose_x_0, transpose_y = inner_1_transpose_y_0, x = var_191, y = v_1)[name = tensor("inner_1")]; - tensor var_193_transpose_x_0 = const()[name = tensor("op_193_transpose_x_0"), val = tensor(false)]; - tensor var_193_transpose_y_0 = const()[name = tensor("op_193_transpose_y_0"), val = tensor(false)]; - tensor var_193 = matmul(transpose_x = var_193_transpose_x_0, transpose_y = var_193_transpose_y_0, x = q_1, y = prev_kv_1)[name = tensor("op_193")]; - tensor var_194 = real_div(x = sqrt_s0_1, y = sqrt_s_t_1)[name = tensor("op_194")]; - tensor var_195 = const()[name = tensor("op_195"), val = tensor([1, 1, 2, 1])]; - tensor var_196 = reshape(shape = var_195, x = var_194)[name = tensor("op_196")]; - tensor cross_1 = mul(x = var_193, y = var_196)[name = tensor("cross_1")]; + tensor v_1 = transpose(perm = v_1_perm_0, x = var_214)[name = tensor("transpose_56")]; + tensor inner_1 = matmul(transpose_x = inner_1_transpose_x_0, transpose_y = inner_1_transpose_y_0, x = var_227, y = v_1)[name = tensor("inner_1")]; + tensor var_229_transpose_x_0 = const()[name = tensor("op_229_transpose_x_0"), val = tensor(false)]; + tensor var_229_transpose_y_0 = const()[name = tensor("op_229_transpose_y_0"), val = tensor(false)]; + tensor var_229 = matmul(transpose_x = var_229_transpose_x_0, transpose_y = var_229_transpose_y_0, x = q_1, y = prev_kv_1)[name = tensor("op_229")]; + tensor var_230 = real_div(x = sqrt_s0_1, y = sqrt_s_t_1)[name = tensor("op_230")]; + tensor var_231 = const()[name = tensor("op_231"), val = tensor([1, 1, 2, 1])]; + tensor var_232 = reshape(shape = var_231, x = var_230)[name = tensor("op_232")]; + tensor cross_1 = mul(x = var_229, y = var_232)[name = tensor("cross_1")]; tensor out_1 = add(x = inner_1, y = cross_1)[name = tensor("out_1")]; - tensor var_199 = mul(x = prev_kv_1, y = sqrt_s0_1)[name = tensor("op_199")]; - tensor var_201_transpose_x_1 = const()[name = tensor("op_201_transpose_x_1"), val = tensor(true)]; - tensor var_201_transpose_y_1 = const()[name = tensor("op_201_transpose_y_1"), val = tensor(false)]; - tensor var_201 = matmul(transpose_x = var_201_transpose_x_1, transpose_y = var_201_transpose_y_1, x = k_1, y = v_1)[name = tensor("op_201")]; - tensor new_kv_unnorm_1 = add(x = var_199, y = var_201)[name = tensor("new_kv_unnorm_1")]; - tensor var_203 = const()[name = tensor("op_203"), val = tensor(0x1p+1)]; - tensor new_scale_1 = add(x = prev_scale_1, y = var_203)[name = tensor("new_scale_1")]; - tensor var_205 = sqrt(x = new_scale_1)[name = tensor("op_205")]; - tensor var_206 = real_div(x = new_kv_unnorm_1, y = var_205)[name = tensor("op_206")]; - tensor var_207_perm_0 = const()[name = tensor("op_207_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_235 = mul(x = prev_kv_1, y = sqrt_s0_1)[name = tensor("op_235")]; + tensor var_237_transpose_x_1 = const()[name = tensor("op_237_transpose_x_1"), val = tensor(true)]; + tensor var_237_transpose_y_1 = const()[name = tensor("op_237_transpose_y_1"), val = tensor(false)]; + tensor var_237 = matmul(transpose_x = var_237_transpose_x_1, transpose_y = var_237_transpose_y_1, x = k_1, y = v_1)[name = tensor("op_237")]; + tensor new_kv_unnorm_1 = add(x = var_235, y = var_237)[name = tensor("new_kv_unnorm_1")]; + tensor var_239 = const()[name = tensor("op_239"), val = tensor(0x1p+1)]; + tensor new_scale_1 = add(x = prev_scale_1, y = var_239)[name = tensor("new_scale_1")]; + tensor var_241 = sqrt(x = new_scale_1)[name = tensor("op_241")]; + tensor var_242 = real_div(x = new_kv_unnorm_1, y = var_241)[name = tensor("op_242")]; + tensor var_243_perm_0 = const()[name = tensor("op_243_perm_0"), val = tensor([0, 2, 1, 3])]; tensor out_3_axes_0 = const()[name = tensor("out_3_axes_0"), val = tensor([-1])]; - tensor var_207 = transpose(perm = var_207_perm_0, x = out_1)[name = tensor("transpose_55")]; - tensor out_3 = layer_norm(axes = out_3_axes_0, epsilon = var_17, x = var_207)[name = tensor("out_3")]; - tensor var_211 = const()[name = tensor("op_211"), val = tensor([1, 2, 256])]; - tensor out_5 = reshape(shape = var_211, x = out_3)[name = tensor("out_5")]; - tensor var_213 = silu(x = input_17)[name = tensor("op_213")]; - tensor input_19 = mul(x = var_213, y = out_5)[name = tensor("input_19")]; - tensor ret_out_1 = linear(bias = encoder_out_proj_0_bias, weight = encoder_out_proj_0_weight, x = input_19)[name = tensor("linear_7")]; - tensor x_3 = add(x = input_13, y = ret_out_1)[name = tensor("x_3")]; + tensor var_243 = transpose(perm = var_243_perm_0, x = out_1)[name = tensor("transpose_55")]; + tensor out_3 = layer_norm(axes = out_3_axes_0, epsilon = var_54, x = var_243)[name = tensor("out_3")]; + tensor var_247 = const()[name = tensor("op_247"), val = tensor([1, 2, 256])]; + tensor out_5 = reshape(shape = var_247, x = out_3)[name = tensor("out_5")]; + tensor var_249 = silu(x = input_19)[name = tensor("op_249")]; + tensor input_21 = mul(x = var_249, y = out_5)[name = tensor("input_21")]; + tensor ret_out_1 = linear(bias = inner_encoder_out_proj_0_bias, weight = inner_encoder_out_proj_0_weight, x = input_21)[name = tensor("linear_7")]; + tensor x_3 = add(x = input_15, y = ret_out_1)[name = tensor("x_3")]; tensor window_1_begin_0 = const()[name = tensor("window_1_begin_0"), val = tensor([0, 0, 0, 0])]; tensor window_1_end_0 = const()[name = tensor("window_1_end_0"), val = tensor([1, 1, 16, 256])]; tensor window_1_end_mask_0 = const()[name = tensor("window_1_end_mask_0"), val = tensor([false, true, true, true])]; tensor window_1_squeeze_mask_0 = const()[name = tensor("window_1_squeeze_mask_0"), val = tensor([true, false, false, false])]; tensor window_1 = slice_by_index(begin = window_1_begin_0, end = window_1_end_0, end_mask = window_1_end_mask_0, squeeze_mask = window_1_squeeze_mask_0, x = enc_conv_cache)[name = tensor("window_1")]; - tensor var_221_begin_0 = const()[name = tensor("op_221_begin_0"), val = tensor([0, 0, 0])]; - tensor var_221_end_0 = const()[name = tensor("op_221_end_0"), val = tensor([1, 1, 256])]; - tensor var_221_end_mask_0 = const()[name = tensor("op_221_end_mask_0"), val = tensor([true, false, true])]; - tensor var_221 = slice_by_index(begin = var_221_begin_0, end = var_221_end_0, end_mask = var_221_end_mask_0, x = x_3)[name = tensor("op_221")]; - tensor var_224_begin_0 = const()[name = tensor("op_224_begin_0"), val = tensor([0, 1, 0])]; - tensor var_224_end_0 = const()[name = tensor("op_224_end_0"), val = tensor([1, 16, 256])]; - tensor var_224_end_mask_0 = const()[name = tensor("op_224_end_mask_0"), val = tensor([true, true, true])]; - tensor var_224 = slice_by_index(begin = var_224_begin_0, end = var_224_end_0, end_mask = var_224_end_mask_0, x = window_1)[name = tensor("op_224")]; + tensor var_257_begin_0 = const()[name = tensor("op_257_begin_0"), val = tensor([0, 0, 0])]; + tensor var_257_end_0 = const()[name = tensor("op_257_end_0"), val = tensor([1, 1, 256])]; + tensor var_257_end_mask_0 = const()[name = tensor("op_257_end_mask_0"), val = tensor([true, false, true])]; + tensor var_257 = slice_by_index(begin = var_257_begin_0, end = var_257_end_0, end_mask = var_257_end_mask_0, x = x_3)[name = tensor("op_257")]; + tensor var_260_begin_0 = const()[name = tensor("op_260_begin_0"), val = tensor([0, 1, 0])]; + tensor var_260_end_0 = const()[name = tensor("op_260_end_0"), val = tensor([1, 16, 256])]; + tensor var_260_end_mask_0 = const()[name = tensor("op_260_end_mask_0"), val = tensor([true, true, true])]; + tensor var_260 = slice_by_index(begin = var_260_begin_0, end = var_260_end_0, end_mask = var_260_end_mask_0, x = window_1)[name = tensor("op_260")]; tensor window_3_interleave_0 = const()[name = tensor("window_3_interleave_0"), val = tensor(false)]; - tensor window_3 = concat(axis = var_26, interleave = window_3_interleave_0, values = (var_224, var_221))[name = tensor("window_3")]; - tensor var_229_begin_0 = const()[name = tensor("op_229_begin_0"), val = tensor([0, 1, 0])]; - tensor var_229_end_0 = const()[name = tensor("op_229_end_0"), val = tensor([1, 1, 256])]; - tensor var_229_end_mask_0 = const()[name = tensor("op_229_end_mask_0"), val = tensor([true, true, true])]; - tensor var_229 = slice_by_index(begin = var_229_begin_0, end = var_229_end_0, end_mask = var_229_end_mask_0, x = x_3)[name = tensor("op_229")]; - tensor var_232_begin_0 = const()[name = tensor("op_232_begin_0"), val = tensor([0, 1, 0])]; - tensor var_232_end_0 = const()[name = tensor("op_232_end_0"), val = tensor([1, 16, 256])]; - tensor var_232_end_mask_0 = const()[name = tensor("op_232_end_mask_0"), val = tensor([true, true, true])]; - tensor var_232 = slice_by_index(begin = var_232_begin_0, end = var_232_end_0, end_mask = var_232_end_mask_0, x = window_3)[name = tensor("op_232")]; + tensor window_3 = concat(axis = var_62, interleave = window_3_interleave_0, values = (var_260, var_257))[name = tensor("window_3")]; + tensor var_265_begin_0 = const()[name = tensor("op_265_begin_0"), val = tensor([0, 1, 0])]; + tensor var_265_end_0 = const()[name = tensor("op_265_end_0"), val = tensor([1, 1, 256])]; + tensor var_265_end_mask_0 = const()[name = tensor("op_265_end_mask_0"), val = tensor([true, true, true])]; + tensor var_265 = slice_by_index(begin = var_265_begin_0, end = var_265_end_0, end_mask = var_265_end_mask_0, x = x_3)[name = tensor("op_265")]; + tensor var_268_begin_0 = const()[name = tensor("op_268_begin_0"), val = tensor([0, 1, 0])]; + tensor var_268_end_0 = const()[name = tensor("op_268_end_0"), val = tensor([1, 16, 256])]; + tensor var_268_end_mask_0 = const()[name = tensor("op_268_end_mask_0"), val = tensor([true, true, true])]; + tensor var_268 = slice_by_index(begin = var_268_begin_0, end = var_268_end_0, end_mask = var_268_end_mask_0, x = window_3)[name = tensor("op_268")]; tensor window_5_interleave_0 = const()[name = tensor("window_5_interleave_0"), val = tensor(false)]; - tensor window_5 = concat(axis = var_26, interleave = window_5_interleave_0, values = (var_232, var_229))[name = tensor("window_5")]; - tensor input_21_interleave_0 = const()[name = tensor("input_21_interleave_0"), val = tensor(false)]; - tensor input_21 = concat(axis = var_23, interleave = input_21_interleave_0, values = (window_3, window_5))[name = tensor("input_21")]; + tensor window_5 = concat(axis = var_62, interleave = window_5_interleave_0, values = (var_268, var_265))[name = tensor("window_5")]; + tensor input_23_interleave_0 = const()[name = tensor("input_23_interleave_0"), val = tensor(false)]; + tensor input_23 = concat(axis = var_49, interleave = input_23_interleave_0, values = (window_3, window_5))[name = tensor("input_23")]; tensor x_5_axes_0 = const()[name = tensor("x_5_axes_0"), val = tensor([-1])]; - tensor x_5 = layer_norm(axes = x_5_axes_0, beta = encoder_conv_module_0_sequential_0_bias, epsilon = var_28, gamma = encoder_conv_module_0_sequential_0_weight, x = input_21)[name = tensor("x_5")]; - tensor input_23_perm_0 = const()[name = tensor("input_23_perm_0"), val = tensor([0, 2, 1])]; + tensor x_5 = layer_norm(axes = x_5_axes_0, beta = inner_encoder_conv_module_0_sequential_0_bias, epsilon = var_46, gamma = inner_encoder_conv_module_0_sequential_0_weight, x = input_23)[name = tensor("x_5")]; + tensor input_25_perm_0 = const()[name = tensor("input_25_perm_0"), val = tensor([0, 2, 1])]; tensor inputs_3_pad_type_0 = const()[name = tensor("inputs_3_pad_type_0"), val = tensor("valid")]; tensor inputs_3_strides_0 = const()[name = tensor("inputs_3_strides_0"), val = tensor([1])]; tensor inputs_3_pad_0 = const()[name = tensor("inputs_3_pad_0"), val = tensor([0, 0])]; tensor inputs_3_dilations_0 = const()[name = tensor("inputs_3_dilations_0"), val = tensor([1])]; tensor inputs_3_groups_0 = const()[name = tensor("inputs_3_groups_0"), val = tensor(1)]; - tensor input_23 = transpose(perm = input_23_perm_0, x = x_5)[name = tensor("transpose_54")]; - tensor inputs_3 = conv(bias = encoder_conv_module_0_sequential_2_conv_bias, dilations = inputs_3_dilations_0, groups = inputs_3_groups_0, pad = inputs_3_pad_0, pad_type = inputs_3_pad_type_0, strides = inputs_3_strides_0, weight = encoder_conv_module_0_sequential_2_conv_weight, x = input_23)[name = tensor("inputs_3")]; - tensor var_257_split_sizes_0 = const()[name = tensor("op_257_split_sizes_0"), val = tensor([256, 256])]; - tensor var_257_axis_0 = const()[name = tensor("op_257_axis_0"), val = tensor(1)]; - tensor var_257_0, tensor var_257_1 = split(axis = var_257_axis_0, split_sizes = var_257_split_sizes_0, x = inputs_3)[name = tensor("op_257")]; - tensor var_259 = sigmoid(x = var_257_1)[name = tensor("op_259")]; - tensor inputs_5 = mul(x = var_257_0, y = var_259)[name = tensor("inputs_5")]; + tensor input_25 = transpose(perm = input_25_perm_0, x = x_5)[name = tensor("transpose_54")]; + tensor inputs_3 = conv(bias = inner_encoder_conv_module_0_sequential_2_conv_bias, dilations = inputs_3_dilations_0, groups = inputs_3_groups_0, pad = inputs_3_pad_0, pad_type = inputs_3_pad_type_0, strides = inputs_3_strides_0, weight = inner_encoder_conv_module_0_sequential_2_conv_weight, x = input_25)[name = tensor("inputs_3")]; + tensor var_293_split_sizes_0 = const()[name = tensor("op_293_split_sizes_0"), val = tensor([256, 256])]; + tensor var_293_axis_0 = const()[name = tensor("op_293_axis_0"), val = tensor(1)]; + tensor var_293_0, tensor var_293_1 = split(axis = var_293_axis_0, split_sizes = var_293_split_sizes_0, x = inputs_3)[name = tensor("op_293")]; + tensor var_295 = sigmoid(x = var_293_1)[name = tensor("op_295")]; + tensor inputs_5 = mul(x = var_293_0, y = var_295)[name = tensor("inputs_5")]; tensor outputs_aug_1_pad_type_0 = const()[name = tensor("outputs_aug_1_pad_type_0"), val = tensor("custom")]; tensor outputs_aug_1_pad_0 = const()[name = tensor("outputs_aug_1_pad_0"), val = tensor([15, 15])]; tensor outputs_aug_1_groups_0 = const()[name = tensor("outputs_aug_1_groups_0"), val = tensor(256)]; tensor outputs_aug_1_strides_0 = const()[name = tensor("outputs_aug_1_strides_0"), val = tensor([1])]; tensor outputs_aug_1_dilations_0 = const()[name = tensor("outputs_aug_1_dilations_0"), val = tensor([1])]; - tensor outputs_aug_1 = conv(dilations = outputs_aug_1_dilations_0, groups = outputs_aug_1_groups_0, pad = outputs_aug_1_pad_0, pad_type = outputs_aug_1_pad_type_0, strides = outputs_aug_1_strides_0, weight = encoder_conv_module_0_sequential_4_conv_weight, x = inputs_5)[name = tensor("outputs_aug_1")]; - tensor input_25_begin_0 = const()[name = tensor("input_25_begin_0"), val = tensor([0, 0, 0])]; - tensor input_25_end_0 = const()[name = tensor("input_25_end_0"), val = tensor([2, 256, 16])]; - tensor input_25_end_mask_0 = const()[name = tensor("input_25_end_mask_0"), val = tensor([true, true, false])]; - tensor input_25 = slice_by_index(begin = input_25_begin_0, end = input_25_end_0, end_mask = input_25_end_mask_0, x = outputs_aug_1)[name = tensor("input_25")]; - tensor inputs_7 = batch_norm(beta = encoder_conv_module_0_sequential_5_bias, epsilon = var_28, gamma = encoder_conv_module_0_sequential_5_weight, mean = encoder_conv_module_0_sequential_5_running_mean, variance = encoder_conv_module_0_sequential_5_running_var, x = input_25)[name = tensor("inputs_7")]; - tensor input_27 = silu(x = inputs_7)[name = tensor("input_27")]; - tensor input_29_pad_type_0 = const()[name = tensor("input_29_pad_type_0"), val = tensor("valid")]; - tensor input_29_strides_0 = const()[name = tensor("input_29_strides_0"), val = tensor([1])]; - tensor input_29_pad_0 = const()[name = tensor("input_29_pad_0"), val = tensor([0, 0])]; - tensor input_29_dilations_0 = const()[name = tensor("input_29_dilations_0"), val = tensor([1])]; - tensor input_29_groups_0 = const()[name = tensor("input_29_groups_0"), val = tensor(1)]; - tensor input_29 = conv(bias = encoder_conv_module_0_sequential_7_conv_bias, dilations = input_29_dilations_0, groups = input_29_groups_0, pad = input_29_pad_0, pad_type = input_29_pad_type_0, strides = input_29_strides_0, weight = encoder_conv_module_0_sequential_7_conv_weight, x = input_27)[name = tensor("input_29")]; + tensor outputs_aug_1 = conv(dilations = outputs_aug_1_dilations_0, groups = outputs_aug_1_groups_0, pad = outputs_aug_1_pad_0, pad_type = outputs_aug_1_pad_type_0, strides = outputs_aug_1_strides_0, weight = inner_encoder_conv_module_0_sequential_4_conv_weight, x = inputs_5)[name = tensor("outputs_aug_1")]; + tensor input_27_begin_0 = const()[name = tensor("input_27_begin_0"), val = tensor([0, 0, 0])]; + tensor input_27_end_0 = const()[name = tensor("input_27_end_0"), val = tensor([2, 256, 16])]; + tensor input_27_end_mask_0 = const()[name = tensor("input_27_end_mask_0"), val = tensor([true, true, false])]; + tensor input_27 = slice_by_index(begin = input_27_begin_0, end = input_27_end_0, end_mask = input_27_end_mask_0, x = outputs_aug_1)[name = tensor("input_27")]; + tensor inputs_7 = batch_norm(beta = inner_encoder_conv_module_0_sequential_5_bias, epsilon = var_46, gamma = inner_encoder_conv_module_0_sequential_5_weight, mean = inner_encoder_conv_module_0_sequential_5_running_mean, variance = inner_encoder_conv_module_0_sequential_5_running_var, x = input_27)[name = tensor("inputs_7")]; + tensor input_29 = silu(x = inputs_7)[name = tensor("input_29")]; + tensor input_31_pad_type_0 = const()[name = tensor("input_31_pad_type_0"), val = tensor("valid")]; + tensor input_31_strides_0 = const()[name = tensor("input_31_strides_0"), val = tensor([1])]; + tensor input_31_pad_0 = const()[name = tensor("input_31_pad_0"), val = tensor([0, 0])]; + tensor input_31_dilations_0 = const()[name = tensor("input_31_dilations_0"), val = tensor([1])]; + tensor input_31_groups_0 = const()[name = tensor("input_31_groups_0"), val = tensor(1)]; + tensor input_31 = conv(bias = inner_encoder_conv_module_0_sequential_7_conv_bias, dilations = input_31_dilations_0, groups = input_31_groups_0, pad = input_31_pad_0, pad_type = input_31_pad_type_0, strides = input_31_strides_0, weight = inner_encoder_conv_module_0_sequential_7_conv_weight, x = input_29)[name = tensor("input_31")]; tensor conv_out_1_perm_0 = const()[name = tensor("conv_out_1_perm_0"), val = tensor([0, 2, 1])]; - tensor var_290_begin_0 = const()[name = tensor("op_290_begin_0"), val = tensor([0, -1, 0])]; - tensor var_290_end_0 = const()[name = tensor("op_290_end_0"), val = tensor([2, 16, 256])]; - tensor var_290_end_mask_0 = const()[name = tensor("op_290_end_mask_0"), val = tensor([true, true, true])]; - tensor conv_out_1 = transpose(perm = conv_out_1_perm_0, x = input_29)[name = tensor("transpose_53")]; - tensor var_290 = slice_by_index(begin = var_290_begin_0, end = var_290_end_0, end_mask = var_290_end_mask_0, x = conv_out_1)[name = tensor("op_290")]; - tensor var_292_perm_0 = const()[name = tensor("op_292_perm_0"), val = tensor([1, 0, 2])]; - tensor var_292 = transpose(perm = var_292_perm_0, x = var_290)[name = tensor("transpose_52")]; - tensor input_31 = add(x = x_3, y = var_292)[name = tensor("input_31")]; - tensor input_33_axes_0 = const()[name = tensor("input_33_axes_0"), val = tensor([-1])]; - tensor input_33 = layer_norm(axes = input_33_axes_0, beta = encoder_ffn2_0_module_sequential_0_bias, epsilon = var_28, gamma = encoder_ffn2_0_module_sequential_0_weight, x = input_31)[name = tensor("input_33")]; - tensor inputs_9 = linear(bias = encoder_ffn2_0_module_sequential_1_linear_bias, weight = encoder_ffn2_0_module_sequential_1_linear_weight, x = input_33)[name = tensor("linear_8")]; - tensor input_35 = silu(x = inputs_9)[name = tensor("input_35")]; - tensor input_39 = linear(bias = encoder_ffn2_0_module_sequential_4_linear_bias, weight = encoder_ffn2_0_module_sequential_4_linear_weight, x = input_35)[name = tensor("linear_9")]; - tensor var_315 = const()[name = tensor("op_315"), val = tensor(0x1p-1)]; - tensor var_316 = mul(x = input_39, y = var_315)[name = tensor("op_316")]; - tensor input_41 = add(x = var_316, y = input_31)[name = tensor("input_41")]; - tensor input_43_axes_0 = const()[name = tensor("input_43_axes_0"), val = tensor([-1])]; - tensor input_43 = layer_norm(axes = input_43_axes_0, beta = encoder_layer_norm_0_bias, epsilon = var_28, gamma = encoder_layer_norm_0_weight, x = input_41)[name = tensor("input_43")]; + tensor var_326_begin_0 = const()[name = tensor("op_326_begin_0"), val = tensor([0, -1, 0])]; + tensor var_326_end_0 = const()[name = tensor("op_326_end_0"), val = tensor([2, 16, 256])]; + tensor var_326_end_mask_0 = const()[name = tensor("op_326_end_mask_0"), val = tensor([true, true, true])]; + tensor conv_out_1 = transpose(perm = conv_out_1_perm_0, x = input_31)[name = tensor("transpose_53")]; + tensor var_326 = slice_by_index(begin = var_326_begin_0, end = var_326_end_0, end_mask = var_326_end_mask_0, x = conv_out_1)[name = tensor("op_326")]; + tensor var_328_perm_0 = const()[name = tensor("op_328_perm_0"), val = tensor([1, 0, 2])]; + tensor var_328 = transpose(perm = var_328_perm_0, x = var_326)[name = tensor("transpose_52")]; + tensor input_33 = add(x = x_3, y = var_328)[name = tensor("input_33")]; + tensor input_35_axes_0 = const()[name = tensor("input_35_axes_0"), val = tensor([-1])]; + tensor input_35 = layer_norm(axes = input_35_axes_0, beta = inner_encoder_ffn2_0_module_sequential_0_bias, epsilon = var_46, gamma = inner_encoder_ffn2_0_module_sequential_0_weight, x = input_33)[name = tensor("input_35")]; + tensor inputs_9 = linear(bias = inner_encoder_ffn2_0_module_sequential_1_linear_bias, weight = inner_encoder_ffn2_0_module_sequential_1_linear_weight, x = input_35)[name = tensor("linear_8")]; + tensor input_37 = silu(x = inputs_9)[name = tensor("input_37")]; + tensor input_41 = linear(bias = inner_encoder_ffn2_0_module_sequential_4_linear_bias, weight = inner_encoder_ffn2_0_module_sequential_4_linear_weight, x = input_37)[name = tensor("linear_9")]; + tensor var_351 = const()[name = tensor("op_351"), val = tensor(0x1p-1)]; + tensor var_352 = mul(x = input_41, y = var_351)[name = tensor("op_352")]; + tensor input_43 = add(x = var_352, y = input_33)[name = tensor("input_43")]; tensor input_45_axes_0 = const()[name = tensor("input_45_axes_0"), val = tensor([-1])]; - tensor input_45 = layer_norm(axes = input_45_axes_0, beta = encoder_ffn1_1_module_sequential_0_bias, epsilon = var_28, gamma = encoder_ffn1_1_module_sequential_0_weight, x = input_43)[name = tensor("input_45")]; - tensor inputs_11 = linear(bias = encoder_ffn1_1_module_sequential_1_linear_bias, weight = encoder_ffn1_1_module_sequential_1_linear_weight, x = input_45)[name = tensor("linear_10")]; - tensor input_47 = silu(x = inputs_11)[name = tensor("input_47")]; - tensor input_51 = linear(bias = encoder_ffn1_1_module_sequential_4_linear_bias, weight = encoder_ffn1_1_module_sequential_4_linear_weight, x = input_47)[name = tensor("linear_11")]; - tensor var_345 = const()[name = tensor("op_345"), val = tensor(0x1p-1)]; - tensor var_346 = mul(x = input_51, y = var_345)[name = tensor("op_346")]; - tensor input_53 = add(x = var_346, y = input_43)[name = tensor("input_53")]; + tensor input_45 = layer_norm(axes = input_45_axes_0, beta = inner_encoder_layer_norm_0_bias, epsilon = var_46, gamma = inner_encoder_layer_norm_0_weight, x = input_43)[name = tensor("input_45")]; + tensor input_47_axes_0 = const()[name = tensor("input_47_axes_0"), val = tensor([-1])]; + tensor input_47 = layer_norm(axes = input_47_axes_0, beta = inner_encoder_ffn1_1_module_sequential_0_bias, epsilon = var_46, gamma = inner_encoder_ffn1_1_module_sequential_0_weight, x = input_45)[name = tensor("input_47")]; + tensor inputs_11 = linear(bias = inner_encoder_ffn1_1_module_sequential_1_linear_bias, weight = inner_encoder_ffn1_1_module_sequential_1_linear_weight, x = input_47)[name = tensor("linear_10")]; + tensor input_49 = silu(x = inputs_11)[name = tensor("input_49")]; + tensor input_53 = linear(bias = inner_encoder_ffn1_1_module_sequential_4_linear_bias, weight = inner_encoder_ffn1_1_module_sequential_4_linear_weight, x = input_49)[name = tensor("linear_11")]; + tensor var_381 = const()[name = tensor("op_381"), val = tensor(0x1p-1)]; + tensor var_382 = mul(x = input_53, y = var_381)[name = tensor("op_382")]; + tensor input_55 = add(x = var_382, y = input_45)[name = tensor("input_55")]; tensor x_7_axes_0 = const()[name = tensor("x_7_axes_0"), val = tensor([-1])]; - tensor x_7 = layer_norm(axes = x_7_axes_0, beta = encoder_ret_lns_1_bias, epsilon = var_28, gamma = encoder_ret_lns_1_weight, x = input_53)[name = tensor("x_7")]; + tensor x_7 = layer_norm(axes = x_7_axes_0, beta = inner_encoder_ret_lns_1_bias, epsilon = var_46, gamma = inner_encoder_ret_lns_1_weight, x = input_55)[name = tensor("x_7")]; tensor prev_kv_3_begin_0 = const()[name = tensor("prev_kv_3_begin_0"), val = tensor([1, 0, 0, 0, 0])]; tensor prev_kv_3_end_0 = const()[name = tensor("prev_kv_3_end_0"), val = tensor([2, 1, 4, 64, 64])]; tensor prev_kv_3_end_mask_0 = const()[name = tensor("prev_kv_3_end_mask_0"), val = tensor([false, true, true, true, true])]; @@ -396,153 +410,153 @@ program(1.0) tensor prev_scale_3_end_mask_0 = const()[name = tensor("prev_scale_3_end_mask_0"), val = tensor([false, true])]; tensor prev_scale_3_squeeze_mask_0 = const()[name = tensor("prev_scale_3_squeeze_mask_0"), val = tensor([true, false])]; tensor prev_scale_3 = slice_by_index(begin = prev_scale_3_begin_0, end = prev_scale_3_end_0, end_mask = prev_scale_3_end_mask_0, squeeze_mask = prev_scale_3_squeeze_mask_0, x = enc_scale)[name = tensor("prev_scale_3")]; - tensor var_360 = linear(bias = encoder_q_proj_1_bias, weight = encoder_q_proj_1_weight, x = x_7)[name = tensor("linear_12")]; - tensor var_361 = const()[name = tensor("op_361"), val = tensor([1, 2, 4, 64])]; - tensor var_362 = reshape(shape = var_361, x = var_360)[name = tensor("op_362")]; + tensor var_396 = linear(bias = inner_encoder_q_proj_1_bias, weight = inner_encoder_q_proj_1_weight, x = x_7)[name = tensor("linear_12")]; + tensor var_397 = const()[name = tensor("op_397"), val = tensor([1, 2, 4, 64])]; + tensor var_398 = reshape(shape = var_397, x = var_396)[name = tensor("op_398")]; tensor q_3_perm_0 = const()[name = tensor("q_3_perm_0"), val = tensor([0, 2, 1, 3])]; - tensor var_366 = linear(bias = encoder_k_proj_1_bias, weight = encoder_k_proj_1_weight, x = x_7)[name = tensor("linear_13")]; - tensor var_367 = const()[name = tensor("op_367"), val = tensor(0x1p-3)]; - tensor var_368 = mul(x = var_366, y = var_367)[name = tensor("op_368")]; - tensor var_369 = const()[name = tensor("op_369"), val = tensor([1, 2, 4, 64])]; - tensor var_370 = reshape(shape = var_369, x = var_368)[name = tensor("op_370")]; + tensor var_402 = linear(bias = inner_encoder_k_proj_1_bias, weight = inner_encoder_k_proj_1_weight, x = x_7)[name = tensor("linear_13")]; + tensor var_403 = const()[name = tensor("op_403"), val = tensor(0x1p-3)]; + tensor var_404 = mul(x = var_402, y = var_403)[name = tensor("op_404")]; + tensor var_405 = const()[name = tensor("op_405"), val = tensor([1, 2, 4, 64])]; + tensor var_406 = reshape(shape = var_405, x = var_404)[name = tensor("op_406")]; tensor k_3_perm_0 = const()[name = tensor("k_3_perm_0"), val = tensor([0, 2, 1, 3])]; - tensor var_374 = linear(bias = encoder_v_proj_1_bias, weight = encoder_v_proj_1_weight, x = x_7)[name = tensor("linear_14")]; - tensor var_375 = const()[name = tensor("op_375"), val = tensor([1, 2, 4, 64])]; - tensor var_376 = reshape(shape = var_375, x = var_374)[name = tensor("op_376")]; + tensor var_410 = linear(bias = inner_encoder_v_proj_1_bias, weight = inner_encoder_v_proj_1_weight, x = x_7)[name = tensor("linear_14")]; + tensor var_411 = const()[name = tensor("op_411"), val = tensor([1, 2, 4, 64])]; + tensor var_412 = reshape(shape = var_411, x = var_410)[name = tensor("op_412")]; tensor v_3_perm_0 = const()[name = tensor("v_3_perm_0"), val = tensor([0, 2, 1, 3])]; - tensor input_57 = linear(bias = encoder_g_proj_1_bias, weight = encoder_g_proj_1_weight, x = x_7)[name = tensor("linear_15")]; + tensor input_59 = linear(bias = inner_encoder_g_proj_1_bias, weight = inner_encoder_g_proj_1_weight, x = x_7)[name = tensor("linear_15")]; tensor sqrt_s0_3 = sqrt(x = prev_scale_3)[name = tensor("sqrt_s0_3")]; - tensor s_t_3 = add(x = prev_scale_3, y = encoder__t_index)[name = tensor("s_t_3")]; + tensor s_t_3 = add(x = prev_scale_3, y = inner_encoder__t_index)[name = tensor("s_t_3")]; tensor sqrt_s_t_3 = sqrt(x = s_t_3)[name = tensor("sqrt_s_t_3")]; tensor qk_3_transpose_x_1 = const()[name = tensor("qk_3_transpose_x_1"), val = tensor(false)]; tensor qk_3_transpose_y_1 = const()[name = tensor("qk_3_transpose_y_1"), val = tensor(true)]; - tensor k_3 = transpose(perm = k_3_perm_0, x = var_370)[name = tensor("transpose_50")]; - tensor q_3 = transpose(perm = q_3_perm_0, x = var_362)[name = tensor("transpose_51")]; + tensor k_3 = transpose(perm = k_3_perm_0, x = var_406)[name = tensor("transpose_50")]; + tensor q_3 = transpose(perm = q_3_perm_0, x = var_398)[name = tensor("transpose_51")]; tensor qk_3 = matmul(transpose_x = qk_3_transpose_x_1, transpose_y = qk_3_transpose_y_1, x = q_3, y = k_3)[name = tensor("qk_3")]; - tensor var_386 = const()[name = tensor("op_386"), val = tensor([2, 1])]; - tensor var_387 = reshape(shape = var_386, x = sqrt_s_t_3)[name = tensor("op_387")]; - tensor M_3 = real_div(x = encoder__causal_mask, y = var_387)[name = tensor("M_3")]; - tensor var_389 = mul(x = qk_3, y = M_3)[name = tensor("op_389")]; + tensor var_422 = const()[name = tensor("op_422"), val = tensor([2, 1])]; + tensor var_423 = reshape(shape = var_422, x = sqrt_s_t_3)[name = tensor("op_423")]; + tensor M_3 = real_div(x = inner_encoder__causal_mask, y = var_423)[name = tensor("M_3")]; + tensor var_425 = mul(x = qk_3, y = M_3)[name = tensor("op_425")]; tensor inner_3_transpose_x_0 = const()[name = tensor("inner_3_transpose_x_0"), val = tensor(false)]; tensor inner_3_transpose_y_0 = const()[name = tensor("inner_3_transpose_y_0"), val = tensor(false)]; - tensor v_3 = transpose(perm = v_3_perm_0, x = var_376)[name = tensor("transpose_49")]; - tensor inner_3 = matmul(transpose_x = inner_3_transpose_x_0, transpose_y = inner_3_transpose_y_0, x = var_389, y = v_3)[name = tensor("inner_3")]; - tensor var_391_transpose_x_0 = const()[name = tensor("op_391_transpose_x_0"), val = tensor(false)]; - tensor var_391_transpose_y_0 = const()[name = tensor("op_391_transpose_y_0"), val = tensor(false)]; - tensor var_391 = matmul(transpose_x = var_391_transpose_x_0, transpose_y = var_391_transpose_y_0, x = q_3, y = prev_kv_3)[name = tensor("op_391")]; - tensor var_392 = real_div(x = sqrt_s0_3, y = sqrt_s_t_3)[name = tensor("op_392")]; - tensor var_393 = const()[name = tensor("op_393"), val = tensor([1, 1, 2, 1])]; - tensor var_394 = reshape(shape = var_393, x = var_392)[name = tensor("op_394")]; - tensor cross_3 = mul(x = var_391, y = var_394)[name = tensor("cross_3")]; + tensor v_3 = transpose(perm = v_3_perm_0, x = var_412)[name = tensor("transpose_49")]; + tensor inner_3 = matmul(transpose_x = inner_3_transpose_x_0, transpose_y = inner_3_transpose_y_0, x = var_425, y = v_3)[name = tensor("inner_3")]; + tensor var_427_transpose_x_0 = const()[name = tensor("op_427_transpose_x_0"), val = tensor(false)]; + tensor var_427_transpose_y_0 = const()[name = tensor("op_427_transpose_y_0"), val = tensor(false)]; + tensor var_427 = matmul(transpose_x = var_427_transpose_x_0, transpose_y = var_427_transpose_y_0, x = q_3, y = prev_kv_3)[name = tensor("op_427")]; + tensor var_428 = real_div(x = sqrt_s0_3, y = sqrt_s_t_3)[name = tensor("op_428")]; + tensor var_429 = const()[name = tensor("op_429"), val = tensor([1, 1, 2, 1])]; + tensor var_430 = reshape(shape = var_429, x = var_428)[name = tensor("op_430")]; + tensor cross_3 = mul(x = var_427, y = var_430)[name = tensor("cross_3")]; tensor out_7 = add(x = inner_3, y = cross_3)[name = tensor("out_7")]; - tensor var_397 = mul(x = prev_kv_3, y = sqrt_s0_3)[name = tensor("op_397")]; - tensor var_399_transpose_x_1 = const()[name = tensor("op_399_transpose_x_1"), val = tensor(true)]; - tensor var_399_transpose_y_1 = const()[name = tensor("op_399_transpose_y_1"), val = tensor(false)]; - tensor var_399 = matmul(transpose_x = var_399_transpose_x_1, transpose_y = var_399_transpose_y_1, x = k_3, y = v_3)[name = tensor("op_399")]; - tensor new_kv_unnorm_3 = add(x = var_397, y = var_399)[name = tensor("new_kv_unnorm_3")]; - tensor var_401 = const()[name = tensor("op_401"), val = tensor(0x1p+1)]; - tensor new_scale_3 = add(x = prev_scale_3, y = var_401)[name = tensor("new_scale_3")]; - tensor var_403 = sqrt(x = new_scale_3)[name = tensor("op_403")]; - tensor var_404 = real_div(x = new_kv_unnorm_3, y = var_403)[name = tensor("op_404")]; - tensor var_405_perm_0 = const()[name = tensor("op_405_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_433 = mul(x = prev_kv_3, y = sqrt_s0_3)[name = tensor("op_433")]; + tensor var_435_transpose_x_1 = const()[name = tensor("op_435_transpose_x_1"), val = tensor(true)]; + tensor var_435_transpose_y_1 = const()[name = tensor("op_435_transpose_y_1"), val = tensor(false)]; + tensor var_435 = matmul(transpose_x = var_435_transpose_x_1, transpose_y = var_435_transpose_y_1, x = k_3, y = v_3)[name = tensor("op_435")]; + tensor new_kv_unnorm_3 = add(x = var_433, y = var_435)[name = tensor("new_kv_unnorm_3")]; + tensor var_437 = const()[name = tensor("op_437"), val = tensor(0x1p+1)]; + tensor new_scale_3 = add(x = prev_scale_3, y = var_437)[name = tensor("new_scale_3")]; + tensor var_439 = sqrt(x = new_scale_3)[name = tensor("op_439")]; + tensor var_440 = real_div(x = new_kv_unnorm_3, y = var_439)[name = tensor("op_440")]; + tensor var_441_perm_0 = const()[name = tensor("op_441_perm_0"), val = tensor([0, 2, 1, 3])]; tensor out_9_axes_0 = const()[name = tensor("out_9_axes_0"), val = tensor([-1])]; - tensor var_405 = transpose(perm = var_405_perm_0, x = out_7)[name = tensor("transpose_48")]; - tensor out_9 = layer_norm(axes = out_9_axes_0, epsilon = var_17, x = var_405)[name = tensor("out_9")]; - tensor var_409 = const()[name = tensor("op_409"), val = tensor([1, 2, 256])]; - tensor out_11 = reshape(shape = var_409, x = out_9)[name = tensor("out_11")]; - tensor var_411 = silu(x = input_57)[name = tensor("op_411")]; - tensor input_59 = mul(x = var_411, y = out_11)[name = tensor("input_59")]; - tensor ret_out_3 = linear(bias = encoder_out_proj_1_bias, weight = encoder_out_proj_1_weight, x = input_59)[name = tensor("linear_16")]; - tensor x_9 = add(x = input_53, y = ret_out_3)[name = tensor("x_9")]; + tensor var_441 = transpose(perm = var_441_perm_0, x = out_7)[name = tensor("transpose_48")]; + tensor out_9 = layer_norm(axes = out_9_axes_0, epsilon = var_54, x = var_441)[name = tensor("out_9")]; + tensor var_445 = const()[name = tensor("op_445"), val = tensor([1, 2, 256])]; + tensor out_11 = reshape(shape = var_445, x = out_9)[name = tensor("out_11")]; + tensor var_447 = silu(x = input_59)[name = tensor("op_447")]; + tensor input_61 = mul(x = var_447, y = out_11)[name = tensor("input_61")]; + tensor ret_out_3 = linear(bias = inner_encoder_out_proj_1_bias, weight = inner_encoder_out_proj_1_weight, x = input_61)[name = tensor("linear_16")]; + tensor x_9 = add(x = input_55, y = ret_out_3)[name = tensor("x_9")]; tensor window_7_begin_0 = const()[name = tensor("window_7_begin_0"), val = tensor([1, 0, 0, 0])]; tensor window_7_end_0 = const()[name = tensor("window_7_end_0"), val = tensor([2, 1, 16, 256])]; tensor window_7_end_mask_0 = const()[name = tensor("window_7_end_mask_0"), val = tensor([false, true, true, true])]; tensor window_7_squeeze_mask_0 = const()[name = tensor("window_7_squeeze_mask_0"), val = tensor([true, false, false, false])]; tensor window_7 = slice_by_index(begin = window_7_begin_0, end = window_7_end_0, end_mask = window_7_end_mask_0, squeeze_mask = window_7_squeeze_mask_0, x = enc_conv_cache)[name = tensor("window_7")]; - tensor var_419_begin_0 = const()[name = tensor("op_419_begin_0"), val = tensor([0, 0, 0])]; - tensor var_419_end_0 = const()[name = tensor("op_419_end_0"), val = tensor([1, 1, 256])]; - tensor var_419_end_mask_0 = const()[name = tensor("op_419_end_mask_0"), val = tensor([true, false, true])]; - tensor var_419 = slice_by_index(begin = var_419_begin_0, end = var_419_end_0, end_mask = var_419_end_mask_0, x = x_9)[name = tensor("op_419")]; - tensor var_422_begin_0 = const()[name = tensor("op_422_begin_0"), val = tensor([0, 1, 0])]; - tensor var_422_end_0 = const()[name = tensor("op_422_end_0"), val = tensor([1, 16, 256])]; - tensor var_422_end_mask_0 = const()[name = tensor("op_422_end_mask_0"), val = tensor([true, true, true])]; - tensor var_422 = slice_by_index(begin = var_422_begin_0, end = var_422_end_0, end_mask = var_422_end_mask_0, x = window_7)[name = tensor("op_422")]; + tensor var_455_begin_0 = const()[name = tensor("op_455_begin_0"), val = tensor([0, 0, 0])]; + tensor var_455_end_0 = const()[name = tensor("op_455_end_0"), val = tensor([1, 1, 256])]; + tensor var_455_end_mask_0 = const()[name = tensor("op_455_end_mask_0"), val = tensor([true, false, true])]; + tensor var_455 = slice_by_index(begin = var_455_begin_0, end = var_455_end_0, end_mask = var_455_end_mask_0, x = x_9)[name = tensor("op_455")]; + tensor var_458_begin_0 = const()[name = tensor("op_458_begin_0"), val = tensor([0, 1, 0])]; + tensor var_458_end_0 = const()[name = tensor("op_458_end_0"), val = tensor([1, 16, 256])]; + tensor var_458_end_mask_0 = const()[name = tensor("op_458_end_mask_0"), val = tensor([true, true, true])]; + tensor var_458 = slice_by_index(begin = var_458_begin_0, end = var_458_end_0, end_mask = var_458_end_mask_0, x = window_7)[name = tensor("op_458")]; tensor window_9_interleave_0 = const()[name = tensor("window_9_interleave_0"), val = tensor(false)]; - tensor window_9 = concat(axis = var_26, interleave = window_9_interleave_0, values = (var_422, var_419))[name = tensor("window_9")]; - tensor var_427_begin_0 = const()[name = tensor("op_427_begin_0"), val = tensor([0, 1, 0])]; - tensor var_427_end_0 = const()[name = tensor("op_427_end_0"), val = tensor([1, 1, 256])]; - tensor var_427_end_mask_0 = const()[name = tensor("op_427_end_mask_0"), val = tensor([true, true, true])]; - tensor var_427 = slice_by_index(begin = var_427_begin_0, end = var_427_end_0, end_mask = var_427_end_mask_0, x = x_9)[name = tensor("op_427")]; - tensor var_430_begin_0 = const()[name = tensor("op_430_begin_0"), val = tensor([0, 1, 0])]; - tensor var_430_end_0 = const()[name = tensor("op_430_end_0"), val = tensor([1, 16, 256])]; - tensor var_430_end_mask_0 = const()[name = tensor("op_430_end_mask_0"), val = tensor([true, true, true])]; - tensor var_430 = slice_by_index(begin = var_430_begin_0, end = var_430_end_0, end_mask = var_430_end_mask_0, x = window_9)[name = tensor("op_430")]; + tensor window_9 = concat(axis = var_62, interleave = window_9_interleave_0, values = (var_458, var_455))[name = tensor("window_9")]; + tensor var_463_begin_0 = const()[name = tensor("op_463_begin_0"), val = tensor([0, 1, 0])]; + tensor var_463_end_0 = const()[name = tensor("op_463_end_0"), val = tensor([1, 1, 256])]; + tensor var_463_end_mask_0 = const()[name = tensor("op_463_end_mask_0"), val = tensor([true, true, true])]; + tensor var_463 = slice_by_index(begin = var_463_begin_0, end = var_463_end_0, end_mask = var_463_end_mask_0, x = x_9)[name = tensor("op_463")]; + tensor var_466_begin_0 = const()[name = tensor("op_466_begin_0"), val = tensor([0, 1, 0])]; + tensor var_466_end_0 = const()[name = tensor("op_466_end_0"), val = tensor([1, 16, 256])]; + tensor var_466_end_mask_0 = const()[name = tensor("op_466_end_mask_0"), val = tensor([true, true, true])]; + tensor var_466 = slice_by_index(begin = var_466_begin_0, end = var_466_end_0, end_mask = var_466_end_mask_0, x = window_9)[name = tensor("op_466")]; tensor window_11_interleave_0 = const()[name = tensor("window_11_interleave_0"), val = tensor(false)]; - tensor window_11 = concat(axis = var_26, interleave = window_11_interleave_0, values = (var_430, var_427))[name = tensor("window_11")]; - tensor input_61_interleave_0 = const()[name = tensor("input_61_interleave_0"), val = tensor(false)]; - tensor input_61 = concat(axis = var_23, interleave = input_61_interleave_0, values = (window_9, window_11))[name = tensor("input_61")]; + tensor window_11 = concat(axis = var_62, interleave = window_11_interleave_0, values = (var_466, var_463))[name = tensor("window_11")]; + tensor input_63_interleave_0 = const()[name = tensor("input_63_interleave_0"), val = tensor(false)]; + tensor input_63 = concat(axis = var_49, interleave = input_63_interleave_0, values = (window_9, window_11))[name = tensor("input_63")]; tensor x_11_axes_0 = const()[name = tensor("x_11_axes_0"), val = tensor([-1])]; - tensor x_11 = layer_norm(axes = x_11_axes_0, beta = encoder_conv_module_1_sequential_0_bias, epsilon = var_28, gamma = encoder_conv_module_1_sequential_0_weight, x = input_61)[name = tensor("x_11")]; - tensor input_63_perm_0 = const()[name = tensor("input_63_perm_0"), val = tensor([0, 2, 1])]; + tensor x_11 = layer_norm(axes = x_11_axes_0, beta = inner_encoder_conv_module_1_sequential_0_bias, epsilon = var_46, gamma = inner_encoder_conv_module_1_sequential_0_weight, x = input_63)[name = tensor("x_11")]; + tensor input_65_perm_0 = const()[name = tensor("input_65_perm_0"), val = tensor([0, 2, 1])]; tensor inputs_13_pad_type_0 = const()[name = tensor("inputs_13_pad_type_0"), val = tensor("valid")]; tensor inputs_13_strides_0 = const()[name = tensor("inputs_13_strides_0"), val = tensor([1])]; tensor inputs_13_pad_0 = const()[name = tensor("inputs_13_pad_0"), val = tensor([0, 0])]; tensor inputs_13_dilations_0 = const()[name = tensor("inputs_13_dilations_0"), val = tensor([1])]; tensor inputs_13_groups_0 = const()[name = tensor("inputs_13_groups_0"), val = tensor(1)]; - tensor input_63 = transpose(perm = input_63_perm_0, x = x_11)[name = tensor("transpose_47")]; - tensor inputs_13 = conv(bias = encoder_conv_module_1_sequential_2_conv_bias, dilations = inputs_13_dilations_0, groups = inputs_13_groups_0, pad = inputs_13_pad_0, pad_type = inputs_13_pad_type_0, strides = inputs_13_strides_0, weight = encoder_conv_module_1_sequential_2_conv_weight, x = input_63)[name = tensor("inputs_13")]; - tensor var_455_split_sizes_0 = const()[name = tensor("op_455_split_sizes_0"), val = tensor([256, 256])]; - tensor var_455_axis_0 = const()[name = tensor("op_455_axis_0"), val = tensor(1)]; - tensor var_455_0, tensor var_455_1 = split(axis = var_455_axis_0, split_sizes = var_455_split_sizes_0, x = inputs_13)[name = tensor("op_455")]; - tensor var_457 = sigmoid(x = var_455_1)[name = tensor("op_457")]; - tensor inputs_15 = mul(x = var_455_0, y = var_457)[name = tensor("inputs_15")]; + tensor input_65 = transpose(perm = input_65_perm_0, x = x_11)[name = tensor("transpose_47")]; + tensor inputs_13 = conv(bias = inner_encoder_conv_module_1_sequential_2_conv_bias, dilations = inputs_13_dilations_0, groups = inputs_13_groups_0, pad = inputs_13_pad_0, pad_type = inputs_13_pad_type_0, strides = inputs_13_strides_0, weight = inner_encoder_conv_module_1_sequential_2_conv_weight, x = input_65)[name = tensor("inputs_13")]; + tensor var_491_split_sizes_0 = const()[name = tensor("op_491_split_sizes_0"), val = tensor([256, 256])]; + tensor var_491_axis_0 = const()[name = tensor("op_491_axis_0"), val = tensor(1)]; + tensor var_491_0, tensor var_491_1 = split(axis = var_491_axis_0, split_sizes = var_491_split_sizes_0, x = inputs_13)[name = tensor("op_491")]; + tensor var_493 = sigmoid(x = var_491_1)[name = tensor("op_493")]; + tensor inputs_15 = mul(x = var_491_0, y = var_493)[name = tensor("inputs_15")]; tensor outputs_aug_3_pad_type_0 = const()[name = tensor("outputs_aug_3_pad_type_0"), val = tensor("custom")]; tensor outputs_aug_3_pad_0 = const()[name = tensor("outputs_aug_3_pad_0"), val = tensor([15, 15])]; tensor outputs_aug_3_groups_0 = const()[name = tensor("outputs_aug_3_groups_0"), val = tensor(256)]; tensor outputs_aug_3_strides_0 = const()[name = tensor("outputs_aug_3_strides_0"), val = tensor([1])]; tensor outputs_aug_3_dilations_0 = const()[name = tensor("outputs_aug_3_dilations_0"), val = tensor([1])]; - tensor outputs_aug_3 = conv(dilations = outputs_aug_3_dilations_0, groups = outputs_aug_3_groups_0, pad = outputs_aug_3_pad_0, pad_type = outputs_aug_3_pad_type_0, strides = outputs_aug_3_strides_0, weight = encoder_conv_module_1_sequential_4_conv_weight, x = inputs_15)[name = tensor("outputs_aug_3")]; - tensor input_65_begin_0 = const()[name = tensor("input_65_begin_0"), val = tensor([0, 0, 0])]; - tensor input_65_end_0 = const()[name = tensor("input_65_end_0"), val = tensor([2, 256, 16])]; - tensor input_65_end_mask_0 = const()[name = tensor("input_65_end_mask_0"), val = tensor([true, true, false])]; - tensor input_65 = slice_by_index(begin = input_65_begin_0, end = input_65_end_0, end_mask = input_65_end_mask_0, x = outputs_aug_3)[name = tensor("input_65")]; - tensor inputs_17 = batch_norm(beta = encoder_conv_module_1_sequential_5_bias, epsilon = var_28, gamma = encoder_conv_module_1_sequential_5_weight, mean = encoder_conv_module_1_sequential_5_running_mean, variance = encoder_conv_module_1_sequential_5_running_var, x = input_65)[name = tensor("inputs_17")]; - tensor input_67 = silu(x = inputs_17)[name = tensor("input_67")]; - tensor input_69_pad_type_0 = const()[name = tensor("input_69_pad_type_0"), val = tensor("valid")]; - tensor input_69_strides_0 = const()[name = tensor("input_69_strides_0"), val = tensor([1])]; - tensor input_69_pad_0 = const()[name = tensor("input_69_pad_0"), val = tensor([0, 0])]; - tensor input_69_dilations_0 = const()[name = tensor("input_69_dilations_0"), val = tensor([1])]; - tensor input_69_groups_0 = const()[name = tensor("input_69_groups_0"), val = tensor(1)]; - tensor input_69 = conv(bias = encoder_conv_module_1_sequential_7_conv_bias, dilations = input_69_dilations_0, groups = input_69_groups_0, pad = input_69_pad_0, pad_type = input_69_pad_type_0, strides = input_69_strides_0, weight = encoder_conv_module_1_sequential_7_conv_weight, x = input_67)[name = tensor("input_69")]; + tensor outputs_aug_3 = conv(dilations = outputs_aug_3_dilations_0, groups = outputs_aug_3_groups_0, pad = outputs_aug_3_pad_0, pad_type = outputs_aug_3_pad_type_0, strides = outputs_aug_3_strides_0, weight = inner_encoder_conv_module_1_sequential_4_conv_weight, x = inputs_15)[name = tensor("outputs_aug_3")]; + tensor input_67_begin_0 = const()[name = tensor("input_67_begin_0"), val = tensor([0, 0, 0])]; + tensor input_67_end_0 = const()[name = tensor("input_67_end_0"), val = tensor([2, 256, 16])]; + tensor input_67_end_mask_0 = const()[name = tensor("input_67_end_mask_0"), val = tensor([true, true, false])]; + tensor input_67 = slice_by_index(begin = input_67_begin_0, end = input_67_end_0, end_mask = input_67_end_mask_0, x = outputs_aug_3)[name = tensor("input_67")]; + tensor inputs_17 = batch_norm(beta = inner_encoder_conv_module_1_sequential_5_bias, epsilon = var_46, gamma = inner_encoder_conv_module_1_sequential_5_weight, mean = inner_encoder_conv_module_1_sequential_5_running_mean, variance = inner_encoder_conv_module_1_sequential_5_running_var, x = input_67)[name = tensor("inputs_17")]; + tensor input_69 = silu(x = inputs_17)[name = tensor("input_69")]; + tensor input_71_pad_type_0 = const()[name = tensor("input_71_pad_type_0"), val = tensor("valid")]; + tensor input_71_strides_0 = const()[name = tensor("input_71_strides_0"), val = tensor([1])]; + tensor input_71_pad_0 = const()[name = tensor("input_71_pad_0"), val = tensor([0, 0])]; + tensor input_71_dilations_0 = const()[name = tensor("input_71_dilations_0"), val = tensor([1])]; + tensor input_71_groups_0 = const()[name = tensor("input_71_groups_0"), val = tensor(1)]; + tensor input_71 = conv(bias = inner_encoder_conv_module_1_sequential_7_conv_bias, dilations = input_71_dilations_0, groups = input_71_groups_0, pad = input_71_pad_0, pad_type = input_71_pad_type_0, strides = input_71_strides_0, weight = inner_encoder_conv_module_1_sequential_7_conv_weight, x = input_69)[name = tensor("input_71")]; tensor conv_out_3_perm_0 = const()[name = tensor("conv_out_3_perm_0"), val = tensor([0, 2, 1])]; - tensor var_488_begin_0 = const()[name = tensor("op_488_begin_0"), val = tensor([0, -1, 0])]; - tensor var_488_end_0 = const()[name = tensor("op_488_end_0"), val = tensor([2, 16, 256])]; - tensor var_488_end_mask_0 = const()[name = tensor("op_488_end_mask_0"), val = tensor([true, true, true])]; - tensor conv_out_3 = transpose(perm = conv_out_3_perm_0, x = input_69)[name = tensor("transpose_46")]; - tensor var_488 = slice_by_index(begin = var_488_begin_0, end = var_488_end_0, end_mask = var_488_end_mask_0, x = conv_out_3)[name = tensor("op_488")]; - tensor var_490_perm_0 = const()[name = tensor("op_490_perm_0"), val = tensor([1, 0, 2])]; - tensor var_490 = transpose(perm = var_490_perm_0, x = var_488)[name = tensor("transpose_45")]; - tensor input_71 = add(x = x_9, y = var_490)[name = tensor("input_71")]; - tensor input_73_axes_0 = const()[name = tensor("input_73_axes_0"), val = tensor([-1])]; - tensor input_73 = layer_norm(axes = input_73_axes_0, beta = encoder_ffn2_1_module_sequential_0_bias, epsilon = var_28, gamma = encoder_ffn2_1_module_sequential_0_weight, x = input_71)[name = tensor("input_73")]; - tensor inputs_19 = linear(bias = encoder_ffn2_1_module_sequential_1_linear_bias, weight = encoder_ffn2_1_module_sequential_1_linear_weight, x = input_73)[name = tensor("linear_17")]; - tensor input_75 = silu(x = inputs_19)[name = tensor("input_75")]; - tensor input_79 = linear(bias = encoder_ffn2_1_module_sequential_4_linear_bias, weight = encoder_ffn2_1_module_sequential_4_linear_weight, x = input_75)[name = tensor("linear_18")]; - tensor var_513 = const()[name = tensor("op_513"), val = tensor(0x1p-1)]; - tensor var_514 = mul(x = input_79, y = var_513)[name = tensor("op_514")]; - tensor input_81 = add(x = var_514, y = input_71)[name = tensor("input_81")]; - tensor input_83_axes_0 = const()[name = tensor("input_83_axes_0"), val = tensor([-1])]; - tensor input_83 = layer_norm(axes = input_83_axes_0, beta = encoder_layer_norm_1_bias, epsilon = var_28, gamma = encoder_layer_norm_1_weight, x = input_81)[name = tensor("input_83")]; + tensor var_524_begin_0 = const()[name = tensor("op_524_begin_0"), val = tensor([0, -1, 0])]; + tensor var_524_end_0 = const()[name = tensor("op_524_end_0"), val = tensor([2, 16, 256])]; + tensor var_524_end_mask_0 = const()[name = tensor("op_524_end_mask_0"), val = tensor([true, true, true])]; + tensor conv_out_3 = transpose(perm = conv_out_3_perm_0, x = input_71)[name = tensor("transpose_46")]; + tensor var_524 = slice_by_index(begin = var_524_begin_0, end = var_524_end_0, end_mask = var_524_end_mask_0, x = conv_out_3)[name = tensor("op_524")]; + tensor var_526_perm_0 = const()[name = tensor("op_526_perm_0"), val = tensor([1, 0, 2])]; + tensor var_526 = transpose(perm = var_526_perm_0, x = var_524)[name = tensor("transpose_45")]; + tensor input_73 = add(x = x_9, y = var_526)[name = tensor("input_73")]; + tensor input_75_axes_0 = const()[name = tensor("input_75_axes_0"), val = tensor([-1])]; + tensor input_75 = layer_norm(axes = input_75_axes_0, beta = inner_encoder_ffn2_1_module_sequential_0_bias, epsilon = var_46, gamma = inner_encoder_ffn2_1_module_sequential_0_weight, x = input_73)[name = tensor("input_75")]; + tensor inputs_19 = linear(bias = inner_encoder_ffn2_1_module_sequential_1_linear_bias, weight = inner_encoder_ffn2_1_module_sequential_1_linear_weight, x = input_75)[name = tensor("linear_17")]; + tensor input_77 = silu(x = inputs_19)[name = tensor("input_77")]; + tensor input_81 = linear(bias = inner_encoder_ffn2_1_module_sequential_4_linear_bias, weight = inner_encoder_ffn2_1_module_sequential_4_linear_weight, x = input_77)[name = tensor("linear_18")]; + tensor var_549 = const()[name = tensor("op_549"), val = tensor(0x1p-1)]; + tensor var_550 = mul(x = input_81, y = var_549)[name = tensor("op_550")]; + tensor input_83 = add(x = var_550, y = input_73)[name = tensor("input_83")]; tensor input_85_axes_0 = const()[name = tensor("input_85_axes_0"), val = tensor([-1])]; - tensor input_85 = layer_norm(axes = input_85_axes_0, beta = encoder_ffn1_2_module_sequential_0_bias, epsilon = var_28, gamma = encoder_ffn1_2_module_sequential_0_weight, x = input_83)[name = tensor("input_85")]; - tensor inputs_21 = linear(bias = encoder_ffn1_2_module_sequential_1_linear_bias, weight = encoder_ffn1_2_module_sequential_1_linear_weight, x = input_85)[name = tensor("linear_19")]; - tensor input_87 = silu(x = inputs_21)[name = tensor("input_87")]; - tensor input_91 = linear(bias = encoder_ffn1_2_module_sequential_4_linear_bias, weight = encoder_ffn1_2_module_sequential_4_linear_weight, x = input_87)[name = tensor("linear_20")]; - tensor var_543 = const()[name = tensor("op_543"), val = tensor(0x1p-1)]; - tensor var_544 = mul(x = input_91, y = var_543)[name = tensor("op_544")]; - tensor input_93 = add(x = var_544, y = input_83)[name = tensor("input_93")]; + tensor input_85 = layer_norm(axes = input_85_axes_0, beta = inner_encoder_layer_norm_1_bias, epsilon = var_46, gamma = inner_encoder_layer_norm_1_weight, x = input_83)[name = tensor("input_85")]; + tensor input_87_axes_0 = const()[name = tensor("input_87_axes_0"), val = tensor([-1])]; + tensor input_87 = layer_norm(axes = input_87_axes_0, beta = inner_encoder_ffn1_2_module_sequential_0_bias, epsilon = var_46, gamma = inner_encoder_ffn1_2_module_sequential_0_weight, x = input_85)[name = tensor("input_87")]; + tensor inputs_21 = linear(bias = inner_encoder_ffn1_2_module_sequential_1_linear_bias, weight = inner_encoder_ffn1_2_module_sequential_1_linear_weight, x = input_87)[name = tensor("linear_19")]; + tensor input_89 = silu(x = inputs_21)[name = tensor("input_89")]; + tensor input_93 = linear(bias = inner_encoder_ffn1_2_module_sequential_4_linear_bias, weight = inner_encoder_ffn1_2_module_sequential_4_linear_weight, x = input_89)[name = tensor("linear_20")]; + tensor var_579 = const()[name = tensor("op_579"), val = tensor(0x1p-1)]; + tensor var_580 = mul(x = input_93, y = var_579)[name = tensor("op_580")]; + tensor input_95 = add(x = var_580, y = input_85)[name = tensor("input_95")]; tensor x_13_axes_0 = const()[name = tensor("x_13_axes_0"), val = tensor([-1])]; - tensor x_13 = layer_norm(axes = x_13_axes_0, beta = encoder_ret_lns_2_bias, epsilon = var_28, gamma = encoder_ret_lns_2_weight, x = input_93)[name = tensor("x_13")]; + tensor x_13 = layer_norm(axes = x_13_axes_0, beta = inner_encoder_ret_lns_2_bias, epsilon = var_46, gamma = inner_encoder_ret_lns_2_weight, x = input_95)[name = tensor("x_13")]; tensor prev_kv_5_begin_0 = const()[name = tensor("prev_kv_5_begin_0"), val = tensor([2, 0, 0, 0, 0])]; tensor prev_kv_5_end_0 = const()[name = tensor("prev_kv_5_end_0"), val = tensor([3, 1, 4, 64, 64])]; tensor prev_kv_5_end_mask_0 = const()[name = tensor("prev_kv_5_end_mask_0"), val = tensor([false, true, true, true, true])]; @@ -553,153 +567,153 @@ program(1.0) tensor prev_scale_5_end_mask_0 = const()[name = tensor("prev_scale_5_end_mask_0"), val = tensor([false, true])]; tensor prev_scale_5_squeeze_mask_0 = const()[name = tensor("prev_scale_5_squeeze_mask_0"), val = tensor([true, false])]; tensor prev_scale_5 = slice_by_index(begin = prev_scale_5_begin_0, end = prev_scale_5_end_0, end_mask = prev_scale_5_end_mask_0, squeeze_mask = prev_scale_5_squeeze_mask_0, x = enc_scale)[name = tensor("prev_scale_5")]; - tensor var_558 = linear(bias = encoder_q_proj_2_bias, weight = encoder_q_proj_2_weight, x = x_13)[name = tensor("linear_21")]; - tensor var_559 = const()[name = tensor("op_559"), val = tensor([1, 2, 4, 64])]; - tensor var_560 = reshape(shape = var_559, x = var_558)[name = tensor("op_560")]; + tensor var_594 = linear(bias = inner_encoder_q_proj_2_bias, weight = inner_encoder_q_proj_2_weight, x = x_13)[name = tensor("linear_21")]; + tensor var_595 = const()[name = tensor("op_595"), val = tensor([1, 2, 4, 64])]; + tensor var_596 = reshape(shape = var_595, x = var_594)[name = tensor("op_596")]; tensor q_5_perm_0 = const()[name = tensor("q_5_perm_0"), val = tensor([0, 2, 1, 3])]; - tensor var_564 = linear(bias = encoder_k_proj_2_bias, weight = encoder_k_proj_2_weight, x = x_13)[name = tensor("linear_22")]; - tensor var_565 = const()[name = tensor("op_565"), val = tensor(0x1p-3)]; - tensor var_566 = mul(x = var_564, y = var_565)[name = tensor("op_566")]; - tensor var_567 = const()[name = tensor("op_567"), val = tensor([1, 2, 4, 64])]; - tensor var_568 = reshape(shape = var_567, x = var_566)[name = tensor("op_568")]; + tensor var_600 = linear(bias = inner_encoder_k_proj_2_bias, weight = inner_encoder_k_proj_2_weight, x = x_13)[name = tensor("linear_22")]; + tensor var_601 = const()[name = tensor("op_601"), val = tensor(0x1p-3)]; + tensor var_602 = mul(x = var_600, y = var_601)[name = tensor("op_602")]; + tensor var_603 = const()[name = tensor("op_603"), val = tensor([1, 2, 4, 64])]; + tensor var_604 = reshape(shape = var_603, x = var_602)[name = tensor("op_604")]; tensor k_5_perm_0 = const()[name = tensor("k_5_perm_0"), val = tensor([0, 2, 1, 3])]; - tensor var_572 = linear(bias = encoder_v_proj_2_bias, weight = encoder_v_proj_2_weight, x = x_13)[name = tensor("linear_23")]; - tensor var_573 = const()[name = tensor("op_573"), val = tensor([1, 2, 4, 64])]; - tensor var_574 = reshape(shape = var_573, x = var_572)[name = tensor("op_574")]; + tensor var_608 = linear(bias = inner_encoder_v_proj_2_bias, weight = inner_encoder_v_proj_2_weight, x = x_13)[name = tensor("linear_23")]; + tensor var_609 = const()[name = tensor("op_609"), val = tensor([1, 2, 4, 64])]; + tensor var_610 = reshape(shape = var_609, x = var_608)[name = tensor("op_610")]; tensor v_5_perm_0 = const()[name = tensor("v_5_perm_0"), val = tensor([0, 2, 1, 3])]; - tensor input_97 = linear(bias = encoder_g_proj_2_bias, weight = encoder_g_proj_2_weight, x = x_13)[name = tensor("linear_24")]; + tensor input_99 = linear(bias = inner_encoder_g_proj_2_bias, weight = inner_encoder_g_proj_2_weight, x = x_13)[name = tensor("linear_24")]; tensor sqrt_s0_5 = sqrt(x = prev_scale_5)[name = tensor("sqrt_s0_5")]; - tensor s_t_5 = add(x = prev_scale_5, y = encoder__t_index)[name = tensor("s_t_5")]; + tensor s_t_5 = add(x = prev_scale_5, y = inner_encoder__t_index)[name = tensor("s_t_5")]; tensor sqrt_s_t_5 = sqrt(x = s_t_5)[name = tensor("sqrt_s_t_5")]; tensor qk_5_transpose_x_1 = const()[name = tensor("qk_5_transpose_x_1"), val = tensor(false)]; tensor qk_5_transpose_y_1 = const()[name = tensor("qk_5_transpose_y_1"), val = tensor(true)]; - tensor k_5 = transpose(perm = k_5_perm_0, x = var_568)[name = tensor("transpose_43")]; - tensor q_5 = transpose(perm = q_5_perm_0, x = var_560)[name = tensor("transpose_44")]; + tensor k_5 = transpose(perm = k_5_perm_0, x = var_604)[name = tensor("transpose_43")]; + tensor q_5 = transpose(perm = q_5_perm_0, x = var_596)[name = tensor("transpose_44")]; tensor qk_5 = matmul(transpose_x = qk_5_transpose_x_1, transpose_y = qk_5_transpose_y_1, x = q_5, y = k_5)[name = tensor("qk_5")]; - tensor var_584 = const()[name = tensor("op_584"), val = tensor([2, 1])]; - tensor var_585 = reshape(shape = var_584, x = sqrt_s_t_5)[name = tensor("op_585")]; - tensor M_5 = real_div(x = encoder__causal_mask, y = var_585)[name = tensor("M_5")]; - tensor var_587 = mul(x = qk_5, y = M_5)[name = tensor("op_587")]; + tensor var_620 = const()[name = tensor("op_620"), val = tensor([2, 1])]; + tensor var_621 = reshape(shape = var_620, x = sqrt_s_t_5)[name = tensor("op_621")]; + tensor M_5 = real_div(x = inner_encoder__causal_mask, y = var_621)[name = tensor("M_5")]; + tensor var_623 = mul(x = qk_5, y = M_5)[name = tensor("op_623")]; tensor inner_5_transpose_x_0 = const()[name = tensor("inner_5_transpose_x_0"), val = tensor(false)]; tensor inner_5_transpose_y_0 = const()[name = tensor("inner_5_transpose_y_0"), val = tensor(false)]; - tensor v_5 = transpose(perm = v_5_perm_0, x = var_574)[name = tensor("transpose_42")]; - tensor inner_5 = matmul(transpose_x = inner_5_transpose_x_0, transpose_y = inner_5_transpose_y_0, x = var_587, y = v_5)[name = tensor("inner_5")]; - tensor var_589_transpose_x_0 = const()[name = tensor("op_589_transpose_x_0"), val = tensor(false)]; - tensor var_589_transpose_y_0 = const()[name = tensor("op_589_transpose_y_0"), val = tensor(false)]; - tensor var_589 = matmul(transpose_x = var_589_transpose_x_0, transpose_y = var_589_transpose_y_0, x = q_5, y = prev_kv_5)[name = tensor("op_589")]; - tensor var_590 = real_div(x = sqrt_s0_5, y = sqrt_s_t_5)[name = tensor("op_590")]; - tensor var_591 = const()[name = tensor("op_591"), val = tensor([1, 1, 2, 1])]; - tensor var_592 = reshape(shape = var_591, x = var_590)[name = tensor("op_592")]; - tensor cross_5 = mul(x = var_589, y = var_592)[name = tensor("cross_5")]; + tensor v_5 = transpose(perm = v_5_perm_0, x = var_610)[name = tensor("transpose_42")]; + tensor inner_5 = matmul(transpose_x = inner_5_transpose_x_0, transpose_y = inner_5_transpose_y_0, x = var_623, y = v_5)[name = tensor("inner_5")]; + tensor var_625_transpose_x_0 = const()[name = tensor("op_625_transpose_x_0"), val = tensor(false)]; + tensor var_625_transpose_y_0 = const()[name = tensor("op_625_transpose_y_0"), val = tensor(false)]; + tensor var_625 = matmul(transpose_x = var_625_transpose_x_0, transpose_y = var_625_transpose_y_0, x = q_5, y = prev_kv_5)[name = tensor("op_625")]; + tensor var_626 = real_div(x = sqrt_s0_5, y = sqrt_s_t_5)[name = tensor("op_626")]; + tensor var_627 = const()[name = tensor("op_627"), val = tensor([1, 1, 2, 1])]; + tensor var_628 = reshape(shape = var_627, x = var_626)[name = tensor("op_628")]; + tensor cross_5 = mul(x = var_625, y = var_628)[name = tensor("cross_5")]; tensor out_13 = add(x = inner_5, y = cross_5)[name = tensor("out_13")]; - tensor var_595 = mul(x = prev_kv_5, y = sqrt_s0_5)[name = tensor("op_595")]; - tensor var_597_transpose_x_1 = const()[name = tensor("op_597_transpose_x_1"), val = tensor(true)]; - tensor var_597_transpose_y_1 = const()[name = tensor("op_597_transpose_y_1"), val = tensor(false)]; - tensor var_597 = matmul(transpose_x = var_597_transpose_x_1, transpose_y = var_597_transpose_y_1, x = k_5, y = v_5)[name = tensor("op_597")]; - tensor new_kv_unnorm_5 = add(x = var_595, y = var_597)[name = tensor("new_kv_unnorm_5")]; - tensor var_599 = const()[name = tensor("op_599"), val = tensor(0x1p+1)]; - tensor new_scale_5 = add(x = prev_scale_5, y = var_599)[name = tensor("new_scale_5")]; - tensor var_601 = sqrt(x = new_scale_5)[name = tensor("op_601")]; - tensor var_602 = real_div(x = new_kv_unnorm_5, y = var_601)[name = tensor("op_602")]; - tensor var_603_perm_0 = const()[name = tensor("op_603_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_631 = mul(x = prev_kv_5, y = sqrt_s0_5)[name = tensor("op_631")]; + tensor var_633_transpose_x_1 = const()[name = tensor("op_633_transpose_x_1"), val = tensor(true)]; + tensor var_633_transpose_y_1 = const()[name = tensor("op_633_transpose_y_1"), val = tensor(false)]; + tensor var_633 = matmul(transpose_x = var_633_transpose_x_1, transpose_y = var_633_transpose_y_1, x = k_5, y = v_5)[name = tensor("op_633")]; + tensor new_kv_unnorm_5 = add(x = var_631, y = var_633)[name = tensor("new_kv_unnorm_5")]; + tensor var_635 = const()[name = tensor("op_635"), val = tensor(0x1p+1)]; + tensor new_scale_5 = add(x = prev_scale_5, y = var_635)[name = tensor("new_scale_5")]; + tensor var_637 = sqrt(x = new_scale_5)[name = tensor("op_637")]; + tensor var_638 = real_div(x = new_kv_unnorm_5, y = var_637)[name = tensor("op_638")]; + tensor var_639_perm_0 = const()[name = tensor("op_639_perm_0"), val = tensor([0, 2, 1, 3])]; tensor out_15_axes_0 = const()[name = tensor("out_15_axes_0"), val = tensor([-1])]; - tensor var_603 = transpose(perm = var_603_perm_0, x = out_13)[name = tensor("transpose_41")]; - tensor out_15 = layer_norm(axes = out_15_axes_0, epsilon = var_17, x = var_603)[name = tensor("out_15")]; - tensor var_607 = const()[name = tensor("op_607"), val = tensor([1, 2, 256])]; - tensor out_17 = reshape(shape = var_607, x = out_15)[name = tensor("out_17")]; - tensor var_609 = silu(x = input_97)[name = tensor("op_609")]; - tensor input_99 = mul(x = var_609, y = out_17)[name = tensor("input_99")]; - tensor ret_out_5 = linear(bias = encoder_out_proj_2_bias, weight = encoder_out_proj_2_weight, x = input_99)[name = tensor("linear_25")]; - tensor x_15 = add(x = input_93, y = ret_out_5)[name = tensor("x_15")]; + tensor var_639 = transpose(perm = var_639_perm_0, x = out_13)[name = tensor("transpose_41")]; + tensor out_15 = layer_norm(axes = out_15_axes_0, epsilon = var_54, x = var_639)[name = tensor("out_15")]; + tensor var_643 = const()[name = tensor("op_643"), val = tensor([1, 2, 256])]; + tensor out_17 = reshape(shape = var_643, x = out_15)[name = tensor("out_17")]; + tensor var_645 = silu(x = input_99)[name = tensor("op_645")]; + tensor input_101 = mul(x = var_645, y = out_17)[name = tensor("input_101")]; + tensor ret_out_5 = linear(bias = inner_encoder_out_proj_2_bias, weight = inner_encoder_out_proj_2_weight, x = input_101)[name = tensor("linear_25")]; + tensor x_15 = add(x = input_95, y = ret_out_5)[name = tensor("x_15")]; tensor window_13_begin_0 = const()[name = tensor("window_13_begin_0"), val = tensor([2, 0, 0, 0])]; tensor window_13_end_0 = const()[name = tensor("window_13_end_0"), val = tensor([3, 1, 16, 256])]; tensor window_13_end_mask_0 = const()[name = tensor("window_13_end_mask_0"), val = tensor([false, true, true, true])]; tensor window_13_squeeze_mask_0 = const()[name = tensor("window_13_squeeze_mask_0"), val = tensor([true, false, false, false])]; tensor window_13 = slice_by_index(begin = window_13_begin_0, end = window_13_end_0, end_mask = window_13_end_mask_0, squeeze_mask = window_13_squeeze_mask_0, x = enc_conv_cache)[name = tensor("window_13")]; - tensor var_617_begin_0 = const()[name = tensor("op_617_begin_0"), val = tensor([0, 0, 0])]; - tensor var_617_end_0 = const()[name = tensor("op_617_end_0"), val = tensor([1, 1, 256])]; - tensor var_617_end_mask_0 = const()[name = tensor("op_617_end_mask_0"), val = tensor([true, false, true])]; - tensor var_617 = slice_by_index(begin = var_617_begin_0, end = var_617_end_0, end_mask = var_617_end_mask_0, x = x_15)[name = tensor("op_617")]; - tensor var_620_begin_0 = const()[name = tensor("op_620_begin_0"), val = tensor([0, 1, 0])]; - tensor var_620_end_0 = const()[name = tensor("op_620_end_0"), val = tensor([1, 16, 256])]; - tensor var_620_end_mask_0 = const()[name = tensor("op_620_end_mask_0"), val = tensor([true, true, true])]; - tensor var_620 = slice_by_index(begin = var_620_begin_0, end = var_620_end_0, end_mask = var_620_end_mask_0, x = window_13)[name = tensor("op_620")]; + tensor var_653_begin_0 = const()[name = tensor("op_653_begin_0"), val = tensor([0, 0, 0])]; + tensor var_653_end_0 = const()[name = tensor("op_653_end_0"), val = tensor([1, 1, 256])]; + tensor var_653_end_mask_0 = const()[name = tensor("op_653_end_mask_0"), val = tensor([true, false, true])]; + tensor var_653 = slice_by_index(begin = var_653_begin_0, end = var_653_end_0, end_mask = var_653_end_mask_0, x = x_15)[name = tensor("op_653")]; + tensor var_656_begin_0 = const()[name = tensor("op_656_begin_0"), val = tensor([0, 1, 0])]; + tensor var_656_end_0 = const()[name = tensor("op_656_end_0"), val = tensor([1, 16, 256])]; + tensor var_656_end_mask_0 = const()[name = tensor("op_656_end_mask_0"), val = tensor([true, true, true])]; + tensor var_656 = slice_by_index(begin = var_656_begin_0, end = var_656_end_0, end_mask = var_656_end_mask_0, x = window_13)[name = tensor("op_656")]; tensor window_15_interleave_0 = const()[name = tensor("window_15_interleave_0"), val = tensor(false)]; - tensor window_15 = concat(axis = var_26, interleave = window_15_interleave_0, values = (var_620, var_617))[name = tensor("window_15")]; - tensor var_625_begin_0 = const()[name = tensor("op_625_begin_0"), val = tensor([0, 1, 0])]; - tensor var_625_end_0 = const()[name = tensor("op_625_end_0"), val = tensor([1, 1, 256])]; - tensor var_625_end_mask_0 = const()[name = tensor("op_625_end_mask_0"), val = tensor([true, true, true])]; - tensor var_625 = slice_by_index(begin = var_625_begin_0, end = var_625_end_0, end_mask = var_625_end_mask_0, x = x_15)[name = tensor("op_625")]; - tensor var_628_begin_0 = const()[name = tensor("op_628_begin_0"), val = tensor([0, 1, 0])]; - tensor var_628_end_0 = const()[name = tensor("op_628_end_0"), val = tensor([1, 16, 256])]; - tensor var_628_end_mask_0 = const()[name = tensor("op_628_end_mask_0"), val = tensor([true, true, true])]; - tensor var_628 = slice_by_index(begin = var_628_begin_0, end = var_628_end_0, end_mask = var_628_end_mask_0, x = window_15)[name = tensor("op_628")]; + tensor window_15 = concat(axis = var_62, interleave = window_15_interleave_0, values = (var_656, var_653))[name = tensor("window_15")]; + tensor var_661_begin_0 = const()[name = tensor("op_661_begin_0"), val = tensor([0, 1, 0])]; + tensor var_661_end_0 = const()[name = tensor("op_661_end_0"), val = tensor([1, 1, 256])]; + tensor var_661_end_mask_0 = const()[name = tensor("op_661_end_mask_0"), val = tensor([true, true, true])]; + tensor var_661 = slice_by_index(begin = var_661_begin_0, end = var_661_end_0, end_mask = var_661_end_mask_0, x = x_15)[name = tensor("op_661")]; + tensor var_664_begin_0 = const()[name = tensor("op_664_begin_0"), val = tensor([0, 1, 0])]; + tensor var_664_end_0 = const()[name = tensor("op_664_end_0"), val = tensor([1, 16, 256])]; + tensor var_664_end_mask_0 = const()[name = tensor("op_664_end_mask_0"), val = tensor([true, true, true])]; + tensor var_664 = slice_by_index(begin = var_664_begin_0, end = var_664_end_0, end_mask = var_664_end_mask_0, x = window_15)[name = tensor("op_664")]; tensor window_17_interleave_0 = const()[name = tensor("window_17_interleave_0"), val = tensor(false)]; - tensor window_17 = concat(axis = var_26, interleave = window_17_interleave_0, values = (var_628, var_625))[name = tensor("window_17")]; - tensor input_101_interleave_0 = const()[name = tensor("input_101_interleave_0"), val = tensor(false)]; - tensor input_101 = concat(axis = var_23, interleave = input_101_interleave_0, values = (window_15, window_17))[name = tensor("input_101")]; + tensor window_17 = concat(axis = var_62, interleave = window_17_interleave_0, values = (var_664, var_661))[name = tensor("window_17")]; + tensor input_103_interleave_0 = const()[name = tensor("input_103_interleave_0"), val = tensor(false)]; + tensor input_103 = concat(axis = var_49, interleave = input_103_interleave_0, values = (window_15, window_17))[name = tensor("input_103")]; tensor x_17_axes_0 = const()[name = tensor("x_17_axes_0"), val = tensor([-1])]; - tensor x_17 = layer_norm(axes = x_17_axes_0, beta = encoder_conv_module_2_sequential_0_bias, epsilon = var_28, gamma = encoder_conv_module_2_sequential_0_weight, x = input_101)[name = tensor("x_17")]; - tensor input_103_perm_0 = const()[name = tensor("input_103_perm_0"), val = tensor([0, 2, 1])]; + tensor x_17 = layer_norm(axes = x_17_axes_0, beta = inner_encoder_conv_module_2_sequential_0_bias, epsilon = var_46, gamma = inner_encoder_conv_module_2_sequential_0_weight, x = input_103)[name = tensor("x_17")]; + tensor input_105_perm_0 = const()[name = tensor("input_105_perm_0"), val = tensor([0, 2, 1])]; tensor inputs_23_pad_type_0 = const()[name = tensor("inputs_23_pad_type_0"), val = tensor("valid")]; tensor inputs_23_strides_0 = const()[name = tensor("inputs_23_strides_0"), val = tensor([1])]; tensor inputs_23_pad_0 = const()[name = tensor("inputs_23_pad_0"), val = tensor([0, 0])]; tensor inputs_23_dilations_0 = const()[name = tensor("inputs_23_dilations_0"), val = tensor([1])]; tensor inputs_23_groups_0 = const()[name = tensor("inputs_23_groups_0"), val = tensor(1)]; - tensor input_103 = transpose(perm = input_103_perm_0, x = x_17)[name = tensor("transpose_40")]; - tensor inputs_23 = conv(bias = encoder_conv_module_2_sequential_2_conv_bias, dilations = inputs_23_dilations_0, groups = inputs_23_groups_0, pad = inputs_23_pad_0, pad_type = inputs_23_pad_type_0, strides = inputs_23_strides_0, weight = encoder_conv_module_2_sequential_2_conv_weight, x = input_103)[name = tensor("inputs_23")]; - tensor var_653_split_sizes_0 = const()[name = tensor("op_653_split_sizes_0"), val = tensor([256, 256])]; - tensor var_653_axis_0 = const()[name = tensor("op_653_axis_0"), val = tensor(1)]; - tensor var_653_0, tensor var_653_1 = split(axis = var_653_axis_0, split_sizes = var_653_split_sizes_0, x = inputs_23)[name = tensor("op_653")]; - tensor var_655 = sigmoid(x = var_653_1)[name = tensor("op_655")]; - tensor inputs_25 = mul(x = var_653_0, y = var_655)[name = tensor("inputs_25")]; + tensor input_105 = transpose(perm = input_105_perm_0, x = x_17)[name = tensor("transpose_40")]; + tensor inputs_23 = conv(bias = inner_encoder_conv_module_2_sequential_2_conv_bias, dilations = inputs_23_dilations_0, groups = inputs_23_groups_0, pad = inputs_23_pad_0, pad_type = inputs_23_pad_type_0, strides = inputs_23_strides_0, weight = inner_encoder_conv_module_2_sequential_2_conv_weight, x = input_105)[name = tensor("inputs_23")]; + tensor var_689_split_sizes_0 = const()[name = tensor("op_689_split_sizes_0"), val = tensor([256, 256])]; + tensor var_689_axis_0 = const()[name = tensor("op_689_axis_0"), val = tensor(1)]; + tensor var_689_0, tensor var_689_1 = split(axis = var_689_axis_0, split_sizes = var_689_split_sizes_0, x = inputs_23)[name = tensor("op_689")]; + tensor var_691 = sigmoid(x = var_689_1)[name = tensor("op_691")]; + tensor inputs_25 = mul(x = var_689_0, y = var_691)[name = tensor("inputs_25")]; tensor outputs_aug_5_pad_type_0 = const()[name = tensor("outputs_aug_5_pad_type_0"), val = tensor("custom")]; tensor outputs_aug_5_pad_0 = const()[name = tensor("outputs_aug_5_pad_0"), val = tensor([15, 15])]; tensor outputs_aug_5_groups_0 = const()[name = tensor("outputs_aug_5_groups_0"), val = tensor(256)]; tensor outputs_aug_5_strides_0 = const()[name = tensor("outputs_aug_5_strides_0"), val = tensor([1])]; tensor outputs_aug_5_dilations_0 = const()[name = tensor("outputs_aug_5_dilations_0"), val = tensor([1])]; - tensor outputs_aug_5 = conv(dilations = outputs_aug_5_dilations_0, groups = outputs_aug_5_groups_0, pad = outputs_aug_5_pad_0, pad_type = outputs_aug_5_pad_type_0, strides = outputs_aug_5_strides_0, weight = encoder_conv_module_2_sequential_4_conv_weight, x = inputs_25)[name = tensor("outputs_aug_5")]; - tensor input_105_begin_0 = const()[name = tensor("input_105_begin_0"), val = tensor([0, 0, 0])]; - tensor input_105_end_0 = const()[name = tensor("input_105_end_0"), val = tensor([2, 256, 16])]; - tensor input_105_end_mask_0 = const()[name = tensor("input_105_end_mask_0"), val = tensor([true, true, false])]; - tensor input_105 = slice_by_index(begin = input_105_begin_0, end = input_105_end_0, end_mask = input_105_end_mask_0, x = outputs_aug_5)[name = tensor("input_105")]; - tensor inputs_27 = batch_norm(beta = encoder_conv_module_2_sequential_5_bias, epsilon = var_28, gamma = encoder_conv_module_2_sequential_5_weight, mean = encoder_conv_module_2_sequential_5_running_mean, variance = encoder_conv_module_2_sequential_5_running_var, x = input_105)[name = tensor("inputs_27")]; - tensor input_107 = silu(x = inputs_27)[name = tensor("input_107")]; - tensor input_109_pad_type_0 = const()[name = tensor("input_109_pad_type_0"), val = tensor("valid")]; - tensor input_109_strides_0 = const()[name = tensor("input_109_strides_0"), val = tensor([1])]; - tensor input_109_pad_0 = const()[name = tensor("input_109_pad_0"), val = tensor([0, 0])]; - tensor input_109_dilations_0 = const()[name = tensor("input_109_dilations_0"), val = tensor([1])]; - tensor input_109_groups_0 = const()[name = tensor("input_109_groups_0"), val = tensor(1)]; - tensor input_109 = conv(bias = encoder_conv_module_2_sequential_7_conv_bias, dilations = input_109_dilations_0, groups = input_109_groups_0, pad = input_109_pad_0, pad_type = input_109_pad_type_0, strides = input_109_strides_0, weight = encoder_conv_module_2_sequential_7_conv_weight, x = input_107)[name = tensor("input_109")]; + tensor outputs_aug_5 = conv(dilations = outputs_aug_5_dilations_0, groups = outputs_aug_5_groups_0, pad = outputs_aug_5_pad_0, pad_type = outputs_aug_5_pad_type_0, strides = outputs_aug_5_strides_0, weight = inner_encoder_conv_module_2_sequential_4_conv_weight, x = inputs_25)[name = tensor("outputs_aug_5")]; + tensor input_107_begin_0 = const()[name = tensor("input_107_begin_0"), val = tensor([0, 0, 0])]; + tensor input_107_end_0 = const()[name = tensor("input_107_end_0"), val = tensor([2, 256, 16])]; + tensor input_107_end_mask_0 = const()[name = tensor("input_107_end_mask_0"), val = tensor([true, true, false])]; + tensor input_107 = slice_by_index(begin = input_107_begin_0, end = input_107_end_0, end_mask = input_107_end_mask_0, x = outputs_aug_5)[name = tensor("input_107")]; + tensor inputs_27 = batch_norm(beta = inner_encoder_conv_module_2_sequential_5_bias, epsilon = var_46, gamma = inner_encoder_conv_module_2_sequential_5_weight, mean = inner_encoder_conv_module_2_sequential_5_running_mean, variance = inner_encoder_conv_module_2_sequential_5_running_var, x = input_107)[name = tensor("inputs_27")]; + tensor input_109 = silu(x = inputs_27)[name = tensor("input_109")]; + tensor input_111_pad_type_0 = const()[name = tensor("input_111_pad_type_0"), val = tensor("valid")]; + tensor input_111_strides_0 = const()[name = tensor("input_111_strides_0"), val = tensor([1])]; + tensor input_111_pad_0 = const()[name = tensor("input_111_pad_0"), val = tensor([0, 0])]; + tensor input_111_dilations_0 = const()[name = tensor("input_111_dilations_0"), val = tensor([1])]; + tensor input_111_groups_0 = const()[name = tensor("input_111_groups_0"), val = tensor(1)]; + tensor input_111 = conv(bias = inner_encoder_conv_module_2_sequential_7_conv_bias, dilations = input_111_dilations_0, groups = input_111_groups_0, pad = input_111_pad_0, pad_type = input_111_pad_type_0, strides = input_111_strides_0, weight = inner_encoder_conv_module_2_sequential_7_conv_weight, x = input_109)[name = tensor("input_111")]; tensor conv_out_5_perm_0 = const()[name = tensor("conv_out_5_perm_0"), val = tensor([0, 2, 1])]; - tensor var_686_begin_0 = const()[name = tensor("op_686_begin_0"), val = tensor([0, -1, 0])]; - tensor var_686_end_0 = const()[name = tensor("op_686_end_0"), val = tensor([2, 16, 256])]; - tensor var_686_end_mask_0 = const()[name = tensor("op_686_end_mask_0"), val = tensor([true, true, true])]; - tensor conv_out_5 = transpose(perm = conv_out_5_perm_0, x = input_109)[name = tensor("transpose_39")]; - tensor var_686 = slice_by_index(begin = var_686_begin_0, end = var_686_end_0, end_mask = var_686_end_mask_0, x = conv_out_5)[name = tensor("op_686")]; - tensor var_688_perm_0 = const()[name = tensor("op_688_perm_0"), val = tensor([1, 0, 2])]; - tensor var_688 = transpose(perm = var_688_perm_0, x = var_686)[name = tensor("transpose_38")]; - tensor input_111 = add(x = x_15, y = var_688)[name = tensor("input_111")]; - tensor input_113_axes_0 = const()[name = tensor("input_113_axes_0"), val = tensor([-1])]; - tensor input_113 = layer_norm(axes = input_113_axes_0, beta = encoder_ffn2_2_module_sequential_0_bias, epsilon = var_28, gamma = encoder_ffn2_2_module_sequential_0_weight, x = input_111)[name = tensor("input_113")]; - tensor inputs_29 = linear(bias = encoder_ffn2_2_module_sequential_1_linear_bias, weight = encoder_ffn2_2_module_sequential_1_linear_weight, x = input_113)[name = tensor("linear_26")]; - tensor input_115 = silu(x = inputs_29)[name = tensor("input_115")]; - tensor input_119 = linear(bias = encoder_ffn2_2_module_sequential_4_linear_bias, weight = encoder_ffn2_2_module_sequential_4_linear_weight, x = input_115)[name = tensor("linear_27")]; - tensor var_711 = const()[name = tensor("op_711"), val = tensor(0x1p-1)]; - tensor var_712 = mul(x = input_119, y = var_711)[name = tensor("op_712")]; - tensor input_121 = add(x = var_712, y = input_111)[name = tensor("input_121")]; - tensor input_123_axes_0 = const()[name = tensor("input_123_axes_0"), val = tensor([-1])]; - tensor input_123 = layer_norm(axes = input_123_axes_0, beta = encoder_layer_norm_2_bias, epsilon = var_28, gamma = encoder_layer_norm_2_weight, x = input_121)[name = tensor("input_123")]; + tensor var_722_begin_0 = const()[name = tensor("op_722_begin_0"), val = tensor([0, -1, 0])]; + tensor var_722_end_0 = const()[name = tensor("op_722_end_0"), val = tensor([2, 16, 256])]; + tensor var_722_end_mask_0 = const()[name = tensor("op_722_end_mask_0"), val = tensor([true, true, true])]; + tensor conv_out_5 = transpose(perm = conv_out_5_perm_0, x = input_111)[name = tensor("transpose_39")]; + tensor var_722 = slice_by_index(begin = var_722_begin_0, end = var_722_end_0, end_mask = var_722_end_mask_0, x = conv_out_5)[name = tensor("op_722")]; + tensor var_724_perm_0 = const()[name = tensor("op_724_perm_0"), val = tensor([1, 0, 2])]; + tensor var_724 = transpose(perm = var_724_perm_0, x = var_722)[name = tensor("transpose_38")]; + tensor input_113 = add(x = x_15, y = var_724)[name = tensor("input_113")]; + tensor input_115_axes_0 = const()[name = tensor("input_115_axes_0"), val = tensor([-1])]; + tensor input_115 = layer_norm(axes = input_115_axes_0, beta = inner_encoder_ffn2_2_module_sequential_0_bias, epsilon = var_46, gamma = inner_encoder_ffn2_2_module_sequential_0_weight, x = input_113)[name = tensor("input_115")]; + tensor inputs_29 = linear(bias = inner_encoder_ffn2_2_module_sequential_1_linear_bias, weight = inner_encoder_ffn2_2_module_sequential_1_linear_weight, x = input_115)[name = tensor("linear_26")]; + tensor input_117 = silu(x = inputs_29)[name = tensor("input_117")]; + tensor input_121 = linear(bias = inner_encoder_ffn2_2_module_sequential_4_linear_bias, weight = inner_encoder_ffn2_2_module_sequential_4_linear_weight, x = input_117)[name = tensor("linear_27")]; + tensor var_747 = const()[name = tensor("op_747"), val = tensor(0x1p-1)]; + tensor var_748 = mul(x = input_121, y = var_747)[name = tensor("op_748")]; + tensor input_123 = add(x = var_748, y = input_113)[name = tensor("input_123")]; tensor input_125_axes_0 = const()[name = tensor("input_125_axes_0"), val = tensor([-1])]; - tensor input_125 = layer_norm(axes = input_125_axes_0, beta = encoder_ffn1_3_module_sequential_0_bias, epsilon = var_28, gamma = encoder_ffn1_3_module_sequential_0_weight, x = input_123)[name = tensor("input_125")]; - tensor inputs_31 = linear(bias = encoder_ffn1_3_module_sequential_1_linear_bias, weight = encoder_ffn1_3_module_sequential_1_linear_weight, x = input_125)[name = tensor("linear_28")]; - tensor input_127 = silu(x = inputs_31)[name = tensor("input_127")]; - tensor input_131 = linear(bias = encoder_ffn1_3_module_sequential_4_linear_bias, weight = encoder_ffn1_3_module_sequential_4_linear_weight, x = input_127)[name = tensor("linear_29")]; - tensor var_741 = const()[name = tensor("op_741"), val = tensor(0x1p-1)]; - tensor var_742 = mul(x = input_131, y = var_741)[name = tensor("op_742")]; - tensor input_133 = add(x = var_742, y = input_123)[name = tensor("input_133")]; + tensor input_125 = layer_norm(axes = input_125_axes_0, beta = inner_encoder_layer_norm_2_bias, epsilon = var_46, gamma = inner_encoder_layer_norm_2_weight, x = input_123)[name = tensor("input_125")]; + tensor input_127_axes_0 = const()[name = tensor("input_127_axes_0"), val = tensor([-1])]; + tensor input_127 = layer_norm(axes = input_127_axes_0, beta = inner_encoder_ffn1_3_module_sequential_0_bias, epsilon = var_46, gamma = inner_encoder_ffn1_3_module_sequential_0_weight, x = input_125)[name = tensor("input_127")]; + tensor inputs_31 = linear(bias = inner_encoder_ffn1_3_module_sequential_1_linear_bias, weight = inner_encoder_ffn1_3_module_sequential_1_linear_weight, x = input_127)[name = tensor("linear_28")]; + tensor input_129 = silu(x = inputs_31)[name = tensor("input_129")]; + tensor input_133 = linear(bias = inner_encoder_ffn1_3_module_sequential_4_linear_bias, weight = inner_encoder_ffn1_3_module_sequential_4_linear_weight, x = input_129)[name = tensor("linear_29")]; + tensor var_777 = const()[name = tensor("op_777"), val = tensor(0x1p-1)]; + tensor var_778 = mul(x = input_133, y = var_777)[name = tensor("op_778")]; + tensor input_135 = add(x = var_778, y = input_125)[name = tensor("input_135")]; tensor x_19_axes_0 = const()[name = tensor("x_19_axes_0"), val = tensor([-1])]; - tensor x_19 = layer_norm(axes = x_19_axes_0, beta = encoder_ret_lns_3_bias, epsilon = var_28, gamma = encoder_ret_lns_3_weight, x = input_133)[name = tensor("x_19")]; + tensor x_19 = layer_norm(axes = x_19_axes_0, beta = inner_encoder_ret_lns_3_bias, epsilon = var_46, gamma = inner_encoder_ret_lns_3_weight, x = input_135)[name = tensor("x_19")]; tensor prev_kv_7_begin_0 = const()[name = tensor("prev_kv_7_begin_0"), val = tensor([3, 0, 0, 0, 0])]; tensor prev_kv_7_end_0 = const()[name = tensor("prev_kv_7_end_0"), val = tensor([4, 1, 4, 64, 64])]; tensor prev_kv_7_end_mask_0 = const()[name = tensor("prev_kv_7_end_mask_0"), val = tensor([false, true, true, true, true])]; @@ -710,189 +724,182 @@ program(1.0) tensor prev_scale_7_end_mask_0 = const()[name = tensor("prev_scale_7_end_mask_0"), val = tensor([false, true])]; tensor prev_scale_7_squeeze_mask_0 = const()[name = tensor("prev_scale_7_squeeze_mask_0"), val = tensor([true, false])]; tensor prev_scale_7 = slice_by_index(begin = prev_scale_7_begin_0, end = prev_scale_7_end_0, end_mask = prev_scale_7_end_mask_0, squeeze_mask = prev_scale_7_squeeze_mask_0, x = enc_scale)[name = tensor("prev_scale_7")]; - tensor var_756 = linear(bias = encoder_q_proj_3_bias, weight = encoder_q_proj_3_weight, x = x_19)[name = tensor("linear_30")]; - tensor var_757 = const()[name = tensor("op_757"), val = tensor([1, 2, 4, 64])]; - tensor var_758 = reshape(shape = var_757, x = var_756)[name = tensor("op_758")]; + tensor var_792 = linear(bias = inner_encoder_q_proj_3_bias, weight = inner_encoder_q_proj_3_weight, x = x_19)[name = tensor("linear_30")]; + tensor var_793 = const()[name = tensor("op_793"), val = tensor([1, 2, 4, 64])]; + tensor var_794 = reshape(shape = var_793, x = var_792)[name = tensor("op_794")]; tensor q_7_perm_0 = const()[name = tensor("q_7_perm_0"), val = tensor([0, 2, 1, 3])]; - tensor var_762 = linear(bias = encoder_k_proj_3_bias, weight = encoder_k_proj_3_weight, x = x_19)[name = tensor("linear_31")]; - tensor var_763 = const()[name = tensor("op_763"), val = tensor(0x1p-3)]; - tensor var_764 = mul(x = var_762, y = var_763)[name = tensor("op_764")]; - tensor var_765 = const()[name = tensor("op_765"), val = tensor([1, 2, 4, 64])]; - tensor var_766 = reshape(shape = var_765, x = var_764)[name = tensor("op_766")]; + tensor var_798 = linear(bias = inner_encoder_k_proj_3_bias, weight = inner_encoder_k_proj_3_weight, x = x_19)[name = tensor("linear_31")]; + tensor var_799 = const()[name = tensor("op_799"), val = tensor(0x1p-3)]; + tensor var_800 = mul(x = var_798, y = var_799)[name = tensor("op_800")]; + tensor var_801 = const()[name = tensor("op_801"), val = tensor([1, 2, 4, 64])]; + tensor var_802 = reshape(shape = var_801, x = var_800)[name = tensor("op_802")]; tensor k_7_perm_0 = const()[name = tensor("k_7_perm_0"), val = tensor([0, 2, 1, 3])]; - tensor var_770 = linear(bias = encoder_v_proj_3_bias, weight = encoder_v_proj_3_weight, x = x_19)[name = tensor("linear_32")]; - tensor var_771 = const()[name = tensor("op_771"), val = tensor([1, 2, 4, 64])]; - tensor var_772 = reshape(shape = var_771, x = var_770)[name = tensor("op_772")]; + tensor var_806 = linear(bias = inner_encoder_v_proj_3_bias, weight = inner_encoder_v_proj_3_weight, x = x_19)[name = tensor("linear_32")]; + tensor var_807 = const()[name = tensor("op_807"), val = tensor([1, 2, 4, 64])]; + tensor var_808 = reshape(shape = var_807, x = var_806)[name = tensor("op_808")]; tensor v_7_perm_0 = const()[name = tensor("v_7_perm_0"), val = tensor([0, 2, 1, 3])]; - tensor input_137 = linear(bias = encoder_g_proj_3_bias, weight = encoder_g_proj_3_weight, x = x_19)[name = tensor("linear_33")]; + tensor input_139 = linear(bias = inner_encoder_g_proj_3_bias, weight = inner_encoder_g_proj_3_weight, x = x_19)[name = tensor("linear_33")]; tensor sqrt_s0_7 = sqrt(x = prev_scale_7)[name = tensor("sqrt_s0_7")]; - tensor s_t_7 = add(x = prev_scale_7, y = encoder__t_index)[name = tensor("s_t_7")]; + tensor s_t_7 = add(x = prev_scale_7, y = inner_encoder__t_index)[name = tensor("s_t_7")]; tensor sqrt_s_t_7 = sqrt(x = s_t_7)[name = tensor("sqrt_s_t_7")]; tensor qk_7_transpose_x_1 = const()[name = tensor("qk_7_transpose_x_1"), val = tensor(false)]; tensor qk_7_transpose_y_1 = const()[name = tensor("qk_7_transpose_y_1"), val = tensor(true)]; - tensor k_7 = transpose(perm = k_7_perm_0, x = var_766)[name = tensor("transpose_36")]; - tensor q_7 = transpose(perm = q_7_perm_0, x = var_758)[name = tensor("transpose_37")]; + tensor k_7 = transpose(perm = k_7_perm_0, x = var_802)[name = tensor("transpose_36")]; + tensor q_7 = transpose(perm = q_7_perm_0, x = var_794)[name = tensor("transpose_37")]; tensor qk_7 = matmul(transpose_x = qk_7_transpose_x_1, transpose_y = qk_7_transpose_y_1, x = q_7, y = k_7)[name = tensor("qk_7")]; - tensor var_782 = const()[name = tensor("op_782"), val = tensor([2, 1])]; - tensor var_783 = reshape(shape = var_782, x = sqrt_s_t_7)[name = tensor("op_783")]; - tensor M_7 = real_div(x = encoder__causal_mask, y = var_783)[name = tensor("M_7")]; - tensor var_785 = mul(x = qk_7, y = M_7)[name = tensor("op_785")]; + tensor var_818 = const()[name = tensor("op_818"), val = tensor([2, 1])]; + tensor var_819 = reshape(shape = var_818, x = sqrt_s_t_7)[name = tensor("op_819")]; + tensor M_7 = real_div(x = inner_encoder__causal_mask, y = var_819)[name = tensor("M_7")]; + tensor var_821 = mul(x = qk_7, y = M_7)[name = tensor("op_821")]; tensor inner_7_transpose_x_0 = const()[name = tensor("inner_7_transpose_x_0"), val = tensor(false)]; tensor inner_7_transpose_y_0 = const()[name = tensor("inner_7_transpose_y_0"), val = tensor(false)]; - tensor v_7 = transpose(perm = v_7_perm_0, x = var_772)[name = tensor("transpose_35")]; - tensor inner_7 = matmul(transpose_x = inner_7_transpose_x_0, transpose_y = inner_7_transpose_y_0, x = var_785, y = v_7)[name = tensor("inner_7")]; - tensor var_787_transpose_x_0 = const()[name = tensor("op_787_transpose_x_0"), val = tensor(false)]; - tensor var_787_transpose_y_0 = const()[name = tensor("op_787_transpose_y_0"), val = tensor(false)]; - tensor var_787 = matmul(transpose_x = var_787_transpose_x_0, transpose_y = var_787_transpose_y_0, x = q_7, y = prev_kv_7)[name = tensor("op_787")]; - tensor var_788 = real_div(x = sqrt_s0_7, y = sqrt_s_t_7)[name = tensor("op_788")]; - tensor var_789 = const()[name = tensor("op_789"), val = tensor([1, 1, 2, 1])]; - tensor var_790 = reshape(shape = var_789, x = var_788)[name = tensor("op_790")]; - tensor cross_7 = mul(x = var_787, y = var_790)[name = tensor("cross_7")]; + tensor v_7 = transpose(perm = v_7_perm_0, x = var_808)[name = tensor("transpose_35")]; + tensor inner_7 = matmul(transpose_x = inner_7_transpose_x_0, transpose_y = inner_7_transpose_y_0, x = var_821, y = v_7)[name = tensor("inner_7")]; + tensor var_823_transpose_x_0 = const()[name = tensor("op_823_transpose_x_0"), val = tensor(false)]; + tensor var_823_transpose_y_0 = const()[name = tensor("op_823_transpose_y_0"), val = tensor(false)]; + tensor var_823 = matmul(transpose_x = var_823_transpose_x_0, transpose_y = var_823_transpose_y_0, x = q_7, y = prev_kv_7)[name = tensor("op_823")]; + tensor var_824 = real_div(x = sqrt_s0_7, y = sqrt_s_t_7)[name = tensor("op_824")]; + tensor var_825 = const()[name = tensor("op_825"), val = tensor([1, 1, 2, 1])]; + tensor var_826 = reshape(shape = var_825, x = var_824)[name = tensor("op_826")]; + tensor cross_7 = mul(x = var_823, y = var_826)[name = tensor("cross_7")]; tensor out_19 = add(x = inner_7, y = cross_7)[name = tensor("out_19")]; - tensor var_793 = mul(x = prev_kv_7, y = sqrt_s0_7)[name = tensor("op_793")]; - tensor var_795_transpose_x_1 = const()[name = tensor("op_795_transpose_x_1"), val = tensor(true)]; - tensor var_795_transpose_y_1 = const()[name = tensor("op_795_transpose_y_1"), val = tensor(false)]; - tensor var_795 = matmul(transpose_x = var_795_transpose_x_1, transpose_y = var_795_transpose_y_1, x = k_7, y = v_7)[name = tensor("op_795")]; - tensor new_kv_unnorm_7 = add(x = var_793, y = var_795)[name = tensor("new_kv_unnorm_7")]; - tensor var_797 = const()[name = tensor("op_797"), val = tensor(0x1p+1)]; - tensor new_scale_7 = add(x = prev_scale_7, y = var_797)[name = tensor("new_scale_7")]; - tensor var_799 = sqrt(x = new_scale_7)[name = tensor("op_799")]; - tensor nkv_1 = real_div(x = new_kv_unnorm_7, y = var_799)[name = tensor("nkv_1")]; - tensor var_801_perm_0 = const()[name = tensor("op_801_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_829 = mul(x = prev_kv_7, y = sqrt_s0_7)[name = tensor("op_829")]; + tensor var_831_transpose_x_1 = const()[name = tensor("op_831_transpose_x_1"), val = tensor(true)]; + tensor var_831_transpose_y_1 = const()[name = tensor("op_831_transpose_y_1"), val = tensor(false)]; + tensor var_831 = matmul(transpose_x = var_831_transpose_x_1, transpose_y = var_831_transpose_y_1, x = k_7, y = v_7)[name = tensor("op_831")]; + tensor new_kv_unnorm_7 = add(x = var_829, y = var_831)[name = tensor("new_kv_unnorm_7")]; + tensor var_833 = const()[name = tensor("op_833"), val = tensor(0x1p+1)]; + tensor new_scale_7 = add(x = prev_scale_7, y = var_833)[name = tensor("new_scale_7")]; + tensor var_835 = sqrt(x = new_scale_7)[name = tensor("op_835")]; + tensor nkv_1 = real_div(x = new_kv_unnorm_7, y = var_835)[name = tensor("nkv_1")]; + tensor var_837_perm_0 = const()[name = tensor("op_837_perm_0"), val = tensor([0, 2, 1, 3])]; tensor out_21_axes_0 = const()[name = tensor("out_21_axes_0"), val = tensor([-1])]; - tensor var_801 = transpose(perm = var_801_perm_0, x = out_19)[name = tensor("transpose_34")]; - tensor out_21 = layer_norm(axes = out_21_axes_0, epsilon = var_17, x = var_801)[name = tensor("out_21")]; - tensor var_805 = const()[name = tensor("op_805"), val = tensor([1, 2, 256])]; - tensor out_23 = reshape(shape = var_805, x = out_21)[name = tensor("out_23")]; - tensor var_807 = silu(x = input_137)[name = tensor("op_807")]; - tensor input_139 = mul(x = var_807, y = out_23)[name = tensor("input_139")]; - tensor ret_out_7 = linear(bias = encoder_out_proj_3_bias, weight = encoder_out_proj_3_weight, x = input_139)[name = tensor("linear_34")]; - tensor x_21 = add(x = input_133, y = ret_out_7)[name = tensor("x_21")]; + tensor var_837 = transpose(perm = var_837_perm_0, x = out_19)[name = tensor("transpose_34")]; + tensor out_21 = layer_norm(axes = out_21_axes_0, epsilon = var_54, x = var_837)[name = tensor("out_21")]; + tensor var_841 = const()[name = tensor("op_841"), val = tensor([1, 2, 256])]; + tensor out_23 = reshape(shape = var_841, x = out_21)[name = tensor("out_23")]; + tensor var_843 = silu(x = input_139)[name = tensor("op_843")]; + tensor input_141 = mul(x = var_843, y = out_23)[name = tensor("input_141")]; + tensor ret_out_7 = linear(bias = inner_encoder_out_proj_3_bias, weight = inner_encoder_out_proj_3_weight, x = input_141)[name = tensor("linear_34")]; + tensor x_21 = add(x = input_135, y = ret_out_7)[name = tensor("x_21")]; tensor window_19_begin_0 = const()[name = tensor("window_19_begin_0"), val = tensor([3, 0, 0, 0])]; tensor window_19_end_0 = const()[name = tensor("window_19_end_0"), val = tensor([4, 1, 16, 256])]; tensor window_19_end_mask_0 = const()[name = tensor("window_19_end_mask_0"), val = tensor([false, true, true, true])]; tensor window_19_squeeze_mask_0 = const()[name = tensor("window_19_squeeze_mask_0"), val = tensor([true, false, false, false])]; tensor window_19 = slice_by_index(begin = window_19_begin_0, end = window_19_end_0, end_mask = window_19_end_mask_0, squeeze_mask = window_19_squeeze_mask_0, x = enc_conv_cache)[name = tensor("window_19")]; - tensor var_815_begin_0 = const()[name = tensor("op_815_begin_0"), val = tensor([0, 0, 0])]; - tensor var_815_end_0 = const()[name = tensor("op_815_end_0"), val = tensor([1, 1, 256])]; - tensor var_815_end_mask_0 = const()[name = tensor("op_815_end_mask_0"), val = tensor([true, false, true])]; - tensor var_815 = slice_by_index(begin = var_815_begin_0, end = var_815_end_0, end_mask = var_815_end_mask_0, x = x_21)[name = tensor("op_815")]; - tensor var_818_begin_0 = const()[name = tensor("op_818_begin_0"), val = tensor([0, 1, 0])]; - tensor var_818_end_0 = const()[name = tensor("op_818_end_0"), val = tensor([1, 16, 256])]; - tensor var_818_end_mask_0 = const()[name = tensor("op_818_end_mask_0"), val = tensor([true, true, true])]; - tensor var_818 = slice_by_index(begin = var_818_begin_0, end = var_818_end_0, end_mask = var_818_end_mask_0, x = window_19)[name = tensor("op_818")]; + tensor var_851_begin_0 = const()[name = tensor("op_851_begin_0"), val = tensor([0, 0, 0])]; + tensor var_851_end_0 = const()[name = tensor("op_851_end_0"), val = tensor([1, 1, 256])]; + tensor var_851_end_mask_0 = const()[name = tensor("op_851_end_mask_0"), val = tensor([true, false, true])]; + tensor var_851 = slice_by_index(begin = var_851_begin_0, end = var_851_end_0, end_mask = var_851_end_mask_0, x = x_21)[name = tensor("op_851")]; + tensor var_854_begin_0 = const()[name = tensor("op_854_begin_0"), val = tensor([0, 1, 0])]; + tensor var_854_end_0 = const()[name = tensor("op_854_end_0"), val = tensor([1, 16, 256])]; + tensor var_854_end_mask_0 = const()[name = tensor("op_854_end_mask_0"), val = tensor([true, true, true])]; + tensor var_854 = slice_by_index(begin = var_854_begin_0, end = var_854_end_0, end_mask = var_854_end_mask_0, x = window_19)[name = tensor("op_854")]; tensor window_21_interleave_0 = const()[name = tensor("window_21_interleave_0"), val = tensor(false)]; - tensor window_21 = concat(axis = var_26, interleave = window_21_interleave_0, values = (var_818, var_815))[name = tensor("window_21")]; - tensor var_823_begin_0 = const()[name = tensor("op_823_begin_0"), val = tensor([0, 1, 0])]; - tensor var_823_end_0 = const()[name = tensor("op_823_end_0"), val = tensor([1, 1, 256])]; - tensor var_823_end_mask_0 = const()[name = tensor("op_823_end_mask_0"), val = tensor([true, true, true])]; - tensor var_823 = slice_by_index(begin = var_823_begin_0, end = var_823_end_0, end_mask = var_823_end_mask_0, x = x_21)[name = tensor("op_823")]; - tensor var_826_begin_0 = const()[name = tensor("op_826_begin_0"), val = tensor([0, 1, 0])]; - tensor var_826_end_0 = const()[name = tensor("op_826_end_0"), val = tensor([1, 16, 256])]; - tensor var_826_end_mask_0 = const()[name = tensor("op_826_end_mask_0"), val = tensor([true, true, true])]; - tensor var_826 = slice_by_index(begin = var_826_begin_0, end = var_826_end_0, end_mask = var_826_end_mask_0, x = window_21)[name = tensor("op_826")]; + tensor window_21 = concat(axis = var_62, interleave = window_21_interleave_0, values = (var_854, var_851))[name = tensor("window_21")]; + tensor var_859_begin_0 = const()[name = tensor("op_859_begin_0"), val = tensor([0, 1, 0])]; + tensor var_859_end_0 = const()[name = tensor("op_859_end_0"), val = tensor([1, 1, 256])]; + tensor var_859_end_mask_0 = const()[name = tensor("op_859_end_mask_0"), val = tensor([true, true, true])]; + tensor var_859 = slice_by_index(begin = var_859_begin_0, end = var_859_end_0, end_mask = var_859_end_mask_0, x = x_21)[name = tensor("op_859")]; + tensor var_862_begin_0 = const()[name = tensor("op_862_begin_0"), val = tensor([0, 1, 0])]; + tensor var_862_end_0 = const()[name = tensor("op_862_end_0"), val = tensor([1, 16, 256])]; + tensor var_862_end_mask_0 = const()[name = tensor("op_862_end_mask_0"), val = tensor([true, true, true])]; + tensor var_862 = slice_by_index(begin = var_862_begin_0, end = var_862_end_0, end_mask = var_862_end_mask_0, x = window_21)[name = tensor("op_862")]; tensor window_interleave_0 = const()[name = tensor("window_interleave_0"), val = tensor(false)]; - tensor window = concat(axis = var_26, interleave = window_interleave_0, values = (var_826, var_823))[name = tensor("window")]; - tensor input_141_interleave_0 = const()[name = tensor("input_141_interleave_0"), val = tensor(false)]; - tensor input_141 = concat(axis = var_23, interleave = input_141_interleave_0, values = (window_21, window))[name = tensor("input_141")]; + tensor window = concat(axis = var_62, interleave = window_interleave_0, values = (var_862, var_859))[name = tensor("window")]; + tensor input_143_interleave_0 = const()[name = tensor("input_143_interleave_0"), val = tensor(false)]; + tensor input_143 = concat(axis = var_49, interleave = input_143_interleave_0, values = (window_21, window))[name = tensor("input_143")]; tensor x_23_axes_0 = const()[name = tensor("x_23_axes_0"), val = tensor([-1])]; - tensor x_23 = layer_norm(axes = x_23_axes_0, beta = encoder_conv_module_3_sequential_0_bias, epsilon = var_28, gamma = encoder_conv_module_3_sequential_0_weight, x = input_141)[name = tensor("x_23")]; - tensor input_143_perm_0 = const()[name = tensor("input_143_perm_0"), val = tensor([0, 2, 1])]; + tensor x_23 = layer_norm(axes = x_23_axes_0, beta = inner_encoder_conv_module_3_sequential_0_bias, epsilon = var_46, gamma = inner_encoder_conv_module_3_sequential_0_weight, x = input_143)[name = tensor("x_23")]; + tensor input_145_perm_0 = const()[name = tensor("input_145_perm_0"), val = tensor([0, 2, 1])]; tensor inputs_33_pad_type_0 = const()[name = tensor("inputs_33_pad_type_0"), val = tensor("valid")]; tensor inputs_33_strides_0 = const()[name = tensor("inputs_33_strides_0"), val = tensor([1])]; tensor inputs_33_pad_0 = const()[name = tensor("inputs_33_pad_0"), val = tensor([0, 0])]; tensor inputs_33_dilations_0 = const()[name = tensor("inputs_33_dilations_0"), val = tensor([1])]; tensor inputs_33_groups_0 = const()[name = tensor("inputs_33_groups_0"), val = tensor(1)]; - tensor input_143 = transpose(perm = input_143_perm_0, x = x_23)[name = tensor("transpose_33")]; - tensor inputs_33 = conv(bias = encoder_conv_module_3_sequential_2_conv_bias, dilations = inputs_33_dilations_0, groups = inputs_33_groups_0, pad = inputs_33_pad_0, pad_type = inputs_33_pad_type_0, strides = inputs_33_strides_0, weight = encoder_conv_module_3_sequential_2_conv_weight, x = input_143)[name = tensor("inputs_33")]; - tensor var_851_split_sizes_0 = const()[name = tensor("op_851_split_sizes_0"), val = tensor([256, 256])]; - tensor var_851_axis_0 = const()[name = tensor("op_851_axis_0"), val = tensor(1)]; - tensor var_851_0, tensor var_851_1 = split(axis = var_851_axis_0, split_sizes = var_851_split_sizes_0, x = inputs_33)[name = tensor("op_851")]; - tensor var_853 = sigmoid(x = var_851_1)[name = tensor("op_853")]; - tensor inputs_35 = mul(x = var_851_0, y = var_853)[name = tensor("inputs_35")]; + tensor input_145 = transpose(perm = input_145_perm_0, x = x_23)[name = tensor("transpose_33")]; + tensor inputs_33 = conv(bias = inner_encoder_conv_module_3_sequential_2_conv_bias, dilations = inputs_33_dilations_0, groups = inputs_33_groups_0, pad = inputs_33_pad_0, pad_type = inputs_33_pad_type_0, strides = inputs_33_strides_0, weight = inner_encoder_conv_module_3_sequential_2_conv_weight, x = input_145)[name = tensor("inputs_33")]; + tensor var_887_split_sizes_0 = const()[name = tensor("op_887_split_sizes_0"), val = tensor([256, 256])]; + tensor var_887_axis_0 = const()[name = tensor("op_887_axis_0"), val = tensor(1)]; + tensor var_887_0, tensor var_887_1 = split(axis = var_887_axis_0, split_sizes = var_887_split_sizes_0, x = inputs_33)[name = tensor("op_887")]; + tensor var_889 = sigmoid(x = var_887_1)[name = tensor("op_889")]; + tensor inputs_35 = mul(x = var_887_0, y = var_889)[name = tensor("inputs_35")]; tensor outputs_aug_pad_type_0 = const()[name = tensor("outputs_aug_pad_type_0"), val = tensor("custom")]; tensor outputs_aug_pad_0 = const()[name = tensor("outputs_aug_pad_0"), val = tensor([15, 15])]; tensor outputs_aug_groups_0 = const()[name = tensor("outputs_aug_groups_0"), val = tensor(256)]; tensor outputs_aug_strides_0 = const()[name = tensor("outputs_aug_strides_0"), val = tensor([1])]; tensor outputs_aug_dilations_0 = const()[name = tensor("outputs_aug_dilations_0"), val = tensor([1])]; - tensor outputs_aug = conv(dilations = outputs_aug_dilations_0, groups = outputs_aug_groups_0, pad = outputs_aug_pad_0, pad_type = outputs_aug_pad_type_0, strides = outputs_aug_strides_0, weight = encoder_conv_module_3_sequential_4_conv_weight, x = inputs_35)[name = tensor("outputs_aug")]; - tensor input_145_begin_0 = const()[name = tensor("input_145_begin_0"), val = tensor([0, 0, 0])]; - tensor input_145_end_0 = const()[name = tensor("input_145_end_0"), val = tensor([2, 256, 16])]; - tensor input_145_end_mask_0 = const()[name = tensor("input_145_end_mask_0"), val = tensor([true, true, false])]; - tensor input_145 = slice_by_index(begin = input_145_begin_0, end = input_145_end_0, end_mask = input_145_end_mask_0, x = outputs_aug)[name = tensor("input_145")]; - tensor inputs_37 = batch_norm(beta = encoder_conv_module_3_sequential_5_bias, epsilon = var_28, gamma = encoder_conv_module_3_sequential_5_weight, mean = encoder_conv_module_3_sequential_5_running_mean, variance = encoder_conv_module_3_sequential_5_running_var, x = input_145)[name = tensor("inputs_37")]; - tensor input_147 = silu(x = inputs_37)[name = tensor("input_147")]; - tensor input_149_pad_type_0 = const()[name = tensor("input_149_pad_type_0"), val = tensor("valid")]; - tensor input_149_strides_0 = const()[name = tensor("input_149_strides_0"), val = tensor([1])]; - tensor input_149_pad_0 = const()[name = tensor("input_149_pad_0"), val = tensor([0, 0])]; - tensor input_149_dilations_0 = const()[name = tensor("input_149_dilations_0"), val = tensor([1])]; - tensor input_149_groups_0 = const()[name = tensor("input_149_groups_0"), val = tensor(1)]; - tensor input_149 = conv(bias = encoder_conv_module_3_sequential_7_conv_bias, dilations = input_149_dilations_0, groups = input_149_groups_0, pad = input_149_pad_0, pad_type = input_149_pad_type_0, strides = input_149_strides_0, weight = encoder_conv_module_3_sequential_7_conv_weight, x = input_147)[name = tensor("input_149")]; + tensor outputs_aug = conv(dilations = outputs_aug_dilations_0, groups = outputs_aug_groups_0, pad = outputs_aug_pad_0, pad_type = outputs_aug_pad_type_0, strides = outputs_aug_strides_0, weight = inner_encoder_conv_module_3_sequential_4_conv_weight, x = inputs_35)[name = tensor("outputs_aug")]; + tensor input_147_begin_0 = const()[name = tensor("input_147_begin_0"), val = tensor([0, 0, 0])]; + tensor input_147_end_0 = const()[name = tensor("input_147_end_0"), val = tensor([2, 256, 16])]; + tensor input_147_end_mask_0 = const()[name = tensor("input_147_end_mask_0"), val = tensor([true, true, false])]; + tensor input_147 = slice_by_index(begin = input_147_begin_0, end = input_147_end_0, end_mask = input_147_end_mask_0, x = outputs_aug)[name = tensor("input_147")]; + tensor inputs_37 = batch_norm(beta = inner_encoder_conv_module_3_sequential_5_bias, epsilon = var_46, gamma = inner_encoder_conv_module_3_sequential_5_weight, mean = inner_encoder_conv_module_3_sequential_5_running_mean, variance = inner_encoder_conv_module_3_sequential_5_running_var, x = input_147)[name = tensor("inputs_37")]; + tensor input_149 = silu(x = inputs_37)[name = tensor("input_149")]; + tensor input_151_pad_type_0 = const()[name = tensor("input_151_pad_type_0"), val = tensor("valid")]; + tensor input_151_strides_0 = const()[name = tensor("input_151_strides_0"), val = tensor([1])]; + tensor input_151_pad_0 = const()[name = tensor("input_151_pad_0"), val = tensor([0, 0])]; + tensor input_151_dilations_0 = const()[name = tensor("input_151_dilations_0"), val = tensor([1])]; + tensor input_151_groups_0 = const()[name = tensor("input_151_groups_0"), val = tensor(1)]; + tensor input_151 = conv(bias = inner_encoder_conv_module_3_sequential_7_conv_bias, dilations = input_151_dilations_0, groups = input_151_groups_0, pad = input_151_pad_0, pad_type = input_151_pad_type_0, strides = input_151_strides_0, weight = inner_encoder_conv_module_3_sequential_7_conv_weight, x = input_149)[name = tensor("input_151")]; tensor conv_out_7_perm_0 = const()[name = tensor("conv_out_7_perm_0"), val = tensor([0, 2, 1])]; - tensor var_884_begin_0 = const()[name = tensor("op_884_begin_0"), val = tensor([0, -1, 0])]; - tensor var_884_end_0 = const()[name = tensor("op_884_end_0"), val = tensor([2, 16, 256])]; - tensor var_884_end_mask_0 = const()[name = tensor("op_884_end_mask_0"), val = tensor([true, true, true])]; - tensor conv_out_7 = transpose(perm = conv_out_7_perm_0, x = input_149)[name = tensor("transpose_32")]; - tensor var_884 = slice_by_index(begin = var_884_begin_0, end = var_884_end_0, end_mask = var_884_end_mask_0, x = conv_out_7)[name = tensor("op_884")]; - tensor var_886_perm_0 = const()[name = tensor("op_886_perm_0"), val = tensor([1, 0, 2])]; - tensor var_886 = transpose(perm = var_886_perm_0, x = var_884)[name = tensor("transpose_31")]; - tensor input_151 = add(x = x_21, y = var_886)[name = tensor("input_151")]; - tensor input_153_axes_0 = const()[name = tensor("input_153_axes_0"), val = tensor([-1])]; - tensor input_153 = layer_norm(axes = input_153_axes_0, beta = encoder_ffn2_3_module_sequential_0_bias, epsilon = var_28, gamma = encoder_ffn2_3_module_sequential_0_weight, x = input_151)[name = tensor("input_153")]; - tensor inputs = linear(bias = encoder_ffn2_3_module_sequential_1_linear_bias, weight = encoder_ffn2_3_module_sequential_1_linear_weight, x = input_153)[name = tensor("linear_35")]; - tensor input_155 = silu(x = inputs)[name = tensor("input_155")]; - tensor input_159 = linear(bias = encoder_ffn2_3_module_sequential_4_linear_bias, weight = encoder_ffn2_3_module_sequential_4_linear_weight, x = input_155)[name = tensor("linear_36")]; - tensor var_909 = const()[name = tensor("op_909"), val = tensor(0x1p-1)]; - tensor var_910 = mul(x = input_159, y = var_909)[name = tensor("op_910")]; - tensor input_161 = add(x = var_910, y = input_151)[name = tensor("input_161")]; + tensor var_920_begin_0 = const()[name = tensor("op_920_begin_0"), val = tensor([0, -1, 0])]; + tensor var_920_end_0 = const()[name = tensor("op_920_end_0"), val = tensor([2, 16, 256])]; + tensor var_920_end_mask_0 = const()[name = tensor("op_920_end_mask_0"), val = tensor([true, true, true])]; + tensor conv_out_7 = transpose(perm = conv_out_7_perm_0, x = input_151)[name = tensor("transpose_32")]; + tensor var_920 = slice_by_index(begin = var_920_begin_0, end = var_920_end_0, end_mask = var_920_end_mask_0, x = conv_out_7)[name = tensor("op_920")]; + tensor var_922_perm_0 = const()[name = tensor("op_922_perm_0"), val = tensor([1, 0, 2])]; + tensor var_922 = transpose(perm = var_922_perm_0, x = var_920)[name = tensor("transpose_31")]; + tensor input_153 = add(x = x_21, y = var_922)[name = tensor("input_153")]; + tensor input_155_axes_0 = const()[name = tensor("input_155_axes_0"), val = tensor([-1])]; + tensor input_155 = layer_norm(axes = input_155_axes_0, beta = inner_encoder_ffn2_3_module_sequential_0_bias, epsilon = var_46, gamma = inner_encoder_ffn2_3_module_sequential_0_weight, x = input_153)[name = tensor("input_155")]; + tensor inputs = linear(bias = inner_encoder_ffn2_3_module_sequential_1_linear_bias, weight = inner_encoder_ffn2_3_module_sequential_1_linear_weight, x = input_155)[name = tensor("linear_35")]; + tensor input_157 = silu(x = inputs)[name = tensor("input_157")]; + tensor input_161 = linear(bias = inner_encoder_ffn2_3_module_sequential_4_linear_bias, weight = inner_encoder_ffn2_3_module_sequential_4_linear_weight, x = input_157)[name = tensor("linear_36")]; + tensor var_945 = const()[name = tensor("op_945"), val = tensor(0x1p-1)]; + tensor var_946 = mul(x = input_161, y = var_945)[name = tensor("op_946")]; + tensor input_163 = add(x = var_946, y = input_153)[name = tensor("input_163")]; tensor x_25_axes_0 = const()[name = tensor("x_25_axes_0"), val = tensor([-1])]; - tensor x_25 = layer_norm(axes = x_25_axes_0, beta = encoder_layer_norm_3_bias, epsilon = var_28, gamma = encoder_layer_norm_3_weight, x = input_161)[name = tensor("x_25")]; + tensor x_25 = layer_norm(axes = x_25_axes_0, beta = inner_encoder_layer_norm_3_bias, epsilon = var_46, gamma = inner_encoder_layer_norm_3_weight, x = input_163)[name = tensor("x_25")]; tensor x_ct_perm_0 = const()[name = tensor("x_ct_perm_0"), val = tensor([0, 2, 1])]; tensor cat_interleave_0 = const()[name = tensor("cat_interleave_0"), val = tensor(false)]; tensor x_ct = transpose(perm = x_ct_perm_0, x = x_25)[name = tensor("transpose_30")]; - tensor cat = concat(axis = var_20, interleave = cat_interleave_0, values = (cnn_window, x_ct))[name = tensor("cat")]; + tensor cat = concat(axis = var_51, interleave = cat_interleave_0, values = (cnn_window, x_ct))[name = tensor("cat")]; tensor conv_out_pad_type_0 = const()[name = tensor("conv_out_pad_type_0"), val = tensor("valid")]; tensor conv_out_strides_0 = const()[name = tensor("conv_out_strides_0"), val = tensor([1])]; tensor conv_out_pad_0 = const()[name = tensor("conv_out_pad_0"), val = tensor([0, 0])]; tensor conv_out_dilations_0 = const()[name = tensor("conv_out_dilations_0"), val = tensor([1])]; tensor conv_out_groups_0 = const()[name = tensor("conv_out_groups_0"), val = tensor(1)]; - tensor conv_out = conv(bias = encoder_cnn_bias, dilations = conv_out_dilations_0, groups = conv_out_groups_0, pad = conv_out_pad_0, pad_type = conv_out_pad_type_0, strides = conv_out_strides_0, weight = encoder_cnn_weight, x = cat)[name = tensor("conv_out")]; - tensor var_928_begin_0 = const()[name = tensor("op_928_begin_0"), val = tensor([0, 0, 2])]; - tensor var_928_end_0 = const()[name = tensor("op_928_end_0"), val = tensor([1, 256, 20])]; - tensor var_928_end_mask_0 = const()[name = tensor("op_928_end_mask_0"), val = tensor([true, true, true])]; - tensor cnn_window_new = slice_by_index(begin = var_928_begin_0, end = var_928_end_0, end_mask = var_928_end_mask_0, x = cat)[name = tensor("op_928")]; - tensor input_163_perm_0 = const()[name = tensor("input_163_perm_0"), val = tensor([0, 2, 1])]; - tensor var_930 = const()[name = tensor("op_930"), val = tensor([-1])]; - tensor input_163 = transpose(perm = input_163_perm_0, x = conv_out)[name = tensor("transpose_29")]; - tensor var_931 = reduce_l2_norm(axes = var_930, keep_dims = var_29, x = input_163)[name = tensor("op_931")]; + tensor conv_out = conv(bias = inner_encoder_cnn_bias, dilations = conv_out_dilations_0, groups = conv_out_groups_0, pad = conv_out_pad_0, pad_type = conv_out_pad_type_0, strides = conv_out_strides_0, weight = inner_encoder_cnn_weight, x = cat)[name = tensor("conv_out")]; + tensor var_964_begin_0 = const()[name = tensor("op_964_begin_0"), val = tensor([0, 0, 2])]; + tensor var_964_end_0 = const()[name = tensor("op_964_end_0"), val = tensor([1, 256, 20])]; + tensor var_964_end_mask_0 = const()[name = tensor("op_964_end_mask_0"), val = tensor([true, true, true])]; + tensor cnn_window_new = slice_by_index(begin = var_964_begin_0, end = var_964_end_0, end_mask = var_964_end_mask_0, x = cat)[name = tensor("op_964")]; + tensor input_165_perm_0 = const()[name = tensor("input_165_perm_0"), val = tensor([0, 2, 1])]; + tensor var_966 = const()[name = tensor("op_966"), val = tensor([-1])]; + tensor input_165 = transpose(perm = input_165_perm_0, x = conv_out)[name = tensor("transpose_29")]; + tensor var_967 = reduce_l2_norm(axes = var_966, keep_dims = var_45, x = input_165)[name = tensor("op_967")]; tensor const_12 = const()[name = tensor("const_12"), val = tensor(0x1.fffffep+127)]; - tensor clip_0 = clip(alpha = var_11, beta = const_12, x = var_931)[name = tensor("clip_0")]; - tensor emb = real_div(x = input_163, y = clip_0)[name = tensor("emb")]; - tensor var_935_axis_0 = const()[name = tensor("op_935_axis_0"), val = tensor(0)]; - tensor enc_kv_new = stack(axis = var_935_axis_0, values = (var_206, var_404, var_602, nkv_1))[name = tensor("op_935")]; - tensor var_937_axis_0 = const()[name = tensor("op_937_axis_0"), val = tensor(0)]; - tensor enc_scale_new = stack(axis = var_937_axis_0, values = (new_scale_1, new_scale_3, new_scale_5, new_scale_7))[name = tensor("op_937")]; - tensor var_939_axis_0 = const()[name = tensor("op_939_axis_0"), val = tensor(0)]; - tensor enc_conv_cache_new = stack(axis = var_939_axis_0, values = (window_5, window_11, window_17, window))[name = tensor("op_939")]; - tensor var_948 = const()[name = tensor("op_948"), val = tensor(0x1.5798eep-27)]; - tensor var_953 = const()[name = tensor("op_953"), val = tensor(0x1.4f8b58p-17)]; - tensor var_955 = const()[name = tensor("op_955"), val = tensor(0x1.0c6f7ap-20)]; - tensor var_956 = const()[name = tensor("op_956"), val = tensor(true)]; - tensor var_958 = const()[name = tensor("op_958"), val = tensor(0x1p+0)]; - tensor var_962 = const()[name = tensor("op_962"), val = tensor(-1)]; - tensor var_968 = const()[name = tensor("op_968"), val = tensor(0)]; + tensor clip_0 = clip(alpha = var_59, beta = const_12, x = var_967)[name = tensor("clip_0")]; + tensor emb = real_div(x = input_165, y = clip_0)[name = tensor("emb")]; + tensor var_971_axis_0 = const()[name = tensor("op_971_axis_0"), val = tensor(0)]; + tensor enc_kv_new = stack(axis = var_971_axis_0, values = (var_242, var_440, var_638, nkv_1))[name = tensor("op_971")]; + tensor var_973_axis_0 = const()[name = tensor("op_973_axis_0"), val = tensor(0)]; + tensor enc_scale_new = stack(axis = var_973_axis_0, values = (new_scale_1, new_scale_3, new_scale_5, new_scale_7))[name = tensor("op_973")]; + tensor var_975_axis_0 = const()[name = tensor("op_975_axis_0"), val = tensor(0)]; + tensor enc_conv_cache_new = stack(axis = var_975_axis_0, values = (window_5, window_11, window_17, window))[name = tensor("op_975")]; tensor pos = const()[name = tensor("pos"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(44395328)))]; - tensor var_1030_axes_0 = const()[name = tensor("op_1030_axes_0"), val = tensor([2])]; - tensor var_1030 = expand_dims(axes = var_1030_axes_0, x = emb)[name = tensor("op_1030")]; + tensor var_1043_axes_0 = const()[name = tensor("op_1043_axes_0"), val = tensor([2])]; + tensor var_1043 = expand_dims(axes = var_1043_axes_0, x = emb)[name = tensor("op_1043")]; tensor emb_exp_reps_0 = const()[name = tensor("emb_exp_reps_0"), val = tensor([1, 1, 9, 1])]; - tensor emb_exp = tile(reps = emb_exp_reps_0, x = var_1030)[name = tensor("emb_exp")]; - tensor input_165_interleave_0 = const()[name = tensor("input_165_interleave_0"), val = tensor(false)]; - tensor input_165 = concat(axis = var_962, interleave = input_165_interleave_0, values = (emb_exp, pos))[name = tensor("input_165")]; - tensor x_27 = linear(bias = decoder_convert_bias, weight = decoder_convert_weight, x = input_165)[name = tensor("linear_37")]; - tensor var_1038_perm_0 = const()[name = tensor("op_1038_perm_0"), val = tensor([0, 2, 1, 3])]; - tensor var_1042 = const()[name = tensor("op_1042"), val = tensor([9, 2, 256])]; - tensor var_1038 = transpose(perm = var_1038_perm_0, x = x_27)[name = tensor("transpose_28")]; - tensor x_29 = reshape(shape = var_1042, x = var_1038)[name = tensor("x_29")]; + tensor emb_exp = tile(reps = emb_exp_reps_0, x = var_1043)[name = tensor("emb_exp")]; + tensor input_167_interleave_0 = const()[name = tensor("input_167_interleave_0"), val = tensor(false)]; + tensor input_167 = concat(axis = var_52, interleave = input_167_interleave_0, values = (emb_exp, pos))[name = tensor("input_167")]; + tensor x_27 = linear(bias = inner_decoder_convert_bias, weight = inner_decoder_convert_weight, x = input_167)[name = tensor("linear_37")]; + tensor var_1051_perm_0 = const()[name = tensor("op_1051_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_1055 = const()[name = tensor("op_1055"), val = tensor([9, 2, 256])]; + tensor var_1051 = transpose(perm = var_1051_perm_0, x = x_27)[name = tensor("transpose_28")]; + tensor x_29 = reshape(shape = var_1055, x = var_1051)[name = tensor("x_29")]; tensor prev_kv_9_begin_0 = const()[name = tensor("prev_kv_9_begin_0"), val = tensor([0, 0, 0, 0, 0])]; tensor prev_kv_9_end_0 = const()[name = tensor("prev_kv_9_end_0"), val = tensor([1, 9, 4, 64, 64])]; tensor prev_kv_9_end_mask_0 = const()[name = tensor("prev_kv_9_end_mask_0"), val = tensor([false, true, true, true, true])]; @@ -903,132 +910,132 @@ program(1.0) tensor prev_scale_9_end_mask_0 = const()[name = tensor("prev_scale_9_end_mask_0"), val = tensor([false, true])]; tensor prev_scale_9_squeeze_mask_0 = const()[name = tensor("prev_scale_9_squeeze_mask_0"), val = tensor([true, false])]; tensor prev_scale_9 = slice_by_index(begin = prev_scale_9_begin_0, end = prev_scale_9_end_0, end_mask = prev_scale_9_end_mask_0, squeeze_mask = prev_scale_9_squeeze_mask_0, x = dec_scale)[name = tensor("prev_scale_9")]; - tensor var_1050 = linear(bias = decoder_q_proj_0_bias, weight = decoder_q_proj_0_weight, x = x_29)[name = tensor("linear_38")]; - tensor var_1051 = const()[name = tensor("op_1051"), val = tensor([9, 2, 4, 64])]; - tensor var_1052 = reshape(shape = var_1051, x = var_1050)[name = tensor("op_1052")]; + tensor var_1063 = linear(bias = inner_decoder_q_proj_0_bias, weight = inner_decoder_q_proj_0_weight, x = x_29)[name = tensor("linear_38")]; + tensor var_1064 = const()[name = tensor("op_1064"), val = tensor([9, 2, 4, 64])]; + tensor var_1065 = reshape(shape = var_1064, x = var_1063)[name = tensor("op_1065")]; tensor q_9_perm_0 = const()[name = tensor("q_9_perm_0"), val = tensor([0, 2, 1, 3])]; - tensor var_1056 = linear(bias = decoder_k_proj_0_bias, weight = decoder_k_proj_0_weight, x = x_29)[name = tensor("linear_39")]; - tensor var_1057 = const()[name = tensor("op_1057"), val = tensor(0x1p-3)]; - tensor var_1058 = mul(x = var_1056, y = var_1057)[name = tensor("op_1058")]; - tensor var_1059 = const()[name = tensor("op_1059"), val = tensor([9, 2, 4, 64])]; - tensor var_1060 = reshape(shape = var_1059, x = var_1058)[name = tensor("op_1060")]; + tensor var_1069 = linear(bias = inner_decoder_k_proj_0_bias, weight = inner_decoder_k_proj_0_weight, x = x_29)[name = tensor("linear_39")]; + tensor var_1070 = const()[name = tensor("op_1070"), val = tensor(0x1p-3)]; + tensor var_1071 = mul(x = var_1069, y = var_1070)[name = tensor("op_1071")]; + tensor var_1072 = const()[name = tensor("op_1072"), val = tensor([9, 2, 4, 64])]; + tensor var_1073 = reshape(shape = var_1072, x = var_1071)[name = tensor("op_1073")]; tensor k_9_perm_0 = const()[name = tensor("k_9_perm_0"), val = tensor([0, 2, 1, 3])]; - tensor var_1064 = linear(bias = decoder_v_proj_0_bias, weight = decoder_v_proj_0_weight, x = x_29)[name = tensor("linear_40")]; - tensor var_1065 = const()[name = tensor("op_1065"), val = tensor([9, 2, 4, 64])]; - tensor var_1066 = reshape(shape = var_1065, x = var_1064)[name = tensor("op_1066")]; + tensor var_1077 = linear(bias = inner_decoder_v_proj_0_bias, weight = inner_decoder_v_proj_0_weight, x = x_29)[name = tensor("linear_40")]; + tensor var_1078 = const()[name = tensor("op_1078"), val = tensor([9, 2, 4, 64])]; + tensor var_1079 = reshape(shape = var_1078, x = var_1077)[name = tensor("op_1079")]; tensor v_9_perm_0 = const()[name = tensor("v_9_perm_0"), val = tensor([0, 2, 1, 3])]; - tensor input_169 = linear(bias = decoder_g_proj_0_bias, weight = decoder_g_proj_0_weight, x = x_29)[name = tensor("linear_41")]; + tensor input_171 = linear(bias = inner_decoder_g_proj_0_bias, weight = inner_decoder_g_proj_0_weight, x = x_29)[name = tensor("linear_41")]; tensor cumsum_mask_1_exclusive_0 = const()[name = tensor("cumsum_mask_1_exclusive_0"), val = tensor(false)]; tensor cumsum_mask_1_reverse_0 = const()[name = tensor("cumsum_mask_1_reverse_0"), val = tensor(false)]; - tensor cumsum_mask_1 = cumsum(axis = var_968, exclusive = cumsum_mask_1_exclusive_0, reverse = cumsum_mask_1_reverse_0, x = valid_mask)[name = tensor("cumsum_mask_1")]; + tensor cumsum_mask_1 = cumsum(axis = var_49, exclusive = cumsum_mask_1_exclusive_0, reverse = cumsum_mask_1_reverse_0, x = valid_mask)[name = tensor("cumsum_mask_1")]; tensor sqrt_s0_9 = sqrt(x = prev_scale_9)[name = tensor("sqrt_s0_9")]; tensor s_t_9 = add(x = prev_scale_9, y = cumsum_mask_1)[name = tensor("s_t_9")]; tensor const_20 = const()[name = tensor("const_20"), val = tensor(0x1.fffffep+127)]; - tensor clip_1 = clip(alpha = var_958, beta = const_20, x = s_t_9)[name = tensor("clip_1")]; + tensor clip_1 = clip(alpha = var_39, beta = const_20, x = s_t_9)[name = tensor("clip_1")]; tensor sqrt_s_t_9 = sqrt(x = clip_1)[name = tensor("sqrt_s_t_9")]; tensor qk_9_transpose_x_1 = const()[name = tensor("qk_9_transpose_x_1"), val = tensor(false)]; tensor qk_9_transpose_y_1 = const()[name = tensor("qk_9_transpose_y_1"), val = tensor(true)]; - tensor k_9 = transpose(perm = k_9_perm_0, x = var_1060)[name = tensor("transpose_26")]; - tensor q_9 = transpose(perm = q_9_perm_0, x = var_1052)[name = tensor("transpose_27")]; + tensor k_9 = transpose(perm = k_9_perm_0, x = var_1073)[name = tensor("transpose_26")]; + tensor q_9 = transpose(perm = q_9_perm_0, x = var_1065)[name = tensor("transpose_27")]; tensor qk_9 = matmul(transpose_x = qk_9_transpose_x_1, transpose_y = qk_9_transpose_y_1, x = q_9, y = k_9)[name = tensor("qk_9")]; - tensor var_1078 = const()[name = tensor("op_1078"), val = tensor([1, 2])]; - tensor var_1079 = reshape(shape = var_1078, x = valid_mask)[name = tensor("op_1079")]; - tensor causal_with_valid_1 = mul(x = decoder__causal_mask, y = var_1079)[name = tensor("causal_with_valid_1")]; - tensor var_1081 = const()[name = tensor("op_1081"), val = tensor([2, 1])]; - tensor var_1082 = reshape(shape = var_1081, x = sqrt_s_t_9)[name = tensor("op_1082")]; - tensor M_9 = real_div(x = causal_with_valid_1, y = var_1082)[name = tensor("M_9")]; - tensor var_1084 = mul(x = qk_9, y = M_9)[name = tensor("op_1084")]; + tensor var_1091 = const()[name = tensor("op_1091"), val = tensor([1, 2])]; + tensor var_1092 = reshape(shape = var_1091, x = valid_mask)[name = tensor("op_1092")]; + tensor causal_with_valid_1 = mul(x = inner_decoder__causal_mask, y = var_1092)[name = tensor("causal_with_valid_1")]; + tensor var_1094 = const()[name = tensor("op_1094"), val = tensor([2, 1])]; + tensor var_1095 = reshape(shape = var_1094, x = sqrt_s_t_9)[name = tensor("op_1095")]; + tensor M_9 = real_div(x = causal_with_valid_1, y = var_1095)[name = tensor("M_9")]; + tensor var_1097 = mul(x = qk_9, y = M_9)[name = tensor("op_1097")]; tensor inner_9_transpose_x_0 = const()[name = tensor("inner_9_transpose_x_0"), val = tensor(false)]; tensor inner_9_transpose_y_0 = const()[name = tensor("inner_9_transpose_y_0"), val = tensor(false)]; - tensor v_9 = transpose(perm = v_9_perm_0, x = var_1066)[name = tensor("transpose_25")]; - tensor inner_9 = matmul(transpose_x = inner_9_transpose_x_0, transpose_y = inner_9_transpose_y_0, x = var_1084, y = v_9)[name = tensor("inner_9")]; - tensor var_1086_transpose_x_0 = const()[name = tensor("op_1086_transpose_x_0"), val = tensor(false)]; - tensor var_1086_transpose_y_0 = const()[name = tensor("op_1086_transpose_y_0"), val = tensor(false)]; - tensor var_1086 = matmul(transpose_x = var_1086_transpose_x_0, transpose_y = var_1086_transpose_y_0, x = q_9, y = prev_kv_9)[name = tensor("op_1086")]; - tensor var_1087 = real_div(x = sqrt_s0_9, y = sqrt_s_t_9)[name = tensor("op_1087")]; - tensor var_1088 = const()[name = tensor("op_1088"), val = tensor([1, 1, 2, 1])]; - tensor var_1089 = reshape(shape = var_1088, x = var_1087)[name = tensor("op_1089")]; - tensor cross_9 = mul(x = var_1086, y = var_1089)[name = tensor("cross_9")]; + tensor v_9 = transpose(perm = v_9_perm_0, x = var_1079)[name = tensor("transpose_25")]; + tensor inner_9 = matmul(transpose_x = inner_9_transpose_x_0, transpose_y = inner_9_transpose_y_0, x = var_1097, y = v_9)[name = tensor("inner_9")]; + tensor var_1099_transpose_x_0 = const()[name = tensor("op_1099_transpose_x_0"), val = tensor(false)]; + tensor var_1099_transpose_y_0 = const()[name = tensor("op_1099_transpose_y_0"), val = tensor(false)]; + tensor var_1099 = matmul(transpose_x = var_1099_transpose_x_0, transpose_y = var_1099_transpose_y_0, x = q_9, y = prev_kv_9)[name = tensor("op_1099")]; + tensor var_1100 = real_div(x = sqrt_s0_9, y = sqrt_s_t_9)[name = tensor("op_1100")]; + tensor var_1101 = const()[name = tensor("op_1101"), val = tensor([1, 1, 2, 1])]; + tensor var_1102 = reshape(shape = var_1101, x = var_1100)[name = tensor("op_1102")]; + tensor cross_9 = mul(x = var_1099, y = var_1102)[name = tensor("cross_9")]; tensor out_25 = add(x = inner_9, y = cross_9)[name = tensor("out_25")]; - tensor var_1092 = const()[name = tensor("op_1092"), val = tensor([1, 1, 2, 1])]; - tensor var_1093 = reshape(shape = var_1092, x = valid_mask)[name = tensor("op_1093")]; - tensor v_masked_1 = mul(x = v_9, y = var_1093)[name = tensor("v_masked_1")]; - tensor var_1095 = mul(x = prev_kv_9, y = sqrt_s0_9)[name = tensor("op_1095")]; - tensor var_1097_transpose_x_1 = const()[name = tensor("op_1097_transpose_x_1"), val = tensor(true)]; - tensor var_1097_transpose_y_1 = const()[name = tensor("op_1097_transpose_y_1"), val = tensor(false)]; - tensor var_1097 = matmul(transpose_x = var_1097_transpose_x_1, transpose_y = var_1097_transpose_y_1, x = k_9, y = v_masked_1)[name = tensor("op_1097")]; - tensor new_kv_unnorm_9 = add(x = var_1095, y = var_1097)[name = tensor("new_kv_unnorm_9")]; - tensor var_1099_keep_dims_0 = const()[name = tensor("op_1099_keep_dims_0"), val = tensor(false)]; - tensor var_1099 = reduce_sum(keep_dims = var_1099_keep_dims_0, x = valid_mask)[name = tensor("op_1099")]; - tensor var_1100 = const()[name = tensor("op_1100"), val = tensor([1])]; - tensor var_1101 = reshape(shape = var_1100, x = var_1099)[name = tensor("op_1101")]; - tensor new_scale_9 = add(x = prev_scale_9, y = var_1101)[name = tensor("new_scale_9")]; + tensor var_1105 = const()[name = tensor("op_1105"), val = tensor([1, 1, 2, 1])]; + tensor var_1106 = reshape(shape = var_1105, x = valid_mask)[name = tensor("op_1106")]; + tensor v_masked_1 = mul(x = v_9, y = var_1106)[name = tensor("v_masked_1")]; + tensor var_1108 = mul(x = prev_kv_9, y = sqrt_s0_9)[name = tensor("op_1108")]; + tensor var_1110_transpose_x_1 = const()[name = tensor("op_1110_transpose_x_1"), val = tensor(true)]; + tensor var_1110_transpose_y_1 = const()[name = tensor("op_1110_transpose_y_1"), val = tensor(false)]; + tensor var_1110 = matmul(transpose_x = var_1110_transpose_x_1, transpose_y = var_1110_transpose_y_1, x = k_9, y = v_masked_1)[name = tensor("op_1110")]; + tensor new_kv_unnorm_9 = add(x = var_1108, y = var_1110)[name = tensor("new_kv_unnorm_9")]; + tensor var_1112_keep_dims_0 = const()[name = tensor("op_1112_keep_dims_0"), val = tensor(false)]; + tensor var_1112 = reduce_sum(keep_dims = var_1112_keep_dims_0, x = valid_mask)[name = tensor("op_1112")]; + tensor var_1113 = const()[name = tensor("op_1113"), val = tensor([1])]; + tensor var_1114 = reshape(shape = var_1113, x = var_1112)[name = tensor("op_1114")]; + tensor new_scale_9 = add(x = prev_scale_9, y = var_1114)[name = tensor("new_scale_9")]; tensor const_21 = const()[name = tensor("const_21"), val = tensor(0x1.fffffep+127)]; - tensor clip_2 = clip(alpha = var_958, beta = const_21, x = new_scale_9)[name = tensor("clip_2")]; + tensor clip_2 = clip(alpha = var_39, beta = const_21, x = new_scale_9)[name = tensor("clip_2")]; tensor sqrt_new_scale_1 = sqrt(x = clip_2)[name = tensor("sqrt_new_scale_1")]; - tensor var_1105 = real_div(x = new_kv_unnorm_9, y = sqrt_new_scale_1)[name = tensor("op_1105")]; - tensor var_1106_perm_0 = const()[name = tensor("op_1106_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_1118 = real_div(x = new_kv_unnorm_9, y = sqrt_new_scale_1)[name = tensor("op_1118")]; + tensor var_1119_perm_0 = const()[name = tensor("op_1119_perm_0"), val = tensor([0, 2, 1, 3])]; tensor out_27_axes_0 = const()[name = tensor("out_27_axes_0"), val = tensor([-1])]; - tensor var_1106 = transpose(perm = var_1106_perm_0, x = out_25)[name = tensor("transpose_24")]; - tensor out_27 = layer_norm(axes = out_27_axes_0, epsilon = var_955, x = var_1106)[name = tensor("out_27")]; - tensor var_1110 = const()[name = tensor("op_1110"), val = tensor([9, 2, 256])]; - tensor out_29 = reshape(shape = var_1110, x = out_27)[name = tensor("out_29")]; - tensor var_1112 = silu(x = input_169)[name = tensor("op_1112")]; - tensor input_171 = mul(x = var_1112, y = out_29)[name = tensor("input_171")]; - tensor ret_out_9 = linear(bias = decoder_out_proj_0_bias, weight = decoder_out_proj_0_weight, x = input_171)[name = tensor("linear_42")]; - tensor input_173 = add(x = x_29, y = ret_out_9)[name = tensor("input_173")]; + tensor var_1119 = transpose(perm = var_1119_perm_0, x = out_25)[name = tensor("transpose_24")]; + tensor out_27 = layer_norm(axes = out_27_axes_0, epsilon = var_54, x = var_1119)[name = tensor("out_27")]; + tensor var_1123 = const()[name = tensor("op_1123"), val = tensor([9, 2, 256])]; + tensor out_29 = reshape(shape = var_1123, x = out_27)[name = tensor("out_29")]; + tensor var_1125 = silu(x = input_171)[name = tensor("op_1125")]; + tensor input_173 = mul(x = var_1125, y = out_29)[name = tensor("input_173")]; + tensor ret_out_9 = linear(bias = inner_decoder_out_proj_0_bias, weight = inner_decoder_out_proj_0_weight, x = input_173)[name = tensor("linear_42")]; + tensor input_175 = add(x = x_29, y = ret_out_9)[name = tensor("input_175")]; tensor xt_1_axes_0 = const()[name = tensor("xt_1_axes_0"), val = tensor([-1])]; - tensor xt_1 = layer_norm(axes = xt_1_axes_0, beta = decoder_norm11_0_bias, epsilon = var_953, gamma = decoder_norm11_0_weight, x = input_173)[name = tensor("xt_1")]; - tensor var_1122 = const()[name = tensor("op_1122"), val = tensor([1, 9, 2, 256])]; - tensor var_1123 = reshape(shape = var_1122, x = xt_1)[name = tensor("op_1123")]; - tensor var_1124_perm_0 = const()[name = tensor("op_1124_perm_0"), val = tensor([0, 2, 1, 3])]; - tensor var_1127 = const()[name = tensor("op_1127"), val = tensor([2, 9, 256])]; - tensor var_1124 = transpose(perm = var_1124_perm_0, x = var_1123)[name = tensor("transpose_23")]; - tensor query_1 = reshape(shape = var_1127, x = var_1124)[name = tensor("query_1")]; + tensor xt_1 = layer_norm(axes = xt_1_axes_0, beta = inner_decoder_norm11_0_bias, epsilon = var_46, gamma = inner_decoder_norm11_0_weight, x = input_175)[name = tensor("xt_1")]; + tensor var_1135 = const()[name = tensor("op_1135"), val = tensor([1, 9, 2, 256])]; + tensor var_1136 = reshape(shape = var_1135, x = xt_1)[name = tensor("op_1136")]; + tensor var_1137_perm_0 = const()[name = tensor("op_1137_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_1140 = const()[name = tensor("op_1140"), val = tensor([2, 9, 256])]; + tensor var_1137 = transpose(perm = var_1137_perm_0, x = var_1136)[name = tensor("transpose_23")]; + tensor query_1 = reshape(shape = var_1140, x = var_1137)[name = tensor("query_1")]; tensor query_3_perm_0 = const()[name = tensor("query_3_perm_0"), val = tensor([1, 0, 2])]; tensor query_3 = transpose(perm = query_3_perm_0, x = query_1)[name = tensor("transpose_22")]; - tensor var_1150 = linear(bias = decoder_self_attn2_0_in_proj_bias, weight = decoder_self_attn2_0_in_proj_weight, x = query_3)[name = tensor("linear_43")]; + tensor var_1163 = linear(bias = inner_decoder_self_attn2_0_in_proj_bias, weight = inner_decoder_self_attn2_0_in_proj_weight, x = query_3)[name = tensor("linear_43")]; tensor concat_1 = const()[name = tensor("concat_1"), val = tensor([9, 2, 3, 256])]; - tensor var_1152 = reshape(shape = concat_1, x = var_1150)[name = tensor("op_1152")]; - tensor var_1153_axes_0 = const()[name = tensor("op_1153_axes_0"), val = tensor([0])]; - tensor var_1153 = expand_dims(axes = var_1153_axes_0, x = var_1152)[name = tensor("op_1153")]; - tensor var_1154_perm_0 = const()[name = tensor("op_1154_perm_0"), val = tensor([-2, 1, 2, 0, 4])]; - tensor var_1155_axes_0 = const()[name = tensor("op_1155_axes_0"), val = tensor([-2])]; - tensor var_1154 = transpose(perm = var_1154_perm_0, x = var_1153)[name = tensor("transpose_21")]; - tensor var_1155 = squeeze(axes = var_1155_axes_0, x = var_1154)[name = tensor("op_1155")]; + tensor var_1165 = reshape(shape = concat_1, x = var_1163)[name = tensor("op_1165")]; + tensor var_1166_axes_0 = const()[name = tensor("op_1166_axes_0"), val = tensor([0])]; + tensor var_1166 = expand_dims(axes = var_1166_axes_0, x = var_1165)[name = tensor("op_1166")]; + tensor var_1167_perm_0 = const()[name = tensor("op_1167_perm_0"), val = tensor([-2, 1, 2, 0, 4])]; + tensor var_1168_axes_0 = const()[name = tensor("op_1168_axes_0"), val = tensor([-2])]; + tensor var_1167 = transpose(perm = var_1167_perm_0, x = var_1166)[name = tensor("transpose_21")]; + tensor var_1168 = squeeze(axes = var_1168_axes_0, x = var_1167)[name = tensor("op_1168")]; tensor q_11_begin_0 = const()[name = tensor("q_11_begin_0"), val = tensor([0, 0, 0, 0])]; tensor q_11_end_0 = const()[name = tensor("q_11_end_0"), val = tensor([1, 9, 2, 256])]; tensor q_11_end_mask_0 = const()[name = tensor("q_11_end_mask_0"), val = tensor([false, true, true, true])]; tensor q_11_squeeze_mask_0 = const()[name = tensor("q_11_squeeze_mask_0"), val = tensor([true, false, false, false])]; - tensor q_11 = slice_by_index(begin = q_11_begin_0, end = q_11_end_0, end_mask = q_11_end_mask_0, squeeze_mask = q_11_squeeze_mask_0, x = var_1155)[name = tensor("q_11")]; + tensor q_11 = slice_by_index(begin = q_11_begin_0, end = q_11_end_0, end_mask = q_11_end_mask_0, squeeze_mask = q_11_squeeze_mask_0, x = var_1168)[name = tensor("q_11")]; tensor k_11_begin_0 = const()[name = tensor("k_11_begin_0"), val = tensor([1, 0, 0, 0])]; tensor k_11_end_0 = const()[name = tensor("k_11_end_0"), val = tensor([2, 9, 2, 256])]; tensor k_11_end_mask_0 = const()[name = tensor("k_11_end_mask_0"), val = tensor([false, true, true, true])]; tensor k_11_squeeze_mask_0 = const()[name = tensor("k_11_squeeze_mask_0"), val = tensor([true, false, false, false])]; - tensor k_11 = slice_by_index(begin = k_11_begin_0, end = k_11_end_0, end_mask = k_11_end_mask_0, squeeze_mask = k_11_squeeze_mask_0, x = var_1155)[name = tensor("k_11")]; + tensor k_11 = slice_by_index(begin = k_11_begin_0, end = k_11_end_0, end_mask = k_11_end_mask_0, squeeze_mask = k_11_squeeze_mask_0, x = var_1168)[name = tensor("k_11")]; tensor v_11_begin_0 = const()[name = tensor("v_11_begin_0"), val = tensor([2, 0, 0, 0])]; tensor v_11_end_0 = const()[name = tensor("v_11_end_0"), val = tensor([3, 9, 2, 256])]; tensor v_11_end_mask_0 = const()[name = tensor("v_11_end_mask_0"), val = tensor([false, true, true, true])]; tensor v_11_squeeze_mask_0 = const()[name = tensor("v_11_squeeze_mask_0"), val = tensor([true, false, false, false])]; - tensor v_11 = slice_by_index(begin = v_11_begin_0, end = v_11_end_0, end_mask = v_11_end_mask_0, squeeze_mask = v_11_squeeze_mask_0, x = var_1155)[name = tensor("v_11")]; - tensor var_1163 = const()[name = tensor("op_1163"), val = tensor([9, 8, 64])]; - tensor var_1164 = reshape(shape = var_1163, x = q_11)[name = tensor("op_1164")]; + tensor v_11 = slice_by_index(begin = v_11_begin_0, end = v_11_end_0, end_mask = v_11_end_mask_0, squeeze_mask = v_11_squeeze_mask_0, x = var_1168)[name = tensor("v_11")]; + tensor var_1176 = const()[name = tensor("op_1176"), val = tensor([9, 8, 64])]; + tensor var_1177 = reshape(shape = var_1176, x = q_11)[name = tensor("op_1177")]; tensor q_13_perm_0 = const()[name = tensor("q_13_perm_0"), val = tensor([1, 0, 2])]; - tensor var_1170 = const()[name = tensor("op_1170"), val = tensor([9, 8, 64])]; - tensor var_1171 = reshape(shape = var_1170, x = k_11)[name = tensor("op_1171")]; + tensor var_1183 = const()[name = tensor("op_1183"), val = tensor([9, 8, 64])]; + tensor var_1184 = reshape(shape = var_1183, x = k_11)[name = tensor("op_1184")]; tensor k_13_perm_0 = const()[name = tensor("k_13_perm_0"), val = tensor([1, 0, 2])]; - tensor var_1177 = const()[name = tensor("op_1177"), val = tensor([9, 8, 64])]; - tensor var_1178 = reshape(shape = var_1177, x = v_11)[name = tensor("op_1178")]; + tensor var_1190 = const()[name = tensor("op_1190"), val = tensor([9, 8, 64])]; + tensor var_1191 = reshape(shape = var_1190, x = v_11)[name = tensor("op_1191")]; tensor v_13_perm_0 = const()[name = tensor("v_13_perm_0"), val = tensor([1, 0, 2])]; - tensor var_1181 = const()[name = tensor("op_1181"), val = tensor([2, 4, 9, 64])]; - tensor q_13 = transpose(perm = q_13_perm_0, x = var_1164)[name = tensor("transpose_20")]; - tensor q_15 = reshape(shape = var_1181, x = q_13)[name = tensor("q_15")]; - tensor var_1183 = const()[name = tensor("op_1183"), val = tensor([2, 4, 9, 64])]; - tensor k_13 = transpose(perm = k_13_perm_0, x = var_1171)[name = tensor("transpose_19")]; - tensor k_15 = reshape(shape = var_1183, x = k_13)[name = tensor("k_15")]; - tensor var_1185 = const()[name = tensor("op_1185"), val = tensor([2, 4, 9, 64])]; - tensor v_13 = transpose(perm = v_13_perm_0, x = var_1178)[name = tensor("transpose_18")]; - tensor v_15 = reshape(shape = var_1185, x = v_13)[name = tensor("v_15")]; + tensor var_1194 = const()[name = tensor("op_1194"), val = tensor([2, 4, 9, 64])]; + tensor q_13 = transpose(perm = q_13_perm_0, x = var_1177)[name = tensor("transpose_20")]; + tensor q_15 = reshape(shape = var_1194, x = q_13)[name = tensor("q_15")]; + tensor var_1196 = const()[name = tensor("op_1196"), val = tensor([2, 4, 9, 64])]; + tensor k_13 = transpose(perm = k_13_perm_0, x = var_1184)[name = tensor("transpose_19")]; + tensor k_15 = reshape(shape = var_1196, x = k_13)[name = tensor("k_15")]; + tensor var_1198 = const()[name = tensor("op_1198"), val = tensor([2, 4, 9, 64])]; + tensor v_13 = transpose(perm = v_13_perm_0, x = var_1191)[name = tensor("transpose_18")]; + tensor v_15 = reshape(shape = var_1198, x = v_13)[name = tensor("v_15")]; tensor mul_1_y_0 = const()[name = tensor("mul_1_y_0"), val = tensor(0x1p-3)]; tensor mul_1 = mul(x = q_15, y = mul_1_y_0)[name = tensor("mul_1")]; tensor matmul_0_transpose_y_0 = const()[name = tensor("matmul_0_transpose_y_0"), val = tensor(true)]; @@ -1039,30 +1046,30 @@ program(1.0) tensor attn_output_1_transpose_x_0 = const()[name = tensor("attn_output_1_transpose_x_0"), val = tensor(false)]; tensor attn_output_1_transpose_y_0 = const()[name = tensor("attn_output_1_transpose_y_0"), val = tensor(false)]; tensor attn_output_1 = matmul(transpose_x = attn_output_1_transpose_x_0, transpose_y = attn_output_1_transpose_y_0, x = softmax_0, y = v_15)[name = tensor("attn_output_1")]; - tensor var_1188 = const()[name = tensor("op_1188"), val = tensor([2, 0, 1, 3])]; - tensor var_1193 = const()[name = tensor("op_1193"), val = tensor([18, 256])]; - tensor var_1189 = transpose(perm = var_1188, x = attn_output_1)[name = tensor("transpose_17")]; - tensor attn_output_3 = reshape(shape = var_1193, x = var_1189)[name = tensor("attn_output_3")]; - tensor attn_output_5 = linear(bias = decoder_self_attn2_0_out_proj_bias, weight = decoder_self_attn2_0_out_proj_weight, x = attn_output_3)[name = tensor("linear_44")]; - tensor var_1197 = const()[name = tensor("op_1197"), val = tensor([9, 2, 256])]; - tensor attn_output_7 = reshape(shape = var_1197, x = attn_output_5)[name = tensor("attn_output_7")]; + tensor var_1201 = const()[name = tensor("op_1201"), val = tensor([2, 0, 1, 3])]; + tensor var_1206 = const()[name = tensor("op_1206"), val = tensor([18, 256])]; + tensor var_1202 = transpose(perm = var_1201, x = attn_output_1)[name = tensor("transpose_17")]; + tensor attn_output_3 = reshape(shape = var_1206, x = var_1202)[name = tensor("attn_output_3")]; + tensor attn_output_5 = linear(bias = inner_decoder_self_attn2_0_out_proj_bias, weight = inner_decoder_self_attn2_0_out_proj_weight, x = attn_output_3)[name = tensor("linear_44")]; + tensor var_1210 = const()[name = tensor("op_1210"), val = tensor([9, 2, 256])]; + tensor attn_output_7 = reshape(shape = var_1210, x = attn_output_5)[name = tensor("attn_output_7")]; tensor sa2_out_1_perm_0 = const()[name = tensor("sa2_out_1_perm_0"), val = tensor([1, 0, 2])]; tensor sa2_out_1 = transpose(perm = sa2_out_1_perm_0, x = attn_output_7)[name = tensor("transpose_16")]; - tensor input_175 = add(x = query_1, y = sa2_out_1)[name = tensor("input_175")]; - tensor input_177_axes_0 = const()[name = tensor("input_177_axes_0"), val = tensor([-1])]; - tensor input_177 = layer_norm(axes = input_177_axes_0, beta = decoder_norm21_0_bias, epsilon = var_953, gamma = decoder_norm21_0_weight, x = input_175)[name = tensor("input_177")]; - tensor input_179 = linear(bias = decoder_linear1_0_bias, weight = decoder_linear1_0_weight, x = input_177)[name = tensor("linear_45")]; - tensor input_181 = relu(x = input_179)[name = tensor("input_181")]; - tensor ffn_out_1 = linear(bias = decoder_linear2_0_bias, weight = decoder_linear2_0_weight, x = input_181)[name = tensor("linear_46")]; - tensor input_183 = add(x = input_177, y = ffn_out_1)[name = tensor("input_183")]; + tensor input_177 = add(x = query_1, y = sa2_out_1)[name = tensor("input_177")]; + tensor input_179_axes_0 = const()[name = tensor("input_179_axes_0"), val = tensor([-1])]; + tensor input_179 = layer_norm(axes = input_179_axes_0, beta = inner_decoder_norm21_0_bias, epsilon = var_46, gamma = inner_decoder_norm21_0_weight, x = input_177)[name = tensor("input_179")]; + tensor input_181 = linear(bias = inner_decoder_linear1_0_bias, weight = inner_decoder_linear1_0_weight, x = input_179)[name = tensor("linear_45")]; + tensor input_183 = relu(x = input_181)[name = tensor("input_183")]; + tensor ffn_out_1 = linear(bias = inner_decoder_linear2_0_bias, weight = inner_decoder_linear2_0_weight, x = input_183)[name = tensor("linear_46")]; + tensor input_185 = add(x = input_179, y = ffn_out_1)[name = tensor("input_185")]; tensor xt_3_axes_0 = const()[name = tensor("xt_3_axes_0"), val = tensor([-1])]; - tensor xt_3 = layer_norm(axes = xt_3_axes_0, beta = decoder_norm22_0_bias, epsilon = var_953, gamma = decoder_norm22_0_weight, x = input_183)[name = tensor("xt_3")]; - tensor var_1217 = const()[name = tensor("op_1217"), val = tensor([1, 2, 9, 256])]; - tensor x_31 = reshape(shape = var_1217, x = xt_3)[name = tensor("x_31")]; - tensor var_1219_perm_0 = const()[name = tensor("op_1219_perm_0"), val = tensor([0, 2, 1, 3])]; - tensor var_1223 = const()[name = tensor("op_1223"), val = tensor([9, 2, 256])]; - tensor var_1219 = transpose(perm = var_1219_perm_0, x = x_31)[name = tensor("transpose_15")]; - tensor x = reshape(shape = var_1223, x = var_1219)[name = tensor("x")]; + tensor xt_3 = layer_norm(axes = xt_3_axes_0, beta = inner_decoder_norm22_0_bias, epsilon = var_46, gamma = inner_decoder_norm22_0_weight, x = input_185)[name = tensor("xt_3")]; + tensor var_1230 = const()[name = tensor("op_1230"), val = tensor([1, 2, 9, 256])]; + tensor x_31 = reshape(shape = var_1230, x = xt_3)[name = tensor("x_31")]; + tensor var_1232_perm_0 = const()[name = tensor("op_1232_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_1236 = const()[name = tensor("op_1236"), val = tensor([9, 2, 256])]; + tensor var_1232 = transpose(perm = var_1232_perm_0, x = x_31)[name = tensor("transpose_15")]; + tensor x = reshape(shape = var_1236, x = var_1232)[name = tensor("x")]; tensor prev_kv_begin_0 = const()[name = tensor("prev_kv_begin_0"), val = tensor([1, 0, 0, 0, 0])]; tensor prev_kv_end_0 = const()[name = tensor("prev_kv_end_0"), val = tensor([2, 9, 4, 64, 64])]; tensor prev_kv_end_mask_0 = const()[name = tensor("prev_kv_end_mask_0"), val = tensor([false, true, true, true, true])]; @@ -1073,120 +1080,120 @@ program(1.0) tensor prev_scale_end_mask_0 = const()[name = tensor("prev_scale_end_mask_0"), val = tensor([false, true])]; tensor prev_scale_squeeze_mask_0 = const()[name = tensor("prev_scale_squeeze_mask_0"), val = tensor([true, false])]; tensor prev_scale = slice_by_index(begin = prev_scale_begin_0, end = prev_scale_end_0, end_mask = prev_scale_end_mask_0, squeeze_mask = prev_scale_squeeze_mask_0, x = dec_scale)[name = tensor("prev_scale")]; - tensor var_1231 = linear(bias = decoder_q_proj_1_bias, weight = decoder_q_proj_1_weight, x = x)[name = tensor("linear_47")]; - tensor var_1232 = const()[name = tensor("op_1232"), val = tensor([9, 2, 4, 64])]; - tensor var_1233 = reshape(shape = var_1232, x = var_1231)[name = tensor("op_1233")]; + tensor var_1244 = linear(bias = inner_decoder_q_proj_1_bias, weight = inner_decoder_q_proj_1_weight, x = x)[name = tensor("linear_47")]; + tensor var_1245 = const()[name = tensor("op_1245"), val = tensor([9, 2, 4, 64])]; + tensor var_1246 = reshape(shape = var_1245, x = var_1244)[name = tensor("op_1246")]; tensor q_17_perm_0 = const()[name = tensor("q_17_perm_0"), val = tensor([0, 2, 1, 3])]; - tensor var_1237 = linear(bias = decoder_k_proj_1_bias, weight = decoder_k_proj_1_weight, x = x)[name = tensor("linear_48")]; - tensor var_1238 = const()[name = tensor("op_1238"), val = tensor(0x1p-3)]; - tensor var_1239 = mul(x = var_1237, y = var_1238)[name = tensor("op_1239")]; - tensor var_1240 = const()[name = tensor("op_1240"), val = tensor([9, 2, 4, 64])]; - tensor var_1241 = reshape(shape = var_1240, x = var_1239)[name = tensor("op_1241")]; + tensor var_1250 = linear(bias = inner_decoder_k_proj_1_bias, weight = inner_decoder_k_proj_1_weight, x = x)[name = tensor("linear_48")]; + tensor var_1251 = const()[name = tensor("op_1251"), val = tensor(0x1p-3)]; + tensor var_1252 = mul(x = var_1250, y = var_1251)[name = tensor("op_1252")]; + tensor var_1253 = const()[name = tensor("op_1253"), val = tensor([9, 2, 4, 64])]; + tensor var_1254 = reshape(shape = var_1253, x = var_1252)[name = tensor("op_1254")]; tensor k_17_perm_0 = const()[name = tensor("k_17_perm_0"), val = tensor([0, 2, 1, 3])]; - tensor var_1245 = linear(bias = decoder_v_proj_1_bias, weight = decoder_v_proj_1_weight, x = x)[name = tensor("linear_49")]; - tensor var_1246 = const()[name = tensor("op_1246"), val = tensor([9, 2, 4, 64])]; - tensor var_1247 = reshape(shape = var_1246, x = var_1245)[name = tensor("op_1247")]; + tensor var_1258 = linear(bias = inner_decoder_v_proj_1_bias, weight = inner_decoder_v_proj_1_weight, x = x)[name = tensor("linear_49")]; + tensor var_1259 = const()[name = tensor("op_1259"), val = tensor([9, 2, 4, 64])]; + tensor var_1260 = reshape(shape = var_1259, x = var_1258)[name = tensor("op_1260")]; tensor v_17_perm_0 = const()[name = tensor("v_17_perm_0"), val = tensor([0, 2, 1, 3])]; - tensor input_187 = linear(bias = decoder_g_proj_1_bias, weight = decoder_g_proj_1_weight, x = x)[name = tensor("linear_50")]; + tensor input_189 = linear(bias = inner_decoder_g_proj_1_bias, weight = inner_decoder_g_proj_1_weight, x = x)[name = tensor("linear_50")]; tensor sqrt_s0 = sqrt(x = prev_scale)[name = tensor("sqrt_s0")]; tensor s_t = add(x = prev_scale, y = cumsum_mask_1)[name = tensor("s_t")]; tensor const_32 = const()[name = tensor("const_32"), val = tensor(0x1.fffffep+127)]; - tensor clip_3 = clip(alpha = var_958, beta = const_32, x = s_t)[name = tensor("clip_3")]; + tensor clip_3 = clip(alpha = var_39, beta = const_32, x = s_t)[name = tensor("clip_3")]; tensor sqrt_s_t = sqrt(x = clip_3)[name = tensor("sqrt_s_t")]; tensor qk_transpose_x_1 = const()[name = tensor("qk_transpose_x_1"), val = tensor(false)]; tensor qk_transpose_y_1 = const()[name = tensor("qk_transpose_y_1"), val = tensor(true)]; - tensor k_17 = transpose(perm = k_17_perm_0, x = var_1241)[name = tensor("transpose_13")]; - tensor q_17 = transpose(perm = q_17_perm_0, x = var_1233)[name = tensor("transpose_14")]; + tensor k_17 = transpose(perm = k_17_perm_0, x = var_1254)[name = tensor("transpose_13")]; + tensor q_17 = transpose(perm = q_17_perm_0, x = var_1246)[name = tensor("transpose_14")]; tensor qk = matmul(transpose_x = qk_transpose_x_1, transpose_y = qk_transpose_y_1, x = q_17, y = k_17)[name = tensor("qk")]; - tensor var_1262 = const()[name = tensor("op_1262"), val = tensor([2, 1])]; - tensor var_1263 = reshape(shape = var_1262, x = sqrt_s_t)[name = tensor("op_1263")]; - tensor M = real_div(x = causal_with_valid_1, y = var_1263)[name = tensor("M")]; - tensor var_1265 = mul(x = qk, y = M)[name = tensor("op_1265")]; - tensor inner_transpose_x_0 = const()[name = tensor("inner_transpose_x_0"), val = tensor(false)]; - tensor inner_transpose_y_0 = const()[name = tensor("inner_transpose_y_0"), val = tensor(false)]; - tensor v_17 = transpose(perm = v_17_perm_0, x = var_1247)[name = tensor("transpose_12")]; - tensor inner = matmul(transpose_x = inner_transpose_x_0, transpose_y = inner_transpose_y_0, x = var_1265, y = v_17)[name = tensor("inner")]; - tensor var_1267_transpose_x_0 = const()[name = tensor("op_1267_transpose_x_0"), val = tensor(false)]; - tensor var_1267_transpose_y_0 = const()[name = tensor("op_1267_transpose_y_0"), val = tensor(false)]; - tensor var_1267 = matmul(transpose_x = var_1267_transpose_x_0, transpose_y = var_1267_transpose_y_0, x = q_17, y = prev_kv)[name = tensor("op_1267")]; - tensor var_1268 = real_div(x = sqrt_s0, y = sqrt_s_t)[name = tensor("op_1268")]; - tensor var_1269 = const()[name = tensor("op_1269"), val = tensor([1, 1, 2, 1])]; - tensor var_1270 = reshape(shape = var_1269, x = var_1268)[name = tensor("op_1270")]; - tensor cross = mul(x = var_1267, y = var_1270)[name = tensor("cross")]; - tensor out_31 = add(x = inner, y = cross)[name = tensor("out_31")]; - tensor v_masked = mul(x = v_17, y = var_1093)[name = tensor("v_masked")]; - tensor var_1276 = mul(x = prev_kv, y = sqrt_s0)[name = tensor("op_1276")]; - tensor var_1278_transpose_x_1 = const()[name = tensor("op_1278_transpose_x_1"), val = tensor(true)]; - tensor var_1278_transpose_y_1 = const()[name = tensor("op_1278_transpose_y_1"), val = tensor(false)]; - tensor var_1278 = matmul(transpose_x = var_1278_transpose_x_1, transpose_y = var_1278_transpose_y_1, x = k_17, y = v_masked)[name = tensor("op_1278")]; - tensor new_kv_unnorm = add(x = var_1276, y = var_1278)[name = tensor("new_kv_unnorm")]; - tensor new_scale = add(x = prev_scale, y = var_1101)[name = tensor("new_scale")]; + tensor var_1275 = const()[name = tensor("op_1275"), val = tensor([2, 1])]; + tensor var_1276 = reshape(shape = var_1275, x = sqrt_s_t)[name = tensor("op_1276")]; + tensor M = real_div(x = causal_with_valid_1, y = var_1276)[name = tensor("M")]; + tensor var_1278 = mul(x = qk, y = M)[name = tensor("op_1278")]; + tensor inner_11_transpose_x_0 = const()[name = tensor("inner_11_transpose_x_0"), val = tensor(false)]; + tensor inner_11_transpose_y_0 = const()[name = tensor("inner_11_transpose_y_0"), val = tensor(false)]; + tensor v_17 = transpose(perm = v_17_perm_0, x = var_1260)[name = tensor("transpose_12")]; + tensor inner_11 = matmul(transpose_x = inner_11_transpose_x_0, transpose_y = inner_11_transpose_y_0, x = var_1278, y = v_17)[name = tensor("inner_11")]; + tensor var_1280_transpose_x_0 = const()[name = tensor("op_1280_transpose_x_0"), val = tensor(false)]; + tensor var_1280_transpose_y_0 = const()[name = tensor("op_1280_transpose_y_0"), val = tensor(false)]; + tensor var_1280 = matmul(transpose_x = var_1280_transpose_x_0, transpose_y = var_1280_transpose_y_0, x = q_17, y = prev_kv)[name = tensor("op_1280")]; + tensor var_1281 = real_div(x = sqrt_s0, y = sqrt_s_t)[name = tensor("op_1281")]; + tensor var_1282 = const()[name = tensor("op_1282"), val = tensor([1, 1, 2, 1])]; + tensor var_1283 = reshape(shape = var_1282, x = var_1281)[name = tensor("op_1283")]; + tensor cross = mul(x = var_1280, y = var_1283)[name = tensor("cross")]; + tensor out_31 = add(x = inner_11, y = cross)[name = tensor("out_31")]; + tensor v_masked = mul(x = v_17, y = var_1106)[name = tensor("v_masked")]; + tensor var_1289 = mul(x = prev_kv, y = sqrt_s0)[name = tensor("op_1289")]; + tensor var_1291_transpose_x_1 = const()[name = tensor("op_1291_transpose_x_1"), val = tensor(true)]; + tensor var_1291_transpose_y_1 = const()[name = tensor("op_1291_transpose_y_1"), val = tensor(false)]; + tensor var_1291 = matmul(transpose_x = var_1291_transpose_x_1, transpose_y = var_1291_transpose_y_1, x = k_17, y = v_masked)[name = tensor("op_1291")]; + tensor new_kv_unnorm = add(x = var_1289, y = var_1291)[name = tensor("new_kv_unnorm")]; + tensor new_scale = add(x = prev_scale, y = var_1114)[name = tensor("new_scale")]; tensor const_33 = const()[name = tensor("const_33"), val = tensor(0x1.fffffep+127)]; - tensor clip_4 = clip(alpha = var_958, beta = const_33, x = new_scale)[name = tensor("clip_4")]; + tensor clip_4 = clip(alpha = var_39, beta = const_33, x = new_scale)[name = tensor("clip_4")]; tensor sqrt_new_scale = sqrt(x = clip_4)[name = tensor("sqrt_new_scale")]; tensor nkv = real_div(x = new_kv_unnorm, y = sqrt_new_scale)[name = tensor("nkv")]; - tensor var_1287_perm_0 = const()[name = tensor("op_1287_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_1300_perm_0 = const()[name = tensor("op_1300_perm_0"), val = tensor([0, 2, 1, 3])]; tensor out_33_axes_0 = const()[name = tensor("out_33_axes_0"), val = tensor([-1])]; - tensor var_1287 = transpose(perm = var_1287_perm_0, x = out_31)[name = tensor("transpose_11")]; - tensor out_33 = layer_norm(axes = out_33_axes_0, epsilon = var_955, x = var_1287)[name = tensor("out_33")]; - tensor var_1291 = const()[name = tensor("op_1291"), val = tensor([9, 2, 256])]; - tensor out = reshape(shape = var_1291, x = out_33)[name = tensor("out")]; - tensor var_1293 = silu(x = input_187)[name = tensor("op_1293")]; - tensor input_189 = mul(x = var_1293, y = out)[name = tensor("input_189")]; - tensor ret_out = linear(bias = decoder_out_proj_1_bias, weight = decoder_out_proj_1_weight, x = input_189)[name = tensor("linear_51")]; - tensor input_191 = add(x = x, y = ret_out)[name = tensor("input_191")]; + tensor var_1300 = transpose(perm = var_1300_perm_0, x = out_31)[name = tensor("transpose_11")]; + tensor out_33 = layer_norm(axes = out_33_axes_0, epsilon = var_54, x = var_1300)[name = tensor("out_33")]; + tensor var_1304 = const()[name = tensor("op_1304"), val = tensor([9, 2, 256])]; + tensor out = reshape(shape = var_1304, x = out_33)[name = tensor("out")]; + tensor var_1306 = silu(x = input_189)[name = tensor("op_1306")]; + tensor input_191 = mul(x = var_1306, y = out)[name = tensor("input_191")]; + tensor ret_out = linear(bias = inner_decoder_out_proj_1_bias, weight = inner_decoder_out_proj_1_weight, x = input_191)[name = tensor("linear_51")]; + tensor input_193 = add(x = x, y = ret_out)[name = tensor("input_193")]; tensor xt_5_axes_0 = const()[name = tensor("xt_5_axes_0"), val = tensor([-1])]; - tensor xt_5 = layer_norm(axes = xt_5_axes_0, beta = decoder_norm11_1_bias, epsilon = var_953, gamma = decoder_norm11_1_weight, x = input_191)[name = tensor("xt_5")]; - tensor var_1303 = const()[name = tensor("op_1303"), val = tensor([1, 9, 2, 256])]; - tensor var_1304 = reshape(shape = var_1303, x = xt_5)[name = tensor("op_1304")]; - tensor var_1305_perm_0 = const()[name = tensor("op_1305_perm_0"), val = tensor([0, 2, 1, 3])]; - tensor var_1308 = const()[name = tensor("op_1308"), val = tensor([2, 9, 256])]; - tensor var_1305 = transpose(perm = var_1305_perm_0, x = var_1304)[name = tensor("transpose_10")]; - tensor query_5 = reshape(shape = var_1308, x = var_1305)[name = tensor("query_5")]; + tensor xt_5 = layer_norm(axes = xt_5_axes_0, beta = inner_decoder_norm11_1_bias, epsilon = var_46, gamma = inner_decoder_norm11_1_weight, x = input_193)[name = tensor("xt_5")]; + tensor var_1316 = const()[name = tensor("op_1316"), val = tensor([1, 9, 2, 256])]; + tensor var_1317 = reshape(shape = var_1316, x = xt_5)[name = tensor("op_1317")]; + tensor var_1318_perm_0 = const()[name = tensor("op_1318_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_1321 = const()[name = tensor("op_1321"), val = tensor([2, 9, 256])]; + tensor var_1318 = transpose(perm = var_1318_perm_0, x = var_1317)[name = tensor("transpose_10")]; + tensor query_5 = reshape(shape = var_1321, x = var_1318)[name = tensor("query_5")]; tensor query_perm_0 = const()[name = tensor("query_perm_0"), val = tensor([1, 0, 2])]; tensor query = transpose(perm = query_perm_0, x = query_5)[name = tensor("transpose_9")]; - tensor var_1331 = linear(bias = decoder_self_attn2_1_in_proj_bias, weight = decoder_self_attn2_1_in_proj_weight, x = query)[name = tensor("linear_52")]; + tensor var_1344 = linear(bias = inner_decoder_self_attn2_1_in_proj_bias, weight = inner_decoder_self_attn2_1_in_proj_weight, x = query)[name = tensor("linear_52")]; tensor concat_2 = const()[name = tensor("concat_2"), val = tensor([9, 2, 3, 256])]; - tensor var_1333 = reshape(shape = concat_2, x = var_1331)[name = tensor("op_1333")]; - tensor var_1334_axes_0 = const()[name = tensor("op_1334_axes_0"), val = tensor([0])]; - tensor var_1334 = expand_dims(axes = var_1334_axes_0, x = var_1333)[name = tensor("op_1334")]; - tensor var_1335_perm_0 = const()[name = tensor("op_1335_perm_0"), val = tensor([-2, 1, 2, 0, 4])]; - tensor var_1336_axes_0 = const()[name = tensor("op_1336_axes_0"), val = tensor([-2])]; - tensor var_1335 = transpose(perm = var_1335_perm_0, x = var_1334)[name = tensor("transpose_8")]; - tensor var_1336 = squeeze(axes = var_1336_axes_0, x = var_1335)[name = tensor("op_1336")]; + tensor var_1346 = reshape(shape = concat_2, x = var_1344)[name = tensor("op_1346")]; + tensor var_1347_axes_0 = const()[name = tensor("op_1347_axes_0"), val = tensor([0])]; + tensor var_1347 = expand_dims(axes = var_1347_axes_0, x = var_1346)[name = tensor("op_1347")]; + tensor var_1348_perm_0 = const()[name = tensor("op_1348_perm_0"), val = tensor([-2, 1, 2, 0, 4])]; + tensor var_1349_axes_0 = const()[name = tensor("op_1349_axes_0"), val = tensor([-2])]; + tensor var_1348 = transpose(perm = var_1348_perm_0, x = var_1347)[name = tensor("transpose_8")]; + tensor var_1349 = squeeze(axes = var_1349_axes_0, x = var_1348)[name = tensor("op_1349")]; tensor q_19_begin_0 = const()[name = tensor("q_19_begin_0"), val = tensor([0, 0, 0, 0])]; tensor q_19_end_0 = const()[name = tensor("q_19_end_0"), val = tensor([1, 9, 2, 256])]; tensor q_19_end_mask_0 = const()[name = tensor("q_19_end_mask_0"), val = tensor([false, true, true, true])]; tensor q_19_squeeze_mask_0 = const()[name = tensor("q_19_squeeze_mask_0"), val = tensor([true, false, false, false])]; - tensor q_19 = slice_by_index(begin = q_19_begin_0, end = q_19_end_0, end_mask = q_19_end_mask_0, squeeze_mask = q_19_squeeze_mask_0, x = var_1336)[name = tensor("q_19")]; + tensor q_19 = slice_by_index(begin = q_19_begin_0, end = q_19_end_0, end_mask = q_19_end_mask_0, squeeze_mask = q_19_squeeze_mask_0, x = var_1349)[name = tensor("q_19")]; tensor k_19_begin_0 = const()[name = tensor("k_19_begin_0"), val = tensor([1, 0, 0, 0])]; tensor k_19_end_0 = const()[name = tensor("k_19_end_0"), val = tensor([2, 9, 2, 256])]; tensor k_19_end_mask_0 = const()[name = tensor("k_19_end_mask_0"), val = tensor([false, true, true, true])]; tensor k_19_squeeze_mask_0 = const()[name = tensor("k_19_squeeze_mask_0"), val = tensor([true, false, false, false])]; - tensor k_19 = slice_by_index(begin = k_19_begin_0, end = k_19_end_0, end_mask = k_19_end_mask_0, squeeze_mask = k_19_squeeze_mask_0, x = var_1336)[name = tensor("k_19")]; + tensor k_19 = slice_by_index(begin = k_19_begin_0, end = k_19_end_0, end_mask = k_19_end_mask_0, squeeze_mask = k_19_squeeze_mask_0, x = var_1349)[name = tensor("k_19")]; tensor v_19_begin_0 = const()[name = tensor("v_19_begin_0"), val = tensor([2, 0, 0, 0])]; tensor v_19_end_0 = const()[name = tensor("v_19_end_0"), val = tensor([3, 9, 2, 256])]; tensor v_19_end_mask_0 = const()[name = tensor("v_19_end_mask_0"), val = tensor([false, true, true, true])]; tensor v_19_squeeze_mask_0 = const()[name = tensor("v_19_squeeze_mask_0"), val = tensor([true, false, false, false])]; - tensor v_19 = slice_by_index(begin = v_19_begin_0, end = v_19_end_0, end_mask = v_19_end_mask_0, squeeze_mask = v_19_squeeze_mask_0, x = var_1336)[name = tensor("v_19")]; - tensor var_1344 = const()[name = tensor("op_1344"), val = tensor([9, 8, 64])]; - tensor var_1345 = reshape(shape = var_1344, x = q_19)[name = tensor("op_1345")]; + tensor v_19 = slice_by_index(begin = v_19_begin_0, end = v_19_end_0, end_mask = v_19_end_mask_0, squeeze_mask = v_19_squeeze_mask_0, x = var_1349)[name = tensor("v_19")]; + tensor var_1357 = const()[name = tensor("op_1357"), val = tensor([9, 8, 64])]; + tensor var_1358 = reshape(shape = var_1357, x = q_19)[name = tensor("op_1358")]; tensor q_21_perm_0 = const()[name = tensor("q_21_perm_0"), val = tensor([1, 0, 2])]; - tensor var_1351 = const()[name = tensor("op_1351"), val = tensor([9, 8, 64])]; - tensor var_1352 = reshape(shape = var_1351, x = k_19)[name = tensor("op_1352")]; + tensor var_1364 = const()[name = tensor("op_1364"), val = tensor([9, 8, 64])]; + tensor var_1365 = reshape(shape = var_1364, x = k_19)[name = tensor("op_1365")]; tensor k_21_perm_0 = const()[name = tensor("k_21_perm_0"), val = tensor([1, 0, 2])]; - tensor var_1358 = const()[name = tensor("op_1358"), val = tensor([9, 8, 64])]; - tensor var_1359 = reshape(shape = var_1358, x = v_19)[name = tensor("op_1359")]; + tensor var_1371 = const()[name = tensor("op_1371"), val = tensor([9, 8, 64])]; + tensor var_1372 = reshape(shape = var_1371, x = v_19)[name = tensor("op_1372")]; tensor v_21_perm_0 = const()[name = tensor("v_21_perm_0"), val = tensor([1, 0, 2])]; - tensor var_1362 = const()[name = tensor("op_1362"), val = tensor([2, 4, 9, 64])]; - tensor q_21 = transpose(perm = q_21_perm_0, x = var_1345)[name = tensor("transpose_7")]; - tensor q = reshape(shape = var_1362, x = q_21)[name = tensor("q")]; - tensor var_1364 = const()[name = tensor("op_1364"), val = tensor([2, 4, 9, 64])]; - tensor k_21 = transpose(perm = k_21_perm_0, x = var_1352)[name = tensor("transpose_6")]; - tensor k = reshape(shape = var_1364, x = k_21)[name = tensor("k")]; - tensor var_1366 = const()[name = tensor("op_1366"), val = tensor([2, 4, 9, 64])]; - tensor v_21 = transpose(perm = v_21_perm_0, x = var_1359)[name = tensor("transpose_5")]; - tensor v = reshape(shape = var_1366, x = v_21)[name = tensor("v")]; + tensor var_1375 = const()[name = tensor("op_1375"), val = tensor([2, 4, 9, 64])]; + tensor q_21 = transpose(perm = q_21_perm_0, x = var_1358)[name = tensor("transpose_7")]; + tensor q = reshape(shape = var_1375, x = q_21)[name = tensor("q")]; + tensor var_1377 = const()[name = tensor("op_1377"), val = tensor([2, 4, 9, 64])]; + tensor k_21 = transpose(perm = k_21_perm_0, x = var_1365)[name = tensor("transpose_6")]; + tensor k = reshape(shape = var_1377, x = k_21)[name = tensor("k")]; + tensor var_1379 = const()[name = tensor("op_1379"), val = tensor([2, 4, 9, 64])]; + tensor v_21 = transpose(perm = v_21_perm_0, x = var_1372)[name = tensor("transpose_5")]; + tensor v = reshape(shape = var_1379, x = v_21)[name = tensor("v")]; tensor mul_3_y_0 = const()[name = tensor("mul_3_y_0"), val = tensor(0x1p-3)]; tensor mul_3 = mul(x = q, y = mul_3_y_0)[name = tensor("mul_3")]; tensor matmul_1_transpose_y_0 = const()[name = tensor("matmul_1_transpose_y_0"), val = tensor(true)]; @@ -1197,36 +1204,36 @@ program(1.0) tensor attn_output_9_transpose_x_0 = const()[name = tensor("attn_output_9_transpose_x_0"), val = tensor(false)]; tensor attn_output_9_transpose_y_0 = const()[name = tensor("attn_output_9_transpose_y_0"), val = tensor(false)]; tensor attn_output_9 = matmul(transpose_x = attn_output_9_transpose_x_0, transpose_y = attn_output_9_transpose_y_0, x = softmax_1, y = v)[name = tensor("attn_output_9")]; - tensor var_1369 = const()[name = tensor("op_1369"), val = tensor([2, 0, 1, 3])]; - tensor var_1374 = const()[name = tensor("op_1374"), val = tensor([18, 256])]; - tensor var_1370 = transpose(perm = var_1369, x = attn_output_9)[name = tensor("transpose_4")]; - tensor attn_output_11 = reshape(shape = var_1374, x = var_1370)[name = tensor("attn_output_11")]; - tensor attn_output_13 = linear(bias = decoder_self_attn2_1_out_proj_bias, weight = decoder_self_attn2_1_out_proj_weight, x = attn_output_11)[name = tensor("linear_53")]; - tensor var_1378 = const()[name = tensor("op_1378"), val = tensor([9, 2, 256])]; - tensor attn_output = reshape(shape = var_1378, x = attn_output_13)[name = tensor("attn_output")]; + tensor var_1382 = const()[name = tensor("op_1382"), val = tensor([2, 0, 1, 3])]; + tensor var_1387 = const()[name = tensor("op_1387"), val = tensor([18, 256])]; + tensor var_1383 = transpose(perm = var_1382, x = attn_output_9)[name = tensor("transpose_4")]; + tensor attn_output_11 = reshape(shape = var_1387, x = var_1383)[name = tensor("attn_output_11")]; + tensor attn_output_13 = linear(bias = inner_decoder_self_attn2_1_out_proj_bias, weight = inner_decoder_self_attn2_1_out_proj_weight, x = attn_output_11)[name = tensor("linear_53")]; + tensor var_1391 = const()[name = tensor("op_1391"), val = tensor([9, 2, 256])]; + tensor attn_output = reshape(shape = var_1391, x = attn_output_13)[name = tensor("attn_output")]; tensor sa2_out_perm_0 = const()[name = tensor("sa2_out_perm_0"), val = tensor([1, 0, 2])]; tensor sa2_out = transpose(perm = sa2_out_perm_0, x = attn_output)[name = tensor("transpose_3")]; - tensor input_193 = add(x = query_5, y = sa2_out)[name = tensor("input_193")]; - tensor input_195_axes_0 = const()[name = tensor("input_195_axes_0"), val = tensor([-1])]; - tensor input_195 = layer_norm(axes = input_195_axes_0, beta = decoder_norm21_1_bias, epsilon = var_953, gamma = decoder_norm21_1_weight, x = input_193)[name = tensor("input_195")]; - tensor input_197 = linear(bias = decoder_linear1_1_bias, weight = decoder_linear1_1_weight, x = input_195)[name = tensor("linear_54")]; - tensor input_199 = relu(x = input_197)[name = tensor("input_199")]; - tensor ffn_out = linear(bias = decoder_linear2_1_bias, weight = decoder_linear2_1_weight, x = input_199)[name = tensor("linear_55")]; - tensor input_201 = add(x = input_195, y = ffn_out)[name = tensor("input_201")]; + tensor input_195 = add(x = query_5, y = sa2_out)[name = tensor("input_195")]; + tensor input_197_axes_0 = const()[name = tensor("input_197_axes_0"), val = tensor([-1])]; + tensor input_197 = layer_norm(axes = input_197_axes_0, beta = inner_decoder_norm21_1_bias, epsilon = var_46, gamma = inner_decoder_norm21_1_weight, x = input_195)[name = tensor("input_197")]; + tensor input_199 = linear(bias = inner_decoder_linear1_1_bias, weight = inner_decoder_linear1_1_weight, x = input_197)[name = tensor("linear_54")]; + tensor input_201 = relu(x = input_199)[name = tensor("input_201")]; + tensor ffn_out = linear(bias = inner_decoder_linear2_1_bias, weight = inner_decoder_linear2_1_weight, x = input_201)[name = tensor("linear_55")]; + tensor input_203 = add(x = input_197, y = ffn_out)[name = tensor("input_203")]; tensor xt_axes_0 = const()[name = tensor("xt_axes_0"), val = tensor([-1])]; - tensor xt = layer_norm(axes = xt_axes_0, beta = decoder_norm22_1_bias, epsilon = var_953, gamma = decoder_norm22_1_weight, x = input_201)[name = tensor("xt")]; - tensor var_1398 = const()[name = tensor("op_1398"), val = tensor([1, 2, 9, 256])]; - tensor input = reshape(shape = var_1398, x = xt)[name = tensor("input")]; - tensor var_1400 = const()[name = tensor("op_1400"), val = tensor([-1])]; - tensor var_1401 = reduce_l2_norm(axes = var_1400, keep_dims = var_956, x = input)[name = tensor("op_1401")]; + tensor xt = layer_norm(axes = xt_axes_0, beta = inner_decoder_norm22_1_bias, epsilon = var_46, gamma = inner_decoder_norm22_1_weight, x = input_203)[name = tensor("xt")]; + tensor var_1411 = const()[name = tensor("op_1411"), val = tensor([1, 2, 9, 256])]; + tensor input = reshape(shape = var_1411, x = xt)[name = tensor("input")]; + tensor var_1413 = const()[name = tensor("op_1413"), val = tensor([-1])]; + tensor var_1414 = reduce_l2_norm(axes = var_1413, keep_dims = var_45, x = input)[name = tensor("op_1414")]; tensor const_42 = const()[name = tensor("const_42"), val = tensor(0x1.fffffep+127)]; - tensor clip_5 = clip(alpha = var_948, beta = const_42, x = var_1401)[name = tensor("clip_5")]; - tensor var_1403 = real_div(x = input, y = clip_5)[name = tensor("op_1403")]; + tensor clip_5 = clip(alpha = var_59, beta = const_42, x = var_1414)[name = tensor("clip_5")]; + tensor var_1416 = real_div(x = input, y = clip_5)[name = tensor("op_1416")]; tensor concat_6 = const()[name = tensor("concat_6"), val = tensor([2, 1, 256])]; tensor reshape_0 = reshape(shape = concat_6, x = emb)[name = tensor("reshape_0")]; tensor transpose_1_perm_0 = const()[name = tensor("transpose_1_perm_0"), val = tensor([0, 1, 3, 2])]; tensor concat_7 = const()[name = tensor("concat_7"), val = tensor([2, 256, 9])]; - tensor transpose_1 = transpose(perm = transpose_1_perm_0, x = var_1403)[name = tensor("transpose_2")]; + tensor transpose_1 = transpose(perm = transpose_1_perm_0, x = var_1416)[name = tensor("transpose_2")]; tensor reshape_1 = reshape(shape = concat_7, x = transpose_1)[name = tensor("reshape_1")]; tensor matmul_2_transpose_x_0 = const()[name = tensor("matmul_2_transpose_x_0"), val = tensor(false)]; tensor matmul_2_transpose_y_0 = const()[name = tensor("matmul_2_transpose_y_0"), val = tensor(false)]; @@ -1237,10 +1244,10 @@ program(1.0) tensor output_end_0 = const()[name = tensor("output_end_0"), val = tensor([1, 2, 8])]; tensor output_end_mask_0 = const()[name = tensor("output_end_mask_0"), val = tensor([true, true, false])]; tensor output = slice_by_index(begin = output_begin_0, end = output_end_0, end_mask = output_end_mask_0, x = reshape_2)[name = tensor("output")]; - tensor probs = sigmoid(x = output)[name = tensor("op_1407")]; - tensor var_1409_axis_0 = const()[name = tensor("op_1409_axis_0"), val = tensor(0)]; - tensor dec_kv_new = stack(axis = var_1409_axis_0, values = (var_1105, nkv))[name = tensor("op_1409")]; - tensor var_1411_axis_0 = const()[name = tensor("op_1411_axis_0"), val = tensor(0)]; - tensor dec_scale_new = stack(axis = var_1411_axis_0, values = (new_scale_9, new_scale))[name = tensor("op_1411")]; + tensor probs = sigmoid(x = output)[name = tensor("op_1420")]; + tensor var_1422_axis_0 = const()[name = tensor("op_1422_axis_0"), val = tensor(0)]; + tensor dec_kv_new = stack(axis = var_1422_axis_0, values = (var_1118, nkv))[name = tensor("op_1422")]; + tensor var_1424_axis_0 = const()[name = tensor("op_1424_axis_0"), val = tensor(0)]; + tensor dec_scale_new = stack(axis = var_1424_axis_0, values = (new_scale_9, new_scale))[name = tensor("op_1424")]; } -> (probs, enc_kv_new, enc_scale_new, enc_conv_cache_new, cnn_window_new, dec_kv_new, dec_scale_new); } \ No newline at end of file diff --git a/optimized/ch/200ms/ls_eend_ch_200ms.mlpackage/Data/com.apple.CoreML/model.mlmodel b/optimized/ch/200ms/ls_eend_ch_200ms.mlpackage/Data/com.apple.CoreML/model.mlmodel index bee7293d1930f8df17e82f398224b087ead0eed6..b0b764951a916e183e0396624d5f52625082c014 100644 --- a/optimized/ch/200ms/ls_eend_ch_200ms.mlpackage/Data/com.apple.CoreML/model.mlmodel +++ b/optimized/ch/200ms/ls_eend_ch_200ms.mlpackage/Data/com.apple.CoreML/model.mlmodel @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:4f15d275fba79646bf8e6b1acd94f18b344b4ba451e802249a4814d7601f426f -size 179867 +oid sha256:c2fdc8691b90d30c33d9e6eeba0997e644c82adc5af2f87dd493ba202c0ffc11 +size 184847 diff --git a/optimized/ch/200ms/ls_eend_ch_200ms.mlpackage/Manifest.json b/optimized/ch/200ms/ls_eend_ch_200ms.mlpackage/Manifest.json index cd59efc886fa4d1cfef4032c5411dbb7569655b1..d53e7f18f986fec2a3e83fc62b0c87de4207fbbd 100644 --- a/optimized/ch/200ms/ls_eend_ch_200ms.mlpackage/Manifest.json +++ b/optimized/ch/200ms/ls_eend_ch_200ms.mlpackage/Manifest.json @@ -1,18 +1,18 @@ { "fileFormatVersion": "1.0.0", "itemInfoEntries": { - "5C0B804E-E764-41B0-9747-F2E8A83C47B7": { - "author": "com.apple.CoreML", - "description": "CoreML Model Specification", - "name": "model.mlmodel", - "path": "com.apple.CoreML/model.mlmodel" - }, - "9610BCAB-14A1-4087-9959-ED1CB79286B1": { + "857B2C68-68CE-48B9-8135-33407386BD83": { "author": "com.apple.CoreML", "description": "CoreML Model Weights", "name": "weights", "path": "com.apple.CoreML/weights" + }, + "D82F93A9-BACD-4FC2-824C-63C27C744141": { + "author": "com.apple.CoreML", + "description": "CoreML Model Specification", + "name": "model.mlmodel", + "path": "com.apple.CoreML/model.mlmodel" } }, - "rootModelIdentifier": "5C0B804E-E764-41B0-9747-F2E8A83C47B7" + "rootModelIdentifier": "D82F93A9-BACD-4FC2-824C-63C27C744141" } diff --git a/optimized/ch/300ms/ls_eend_ch_300ms.mlmodelc/analytics/coremldata.bin b/optimized/ch/300ms/ls_eend_ch_300ms.mlmodelc/analytics/coremldata.bin index d8f91c11eb8b5980108c3d2fdd2be965c635d965..2af7b7c943db8138fbdd9decacfd9fbee182def8 100644 --- a/optimized/ch/300ms/ls_eend_ch_300ms.mlmodelc/analytics/coremldata.bin +++ b/optimized/ch/300ms/ls_eend_ch_300ms.mlmodelc/analytics/coremldata.bin @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:ceb496d3598e7070ae0538a32010b7e9d3ffdd68c9c569f14b6987731dc9f1ad +oid sha256:266d247cb95834ecbe57555a6d80a22fe1fd9858d3ac21bbadbd681e462cf2e1 size 243 diff --git a/optimized/ch/300ms/ls_eend_ch_300ms.mlmodelc/coremldata.bin b/optimized/ch/300ms/ls_eend_ch_300ms.mlmodelc/coremldata.bin index 58a2886f988ba53cdc741a3c5780fe599407acd9..c772e73a83c2eb579b7a2e6be9a1f20af803b388 100644 --- a/optimized/ch/300ms/ls_eend_ch_300ms.mlmodelc/coremldata.bin +++ b/optimized/ch/300ms/ls_eend_ch_300ms.mlmodelc/coremldata.bin @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:595472cafa7ff618c858be86076cfe0e894ffc64361b5318c631e631bd0a0669 -size 1301 +oid sha256:667d6e71963d032f6ece74f64e2e61c6391d8b221cffef41062b74c3268b58c9 +size 1404 diff --git a/optimized/ch/300ms/ls_eend_ch_300ms.mlmodelc/metadata.json b/optimized/ch/300ms/ls_eend_ch_300ms.mlmodelc/metadata.json index 637d583fb946317262e183e23421f1a720e2a1b1..90bda323158607a50918acb5f9fa5f44cfb1d2a9 100644 --- a/optimized/ch/300ms/ls_eend_ch_300ms.mlmodelc/metadata.json +++ b/optimized/ch/300ms/ls_eend_ch_300ms.mlmodelc/metadata.json @@ -1,7 +1,7 @@ [ { "metadataOutputVersion" : "3.0", - "shortDescription" : "LS-EEND CALLHOME streaming diarizer (pipeline, T=3, max_speakers=7)", + "shortDescription" : "LS-EEND CALLHOME streaming diarizer (pipeline, T=3, max_speakers=7, layout=raw_mel, cu=all)", "outputSchema" : [ { "hasShapeFlexibility" : "0", @@ -81,7 +81,7 @@ "specificationVersion" : 8, "mlProgramOperationTypeHistogram" : { "Ios16.reduceL2Norm" : 2, - "Ios17.reshape" : 68, + "Ios17.reshape" : 69, "Ios16.softmax" : 2, "Ios17.matmul" : 29, "Ios17.transpose" : 57, @@ -89,7 +89,7 @@ "Ios17.expandDims" : 3, "Ios17.add" : 46, "Ios16.sigmoid" : 5, - "Ios17.sliceByIndex" : 56, + "Ios17.sliceByIndex" : 59, "Tile" : 1, "Ios16.reduceSum" : 1, "Ios17.squeeze" : 2, @@ -101,7 +101,7 @@ "Ios16.silu" : 18, "Ios17.realDiv" : 20, "Ios17.linear" : 56, - "Stack" : 5, + "Stack" : 6, "Ios17.concat" : 18, "Ios16.relu" : 2, "Ios16.cumsum" : 1, @@ -128,9 +128,9 @@ "hasShapeFlexibility" : "0", "isOptional" : "0", "dataType" : "Float32", - "formattedType" : "MultiArray (Float32 1 × 3 × 345)", + "formattedType" : "MultiArray (Float32 1 × 35 × 23)", "shortDescription" : "", - "shape" : "[1, 3, 345]", + "shape" : "[1, 35, 23]", "name" : "features", "type" : "MultiArray" }, @@ -206,8 +206,8 @@ } ], "userDefinedMetadata" : { - "com.github.apple.coremltools.conversion_date" : "2026-04-16", - "config" : "{\"model_name\": \"ch\", \"model_label\": \"CALLHOME\", \"variant\": \"pipeline\", \"chunk_size\": 3, \"step_duration_ms\": 300, \"frame_duration_ms\": 100, \"frame_duration_seconds\": 0.1, \"max_speakers\": 7, \"max_nspks\": 9, \"n_units\": 256, \"n_heads\": 4, \"enc_n_layers\": 4, \"dec_n_layers\": 2, \"conv_kernel_size\": 16, \"conv_delay\": 9, \"sample_rate\": 8000, \"win_length\": 200, \"hop_length\": 80, \"n_mels\": 23, \"context_size\": 7, \"subsampling\": 10, \"feat_type\": \"logmel23_cummn\", \"pure_roll\": true}", + "com.github.apple.coremltools.conversion_date" : "2026-04-18", + "config" : "{\"model_name\": \"ch\", \"model_label\": \"CALLHOME\", \"variant\": \"pipeline\", \"chunk_size\": 3, \"step_duration_ms\": 300, \"frame_duration_ms\": 100, \"frame_duration_seconds\": 0.1, \"max_speakers\": 7, \"max_nspks\": 9, \"n_units\": 256, \"n_heads\": 4, \"enc_n_layers\": 4, \"dec_n_layers\": 2, \"conv_kernel_size\": 16, \"conv_delay\": 9, \"sample_rate\": 8000, \"win_length\": 200, \"hop_length\": 80, \"n_mels\": 23, \"context_size\": 7, \"subsampling\": 10, \"feat_type\": \"logmel23_cummn\", \"pure_roll\": true, \"input_layout\": \"raw_mel\", \"compute_units_export\": \"all\", \"raw_mel_length\": 35}", "com.github.apple.coremltools.source" : "torch==2.6.0", "com.github.apple.coremltools.version" : "9.0", "com.github.apple.coremltools.source_dialect" : "TorchScript" diff --git a/optimized/ch/300ms/ls_eend_ch_300ms.mlmodelc/model.mil b/optimized/ch/300ms/ls_eend_ch_300ms.mlmodelc/model.mil index 5eef33ebf1d0746b1ddb3ff22edac4fc03ac7b54..bef23272c0c03dca7b3c2e0e89406d225ef31cf4 100644 --- a/optimized/ch/300ms/ls_eend_ch_300ms.mlmodelc/model.mil +++ b/optimized/ch/300ms/ls_eend_ch_300ms.mlmodelc/model.mil @@ -1,234 +1,252 @@ program(1.0) [buildInfo = dict, tensor>({{"coremlc-component-MIL", "3510.2.1"}, {"coremlc-version", "3500.32.1"}, {"coremltools-component-torch", "2.6.0"}, {"coremltools-source-dialect", "TorchScript"}, {"coremltools-version", "9.0"}})] { - func main(tensor cnn_window, tensor dec_kv, tensor dec_scale, tensor enc_conv_cache, tensor enc_kv, tensor enc_scale, tensor features, tensor valid_mask) { - tensor encoder_cnn_bias = const()[name = tensor("encoder_cnn_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(64)))]; - tensor encoder_cnn_weight = const()[name = tensor("encoder_cnn_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1152)))]; - tensor encoder__causal_mask = const()[name = tensor("encoder__causal_mask"), val = tensor([[0x1p+0, 0x0p+0, 0x0p+0], [0x1p+0, 0x1p+0, 0x0p+0], [0x1p+0, 0x1p+0, 0x1p+0]])]; - tensor encoder__t_index = const()[name = tensor("encoder__t_index"), val = tensor([0x1p+0, 0x1p+1, 0x1.8p+1])]; - tensor encoder_input_projection_linear_bias = const()[name = tensor("encoder_input_projection_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4981952)))]; - tensor encoder_input_projection_linear_weight = const()[name = tensor("encoder_input_projection_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4983040)))]; - tensor encoder_pre_layer_norm_bias = const()[name = tensor("encoder_pre_layer_norm_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(5336384)))]; - tensor encoder_pre_layer_norm_weight = const()[name = tensor("encoder_pre_layer_norm_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(5337472)))]; - tensor encoder_ffn1_0_module_sequential_0_bias = const()[name = tensor("encoder_ffn1_0_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(5338560)))]; - tensor encoder_ffn1_0_module_sequential_0_weight = const()[name = tensor("encoder_ffn1_0_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(5339648)))]; - tensor encoder_ffn1_0_module_sequential_1_linear_bias = const()[name = tensor("encoder_ffn1_0_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(5340736)))]; - tensor encoder_ffn1_0_module_sequential_1_linear_weight = const()[name = tensor("encoder_ffn1_0_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(5344896)))]; - tensor encoder_ffn1_0_module_sequential_4_linear_bias = const()[name = tensor("encoder_ffn1_0_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(6393536)))]; - tensor encoder_ffn1_0_module_sequential_4_linear_weight = const()[name = tensor("encoder_ffn1_0_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(6394624)))]; - tensor encoder_ret_lns_0_bias = const()[name = tensor("encoder_ret_lns_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7443264)))]; - tensor encoder_ret_lns_0_weight = const()[name = tensor("encoder_ret_lns_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7444352)))]; - tensor encoder_q_proj_0_bias = const()[name = tensor("encoder_q_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7445440)))]; - tensor encoder_q_proj_0_weight = const()[name = tensor("encoder_q_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7446528)))]; - tensor encoder_k_proj_0_bias = const()[name = tensor("encoder_k_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7708736)))]; - tensor encoder_k_proj_0_weight = const()[name = tensor("encoder_k_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7709824)))]; - tensor encoder_v_proj_0_bias = const()[name = tensor("encoder_v_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7972032)))]; - tensor encoder_v_proj_0_weight = const()[name = tensor("encoder_v_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7973120)))]; - tensor encoder_g_proj_0_bias = const()[name = tensor("encoder_g_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8235328)))]; - tensor encoder_g_proj_0_weight = const()[name = tensor("encoder_g_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8236416)))]; - tensor encoder_out_proj_0_bias = const()[name = tensor("encoder_out_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8498624)))]; - tensor encoder_out_proj_0_weight = const()[name = tensor("encoder_out_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8499712)))]; - tensor encoder_conv_module_0_sequential_0_bias = const()[name = tensor("encoder_conv_module_0_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8761920)))]; - tensor encoder_conv_module_0_sequential_0_weight = const()[name = tensor("encoder_conv_module_0_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8763008)))]; - tensor encoder_conv_module_0_sequential_2_conv_bias = const()[name = tensor("encoder_conv_module_0_sequential_2_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8764096)))]; - tensor encoder_conv_module_0_sequential_2_conv_weight = const()[name = tensor("encoder_conv_module_0_sequential_2_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8766208)))]; - tensor encoder_conv_module_0_sequential_4_conv_weight = const()[name = tensor("encoder_conv_module_0_sequential_4_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9290560)))]; - tensor encoder_conv_module_0_sequential_5_running_var = const()[name = tensor("encoder_conv_module_0_sequential_5_running_var"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9307008)))]; - tensor encoder_conv_module_0_sequential_5_running_mean = const()[name = tensor("encoder_conv_module_0_sequential_5_running_mean"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9308096)))]; - tensor encoder_conv_module_0_sequential_5_bias = const()[name = tensor("encoder_conv_module_0_sequential_5_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9309184)))]; - tensor encoder_conv_module_0_sequential_5_weight = const()[name = tensor("encoder_conv_module_0_sequential_5_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9310272)))]; - tensor encoder_conv_module_0_sequential_7_conv_bias = const()[name = tensor("encoder_conv_module_0_sequential_7_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9311360)))]; - tensor encoder_conv_module_0_sequential_7_conv_weight = const()[name = tensor("encoder_conv_module_0_sequential_7_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9312448)))]; - tensor encoder_ffn2_0_module_sequential_0_bias = const()[name = tensor("encoder_ffn2_0_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9574656)))]; - tensor encoder_ffn2_0_module_sequential_0_weight = const()[name = tensor("encoder_ffn2_0_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9575744)))]; - tensor encoder_ffn2_0_module_sequential_1_linear_bias = const()[name = tensor("encoder_ffn2_0_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9576832)))]; - tensor encoder_ffn2_0_module_sequential_1_linear_weight = const()[name = tensor("encoder_ffn2_0_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9580992)))]; - tensor encoder_ffn2_0_module_sequential_4_linear_bias = const()[name = tensor("encoder_ffn2_0_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(10629632)))]; - tensor encoder_ffn2_0_module_sequential_4_linear_weight = const()[name = tensor("encoder_ffn2_0_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(10630720)))]; - tensor encoder_layer_norm_0_bias = const()[name = tensor("encoder_layer_norm_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11679360)))]; - tensor encoder_layer_norm_0_weight = const()[name = tensor("encoder_layer_norm_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11680448)))]; - tensor encoder_ffn1_1_module_sequential_0_bias = const()[name = tensor("encoder_ffn1_1_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11681536)))]; - tensor encoder_ffn1_1_module_sequential_0_weight = const()[name = tensor("encoder_ffn1_1_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11682624)))]; - tensor encoder_ffn1_1_module_sequential_1_linear_bias = const()[name = tensor("encoder_ffn1_1_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11683712)))]; - tensor encoder_ffn1_1_module_sequential_1_linear_weight = const()[name = tensor("encoder_ffn1_1_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11687872)))]; - tensor encoder_ffn1_1_module_sequential_4_linear_bias = const()[name = tensor("encoder_ffn1_1_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(12736512)))]; - tensor encoder_ffn1_1_module_sequential_4_linear_weight = const()[name = tensor("encoder_ffn1_1_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(12737600)))]; - tensor encoder_ret_lns_1_bias = const()[name = tensor("encoder_ret_lns_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(13786240)))]; - tensor encoder_ret_lns_1_weight = const()[name = tensor("encoder_ret_lns_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(13787328)))]; - tensor encoder_q_proj_1_bias = const()[name = tensor("encoder_q_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(13788416)))]; - tensor encoder_q_proj_1_weight = const()[name = tensor("encoder_q_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(13789504)))]; - tensor encoder_k_proj_1_bias = const()[name = tensor("encoder_k_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14051712)))]; - tensor encoder_k_proj_1_weight = const()[name = tensor("encoder_k_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14052800)))]; - tensor encoder_v_proj_1_bias = const()[name = tensor("encoder_v_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14315008)))]; - tensor encoder_v_proj_1_weight = const()[name = tensor("encoder_v_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14316096)))]; - tensor encoder_g_proj_1_bias = const()[name = tensor("encoder_g_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14578304)))]; - tensor encoder_g_proj_1_weight = const()[name = tensor("encoder_g_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14579392)))]; - tensor encoder_out_proj_1_bias = const()[name = tensor("encoder_out_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14841600)))]; - tensor encoder_out_proj_1_weight = const()[name = tensor("encoder_out_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14842688)))]; - tensor encoder_conv_module_1_sequential_0_bias = const()[name = tensor("encoder_conv_module_1_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15104896)))]; - tensor encoder_conv_module_1_sequential_0_weight = const()[name = tensor("encoder_conv_module_1_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15105984)))]; - tensor encoder_conv_module_1_sequential_2_conv_bias = const()[name = tensor("encoder_conv_module_1_sequential_2_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15107072)))]; - tensor encoder_conv_module_1_sequential_2_conv_weight = const()[name = tensor("encoder_conv_module_1_sequential_2_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15109184)))]; - tensor encoder_conv_module_1_sequential_4_conv_weight = const()[name = tensor("encoder_conv_module_1_sequential_4_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15633536)))]; - tensor encoder_conv_module_1_sequential_5_running_var = const()[name = tensor("encoder_conv_module_1_sequential_5_running_var"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15649984)))]; - tensor encoder_conv_module_1_sequential_5_running_mean = const()[name = tensor("encoder_conv_module_1_sequential_5_running_mean"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15651072)))]; - tensor encoder_conv_module_1_sequential_5_bias = const()[name = tensor("encoder_conv_module_1_sequential_5_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15652160)))]; - tensor encoder_conv_module_1_sequential_5_weight = const()[name = tensor("encoder_conv_module_1_sequential_5_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15653248)))]; - tensor encoder_conv_module_1_sequential_7_conv_bias = const()[name = tensor("encoder_conv_module_1_sequential_7_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15654336)))]; - tensor encoder_conv_module_1_sequential_7_conv_weight = const()[name = tensor("encoder_conv_module_1_sequential_7_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15655424)))]; - tensor encoder_ffn2_1_module_sequential_0_bias = const()[name = tensor("encoder_ffn2_1_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15917632)))]; - tensor encoder_ffn2_1_module_sequential_0_weight = const()[name = tensor("encoder_ffn2_1_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15918720)))]; - tensor encoder_ffn2_1_module_sequential_1_linear_bias = const()[name = tensor("encoder_ffn2_1_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15919808)))]; - tensor encoder_ffn2_1_module_sequential_1_linear_weight = const()[name = tensor("encoder_ffn2_1_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15923968)))]; - tensor encoder_ffn2_1_module_sequential_4_linear_bias = const()[name = tensor("encoder_ffn2_1_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(16972608)))]; - tensor encoder_ffn2_1_module_sequential_4_linear_weight = const()[name = tensor("encoder_ffn2_1_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(16973696)))]; - tensor encoder_layer_norm_1_bias = const()[name = tensor("encoder_layer_norm_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18022336)))]; - tensor encoder_layer_norm_1_weight = const()[name = tensor("encoder_layer_norm_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18023424)))]; - tensor encoder_ffn1_2_module_sequential_0_bias = const()[name = tensor("encoder_ffn1_2_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18024512)))]; - tensor encoder_ffn1_2_module_sequential_0_weight = const()[name = tensor("encoder_ffn1_2_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18025600)))]; - tensor encoder_ffn1_2_module_sequential_1_linear_bias = const()[name = tensor("encoder_ffn1_2_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18026688)))]; - tensor encoder_ffn1_2_module_sequential_1_linear_weight = const()[name = tensor("encoder_ffn1_2_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18030848)))]; - tensor encoder_ffn1_2_module_sequential_4_linear_bias = const()[name = tensor("encoder_ffn1_2_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(19079488)))]; - tensor encoder_ffn1_2_module_sequential_4_linear_weight = const()[name = tensor("encoder_ffn1_2_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(19080576)))]; - tensor encoder_ret_lns_2_bias = const()[name = tensor("encoder_ret_lns_2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20129216)))]; - tensor encoder_ret_lns_2_weight = const()[name = tensor("encoder_ret_lns_2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20130304)))]; - tensor encoder_q_proj_2_bias = const()[name = tensor("encoder_q_proj_2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20131392)))]; - tensor encoder_q_proj_2_weight = const()[name = tensor("encoder_q_proj_2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20132480)))]; - tensor encoder_k_proj_2_bias = const()[name = tensor("encoder_k_proj_2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20394688)))]; - tensor encoder_k_proj_2_weight = const()[name = tensor("encoder_k_proj_2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20395776)))]; - tensor encoder_v_proj_2_bias = const()[name = tensor("encoder_v_proj_2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20657984)))]; - tensor encoder_v_proj_2_weight = const()[name = tensor("encoder_v_proj_2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20659072)))]; - tensor encoder_g_proj_2_bias = const()[name = tensor("encoder_g_proj_2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20921280)))]; - tensor encoder_g_proj_2_weight = const()[name = tensor("encoder_g_proj_2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20922368)))]; - tensor encoder_out_proj_2_bias = const()[name = tensor("encoder_out_proj_2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21184576)))]; - tensor encoder_out_proj_2_weight = const()[name = tensor("encoder_out_proj_2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21185664)))]; - tensor encoder_conv_module_2_sequential_0_bias = const()[name = tensor("encoder_conv_module_2_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21447872)))]; - tensor encoder_conv_module_2_sequential_0_weight = const()[name = tensor("encoder_conv_module_2_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21448960)))]; - tensor encoder_conv_module_2_sequential_2_conv_bias = const()[name = tensor("encoder_conv_module_2_sequential_2_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21450048)))]; - tensor encoder_conv_module_2_sequential_2_conv_weight = const()[name = tensor("encoder_conv_module_2_sequential_2_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21452160)))]; - tensor encoder_conv_module_2_sequential_4_conv_weight = const()[name = tensor("encoder_conv_module_2_sequential_4_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21976512)))]; - tensor encoder_conv_module_2_sequential_5_running_var = const()[name = tensor("encoder_conv_module_2_sequential_5_running_var"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21992960)))]; - tensor encoder_conv_module_2_sequential_5_running_mean = const()[name = tensor("encoder_conv_module_2_sequential_5_running_mean"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21994048)))]; - tensor encoder_conv_module_2_sequential_5_bias = const()[name = tensor("encoder_conv_module_2_sequential_5_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21995136)))]; - tensor encoder_conv_module_2_sequential_5_weight = const()[name = tensor("encoder_conv_module_2_sequential_5_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21996224)))]; - tensor encoder_conv_module_2_sequential_7_conv_bias = const()[name = tensor("encoder_conv_module_2_sequential_7_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21997312)))]; - tensor encoder_conv_module_2_sequential_7_conv_weight = const()[name = tensor("encoder_conv_module_2_sequential_7_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21998400)))]; - tensor encoder_ffn2_2_module_sequential_0_bias = const()[name = tensor("encoder_ffn2_2_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(22260608)))]; - tensor encoder_ffn2_2_module_sequential_0_weight = const()[name = tensor("encoder_ffn2_2_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(22261696)))]; - tensor encoder_ffn2_2_module_sequential_1_linear_bias = const()[name = tensor("encoder_ffn2_2_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(22262784)))]; - tensor encoder_ffn2_2_module_sequential_1_linear_weight = const()[name = tensor("encoder_ffn2_2_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(22266944)))]; - tensor encoder_ffn2_2_module_sequential_4_linear_bias = const()[name = tensor("encoder_ffn2_2_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(23315584)))]; - tensor encoder_ffn2_2_module_sequential_4_linear_weight = const()[name = tensor("encoder_ffn2_2_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(23316672)))]; - tensor encoder_layer_norm_2_bias = const()[name = tensor("encoder_layer_norm_2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(24365312)))]; - tensor encoder_layer_norm_2_weight = const()[name = tensor("encoder_layer_norm_2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(24366400)))]; - tensor encoder_ffn1_3_module_sequential_0_bias = const()[name = tensor("encoder_ffn1_3_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(24367488)))]; - tensor encoder_ffn1_3_module_sequential_0_weight = const()[name = tensor("encoder_ffn1_3_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(24368576)))]; - tensor encoder_ffn1_3_module_sequential_1_linear_bias = const()[name = tensor("encoder_ffn1_3_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(24369664)))]; - tensor encoder_ffn1_3_module_sequential_1_linear_weight = const()[name = tensor("encoder_ffn1_3_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(24373824)))]; - tensor encoder_ffn1_3_module_sequential_4_linear_bias = const()[name = tensor("encoder_ffn1_3_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(25422464)))]; - tensor encoder_ffn1_3_module_sequential_4_linear_weight = const()[name = tensor("encoder_ffn1_3_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(25423552)))]; - tensor encoder_ret_lns_3_bias = const()[name = tensor("encoder_ret_lns_3_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(26472192)))]; - tensor encoder_ret_lns_3_weight = const()[name = tensor("encoder_ret_lns_3_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(26473280)))]; - tensor encoder_q_proj_3_bias = const()[name = tensor("encoder_q_proj_3_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(26474368)))]; - tensor encoder_q_proj_3_weight = const()[name = tensor("encoder_q_proj_3_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(26475456)))]; - tensor encoder_k_proj_3_bias = const()[name = tensor("encoder_k_proj_3_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(26737664)))]; - tensor encoder_k_proj_3_weight = const()[name = tensor("encoder_k_proj_3_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(26738752)))]; - tensor encoder_v_proj_3_bias = const()[name = tensor("encoder_v_proj_3_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27000960)))]; - tensor encoder_v_proj_3_weight = const()[name = tensor("encoder_v_proj_3_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27002048)))]; - tensor encoder_g_proj_3_bias = const()[name = tensor("encoder_g_proj_3_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27264256)))]; - tensor encoder_g_proj_3_weight = const()[name = tensor("encoder_g_proj_3_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27265344)))]; - tensor encoder_out_proj_3_bias = const()[name = tensor("encoder_out_proj_3_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27527552)))]; - tensor encoder_out_proj_3_weight = const()[name = tensor("encoder_out_proj_3_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27528640)))]; - tensor encoder_conv_module_3_sequential_0_bias = const()[name = tensor("encoder_conv_module_3_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27790848)))]; - tensor encoder_conv_module_3_sequential_0_weight = const()[name = tensor("encoder_conv_module_3_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27791936)))]; - tensor encoder_conv_module_3_sequential_2_conv_bias = const()[name = tensor("encoder_conv_module_3_sequential_2_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27793024)))]; - tensor encoder_conv_module_3_sequential_2_conv_weight = const()[name = tensor("encoder_conv_module_3_sequential_2_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27795136)))]; - tensor encoder_conv_module_3_sequential_4_conv_weight = const()[name = tensor("encoder_conv_module_3_sequential_4_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28319488)))]; - tensor encoder_conv_module_3_sequential_5_running_var = const()[name = tensor("encoder_conv_module_3_sequential_5_running_var"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28335936)))]; - tensor encoder_conv_module_3_sequential_5_running_mean = const()[name = tensor("encoder_conv_module_3_sequential_5_running_mean"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28337024)))]; - tensor encoder_conv_module_3_sequential_5_bias = const()[name = tensor("encoder_conv_module_3_sequential_5_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28338112)))]; - tensor encoder_conv_module_3_sequential_5_weight = const()[name = tensor("encoder_conv_module_3_sequential_5_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28339200)))]; - tensor encoder_conv_module_3_sequential_7_conv_bias = const()[name = tensor("encoder_conv_module_3_sequential_7_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28340288)))]; - tensor encoder_conv_module_3_sequential_7_conv_weight = const()[name = tensor("encoder_conv_module_3_sequential_7_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28341376)))]; - tensor encoder_ffn2_3_module_sequential_0_bias = const()[name = tensor("encoder_ffn2_3_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28603584)))]; - tensor encoder_ffn2_3_module_sequential_0_weight = const()[name = tensor("encoder_ffn2_3_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28604672)))]; - tensor encoder_ffn2_3_module_sequential_1_linear_bias = const()[name = tensor("encoder_ffn2_3_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28605760)))]; - tensor encoder_ffn2_3_module_sequential_1_linear_weight = const()[name = tensor("encoder_ffn2_3_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28609920)))]; - tensor encoder_ffn2_3_module_sequential_4_linear_bias = const()[name = tensor("encoder_ffn2_3_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(29658560)))]; - tensor encoder_ffn2_3_module_sequential_4_linear_weight = const()[name = tensor("encoder_ffn2_3_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(29659648)))]; - tensor encoder_layer_norm_3_bias = const()[name = tensor("encoder_layer_norm_3_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(30708288)))]; - tensor encoder_layer_norm_3_weight = const()[name = tensor("encoder_layer_norm_3_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(30709376)))]; - tensor decoder__causal_mask = const()[name = tensor("decoder__causal_mask"), val = tensor([[0x1p+0, 0x0p+0, 0x0p+0], [0x1p+0, 0x1p+0, 0x0p+0], [0x1p+0, 0x1p+0, 0x1p+0]])]; - tensor decoder_convert_bias = const()[name = tensor("decoder_convert_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(30710464)))]; - tensor decoder_convert_weight = const()[name = tensor("decoder_convert_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(30711552)))]; - tensor decoder_q_proj_0_bias = const()[name = tensor("decoder_q_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(31235904)))]; - tensor decoder_q_proj_0_weight = const()[name = tensor("decoder_q_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(31236992)))]; - tensor decoder_k_proj_0_bias = const()[name = tensor("decoder_k_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(31499200)))]; - tensor decoder_k_proj_0_weight = const()[name = tensor("decoder_k_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(31500288)))]; - tensor decoder_v_proj_0_bias = const()[name = tensor("decoder_v_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(31762496)))]; - tensor decoder_v_proj_0_weight = const()[name = tensor("decoder_v_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(31763584)))]; - tensor decoder_g_proj_0_bias = const()[name = tensor("decoder_g_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32025792)))]; - tensor decoder_g_proj_0_weight = const()[name = tensor("decoder_g_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32026880)))]; - tensor decoder_out_proj_0_bias = const()[name = tensor("decoder_out_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32289088)))]; - tensor decoder_out_proj_0_weight = const()[name = tensor("decoder_out_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32290176)))]; - tensor decoder_norm11_0_bias = const()[name = tensor("decoder_norm11_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32552384)))]; - tensor decoder_norm11_0_weight = const()[name = tensor("decoder_norm11_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32553472)))]; - tensor decoder_self_attn2_0_out_proj_bias = const()[name = tensor("decoder_self_attn2_0_out_proj_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32554560)))]; - tensor decoder_self_attn2_0_out_proj_weight = const()[name = tensor("decoder_self_attn2_0_out_proj_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32555648)))]; - tensor decoder_self_attn2_0_in_proj_bias = const()[name = tensor("decoder_self_attn2_0_in_proj_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32817856)))]; - tensor decoder_self_attn2_0_in_proj_weight = const()[name = tensor("decoder_self_attn2_0_in_proj_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32820992)))]; - tensor decoder_norm21_0_bias = const()[name = tensor("decoder_norm21_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(33607488)))]; - tensor decoder_norm21_0_weight = const()[name = tensor("decoder_norm21_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(33608576)))]; - tensor decoder_linear1_0_bias = const()[name = tensor("decoder_linear1_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(33609664)))]; - tensor decoder_linear1_0_weight = const()[name = tensor("decoder_linear1_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(33617920)))]; - tensor decoder_linear2_0_bias = const()[name = tensor("decoder_linear2_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(35715136)))]; - tensor decoder_linear2_0_weight = const()[name = tensor("decoder_linear2_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(35716224)))]; - tensor decoder_norm22_0_bias = const()[name = tensor("decoder_norm22_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(37813440)))]; - tensor decoder_norm22_0_weight = const()[name = tensor("decoder_norm22_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(37814528)))]; - tensor decoder_q_proj_1_bias = const()[name = tensor("decoder_q_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(37815616)))]; - tensor decoder_q_proj_1_weight = const()[name = tensor("decoder_q_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(37816704)))]; - tensor decoder_k_proj_1_bias = const()[name = tensor("decoder_k_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38078912)))]; - tensor decoder_k_proj_1_weight = const()[name = tensor("decoder_k_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38080000)))]; - tensor decoder_v_proj_1_bias = const()[name = tensor("decoder_v_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38342208)))]; - tensor decoder_v_proj_1_weight = const()[name = tensor("decoder_v_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38343296)))]; - tensor decoder_g_proj_1_bias = const()[name = tensor("decoder_g_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38605504)))]; - tensor decoder_g_proj_1_weight = const()[name = tensor("decoder_g_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38606592)))]; - tensor decoder_out_proj_1_bias = const()[name = tensor("decoder_out_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38868800)))]; - tensor decoder_out_proj_1_weight = const()[name = tensor("decoder_out_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38869888)))]; - tensor decoder_norm11_1_bias = const()[name = tensor("decoder_norm11_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(39132096)))]; - tensor decoder_norm11_1_weight = const()[name = tensor("decoder_norm11_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(39133184)))]; - tensor decoder_self_attn2_1_out_proj_bias = const()[name = tensor("decoder_self_attn2_1_out_proj_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(39134272)))]; - tensor decoder_self_attn2_1_out_proj_weight = const()[name = tensor("decoder_self_attn2_1_out_proj_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(39135360)))]; - tensor decoder_self_attn2_1_in_proj_bias = const()[name = tensor("decoder_self_attn2_1_in_proj_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(39397568)))]; - tensor decoder_self_attn2_1_in_proj_weight = const()[name = tensor("decoder_self_attn2_1_in_proj_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(39400704)))]; - tensor decoder_norm21_1_bias = const()[name = tensor("decoder_norm21_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(40187200)))]; - tensor decoder_norm21_1_weight = const()[name = tensor("decoder_norm21_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(40188288)))]; - tensor decoder_linear1_1_bias = const()[name = tensor("decoder_linear1_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(40189376)))]; - tensor decoder_linear1_1_weight = const()[name = tensor("decoder_linear1_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(40197632)))]; - tensor decoder_linear2_1_bias = const()[name = tensor("decoder_linear2_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(42294848)))]; - tensor decoder_linear2_1_weight = const()[name = tensor("decoder_linear2_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(42295936)))]; - tensor decoder_norm22_1_bias = const()[name = tensor("decoder_norm22_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(44393152)))]; - tensor decoder_norm22_1_weight = const()[name = tensor("decoder_norm22_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(44394240)))]; - tensor var_11 = const()[name = tensor("op_11"), val = tensor(0x1.5798eep-27)]; - tensor var_17 = const()[name = tensor("op_17"), val = tensor(0x1.0c6f7ap-20)]; - tensor var_20 = const()[name = tensor("op_20"), val = tensor(2)]; - tensor var_23 = const()[name = tensor("op_23"), val = tensor(0)]; - tensor var_26 = const()[name = tensor("op_26"), val = tensor(1)]; - tensor var_28 = const()[name = tensor("op_28"), val = tensor(0x1.4f8b58p-17)]; - tensor var_29 = const()[name = tensor("op_29"), val = tensor(true)]; - tensor input_1 = linear(bias = encoder_input_projection_linear_bias, weight = encoder_input_projection_linear_weight, x = features)[name = tensor("linear_0")]; - tensor input_3_axes_0 = const()[name = tensor("input_3_axes_0"), val = tensor([-1])]; - tensor input_3 = layer_norm(axes = input_3_axes_0, beta = encoder_pre_layer_norm_bias, epsilon = var_28, gamma = encoder_pre_layer_norm_weight, x = input_1)[name = tensor("input_3")]; + func main(tensor cnn_window, tensor dec_kv, tensor dec_scale, tensor enc_conv_cache, tensor enc_kv, tensor enc_scale, tensor features, tensor valid_mask) { + tensor inner_encoder_cnn_bias = const()[name = tensor("inner_encoder_cnn_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(64)))]; + tensor inner_encoder_cnn_weight = const()[name = tensor("inner_encoder_cnn_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1152)))]; + tensor inner_encoder__causal_mask = const()[name = tensor("inner_encoder__causal_mask"), val = tensor([[0x1p+0, 0x0p+0, 0x0p+0], [0x1p+0, 0x1p+0, 0x0p+0], [0x1p+0, 0x1p+0, 0x1p+0]])]; + tensor inner_encoder__t_index = const()[name = tensor("inner_encoder__t_index"), val = tensor([0x1p+0, 0x1p+1, 0x1.8p+1])]; + tensor inner_encoder_input_projection_linear_bias = const()[name = tensor("inner_encoder_input_projection_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4981952)))]; + tensor inner_encoder_input_projection_linear_weight = const()[name = tensor("inner_encoder_input_projection_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4983040)))]; + tensor inner_encoder_pre_layer_norm_bias = const()[name = tensor("inner_encoder_pre_layer_norm_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(5336384)))]; + tensor inner_encoder_pre_layer_norm_weight = const()[name = tensor("inner_encoder_pre_layer_norm_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(5337472)))]; + tensor inner_encoder_ffn1_0_module_sequential_0_bias = const()[name = tensor("inner_encoder_ffn1_0_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(5338560)))]; + tensor inner_encoder_ffn1_0_module_sequential_0_weight = const()[name = tensor("inner_encoder_ffn1_0_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(5339648)))]; + tensor inner_encoder_ffn1_0_module_sequential_1_linear_bias = const()[name = tensor("inner_encoder_ffn1_0_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(5340736)))]; + tensor inner_encoder_ffn1_0_module_sequential_1_linear_weight = const()[name = tensor("inner_encoder_ffn1_0_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(5344896)))]; + tensor inner_encoder_ffn1_0_module_sequential_4_linear_bias = const()[name = tensor("inner_encoder_ffn1_0_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(6393536)))]; + tensor inner_encoder_ffn1_0_module_sequential_4_linear_weight = const()[name = tensor("inner_encoder_ffn1_0_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(6394624)))]; + tensor inner_encoder_ret_lns_0_bias = const()[name = tensor("inner_encoder_ret_lns_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7443264)))]; + tensor inner_encoder_ret_lns_0_weight = const()[name = tensor("inner_encoder_ret_lns_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7444352)))]; + tensor inner_encoder_q_proj_0_bias = const()[name = tensor("inner_encoder_q_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7445440)))]; + tensor inner_encoder_q_proj_0_weight = const()[name = tensor("inner_encoder_q_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7446528)))]; + tensor inner_encoder_k_proj_0_bias = const()[name = tensor("inner_encoder_k_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7708736)))]; + tensor inner_encoder_k_proj_0_weight = const()[name = tensor("inner_encoder_k_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7709824)))]; + tensor inner_encoder_v_proj_0_bias = const()[name = tensor("inner_encoder_v_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7972032)))]; + tensor inner_encoder_v_proj_0_weight = const()[name = tensor("inner_encoder_v_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7973120)))]; + tensor inner_encoder_g_proj_0_bias = const()[name = tensor("inner_encoder_g_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8235328)))]; + tensor inner_encoder_g_proj_0_weight = const()[name = tensor("inner_encoder_g_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8236416)))]; + tensor inner_encoder_out_proj_0_bias = const()[name = tensor("inner_encoder_out_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8498624)))]; + tensor inner_encoder_out_proj_0_weight = const()[name = tensor("inner_encoder_out_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8499712)))]; + tensor inner_encoder_conv_module_0_sequential_0_bias = const()[name = tensor("inner_encoder_conv_module_0_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8761920)))]; + tensor inner_encoder_conv_module_0_sequential_0_weight = const()[name = tensor("inner_encoder_conv_module_0_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8763008)))]; + tensor inner_encoder_conv_module_0_sequential_2_conv_bias = const()[name = tensor("inner_encoder_conv_module_0_sequential_2_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8764096)))]; + tensor inner_encoder_conv_module_0_sequential_2_conv_weight = const()[name = tensor("inner_encoder_conv_module_0_sequential_2_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8766208)))]; + tensor inner_encoder_conv_module_0_sequential_4_conv_weight = const()[name = tensor("inner_encoder_conv_module_0_sequential_4_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9290560)))]; + tensor inner_encoder_conv_module_0_sequential_5_running_var = const()[name = tensor("inner_encoder_conv_module_0_sequential_5_running_var"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9307008)))]; + tensor inner_encoder_conv_module_0_sequential_5_running_mean = const()[name = tensor("inner_encoder_conv_module_0_sequential_5_running_mean"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9308096)))]; + tensor inner_encoder_conv_module_0_sequential_5_bias = const()[name = tensor("inner_encoder_conv_module_0_sequential_5_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9309184)))]; + tensor inner_encoder_conv_module_0_sequential_5_weight = const()[name = tensor("inner_encoder_conv_module_0_sequential_5_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9310272)))]; + tensor inner_encoder_conv_module_0_sequential_7_conv_bias = const()[name = tensor("inner_encoder_conv_module_0_sequential_7_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9311360)))]; + tensor inner_encoder_conv_module_0_sequential_7_conv_weight = const()[name = tensor("inner_encoder_conv_module_0_sequential_7_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9312448)))]; + tensor inner_encoder_ffn2_0_module_sequential_0_bias = const()[name = tensor("inner_encoder_ffn2_0_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9574656)))]; + tensor inner_encoder_ffn2_0_module_sequential_0_weight = const()[name = tensor("inner_encoder_ffn2_0_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9575744)))]; + tensor inner_encoder_ffn2_0_module_sequential_1_linear_bias = const()[name = tensor("inner_encoder_ffn2_0_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9576832)))]; + tensor inner_encoder_ffn2_0_module_sequential_1_linear_weight = const()[name = tensor("inner_encoder_ffn2_0_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9580992)))]; + tensor inner_encoder_ffn2_0_module_sequential_4_linear_bias = const()[name = tensor("inner_encoder_ffn2_0_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(10629632)))]; + tensor inner_encoder_ffn2_0_module_sequential_4_linear_weight = const()[name = tensor("inner_encoder_ffn2_0_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(10630720)))]; + tensor inner_encoder_layer_norm_0_bias = const()[name = tensor("inner_encoder_layer_norm_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11679360)))]; + tensor inner_encoder_layer_norm_0_weight = const()[name = tensor("inner_encoder_layer_norm_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11680448)))]; + tensor inner_encoder_ffn1_1_module_sequential_0_bias = const()[name = tensor("inner_encoder_ffn1_1_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11681536)))]; + tensor inner_encoder_ffn1_1_module_sequential_0_weight = const()[name = tensor("inner_encoder_ffn1_1_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11682624)))]; + tensor inner_encoder_ffn1_1_module_sequential_1_linear_bias = const()[name = tensor("inner_encoder_ffn1_1_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11683712)))]; + tensor inner_encoder_ffn1_1_module_sequential_1_linear_weight = const()[name = tensor("inner_encoder_ffn1_1_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11687872)))]; + tensor inner_encoder_ffn1_1_module_sequential_4_linear_bias = const()[name = tensor("inner_encoder_ffn1_1_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(12736512)))]; + tensor inner_encoder_ffn1_1_module_sequential_4_linear_weight = const()[name = tensor("inner_encoder_ffn1_1_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(12737600)))]; + tensor inner_encoder_ret_lns_1_bias = const()[name = tensor("inner_encoder_ret_lns_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(13786240)))]; + tensor inner_encoder_ret_lns_1_weight = const()[name = tensor("inner_encoder_ret_lns_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(13787328)))]; + tensor inner_encoder_q_proj_1_bias = const()[name = tensor("inner_encoder_q_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(13788416)))]; + tensor inner_encoder_q_proj_1_weight = const()[name = tensor("inner_encoder_q_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(13789504)))]; + tensor inner_encoder_k_proj_1_bias = const()[name = tensor("inner_encoder_k_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14051712)))]; + tensor inner_encoder_k_proj_1_weight = const()[name = tensor("inner_encoder_k_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14052800)))]; + tensor inner_encoder_v_proj_1_bias = const()[name = tensor("inner_encoder_v_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14315008)))]; + tensor inner_encoder_v_proj_1_weight = const()[name = tensor("inner_encoder_v_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14316096)))]; + tensor inner_encoder_g_proj_1_bias = const()[name = tensor("inner_encoder_g_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14578304)))]; + tensor inner_encoder_g_proj_1_weight = const()[name = tensor("inner_encoder_g_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14579392)))]; + tensor inner_encoder_out_proj_1_bias = const()[name = tensor("inner_encoder_out_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14841600)))]; + tensor inner_encoder_out_proj_1_weight = const()[name = tensor("inner_encoder_out_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14842688)))]; + tensor inner_encoder_conv_module_1_sequential_0_bias = const()[name = tensor("inner_encoder_conv_module_1_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15104896)))]; + tensor inner_encoder_conv_module_1_sequential_0_weight = const()[name = tensor("inner_encoder_conv_module_1_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15105984)))]; + tensor inner_encoder_conv_module_1_sequential_2_conv_bias = const()[name = tensor("inner_encoder_conv_module_1_sequential_2_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15107072)))]; + tensor inner_encoder_conv_module_1_sequential_2_conv_weight = const()[name = tensor("inner_encoder_conv_module_1_sequential_2_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15109184)))]; + tensor inner_encoder_conv_module_1_sequential_4_conv_weight = const()[name = tensor("inner_encoder_conv_module_1_sequential_4_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15633536)))]; + tensor inner_encoder_conv_module_1_sequential_5_running_var = const()[name = tensor("inner_encoder_conv_module_1_sequential_5_running_var"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15649984)))]; + tensor inner_encoder_conv_module_1_sequential_5_running_mean = const()[name = tensor("inner_encoder_conv_module_1_sequential_5_running_mean"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15651072)))]; + tensor inner_encoder_conv_module_1_sequential_5_bias = const()[name = tensor("inner_encoder_conv_module_1_sequential_5_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15652160)))]; + tensor inner_encoder_conv_module_1_sequential_5_weight = const()[name = tensor("inner_encoder_conv_module_1_sequential_5_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15653248)))]; + tensor inner_encoder_conv_module_1_sequential_7_conv_bias = const()[name = tensor("inner_encoder_conv_module_1_sequential_7_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15654336)))]; + tensor inner_encoder_conv_module_1_sequential_7_conv_weight = const()[name = tensor("inner_encoder_conv_module_1_sequential_7_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15655424)))]; + tensor inner_encoder_ffn2_1_module_sequential_0_bias = const()[name = tensor("inner_encoder_ffn2_1_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15917632)))]; + tensor inner_encoder_ffn2_1_module_sequential_0_weight = const()[name = tensor("inner_encoder_ffn2_1_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15918720)))]; + tensor inner_encoder_ffn2_1_module_sequential_1_linear_bias = const()[name = tensor("inner_encoder_ffn2_1_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15919808)))]; + tensor inner_encoder_ffn2_1_module_sequential_1_linear_weight = const()[name = tensor("inner_encoder_ffn2_1_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15923968)))]; + tensor inner_encoder_ffn2_1_module_sequential_4_linear_bias = const()[name = tensor("inner_encoder_ffn2_1_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(16972608)))]; + tensor inner_encoder_ffn2_1_module_sequential_4_linear_weight = const()[name = tensor("inner_encoder_ffn2_1_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(16973696)))]; + tensor inner_encoder_layer_norm_1_bias = const()[name = tensor("inner_encoder_layer_norm_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18022336)))]; + tensor inner_encoder_layer_norm_1_weight = const()[name = tensor("inner_encoder_layer_norm_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18023424)))]; + tensor inner_encoder_ffn1_2_module_sequential_0_bias = const()[name = tensor("inner_encoder_ffn1_2_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18024512)))]; + tensor inner_encoder_ffn1_2_module_sequential_0_weight = const()[name = tensor("inner_encoder_ffn1_2_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18025600)))]; + tensor inner_encoder_ffn1_2_module_sequential_1_linear_bias = const()[name = tensor("inner_encoder_ffn1_2_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18026688)))]; + tensor inner_encoder_ffn1_2_module_sequential_1_linear_weight = const()[name = tensor("inner_encoder_ffn1_2_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18030848)))]; + tensor inner_encoder_ffn1_2_module_sequential_4_linear_bias = const()[name = tensor("inner_encoder_ffn1_2_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(19079488)))]; + tensor inner_encoder_ffn1_2_module_sequential_4_linear_weight = const()[name = tensor("inner_encoder_ffn1_2_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(19080576)))]; + tensor inner_encoder_ret_lns_2_bias = const()[name = tensor("inner_encoder_ret_lns_2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20129216)))]; + tensor inner_encoder_ret_lns_2_weight = const()[name = tensor("inner_encoder_ret_lns_2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20130304)))]; + tensor inner_encoder_q_proj_2_bias = const()[name = tensor("inner_encoder_q_proj_2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20131392)))]; + tensor inner_encoder_q_proj_2_weight = const()[name = tensor("inner_encoder_q_proj_2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20132480)))]; + tensor inner_encoder_k_proj_2_bias = const()[name = tensor("inner_encoder_k_proj_2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20394688)))]; + tensor inner_encoder_k_proj_2_weight = const()[name = tensor("inner_encoder_k_proj_2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20395776)))]; + tensor inner_encoder_v_proj_2_bias = const()[name = tensor("inner_encoder_v_proj_2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20657984)))]; + tensor inner_encoder_v_proj_2_weight = const()[name = tensor("inner_encoder_v_proj_2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20659072)))]; + tensor inner_encoder_g_proj_2_bias = const()[name = tensor("inner_encoder_g_proj_2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20921280)))]; + tensor inner_encoder_g_proj_2_weight = const()[name = tensor("inner_encoder_g_proj_2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20922368)))]; + tensor inner_encoder_out_proj_2_bias = const()[name = tensor("inner_encoder_out_proj_2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21184576)))]; + tensor inner_encoder_out_proj_2_weight = const()[name = tensor("inner_encoder_out_proj_2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21185664)))]; + tensor inner_encoder_conv_module_2_sequential_0_bias = const()[name = tensor("inner_encoder_conv_module_2_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21447872)))]; + tensor inner_encoder_conv_module_2_sequential_0_weight = const()[name = tensor("inner_encoder_conv_module_2_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21448960)))]; + tensor inner_encoder_conv_module_2_sequential_2_conv_bias = const()[name = tensor("inner_encoder_conv_module_2_sequential_2_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21450048)))]; + tensor inner_encoder_conv_module_2_sequential_2_conv_weight = const()[name = tensor("inner_encoder_conv_module_2_sequential_2_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21452160)))]; + tensor inner_encoder_conv_module_2_sequential_4_conv_weight = const()[name = tensor("inner_encoder_conv_module_2_sequential_4_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21976512)))]; + tensor inner_encoder_conv_module_2_sequential_5_running_var = const()[name = tensor("inner_encoder_conv_module_2_sequential_5_running_var"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21992960)))]; + tensor inner_encoder_conv_module_2_sequential_5_running_mean = const()[name = tensor("inner_encoder_conv_module_2_sequential_5_running_mean"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21994048)))]; + tensor inner_encoder_conv_module_2_sequential_5_bias = const()[name = tensor("inner_encoder_conv_module_2_sequential_5_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21995136)))]; + tensor inner_encoder_conv_module_2_sequential_5_weight = const()[name = tensor("inner_encoder_conv_module_2_sequential_5_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21996224)))]; + tensor inner_encoder_conv_module_2_sequential_7_conv_bias = const()[name = tensor("inner_encoder_conv_module_2_sequential_7_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21997312)))]; + tensor inner_encoder_conv_module_2_sequential_7_conv_weight = const()[name = tensor("inner_encoder_conv_module_2_sequential_7_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21998400)))]; + tensor inner_encoder_ffn2_2_module_sequential_0_bias = const()[name = tensor("inner_encoder_ffn2_2_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(22260608)))]; + tensor inner_encoder_ffn2_2_module_sequential_0_weight = const()[name = tensor("inner_encoder_ffn2_2_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(22261696)))]; + tensor inner_encoder_ffn2_2_module_sequential_1_linear_bias = const()[name = tensor("inner_encoder_ffn2_2_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(22262784)))]; + tensor inner_encoder_ffn2_2_module_sequential_1_linear_weight = const()[name = tensor("inner_encoder_ffn2_2_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(22266944)))]; + tensor inner_encoder_ffn2_2_module_sequential_4_linear_bias = const()[name = tensor("inner_encoder_ffn2_2_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(23315584)))]; + tensor inner_encoder_ffn2_2_module_sequential_4_linear_weight = const()[name = tensor("inner_encoder_ffn2_2_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(23316672)))]; + tensor inner_encoder_layer_norm_2_bias = const()[name = tensor("inner_encoder_layer_norm_2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(24365312)))]; + tensor inner_encoder_layer_norm_2_weight = const()[name = tensor("inner_encoder_layer_norm_2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(24366400)))]; + tensor inner_encoder_ffn1_3_module_sequential_0_bias = const()[name = tensor("inner_encoder_ffn1_3_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(24367488)))]; + tensor inner_encoder_ffn1_3_module_sequential_0_weight = const()[name = tensor("inner_encoder_ffn1_3_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(24368576)))]; + tensor inner_encoder_ffn1_3_module_sequential_1_linear_bias = const()[name = tensor("inner_encoder_ffn1_3_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(24369664)))]; + tensor inner_encoder_ffn1_3_module_sequential_1_linear_weight = const()[name = tensor("inner_encoder_ffn1_3_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(24373824)))]; + tensor inner_encoder_ffn1_3_module_sequential_4_linear_bias = const()[name = tensor("inner_encoder_ffn1_3_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(25422464)))]; + tensor inner_encoder_ffn1_3_module_sequential_4_linear_weight = const()[name = tensor("inner_encoder_ffn1_3_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(25423552)))]; + tensor inner_encoder_ret_lns_3_bias = const()[name = tensor("inner_encoder_ret_lns_3_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(26472192)))]; + tensor inner_encoder_ret_lns_3_weight = const()[name = tensor("inner_encoder_ret_lns_3_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(26473280)))]; + tensor inner_encoder_q_proj_3_bias = const()[name = tensor("inner_encoder_q_proj_3_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(26474368)))]; + tensor inner_encoder_q_proj_3_weight = const()[name = tensor("inner_encoder_q_proj_3_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(26475456)))]; + tensor inner_encoder_k_proj_3_bias = const()[name = tensor("inner_encoder_k_proj_3_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(26737664)))]; + tensor inner_encoder_k_proj_3_weight = const()[name = tensor("inner_encoder_k_proj_3_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(26738752)))]; + tensor inner_encoder_v_proj_3_bias = const()[name = tensor("inner_encoder_v_proj_3_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27000960)))]; + tensor inner_encoder_v_proj_3_weight = const()[name = tensor("inner_encoder_v_proj_3_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27002048)))]; + tensor inner_encoder_g_proj_3_bias = const()[name = tensor("inner_encoder_g_proj_3_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27264256)))]; + tensor inner_encoder_g_proj_3_weight = const()[name = tensor("inner_encoder_g_proj_3_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27265344)))]; + tensor inner_encoder_out_proj_3_bias = const()[name = tensor("inner_encoder_out_proj_3_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27527552)))]; + tensor inner_encoder_out_proj_3_weight = const()[name = tensor("inner_encoder_out_proj_3_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27528640)))]; + tensor inner_encoder_conv_module_3_sequential_0_bias = const()[name = tensor("inner_encoder_conv_module_3_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27790848)))]; + tensor inner_encoder_conv_module_3_sequential_0_weight = const()[name = tensor("inner_encoder_conv_module_3_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27791936)))]; + tensor inner_encoder_conv_module_3_sequential_2_conv_bias = const()[name = tensor("inner_encoder_conv_module_3_sequential_2_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27793024)))]; + tensor inner_encoder_conv_module_3_sequential_2_conv_weight = const()[name = tensor("inner_encoder_conv_module_3_sequential_2_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27795136)))]; + tensor inner_encoder_conv_module_3_sequential_4_conv_weight = const()[name = tensor("inner_encoder_conv_module_3_sequential_4_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28319488)))]; + tensor inner_encoder_conv_module_3_sequential_5_running_var = const()[name = tensor("inner_encoder_conv_module_3_sequential_5_running_var"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28335936)))]; + tensor inner_encoder_conv_module_3_sequential_5_running_mean = const()[name = tensor("inner_encoder_conv_module_3_sequential_5_running_mean"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28337024)))]; + tensor inner_encoder_conv_module_3_sequential_5_bias = const()[name = tensor("inner_encoder_conv_module_3_sequential_5_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28338112)))]; + tensor inner_encoder_conv_module_3_sequential_5_weight = const()[name = tensor("inner_encoder_conv_module_3_sequential_5_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28339200)))]; + tensor inner_encoder_conv_module_3_sequential_7_conv_bias = const()[name = tensor("inner_encoder_conv_module_3_sequential_7_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28340288)))]; + tensor inner_encoder_conv_module_3_sequential_7_conv_weight = const()[name = tensor("inner_encoder_conv_module_3_sequential_7_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28341376)))]; + tensor inner_encoder_ffn2_3_module_sequential_0_bias = const()[name = tensor("inner_encoder_ffn2_3_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28603584)))]; + tensor inner_encoder_ffn2_3_module_sequential_0_weight = const()[name = tensor("inner_encoder_ffn2_3_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28604672)))]; + tensor inner_encoder_ffn2_3_module_sequential_1_linear_bias = const()[name = tensor("inner_encoder_ffn2_3_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28605760)))]; + tensor inner_encoder_ffn2_3_module_sequential_1_linear_weight = const()[name = tensor("inner_encoder_ffn2_3_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28609920)))]; + tensor inner_encoder_ffn2_3_module_sequential_4_linear_bias = const()[name = tensor("inner_encoder_ffn2_3_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(29658560)))]; + tensor inner_encoder_ffn2_3_module_sequential_4_linear_weight = const()[name = tensor("inner_encoder_ffn2_3_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(29659648)))]; + tensor inner_encoder_layer_norm_3_bias = const()[name = tensor("inner_encoder_layer_norm_3_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(30708288)))]; + tensor inner_encoder_layer_norm_3_weight = const()[name = tensor("inner_encoder_layer_norm_3_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(30709376)))]; + tensor inner_decoder__causal_mask = const()[name = tensor("inner_decoder__causal_mask"), val = tensor([[0x1p+0, 0x0p+0, 0x0p+0], [0x1p+0, 0x1p+0, 0x0p+0], [0x1p+0, 0x1p+0, 0x1p+0]])]; + tensor inner_decoder_convert_bias = const()[name = tensor("inner_decoder_convert_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(30710464)))]; + tensor inner_decoder_convert_weight = const()[name = tensor("inner_decoder_convert_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(30711552)))]; + tensor inner_decoder_q_proj_0_bias = const()[name = tensor("inner_decoder_q_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(31235904)))]; + tensor inner_decoder_q_proj_0_weight = const()[name = tensor("inner_decoder_q_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(31236992)))]; + tensor inner_decoder_k_proj_0_bias = const()[name = tensor("inner_decoder_k_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(31499200)))]; + tensor inner_decoder_k_proj_0_weight = const()[name = tensor("inner_decoder_k_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(31500288)))]; + tensor inner_decoder_v_proj_0_bias = const()[name = tensor("inner_decoder_v_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(31762496)))]; + tensor inner_decoder_v_proj_0_weight = const()[name = tensor("inner_decoder_v_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(31763584)))]; + tensor inner_decoder_g_proj_0_bias = const()[name = tensor("inner_decoder_g_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32025792)))]; + tensor inner_decoder_g_proj_0_weight = const()[name = tensor("inner_decoder_g_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32026880)))]; + tensor inner_decoder_out_proj_0_bias = const()[name = tensor("inner_decoder_out_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32289088)))]; + tensor inner_decoder_out_proj_0_weight = const()[name = tensor("inner_decoder_out_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32290176)))]; + tensor inner_decoder_norm11_0_bias = const()[name = tensor("inner_decoder_norm11_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32552384)))]; + tensor inner_decoder_norm11_0_weight = const()[name = tensor("inner_decoder_norm11_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32553472)))]; + tensor inner_decoder_self_attn2_0_out_proj_bias = const()[name = tensor("inner_decoder_self_attn2_0_out_proj_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32554560)))]; + tensor inner_decoder_self_attn2_0_out_proj_weight = const()[name = tensor("inner_decoder_self_attn2_0_out_proj_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32555648)))]; + tensor inner_decoder_self_attn2_0_in_proj_bias = const()[name = tensor("inner_decoder_self_attn2_0_in_proj_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32817856)))]; + tensor inner_decoder_self_attn2_0_in_proj_weight = const()[name = tensor("inner_decoder_self_attn2_0_in_proj_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32820992)))]; + tensor inner_decoder_norm21_0_bias = const()[name = tensor("inner_decoder_norm21_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(33607488)))]; + tensor inner_decoder_norm21_0_weight = const()[name = tensor("inner_decoder_norm21_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(33608576)))]; + tensor inner_decoder_linear1_0_bias = const()[name = tensor("inner_decoder_linear1_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(33609664)))]; + tensor inner_decoder_linear1_0_weight = const()[name = tensor("inner_decoder_linear1_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(33617920)))]; + tensor inner_decoder_linear2_0_bias = const()[name = tensor("inner_decoder_linear2_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(35715136)))]; + tensor inner_decoder_linear2_0_weight = const()[name = tensor("inner_decoder_linear2_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(35716224)))]; + tensor inner_decoder_norm22_0_bias = const()[name = tensor("inner_decoder_norm22_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(37813440)))]; + tensor inner_decoder_norm22_0_weight = const()[name = tensor("inner_decoder_norm22_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(37814528)))]; + tensor inner_decoder_q_proj_1_bias = const()[name = tensor("inner_decoder_q_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(37815616)))]; + tensor inner_decoder_q_proj_1_weight = const()[name = tensor("inner_decoder_q_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(37816704)))]; + tensor inner_decoder_k_proj_1_bias = const()[name = tensor("inner_decoder_k_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38078912)))]; + tensor inner_decoder_k_proj_1_weight = const()[name = tensor("inner_decoder_k_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38080000)))]; + tensor inner_decoder_v_proj_1_bias = const()[name = tensor("inner_decoder_v_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38342208)))]; + tensor inner_decoder_v_proj_1_weight = const()[name = tensor("inner_decoder_v_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38343296)))]; + tensor inner_decoder_g_proj_1_bias = const()[name = tensor("inner_decoder_g_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38605504)))]; + tensor inner_decoder_g_proj_1_weight = const()[name = tensor("inner_decoder_g_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38606592)))]; + tensor inner_decoder_out_proj_1_bias = const()[name = tensor("inner_decoder_out_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38868800)))]; + tensor inner_decoder_out_proj_1_weight = const()[name = tensor("inner_decoder_out_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38869888)))]; + tensor inner_decoder_norm11_1_bias = const()[name = tensor("inner_decoder_norm11_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(39132096)))]; + tensor inner_decoder_norm11_1_weight = const()[name = tensor("inner_decoder_norm11_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(39133184)))]; + tensor inner_decoder_self_attn2_1_out_proj_bias = const()[name = tensor("inner_decoder_self_attn2_1_out_proj_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(39134272)))]; + tensor inner_decoder_self_attn2_1_out_proj_weight = const()[name = tensor("inner_decoder_self_attn2_1_out_proj_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(39135360)))]; + tensor inner_decoder_self_attn2_1_in_proj_bias = const()[name = tensor("inner_decoder_self_attn2_1_in_proj_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(39397568)))]; + tensor inner_decoder_self_attn2_1_in_proj_weight = const()[name = tensor("inner_decoder_self_attn2_1_in_proj_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(39400704)))]; + tensor inner_decoder_norm21_1_bias = const()[name = tensor("inner_decoder_norm21_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(40187200)))]; + tensor inner_decoder_norm21_1_weight = const()[name = tensor("inner_decoder_norm21_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(40188288)))]; + tensor inner_decoder_linear1_1_bias = const()[name = tensor("inner_decoder_linear1_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(40189376)))]; + tensor inner_decoder_linear1_1_weight = const()[name = tensor("inner_decoder_linear1_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(40197632)))]; + tensor inner_decoder_linear2_1_bias = const()[name = tensor("inner_decoder_linear2_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(42294848)))]; + tensor inner_decoder_linear2_1_weight = const()[name = tensor("inner_decoder_linear2_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(42295936)))]; + tensor inner_decoder_norm22_1_bias = const()[name = tensor("inner_decoder_norm22_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(44393152)))]; + tensor inner_decoder_norm22_1_weight = const()[name = tensor("inner_decoder_norm22_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(44394240)))]; + tensor var_19_begin_0 = const()[name = tensor("op_19_begin_0"), val = tensor([0, 0, 0])]; + tensor var_19_end_0 = const()[name = tensor("op_19_end_0"), val = tensor([1, 15, 23])]; + tensor var_19_end_mask_0 = const()[name = tensor("op_19_end_mask_0"), val = tensor([true, false, true])]; + tensor var_19 = slice_by_index(begin = var_19_begin_0, end = var_19_end_0, end_mask = var_19_end_mask_0, x = features)[name = tensor("op_19")]; + tensor var_29_begin_0 = const()[name = tensor("op_29_begin_0"), val = tensor([0, 10, 0])]; + tensor var_29_end_0 = const()[name = tensor("op_29_end_0"), val = tensor([1, 25, 23])]; + tensor var_29_end_mask_0 = const()[name = tensor("op_29_end_mask_0"), val = tensor([true, false, true])]; + tensor var_29 = slice_by_index(begin = var_29_begin_0, end = var_29_end_0, end_mask = var_29_end_mask_0, x = features)[name = tensor("op_29")]; + tensor var_39_begin_0 = const()[name = tensor("op_39_begin_0"), val = tensor([0, 20, 0])]; + tensor var_39_end_0 = const()[name = tensor("op_39_end_0"), val = tensor([1, 1, 23])]; + tensor var_39_end_mask_0 = const()[name = tensor("op_39_end_mask_0"), val = tensor([true, true, true])]; + tensor var_39 = slice_by_index(begin = var_39_begin_0, end = var_39_end_0, end_mask = var_39_end_mask_0, x = features)[name = tensor("op_39")]; + tensor stacked_axis_0 = const()[name = tensor("stacked_axis_0"), val = tensor(1)]; + tensor stacked = stack(axis = stacked_axis_0, values = (var_19, var_29, var_39))[name = tensor("stacked")]; + tensor var_46 = const()[name = tensor("op_46"), val = tensor([1, 3, 345])]; + tensor input_1 = reshape(shape = var_46, x = stacked)[name = tensor("input_1")]; + tensor var_49 = const()[name = tensor("op_49"), val = tensor(0x1p+0)]; + tensor var_55 = const()[name = tensor("op_55"), val = tensor(true)]; + tensor var_56 = const()[name = tensor("op_56"), val = tensor(0x1.4f8b58p-17)]; + tensor var_59 = const()[name = tensor("op_59"), val = tensor(0)]; + tensor var_61 = const()[name = tensor("op_61"), val = tensor(2)]; + tensor var_62 = const()[name = tensor("op_62"), val = tensor(-1)]; + tensor var_64 = const()[name = tensor("op_64"), val = tensor(0x1.0c6f7ap-20)]; + tensor var_69 = const()[name = tensor("op_69"), val = tensor(0x1.5798eep-27)]; + tensor var_72 = const()[name = tensor("op_72"), val = tensor(1)]; + tensor input_3 = linear(bias = inner_encoder_input_projection_linear_bias, weight = inner_encoder_input_projection_linear_weight, x = input_1)[name = tensor("linear_0")]; tensor input_5_axes_0 = const()[name = tensor("input_5_axes_0"), val = tensor([-1])]; - tensor input_5 = layer_norm(axes = input_5_axes_0, beta = encoder_ffn1_0_module_sequential_0_bias, epsilon = var_28, gamma = encoder_ffn1_0_module_sequential_0_weight, x = input_3)[name = tensor("input_5")]; - tensor inputs_1 = linear(bias = encoder_ffn1_0_module_sequential_1_linear_bias, weight = encoder_ffn1_0_module_sequential_1_linear_weight, x = input_5)[name = tensor("linear_1")]; - tensor input_7 = silu(x = inputs_1)[name = tensor("input_7")]; - tensor input_11 = linear(bias = encoder_ffn1_0_module_sequential_4_linear_bias, weight = encoder_ffn1_0_module_sequential_4_linear_weight, x = input_7)[name = tensor("linear_2")]; - tensor var_147 = const()[name = tensor("op_147"), val = tensor(0x1p-1)]; - tensor var_148 = mul(x = input_11, y = var_147)[name = tensor("op_148")]; - tensor input_13 = add(x = var_148, y = input_3)[name = tensor("input_13")]; + tensor input_5 = layer_norm(axes = input_5_axes_0, beta = inner_encoder_pre_layer_norm_bias, epsilon = var_56, gamma = inner_encoder_pre_layer_norm_weight, x = input_3)[name = tensor("input_5")]; + tensor input_7_axes_0 = const()[name = tensor("input_7_axes_0"), val = tensor([-1])]; + tensor input_7 = layer_norm(axes = input_7_axes_0, beta = inner_encoder_ffn1_0_module_sequential_0_bias, epsilon = var_56, gamma = inner_encoder_ffn1_0_module_sequential_0_weight, x = input_5)[name = tensor("input_7")]; + tensor inputs_1 = linear(bias = inner_encoder_ffn1_0_module_sequential_1_linear_bias, weight = inner_encoder_ffn1_0_module_sequential_1_linear_weight, x = input_7)[name = tensor("linear_1")]; + tensor input_9 = silu(x = inputs_1)[name = tensor("input_9")]; + tensor input_13 = linear(bias = inner_encoder_ffn1_0_module_sequential_4_linear_bias, weight = inner_encoder_ffn1_0_module_sequential_4_linear_weight, x = input_9)[name = tensor("linear_2")]; + tensor var_193 = const()[name = tensor("op_193"), val = tensor(0x1p-1)]; + tensor var_194 = mul(x = input_13, y = var_193)[name = tensor("op_194")]; + tensor input_15 = add(x = var_194, y = input_5)[name = tensor("input_15")]; tensor x_1_axes_0 = const()[name = tensor("x_1_axes_0"), val = tensor([-1])]; - tensor x_1 = layer_norm(axes = x_1_axes_0, beta = encoder_ret_lns_0_bias, epsilon = var_28, gamma = encoder_ret_lns_0_weight, x = input_13)[name = tensor("x_1")]; + tensor x_1 = layer_norm(axes = x_1_axes_0, beta = inner_encoder_ret_lns_0_bias, epsilon = var_56, gamma = inner_encoder_ret_lns_0_weight, x = input_15)[name = tensor("x_1")]; tensor prev_kv_1_begin_0 = const()[name = tensor("prev_kv_1_begin_0"), val = tensor([0, 0, 0, 0, 0])]; tensor prev_kv_1_end_0 = const()[name = tensor("prev_kv_1_end_0"), val = tensor([1, 1, 4, 64, 64])]; tensor prev_kv_1_end_mask_0 = const()[name = tensor("prev_kv_1_end_mask_0"), val = tensor([false, true, true, true, true])]; @@ -239,163 +257,163 @@ program(1.0) tensor prev_scale_1_end_mask_0 = const()[name = tensor("prev_scale_1_end_mask_0"), val = tensor([false, true])]; tensor prev_scale_1_squeeze_mask_0 = const()[name = tensor("prev_scale_1_squeeze_mask_0"), val = tensor([true, false])]; tensor prev_scale_1 = slice_by_index(begin = prev_scale_1_begin_0, end = prev_scale_1_end_0, end_mask = prev_scale_1_end_mask_0, squeeze_mask = prev_scale_1_squeeze_mask_0, x = enc_scale)[name = tensor("prev_scale_1")]; - tensor var_162 = linear(bias = encoder_q_proj_0_bias, weight = encoder_q_proj_0_weight, x = x_1)[name = tensor("linear_3")]; - tensor var_163 = const()[name = tensor("op_163"), val = tensor([1, 3, 4, 64])]; - tensor var_164 = reshape(shape = var_163, x = var_162)[name = tensor("op_164")]; + tensor var_208 = linear(bias = inner_encoder_q_proj_0_bias, weight = inner_encoder_q_proj_0_weight, x = x_1)[name = tensor("linear_3")]; + tensor var_209 = const()[name = tensor("op_209"), val = tensor([1, 3, 4, 64])]; + tensor var_210 = reshape(shape = var_209, x = var_208)[name = tensor("op_210")]; tensor q_1_perm_0 = const()[name = tensor("q_1_perm_0"), val = tensor([0, 2, 1, 3])]; - tensor var_168 = linear(bias = encoder_k_proj_0_bias, weight = encoder_k_proj_0_weight, x = x_1)[name = tensor("linear_4")]; - tensor var_169 = const()[name = tensor("op_169"), val = tensor(0x1p-3)]; - tensor var_170 = mul(x = var_168, y = var_169)[name = tensor("op_170")]; - tensor var_171 = const()[name = tensor("op_171"), val = tensor([1, 3, 4, 64])]; - tensor var_172 = reshape(shape = var_171, x = var_170)[name = tensor("op_172")]; + tensor var_214 = linear(bias = inner_encoder_k_proj_0_bias, weight = inner_encoder_k_proj_0_weight, x = x_1)[name = tensor("linear_4")]; + tensor var_215 = const()[name = tensor("op_215"), val = tensor(0x1p-3)]; + tensor var_216 = mul(x = var_214, y = var_215)[name = tensor("op_216")]; + tensor var_217 = const()[name = tensor("op_217"), val = tensor([1, 3, 4, 64])]; + tensor var_218 = reshape(shape = var_217, x = var_216)[name = tensor("op_218")]; tensor k_1_perm_0 = const()[name = tensor("k_1_perm_0"), val = tensor([0, 2, 1, 3])]; - tensor var_176 = linear(bias = encoder_v_proj_0_bias, weight = encoder_v_proj_0_weight, x = x_1)[name = tensor("linear_5")]; - tensor var_177 = const()[name = tensor("op_177"), val = tensor([1, 3, 4, 64])]; - tensor var_178 = reshape(shape = var_177, x = var_176)[name = tensor("op_178")]; + tensor var_222 = linear(bias = inner_encoder_v_proj_0_bias, weight = inner_encoder_v_proj_0_weight, x = x_1)[name = tensor("linear_5")]; + tensor var_223 = const()[name = tensor("op_223"), val = tensor([1, 3, 4, 64])]; + tensor var_224 = reshape(shape = var_223, x = var_222)[name = tensor("op_224")]; tensor v_1_perm_0 = const()[name = tensor("v_1_perm_0"), val = tensor([0, 2, 1, 3])]; - tensor input_17 = linear(bias = encoder_g_proj_0_bias, weight = encoder_g_proj_0_weight, x = x_1)[name = tensor("linear_6")]; + tensor input_19 = linear(bias = inner_encoder_g_proj_0_bias, weight = inner_encoder_g_proj_0_weight, x = x_1)[name = tensor("linear_6")]; tensor sqrt_s0_1 = sqrt(x = prev_scale_1)[name = tensor("sqrt_s0_1")]; - tensor s_t_1 = add(x = prev_scale_1, y = encoder__t_index)[name = tensor("s_t_1")]; + tensor s_t_1 = add(x = prev_scale_1, y = inner_encoder__t_index)[name = tensor("s_t_1")]; tensor sqrt_s_t_1 = sqrt(x = s_t_1)[name = tensor("sqrt_s_t_1")]; tensor qk_1_transpose_x_1 = const()[name = tensor("qk_1_transpose_x_1"), val = tensor(false)]; tensor qk_1_transpose_y_1 = const()[name = tensor("qk_1_transpose_y_1"), val = tensor(true)]; - tensor k_1 = transpose(perm = k_1_perm_0, x = var_172)[name = tensor("transpose_57")]; - tensor q_1 = transpose(perm = q_1_perm_0, x = var_164)[name = tensor("transpose_58")]; + tensor k_1 = transpose(perm = k_1_perm_0, x = var_218)[name = tensor("transpose_57")]; + tensor q_1 = transpose(perm = q_1_perm_0, x = var_210)[name = tensor("transpose_58")]; tensor qk_1 = matmul(transpose_x = qk_1_transpose_x_1, transpose_y = qk_1_transpose_y_1, x = q_1, y = k_1)[name = tensor("qk_1")]; - tensor var_188 = const()[name = tensor("op_188"), val = tensor([3, 1])]; - tensor var_189 = reshape(shape = var_188, x = sqrt_s_t_1)[name = tensor("op_189")]; - tensor M_1 = real_div(x = encoder__causal_mask, y = var_189)[name = tensor("M_1")]; - tensor var_191 = mul(x = qk_1, y = M_1)[name = tensor("op_191")]; + tensor var_234 = const()[name = tensor("op_234"), val = tensor([3, 1])]; + tensor var_235 = reshape(shape = var_234, x = sqrt_s_t_1)[name = tensor("op_235")]; + tensor M_1 = real_div(x = inner_encoder__causal_mask, y = var_235)[name = tensor("M_1")]; + tensor var_237 = mul(x = qk_1, y = M_1)[name = tensor("op_237")]; tensor inner_1_transpose_x_0 = const()[name = tensor("inner_1_transpose_x_0"), val = tensor(false)]; tensor inner_1_transpose_y_0 = const()[name = tensor("inner_1_transpose_y_0"), val = tensor(false)]; - tensor v_1 = transpose(perm = v_1_perm_0, x = var_178)[name = tensor("transpose_56")]; - tensor inner_1 = matmul(transpose_x = inner_1_transpose_x_0, transpose_y = inner_1_transpose_y_0, x = var_191, y = v_1)[name = tensor("inner_1")]; - tensor var_193_transpose_x_0 = const()[name = tensor("op_193_transpose_x_0"), val = tensor(false)]; - tensor var_193_transpose_y_0 = const()[name = tensor("op_193_transpose_y_0"), val = tensor(false)]; - tensor var_193 = matmul(transpose_x = var_193_transpose_x_0, transpose_y = var_193_transpose_y_0, x = q_1, y = prev_kv_1)[name = tensor("op_193")]; - tensor var_194 = real_div(x = sqrt_s0_1, y = sqrt_s_t_1)[name = tensor("op_194")]; - tensor var_195 = const()[name = tensor("op_195"), val = tensor([1, 1, 3, 1])]; - tensor var_196 = reshape(shape = var_195, x = var_194)[name = tensor("op_196")]; - tensor cross_1 = mul(x = var_193, y = var_196)[name = tensor("cross_1")]; + tensor v_1 = transpose(perm = v_1_perm_0, x = var_224)[name = tensor("transpose_56")]; + tensor inner_1 = matmul(transpose_x = inner_1_transpose_x_0, transpose_y = inner_1_transpose_y_0, x = var_237, y = v_1)[name = tensor("inner_1")]; + tensor var_239_transpose_x_0 = const()[name = tensor("op_239_transpose_x_0"), val = tensor(false)]; + tensor var_239_transpose_y_0 = const()[name = tensor("op_239_transpose_y_0"), val = tensor(false)]; + tensor var_239 = matmul(transpose_x = var_239_transpose_x_0, transpose_y = var_239_transpose_y_0, x = q_1, y = prev_kv_1)[name = tensor("op_239")]; + tensor var_240 = real_div(x = sqrt_s0_1, y = sqrt_s_t_1)[name = tensor("op_240")]; + tensor var_241 = const()[name = tensor("op_241"), val = tensor([1, 1, 3, 1])]; + tensor var_242 = reshape(shape = var_241, x = var_240)[name = tensor("op_242")]; + tensor cross_1 = mul(x = var_239, y = var_242)[name = tensor("cross_1")]; tensor out_1 = add(x = inner_1, y = cross_1)[name = tensor("out_1")]; - tensor var_199 = mul(x = prev_kv_1, y = sqrt_s0_1)[name = tensor("op_199")]; - tensor var_201_transpose_x_1 = const()[name = tensor("op_201_transpose_x_1"), val = tensor(true)]; - tensor var_201_transpose_y_1 = const()[name = tensor("op_201_transpose_y_1"), val = tensor(false)]; - tensor var_201 = matmul(transpose_x = var_201_transpose_x_1, transpose_y = var_201_transpose_y_1, x = k_1, y = v_1)[name = tensor("op_201")]; - tensor new_kv_unnorm_1 = add(x = var_199, y = var_201)[name = tensor("new_kv_unnorm_1")]; - tensor var_203 = const()[name = tensor("op_203"), val = tensor(0x1.8p+1)]; - tensor new_scale_1 = add(x = prev_scale_1, y = var_203)[name = tensor("new_scale_1")]; - tensor var_205 = sqrt(x = new_scale_1)[name = tensor("op_205")]; - tensor var_206 = real_div(x = new_kv_unnorm_1, y = var_205)[name = tensor("op_206")]; - tensor var_207_perm_0 = const()[name = tensor("op_207_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_245 = mul(x = prev_kv_1, y = sqrt_s0_1)[name = tensor("op_245")]; + tensor var_247_transpose_x_1 = const()[name = tensor("op_247_transpose_x_1"), val = tensor(true)]; + tensor var_247_transpose_y_1 = const()[name = tensor("op_247_transpose_y_1"), val = tensor(false)]; + tensor var_247 = matmul(transpose_x = var_247_transpose_x_1, transpose_y = var_247_transpose_y_1, x = k_1, y = v_1)[name = tensor("op_247")]; + tensor new_kv_unnorm_1 = add(x = var_245, y = var_247)[name = tensor("new_kv_unnorm_1")]; + tensor var_249 = const()[name = tensor("op_249"), val = tensor(0x1.8p+1)]; + tensor new_scale_1 = add(x = prev_scale_1, y = var_249)[name = tensor("new_scale_1")]; + tensor var_251 = sqrt(x = new_scale_1)[name = tensor("op_251")]; + tensor var_252 = real_div(x = new_kv_unnorm_1, y = var_251)[name = tensor("op_252")]; + tensor var_253_perm_0 = const()[name = tensor("op_253_perm_0"), val = tensor([0, 2, 1, 3])]; tensor out_3_axes_0 = const()[name = tensor("out_3_axes_0"), val = tensor([-1])]; - tensor var_207 = transpose(perm = var_207_perm_0, x = out_1)[name = tensor("transpose_55")]; - tensor out_3 = layer_norm(axes = out_3_axes_0, epsilon = var_17, x = var_207)[name = tensor("out_3")]; - tensor var_211 = const()[name = tensor("op_211"), val = tensor([1, 3, 256])]; - tensor out_5 = reshape(shape = var_211, x = out_3)[name = tensor("out_5")]; - tensor var_213 = silu(x = input_17)[name = tensor("op_213")]; - tensor input_19 = mul(x = var_213, y = out_5)[name = tensor("input_19")]; - tensor ret_out_1 = linear(bias = encoder_out_proj_0_bias, weight = encoder_out_proj_0_weight, x = input_19)[name = tensor("linear_7")]; - tensor x_3 = add(x = input_13, y = ret_out_1)[name = tensor("x_3")]; + tensor var_253 = transpose(perm = var_253_perm_0, x = out_1)[name = tensor("transpose_55")]; + tensor out_3 = layer_norm(axes = out_3_axes_0, epsilon = var_64, x = var_253)[name = tensor("out_3")]; + tensor var_257 = const()[name = tensor("op_257"), val = tensor([1, 3, 256])]; + tensor out_5 = reshape(shape = var_257, x = out_3)[name = tensor("out_5")]; + tensor var_259 = silu(x = input_19)[name = tensor("op_259")]; + tensor input_21 = mul(x = var_259, y = out_5)[name = tensor("input_21")]; + tensor ret_out_1 = linear(bias = inner_encoder_out_proj_0_bias, weight = inner_encoder_out_proj_0_weight, x = input_21)[name = tensor("linear_7")]; + tensor x_3 = add(x = input_15, y = ret_out_1)[name = tensor("x_3")]; tensor window_1_begin_0 = const()[name = tensor("window_1_begin_0"), val = tensor([0, 0, 0, 0])]; tensor window_1_end_0 = const()[name = tensor("window_1_end_0"), val = tensor([1, 1, 16, 256])]; tensor window_1_end_mask_0 = const()[name = tensor("window_1_end_mask_0"), val = tensor([false, true, true, true])]; tensor window_1_squeeze_mask_0 = const()[name = tensor("window_1_squeeze_mask_0"), val = tensor([true, false, false, false])]; tensor window_1 = slice_by_index(begin = window_1_begin_0, end = window_1_end_0, end_mask = window_1_end_mask_0, squeeze_mask = window_1_squeeze_mask_0, x = enc_conv_cache)[name = tensor("window_1")]; - tensor var_221_begin_0 = const()[name = tensor("op_221_begin_0"), val = tensor([0, 0, 0])]; - tensor var_221_end_0 = const()[name = tensor("op_221_end_0"), val = tensor([1, 1, 256])]; - tensor var_221_end_mask_0 = const()[name = tensor("op_221_end_mask_0"), val = tensor([true, false, true])]; - tensor var_221 = slice_by_index(begin = var_221_begin_0, end = var_221_end_0, end_mask = var_221_end_mask_0, x = x_3)[name = tensor("op_221")]; - tensor var_224_begin_0 = const()[name = tensor("op_224_begin_0"), val = tensor([0, 1, 0])]; - tensor var_224_end_0 = const()[name = tensor("op_224_end_0"), val = tensor([1, 16, 256])]; - tensor var_224_end_mask_0 = const()[name = tensor("op_224_end_mask_0"), val = tensor([true, true, true])]; - tensor var_224 = slice_by_index(begin = var_224_begin_0, end = var_224_end_0, end_mask = var_224_end_mask_0, x = window_1)[name = tensor("op_224")]; + tensor var_267_begin_0 = const()[name = tensor("op_267_begin_0"), val = tensor([0, 0, 0])]; + tensor var_267_end_0 = const()[name = tensor("op_267_end_0"), val = tensor([1, 1, 256])]; + tensor var_267_end_mask_0 = const()[name = tensor("op_267_end_mask_0"), val = tensor([true, false, true])]; + tensor var_267 = slice_by_index(begin = var_267_begin_0, end = var_267_end_0, end_mask = var_267_end_mask_0, x = x_3)[name = tensor("op_267")]; + tensor var_270_begin_0 = const()[name = tensor("op_270_begin_0"), val = tensor([0, 1, 0])]; + tensor var_270_end_0 = const()[name = tensor("op_270_end_0"), val = tensor([1, 16, 256])]; + tensor var_270_end_mask_0 = const()[name = tensor("op_270_end_mask_0"), val = tensor([true, true, true])]; + tensor var_270 = slice_by_index(begin = var_270_begin_0, end = var_270_end_0, end_mask = var_270_end_mask_0, x = window_1)[name = tensor("op_270")]; tensor window_3_interleave_0 = const()[name = tensor("window_3_interleave_0"), val = tensor(false)]; - tensor window_3 = concat(axis = var_26, interleave = window_3_interleave_0, values = (var_224, var_221))[name = tensor("window_3")]; - tensor var_229_begin_0 = const()[name = tensor("op_229_begin_0"), val = tensor([0, 1, 0])]; - tensor var_229_end_0 = const()[name = tensor("op_229_end_0"), val = tensor([1, 2, 256])]; - tensor var_229_end_mask_0 = const()[name = tensor("op_229_end_mask_0"), val = tensor([true, false, true])]; - tensor var_229 = slice_by_index(begin = var_229_begin_0, end = var_229_end_0, end_mask = var_229_end_mask_0, x = x_3)[name = tensor("op_229")]; - tensor var_232_begin_0 = const()[name = tensor("op_232_begin_0"), val = tensor([0, 1, 0])]; - tensor var_232_end_0 = const()[name = tensor("op_232_end_0"), val = tensor([1, 16, 256])]; - tensor var_232_end_mask_0 = const()[name = tensor("op_232_end_mask_0"), val = tensor([true, true, true])]; - tensor var_232 = slice_by_index(begin = var_232_begin_0, end = var_232_end_0, end_mask = var_232_end_mask_0, x = window_3)[name = tensor("op_232")]; + tensor window_3 = concat(axis = var_72, interleave = window_3_interleave_0, values = (var_270, var_267))[name = tensor("window_3")]; + tensor var_275_begin_0 = const()[name = tensor("op_275_begin_0"), val = tensor([0, 1, 0])]; + tensor var_275_end_0 = const()[name = tensor("op_275_end_0"), val = tensor([1, 2, 256])]; + tensor var_275_end_mask_0 = const()[name = tensor("op_275_end_mask_0"), val = tensor([true, false, true])]; + tensor var_275 = slice_by_index(begin = var_275_begin_0, end = var_275_end_0, end_mask = var_275_end_mask_0, x = x_3)[name = tensor("op_275")]; + tensor var_278_begin_0 = const()[name = tensor("op_278_begin_0"), val = tensor([0, 1, 0])]; + tensor var_278_end_0 = const()[name = tensor("op_278_end_0"), val = tensor([1, 16, 256])]; + tensor var_278_end_mask_0 = const()[name = tensor("op_278_end_mask_0"), val = tensor([true, true, true])]; + tensor var_278 = slice_by_index(begin = var_278_begin_0, end = var_278_end_0, end_mask = var_278_end_mask_0, x = window_3)[name = tensor("op_278")]; tensor window_5_interleave_0 = const()[name = tensor("window_5_interleave_0"), val = tensor(false)]; - tensor window_5 = concat(axis = var_26, interleave = window_5_interleave_0, values = (var_232, var_229))[name = tensor("window_5")]; - tensor var_237_begin_0 = const()[name = tensor("op_237_begin_0"), val = tensor([0, 2, 0])]; - tensor var_237_end_0 = const()[name = tensor("op_237_end_0"), val = tensor([1, 1, 256])]; - tensor var_237_end_mask_0 = const()[name = tensor("op_237_end_mask_0"), val = tensor([true, true, true])]; - tensor var_237 = slice_by_index(begin = var_237_begin_0, end = var_237_end_0, end_mask = var_237_end_mask_0, x = x_3)[name = tensor("op_237")]; - tensor var_240_begin_0 = const()[name = tensor("op_240_begin_0"), val = tensor([0, 1, 0])]; - tensor var_240_end_0 = const()[name = tensor("op_240_end_0"), val = tensor([1, 16, 256])]; - tensor var_240_end_mask_0 = const()[name = tensor("op_240_end_mask_0"), val = tensor([true, true, true])]; - tensor var_240 = slice_by_index(begin = var_240_begin_0, end = var_240_end_0, end_mask = var_240_end_mask_0, x = window_5)[name = tensor("op_240")]; + tensor window_5 = concat(axis = var_72, interleave = window_5_interleave_0, values = (var_278, var_275))[name = tensor("window_5")]; + tensor var_283_begin_0 = const()[name = tensor("op_283_begin_0"), val = tensor([0, 2, 0])]; + tensor var_283_end_0 = const()[name = tensor("op_283_end_0"), val = tensor([1, 1, 256])]; + tensor var_283_end_mask_0 = const()[name = tensor("op_283_end_mask_0"), val = tensor([true, true, true])]; + tensor var_283 = slice_by_index(begin = var_283_begin_0, end = var_283_end_0, end_mask = var_283_end_mask_0, x = x_3)[name = tensor("op_283")]; + tensor var_286_begin_0 = const()[name = tensor("op_286_begin_0"), val = tensor([0, 1, 0])]; + tensor var_286_end_0 = const()[name = tensor("op_286_end_0"), val = tensor([1, 16, 256])]; + tensor var_286_end_mask_0 = const()[name = tensor("op_286_end_mask_0"), val = tensor([true, true, true])]; + tensor var_286 = slice_by_index(begin = var_286_begin_0, end = var_286_end_0, end_mask = var_286_end_mask_0, x = window_5)[name = tensor("op_286")]; tensor window_7_interleave_0 = const()[name = tensor("window_7_interleave_0"), val = tensor(false)]; - tensor window_7 = concat(axis = var_26, interleave = window_7_interleave_0, values = (var_240, var_237))[name = tensor("window_7")]; - tensor input_21_interleave_0 = const()[name = tensor("input_21_interleave_0"), val = tensor(false)]; - tensor input_21 = concat(axis = var_23, interleave = input_21_interleave_0, values = (window_3, window_5, window_7))[name = tensor("input_21")]; + tensor window_7 = concat(axis = var_72, interleave = window_7_interleave_0, values = (var_286, var_283))[name = tensor("window_7")]; + tensor input_23_interleave_0 = const()[name = tensor("input_23_interleave_0"), val = tensor(false)]; + tensor input_23 = concat(axis = var_59, interleave = input_23_interleave_0, values = (window_3, window_5, window_7))[name = tensor("input_23")]; tensor x_5_axes_0 = const()[name = tensor("x_5_axes_0"), val = tensor([-1])]; - tensor x_5 = layer_norm(axes = x_5_axes_0, beta = encoder_conv_module_0_sequential_0_bias, epsilon = var_28, gamma = encoder_conv_module_0_sequential_0_weight, x = input_21)[name = tensor("x_5")]; - tensor input_23_perm_0 = const()[name = tensor("input_23_perm_0"), val = tensor([0, 2, 1])]; + tensor x_5 = layer_norm(axes = x_5_axes_0, beta = inner_encoder_conv_module_0_sequential_0_bias, epsilon = var_56, gamma = inner_encoder_conv_module_0_sequential_0_weight, x = input_23)[name = tensor("x_5")]; + tensor input_25_perm_0 = const()[name = tensor("input_25_perm_0"), val = tensor([0, 2, 1])]; tensor inputs_3_pad_type_0 = const()[name = tensor("inputs_3_pad_type_0"), val = tensor("valid")]; tensor inputs_3_strides_0 = const()[name = tensor("inputs_3_strides_0"), val = tensor([1])]; tensor inputs_3_pad_0 = const()[name = tensor("inputs_3_pad_0"), val = tensor([0, 0])]; tensor inputs_3_dilations_0 = const()[name = tensor("inputs_3_dilations_0"), val = tensor([1])]; tensor inputs_3_groups_0 = const()[name = tensor("inputs_3_groups_0"), val = tensor(1)]; - tensor input_23 = transpose(perm = input_23_perm_0, x = x_5)[name = tensor("transpose_54")]; - tensor inputs_3 = conv(bias = encoder_conv_module_0_sequential_2_conv_bias, dilations = inputs_3_dilations_0, groups = inputs_3_groups_0, pad = inputs_3_pad_0, pad_type = inputs_3_pad_type_0, strides = inputs_3_strides_0, weight = encoder_conv_module_0_sequential_2_conv_weight, x = input_23)[name = tensor("inputs_3")]; - tensor var_265_split_sizes_0 = const()[name = tensor("op_265_split_sizes_0"), val = tensor([256, 256])]; - tensor var_265_axis_0 = const()[name = tensor("op_265_axis_0"), val = tensor(1)]; - tensor var_265_0, tensor var_265_1 = split(axis = var_265_axis_0, split_sizes = var_265_split_sizes_0, x = inputs_3)[name = tensor("op_265")]; - tensor var_267 = sigmoid(x = var_265_1)[name = tensor("op_267")]; - tensor inputs_5 = mul(x = var_265_0, y = var_267)[name = tensor("inputs_5")]; + tensor input_25 = transpose(perm = input_25_perm_0, x = x_5)[name = tensor("transpose_54")]; + tensor inputs_3 = conv(bias = inner_encoder_conv_module_0_sequential_2_conv_bias, dilations = inputs_3_dilations_0, groups = inputs_3_groups_0, pad = inputs_3_pad_0, pad_type = inputs_3_pad_type_0, strides = inputs_3_strides_0, weight = inner_encoder_conv_module_0_sequential_2_conv_weight, x = input_25)[name = tensor("inputs_3")]; + tensor var_311_split_sizes_0 = const()[name = tensor("op_311_split_sizes_0"), val = tensor([256, 256])]; + tensor var_311_axis_0 = const()[name = tensor("op_311_axis_0"), val = tensor(1)]; + tensor var_311_0, tensor var_311_1 = split(axis = var_311_axis_0, split_sizes = var_311_split_sizes_0, x = inputs_3)[name = tensor("op_311")]; + tensor var_313 = sigmoid(x = var_311_1)[name = tensor("op_313")]; + tensor inputs_5 = mul(x = var_311_0, y = var_313)[name = tensor("inputs_5")]; tensor outputs_aug_1_pad_type_0 = const()[name = tensor("outputs_aug_1_pad_type_0"), val = tensor("custom")]; tensor outputs_aug_1_pad_0 = const()[name = tensor("outputs_aug_1_pad_0"), val = tensor([15, 15])]; tensor outputs_aug_1_groups_0 = const()[name = tensor("outputs_aug_1_groups_0"), val = tensor(256)]; tensor outputs_aug_1_strides_0 = const()[name = tensor("outputs_aug_1_strides_0"), val = tensor([1])]; tensor outputs_aug_1_dilations_0 = const()[name = tensor("outputs_aug_1_dilations_0"), val = tensor([1])]; - tensor outputs_aug_1 = conv(dilations = outputs_aug_1_dilations_0, groups = outputs_aug_1_groups_0, pad = outputs_aug_1_pad_0, pad_type = outputs_aug_1_pad_type_0, strides = outputs_aug_1_strides_0, weight = encoder_conv_module_0_sequential_4_conv_weight, x = inputs_5)[name = tensor("outputs_aug_1")]; - tensor input_25_begin_0 = const()[name = tensor("input_25_begin_0"), val = tensor([0, 0, 0])]; - tensor input_25_end_0 = const()[name = tensor("input_25_end_0"), val = tensor([3, 256, 16])]; - tensor input_25_end_mask_0 = const()[name = tensor("input_25_end_mask_0"), val = tensor([true, true, false])]; - tensor input_25 = slice_by_index(begin = input_25_begin_0, end = input_25_end_0, end_mask = input_25_end_mask_0, x = outputs_aug_1)[name = tensor("input_25")]; - tensor inputs_7 = batch_norm(beta = encoder_conv_module_0_sequential_5_bias, epsilon = var_28, gamma = encoder_conv_module_0_sequential_5_weight, mean = encoder_conv_module_0_sequential_5_running_mean, variance = encoder_conv_module_0_sequential_5_running_var, x = input_25)[name = tensor("inputs_7")]; - tensor input_27 = silu(x = inputs_7)[name = tensor("input_27")]; - tensor input_29_pad_type_0 = const()[name = tensor("input_29_pad_type_0"), val = tensor("valid")]; - tensor input_29_strides_0 = const()[name = tensor("input_29_strides_0"), val = tensor([1])]; - tensor input_29_pad_0 = const()[name = tensor("input_29_pad_0"), val = tensor([0, 0])]; - tensor input_29_dilations_0 = const()[name = tensor("input_29_dilations_0"), val = tensor([1])]; - tensor input_29_groups_0 = const()[name = tensor("input_29_groups_0"), val = tensor(1)]; - tensor input_29 = conv(bias = encoder_conv_module_0_sequential_7_conv_bias, dilations = input_29_dilations_0, groups = input_29_groups_0, pad = input_29_pad_0, pad_type = input_29_pad_type_0, strides = input_29_strides_0, weight = encoder_conv_module_0_sequential_7_conv_weight, x = input_27)[name = tensor("input_29")]; + tensor outputs_aug_1 = conv(dilations = outputs_aug_1_dilations_0, groups = outputs_aug_1_groups_0, pad = outputs_aug_1_pad_0, pad_type = outputs_aug_1_pad_type_0, strides = outputs_aug_1_strides_0, weight = inner_encoder_conv_module_0_sequential_4_conv_weight, x = inputs_5)[name = tensor("outputs_aug_1")]; + tensor input_27_begin_0 = const()[name = tensor("input_27_begin_0"), val = tensor([0, 0, 0])]; + tensor input_27_end_0 = const()[name = tensor("input_27_end_0"), val = tensor([3, 256, 16])]; + tensor input_27_end_mask_0 = const()[name = tensor("input_27_end_mask_0"), val = tensor([true, true, false])]; + tensor input_27 = slice_by_index(begin = input_27_begin_0, end = input_27_end_0, end_mask = input_27_end_mask_0, x = outputs_aug_1)[name = tensor("input_27")]; + tensor inputs_7 = batch_norm(beta = inner_encoder_conv_module_0_sequential_5_bias, epsilon = var_56, gamma = inner_encoder_conv_module_0_sequential_5_weight, mean = inner_encoder_conv_module_0_sequential_5_running_mean, variance = inner_encoder_conv_module_0_sequential_5_running_var, x = input_27)[name = tensor("inputs_7")]; + tensor input_29 = silu(x = inputs_7)[name = tensor("input_29")]; + tensor input_31_pad_type_0 = const()[name = tensor("input_31_pad_type_0"), val = tensor("valid")]; + tensor input_31_strides_0 = const()[name = tensor("input_31_strides_0"), val = tensor([1])]; + tensor input_31_pad_0 = const()[name = tensor("input_31_pad_0"), val = tensor([0, 0])]; + tensor input_31_dilations_0 = const()[name = tensor("input_31_dilations_0"), val = tensor([1])]; + tensor input_31_groups_0 = const()[name = tensor("input_31_groups_0"), val = tensor(1)]; + tensor input_31 = conv(bias = inner_encoder_conv_module_0_sequential_7_conv_bias, dilations = input_31_dilations_0, groups = input_31_groups_0, pad = input_31_pad_0, pad_type = input_31_pad_type_0, strides = input_31_strides_0, weight = inner_encoder_conv_module_0_sequential_7_conv_weight, x = input_29)[name = tensor("input_31")]; tensor conv_out_1_perm_0 = const()[name = tensor("conv_out_1_perm_0"), val = tensor([0, 2, 1])]; - tensor var_298_begin_0 = const()[name = tensor("op_298_begin_0"), val = tensor([0, -1, 0])]; - tensor var_298_end_0 = const()[name = tensor("op_298_end_0"), val = tensor([3, 16, 256])]; - tensor var_298_end_mask_0 = const()[name = tensor("op_298_end_mask_0"), val = tensor([true, true, true])]; - tensor conv_out_1 = transpose(perm = conv_out_1_perm_0, x = input_29)[name = tensor("transpose_53")]; - tensor var_298 = slice_by_index(begin = var_298_begin_0, end = var_298_end_0, end_mask = var_298_end_mask_0, x = conv_out_1)[name = tensor("op_298")]; - tensor var_300_perm_0 = const()[name = tensor("op_300_perm_0"), val = tensor([1, 0, 2])]; - tensor var_300 = transpose(perm = var_300_perm_0, x = var_298)[name = tensor("transpose_52")]; - tensor input_31 = add(x = x_3, y = var_300)[name = tensor("input_31")]; - tensor input_33_axes_0 = const()[name = tensor("input_33_axes_0"), val = tensor([-1])]; - tensor input_33 = layer_norm(axes = input_33_axes_0, beta = encoder_ffn2_0_module_sequential_0_bias, epsilon = var_28, gamma = encoder_ffn2_0_module_sequential_0_weight, x = input_31)[name = tensor("input_33")]; - tensor inputs_9 = linear(bias = encoder_ffn2_0_module_sequential_1_linear_bias, weight = encoder_ffn2_0_module_sequential_1_linear_weight, x = input_33)[name = tensor("linear_8")]; - tensor input_35 = silu(x = inputs_9)[name = tensor("input_35")]; - tensor input_39 = linear(bias = encoder_ffn2_0_module_sequential_4_linear_bias, weight = encoder_ffn2_0_module_sequential_4_linear_weight, x = input_35)[name = tensor("linear_9")]; - tensor var_323 = const()[name = tensor("op_323"), val = tensor(0x1p-1)]; - tensor var_324 = mul(x = input_39, y = var_323)[name = tensor("op_324")]; - tensor input_41 = add(x = var_324, y = input_31)[name = tensor("input_41")]; - tensor input_43_axes_0 = const()[name = tensor("input_43_axes_0"), val = tensor([-1])]; - tensor input_43 = layer_norm(axes = input_43_axes_0, beta = encoder_layer_norm_0_bias, epsilon = var_28, gamma = encoder_layer_norm_0_weight, x = input_41)[name = tensor("input_43")]; + tensor var_344_begin_0 = const()[name = tensor("op_344_begin_0"), val = tensor([0, -1, 0])]; + tensor var_344_end_0 = const()[name = tensor("op_344_end_0"), val = tensor([3, 16, 256])]; + tensor var_344_end_mask_0 = const()[name = tensor("op_344_end_mask_0"), val = tensor([true, true, true])]; + tensor conv_out_1 = transpose(perm = conv_out_1_perm_0, x = input_31)[name = tensor("transpose_53")]; + tensor var_344 = slice_by_index(begin = var_344_begin_0, end = var_344_end_0, end_mask = var_344_end_mask_0, x = conv_out_1)[name = tensor("op_344")]; + tensor var_346_perm_0 = const()[name = tensor("op_346_perm_0"), val = tensor([1, 0, 2])]; + tensor var_346 = transpose(perm = var_346_perm_0, x = var_344)[name = tensor("transpose_52")]; + tensor input_33 = add(x = x_3, y = var_346)[name = tensor("input_33")]; + tensor input_35_axes_0 = const()[name = tensor("input_35_axes_0"), val = tensor([-1])]; + tensor input_35 = layer_norm(axes = input_35_axes_0, beta = inner_encoder_ffn2_0_module_sequential_0_bias, epsilon = var_56, gamma = inner_encoder_ffn2_0_module_sequential_0_weight, x = input_33)[name = tensor("input_35")]; + tensor inputs_9 = linear(bias = inner_encoder_ffn2_0_module_sequential_1_linear_bias, weight = inner_encoder_ffn2_0_module_sequential_1_linear_weight, x = input_35)[name = tensor("linear_8")]; + tensor input_37 = silu(x = inputs_9)[name = tensor("input_37")]; + tensor input_41 = linear(bias = inner_encoder_ffn2_0_module_sequential_4_linear_bias, weight = inner_encoder_ffn2_0_module_sequential_4_linear_weight, x = input_37)[name = tensor("linear_9")]; + tensor var_369 = const()[name = tensor("op_369"), val = tensor(0x1p-1)]; + tensor var_370 = mul(x = input_41, y = var_369)[name = tensor("op_370")]; + tensor input_43 = add(x = var_370, y = input_33)[name = tensor("input_43")]; tensor input_45_axes_0 = const()[name = tensor("input_45_axes_0"), val = tensor([-1])]; - tensor input_45 = layer_norm(axes = input_45_axes_0, beta = encoder_ffn1_1_module_sequential_0_bias, epsilon = var_28, gamma = encoder_ffn1_1_module_sequential_0_weight, x = input_43)[name = tensor("input_45")]; - tensor inputs_11 = linear(bias = encoder_ffn1_1_module_sequential_1_linear_bias, weight = encoder_ffn1_1_module_sequential_1_linear_weight, x = input_45)[name = tensor("linear_10")]; - tensor input_47 = silu(x = inputs_11)[name = tensor("input_47")]; - tensor input_51 = linear(bias = encoder_ffn1_1_module_sequential_4_linear_bias, weight = encoder_ffn1_1_module_sequential_4_linear_weight, x = input_47)[name = tensor("linear_11")]; - tensor var_353 = const()[name = tensor("op_353"), val = tensor(0x1p-1)]; - tensor var_354 = mul(x = input_51, y = var_353)[name = tensor("op_354")]; - tensor input_53 = add(x = var_354, y = input_43)[name = tensor("input_53")]; + tensor input_45 = layer_norm(axes = input_45_axes_0, beta = inner_encoder_layer_norm_0_bias, epsilon = var_56, gamma = inner_encoder_layer_norm_0_weight, x = input_43)[name = tensor("input_45")]; + tensor input_47_axes_0 = const()[name = tensor("input_47_axes_0"), val = tensor([-1])]; + tensor input_47 = layer_norm(axes = input_47_axes_0, beta = inner_encoder_ffn1_1_module_sequential_0_bias, epsilon = var_56, gamma = inner_encoder_ffn1_1_module_sequential_0_weight, x = input_45)[name = tensor("input_47")]; + tensor inputs_11 = linear(bias = inner_encoder_ffn1_1_module_sequential_1_linear_bias, weight = inner_encoder_ffn1_1_module_sequential_1_linear_weight, x = input_47)[name = tensor("linear_10")]; + tensor input_49 = silu(x = inputs_11)[name = tensor("input_49")]; + tensor input_53 = linear(bias = inner_encoder_ffn1_1_module_sequential_4_linear_bias, weight = inner_encoder_ffn1_1_module_sequential_4_linear_weight, x = input_49)[name = tensor("linear_11")]; + tensor var_399 = const()[name = tensor("op_399"), val = tensor(0x1p-1)]; + tensor var_400 = mul(x = input_53, y = var_399)[name = tensor("op_400")]; + tensor input_55 = add(x = var_400, y = input_45)[name = tensor("input_55")]; tensor x_7_axes_0 = const()[name = tensor("x_7_axes_0"), val = tensor([-1])]; - tensor x_7 = layer_norm(axes = x_7_axes_0, beta = encoder_ret_lns_1_bias, epsilon = var_28, gamma = encoder_ret_lns_1_weight, x = input_53)[name = tensor("x_7")]; + tensor x_7 = layer_norm(axes = x_7_axes_0, beta = inner_encoder_ret_lns_1_bias, epsilon = var_56, gamma = inner_encoder_ret_lns_1_weight, x = input_55)[name = tensor("x_7")]; tensor prev_kv_3_begin_0 = const()[name = tensor("prev_kv_3_begin_0"), val = tensor([1, 0, 0, 0, 0])]; tensor prev_kv_3_end_0 = const()[name = tensor("prev_kv_3_end_0"), val = tensor([2, 1, 4, 64, 64])]; tensor prev_kv_3_end_mask_0 = const()[name = tensor("prev_kv_3_end_mask_0"), val = tensor([false, true, true, true, true])]; @@ -406,163 +424,163 @@ program(1.0) tensor prev_scale_3_end_mask_0 = const()[name = tensor("prev_scale_3_end_mask_0"), val = tensor([false, true])]; tensor prev_scale_3_squeeze_mask_0 = const()[name = tensor("prev_scale_3_squeeze_mask_0"), val = tensor([true, false])]; tensor prev_scale_3 = slice_by_index(begin = prev_scale_3_begin_0, end = prev_scale_3_end_0, end_mask = prev_scale_3_end_mask_0, squeeze_mask = prev_scale_3_squeeze_mask_0, x = enc_scale)[name = tensor("prev_scale_3")]; - tensor var_368 = linear(bias = encoder_q_proj_1_bias, weight = encoder_q_proj_1_weight, x = x_7)[name = tensor("linear_12")]; - tensor var_369 = const()[name = tensor("op_369"), val = tensor([1, 3, 4, 64])]; - tensor var_370 = reshape(shape = var_369, x = var_368)[name = tensor("op_370")]; + tensor var_414 = linear(bias = inner_encoder_q_proj_1_bias, weight = inner_encoder_q_proj_1_weight, x = x_7)[name = tensor("linear_12")]; + tensor var_415 = const()[name = tensor("op_415"), val = tensor([1, 3, 4, 64])]; + tensor var_416 = reshape(shape = var_415, x = var_414)[name = tensor("op_416")]; tensor q_3_perm_0 = const()[name = tensor("q_3_perm_0"), val = tensor([0, 2, 1, 3])]; - tensor var_374 = linear(bias = encoder_k_proj_1_bias, weight = encoder_k_proj_1_weight, x = x_7)[name = tensor("linear_13")]; - tensor var_375 = const()[name = tensor("op_375"), val = tensor(0x1p-3)]; - tensor var_376 = mul(x = var_374, y = var_375)[name = tensor("op_376")]; - tensor var_377 = const()[name = tensor("op_377"), val = tensor([1, 3, 4, 64])]; - tensor var_378 = reshape(shape = var_377, x = var_376)[name = tensor("op_378")]; + tensor var_420 = linear(bias = inner_encoder_k_proj_1_bias, weight = inner_encoder_k_proj_1_weight, x = x_7)[name = tensor("linear_13")]; + tensor var_421 = const()[name = tensor("op_421"), val = tensor(0x1p-3)]; + tensor var_422 = mul(x = var_420, y = var_421)[name = tensor("op_422")]; + tensor var_423 = const()[name = tensor("op_423"), val = tensor([1, 3, 4, 64])]; + tensor var_424 = reshape(shape = var_423, x = var_422)[name = tensor("op_424")]; tensor k_3_perm_0 = const()[name = tensor("k_3_perm_0"), val = tensor([0, 2, 1, 3])]; - tensor var_382 = linear(bias = encoder_v_proj_1_bias, weight = encoder_v_proj_1_weight, x = x_7)[name = tensor("linear_14")]; - tensor var_383 = const()[name = tensor("op_383"), val = tensor([1, 3, 4, 64])]; - tensor var_384 = reshape(shape = var_383, x = var_382)[name = tensor("op_384")]; + tensor var_428 = linear(bias = inner_encoder_v_proj_1_bias, weight = inner_encoder_v_proj_1_weight, x = x_7)[name = tensor("linear_14")]; + tensor var_429 = const()[name = tensor("op_429"), val = tensor([1, 3, 4, 64])]; + tensor var_430 = reshape(shape = var_429, x = var_428)[name = tensor("op_430")]; tensor v_3_perm_0 = const()[name = tensor("v_3_perm_0"), val = tensor([0, 2, 1, 3])]; - tensor input_57 = linear(bias = encoder_g_proj_1_bias, weight = encoder_g_proj_1_weight, x = x_7)[name = tensor("linear_15")]; + tensor input_59 = linear(bias = inner_encoder_g_proj_1_bias, weight = inner_encoder_g_proj_1_weight, x = x_7)[name = tensor("linear_15")]; tensor sqrt_s0_3 = sqrt(x = prev_scale_3)[name = tensor("sqrt_s0_3")]; - tensor s_t_3 = add(x = prev_scale_3, y = encoder__t_index)[name = tensor("s_t_3")]; + tensor s_t_3 = add(x = prev_scale_3, y = inner_encoder__t_index)[name = tensor("s_t_3")]; tensor sqrt_s_t_3 = sqrt(x = s_t_3)[name = tensor("sqrt_s_t_3")]; tensor qk_3_transpose_x_1 = const()[name = tensor("qk_3_transpose_x_1"), val = tensor(false)]; tensor qk_3_transpose_y_1 = const()[name = tensor("qk_3_transpose_y_1"), val = tensor(true)]; - tensor k_3 = transpose(perm = k_3_perm_0, x = var_378)[name = tensor("transpose_50")]; - tensor q_3 = transpose(perm = q_3_perm_0, x = var_370)[name = tensor("transpose_51")]; + tensor k_3 = transpose(perm = k_3_perm_0, x = var_424)[name = tensor("transpose_50")]; + tensor q_3 = transpose(perm = q_3_perm_0, x = var_416)[name = tensor("transpose_51")]; tensor qk_3 = matmul(transpose_x = qk_3_transpose_x_1, transpose_y = qk_3_transpose_y_1, x = q_3, y = k_3)[name = tensor("qk_3")]; - tensor var_394 = const()[name = tensor("op_394"), val = tensor([3, 1])]; - tensor var_395 = reshape(shape = var_394, x = sqrt_s_t_3)[name = tensor("op_395")]; - tensor M_3 = real_div(x = encoder__causal_mask, y = var_395)[name = tensor("M_3")]; - tensor var_397 = mul(x = qk_3, y = M_3)[name = tensor("op_397")]; + tensor var_440 = const()[name = tensor("op_440"), val = tensor([3, 1])]; + tensor var_441 = reshape(shape = var_440, x = sqrt_s_t_3)[name = tensor("op_441")]; + tensor M_3 = real_div(x = inner_encoder__causal_mask, y = var_441)[name = tensor("M_3")]; + tensor var_443 = mul(x = qk_3, y = M_3)[name = tensor("op_443")]; tensor inner_3_transpose_x_0 = const()[name = tensor("inner_3_transpose_x_0"), val = tensor(false)]; tensor inner_3_transpose_y_0 = const()[name = tensor("inner_3_transpose_y_0"), val = tensor(false)]; - tensor v_3 = transpose(perm = v_3_perm_0, x = var_384)[name = tensor("transpose_49")]; - tensor inner_3 = matmul(transpose_x = inner_3_transpose_x_0, transpose_y = inner_3_transpose_y_0, x = var_397, y = v_3)[name = tensor("inner_3")]; - tensor var_399_transpose_x_0 = const()[name = tensor("op_399_transpose_x_0"), val = tensor(false)]; - tensor var_399_transpose_y_0 = const()[name = tensor("op_399_transpose_y_0"), val = tensor(false)]; - tensor var_399 = matmul(transpose_x = var_399_transpose_x_0, transpose_y = var_399_transpose_y_0, x = q_3, y = prev_kv_3)[name = tensor("op_399")]; - tensor var_400 = real_div(x = sqrt_s0_3, y = sqrt_s_t_3)[name = tensor("op_400")]; - tensor var_401 = const()[name = tensor("op_401"), val = tensor([1, 1, 3, 1])]; - tensor var_402 = reshape(shape = var_401, x = var_400)[name = tensor("op_402")]; - tensor cross_3 = mul(x = var_399, y = var_402)[name = tensor("cross_3")]; + tensor v_3 = transpose(perm = v_3_perm_0, x = var_430)[name = tensor("transpose_49")]; + tensor inner_3 = matmul(transpose_x = inner_3_transpose_x_0, transpose_y = inner_3_transpose_y_0, x = var_443, y = v_3)[name = tensor("inner_3")]; + tensor var_445_transpose_x_0 = const()[name = tensor("op_445_transpose_x_0"), val = tensor(false)]; + tensor var_445_transpose_y_0 = const()[name = tensor("op_445_transpose_y_0"), val = tensor(false)]; + tensor var_445 = matmul(transpose_x = var_445_transpose_x_0, transpose_y = var_445_transpose_y_0, x = q_3, y = prev_kv_3)[name = tensor("op_445")]; + tensor var_446 = real_div(x = sqrt_s0_3, y = sqrt_s_t_3)[name = tensor("op_446")]; + tensor var_447 = const()[name = tensor("op_447"), val = tensor([1, 1, 3, 1])]; + tensor var_448 = reshape(shape = var_447, x = var_446)[name = tensor("op_448")]; + tensor cross_3 = mul(x = var_445, y = var_448)[name = tensor("cross_3")]; tensor out_7 = add(x = inner_3, y = cross_3)[name = tensor("out_7")]; - tensor var_405 = mul(x = prev_kv_3, y = sqrt_s0_3)[name = tensor("op_405")]; - tensor var_407_transpose_x_1 = const()[name = tensor("op_407_transpose_x_1"), val = tensor(true)]; - tensor var_407_transpose_y_1 = const()[name = tensor("op_407_transpose_y_1"), val = tensor(false)]; - tensor var_407 = matmul(transpose_x = var_407_transpose_x_1, transpose_y = var_407_transpose_y_1, x = k_3, y = v_3)[name = tensor("op_407")]; - tensor new_kv_unnorm_3 = add(x = var_405, y = var_407)[name = tensor("new_kv_unnorm_3")]; - tensor var_409 = const()[name = tensor("op_409"), val = tensor(0x1.8p+1)]; - tensor new_scale_3 = add(x = prev_scale_3, y = var_409)[name = tensor("new_scale_3")]; - tensor var_411 = sqrt(x = new_scale_3)[name = tensor("op_411")]; - tensor var_412 = real_div(x = new_kv_unnorm_3, y = var_411)[name = tensor("op_412")]; - tensor var_413_perm_0 = const()[name = tensor("op_413_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_451 = mul(x = prev_kv_3, y = sqrt_s0_3)[name = tensor("op_451")]; + tensor var_453_transpose_x_1 = const()[name = tensor("op_453_transpose_x_1"), val = tensor(true)]; + tensor var_453_transpose_y_1 = const()[name = tensor("op_453_transpose_y_1"), val = tensor(false)]; + tensor var_453 = matmul(transpose_x = var_453_transpose_x_1, transpose_y = var_453_transpose_y_1, x = k_3, y = v_3)[name = tensor("op_453")]; + tensor new_kv_unnorm_3 = add(x = var_451, y = var_453)[name = tensor("new_kv_unnorm_3")]; + tensor var_455 = const()[name = tensor("op_455"), val = tensor(0x1.8p+1)]; + tensor new_scale_3 = add(x = prev_scale_3, y = var_455)[name = tensor("new_scale_3")]; + tensor var_457 = sqrt(x = new_scale_3)[name = tensor("op_457")]; + tensor var_458 = real_div(x = new_kv_unnorm_3, y = var_457)[name = tensor("op_458")]; + tensor var_459_perm_0 = const()[name = tensor("op_459_perm_0"), val = tensor([0, 2, 1, 3])]; tensor out_9_axes_0 = const()[name = tensor("out_9_axes_0"), val = tensor([-1])]; - tensor var_413 = transpose(perm = var_413_perm_0, x = out_7)[name = tensor("transpose_48")]; - tensor out_9 = layer_norm(axes = out_9_axes_0, epsilon = var_17, x = var_413)[name = tensor("out_9")]; - tensor var_417 = const()[name = tensor("op_417"), val = tensor([1, 3, 256])]; - tensor out_11 = reshape(shape = var_417, x = out_9)[name = tensor("out_11")]; - tensor var_419 = silu(x = input_57)[name = tensor("op_419")]; - tensor input_59 = mul(x = var_419, y = out_11)[name = tensor("input_59")]; - tensor ret_out_3 = linear(bias = encoder_out_proj_1_bias, weight = encoder_out_proj_1_weight, x = input_59)[name = tensor("linear_16")]; - tensor x_9 = add(x = input_53, y = ret_out_3)[name = tensor("x_9")]; + tensor var_459 = transpose(perm = var_459_perm_0, x = out_7)[name = tensor("transpose_48")]; + tensor out_9 = layer_norm(axes = out_9_axes_0, epsilon = var_64, x = var_459)[name = tensor("out_9")]; + tensor var_463 = const()[name = tensor("op_463"), val = tensor([1, 3, 256])]; + tensor out_11 = reshape(shape = var_463, x = out_9)[name = tensor("out_11")]; + tensor var_465 = silu(x = input_59)[name = tensor("op_465")]; + tensor input_61 = mul(x = var_465, y = out_11)[name = tensor("input_61")]; + tensor ret_out_3 = linear(bias = inner_encoder_out_proj_1_bias, weight = inner_encoder_out_proj_1_weight, x = input_61)[name = tensor("linear_16")]; + tensor x_9 = add(x = input_55, y = ret_out_3)[name = tensor("x_9")]; tensor window_9_begin_0 = const()[name = tensor("window_9_begin_0"), val = tensor([1, 0, 0, 0])]; tensor window_9_end_0 = const()[name = tensor("window_9_end_0"), val = tensor([2, 1, 16, 256])]; tensor window_9_end_mask_0 = const()[name = tensor("window_9_end_mask_0"), val = tensor([false, true, true, true])]; tensor window_9_squeeze_mask_0 = const()[name = tensor("window_9_squeeze_mask_0"), val = tensor([true, false, false, false])]; tensor window_9 = slice_by_index(begin = window_9_begin_0, end = window_9_end_0, end_mask = window_9_end_mask_0, squeeze_mask = window_9_squeeze_mask_0, x = enc_conv_cache)[name = tensor("window_9")]; - tensor var_427_begin_0 = const()[name = tensor("op_427_begin_0"), val = tensor([0, 0, 0])]; - tensor var_427_end_0 = const()[name = tensor("op_427_end_0"), val = tensor([1, 1, 256])]; - tensor var_427_end_mask_0 = const()[name = tensor("op_427_end_mask_0"), val = tensor([true, false, true])]; - tensor var_427 = slice_by_index(begin = var_427_begin_0, end = var_427_end_0, end_mask = var_427_end_mask_0, x = x_9)[name = tensor("op_427")]; - tensor var_430_begin_0 = const()[name = tensor("op_430_begin_0"), val = tensor([0, 1, 0])]; - tensor var_430_end_0 = const()[name = tensor("op_430_end_0"), val = tensor([1, 16, 256])]; - tensor var_430_end_mask_0 = const()[name = tensor("op_430_end_mask_0"), val = tensor([true, true, true])]; - tensor var_430 = slice_by_index(begin = var_430_begin_0, end = var_430_end_0, end_mask = var_430_end_mask_0, x = window_9)[name = tensor("op_430")]; + tensor var_473_begin_0 = const()[name = tensor("op_473_begin_0"), val = tensor([0, 0, 0])]; + tensor var_473_end_0 = const()[name = tensor("op_473_end_0"), val = tensor([1, 1, 256])]; + tensor var_473_end_mask_0 = const()[name = tensor("op_473_end_mask_0"), val = tensor([true, false, true])]; + tensor var_473 = slice_by_index(begin = var_473_begin_0, end = var_473_end_0, end_mask = var_473_end_mask_0, x = x_9)[name = tensor("op_473")]; + tensor var_476_begin_0 = const()[name = tensor("op_476_begin_0"), val = tensor([0, 1, 0])]; + tensor var_476_end_0 = const()[name = tensor("op_476_end_0"), val = tensor([1, 16, 256])]; + tensor var_476_end_mask_0 = const()[name = tensor("op_476_end_mask_0"), val = tensor([true, true, true])]; + tensor var_476 = slice_by_index(begin = var_476_begin_0, end = var_476_end_0, end_mask = var_476_end_mask_0, x = window_9)[name = tensor("op_476")]; tensor window_11_interleave_0 = const()[name = tensor("window_11_interleave_0"), val = tensor(false)]; - tensor window_11 = concat(axis = var_26, interleave = window_11_interleave_0, values = (var_430, var_427))[name = tensor("window_11")]; - tensor var_435_begin_0 = const()[name = tensor("op_435_begin_0"), val = tensor([0, 1, 0])]; - tensor var_435_end_0 = const()[name = tensor("op_435_end_0"), val = tensor([1, 2, 256])]; - tensor var_435_end_mask_0 = const()[name = tensor("op_435_end_mask_0"), val = tensor([true, false, true])]; - tensor var_435 = slice_by_index(begin = var_435_begin_0, end = var_435_end_0, end_mask = var_435_end_mask_0, x = x_9)[name = tensor("op_435")]; - tensor var_438_begin_0 = const()[name = tensor("op_438_begin_0"), val = tensor([0, 1, 0])]; - tensor var_438_end_0 = const()[name = tensor("op_438_end_0"), val = tensor([1, 16, 256])]; - tensor var_438_end_mask_0 = const()[name = tensor("op_438_end_mask_0"), val = tensor([true, true, true])]; - tensor var_438 = slice_by_index(begin = var_438_begin_0, end = var_438_end_0, end_mask = var_438_end_mask_0, x = window_11)[name = tensor("op_438")]; + tensor window_11 = concat(axis = var_72, interleave = window_11_interleave_0, values = (var_476, var_473))[name = tensor("window_11")]; + tensor var_481_begin_0 = const()[name = tensor("op_481_begin_0"), val = tensor([0, 1, 0])]; + tensor var_481_end_0 = const()[name = tensor("op_481_end_0"), val = tensor([1, 2, 256])]; + tensor var_481_end_mask_0 = const()[name = tensor("op_481_end_mask_0"), val = tensor([true, false, true])]; + tensor var_481 = slice_by_index(begin = var_481_begin_0, end = var_481_end_0, end_mask = var_481_end_mask_0, x = x_9)[name = tensor("op_481")]; + tensor var_484_begin_0 = const()[name = tensor("op_484_begin_0"), val = tensor([0, 1, 0])]; + tensor var_484_end_0 = const()[name = tensor("op_484_end_0"), val = tensor([1, 16, 256])]; + tensor var_484_end_mask_0 = const()[name = tensor("op_484_end_mask_0"), val = tensor([true, true, true])]; + tensor var_484 = slice_by_index(begin = var_484_begin_0, end = var_484_end_0, end_mask = var_484_end_mask_0, x = window_11)[name = tensor("op_484")]; tensor window_13_interleave_0 = const()[name = tensor("window_13_interleave_0"), val = tensor(false)]; - tensor window_13 = concat(axis = var_26, interleave = window_13_interleave_0, values = (var_438, var_435))[name = tensor("window_13")]; - tensor var_443_begin_0 = const()[name = tensor("op_443_begin_0"), val = tensor([0, 2, 0])]; - tensor var_443_end_0 = const()[name = tensor("op_443_end_0"), val = tensor([1, 1, 256])]; - tensor var_443_end_mask_0 = const()[name = tensor("op_443_end_mask_0"), val = tensor([true, true, true])]; - tensor var_443 = slice_by_index(begin = var_443_begin_0, end = var_443_end_0, end_mask = var_443_end_mask_0, x = x_9)[name = tensor("op_443")]; - tensor var_446_begin_0 = const()[name = tensor("op_446_begin_0"), val = tensor([0, 1, 0])]; - tensor var_446_end_0 = const()[name = tensor("op_446_end_0"), val = tensor([1, 16, 256])]; - tensor var_446_end_mask_0 = const()[name = tensor("op_446_end_mask_0"), val = tensor([true, true, true])]; - tensor var_446 = slice_by_index(begin = var_446_begin_0, end = var_446_end_0, end_mask = var_446_end_mask_0, x = window_13)[name = tensor("op_446")]; + tensor window_13 = concat(axis = var_72, interleave = window_13_interleave_0, values = (var_484, var_481))[name = tensor("window_13")]; + tensor var_489_begin_0 = const()[name = tensor("op_489_begin_0"), val = tensor([0, 2, 0])]; + tensor var_489_end_0 = const()[name = tensor("op_489_end_0"), val = tensor([1, 1, 256])]; + tensor var_489_end_mask_0 = const()[name = tensor("op_489_end_mask_0"), val = tensor([true, true, true])]; + tensor var_489 = slice_by_index(begin = var_489_begin_0, end = var_489_end_0, end_mask = var_489_end_mask_0, x = x_9)[name = tensor("op_489")]; + tensor var_492_begin_0 = const()[name = tensor("op_492_begin_0"), val = tensor([0, 1, 0])]; + tensor var_492_end_0 = const()[name = tensor("op_492_end_0"), val = tensor([1, 16, 256])]; + tensor var_492_end_mask_0 = const()[name = tensor("op_492_end_mask_0"), val = tensor([true, true, true])]; + tensor var_492 = slice_by_index(begin = var_492_begin_0, end = var_492_end_0, end_mask = var_492_end_mask_0, x = window_13)[name = tensor("op_492")]; tensor window_15_interleave_0 = const()[name = tensor("window_15_interleave_0"), val = tensor(false)]; - tensor window_15 = concat(axis = var_26, interleave = window_15_interleave_0, values = (var_446, var_443))[name = tensor("window_15")]; - tensor input_61_interleave_0 = const()[name = tensor("input_61_interleave_0"), val = tensor(false)]; - tensor input_61 = concat(axis = var_23, interleave = input_61_interleave_0, values = (window_11, window_13, window_15))[name = tensor("input_61")]; + tensor window_15 = concat(axis = var_72, interleave = window_15_interleave_0, values = (var_492, var_489))[name = tensor("window_15")]; + tensor input_63_interleave_0 = const()[name = tensor("input_63_interleave_0"), val = tensor(false)]; + tensor input_63 = concat(axis = var_59, interleave = input_63_interleave_0, values = (window_11, window_13, window_15))[name = tensor("input_63")]; tensor x_11_axes_0 = const()[name = tensor("x_11_axes_0"), val = tensor([-1])]; - tensor x_11 = layer_norm(axes = x_11_axes_0, beta = encoder_conv_module_1_sequential_0_bias, epsilon = var_28, gamma = encoder_conv_module_1_sequential_0_weight, x = input_61)[name = tensor("x_11")]; - tensor input_63_perm_0 = const()[name = tensor("input_63_perm_0"), val = tensor([0, 2, 1])]; + tensor x_11 = layer_norm(axes = x_11_axes_0, beta = inner_encoder_conv_module_1_sequential_0_bias, epsilon = var_56, gamma = inner_encoder_conv_module_1_sequential_0_weight, x = input_63)[name = tensor("x_11")]; + tensor input_65_perm_0 = const()[name = tensor("input_65_perm_0"), val = tensor([0, 2, 1])]; tensor inputs_13_pad_type_0 = const()[name = tensor("inputs_13_pad_type_0"), val = tensor("valid")]; tensor inputs_13_strides_0 = const()[name = tensor("inputs_13_strides_0"), val = tensor([1])]; tensor inputs_13_pad_0 = const()[name = tensor("inputs_13_pad_0"), val = tensor([0, 0])]; tensor inputs_13_dilations_0 = const()[name = tensor("inputs_13_dilations_0"), val = tensor([1])]; tensor inputs_13_groups_0 = const()[name = tensor("inputs_13_groups_0"), val = tensor(1)]; - tensor input_63 = transpose(perm = input_63_perm_0, x = x_11)[name = tensor("transpose_47")]; - tensor inputs_13 = conv(bias = encoder_conv_module_1_sequential_2_conv_bias, dilations = inputs_13_dilations_0, groups = inputs_13_groups_0, pad = inputs_13_pad_0, pad_type = inputs_13_pad_type_0, strides = inputs_13_strides_0, weight = encoder_conv_module_1_sequential_2_conv_weight, x = input_63)[name = tensor("inputs_13")]; - tensor var_471_split_sizes_0 = const()[name = tensor("op_471_split_sizes_0"), val = tensor([256, 256])]; - tensor var_471_axis_0 = const()[name = tensor("op_471_axis_0"), val = tensor(1)]; - tensor var_471_0, tensor var_471_1 = split(axis = var_471_axis_0, split_sizes = var_471_split_sizes_0, x = inputs_13)[name = tensor("op_471")]; - tensor var_473 = sigmoid(x = var_471_1)[name = tensor("op_473")]; - tensor inputs_15 = mul(x = var_471_0, y = var_473)[name = tensor("inputs_15")]; + tensor input_65 = transpose(perm = input_65_perm_0, x = x_11)[name = tensor("transpose_47")]; + tensor inputs_13 = conv(bias = inner_encoder_conv_module_1_sequential_2_conv_bias, dilations = inputs_13_dilations_0, groups = inputs_13_groups_0, pad = inputs_13_pad_0, pad_type = inputs_13_pad_type_0, strides = inputs_13_strides_0, weight = inner_encoder_conv_module_1_sequential_2_conv_weight, x = input_65)[name = tensor("inputs_13")]; + tensor var_517_split_sizes_0 = const()[name = tensor("op_517_split_sizes_0"), val = tensor([256, 256])]; + tensor var_517_axis_0 = const()[name = tensor("op_517_axis_0"), val = tensor(1)]; + tensor var_517_0, tensor var_517_1 = split(axis = var_517_axis_0, split_sizes = var_517_split_sizes_0, x = inputs_13)[name = tensor("op_517")]; + tensor var_519 = sigmoid(x = var_517_1)[name = tensor("op_519")]; + tensor inputs_15 = mul(x = var_517_0, y = var_519)[name = tensor("inputs_15")]; tensor outputs_aug_3_pad_type_0 = const()[name = tensor("outputs_aug_3_pad_type_0"), val = tensor("custom")]; tensor outputs_aug_3_pad_0 = const()[name = tensor("outputs_aug_3_pad_0"), val = tensor([15, 15])]; tensor outputs_aug_3_groups_0 = const()[name = tensor("outputs_aug_3_groups_0"), val = tensor(256)]; tensor outputs_aug_3_strides_0 = const()[name = tensor("outputs_aug_3_strides_0"), val = tensor([1])]; tensor outputs_aug_3_dilations_0 = const()[name = tensor("outputs_aug_3_dilations_0"), val = tensor([1])]; - tensor outputs_aug_3 = conv(dilations = outputs_aug_3_dilations_0, groups = outputs_aug_3_groups_0, pad = outputs_aug_3_pad_0, pad_type = outputs_aug_3_pad_type_0, strides = outputs_aug_3_strides_0, weight = encoder_conv_module_1_sequential_4_conv_weight, x = inputs_15)[name = tensor("outputs_aug_3")]; - tensor input_65_begin_0 = const()[name = tensor("input_65_begin_0"), val = tensor([0, 0, 0])]; - tensor input_65_end_0 = const()[name = tensor("input_65_end_0"), val = tensor([3, 256, 16])]; - tensor input_65_end_mask_0 = const()[name = tensor("input_65_end_mask_0"), val = tensor([true, true, false])]; - tensor input_65 = slice_by_index(begin = input_65_begin_0, end = input_65_end_0, end_mask = input_65_end_mask_0, x = outputs_aug_3)[name = tensor("input_65")]; - tensor inputs_17 = batch_norm(beta = encoder_conv_module_1_sequential_5_bias, epsilon = var_28, gamma = encoder_conv_module_1_sequential_5_weight, mean = encoder_conv_module_1_sequential_5_running_mean, variance = encoder_conv_module_1_sequential_5_running_var, x = input_65)[name = tensor("inputs_17")]; - tensor input_67 = silu(x = inputs_17)[name = tensor("input_67")]; - tensor input_69_pad_type_0 = const()[name = tensor("input_69_pad_type_0"), val = tensor("valid")]; - tensor input_69_strides_0 = const()[name = tensor("input_69_strides_0"), val = tensor([1])]; - tensor input_69_pad_0 = const()[name = tensor("input_69_pad_0"), val = tensor([0, 0])]; - tensor input_69_dilations_0 = const()[name = tensor("input_69_dilations_0"), val = tensor([1])]; - tensor input_69_groups_0 = const()[name = tensor("input_69_groups_0"), val = tensor(1)]; - tensor input_69 = conv(bias = encoder_conv_module_1_sequential_7_conv_bias, dilations = input_69_dilations_0, groups = input_69_groups_0, pad = input_69_pad_0, pad_type = input_69_pad_type_0, strides = input_69_strides_0, weight = encoder_conv_module_1_sequential_7_conv_weight, x = input_67)[name = tensor("input_69")]; + tensor outputs_aug_3 = conv(dilations = outputs_aug_3_dilations_0, groups = outputs_aug_3_groups_0, pad = outputs_aug_3_pad_0, pad_type = outputs_aug_3_pad_type_0, strides = outputs_aug_3_strides_0, weight = inner_encoder_conv_module_1_sequential_4_conv_weight, x = inputs_15)[name = tensor("outputs_aug_3")]; + tensor input_67_begin_0 = const()[name = tensor("input_67_begin_0"), val = tensor([0, 0, 0])]; + tensor input_67_end_0 = const()[name = tensor("input_67_end_0"), val = tensor([3, 256, 16])]; + tensor input_67_end_mask_0 = const()[name = tensor("input_67_end_mask_0"), val = tensor([true, true, false])]; + tensor input_67 = slice_by_index(begin = input_67_begin_0, end = input_67_end_0, end_mask = input_67_end_mask_0, x = outputs_aug_3)[name = tensor("input_67")]; + tensor inputs_17 = batch_norm(beta = inner_encoder_conv_module_1_sequential_5_bias, epsilon = var_56, gamma = inner_encoder_conv_module_1_sequential_5_weight, mean = inner_encoder_conv_module_1_sequential_5_running_mean, variance = inner_encoder_conv_module_1_sequential_5_running_var, x = input_67)[name = tensor("inputs_17")]; + tensor input_69 = silu(x = inputs_17)[name = tensor("input_69")]; + tensor input_71_pad_type_0 = const()[name = tensor("input_71_pad_type_0"), val = tensor("valid")]; + tensor input_71_strides_0 = const()[name = tensor("input_71_strides_0"), val = tensor([1])]; + tensor input_71_pad_0 = const()[name = tensor("input_71_pad_0"), val = tensor([0, 0])]; + tensor input_71_dilations_0 = const()[name = tensor("input_71_dilations_0"), val = tensor([1])]; + tensor input_71_groups_0 = const()[name = tensor("input_71_groups_0"), val = tensor(1)]; + tensor input_71 = conv(bias = inner_encoder_conv_module_1_sequential_7_conv_bias, dilations = input_71_dilations_0, groups = input_71_groups_0, pad = input_71_pad_0, pad_type = input_71_pad_type_0, strides = input_71_strides_0, weight = inner_encoder_conv_module_1_sequential_7_conv_weight, x = input_69)[name = tensor("input_71")]; tensor conv_out_3_perm_0 = const()[name = tensor("conv_out_3_perm_0"), val = tensor([0, 2, 1])]; - tensor var_504_begin_0 = const()[name = tensor("op_504_begin_0"), val = tensor([0, -1, 0])]; - tensor var_504_end_0 = const()[name = tensor("op_504_end_0"), val = tensor([3, 16, 256])]; - tensor var_504_end_mask_0 = const()[name = tensor("op_504_end_mask_0"), val = tensor([true, true, true])]; - tensor conv_out_3 = transpose(perm = conv_out_3_perm_0, x = input_69)[name = tensor("transpose_46")]; - tensor var_504 = slice_by_index(begin = var_504_begin_0, end = var_504_end_0, end_mask = var_504_end_mask_0, x = conv_out_3)[name = tensor("op_504")]; - tensor var_506_perm_0 = const()[name = tensor("op_506_perm_0"), val = tensor([1, 0, 2])]; - tensor var_506 = transpose(perm = var_506_perm_0, x = var_504)[name = tensor("transpose_45")]; - tensor input_71 = add(x = x_9, y = var_506)[name = tensor("input_71")]; - tensor input_73_axes_0 = const()[name = tensor("input_73_axes_0"), val = tensor([-1])]; - tensor input_73 = layer_norm(axes = input_73_axes_0, beta = encoder_ffn2_1_module_sequential_0_bias, epsilon = var_28, gamma = encoder_ffn2_1_module_sequential_0_weight, x = input_71)[name = tensor("input_73")]; - tensor inputs_19 = linear(bias = encoder_ffn2_1_module_sequential_1_linear_bias, weight = encoder_ffn2_1_module_sequential_1_linear_weight, x = input_73)[name = tensor("linear_17")]; - tensor input_75 = silu(x = inputs_19)[name = tensor("input_75")]; - tensor input_79 = linear(bias = encoder_ffn2_1_module_sequential_4_linear_bias, weight = encoder_ffn2_1_module_sequential_4_linear_weight, x = input_75)[name = tensor("linear_18")]; - tensor var_529 = const()[name = tensor("op_529"), val = tensor(0x1p-1)]; - tensor var_530 = mul(x = input_79, y = var_529)[name = tensor("op_530")]; - tensor input_81 = add(x = var_530, y = input_71)[name = tensor("input_81")]; - tensor input_83_axes_0 = const()[name = tensor("input_83_axes_0"), val = tensor([-1])]; - tensor input_83 = layer_norm(axes = input_83_axes_0, beta = encoder_layer_norm_1_bias, epsilon = var_28, gamma = encoder_layer_norm_1_weight, x = input_81)[name = tensor("input_83")]; + tensor var_550_begin_0 = const()[name = tensor("op_550_begin_0"), val = tensor([0, -1, 0])]; + tensor var_550_end_0 = const()[name = tensor("op_550_end_0"), val = tensor([3, 16, 256])]; + tensor var_550_end_mask_0 = const()[name = tensor("op_550_end_mask_0"), val = tensor([true, true, true])]; + tensor conv_out_3 = transpose(perm = conv_out_3_perm_0, x = input_71)[name = tensor("transpose_46")]; + tensor var_550 = slice_by_index(begin = var_550_begin_0, end = var_550_end_0, end_mask = var_550_end_mask_0, x = conv_out_3)[name = tensor("op_550")]; + tensor var_552_perm_0 = const()[name = tensor("op_552_perm_0"), val = tensor([1, 0, 2])]; + tensor var_552 = transpose(perm = var_552_perm_0, x = var_550)[name = tensor("transpose_45")]; + tensor input_73 = add(x = x_9, y = var_552)[name = tensor("input_73")]; + tensor input_75_axes_0 = const()[name = tensor("input_75_axes_0"), val = tensor([-1])]; + tensor input_75 = layer_norm(axes = input_75_axes_0, beta = inner_encoder_ffn2_1_module_sequential_0_bias, epsilon = var_56, gamma = inner_encoder_ffn2_1_module_sequential_0_weight, x = input_73)[name = tensor("input_75")]; + tensor inputs_19 = linear(bias = inner_encoder_ffn2_1_module_sequential_1_linear_bias, weight = inner_encoder_ffn2_1_module_sequential_1_linear_weight, x = input_75)[name = tensor("linear_17")]; + tensor input_77 = silu(x = inputs_19)[name = tensor("input_77")]; + tensor input_81 = linear(bias = inner_encoder_ffn2_1_module_sequential_4_linear_bias, weight = inner_encoder_ffn2_1_module_sequential_4_linear_weight, x = input_77)[name = tensor("linear_18")]; + tensor var_575 = const()[name = tensor("op_575"), val = tensor(0x1p-1)]; + tensor var_576 = mul(x = input_81, y = var_575)[name = tensor("op_576")]; + tensor input_83 = add(x = var_576, y = input_73)[name = tensor("input_83")]; tensor input_85_axes_0 = const()[name = tensor("input_85_axes_0"), val = tensor([-1])]; - tensor input_85 = layer_norm(axes = input_85_axes_0, beta = encoder_ffn1_2_module_sequential_0_bias, epsilon = var_28, gamma = encoder_ffn1_2_module_sequential_0_weight, x = input_83)[name = tensor("input_85")]; - tensor inputs_21 = linear(bias = encoder_ffn1_2_module_sequential_1_linear_bias, weight = encoder_ffn1_2_module_sequential_1_linear_weight, x = input_85)[name = tensor("linear_19")]; - tensor input_87 = silu(x = inputs_21)[name = tensor("input_87")]; - tensor input_91 = linear(bias = encoder_ffn1_2_module_sequential_4_linear_bias, weight = encoder_ffn1_2_module_sequential_4_linear_weight, x = input_87)[name = tensor("linear_20")]; - tensor var_559 = const()[name = tensor("op_559"), val = tensor(0x1p-1)]; - tensor var_560 = mul(x = input_91, y = var_559)[name = tensor("op_560")]; - tensor input_93 = add(x = var_560, y = input_83)[name = tensor("input_93")]; + tensor input_85 = layer_norm(axes = input_85_axes_0, beta = inner_encoder_layer_norm_1_bias, epsilon = var_56, gamma = inner_encoder_layer_norm_1_weight, x = input_83)[name = tensor("input_85")]; + tensor input_87_axes_0 = const()[name = tensor("input_87_axes_0"), val = tensor([-1])]; + tensor input_87 = layer_norm(axes = input_87_axes_0, beta = inner_encoder_ffn1_2_module_sequential_0_bias, epsilon = var_56, gamma = inner_encoder_ffn1_2_module_sequential_0_weight, x = input_85)[name = tensor("input_87")]; + tensor inputs_21 = linear(bias = inner_encoder_ffn1_2_module_sequential_1_linear_bias, weight = inner_encoder_ffn1_2_module_sequential_1_linear_weight, x = input_87)[name = tensor("linear_19")]; + tensor input_89 = silu(x = inputs_21)[name = tensor("input_89")]; + tensor input_93 = linear(bias = inner_encoder_ffn1_2_module_sequential_4_linear_bias, weight = inner_encoder_ffn1_2_module_sequential_4_linear_weight, x = input_89)[name = tensor("linear_20")]; + tensor var_605 = const()[name = tensor("op_605"), val = tensor(0x1p-1)]; + tensor var_606 = mul(x = input_93, y = var_605)[name = tensor("op_606")]; + tensor input_95 = add(x = var_606, y = input_85)[name = tensor("input_95")]; tensor x_13_axes_0 = const()[name = tensor("x_13_axes_0"), val = tensor([-1])]; - tensor x_13 = layer_norm(axes = x_13_axes_0, beta = encoder_ret_lns_2_bias, epsilon = var_28, gamma = encoder_ret_lns_2_weight, x = input_93)[name = tensor("x_13")]; + tensor x_13 = layer_norm(axes = x_13_axes_0, beta = inner_encoder_ret_lns_2_bias, epsilon = var_56, gamma = inner_encoder_ret_lns_2_weight, x = input_95)[name = tensor("x_13")]; tensor prev_kv_5_begin_0 = const()[name = tensor("prev_kv_5_begin_0"), val = tensor([2, 0, 0, 0, 0])]; tensor prev_kv_5_end_0 = const()[name = tensor("prev_kv_5_end_0"), val = tensor([3, 1, 4, 64, 64])]; tensor prev_kv_5_end_mask_0 = const()[name = tensor("prev_kv_5_end_mask_0"), val = tensor([false, true, true, true, true])]; @@ -573,163 +591,163 @@ program(1.0) tensor prev_scale_5_end_mask_0 = const()[name = tensor("prev_scale_5_end_mask_0"), val = tensor([false, true])]; tensor prev_scale_5_squeeze_mask_0 = const()[name = tensor("prev_scale_5_squeeze_mask_0"), val = tensor([true, false])]; tensor prev_scale_5 = slice_by_index(begin = prev_scale_5_begin_0, end = prev_scale_5_end_0, end_mask = prev_scale_5_end_mask_0, squeeze_mask = prev_scale_5_squeeze_mask_0, x = enc_scale)[name = tensor("prev_scale_5")]; - tensor var_574 = linear(bias = encoder_q_proj_2_bias, weight = encoder_q_proj_2_weight, x = x_13)[name = tensor("linear_21")]; - tensor var_575 = const()[name = tensor("op_575"), val = tensor([1, 3, 4, 64])]; - tensor var_576 = reshape(shape = var_575, x = var_574)[name = tensor("op_576")]; + tensor var_620 = linear(bias = inner_encoder_q_proj_2_bias, weight = inner_encoder_q_proj_2_weight, x = x_13)[name = tensor("linear_21")]; + tensor var_621 = const()[name = tensor("op_621"), val = tensor([1, 3, 4, 64])]; + tensor var_622 = reshape(shape = var_621, x = var_620)[name = tensor("op_622")]; tensor q_5_perm_0 = const()[name = tensor("q_5_perm_0"), val = tensor([0, 2, 1, 3])]; - tensor var_580 = linear(bias = encoder_k_proj_2_bias, weight = encoder_k_proj_2_weight, x = x_13)[name = tensor("linear_22")]; - tensor var_581 = const()[name = tensor("op_581"), val = tensor(0x1p-3)]; - tensor var_582 = mul(x = var_580, y = var_581)[name = tensor("op_582")]; - tensor var_583 = const()[name = tensor("op_583"), val = tensor([1, 3, 4, 64])]; - tensor var_584 = reshape(shape = var_583, x = var_582)[name = tensor("op_584")]; + tensor var_626 = linear(bias = inner_encoder_k_proj_2_bias, weight = inner_encoder_k_proj_2_weight, x = x_13)[name = tensor("linear_22")]; + tensor var_627 = const()[name = tensor("op_627"), val = tensor(0x1p-3)]; + tensor var_628 = mul(x = var_626, y = var_627)[name = tensor("op_628")]; + tensor var_629 = const()[name = tensor("op_629"), val = tensor([1, 3, 4, 64])]; + tensor var_630 = reshape(shape = var_629, x = var_628)[name = tensor("op_630")]; tensor k_5_perm_0 = const()[name = tensor("k_5_perm_0"), val = tensor([0, 2, 1, 3])]; - tensor var_588 = linear(bias = encoder_v_proj_2_bias, weight = encoder_v_proj_2_weight, x = x_13)[name = tensor("linear_23")]; - tensor var_589 = const()[name = tensor("op_589"), val = tensor([1, 3, 4, 64])]; - tensor var_590 = reshape(shape = var_589, x = var_588)[name = tensor("op_590")]; + tensor var_634 = linear(bias = inner_encoder_v_proj_2_bias, weight = inner_encoder_v_proj_2_weight, x = x_13)[name = tensor("linear_23")]; + tensor var_635 = const()[name = tensor("op_635"), val = tensor([1, 3, 4, 64])]; + tensor var_636 = reshape(shape = var_635, x = var_634)[name = tensor("op_636")]; tensor v_5_perm_0 = const()[name = tensor("v_5_perm_0"), val = tensor([0, 2, 1, 3])]; - tensor input_97 = linear(bias = encoder_g_proj_2_bias, weight = encoder_g_proj_2_weight, x = x_13)[name = tensor("linear_24")]; + tensor input_99 = linear(bias = inner_encoder_g_proj_2_bias, weight = inner_encoder_g_proj_2_weight, x = x_13)[name = tensor("linear_24")]; tensor sqrt_s0_5 = sqrt(x = prev_scale_5)[name = tensor("sqrt_s0_5")]; - tensor s_t_5 = add(x = prev_scale_5, y = encoder__t_index)[name = tensor("s_t_5")]; + tensor s_t_5 = add(x = prev_scale_5, y = inner_encoder__t_index)[name = tensor("s_t_5")]; tensor sqrt_s_t_5 = sqrt(x = s_t_5)[name = tensor("sqrt_s_t_5")]; tensor qk_5_transpose_x_1 = const()[name = tensor("qk_5_transpose_x_1"), val = tensor(false)]; tensor qk_5_transpose_y_1 = const()[name = tensor("qk_5_transpose_y_1"), val = tensor(true)]; - tensor k_5 = transpose(perm = k_5_perm_0, x = var_584)[name = tensor("transpose_43")]; - tensor q_5 = transpose(perm = q_5_perm_0, x = var_576)[name = tensor("transpose_44")]; + tensor k_5 = transpose(perm = k_5_perm_0, x = var_630)[name = tensor("transpose_43")]; + tensor q_5 = transpose(perm = q_5_perm_0, x = var_622)[name = tensor("transpose_44")]; tensor qk_5 = matmul(transpose_x = qk_5_transpose_x_1, transpose_y = qk_5_transpose_y_1, x = q_5, y = k_5)[name = tensor("qk_5")]; - tensor var_600 = const()[name = tensor("op_600"), val = tensor([3, 1])]; - tensor var_601 = reshape(shape = var_600, x = sqrt_s_t_5)[name = tensor("op_601")]; - tensor M_5 = real_div(x = encoder__causal_mask, y = var_601)[name = tensor("M_5")]; - tensor var_603 = mul(x = qk_5, y = M_5)[name = tensor("op_603")]; + tensor var_646 = const()[name = tensor("op_646"), val = tensor([3, 1])]; + tensor var_647 = reshape(shape = var_646, x = sqrt_s_t_5)[name = tensor("op_647")]; + tensor M_5 = real_div(x = inner_encoder__causal_mask, y = var_647)[name = tensor("M_5")]; + tensor var_649 = mul(x = qk_5, y = M_5)[name = tensor("op_649")]; tensor inner_5_transpose_x_0 = const()[name = tensor("inner_5_transpose_x_0"), val = tensor(false)]; tensor inner_5_transpose_y_0 = const()[name = tensor("inner_5_transpose_y_0"), val = tensor(false)]; - tensor v_5 = transpose(perm = v_5_perm_0, x = var_590)[name = tensor("transpose_42")]; - tensor inner_5 = matmul(transpose_x = inner_5_transpose_x_0, transpose_y = inner_5_transpose_y_0, x = var_603, y = v_5)[name = tensor("inner_5")]; - tensor var_605_transpose_x_0 = const()[name = tensor("op_605_transpose_x_0"), val = tensor(false)]; - tensor var_605_transpose_y_0 = const()[name = tensor("op_605_transpose_y_0"), val = tensor(false)]; - tensor var_605 = matmul(transpose_x = var_605_transpose_x_0, transpose_y = var_605_transpose_y_0, x = q_5, y = prev_kv_5)[name = tensor("op_605")]; - tensor var_606 = real_div(x = sqrt_s0_5, y = sqrt_s_t_5)[name = tensor("op_606")]; - tensor var_607 = const()[name = tensor("op_607"), val = tensor([1, 1, 3, 1])]; - tensor var_608 = reshape(shape = var_607, x = var_606)[name = tensor("op_608")]; - tensor cross_5 = mul(x = var_605, y = var_608)[name = tensor("cross_5")]; + tensor v_5 = transpose(perm = v_5_perm_0, x = var_636)[name = tensor("transpose_42")]; + tensor inner_5 = matmul(transpose_x = inner_5_transpose_x_0, transpose_y = inner_5_transpose_y_0, x = var_649, y = v_5)[name = tensor("inner_5")]; + tensor var_651_transpose_x_0 = const()[name = tensor("op_651_transpose_x_0"), val = tensor(false)]; + tensor var_651_transpose_y_0 = const()[name = tensor("op_651_transpose_y_0"), val = tensor(false)]; + tensor var_651 = matmul(transpose_x = var_651_transpose_x_0, transpose_y = var_651_transpose_y_0, x = q_5, y = prev_kv_5)[name = tensor("op_651")]; + tensor var_652 = real_div(x = sqrt_s0_5, y = sqrt_s_t_5)[name = tensor("op_652")]; + tensor var_653 = const()[name = tensor("op_653"), val = tensor([1, 1, 3, 1])]; + tensor var_654 = reshape(shape = var_653, x = var_652)[name = tensor("op_654")]; + tensor cross_5 = mul(x = var_651, y = var_654)[name = tensor("cross_5")]; tensor out_13 = add(x = inner_5, y = cross_5)[name = tensor("out_13")]; - tensor var_611 = mul(x = prev_kv_5, y = sqrt_s0_5)[name = tensor("op_611")]; - tensor var_613_transpose_x_1 = const()[name = tensor("op_613_transpose_x_1"), val = tensor(true)]; - tensor var_613_transpose_y_1 = const()[name = tensor("op_613_transpose_y_1"), val = tensor(false)]; - tensor var_613 = matmul(transpose_x = var_613_transpose_x_1, transpose_y = var_613_transpose_y_1, x = k_5, y = v_5)[name = tensor("op_613")]; - tensor new_kv_unnorm_5 = add(x = var_611, y = var_613)[name = tensor("new_kv_unnorm_5")]; - tensor var_615 = const()[name = tensor("op_615"), val = tensor(0x1.8p+1)]; - tensor new_scale_5 = add(x = prev_scale_5, y = var_615)[name = tensor("new_scale_5")]; - tensor var_617 = sqrt(x = new_scale_5)[name = tensor("op_617")]; - tensor var_618 = real_div(x = new_kv_unnorm_5, y = var_617)[name = tensor("op_618")]; - tensor var_619_perm_0 = const()[name = tensor("op_619_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_657 = mul(x = prev_kv_5, y = sqrt_s0_5)[name = tensor("op_657")]; + tensor var_659_transpose_x_1 = const()[name = tensor("op_659_transpose_x_1"), val = tensor(true)]; + tensor var_659_transpose_y_1 = const()[name = tensor("op_659_transpose_y_1"), val = tensor(false)]; + tensor var_659 = matmul(transpose_x = var_659_transpose_x_1, transpose_y = var_659_transpose_y_1, x = k_5, y = v_5)[name = tensor("op_659")]; + tensor new_kv_unnorm_5 = add(x = var_657, y = var_659)[name = tensor("new_kv_unnorm_5")]; + tensor var_661 = const()[name = tensor("op_661"), val = tensor(0x1.8p+1)]; + tensor new_scale_5 = add(x = prev_scale_5, y = var_661)[name = tensor("new_scale_5")]; + tensor var_663 = sqrt(x = new_scale_5)[name = tensor("op_663")]; + tensor var_664 = real_div(x = new_kv_unnorm_5, y = var_663)[name = tensor("op_664")]; + tensor var_665_perm_0 = const()[name = tensor("op_665_perm_0"), val = tensor([0, 2, 1, 3])]; tensor out_15_axes_0 = const()[name = tensor("out_15_axes_0"), val = tensor([-1])]; - tensor var_619 = transpose(perm = var_619_perm_0, x = out_13)[name = tensor("transpose_41")]; - tensor out_15 = layer_norm(axes = out_15_axes_0, epsilon = var_17, x = var_619)[name = tensor("out_15")]; - tensor var_623 = const()[name = tensor("op_623"), val = tensor([1, 3, 256])]; - tensor out_17 = reshape(shape = var_623, x = out_15)[name = tensor("out_17")]; - tensor var_625 = silu(x = input_97)[name = tensor("op_625")]; - tensor input_99 = mul(x = var_625, y = out_17)[name = tensor("input_99")]; - tensor ret_out_5 = linear(bias = encoder_out_proj_2_bias, weight = encoder_out_proj_2_weight, x = input_99)[name = tensor("linear_25")]; - tensor x_15 = add(x = input_93, y = ret_out_5)[name = tensor("x_15")]; + tensor var_665 = transpose(perm = var_665_perm_0, x = out_13)[name = tensor("transpose_41")]; + tensor out_15 = layer_norm(axes = out_15_axes_0, epsilon = var_64, x = var_665)[name = tensor("out_15")]; + tensor var_669 = const()[name = tensor("op_669"), val = tensor([1, 3, 256])]; + tensor out_17 = reshape(shape = var_669, x = out_15)[name = tensor("out_17")]; + tensor var_671 = silu(x = input_99)[name = tensor("op_671")]; + tensor input_101 = mul(x = var_671, y = out_17)[name = tensor("input_101")]; + tensor ret_out_5 = linear(bias = inner_encoder_out_proj_2_bias, weight = inner_encoder_out_proj_2_weight, x = input_101)[name = tensor("linear_25")]; + tensor x_15 = add(x = input_95, y = ret_out_5)[name = tensor("x_15")]; tensor window_17_begin_0 = const()[name = tensor("window_17_begin_0"), val = tensor([2, 0, 0, 0])]; tensor window_17_end_0 = const()[name = tensor("window_17_end_0"), val = tensor([3, 1, 16, 256])]; tensor window_17_end_mask_0 = const()[name = tensor("window_17_end_mask_0"), val = tensor([false, true, true, true])]; tensor window_17_squeeze_mask_0 = const()[name = tensor("window_17_squeeze_mask_0"), val = tensor([true, false, false, false])]; tensor window_17 = slice_by_index(begin = window_17_begin_0, end = window_17_end_0, end_mask = window_17_end_mask_0, squeeze_mask = window_17_squeeze_mask_0, x = enc_conv_cache)[name = tensor("window_17")]; - tensor var_633_begin_0 = const()[name = tensor("op_633_begin_0"), val = tensor([0, 0, 0])]; - tensor var_633_end_0 = const()[name = tensor("op_633_end_0"), val = tensor([1, 1, 256])]; - tensor var_633_end_mask_0 = const()[name = tensor("op_633_end_mask_0"), val = tensor([true, false, true])]; - tensor var_633 = slice_by_index(begin = var_633_begin_0, end = var_633_end_0, end_mask = var_633_end_mask_0, x = x_15)[name = tensor("op_633")]; - tensor var_636_begin_0 = const()[name = tensor("op_636_begin_0"), val = tensor([0, 1, 0])]; - tensor var_636_end_0 = const()[name = tensor("op_636_end_0"), val = tensor([1, 16, 256])]; - tensor var_636_end_mask_0 = const()[name = tensor("op_636_end_mask_0"), val = tensor([true, true, true])]; - tensor var_636 = slice_by_index(begin = var_636_begin_0, end = var_636_end_0, end_mask = var_636_end_mask_0, x = window_17)[name = tensor("op_636")]; + tensor var_679_begin_0 = const()[name = tensor("op_679_begin_0"), val = tensor([0, 0, 0])]; + tensor var_679_end_0 = const()[name = tensor("op_679_end_0"), val = tensor([1, 1, 256])]; + tensor var_679_end_mask_0 = const()[name = tensor("op_679_end_mask_0"), val = tensor([true, false, true])]; + tensor var_679 = slice_by_index(begin = var_679_begin_0, end = var_679_end_0, end_mask = var_679_end_mask_0, x = x_15)[name = tensor("op_679")]; + tensor var_682_begin_0 = const()[name = tensor("op_682_begin_0"), val = tensor([0, 1, 0])]; + tensor var_682_end_0 = const()[name = tensor("op_682_end_0"), val = tensor([1, 16, 256])]; + tensor var_682_end_mask_0 = const()[name = tensor("op_682_end_mask_0"), val = tensor([true, true, true])]; + tensor var_682 = slice_by_index(begin = var_682_begin_0, end = var_682_end_0, end_mask = var_682_end_mask_0, x = window_17)[name = tensor("op_682")]; tensor window_19_interleave_0 = const()[name = tensor("window_19_interleave_0"), val = tensor(false)]; - tensor window_19 = concat(axis = var_26, interleave = window_19_interleave_0, values = (var_636, var_633))[name = tensor("window_19")]; - tensor var_641_begin_0 = const()[name = tensor("op_641_begin_0"), val = tensor([0, 1, 0])]; - tensor var_641_end_0 = const()[name = tensor("op_641_end_0"), val = tensor([1, 2, 256])]; - tensor var_641_end_mask_0 = const()[name = tensor("op_641_end_mask_0"), val = tensor([true, false, true])]; - tensor var_641 = slice_by_index(begin = var_641_begin_0, end = var_641_end_0, end_mask = var_641_end_mask_0, x = x_15)[name = tensor("op_641")]; - tensor var_644_begin_0 = const()[name = tensor("op_644_begin_0"), val = tensor([0, 1, 0])]; - tensor var_644_end_0 = const()[name = tensor("op_644_end_0"), val = tensor([1, 16, 256])]; - tensor var_644_end_mask_0 = const()[name = tensor("op_644_end_mask_0"), val = tensor([true, true, true])]; - tensor var_644 = slice_by_index(begin = var_644_begin_0, end = var_644_end_0, end_mask = var_644_end_mask_0, x = window_19)[name = tensor("op_644")]; + tensor window_19 = concat(axis = var_72, interleave = window_19_interleave_0, values = (var_682, var_679))[name = tensor("window_19")]; + tensor var_687_begin_0 = const()[name = tensor("op_687_begin_0"), val = tensor([0, 1, 0])]; + tensor var_687_end_0 = const()[name = tensor("op_687_end_0"), val = tensor([1, 2, 256])]; + tensor var_687_end_mask_0 = const()[name = tensor("op_687_end_mask_0"), val = tensor([true, false, true])]; + tensor var_687 = slice_by_index(begin = var_687_begin_0, end = var_687_end_0, end_mask = var_687_end_mask_0, x = x_15)[name = tensor("op_687")]; + tensor var_690_begin_0 = const()[name = tensor("op_690_begin_0"), val = tensor([0, 1, 0])]; + tensor var_690_end_0 = const()[name = tensor("op_690_end_0"), val = tensor([1, 16, 256])]; + tensor var_690_end_mask_0 = const()[name = tensor("op_690_end_mask_0"), val = tensor([true, true, true])]; + tensor var_690 = slice_by_index(begin = var_690_begin_0, end = var_690_end_0, end_mask = var_690_end_mask_0, x = window_19)[name = tensor("op_690")]; tensor window_21_interleave_0 = const()[name = tensor("window_21_interleave_0"), val = tensor(false)]; - tensor window_21 = concat(axis = var_26, interleave = window_21_interleave_0, values = (var_644, var_641))[name = tensor("window_21")]; - tensor var_649_begin_0 = const()[name = tensor("op_649_begin_0"), val = tensor([0, 2, 0])]; - tensor var_649_end_0 = const()[name = tensor("op_649_end_0"), val = tensor([1, 1, 256])]; - tensor var_649_end_mask_0 = const()[name = tensor("op_649_end_mask_0"), val = tensor([true, true, true])]; - tensor var_649 = slice_by_index(begin = var_649_begin_0, end = var_649_end_0, end_mask = var_649_end_mask_0, x = x_15)[name = tensor("op_649")]; - tensor var_652_begin_0 = const()[name = tensor("op_652_begin_0"), val = tensor([0, 1, 0])]; - tensor var_652_end_0 = const()[name = tensor("op_652_end_0"), val = tensor([1, 16, 256])]; - tensor var_652_end_mask_0 = const()[name = tensor("op_652_end_mask_0"), val = tensor([true, true, true])]; - tensor var_652 = slice_by_index(begin = var_652_begin_0, end = var_652_end_0, end_mask = var_652_end_mask_0, x = window_21)[name = tensor("op_652")]; + tensor window_21 = concat(axis = var_72, interleave = window_21_interleave_0, values = (var_690, var_687))[name = tensor("window_21")]; + tensor var_695_begin_0 = const()[name = tensor("op_695_begin_0"), val = tensor([0, 2, 0])]; + tensor var_695_end_0 = const()[name = tensor("op_695_end_0"), val = tensor([1, 1, 256])]; + tensor var_695_end_mask_0 = const()[name = tensor("op_695_end_mask_0"), val = tensor([true, true, true])]; + tensor var_695 = slice_by_index(begin = var_695_begin_0, end = var_695_end_0, end_mask = var_695_end_mask_0, x = x_15)[name = tensor("op_695")]; + tensor var_698_begin_0 = const()[name = tensor("op_698_begin_0"), val = tensor([0, 1, 0])]; + tensor var_698_end_0 = const()[name = tensor("op_698_end_0"), val = tensor([1, 16, 256])]; + tensor var_698_end_mask_0 = const()[name = tensor("op_698_end_mask_0"), val = tensor([true, true, true])]; + tensor var_698 = slice_by_index(begin = var_698_begin_0, end = var_698_end_0, end_mask = var_698_end_mask_0, x = window_21)[name = tensor("op_698")]; tensor window_23_interleave_0 = const()[name = tensor("window_23_interleave_0"), val = tensor(false)]; - tensor window_23 = concat(axis = var_26, interleave = window_23_interleave_0, values = (var_652, var_649))[name = tensor("window_23")]; - tensor input_101_interleave_0 = const()[name = tensor("input_101_interleave_0"), val = tensor(false)]; - tensor input_101 = concat(axis = var_23, interleave = input_101_interleave_0, values = (window_19, window_21, window_23))[name = tensor("input_101")]; + tensor window_23 = concat(axis = var_72, interleave = window_23_interleave_0, values = (var_698, var_695))[name = tensor("window_23")]; + tensor input_103_interleave_0 = const()[name = tensor("input_103_interleave_0"), val = tensor(false)]; + tensor input_103 = concat(axis = var_59, interleave = input_103_interleave_0, values = (window_19, window_21, window_23))[name = tensor("input_103")]; tensor x_17_axes_0 = const()[name = tensor("x_17_axes_0"), val = tensor([-1])]; - tensor x_17 = layer_norm(axes = x_17_axes_0, beta = encoder_conv_module_2_sequential_0_bias, epsilon = var_28, gamma = encoder_conv_module_2_sequential_0_weight, x = input_101)[name = tensor("x_17")]; - tensor input_103_perm_0 = const()[name = tensor("input_103_perm_0"), val = tensor([0, 2, 1])]; + tensor x_17 = layer_norm(axes = x_17_axes_0, beta = inner_encoder_conv_module_2_sequential_0_bias, epsilon = var_56, gamma = inner_encoder_conv_module_2_sequential_0_weight, x = input_103)[name = tensor("x_17")]; + tensor input_105_perm_0 = const()[name = tensor("input_105_perm_0"), val = tensor([0, 2, 1])]; tensor inputs_23_pad_type_0 = const()[name = tensor("inputs_23_pad_type_0"), val = tensor("valid")]; tensor inputs_23_strides_0 = const()[name = tensor("inputs_23_strides_0"), val = tensor([1])]; tensor inputs_23_pad_0 = const()[name = tensor("inputs_23_pad_0"), val = tensor([0, 0])]; tensor inputs_23_dilations_0 = const()[name = tensor("inputs_23_dilations_0"), val = tensor([1])]; tensor inputs_23_groups_0 = const()[name = tensor("inputs_23_groups_0"), val = tensor(1)]; - tensor input_103 = transpose(perm = input_103_perm_0, x = x_17)[name = tensor("transpose_40")]; - tensor inputs_23 = conv(bias = encoder_conv_module_2_sequential_2_conv_bias, dilations = inputs_23_dilations_0, groups = inputs_23_groups_0, pad = inputs_23_pad_0, pad_type = inputs_23_pad_type_0, strides = inputs_23_strides_0, weight = encoder_conv_module_2_sequential_2_conv_weight, x = input_103)[name = tensor("inputs_23")]; - tensor var_677_split_sizes_0 = const()[name = tensor("op_677_split_sizes_0"), val = tensor([256, 256])]; - tensor var_677_axis_0 = const()[name = tensor("op_677_axis_0"), val = tensor(1)]; - tensor var_677_0, tensor var_677_1 = split(axis = var_677_axis_0, split_sizes = var_677_split_sizes_0, x = inputs_23)[name = tensor("op_677")]; - tensor var_679 = sigmoid(x = var_677_1)[name = tensor("op_679")]; - tensor inputs_25 = mul(x = var_677_0, y = var_679)[name = tensor("inputs_25")]; + tensor input_105 = transpose(perm = input_105_perm_0, x = x_17)[name = tensor("transpose_40")]; + tensor inputs_23 = conv(bias = inner_encoder_conv_module_2_sequential_2_conv_bias, dilations = inputs_23_dilations_0, groups = inputs_23_groups_0, pad = inputs_23_pad_0, pad_type = inputs_23_pad_type_0, strides = inputs_23_strides_0, weight = inner_encoder_conv_module_2_sequential_2_conv_weight, x = input_105)[name = tensor("inputs_23")]; + tensor var_723_split_sizes_0 = const()[name = tensor("op_723_split_sizes_0"), val = tensor([256, 256])]; + tensor var_723_axis_0 = const()[name = tensor("op_723_axis_0"), val = tensor(1)]; + tensor var_723_0, tensor var_723_1 = split(axis = var_723_axis_0, split_sizes = var_723_split_sizes_0, x = inputs_23)[name = tensor("op_723")]; + tensor var_725 = sigmoid(x = var_723_1)[name = tensor("op_725")]; + tensor inputs_25 = mul(x = var_723_0, y = var_725)[name = tensor("inputs_25")]; tensor outputs_aug_5_pad_type_0 = const()[name = tensor("outputs_aug_5_pad_type_0"), val = tensor("custom")]; tensor outputs_aug_5_pad_0 = const()[name = tensor("outputs_aug_5_pad_0"), val = tensor([15, 15])]; tensor outputs_aug_5_groups_0 = const()[name = tensor("outputs_aug_5_groups_0"), val = tensor(256)]; tensor outputs_aug_5_strides_0 = const()[name = tensor("outputs_aug_5_strides_0"), val = tensor([1])]; tensor outputs_aug_5_dilations_0 = const()[name = tensor("outputs_aug_5_dilations_0"), val = tensor([1])]; - tensor outputs_aug_5 = conv(dilations = outputs_aug_5_dilations_0, groups = outputs_aug_5_groups_0, pad = outputs_aug_5_pad_0, pad_type = outputs_aug_5_pad_type_0, strides = outputs_aug_5_strides_0, weight = encoder_conv_module_2_sequential_4_conv_weight, x = inputs_25)[name = tensor("outputs_aug_5")]; - tensor input_105_begin_0 = const()[name = tensor("input_105_begin_0"), val = tensor([0, 0, 0])]; - tensor input_105_end_0 = const()[name = tensor("input_105_end_0"), val = tensor([3, 256, 16])]; - tensor input_105_end_mask_0 = const()[name = tensor("input_105_end_mask_0"), val = tensor([true, true, false])]; - tensor input_105 = slice_by_index(begin = input_105_begin_0, end = input_105_end_0, end_mask = input_105_end_mask_0, x = outputs_aug_5)[name = tensor("input_105")]; - tensor inputs_27 = batch_norm(beta = encoder_conv_module_2_sequential_5_bias, epsilon = var_28, gamma = encoder_conv_module_2_sequential_5_weight, mean = encoder_conv_module_2_sequential_5_running_mean, variance = encoder_conv_module_2_sequential_5_running_var, x = input_105)[name = tensor("inputs_27")]; - tensor input_107 = silu(x = inputs_27)[name = tensor("input_107")]; - tensor input_109_pad_type_0 = const()[name = tensor("input_109_pad_type_0"), val = tensor("valid")]; - tensor input_109_strides_0 = const()[name = tensor("input_109_strides_0"), val = tensor([1])]; - tensor input_109_pad_0 = const()[name = tensor("input_109_pad_0"), val = tensor([0, 0])]; - tensor input_109_dilations_0 = const()[name = tensor("input_109_dilations_0"), val = tensor([1])]; - tensor input_109_groups_0 = const()[name = tensor("input_109_groups_0"), val = tensor(1)]; - tensor input_109 = conv(bias = encoder_conv_module_2_sequential_7_conv_bias, dilations = input_109_dilations_0, groups = input_109_groups_0, pad = input_109_pad_0, pad_type = input_109_pad_type_0, strides = input_109_strides_0, weight = encoder_conv_module_2_sequential_7_conv_weight, x = input_107)[name = tensor("input_109")]; + tensor outputs_aug_5 = conv(dilations = outputs_aug_5_dilations_0, groups = outputs_aug_5_groups_0, pad = outputs_aug_5_pad_0, pad_type = outputs_aug_5_pad_type_0, strides = outputs_aug_5_strides_0, weight = inner_encoder_conv_module_2_sequential_4_conv_weight, x = inputs_25)[name = tensor("outputs_aug_5")]; + tensor input_107_begin_0 = const()[name = tensor("input_107_begin_0"), val = tensor([0, 0, 0])]; + tensor input_107_end_0 = const()[name = tensor("input_107_end_0"), val = tensor([3, 256, 16])]; + tensor input_107_end_mask_0 = const()[name = tensor("input_107_end_mask_0"), val = tensor([true, true, false])]; + tensor input_107 = slice_by_index(begin = input_107_begin_0, end = input_107_end_0, end_mask = input_107_end_mask_0, x = outputs_aug_5)[name = tensor("input_107")]; + tensor inputs_27 = batch_norm(beta = inner_encoder_conv_module_2_sequential_5_bias, epsilon = var_56, gamma = inner_encoder_conv_module_2_sequential_5_weight, mean = inner_encoder_conv_module_2_sequential_5_running_mean, variance = inner_encoder_conv_module_2_sequential_5_running_var, x = input_107)[name = tensor("inputs_27")]; + tensor input_109 = silu(x = inputs_27)[name = tensor("input_109")]; + tensor input_111_pad_type_0 = const()[name = tensor("input_111_pad_type_0"), val = tensor("valid")]; + tensor input_111_strides_0 = const()[name = tensor("input_111_strides_0"), val = tensor([1])]; + tensor input_111_pad_0 = const()[name = tensor("input_111_pad_0"), val = tensor([0, 0])]; + tensor input_111_dilations_0 = const()[name = tensor("input_111_dilations_0"), val = tensor([1])]; + tensor input_111_groups_0 = const()[name = tensor("input_111_groups_0"), val = tensor(1)]; + tensor input_111 = conv(bias = inner_encoder_conv_module_2_sequential_7_conv_bias, dilations = input_111_dilations_0, groups = input_111_groups_0, pad = input_111_pad_0, pad_type = input_111_pad_type_0, strides = input_111_strides_0, weight = inner_encoder_conv_module_2_sequential_7_conv_weight, x = input_109)[name = tensor("input_111")]; tensor conv_out_5_perm_0 = const()[name = tensor("conv_out_5_perm_0"), val = tensor([0, 2, 1])]; - tensor var_710_begin_0 = const()[name = tensor("op_710_begin_0"), val = tensor([0, -1, 0])]; - tensor var_710_end_0 = const()[name = tensor("op_710_end_0"), val = tensor([3, 16, 256])]; - tensor var_710_end_mask_0 = const()[name = tensor("op_710_end_mask_0"), val = tensor([true, true, true])]; - tensor conv_out_5 = transpose(perm = conv_out_5_perm_0, x = input_109)[name = tensor("transpose_39")]; - tensor var_710 = slice_by_index(begin = var_710_begin_0, end = var_710_end_0, end_mask = var_710_end_mask_0, x = conv_out_5)[name = tensor("op_710")]; - tensor var_712_perm_0 = const()[name = tensor("op_712_perm_0"), val = tensor([1, 0, 2])]; - tensor var_712 = transpose(perm = var_712_perm_0, x = var_710)[name = tensor("transpose_38")]; - tensor input_111 = add(x = x_15, y = var_712)[name = tensor("input_111")]; - tensor input_113_axes_0 = const()[name = tensor("input_113_axes_0"), val = tensor([-1])]; - tensor input_113 = layer_norm(axes = input_113_axes_0, beta = encoder_ffn2_2_module_sequential_0_bias, epsilon = var_28, gamma = encoder_ffn2_2_module_sequential_0_weight, x = input_111)[name = tensor("input_113")]; - tensor inputs_29 = linear(bias = encoder_ffn2_2_module_sequential_1_linear_bias, weight = encoder_ffn2_2_module_sequential_1_linear_weight, x = input_113)[name = tensor("linear_26")]; - tensor input_115 = silu(x = inputs_29)[name = tensor("input_115")]; - tensor input_119 = linear(bias = encoder_ffn2_2_module_sequential_4_linear_bias, weight = encoder_ffn2_2_module_sequential_4_linear_weight, x = input_115)[name = tensor("linear_27")]; - tensor var_735 = const()[name = tensor("op_735"), val = tensor(0x1p-1)]; - tensor var_736 = mul(x = input_119, y = var_735)[name = tensor("op_736")]; - tensor input_121 = add(x = var_736, y = input_111)[name = tensor("input_121")]; - tensor input_123_axes_0 = const()[name = tensor("input_123_axes_0"), val = tensor([-1])]; - tensor input_123 = layer_norm(axes = input_123_axes_0, beta = encoder_layer_norm_2_bias, epsilon = var_28, gamma = encoder_layer_norm_2_weight, x = input_121)[name = tensor("input_123")]; + tensor var_756_begin_0 = const()[name = tensor("op_756_begin_0"), val = tensor([0, -1, 0])]; + tensor var_756_end_0 = const()[name = tensor("op_756_end_0"), val = tensor([3, 16, 256])]; + tensor var_756_end_mask_0 = const()[name = tensor("op_756_end_mask_0"), val = tensor([true, true, true])]; + tensor conv_out_5 = transpose(perm = conv_out_5_perm_0, x = input_111)[name = tensor("transpose_39")]; + tensor var_756 = slice_by_index(begin = var_756_begin_0, end = var_756_end_0, end_mask = var_756_end_mask_0, x = conv_out_5)[name = tensor("op_756")]; + tensor var_758_perm_0 = const()[name = tensor("op_758_perm_0"), val = tensor([1, 0, 2])]; + tensor var_758 = transpose(perm = var_758_perm_0, x = var_756)[name = tensor("transpose_38")]; + tensor input_113 = add(x = x_15, y = var_758)[name = tensor("input_113")]; + tensor input_115_axes_0 = const()[name = tensor("input_115_axes_0"), val = tensor([-1])]; + tensor input_115 = layer_norm(axes = input_115_axes_0, beta = inner_encoder_ffn2_2_module_sequential_0_bias, epsilon = var_56, gamma = inner_encoder_ffn2_2_module_sequential_0_weight, x = input_113)[name = tensor("input_115")]; + tensor inputs_29 = linear(bias = inner_encoder_ffn2_2_module_sequential_1_linear_bias, weight = inner_encoder_ffn2_2_module_sequential_1_linear_weight, x = input_115)[name = tensor("linear_26")]; + tensor input_117 = silu(x = inputs_29)[name = tensor("input_117")]; + tensor input_121 = linear(bias = inner_encoder_ffn2_2_module_sequential_4_linear_bias, weight = inner_encoder_ffn2_2_module_sequential_4_linear_weight, x = input_117)[name = tensor("linear_27")]; + tensor var_781 = const()[name = tensor("op_781"), val = tensor(0x1p-1)]; + tensor var_782 = mul(x = input_121, y = var_781)[name = tensor("op_782")]; + tensor input_123 = add(x = var_782, y = input_113)[name = tensor("input_123")]; tensor input_125_axes_0 = const()[name = tensor("input_125_axes_0"), val = tensor([-1])]; - tensor input_125 = layer_norm(axes = input_125_axes_0, beta = encoder_ffn1_3_module_sequential_0_bias, epsilon = var_28, gamma = encoder_ffn1_3_module_sequential_0_weight, x = input_123)[name = tensor("input_125")]; - tensor inputs_31 = linear(bias = encoder_ffn1_3_module_sequential_1_linear_bias, weight = encoder_ffn1_3_module_sequential_1_linear_weight, x = input_125)[name = tensor("linear_28")]; - tensor input_127 = silu(x = inputs_31)[name = tensor("input_127")]; - tensor input_131 = linear(bias = encoder_ffn1_3_module_sequential_4_linear_bias, weight = encoder_ffn1_3_module_sequential_4_linear_weight, x = input_127)[name = tensor("linear_29")]; - tensor var_765 = const()[name = tensor("op_765"), val = tensor(0x1p-1)]; - tensor var_766 = mul(x = input_131, y = var_765)[name = tensor("op_766")]; - tensor input_133 = add(x = var_766, y = input_123)[name = tensor("input_133")]; + tensor input_125 = layer_norm(axes = input_125_axes_0, beta = inner_encoder_layer_norm_2_bias, epsilon = var_56, gamma = inner_encoder_layer_norm_2_weight, x = input_123)[name = tensor("input_125")]; + tensor input_127_axes_0 = const()[name = tensor("input_127_axes_0"), val = tensor([-1])]; + tensor input_127 = layer_norm(axes = input_127_axes_0, beta = inner_encoder_ffn1_3_module_sequential_0_bias, epsilon = var_56, gamma = inner_encoder_ffn1_3_module_sequential_0_weight, x = input_125)[name = tensor("input_127")]; + tensor inputs_31 = linear(bias = inner_encoder_ffn1_3_module_sequential_1_linear_bias, weight = inner_encoder_ffn1_3_module_sequential_1_linear_weight, x = input_127)[name = tensor("linear_28")]; + tensor input_129 = silu(x = inputs_31)[name = tensor("input_129")]; + tensor input_133 = linear(bias = inner_encoder_ffn1_3_module_sequential_4_linear_bias, weight = inner_encoder_ffn1_3_module_sequential_4_linear_weight, x = input_129)[name = tensor("linear_29")]; + tensor var_811 = const()[name = tensor("op_811"), val = tensor(0x1p-1)]; + tensor var_812 = mul(x = input_133, y = var_811)[name = tensor("op_812")]; + tensor input_135 = add(x = var_812, y = input_125)[name = tensor("input_135")]; tensor x_19_axes_0 = const()[name = tensor("x_19_axes_0"), val = tensor([-1])]; - tensor x_19 = layer_norm(axes = x_19_axes_0, beta = encoder_ret_lns_3_bias, epsilon = var_28, gamma = encoder_ret_lns_3_weight, x = input_133)[name = tensor("x_19")]; + tensor x_19 = layer_norm(axes = x_19_axes_0, beta = inner_encoder_ret_lns_3_bias, epsilon = var_56, gamma = inner_encoder_ret_lns_3_weight, x = input_135)[name = tensor("x_19")]; tensor prev_kv_7_begin_0 = const()[name = tensor("prev_kv_7_begin_0"), val = tensor([3, 0, 0, 0, 0])]; tensor prev_kv_7_end_0 = const()[name = tensor("prev_kv_7_end_0"), val = tensor([4, 1, 4, 64, 64])]; tensor prev_kv_7_end_mask_0 = const()[name = tensor("prev_kv_7_end_mask_0"), val = tensor([false, true, true, true, true])]; @@ -740,199 +758,192 @@ program(1.0) tensor prev_scale_7_end_mask_0 = const()[name = tensor("prev_scale_7_end_mask_0"), val = tensor([false, true])]; tensor prev_scale_7_squeeze_mask_0 = const()[name = tensor("prev_scale_7_squeeze_mask_0"), val = tensor([true, false])]; tensor prev_scale_7 = slice_by_index(begin = prev_scale_7_begin_0, end = prev_scale_7_end_0, end_mask = prev_scale_7_end_mask_0, squeeze_mask = prev_scale_7_squeeze_mask_0, x = enc_scale)[name = tensor("prev_scale_7")]; - tensor var_780 = linear(bias = encoder_q_proj_3_bias, weight = encoder_q_proj_3_weight, x = x_19)[name = tensor("linear_30")]; - tensor var_781 = const()[name = tensor("op_781"), val = tensor([1, 3, 4, 64])]; - tensor var_782 = reshape(shape = var_781, x = var_780)[name = tensor("op_782")]; + tensor var_826 = linear(bias = inner_encoder_q_proj_3_bias, weight = inner_encoder_q_proj_3_weight, x = x_19)[name = tensor("linear_30")]; + tensor var_827 = const()[name = tensor("op_827"), val = tensor([1, 3, 4, 64])]; + tensor var_828 = reshape(shape = var_827, x = var_826)[name = tensor("op_828")]; tensor q_7_perm_0 = const()[name = tensor("q_7_perm_0"), val = tensor([0, 2, 1, 3])]; - tensor var_786 = linear(bias = encoder_k_proj_3_bias, weight = encoder_k_proj_3_weight, x = x_19)[name = tensor("linear_31")]; - tensor var_787 = const()[name = tensor("op_787"), val = tensor(0x1p-3)]; - tensor var_788 = mul(x = var_786, y = var_787)[name = tensor("op_788")]; - tensor var_789 = const()[name = tensor("op_789"), val = tensor([1, 3, 4, 64])]; - tensor var_790 = reshape(shape = var_789, x = var_788)[name = tensor("op_790")]; + tensor var_832 = linear(bias = inner_encoder_k_proj_3_bias, weight = inner_encoder_k_proj_3_weight, x = x_19)[name = tensor("linear_31")]; + tensor var_833 = const()[name = tensor("op_833"), val = tensor(0x1p-3)]; + tensor var_834 = mul(x = var_832, y = var_833)[name = tensor("op_834")]; + tensor var_835 = const()[name = tensor("op_835"), val = tensor([1, 3, 4, 64])]; + tensor var_836 = reshape(shape = var_835, x = var_834)[name = tensor("op_836")]; tensor k_7_perm_0 = const()[name = tensor("k_7_perm_0"), val = tensor([0, 2, 1, 3])]; - tensor var_794 = linear(bias = encoder_v_proj_3_bias, weight = encoder_v_proj_3_weight, x = x_19)[name = tensor("linear_32")]; - tensor var_795 = const()[name = tensor("op_795"), val = tensor([1, 3, 4, 64])]; - tensor var_796 = reshape(shape = var_795, x = var_794)[name = tensor("op_796")]; + tensor var_840 = linear(bias = inner_encoder_v_proj_3_bias, weight = inner_encoder_v_proj_3_weight, x = x_19)[name = tensor("linear_32")]; + tensor var_841 = const()[name = tensor("op_841"), val = tensor([1, 3, 4, 64])]; + tensor var_842 = reshape(shape = var_841, x = var_840)[name = tensor("op_842")]; tensor v_7_perm_0 = const()[name = tensor("v_7_perm_0"), val = tensor([0, 2, 1, 3])]; - tensor input_137 = linear(bias = encoder_g_proj_3_bias, weight = encoder_g_proj_3_weight, x = x_19)[name = tensor("linear_33")]; + tensor input_139 = linear(bias = inner_encoder_g_proj_3_bias, weight = inner_encoder_g_proj_3_weight, x = x_19)[name = tensor("linear_33")]; tensor sqrt_s0_7 = sqrt(x = prev_scale_7)[name = tensor("sqrt_s0_7")]; - tensor s_t_7 = add(x = prev_scale_7, y = encoder__t_index)[name = tensor("s_t_7")]; + tensor s_t_7 = add(x = prev_scale_7, y = inner_encoder__t_index)[name = tensor("s_t_7")]; tensor sqrt_s_t_7 = sqrt(x = s_t_7)[name = tensor("sqrt_s_t_7")]; tensor qk_7_transpose_x_1 = const()[name = tensor("qk_7_transpose_x_1"), val = tensor(false)]; tensor qk_7_transpose_y_1 = const()[name = tensor("qk_7_transpose_y_1"), val = tensor(true)]; - tensor k_7 = transpose(perm = k_7_perm_0, x = var_790)[name = tensor("transpose_36")]; - tensor q_7 = transpose(perm = q_7_perm_0, x = var_782)[name = tensor("transpose_37")]; + tensor k_7 = transpose(perm = k_7_perm_0, x = var_836)[name = tensor("transpose_36")]; + tensor q_7 = transpose(perm = q_7_perm_0, x = var_828)[name = tensor("transpose_37")]; tensor qk_7 = matmul(transpose_x = qk_7_transpose_x_1, transpose_y = qk_7_transpose_y_1, x = q_7, y = k_7)[name = tensor("qk_7")]; - tensor var_806 = const()[name = tensor("op_806"), val = tensor([3, 1])]; - tensor var_807 = reshape(shape = var_806, x = sqrt_s_t_7)[name = tensor("op_807")]; - tensor M_7 = real_div(x = encoder__causal_mask, y = var_807)[name = tensor("M_7")]; - tensor var_809 = mul(x = qk_7, y = M_7)[name = tensor("op_809")]; + tensor var_852 = const()[name = tensor("op_852"), val = tensor([3, 1])]; + tensor var_853 = reshape(shape = var_852, x = sqrt_s_t_7)[name = tensor("op_853")]; + tensor M_7 = real_div(x = inner_encoder__causal_mask, y = var_853)[name = tensor("M_7")]; + tensor var_855 = mul(x = qk_7, y = M_7)[name = tensor("op_855")]; tensor inner_7_transpose_x_0 = const()[name = tensor("inner_7_transpose_x_0"), val = tensor(false)]; tensor inner_7_transpose_y_0 = const()[name = tensor("inner_7_transpose_y_0"), val = tensor(false)]; - tensor v_7 = transpose(perm = v_7_perm_0, x = var_796)[name = tensor("transpose_35")]; - tensor inner_7 = matmul(transpose_x = inner_7_transpose_x_0, transpose_y = inner_7_transpose_y_0, x = var_809, y = v_7)[name = tensor("inner_7")]; - tensor var_811_transpose_x_0 = const()[name = tensor("op_811_transpose_x_0"), val = tensor(false)]; - tensor var_811_transpose_y_0 = const()[name = tensor("op_811_transpose_y_0"), val = tensor(false)]; - tensor var_811 = matmul(transpose_x = var_811_transpose_x_0, transpose_y = var_811_transpose_y_0, x = q_7, y = prev_kv_7)[name = tensor("op_811")]; - tensor var_812 = real_div(x = sqrt_s0_7, y = sqrt_s_t_7)[name = tensor("op_812")]; - tensor var_813 = const()[name = tensor("op_813"), val = tensor([1, 1, 3, 1])]; - tensor var_814 = reshape(shape = var_813, x = var_812)[name = tensor("op_814")]; - tensor cross_7 = mul(x = var_811, y = var_814)[name = tensor("cross_7")]; + tensor v_7 = transpose(perm = v_7_perm_0, x = var_842)[name = tensor("transpose_35")]; + tensor inner_7 = matmul(transpose_x = inner_7_transpose_x_0, transpose_y = inner_7_transpose_y_0, x = var_855, y = v_7)[name = tensor("inner_7")]; + tensor var_857_transpose_x_0 = const()[name = tensor("op_857_transpose_x_0"), val = tensor(false)]; + tensor var_857_transpose_y_0 = const()[name = tensor("op_857_transpose_y_0"), val = tensor(false)]; + tensor var_857 = matmul(transpose_x = var_857_transpose_x_0, transpose_y = var_857_transpose_y_0, x = q_7, y = prev_kv_7)[name = tensor("op_857")]; + tensor var_858 = real_div(x = sqrt_s0_7, y = sqrt_s_t_7)[name = tensor("op_858")]; + tensor var_859 = const()[name = tensor("op_859"), val = tensor([1, 1, 3, 1])]; + tensor var_860 = reshape(shape = var_859, x = var_858)[name = tensor("op_860")]; + tensor cross_7 = mul(x = var_857, y = var_860)[name = tensor("cross_7")]; tensor out_19 = add(x = inner_7, y = cross_7)[name = tensor("out_19")]; - tensor var_817 = mul(x = prev_kv_7, y = sqrt_s0_7)[name = tensor("op_817")]; - tensor var_819_transpose_x_1 = const()[name = tensor("op_819_transpose_x_1"), val = tensor(true)]; - tensor var_819_transpose_y_1 = const()[name = tensor("op_819_transpose_y_1"), val = tensor(false)]; - tensor var_819 = matmul(transpose_x = var_819_transpose_x_1, transpose_y = var_819_transpose_y_1, x = k_7, y = v_7)[name = tensor("op_819")]; - tensor new_kv_unnorm_7 = add(x = var_817, y = var_819)[name = tensor("new_kv_unnorm_7")]; - tensor var_821 = const()[name = tensor("op_821"), val = tensor(0x1.8p+1)]; - tensor new_scale_7 = add(x = prev_scale_7, y = var_821)[name = tensor("new_scale_7")]; - tensor var_823 = sqrt(x = new_scale_7)[name = tensor("op_823")]; - tensor nkv_1 = real_div(x = new_kv_unnorm_7, y = var_823)[name = tensor("nkv_1")]; - tensor var_825_perm_0 = const()[name = tensor("op_825_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_863 = mul(x = prev_kv_7, y = sqrt_s0_7)[name = tensor("op_863")]; + tensor var_865_transpose_x_1 = const()[name = tensor("op_865_transpose_x_1"), val = tensor(true)]; + tensor var_865_transpose_y_1 = const()[name = tensor("op_865_transpose_y_1"), val = tensor(false)]; + tensor var_865 = matmul(transpose_x = var_865_transpose_x_1, transpose_y = var_865_transpose_y_1, x = k_7, y = v_7)[name = tensor("op_865")]; + tensor new_kv_unnorm_7 = add(x = var_863, y = var_865)[name = tensor("new_kv_unnorm_7")]; + tensor var_867 = const()[name = tensor("op_867"), val = tensor(0x1.8p+1)]; + tensor new_scale_7 = add(x = prev_scale_7, y = var_867)[name = tensor("new_scale_7")]; + tensor var_869 = sqrt(x = new_scale_7)[name = tensor("op_869")]; + tensor nkv_1 = real_div(x = new_kv_unnorm_7, y = var_869)[name = tensor("nkv_1")]; + tensor var_871_perm_0 = const()[name = tensor("op_871_perm_0"), val = tensor([0, 2, 1, 3])]; tensor out_21_axes_0 = const()[name = tensor("out_21_axes_0"), val = tensor([-1])]; - tensor var_825 = transpose(perm = var_825_perm_0, x = out_19)[name = tensor("transpose_34")]; - tensor out_21 = layer_norm(axes = out_21_axes_0, epsilon = var_17, x = var_825)[name = tensor("out_21")]; - tensor var_829 = const()[name = tensor("op_829"), val = tensor([1, 3, 256])]; - tensor out_23 = reshape(shape = var_829, x = out_21)[name = tensor("out_23")]; - tensor var_831 = silu(x = input_137)[name = tensor("op_831")]; - tensor input_139 = mul(x = var_831, y = out_23)[name = tensor("input_139")]; - tensor ret_out_7 = linear(bias = encoder_out_proj_3_bias, weight = encoder_out_proj_3_weight, x = input_139)[name = tensor("linear_34")]; - tensor x_21 = add(x = input_133, y = ret_out_7)[name = tensor("x_21")]; + tensor var_871 = transpose(perm = var_871_perm_0, x = out_19)[name = tensor("transpose_34")]; + tensor out_21 = layer_norm(axes = out_21_axes_0, epsilon = var_64, x = var_871)[name = tensor("out_21")]; + tensor var_875 = const()[name = tensor("op_875"), val = tensor([1, 3, 256])]; + tensor out_23 = reshape(shape = var_875, x = out_21)[name = tensor("out_23")]; + tensor var_877 = silu(x = input_139)[name = tensor("op_877")]; + tensor input_141 = mul(x = var_877, y = out_23)[name = tensor("input_141")]; + tensor ret_out_7 = linear(bias = inner_encoder_out_proj_3_bias, weight = inner_encoder_out_proj_3_weight, x = input_141)[name = tensor("linear_34")]; + tensor x_21 = add(x = input_135, y = ret_out_7)[name = tensor("x_21")]; tensor window_25_begin_0 = const()[name = tensor("window_25_begin_0"), val = tensor([3, 0, 0, 0])]; tensor window_25_end_0 = const()[name = tensor("window_25_end_0"), val = tensor([4, 1, 16, 256])]; tensor window_25_end_mask_0 = const()[name = tensor("window_25_end_mask_0"), val = tensor([false, true, true, true])]; tensor window_25_squeeze_mask_0 = const()[name = tensor("window_25_squeeze_mask_0"), val = tensor([true, false, false, false])]; tensor window_25 = slice_by_index(begin = window_25_begin_0, end = window_25_end_0, end_mask = window_25_end_mask_0, squeeze_mask = window_25_squeeze_mask_0, x = enc_conv_cache)[name = tensor("window_25")]; - tensor var_839_begin_0 = const()[name = tensor("op_839_begin_0"), val = tensor([0, 0, 0])]; - tensor var_839_end_0 = const()[name = tensor("op_839_end_0"), val = tensor([1, 1, 256])]; - tensor var_839_end_mask_0 = const()[name = tensor("op_839_end_mask_0"), val = tensor([true, false, true])]; - tensor var_839 = slice_by_index(begin = var_839_begin_0, end = var_839_end_0, end_mask = var_839_end_mask_0, x = x_21)[name = tensor("op_839")]; - tensor var_842_begin_0 = const()[name = tensor("op_842_begin_0"), val = tensor([0, 1, 0])]; - tensor var_842_end_0 = const()[name = tensor("op_842_end_0"), val = tensor([1, 16, 256])]; - tensor var_842_end_mask_0 = const()[name = tensor("op_842_end_mask_0"), val = tensor([true, true, true])]; - tensor var_842 = slice_by_index(begin = var_842_begin_0, end = var_842_end_0, end_mask = var_842_end_mask_0, x = window_25)[name = tensor("op_842")]; + tensor var_885_begin_0 = const()[name = tensor("op_885_begin_0"), val = tensor([0, 0, 0])]; + tensor var_885_end_0 = const()[name = tensor("op_885_end_0"), val = tensor([1, 1, 256])]; + tensor var_885_end_mask_0 = const()[name = tensor("op_885_end_mask_0"), val = tensor([true, false, true])]; + tensor var_885 = slice_by_index(begin = var_885_begin_0, end = var_885_end_0, end_mask = var_885_end_mask_0, x = x_21)[name = tensor("op_885")]; + tensor var_888_begin_0 = const()[name = tensor("op_888_begin_0"), val = tensor([0, 1, 0])]; + tensor var_888_end_0 = const()[name = tensor("op_888_end_0"), val = tensor([1, 16, 256])]; + tensor var_888_end_mask_0 = const()[name = tensor("op_888_end_mask_0"), val = tensor([true, true, true])]; + tensor var_888 = slice_by_index(begin = var_888_begin_0, end = var_888_end_0, end_mask = var_888_end_mask_0, x = window_25)[name = tensor("op_888")]; tensor window_27_interleave_0 = const()[name = tensor("window_27_interleave_0"), val = tensor(false)]; - tensor window_27 = concat(axis = var_26, interleave = window_27_interleave_0, values = (var_842, var_839))[name = tensor("window_27")]; - tensor var_847_begin_0 = const()[name = tensor("op_847_begin_0"), val = tensor([0, 1, 0])]; - tensor var_847_end_0 = const()[name = tensor("op_847_end_0"), val = tensor([1, 2, 256])]; - tensor var_847_end_mask_0 = const()[name = tensor("op_847_end_mask_0"), val = tensor([true, false, true])]; - tensor var_847 = slice_by_index(begin = var_847_begin_0, end = var_847_end_0, end_mask = var_847_end_mask_0, x = x_21)[name = tensor("op_847")]; - tensor var_850_begin_0 = const()[name = tensor("op_850_begin_0"), val = tensor([0, 1, 0])]; - tensor var_850_end_0 = const()[name = tensor("op_850_end_0"), val = tensor([1, 16, 256])]; - tensor var_850_end_mask_0 = const()[name = tensor("op_850_end_mask_0"), val = tensor([true, true, true])]; - tensor var_850 = slice_by_index(begin = var_850_begin_0, end = var_850_end_0, end_mask = var_850_end_mask_0, x = window_27)[name = tensor("op_850")]; + tensor window_27 = concat(axis = var_72, interleave = window_27_interleave_0, values = (var_888, var_885))[name = tensor("window_27")]; + tensor var_893_begin_0 = const()[name = tensor("op_893_begin_0"), val = tensor([0, 1, 0])]; + tensor var_893_end_0 = const()[name = tensor("op_893_end_0"), val = tensor([1, 2, 256])]; + tensor var_893_end_mask_0 = const()[name = tensor("op_893_end_mask_0"), val = tensor([true, false, true])]; + tensor var_893 = slice_by_index(begin = var_893_begin_0, end = var_893_end_0, end_mask = var_893_end_mask_0, x = x_21)[name = tensor("op_893")]; + tensor var_896_begin_0 = const()[name = tensor("op_896_begin_0"), val = tensor([0, 1, 0])]; + tensor var_896_end_0 = const()[name = tensor("op_896_end_0"), val = tensor([1, 16, 256])]; + tensor var_896_end_mask_0 = const()[name = tensor("op_896_end_mask_0"), val = tensor([true, true, true])]; + tensor var_896 = slice_by_index(begin = var_896_begin_0, end = var_896_end_0, end_mask = var_896_end_mask_0, x = window_27)[name = tensor("op_896")]; tensor window_29_interleave_0 = const()[name = tensor("window_29_interleave_0"), val = tensor(false)]; - tensor window_29 = concat(axis = var_26, interleave = window_29_interleave_0, values = (var_850, var_847))[name = tensor("window_29")]; - tensor var_855_begin_0 = const()[name = tensor("op_855_begin_0"), val = tensor([0, 2, 0])]; - tensor var_855_end_0 = const()[name = tensor("op_855_end_0"), val = tensor([1, 1, 256])]; - tensor var_855_end_mask_0 = const()[name = tensor("op_855_end_mask_0"), val = tensor([true, true, true])]; - tensor var_855 = slice_by_index(begin = var_855_begin_0, end = var_855_end_0, end_mask = var_855_end_mask_0, x = x_21)[name = tensor("op_855")]; - tensor var_858_begin_0 = const()[name = tensor("op_858_begin_0"), val = tensor([0, 1, 0])]; - tensor var_858_end_0 = const()[name = tensor("op_858_end_0"), val = tensor([1, 16, 256])]; - tensor var_858_end_mask_0 = const()[name = tensor("op_858_end_mask_0"), val = tensor([true, true, true])]; - tensor var_858 = slice_by_index(begin = var_858_begin_0, end = var_858_end_0, end_mask = var_858_end_mask_0, x = window_29)[name = tensor("op_858")]; + tensor window_29 = concat(axis = var_72, interleave = window_29_interleave_0, values = (var_896, var_893))[name = tensor("window_29")]; + tensor var_901_begin_0 = const()[name = tensor("op_901_begin_0"), val = tensor([0, 2, 0])]; + tensor var_901_end_0 = const()[name = tensor("op_901_end_0"), val = tensor([1, 1, 256])]; + tensor var_901_end_mask_0 = const()[name = tensor("op_901_end_mask_0"), val = tensor([true, true, true])]; + tensor var_901 = slice_by_index(begin = var_901_begin_0, end = var_901_end_0, end_mask = var_901_end_mask_0, x = x_21)[name = tensor("op_901")]; + tensor var_904_begin_0 = const()[name = tensor("op_904_begin_0"), val = tensor([0, 1, 0])]; + tensor var_904_end_0 = const()[name = tensor("op_904_end_0"), val = tensor([1, 16, 256])]; + tensor var_904_end_mask_0 = const()[name = tensor("op_904_end_mask_0"), val = tensor([true, true, true])]; + tensor var_904 = slice_by_index(begin = var_904_begin_0, end = var_904_end_0, end_mask = var_904_end_mask_0, x = window_29)[name = tensor("op_904")]; tensor window_interleave_0 = const()[name = tensor("window_interleave_0"), val = tensor(false)]; - tensor window = concat(axis = var_26, interleave = window_interleave_0, values = (var_858, var_855))[name = tensor("window")]; - tensor input_141_interleave_0 = const()[name = tensor("input_141_interleave_0"), val = tensor(false)]; - tensor input_141 = concat(axis = var_23, interleave = input_141_interleave_0, values = (window_27, window_29, window))[name = tensor("input_141")]; + tensor window = concat(axis = var_72, interleave = window_interleave_0, values = (var_904, var_901))[name = tensor("window")]; + tensor input_143_interleave_0 = const()[name = tensor("input_143_interleave_0"), val = tensor(false)]; + tensor input_143 = concat(axis = var_59, interleave = input_143_interleave_0, values = (window_27, window_29, window))[name = tensor("input_143")]; tensor x_23_axes_0 = const()[name = tensor("x_23_axes_0"), val = tensor([-1])]; - tensor x_23 = layer_norm(axes = x_23_axes_0, beta = encoder_conv_module_3_sequential_0_bias, epsilon = var_28, gamma = encoder_conv_module_3_sequential_0_weight, x = input_141)[name = tensor("x_23")]; - tensor input_143_perm_0 = const()[name = tensor("input_143_perm_0"), val = tensor([0, 2, 1])]; + tensor x_23 = layer_norm(axes = x_23_axes_0, beta = inner_encoder_conv_module_3_sequential_0_bias, epsilon = var_56, gamma = inner_encoder_conv_module_3_sequential_0_weight, x = input_143)[name = tensor("x_23")]; + tensor input_145_perm_0 = const()[name = tensor("input_145_perm_0"), val = tensor([0, 2, 1])]; tensor inputs_33_pad_type_0 = const()[name = tensor("inputs_33_pad_type_0"), val = tensor("valid")]; tensor inputs_33_strides_0 = const()[name = tensor("inputs_33_strides_0"), val = tensor([1])]; tensor inputs_33_pad_0 = const()[name = tensor("inputs_33_pad_0"), val = tensor([0, 0])]; tensor inputs_33_dilations_0 = const()[name = tensor("inputs_33_dilations_0"), val = tensor([1])]; tensor inputs_33_groups_0 = const()[name = tensor("inputs_33_groups_0"), val = tensor(1)]; - tensor input_143 = transpose(perm = input_143_perm_0, x = x_23)[name = tensor("transpose_33")]; - tensor inputs_33 = conv(bias = encoder_conv_module_3_sequential_2_conv_bias, dilations = inputs_33_dilations_0, groups = inputs_33_groups_0, pad = inputs_33_pad_0, pad_type = inputs_33_pad_type_0, strides = inputs_33_strides_0, weight = encoder_conv_module_3_sequential_2_conv_weight, x = input_143)[name = tensor("inputs_33")]; - tensor var_883_split_sizes_0 = const()[name = tensor("op_883_split_sizes_0"), val = tensor([256, 256])]; - tensor var_883_axis_0 = const()[name = tensor("op_883_axis_0"), val = tensor(1)]; - tensor var_883_0, tensor var_883_1 = split(axis = var_883_axis_0, split_sizes = var_883_split_sizes_0, x = inputs_33)[name = tensor("op_883")]; - tensor var_885 = sigmoid(x = var_883_1)[name = tensor("op_885")]; - tensor inputs_35 = mul(x = var_883_0, y = var_885)[name = tensor("inputs_35")]; + tensor input_145 = transpose(perm = input_145_perm_0, x = x_23)[name = tensor("transpose_33")]; + tensor inputs_33 = conv(bias = inner_encoder_conv_module_3_sequential_2_conv_bias, dilations = inputs_33_dilations_0, groups = inputs_33_groups_0, pad = inputs_33_pad_0, pad_type = inputs_33_pad_type_0, strides = inputs_33_strides_0, weight = inner_encoder_conv_module_3_sequential_2_conv_weight, x = input_145)[name = tensor("inputs_33")]; + tensor var_929_split_sizes_0 = const()[name = tensor("op_929_split_sizes_0"), val = tensor([256, 256])]; + tensor var_929_axis_0 = const()[name = tensor("op_929_axis_0"), val = tensor(1)]; + tensor var_929_0, tensor var_929_1 = split(axis = var_929_axis_0, split_sizes = var_929_split_sizes_0, x = inputs_33)[name = tensor("op_929")]; + tensor var_931 = sigmoid(x = var_929_1)[name = tensor("op_931")]; + tensor inputs_35 = mul(x = var_929_0, y = var_931)[name = tensor("inputs_35")]; tensor outputs_aug_pad_type_0 = const()[name = tensor("outputs_aug_pad_type_0"), val = tensor("custom")]; tensor outputs_aug_pad_0 = const()[name = tensor("outputs_aug_pad_0"), val = tensor([15, 15])]; tensor outputs_aug_groups_0 = const()[name = tensor("outputs_aug_groups_0"), val = tensor(256)]; tensor outputs_aug_strides_0 = const()[name = tensor("outputs_aug_strides_0"), val = tensor([1])]; tensor outputs_aug_dilations_0 = const()[name = tensor("outputs_aug_dilations_0"), val = tensor([1])]; - tensor outputs_aug = conv(dilations = outputs_aug_dilations_0, groups = outputs_aug_groups_0, pad = outputs_aug_pad_0, pad_type = outputs_aug_pad_type_0, strides = outputs_aug_strides_0, weight = encoder_conv_module_3_sequential_4_conv_weight, x = inputs_35)[name = tensor("outputs_aug")]; - tensor input_145_begin_0 = const()[name = tensor("input_145_begin_0"), val = tensor([0, 0, 0])]; - tensor input_145_end_0 = const()[name = tensor("input_145_end_0"), val = tensor([3, 256, 16])]; - tensor input_145_end_mask_0 = const()[name = tensor("input_145_end_mask_0"), val = tensor([true, true, false])]; - tensor input_145 = slice_by_index(begin = input_145_begin_0, end = input_145_end_0, end_mask = input_145_end_mask_0, x = outputs_aug)[name = tensor("input_145")]; - tensor inputs_37 = batch_norm(beta = encoder_conv_module_3_sequential_5_bias, epsilon = var_28, gamma = encoder_conv_module_3_sequential_5_weight, mean = encoder_conv_module_3_sequential_5_running_mean, variance = encoder_conv_module_3_sequential_5_running_var, x = input_145)[name = tensor("inputs_37")]; - tensor input_147 = silu(x = inputs_37)[name = tensor("input_147")]; - tensor input_149_pad_type_0 = const()[name = tensor("input_149_pad_type_0"), val = tensor("valid")]; - tensor input_149_strides_0 = const()[name = tensor("input_149_strides_0"), val = tensor([1])]; - tensor input_149_pad_0 = const()[name = tensor("input_149_pad_0"), val = tensor([0, 0])]; - tensor input_149_dilations_0 = const()[name = tensor("input_149_dilations_0"), val = tensor([1])]; - tensor input_149_groups_0 = const()[name = tensor("input_149_groups_0"), val = tensor(1)]; - tensor input_149 = conv(bias = encoder_conv_module_3_sequential_7_conv_bias, dilations = input_149_dilations_0, groups = input_149_groups_0, pad = input_149_pad_0, pad_type = input_149_pad_type_0, strides = input_149_strides_0, weight = encoder_conv_module_3_sequential_7_conv_weight, x = input_147)[name = tensor("input_149")]; + tensor outputs_aug = conv(dilations = outputs_aug_dilations_0, groups = outputs_aug_groups_0, pad = outputs_aug_pad_0, pad_type = outputs_aug_pad_type_0, strides = outputs_aug_strides_0, weight = inner_encoder_conv_module_3_sequential_4_conv_weight, x = inputs_35)[name = tensor("outputs_aug")]; + tensor input_147_begin_0 = const()[name = tensor("input_147_begin_0"), val = tensor([0, 0, 0])]; + tensor input_147_end_0 = const()[name = tensor("input_147_end_0"), val = tensor([3, 256, 16])]; + tensor input_147_end_mask_0 = const()[name = tensor("input_147_end_mask_0"), val = tensor([true, true, false])]; + tensor input_147 = slice_by_index(begin = input_147_begin_0, end = input_147_end_0, end_mask = input_147_end_mask_0, x = outputs_aug)[name = tensor("input_147")]; + tensor inputs_37 = batch_norm(beta = inner_encoder_conv_module_3_sequential_5_bias, epsilon = var_56, gamma = inner_encoder_conv_module_3_sequential_5_weight, mean = inner_encoder_conv_module_3_sequential_5_running_mean, variance = inner_encoder_conv_module_3_sequential_5_running_var, x = input_147)[name = tensor("inputs_37")]; + tensor input_149 = silu(x = inputs_37)[name = tensor("input_149")]; + tensor input_151_pad_type_0 = const()[name = tensor("input_151_pad_type_0"), val = tensor("valid")]; + tensor input_151_strides_0 = const()[name = tensor("input_151_strides_0"), val = tensor([1])]; + tensor input_151_pad_0 = const()[name = tensor("input_151_pad_0"), val = tensor([0, 0])]; + tensor input_151_dilations_0 = const()[name = tensor("input_151_dilations_0"), val = tensor([1])]; + tensor input_151_groups_0 = const()[name = tensor("input_151_groups_0"), val = tensor(1)]; + tensor input_151 = conv(bias = inner_encoder_conv_module_3_sequential_7_conv_bias, dilations = input_151_dilations_0, groups = input_151_groups_0, pad = input_151_pad_0, pad_type = input_151_pad_type_0, strides = input_151_strides_0, weight = inner_encoder_conv_module_3_sequential_7_conv_weight, x = input_149)[name = tensor("input_151")]; tensor conv_out_7_perm_0 = const()[name = tensor("conv_out_7_perm_0"), val = tensor([0, 2, 1])]; - tensor var_916_begin_0 = const()[name = tensor("op_916_begin_0"), val = tensor([0, -1, 0])]; - tensor var_916_end_0 = const()[name = tensor("op_916_end_0"), val = tensor([3, 16, 256])]; - tensor var_916_end_mask_0 = const()[name = tensor("op_916_end_mask_0"), val = tensor([true, true, true])]; - tensor conv_out_7 = transpose(perm = conv_out_7_perm_0, x = input_149)[name = tensor("transpose_32")]; - tensor var_916 = slice_by_index(begin = var_916_begin_0, end = var_916_end_0, end_mask = var_916_end_mask_0, x = conv_out_7)[name = tensor("op_916")]; - tensor var_918_perm_0 = const()[name = tensor("op_918_perm_0"), val = tensor([1, 0, 2])]; - tensor var_918 = transpose(perm = var_918_perm_0, x = var_916)[name = tensor("transpose_31")]; - tensor input_151 = add(x = x_21, y = var_918)[name = tensor("input_151")]; - tensor input_153_axes_0 = const()[name = tensor("input_153_axes_0"), val = tensor([-1])]; - tensor input_153 = layer_norm(axes = input_153_axes_0, beta = encoder_ffn2_3_module_sequential_0_bias, epsilon = var_28, gamma = encoder_ffn2_3_module_sequential_0_weight, x = input_151)[name = tensor("input_153")]; - tensor inputs = linear(bias = encoder_ffn2_3_module_sequential_1_linear_bias, weight = encoder_ffn2_3_module_sequential_1_linear_weight, x = input_153)[name = tensor("linear_35")]; - tensor input_155 = silu(x = inputs)[name = tensor("input_155")]; - tensor input_159 = linear(bias = encoder_ffn2_3_module_sequential_4_linear_bias, weight = encoder_ffn2_3_module_sequential_4_linear_weight, x = input_155)[name = tensor("linear_36")]; - tensor var_941 = const()[name = tensor("op_941"), val = tensor(0x1p-1)]; - tensor var_942 = mul(x = input_159, y = var_941)[name = tensor("op_942")]; - tensor input_161 = add(x = var_942, y = input_151)[name = tensor("input_161")]; + tensor var_962_begin_0 = const()[name = tensor("op_962_begin_0"), val = tensor([0, -1, 0])]; + tensor var_962_end_0 = const()[name = tensor("op_962_end_0"), val = tensor([3, 16, 256])]; + tensor var_962_end_mask_0 = const()[name = tensor("op_962_end_mask_0"), val = tensor([true, true, true])]; + tensor conv_out_7 = transpose(perm = conv_out_7_perm_0, x = input_151)[name = tensor("transpose_32")]; + tensor var_962 = slice_by_index(begin = var_962_begin_0, end = var_962_end_0, end_mask = var_962_end_mask_0, x = conv_out_7)[name = tensor("op_962")]; + tensor var_964_perm_0 = const()[name = tensor("op_964_perm_0"), val = tensor([1, 0, 2])]; + tensor var_964 = transpose(perm = var_964_perm_0, x = var_962)[name = tensor("transpose_31")]; + tensor input_153 = add(x = x_21, y = var_964)[name = tensor("input_153")]; + tensor input_155_axes_0 = const()[name = tensor("input_155_axes_0"), val = tensor([-1])]; + tensor input_155 = layer_norm(axes = input_155_axes_0, beta = inner_encoder_ffn2_3_module_sequential_0_bias, epsilon = var_56, gamma = inner_encoder_ffn2_3_module_sequential_0_weight, x = input_153)[name = tensor("input_155")]; + tensor inputs = linear(bias = inner_encoder_ffn2_3_module_sequential_1_linear_bias, weight = inner_encoder_ffn2_3_module_sequential_1_linear_weight, x = input_155)[name = tensor("linear_35")]; + tensor input_157 = silu(x = inputs)[name = tensor("input_157")]; + tensor input_161 = linear(bias = inner_encoder_ffn2_3_module_sequential_4_linear_bias, weight = inner_encoder_ffn2_3_module_sequential_4_linear_weight, x = input_157)[name = tensor("linear_36")]; + tensor var_987 = const()[name = tensor("op_987"), val = tensor(0x1p-1)]; + tensor var_988 = mul(x = input_161, y = var_987)[name = tensor("op_988")]; + tensor input_163 = add(x = var_988, y = input_153)[name = tensor("input_163")]; tensor x_25_axes_0 = const()[name = tensor("x_25_axes_0"), val = tensor([-1])]; - tensor x_25 = layer_norm(axes = x_25_axes_0, beta = encoder_layer_norm_3_bias, epsilon = var_28, gamma = encoder_layer_norm_3_weight, x = input_161)[name = tensor("x_25")]; + tensor x_25 = layer_norm(axes = x_25_axes_0, beta = inner_encoder_layer_norm_3_bias, epsilon = var_56, gamma = inner_encoder_layer_norm_3_weight, x = input_163)[name = tensor("x_25")]; tensor x_ct_perm_0 = const()[name = tensor("x_ct_perm_0"), val = tensor([0, 2, 1])]; tensor cat_interleave_0 = const()[name = tensor("cat_interleave_0"), val = tensor(false)]; tensor x_ct = transpose(perm = x_ct_perm_0, x = x_25)[name = tensor("transpose_30")]; - tensor cat = concat(axis = var_20, interleave = cat_interleave_0, values = (cnn_window, x_ct))[name = tensor("cat")]; + tensor cat = concat(axis = var_61, interleave = cat_interleave_0, values = (cnn_window, x_ct))[name = tensor("cat")]; tensor conv_out_pad_type_0 = const()[name = tensor("conv_out_pad_type_0"), val = tensor("valid")]; tensor conv_out_strides_0 = const()[name = tensor("conv_out_strides_0"), val = tensor([1])]; tensor conv_out_pad_0 = const()[name = tensor("conv_out_pad_0"), val = tensor([0, 0])]; tensor conv_out_dilations_0 = const()[name = tensor("conv_out_dilations_0"), val = tensor([1])]; tensor conv_out_groups_0 = const()[name = tensor("conv_out_groups_0"), val = tensor(1)]; - tensor conv_out = conv(bias = encoder_cnn_bias, dilations = conv_out_dilations_0, groups = conv_out_groups_0, pad = conv_out_pad_0, pad_type = conv_out_pad_type_0, strides = conv_out_strides_0, weight = encoder_cnn_weight, x = cat)[name = tensor("conv_out")]; - tensor var_960_begin_0 = const()[name = tensor("op_960_begin_0"), val = tensor([0, 0, 3])]; - tensor var_960_end_0 = const()[name = tensor("op_960_end_0"), val = tensor([1, 256, 21])]; - tensor var_960_end_mask_0 = const()[name = tensor("op_960_end_mask_0"), val = tensor([true, true, true])]; - tensor cnn_window_new = slice_by_index(begin = var_960_begin_0, end = var_960_end_0, end_mask = var_960_end_mask_0, x = cat)[name = tensor("op_960")]; - tensor input_163_perm_0 = const()[name = tensor("input_163_perm_0"), val = tensor([0, 2, 1])]; - tensor var_962 = const()[name = tensor("op_962"), val = tensor([-1])]; - tensor input_163 = transpose(perm = input_163_perm_0, x = conv_out)[name = tensor("transpose_29")]; - tensor var_963 = reduce_l2_norm(axes = var_962, keep_dims = var_29, x = input_163)[name = tensor("op_963")]; + tensor conv_out = conv(bias = inner_encoder_cnn_bias, dilations = conv_out_dilations_0, groups = conv_out_groups_0, pad = conv_out_pad_0, pad_type = conv_out_pad_type_0, strides = conv_out_strides_0, weight = inner_encoder_cnn_weight, x = cat)[name = tensor("conv_out")]; + tensor var_1006_begin_0 = const()[name = tensor("op_1006_begin_0"), val = tensor([0, 0, 3])]; + tensor var_1006_end_0 = const()[name = tensor("op_1006_end_0"), val = tensor([1, 256, 21])]; + tensor var_1006_end_mask_0 = const()[name = tensor("op_1006_end_mask_0"), val = tensor([true, true, true])]; + tensor cnn_window_new = slice_by_index(begin = var_1006_begin_0, end = var_1006_end_0, end_mask = var_1006_end_mask_0, x = cat)[name = tensor("op_1006")]; + tensor input_165_perm_0 = const()[name = tensor("input_165_perm_0"), val = tensor([0, 2, 1])]; + tensor var_1008 = const()[name = tensor("op_1008"), val = tensor([-1])]; + tensor input_165 = transpose(perm = input_165_perm_0, x = conv_out)[name = tensor("transpose_29")]; + tensor var_1009 = reduce_l2_norm(axes = var_1008, keep_dims = var_55, x = input_165)[name = tensor("op_1009")]; tensor const_12 = const()[name = tensor("const_12"), val = tensor(0x1.fffffep+127)]; - tensor clip_0 = clip(alpha = var_11, beta = const_12, x = var_963)[name = tensor("clip_0")]; - tensor emb = real_div(x = input_163, y = clip_0)[name = tensor("emb")]; - tensor var_967_axis_0 = const()[name = tensor("op_967_axis_0"), val = tensor(0)]; - tensor enc_kv_new = stack(axis = var_967_axis_0, values = (var_206, var_412, var_618, nkv_1))[name = tensor("op_967")]; - tensor var_969_axis_0 = const()[name = tensor("op_969_axis_0"), val = tensor(0)]; - tensor enc_scale_new = stack(axis = var_969_axis_0, values = (new_scale_1, new_scale_3, new_scale_5, new_scale_7))[name = tensor("op_969")]; - tensor var_971_axis_0 = const()[name = tensor("op_971_axis_0"), val = tensor(0)]; - tensor enc_conv_cache_new = stack(axis = var_971_axis_0, values = (window_7, window_15, window_23, window))[name = tensor("op_971")]; - tensor var_980 = const()[name = tensor("op_980"), val = tensor(0x1.5798eep-27)]; - tensor var_985 = const()[name = tensor("op_985"), val = tensor(0x1.4f8b58p-17)]; - tensor var_987 = const()[name = tensor("op_987"), val = tensor(0x1.0c6f7ap-20)]; - tensor var_988 = const()[name = tensor("op_988"), val = tensor(true)]; - tensor var_990 = const()[name = tensor("op_990"), val = tensor(0x1p+0)]; - tensor var_994 = const()[name = tensor("op_994"), val = tensor(-1)]; - tensor var_1000 = const()[name = tensor("op_1000"), val = tensor(0)]; + tensor clip_0 = clip(alpha = var_69, beta = const_12, x = var_1009)[name = tensor("clip_0")]; + tensor emb = real_div(x = input_165, y = clip_0)[name = tensor("emb")]; + tensor var_1013_axis_0 = const()[name = tensor("op_1013_axis_0"), val = tensor(0)]; + tensor enc_kv_new = stack(axis = var_1013_axis_0, values = (var_252, var_458, var_664, nkv_1))[name = tensor("op_1013")]; + tensor var_1015_axis_0 = const()[name = tensor("op_1015_axis_0"), val = tensor(0)]; + tensor enc_scale_new = stack(axis = var_1015_axis_0, values = (new_scale_1, new_scale_3, new_scale_5, new_scale_7))[name = tensor("op_1015")]; + tensor var_1017_axis_0 = const()[name = tensor("op_1017_axis_0"), val = tensor(0)]; + tensor enc_conv_cache_new = stack(axis = var_1017_axis_0, values = (window_7, window_15, window_23, window))[name = tensor("op_1017")]; tensor pos = const()[name = tensor("pos"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(44395328)))]; - tensor var_1062_axes_0 = const()[name = tensor("op_1062_axes_0"), val = tensor([2])]; - tensor var_1062 = expand_dims(axes = var_1062_axes_0, x = emb)[name = tensor("op_1062")]; + tensor var_1085_axes_0 = const()[name = tensor("op_1085_axes_0"), val = tensor([2])]; + tensor var_1085 = expand_dims(axes = var_1085_axes_0, x = emb)[name = tensor("op_1085")]; tensor emb_exp_reps_0 = const()[name = tensor("emb_exp_reps_0"), val = tensor([1, 1, 9, 1])]; - tensor emb_exp = tile(reps = emb_exp_reps_0, x = var_1062)[name = tensor("emb_exp")]; - tensor input_165_interleave_0 = const()[name = tensor("input_165_interleave_0"), val = tensor(false)]; - tensor input_165 = concat(axis = var_994, interleave = input_165_interleave_0, values = (emb_exp, pos))[name = tensor("input_165")]; - tensor x_27 = linear(bias = decoder_convert_bias, weight = decoder_convert_weight, x = input_165)[name = tensor("linear_37")]; - tensor var_1070_perm_0 = const()[name = tensor("op_1070_perm_0"), val = tensor([0, 2, 1, 3])]; - tensor var_1074 = const()[name = tensor("op_1074"), val = tensor([9, 3, 256])]; - tensor var_1070 = transpose(perm = var_1070_perm_0, x = x_27)[name = tensor("transpose_28")]; - tensor x_29 = reshape(shape = var_1074, x = var_1070)[name = tensor("x_29")]; + tensor emb_exp = tile(reps = emb_exp_reps_0, x = var_1085)[name = tensor("emb_exp")]; + tensor input_167_interleave_0 = const()[name = tensor("input_167_interleave_0"), val = tensor(false)]; + tensor input_167 = concat(axis = var_62, interleave = input_167_interleave_0, values = (emb_exp, pos))[name = tensor("input_167")]; + tensor x_27 = linear(bias = inner_decoder_convert_bias, weight = inner_decoder_convert_weight, x = input_167)[name = tensor("linear_37")]; + tensor var_1093_perm_0 = const()[name = tensor("op_1093_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_1097 = const()[name = tensor("op_1097"), val = tensor([9, 3, 256])]; + tensor var_1093 = transpose(perm = var_1093_perm_0, x = x_27)[name = tensor("transpose_28")]; + tensor x_29 = reshape(shape = var_1097, x = var_1093)[name = tensor("x_29")]; tensor prev_kv_9_begin_0 = const()[name = tensor("prev_kv_9_begin_0"), val = tensor([0, 0, 0, 0, 0])]; tensor prev_kv_9_end_0 = const()[name = tensor("prev_kv_9_end_0"), val = tensor([1, 9, 4, 64, 64])]; tensor prev_kv_9_end_mask_0 = const()[name = tensor("prev_kv_9_end_mask_0"), val = tensor([false, true, true, true, true])]; @@ -943,132 +954,132 @@ program(1.0) tensor prev_scale_9_end_mask_0 = const()[name = tensor("prev_scale_9_end_mask_0"), val = tensor([false, true])]; tensor prev_scale_9_squeeze_mask_0 = const()[name = tensor("prev_scale_9_squeeze_mask_0"), val = tensor([true, false])]; tensor prev_scale_9 = slice_by_index(begin = prev_scale_9_begin_0, end = prev_scale_9_end_0, end_mask = prev_scale_9_end_mask_0, squeeze_mask = prev_scale_9_squeeze_mask_0, x = dec_scale)[name = tensor("prev_scale_9")]; - tensor var_1082 = linear(bias = decoder_q_proj_0_bias, weight = decoder_q_proj_0_weight, x = x_29)[name = tensor("linear_38")]; - tensor var_1083 = const()[name = tensor("op_1083"), val = tensor([9, 3, 4, 64])]; - tensor var_1084 = reshape(shape = var_1083, x = var_1082)[name = tensor("op_1084")]; + tensor var_1105 = linear(bias = inner_decoder_q_proj_0_bias, weight = inner_decoder_q_proj_0_weight, x = x_29)[name = tensor("linear_38")]; + tensor var_1106 = const()[name = tensor("op_1106"), val = tensor([9, 3, 4, 64])]; + tensor var_1107 = reshape(shape = var_1106, x = var_1105)[name = tensor("op_1107")]; tensor q_9_perm_0 = const()[name = tensor("q_9_perm_0"), val = tensor([0, 2, 1, 3])]; - tensor var_1088 = linear(bias = decoder_k_proj_0_bias, weight = decoder_k_proj_0_weight, x = x_29)[name = tensor("linear_39")]; - tensor var_1089 = const()[name = tensor("op_1089"), val = tensor(0x1p-3)]; - tensor var_1090 = mul(x = var_1088, y = var_1089)[name = tensor("op_1090")]; - tensor var_1091 = const()[name = tensor("op_1091"), val = tensor([9, 3, 4, 64])]; - tensor var_1092 = reshape(shape = var_1091, x = var_1090)[name = tensor("op_1092")]; + tensor var_1111 = linear(bias = inner_decoder_k_proj_0_bias, weight = inner_decoder_k_proj_0_weight, x = x_29)[name = tensor("linear_39")]; + tensor var_1112 = const()[name = tensor("op_1112"), val = tensor(0x1p-3)]; + tensor var_1113 = mul(x = var_1111, y = var_1112)[name = tensor("op_1113")]; + tensor var_1114 = const()[name = tensor("op_1114"), val = tensor([9, 3, 4, 64])]; + tensor var_1115 = reshape(shape = var_1114, x = var_1113)[name = tensor("op_1115")]; tensor k_9_perm_0 = const()[name = tensor("k_9_perm_0"), val = tensor([0, 2, 1, 3])]; - tensor var_1096 = linear(bias = decoder_v_proj_0_bias, weight = decoder_v_proj_0_weight, x = x_29)[name = tensor("linear_40")]; - tensor var_1097 = const()[name = tensor("op_1097"), val = tensor([9, 3, 4, 64])]; - tensor var_1098 = reshape(shape = var_1097, x = var_1096)[name = tensor("op_1098")]; + tensor var_1119 = linear(bias = inner_decoder_v_proj_0_bias, weight = inner_decoder_v_proj_0_weight, x = x_29)[name = tensor("linear_40")]; + tensor var_1120 = const()[name = tensor("op_1120"), val = tensor([9, 3, 4, 64])]; + tensor var_1121 = reshape(shape = var_1120, x = var_1119)[name = tensor("op_1121")]; tensor v_9_perm_0 = const()[name = tensor("v_9_perm_0"), val = tensor([0, 2, 1, 3])]; - tensor input_169 = linear(bias = decoder_g_proj_0_bias, weight = decoder_g_proj_0_weight, x = x_29)[name = tensor("linear_41")]; + tensor input_171 = linear(bias = inner_decoder_g_proj_0_bias, weight = inner_decoder_g_proj_0_weight, x = x_29)[name = tensor("linear_41")]; tensor cumsum_mask_1_exclusive_0 = const()[name = tensor("cumsum_mask_1_exclusive_0"), val = tensor(false)]; tensor cumsum_mask_1_reverse_0 = const()[name = tensor("cumsum_mask_1_reverse_0"), val = tensor(false)]; - tensor cumsum_mask_1 = cumsum(axis = var_1000, exclusive = cumsum_mask_1_exclusive_0, reverse = cumsum_mask_1_reverse_0, x = valid_mask)[name = tensor("cumsum_mask_1")]; + tensor cumsum_mask_1 = cumsum(axis = var_59, exclusive = cumsum_mask_1_exclusive_0, reverse = cumsum_mask_1_reverse_0, x = valid_mask)[name = tensor("cumsum_mask_1")]; tensor sqrt_s0_9 = sqrt(x = prev_scale_9)[name = tensor("sqrt_s0_9")]; tensor s_t_9 = add(x = prev_scale_9, y = cumsum_mask_1)[name = tensor("s_t_9")]; tensor const_20 = const()[name = tensor("const_20"), val = tensor(0x1.fffffep+127)]; - tensor clip_1 = clip(alpha = var_990, beta = const_20, x = s_t_9)[name = tensor("clip_1")]; + tensor clip_1 = clip(alpha = var_49, beta = const_20, x = s_t_9)[name = tensor("clip_1")]; tensor sqrt_s_t_9 = sqrt(x = clip_1)[name = tensor("sqrt_s_t_9")]; tensor qk_9_transpose_x_1 = const()[name = tensor("qk_9_transpose_x_1"), val = tensor(false)]; tensor qk_9_transpose_y_1 = const()[name = tensor("qk_9_transpose_y_1"), val = tensor(true)]; - tensor k_9 = transpose(perm = k_9_perm_0, x = var_1092)[name = tensor("transpose_26")]; - tensor q_9 = transpose(perm = q_9_perm_0, x = var_1084)[name = tensor("transpose_27")]; + tensor k_9 = transpose(perm = k_9_perm_0, x = var_1115)[name = tensor("transpose_26")]; + tensor q_9 = transpose(perm = q_9_perm_0, x = var_1107)[name = tensor("transpose_27")]; tensor qk_9 = matmul(transpose_x = qk_9_transpose_x_1, transpose_y = qk_9_transpose_y_1, x = q_9, y = k_9)[name = tensor("qk_9")]; - tensor var_1110 = const()[name = tensor("op_1110"), val = tensor([1, 3])]; - tensor var_1111 = reshape(shape = var_1110, x = valid_mask)[name = tensor("op_1111")]; - tensor causal_with_valid_1 = mul(x = decoder__causal_mask, y = var_1111)[name = tensor("causal_with_valid_1")]; - tensor var_1113 = const()[name = tensor("op_1113"), val = tensor([3, 1])]; - tensor var_1114 = reshape(shape = var_1113, x = sqrt_s_t_9)[name = tensor("op_1114")]; - tensor M_9 = real_div(x = causal_with_valid_1, y = var_1114)[name = tensor("M_9")]; - tensor var_1116 = mul(x = qk_9, y = M_9)[name = tensor("op_1116")]; + tensor var_1133 = const()[name = tensor("op_1133"), val = tensor([1, 3])]; + tensor var_1134 = reshape(shape = var_1133, x = valid_mask)[name = tensor("op_1134")]; + tensor causal_with_valid_1 = mul(x = inner_decoder__causal_mask, y = var_1134)[name = tensor("causal_with_valid_1")]; + tensor var_1136 = const()[name = tensor("op_1136"), val = tensor([3, 1])]; + tensor var_1137 = reshape(shape = var_1136, x = sqrt_s_t_9)[name = tensor("op_1137")]; + tensor M_9 = real_div(x = causal_with_valid_1, y = var_1137)[name = tensor("M_9")]; + tensor var_1139 = mul(x = qk_9, y = M_9)[name = tensor("op_1139")]; tensor inner_9_transpose_x_0 = const()[name = tensor("inner_9_transpose_x_0"), val = tensor(false)]; tensor inner_9_transpose_y_0 = const()[name = tensor("inner_9_transpose_y_0"), val = tensor(false)]; - tensor v_9 = transpose(perm = v_9_perm_0, x = var_1098)[name = tensor("transpose_25")]; - tensor inner_9 = matmul(transpose_x = inner_9_transpose_x_0, transpose_y = inner_9_transpose_y_0, x = var_1116, y = v_9)[name = tensor("inner_9")]; - tensor var_1118_transpose_x_0 = const()[name = tensor("op_1118_transpose_x_0"), val = tensor(false)]; - tensor var_1118_transpose_y_0 = const()[name = tensor("op_1118_transpose_y_0"), val = tensor(false)]; - tensor var_1118 = matmul(transpose_x = var_1118_transpose_x_0, transpose_y = var_1118_transpose_y_0, x = q_9, y = prev_kv_9)[name = tensor("op_1118")]; - tensor var_1119 = real_div(x = sqrt_s0_9, y = sqrt_s_t_9)[name = tensor("op_1119")]; - tensor var_1120 = const()[name = tensor("op_1120"), val = tensor([1, 1, 3, 1])]; - tensor var_1121 = reshape(shape = var_1120, x = var_1119)[name = tensor("op_1121")]; - tensor cross_9 = mul(x = var_1118, y = var_1121)[name = tensor("cross_9")]; + tensor v_9 = transpose(perm = v_9_perm_0, x = var_1121)[name = tensor("transpose_25")]; + tensor inner_9 = matmul(transpose_x = inner_9_transpose_x_0, transpose_y = inner_9_transpose_y_0, x = var_1139, y = v_9)[name = tensor("inner_9")]; + tensor var_1141_transpose_x_0 = const()[name = tensor("op_1141_transpose_x_0"), val = tensor(false)]; + tensor var_1141_transpose_y_0 = const()[name = tensor("op_1141_transpose_y_0"), val = tensor(false)]; + tensor var_1141 = matmul(transpose_x = var_1141_transpose_x_0, transpose_y = var_1141_transpose_y_0, x = q_9, y = prev_kv_9)[name = tensor("op_1141")]; + tensor var_1142 = real_div(x = sqrt_s0_9, y = sqrt_s_t_9)[name = tensor("op_1142")]; + tensor var_1143 = const()[name = tensor("op_1143"), val = tensor([1, 1, 3, 1])]; + tensor var_1144 = reshape(shape = var_1143, x = var_1142)[name = tensor("op_1144")]; + tensor cross_9 = mul(x = var_1141, y = var_1144)[name = tensor("cross_9")]; tensor out_25 = add(x = inner_9, y = cross_9)[name = tensor("out_25")]; - tensor var_1124 = const()[name = tensor("op_1124"), val = tensor([1, 1, 3, 1])]; - tensor var_1125 = reshape(shape = var_1124, x = valid_mask)[name = tensor("op_1125")]; - tensor v_masked_1 = mul(x = v_9, y = var_1125)[name = tensor("v_masked_1")]; - tensor var_1127 = mul(x = prev_kv_9, y = sqrt_s0_9)[name = tensor("op_1127")]; - tensor var_1129_transpose_x_1 = const()[name = tensor("op_1129_transpose_x_1"), val = tensor(true)]; - tensor var_1129_transpose_y_1 = const()[name = tensor("op_1129_transpose_y_1"), val = tensor(false)]; - tensor var_1129 = matmul(transpose_x = var_1129_transpose_x_1, transpose_y = var_1129_transpose_y_1, x = k_9, y = v_masked_1)[name = tensor("op_1129")]; - tensor new_kv_unnorm_9 = add(x = var_1127, y = var_1129)[name = tensor("new_kv_unnorm_9")]; - tensor var_1131_keep_dims_0 = const()[name = tensor("op_1131_keep_dims_0"), val = tensor(false)]; - tensor var_1131 = reduce_sum(keep_dims = var_1131_keep_dims_0, x = valid_mask)[name = tensor("op_1131")]; - tensor var_1132 = const()[name = tensor("op_1132"), val = tensor([1])]; - tensor var_1133 = reshape(shape = var_1132, x = var_1131)[name = tensor("op_1133")]; - tensor new_scale_9 = add(x = prev_scale_9, y = var_1133)[name = tensor("new_scale_9")]; + tensor var_1147 = const()[name = tensor("op_1147"), val = tensor([1, 1, 3, 1])]; + tensor var_1148 = reshape(shape = var_1147, x = valid_mask)[name = tensor("op_1148")]; + tensor v_masked_1 = mul(x = v_9, y = var_1148)[name = tensor("v_masked_1")]; + tensor var_1150 = mul(x = prev_kv_9, y = sqrt_s0_9)[name = tensor("op_1150")]; + tensor var_1152_transpose_x_1 = const()[name = tensor("op_1152_transpose_x_1"), val = tensor(true)]; + tensor var_1152_transpose_y_1 = const()[name = tensor("op_1152_transpose_y_1"), val = tensor(false)]; + tensor var_1152 = matmul(transpose_x = var_1152_transpose_x_1, transpose_y = var_1152_transpose_y_1, x = k_9, y = v_masked_1)[name = tensor("op_1152")]; + tensor new_kv_unnorm_9 = add(x = var_1150, y = var_1152)[name = tensor("new_kv_unnorm_9")]; + tensor var_1154_keep_dims_0 = const()[name = tensor("op_1154_keep_dims_0"), val = tensor(false)]; + tensor var_1154 = reduce_sum(keep_dims = var_1154_keep_dims_0, x = valid_mask)[name = tensor("op_1154")]; + tensor var_1155 = const()[name = tensor("op_1155"), val = tensor([1])]; + tensor var_1156 = reshape(shape = var_1155, x = var_1154)[name = tensor("op_1156")]; + tensor new_scale_9 = add(x = prev_scale_9, y = var_1156)[name = tensor("new_scale_9")]; tensor const_21 = const()[name = tensor("const_21"), val = tensor(0x1.fffffep+127)]; - tensor clip_2 = clip(alpha = var_990, beta = const_21, x = new_scale_9)[name = tensor("clip_2")]; + tensor clip_2 = clip(alpha = var_49, beta = const_21, x = new_scale_9)[name = tensor("clip_2")]; tensor sqrt_new_scale_1 = sqrt(x = clip_2)[name = tensor("sqrt_new_scale_1")]; - tensor var_1137 = real_div(x = new_kv_unnorm_9, y = sqrt_new_scale_1)[name = tensor("op_1137")]; - tensor var_1138_perm_0 = const()[name = tensor("op_1138_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_1160 = real_div(x = new_kv_unnorm_9, y = sqrt_new_scale_1)[name = tensor("op_1160")]; + tensor var_1161_perm_0 = const()[name = tensor("op_1161_perm_0"), val = tensor([0, 2, 1, 3])]; tensor out_27_axes_0 = const()[name = tensor("out_27_axes_0"), val = tensor([-1])]; - tensor var_1138 = transpose(perm = var_1138_perm_0, x = out_25)[name = tensor("transpose_24")]; - tensor out_27 = layer_norm(axes = out_27_axes_0, epsilon = var_987, x = var_1138)[name = tensor("out_27")]; - tensor var_1142 = const()[name = tensor("op_1142"), val = tensor([9, 3, 256])]; - tensor out_29 = reshape(shape = var_1142, x = out_27)[name = tensor("out_29")]; - tensor var_1144 = silu(x = input_169)[name = tensor("op_1144")]; - tensor input_171 = mul(x = var_1144, y = out_29)[name = tensor("input_171")]; - tensor ret_out_9 = linear(bias = decoder_out_proj_0_bias, weight = decoder_out_proj_0_weight, x = input_171)[name = tensor("linear_42")]; - tensor input_173 = add(x = x_29, y = ret_out_9)[name = tensor("input_173")]; + tensor var_1161 = transpose(perm = var_1161_perm_0, x = out_25)[name = tensor("transpose_24")]; + tensor out_27 = layer_norm(axes = out_27_axes_0, epsilon = var_64, x = var_1161)[name = tensor("out_27")]; + tensor var_1165 = const()[name = tensor("op_1165"), val = tensor([9, 3, 256])]; + tensor out_29 = reshape(shape = var_1165, x = out_27)[name = tensor("out_29")]; + tensor var_1167 = silu(x = input_171)[name = tensor("op_1167")]; + tensor input_173 = mul(x = var_1167, y = out_29)[name = tensor("input_173")]; + tensor ret_out_9 = linear(bias = inner_decoder_out_proj_0_bias, weight = inner_decoder_out_proj_0_weight, x = input_173)[name = tensor("linear_42")]; + tensor input_175 = add(x = x_29, y = ret_out_9)[name = tensor("input_175")]; tensor xt_1_axes_0 = const()[name = tensor("xt_1_axes_0"), val = tensor([-1])]; - tensor xt_1 = layer_norm(axes = xt_1_axes_0, beta = decoder_norm11_0_bias, epsilon = var_985, gamma = decoder_norm11_0_weight, x = input_173)[name = tensor("xt_1")]; - tensor var_1154 = const()[name = tensor("op_1154"), val = tensor([1, 9, 3, 256])]; - tensor var_1155 = reshape(shape = var_1154, x = xt_1)[name = tensor("op_1155")]; - tensor var_1156_perm_0 = const()[name = tensor("op_1156_perm_0"), val = tensor([0, 2, 1, 3])]; - tensor var_1159 = const()[name = tensor("op_1159"), val = tensor([3, 9, 256])]; - tensor var_1156 = transpose(perm = var_1156_perm_0, x = var_1155)[name = tensor("transpose_23")]; - tensor query_1 = reshape(shape = var_1159, x = var_1156)[name = tensor("query_1")]; + tensor xt_1 = layer_norm(axes = xt_1_axes_0, beta = inner_decoder_norm11_0_bias, epsilon = var_56, gamma = inner_decoder_norm11_0_weight, x = input_175)[name = tensor("xt_1")]; + tensor var_1177 = const()[name = tensor("op_1177"), val = tensor([1, 9, 3, 256])]; + tensor var_1178 = reshape(shape = var_1177, x = xt_1)[name = tensor("op_1178")]; + tensor var_1179_perm_0 = const()[name = tensor("op_1179_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_1182 = const()[name = tensor("op_1182"), val = tensor([3, 9, 256])]; + tensor var_1179 = transpose(perm = var_1179_perm_0, x = var_1178)[name = tensor("transpose_23")]; + tensor query_1 = reshape(shape = var_1182, x = var_1179)[name = tensor("query_1")]; tensor query_3_perm_0 = const()[name = tensor("query_3_perm_0"), val = tensor([1, 0, 2])]; tensor query_3 = transpose(perm = query_3_perm_0, x = query_1)[name = tensor("transpose_22")]; - tensor var_1182 = linear(bias = decoder_self_attn2_0_in_proj_bias, weight = decoder_self_attn2_0_in_proj_weight, x = query_3)[name = tensor("linear_43")]; + tensor var_1205 = linear(bias = inner_decoder_self_attn2_0_in_proj_bias, weight = inner_decoder_self_attn2_0_in_proj_weight, x = query_3)[name = tensor("linear_43")]; tensor concat_1 = const()[name = tensor("concat_1"), val = tensor([9, 3, 3, 256])]; - tensor var_1184 = reshape(shape = concat_1, x = var_1182)[name = tensor("op_1184")]; - tensor var_1185_axes_0 = const()[name = tensor("op_1185_axes_0"), val = tensor([0])]; - tensor var_1185 = expand_dims(axes = var_1185_axes_0, x = var_1184)[name = tensor("op_1185")]; - tensor var_1186_perm_0 = const()[name = tensor("op_1186_perm_0"), val = tensor([-2, 1, 2, 0, 4])]; - tensor var_1187_axes_0 = const()[name = tensor("op_1187_axes_0"), val = tensor([-2])]; - tensor var_1186 = transpose(perm = var_1186_perm_0, x = var_1185)[name = tensor("transpose_21")]; - tensor var_1187 = squeeze(axes = var_1187_axes_0, x = var_1186)[name = tensor("op_1187")]; + tensor var_1207 = reshape(shape = concat_1, x = var_1205)[name = tensor("op_1207")]; + tensor var_1208_axes_0 = const()[name = tensor("op_1208_axes_0"), val = tensor([0])]; + tensor var_1208 = expand_dims(axes = var_1208_axes_0, x = var_1207)[name = tensor("op_1208")]; + tensor var_1209_perm_0 = const()[name = tensor("op_1209_perm_0"), val = tensor([-2, 1, 2, 0, 4])]; + tensor var_1210_axes_0 = const()[name = tensor("op_1210_axes_0"), val = tensor([-2])]; + tensor var_1209 = transpose(perm = var_1209_perm_0, x = var_1208)[name = tensor("transpose_21")]; + tensor var_1210 = squeeze(axes = var_1210_axes_0, x = var_1209)[name = tensor("op_1210")]; tensor q_11_begin_0 = const()[name = tensor("q_11_begin_0"), val = tensor([0, 0, 0, 0])]; tensor q_11_end_0 = const()[name = tensor("q_11_end_0"), val = tensor([1, 9, 3, 256])]; tensor q_11_end_mask_0 = const()[name = tensor("q_11_end_mask_0"), val = tensor([false, true, true, true])]; tensor q_11_squeeze_mask_0 = const()[name = tensor("q_11_squeeze_mask_0"), val = tensor([true, false, false, false])]; - tensor q_11 = slice_by_index(begin = q_11_begin_0, end = q_11_end_0, end_mask = q_11_end_mask_0, squeeze_mask = q_11_squeeze_mask_0, x = var_1187)[name = tensor("q_11")]; + tensor q_11 = slice_by_index(begin = q_11_begin_0, end = q_11_end_0, end_mask = q_11_end_mask_0, squeeze_mask = q_11_squeeze_mask_0, x = var_1210)[name = tensor("q_11")]; tensor k_11_begin_0 = const()[name = tensor("k_11_begin_0"), val = tensor([1, 0, 0, 0])]; tensor k_11_end_0 = const()[name = tensor("k_11_end_0"), val = tensor([2, 9, 3, 256])]; tensor k_11_end_mask_0 = const()[name = tensor("k_11_end_mask_0"), val = tensor([false, true, true, true])]; tensor k_11_squeeze_mask_0 = const()[name = tensor("k_11_squeeze_mask_0"), val = tensor([true, false, false, false])]; - tensor k_11 = slice_by_index(begin = k_11_begin_0, end = k_11_end_0, end_mask = k_11_end_mask_0, squeeze_mask = k_11_squeeze_mask_0, x = var_1187)[name = tensor("k_11")]; + tensor k_11 = slice_by_index(begin = k_11_begin_0, end = k_11_end_0, end_mask = k_11_end_mask_0, squeeze_mask = k_11_squeeze_mask_0, x = var_1210)[name = tensor("k_11")]; tensor v_11_begin_0 = const()[name = tensor("v_11_begin_0"), val = tensor([2, 0, 0, 0])]; tensor v_11_end_0 = const()[name = tensor("v_11_end_0"), val = tensor([3, 9, 3, 256])]; tensor v_11_end_mask_0 = const()[name = tensor("v_11_end_mask_0"), val = tensor([false, true, true, true])]; tensor v_11_squeeze_mask_0 = const()[name = tensor("v_11_squeeze_mask_0"), val = tensor([true, false, false, false])]; - tensor v_11 = slice_by_index(begin = v_11_begin_0, end = v_11_end_0, end_mask = v_11_end_mask_0, squeeze_mask = v_11_squeeze_mask_0, x = var_1187)[name = tensor("v_11")]; - tensor var_1195 = const()[name = tensor("op_1195"), val = tensor([9, 12, 64])]; - tensor var_1196 = reshape(shape = var_1195, x = q_11)[name = tensor("op_1196")]; + tensor v_11 = slice_by_index(begin = v_11_begin_0, end = v_11_end_0, end_mask = v_11_end_mask_0, squeeze_mask = v_11_squeeze_mask_0, x = var_1210)[name = tensor("v_11")]; + tensor var_1218 = const()[name = tensor("op_1218"), val = tensor([9, 12, 64])]; + tensor var_1219 = reshape(shape = var_1218, x = q_11)[name = tensor("op_1219")]; tensor q_13_perm_0 = const()[name = tensor("q_13_perm_0"), val = tensor([1, 0, 2])]; - tensor var_1202 = const()[name = tensor("op_1202"), val = tensor([9, 12, 64])]; - tensor var_1203 = reshape(shape = var_1202, x = k_11)[name = tensor("op_1203")]; + tensor var_1225 = const()[name = tensor("op_1225"), val = tensor([9, 12, 64])]; + tensor var_1226 = reshape(shape = var_1225, x = k_11)[name = tensor("op_1226")]; tensor k_13_perm_0 = const()[name = tensor("k_13_perm_0"), val = tensor([1, 0, 2])]; - tensor var_1209 = const()[name = tensor("op_1209"), val = tensor([9, 12, 64])]; - tensor var_1210 = reshape(shape = var_1209, x = v_11)[name = tensor("op_1210")]; + tensor var_1232 = const()[name = tensor("op_1232"), val = tensor([9, 12, 64])]; + tensor var_1233 = reshape(shape = var_1232, x = v_11)[name = tensor("op_1233")]; tensor v_13_perm_0 = const()[name = tensor("v_13_perm_0"), val = tensor([1, 0, 2])]; - tensor var_1213 = const()[name = tensor("op_1213"), val = tensor([3, 4, 9, 64])]; - tensor q_13 = transpose(perm = q_13_perm_0, x = var_1196)[name = tensor("transpose_20")]; - tensor q_15 = reshape(shape = var_1213, x = q_13)[name = tensor("q_15")]; - tensor var_1215 = const()[name = tensor("op_1215"), val = tensor([3, 4, 9, 64])]; - tensor k_13 = transpose(perm = k_13_perm_0, x = var_1203)[name = tensor("transpose_19")]; - tensor k_15 = reshape(shape = var_1215, x = k_13)[name = tensor("k_15")]; - tensor var_1217 = const()[name = tensor("op_1217"), val = tensor([3, 4, 9, 64])]; - tensor v_13 = transpose(perm = v_13_perm_0, x = var_1210)[name = tensor("transpose_18")]; - tensor v_15 = reshape(shape = var_1217, x = v_13)[name = tensor("v_15")]; + tensor var_1236 = const()[name = tensor("op_1236"), val = tensor([3, 4, 9, 64])]; + tensor q_13 = transpose(perm = q_13_perm_0, x = var_1219)[name = tensor("transpose_20")]; + tensor q_15 = reshape(shape = var_1236, x = q_13)[name = tensor("q_15")]; + tensor var_1238 = const()[name = tensor("op_1238"), val = tensor([3, 4, 9, 64])]; + tensor k_13 = transpose(perm = k_13_perm_0, x = var_1226)[name = tensor("transpose_19")]; + tensor k_15 = reshape(shape = var_1238, x = k_13)[name = tensor("k_15")]; + tensor var_1240 = const()[name = tensor("op_1240"), val = tensor([3, 4, 9, 64])]; + tensor v_13 = transpose(perm = v_13_perm_0, x = var_1233)[name = tensor("transpose_18")]; + tensor v_15 = reshape(shape = var_1240, x = v_13)[name = tensor("v_15")]; tensor mul_1_y_0 = const()[name = tensor("mul_1_y_0"), val = tensor(0x1p-3)]; tensor mul_1 = mul(x = q_15, y = mul_1_y_0)[name = tensor("mul_1")]; tensor matmul_0_transpose_y_0 = const()[name = tensor("matmul_0_transpose_y_0"), val = tensor(true)]; @@ -1079,30 +1090,30 @@ program(1.0) tensor attn_output_1_transpose_x_0 = const()[name = tensor("attn_output_1_transpose_x_0"), val = tensor(false)]; tensor attn_output_1_transpose_y_0 = const()[name = tensor("attn_output_1_transpose_y_0"), val = tensor(false)]; tensor attn_output_1 = matmul(transpose_x = attn_output_1_transpose_x_0, transpose_y = attn_output_1_transpose_y_0, x = softmax_0, y = v_15)[name = tensor("attn_output_1")]; - tensor var_1220 = const()[name = tensor("op_1220"), val = tensor([2, 0, 1, 3])]; - tensor var_1225 = const()[name = tensor("op_1225"), val = tensor([27, 256])]; - tensor var_1221 = transpose(perm = var_1220, x = attn_output_1)[name = tensor("transpose_17")]; - tensor attn_output_3 = reshape(shape = var_1225, x = var_1221)[name = tensor("attn_output_3")]; - tensor attn_output_5 = linear(bias = decoder_self_attn2_0_out_proj_bias, weight = decoder_self_attn2_0_out_proj_weight, x = attn_output_3)[name = tensor("linear_44")]; - tensor var_1229 = const()[name = tensor("op_1229"), val = tensor([9, 3, 256])]; - tensor attn_output_7 = reshape(shape = var_1229, x = attn_output_5)[name = tensor("attn_output_7")]; + tensor var_1243 = const()[name = tensor("op_1243"), val = tensor([2, 0, 1, 3])]; + tensor var_1248 = const()[name = tensor("op_1248"), val = tensor([27, 256])]; + tensor var_1244 = transpose(perm = var_1243, x = attn_output_1)[name = tensor("transpose_17")]; + tensor attn_output_3 = reshape(shape = var_1248, x = var_1244)[name = tensor("attn_output_3")]; + tensor attn_output_5 = linear(bias = inner_decoder_self_attn2_0_out_proj_bias, weight = inner_decoder_self_attn2_0_out_proj_weight, x = attn_output_3)[name = tensor("linear_44")]; + tensor var_1252 = const()[name = tensor("op_1252"), val = tensor([9, 3, 256])]; + tensor attn_output_7 = reshape(shape = var_1252, x = attn_output_5)[name = tensor("attn_output_7")]; tensor sa2_out_1_perm_0 = const()[name = tensor("sa2_out_1_perm_0"), val = tensor([1, 0, 2])]; tensor sa2_out_1 = transpose(perm = sa2_out_1_perm_0, x = attn_output_7)[name = tensor("transpose_16")]; - tensor input_175 = add(x = query_1, y = sa2_out_1)[name = tensor("input_175")]; - tensor input_177_axes_0 = const()[name = tensor("input_177_axes_0"), val = tensor([-1])]; - tensor input_177 = layer_norm(axes = input_177_axes_0, beta = decoder_norm21_0_bias, epsilon = var_985, gamma = decoder_norm21_0_weight, x = input_175)[name = tensor("input_177")]; - tensor input_179 = linear(bias = decoder_linear1_0_bias, weight = decoder_linear1_0_weight, x = input_177)[name = tensor("linear_45")]; - tensor input_181 = relu(x = input_179)[name = tensor("input_181")]; - tensor ffn_out_1 = linear(bias = decoder_linear2_0_bias, weight = decoder_linear2_0_weight, x = input_181)[name = tensor("linear_46")]; - tensor input_183 = add(x = input_177, y = ffn_out_1)[name = tensor("input_183")]; + tensor input_177 = add(x = query_1, y = sa2_out_1)[name = tensor("input_177")]; + tensor input_179_axes_0 = const()[name = tensor("input_179_axes_0"), val = tensor([-1])]; + tensor input_179 = layer_norm(axes = input_179_axes_0, beta = inner_decoder_norm21_0_bias, epsilon = var_56, gamma = inner_decoder_norm21_0_weight, x = input_177)[name = tensor("input_179")]; + tensor input_181 = linear(bias = inner_decoder_linear1_0_bias, weight = inner_decoder_linear1_0_weight, x = input_179)[name = tensor("linear_45")]; + tensor input_183 = relu(x = input_181)[name = tensor("input_183")]; + tensor ffn_out_1 = linear(bias = inner_decoder_linear2_0_bias, weight = inner_decoder_linear2_0_weight, x = input_183)[name = tensor("linear_46")]; + tensor input_185 = add(x = input_179, y = ffn_out_1)[name = tensor("input_185")]; tensor xt_3_axes_0 = const()[name = tensor("xt_3_axes_0"), val = tensor([-1])]; - tensor xt_3 = layer_norm(axes = xt_3_axes_0, beta = decoder_norm22_0_bias, epsilon = var_985, gamma = decoder_norm22_0_weight, x = input_183)[name = tensor("xt_3")]; - tensor var_1249 = const()[name = tensor("op_1249"), val = tensor([1, 3, 9, 256])]; - tensor x_31 = reshape(shape = var_1249, x = xt_3)[name = tensor("x_31")]; - tensor var_1251_perm_0 = const()[name = tensor("op_1251_perm_0"), val = tensor([0, 2, 1, 3])]; - tensor var_1255 = const()[name = tensor("op_1255"), val = tensor([9, 3, 256])]; - tensor var_1251 = transpose(perm = var_1251_perm_0, x = x_31)[name = tensor("transpose_15")]; - tensor x = reshape(shape = var_1255, x = var_1251)[name = tensor("x")]; + tensor xt_3 = layer_norm(axes = xt_3_axes_0, beta = inner_decoder_norm22_0_bias, epsilon = var_56, gamma = inner_decoder_norm22_0_weight, x = input_185)[name = tensor("xt_3")]; + tensor var_1272 = const()[name = tensor("op_1272"), val = tensor([1, 3, 9, 256])]; + tensor x_31 = reshape(shape = var_1272, x = xt_3)[name = tensor("x_31")]; + tensor var_1274_perm_0 = const()[name = tensor("op_1274_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_1278 = const()[name = tensor("op_1278"), val = tensor([9, 3, 256])]; + tensor var_1274 = transpose(perm = var_1274_perm_0, x = x_31)[name = tensor("transpose_15")]; + tensor x = reshape(shape = var_1278, x = var_1274)[name = tensor("x")]; tensor prev_kv_begin_0 = const()[name = tensor("prev_kv_begin_0"), val = tensor([1, 0, 0, 0, 0])]; tensor prev_kv_end_0 = const()[name = tensor("prev_kv_end_0"), val = tensor([2, 9, 4, 64, 64])]; tensor prev_kv_end_mask_0 = const()[name = tensor("prev_kv_end_mask_0"), val = tensor([false, true, true, true, true])]; @@ -1113,120 +1124,120 @@ program(1.0) tensor prev_scale_end_mask_0 = const()[name = tensor("prev_scale_end_mask_0"), val = tensor([false, true])]; tensor prev_scale_squeeze_mask_0 = const()[name = tensor("prev_scale_squeeze_mask_0"), val = tensor([true, false])]; tensor prev_scale = slice_by_index(begin = prev_scale_begin_0, end = prev_scale_end_0, end_mask = prev_scale_end_mask_0, squeeze_mask = prev_scale_squeeze_mask_0, x = dec_scale)[name = tensor("prev_scale")]; - tensor var_1263 = linear(bias = decoder_q_proj_1_bias, weight = decoder_q_proj_1_weight, x = x)[name = tensor("linear_47")]; - tensor var_1264 = const()[name = tensor("op_1264"), val = tensor([9, 3, 4, 64])]; - tensor var_1265 = reshape(shape = var_1264, x = var_1263)[name = tensor("op_1265")]; + tensor var_1286 = linear(bias = inner_decoder_q_proj_1_bias, weight = inner_decoder_q_proj_1_weight, x = x)[name = tensor("linear_47")]; + tensor var_1287 = const()[name = tensor("op_1287"), val = tensor([9, 3, 4, 64])]; + tensor var_1288 = reshape(shape = var_1287, x = var_1286)[name = tensor("op_1288")]; tensor q_17_perm_0 = const()[name = tensor("q_17_perm_0"), val = tensor([0, 2, 1, 3])]; - tensor var_1269 = linear(bias = decoder_k_proj_1_bias, weight = decoder_k_proj_1_weight, x = x)[name = tensor("linear_48")]; - tensor var_1270 = const()[name = tensor("op_1270"), val = tensor(0x1p-3)]; - tensor var_1271 = mul(x = var_1269, y = var_1270)[name = tensor("op_1271")]; - tensor var_1272 = const()[name = tensor("op_1272"), val = tensor([9, 3, 4, 64])]; - tensor var_1273 = reshape(shape = var_1272, x = var_1271)[name = tensor("op_1273")]; + tensor var_1292 = linear(bias = inner_decoder_k_proj_1_bias, weight = inner_decoder_k_proj_1_weight, x = x)[name = tensor("linear_48")]; + tensor var_1293 = const()[name = tensor("op_1293"), val = tensor(0x1p-3)]; + tensor var_1294 = mul(x = var_1292, y = var_1293)[name = tensor("op_1294")]; + tensor var_1295 = const()[name = tensor("op_1295"), val = tensor([9, 3, 4, 64])]; + tensor var_1296 = reshape(shape = var_1295, x = var_1294)[name = tensor("op_1296")]; tensor k_17_perm_0 = const()[name = tensor("k_17_perm_0"), val = tensor([0, 2, 1, 3])]; - tensor var_1277 = linear(bias = decoder_v_proj_1_bias, weight = decoder_v_proj_1_weight, x = x)[name = tensor("linear_49")]; - tensor var_1278 = const()[name = tensor("op_1278"), val = tensor([9, 3, 4, 64])]; - tensor var_1279 = reshape(shape = var_1278, x = var_1277)[name = tensor("op_1279")]; + tensor var_1300 = linear(bias = inner_decoder_v_proj_1_bias, weight = inner_decoder_v_proj_1_weight, x = x)[name = tensor("linear_49")]; + tensor var_1301 = const()[name = tensor("op_1301"), val = tensor([9, 3, 4, 64])]; + tensor var_1302 = reshape(shape = var_1301, x = var_1300)[name = tensor("op_1302")]; tensor v_17_perm_0 = const()[name = tensor("v_17_perm_0"), val = tensor([0, 2, 1, 3])]; - tensor input_187 = linear(bias = decoder_g_proj_1_bias, weight = decoder_g_proj_1_weight, x = x)[name = tensor("linear_50")]; + tensor input_189 = linear(bias = inner_decoder_g_proj_1_bias, weight = inner_decoder_g_proj_1_weight, x = x)[name = tensor("linear_50")]; tensor sqrt_s0 = sqrt(x = prev_scale)[name = tensor("sqrt_s0")]; tensor s_t = add(x = prev_scale, y = cumsum_mask_1)[name = tensor("s_t")]; tensor const_32 = const()[name = tensor("const_32"), val = tensor(0x1.fffffep+127)]; - tensor clip_3 = clip(alpha = var_990, beta = const_32, x = s_t)[name = tensor("clip_3")]; + tensor clip_3 = clip(alpha = var_49, beta = const_32, x = s_t)[name = tensor("clip_3")]; tensor sqrt_s_t = sqrt(x = clip_3)[name = tensor("sqrt_s_t")]; tensor qk_transpose_x_1 = const()[name = tensor("qk_transpose_x_1"), val = tensor(false)]; tensor qk_transpose_y_1 = const()[name = tensor("qk_transpose_y_1"), val = tensor(true)]; - tensor k_17 = transpose(perm = k_17_perm_0, x = var_1273)[name = tensor("transpose_13")]; - tensor q_17 = transpose(perm = q_17_perm_0, x = var_1265)[name = tensor("transpose_14")]; + tensor k_17 = transpose(perm = k_17_perm_0, x = var_1296)[name = tensor("transpose_13")]; + tensor q_17 = transpose(perm = q_17_perm_0, x = var_1288)[name = tensor("transpose_14")]; tensor qk = matmul(transpose_x = qk_transpose_x_1, transpose_y = qk_transpose_y_1, x = q_17, y = k_17)[name = tensor("qk")]; - tensor var_1294 = const()[name = tensor("op_1294"), val = tensor([3, 1])]; - tensor var_1295 = reshape(shape = var_1294, x = sqrt_s_t)[name = tensor("op_1295")]; - tensor M = real_div(x = causal_with_valid_1, y = var_1295)[name = tensor("M")]; - tensor var_1297 = mul(x = qk, y = M)[name = tensor("op_1297")]; - tensor inner_transpose_x_0 = const()[name = tensor("inner_transpose_x_0"), val = tensor(false)]; - tensor inner_transpose_y_0 = const()[name = tensor("inner_transpose_y_0"), val = tensor(false)]; - tensor v_17 = transpose(perm = v_17_perm_0, x = var_1279)[name = tensor("transpose_12")]; - tensor inner = matmul(transpose_x = inner_transpose_x_0, transpose_y = inner_transpose_y_0, x = var_1297, y = v_17)[name = tensor("inner")]; - tensor var_1299_transpose_x_0 = const()[name = tensor("op_1299_transpose_x_0"), val = tensor(false)]; - tensor var_1299_transpose_y_0 = const()[name = tensor("op_1299_transpose_y_0"), val = tensor(false)]; - tensor var_1299 = matmul(transpose_x = var_1299_transpose_x_0, transpose_y = var_1299_transpose_y_0, x = q_17, y = prev_kv)[name = tensor("op_1299")]; - tensor var_1300 = real_div(x = sqrt_s0, y = sqrt_s_t)[name = tensor("op_1300")]; - tensor var_1301 = const()[name = tensor("op_1301"), val = tensor([1, 1, 3, 1])]; - tensor var_1302 = reshape(shape = var_1301, x = var_1300)[name = tensor("op_1302")]; - tensor cross = mul(x = var_1299, y = var_1302)[name = tensor("cross")]; - tensor out_31 = add(x = inner, y = cross)[name = tensor("out_31")]; - tensor v_masked = mul(x = v_17, y = var_1125)[name = tensor("v_masked")]; - tensor var_1308 = mul(x = prev_kv, y = sqrt_s0)[name = tensor("op_1308")]; - tensor var_1310_transpose_x_1 = const()[name = tensor("op_1310_transpose_x_1"), val = tensor(true)]; - tensor var_1310_transpose_y_1 = const()[name = tensor("op_1310_transpose_y_1"), val = tensor(false)]; - tensor var_1310 = matmul(transpose_x = var_1310_transpose_x_1, transpose_y = var_1310_transpose_y_1, x = k_17, y = v_masked)[name = tensor("op_1310")]; - tensor new_kv_unnorm = add(x = var_1308, y = var_1310)[name = tensor("new_kv_unnorm")]; - tensor new_scale = add(x = prev_scale, y = var_1133)[name = tensor("new_scale")]; + tensor var_1317 = const()[name = tensor("op_1317"), val = tensor([3, 1])]; + tensor var_1318 = reshape(shape = var_1317, x = sqrt_s_t)[name = tensor("op_1318")]; + tensor M = real_div(x = causal_with_valid_1, y = var_1318)[name = tensor("M")]; + tensor var_1320 = mul(x = qk, y = M)[name = tensor("op_1320")]; + tensor inner_11_transpose_x_0 = const()[name = tensor("inner_11_transpose_x_0"), val = tensor(false)]; + tensor inner_11_transpose_y_0 = const()[name = tensor("inner_11_transpose_y_0"), val = tensor(false)]; + tensor v_17 = transpose(perm = v_17_perm_0, x = var_1302)[name = tensor("transpose_12")]; + tensor inner_11 = matmul(transpose_x = inner_11_transpose_x_0, transpose_y = inner_11_transpose_y_0, x = var_1320, y = v_17)[name = tensor("inner_11")]; + tensor var_1322_transpose_x_0 = const()[name = tensor("op_1322_transpose_x_0"), val = tensor(false)]; + tensor var_1322_transpose_y_0 = const()[name = tensor("op_1322_transpose_y_0"), val = tensor(false)]; + tensor var_1322 = matmul(transpose_x = var_1322_transpose_x_0, transpose_y = var_1322_transpose_y_0, x = q_17, y = prev_kv)[name = tensor("op_1322")]; + tensor var_1323 = real_div(x = sqrt_s0, y = sqrt_s_t)[name = tensor("op_1323")]; + tensor var_1324 = const()[name = tensor("op_1324"), val = tensor([1, 1, 3, 1])]; + tensor var_1325 = reshape(shape = var_1324, x = var_1323)[name = tensor("op_1325")]; + tensor cross = mul(x = var_1322, y = var_1325)[name = tensor("cross")]; + tensor out_31 = add(x = inner_11, y = cross)[name = tensor("out_31")]; + tensor v_masked = mul(x = v_17, y = var_1148)[name = tensor("v_masked")]; + tensor var_1331 = mul(x = prev_kv, y = sqrt_s0)[name = tensor("op_1331")]; + tensor var_1333_transpose_x_1 = const()[name = tensor("op_1333_transpose_x_1"), val = tensor(true)]; + tensor var_1333_transpose_y_1 = const()[name = tensor("op_1333_transpose_y_1"), val = tensor(false)]; + tensor var_1333 = matmul(transpose_x = var_1333_transpose_x_1, transpose_y = var_1333_transpose_y_1, x = k_17, y = v_masked)[name = tensor("op_1333")]; + tensor new_kv_unnorm = add(x = var_1331, y = var_1333)[name = tensor("new_kv_unnorm")]; + tensor new_scale = add(x = prev_scale, y = var_1156)[name = tensor("new_scale")]; tensor const_33 = const()[name = tensor("const_33"), val = tensor(0x1.fffffep+127)]; - tensor clip_4 = clip(alpha = var_990, beta = const_33, x = new_scale)[name = tensor("clip_4")]; + tensor clip_4 = clip(alpha = var_49, beta = const_33, x = new_scale)[name = tensor("clip_4")]; tensor sqrt_new_scale = sqrt(x = clip_4)[name = tensor("sqrt_new_scale")]; tensor nkv = real_div(x = new_kv_unnorm, y = sqrt_new_scale)[name = tensor("nkv")]; - tensor var_1319_perm_0 = const()[name = tensor("op_1319_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_1342_perm_0 = const()[name = tensor("op_1342_perm_0"), val = tensor([0, 2, 1, 3])]; tensor out_33_axes_0 = const()[name = tensor("out_33_axes_0"), val = tensor([-1])]; - tensor var_1319 = transpose(perm = var_1319_perm_0, x = out_31)[name = tensor("transpose_11")]; - tensor out_33 = layer_norm(axes = out_33_axes_0, epsilon = var_987, x = var_1319)[name = tensor("out_33")]; - tensor var_1323 = const()[name = tensor("op_1323"), val = tensor([9, 3, 256])]; - tensor out = reshape(shape = var_1323, x = out_33)[name = tensor("out")]; - tensor var_1325 = silu(x = input_187)[name = tensor("op_1325")]; - tensor input_189 = mul(x = var_1325, y = out)[name = tensor("input_189")]; - tensor ret_out = linear(bias = decoder_out_proj_1_bias, weight = decoder_out_proj_1_weight, x = input_189)[name = tensor("linear_51")]; - tensor input_191 = add(x = x, y = ret_out)[name = tensor("input_191")]; + tensor var_1342 = transpose(perm = var_1342_perm_0, x = out_31)[name = tensor("transpose_11")]; + tensor out_33 = layer_norm(axes = out_33_axes_0, epsilon = var_64, x = var_1342)[name = tensor("out_33")]; + tensor var_1346 = const()[name = tensor("op_1346"), val = tensor([9, 3, 256])]; + tensor out = reshape(shape = var_1346, x = out_33)[name = tensor("out")]; + tensor var_1348 = silu(x = input_189)[name = tensor("op_1348")]; + tensor input_191 = mul(x = var_1348, y = out)[name = tensor("input_191")]; + tensor ret_out = linear(bias = inner_decoder_out_proj_1_bias, weight = inner_decoder_out_proj_1_weight, x = input_191)[name = tensor("linear_51")]; + tensor input_193 = add(x = x, y = ret_out)[name = tensor("input_193")]; tensor xt_5_axes_0 = const()[name = tensor("xt_5_axes_0"), val = tensor([-1])]; - tensor xt_5 = layer_norm(axes = xt_5_axes_0, beta = decoder_norm11_1_bias, epsilon = var_985, gamma = decoder_norm11_1_weight, x = input_191)[name = tensor("xt_5")]; - tensor var_1335 = const()[name = tensor("op_1335"), val = tensor([1, 9, 3, 256])]; - tensor var_1336 = reshape(shape = var_1335, x = xt_5)[name = tensor("op_1336")]; - tensor var_1337_perm_0 = const()[name = tensor("op_1337_perm_0"), val = tensor([0, 2, 1, 3])]; - tensor var_1340 = const()[name = tensor("op_1340"), val = tensor([3, 9, 256])]; - tensor var_1337 = transpose(perm = var_1337_perm_0, x = var_1336)[name = tensor("transpose_10")]; - tensor query_5 = reshape(shape = var_1340, x = var_1337)[name = tensor("query_5")]; + tensor xt_5 = layer_norm(axes = xt_5_axes_0, beta = inner_decoder_norm11_1_bias, epsilon = var_56, gamma = inner_decoder_norm11_1_weight, x = input_193)[name = tensor("xt_5")]; + tensor var_1358 = const()[name = tensor("op_1358"), val = tensor([1, 9, 3, 256])]; + tensor var_1359 = reshape(shape = var_1358, x = xt_5)[name = tensor("op_1359")]; + tensor var_1360_perm_0 = const()[name = tensor("op_1360_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_1363 = const()[name = tensor("op_1363"), val = tensor([3, 9, 256])]; + tensor var_1360 = transpose(perm = var_1360_perm_0, x = var_1359)[name = tensor("transpose_10")]; + tensor query_5 = reshape(shape = var_1363, x = var_1360)[name = tensor("query_5")]; tensor query_perm_0 = const()[name = tensor("query_perm_0"), val = tensor([1, 0, 2])]; tensor query = transpose(perm = query_perm_0, x = query_5)[name = tensor("transpose_9")]; - tensor var_1363 = linear(bias = decoder_self_attn2_1_in_proj_bias, weight = decoder_self_attn2_1_in_proj_weight, x = query)[name = tensor("linear_52")]; + tensor var_1386 = linear(bias = inner_decoder_self_attn2_1_in_proj_bias, weight = inner_decoder_self_attn2_1_in_proj_weight, x = query)[name = tensor("linear_52")]; tensor concat_2 = const()[name = tensor("concat_2"), val = tensor([9, 3, 3, 256])]; - tensor var_1365 = reshape(shape = concat_2, x = var_1363)[name = tensor("op_1365")]; - tensor var_1366_axes_0 = const()[name = tensor("op_1366_axes_0"), val = tensor([0])]; - tensor var_1366 = expand_dims(axes = var_1366_axes_0, x = var_1365)[name = tensor("op_1366")]; - tensor var_1367_perm_0 = const()[name = tensor("op_1367_perm_0"), val = tensor([-2, 1, 2, 0, 4])]; - tensor var_1368_axes_0 = const()[name = tensor("op_1368_axes_0"), val = tensor([-2])]; - tensor var_1367 = transpose(perm = var_1367_perm_0, x = var_1366)[name = tensor("transpose_8")]; - tensor var_1368 = squeeze(axes = var_1368_axes_0, x = var_1367)[name = tensor("op_1368")]; + tensor var_1388 = reshape(shape = concat_2, x = var_1386)[name = tensor("op_1388")]; + tensor var_1389_axes_0 = const()[name = tensor("op_1389_axes_0"), val = tensor([0])]; + tensor var_1389 = expand_dims(axes = var_1389_axes_0, x = var_1388)[name = tensor("op_1389")]; + tensor var_1390_perm_0 = const()[name = tensor("op_1390_perm_0"), val = tensor([-2, 1, 2, 0, 4])]; + tensor var_1391_axes_0 = const()[name = tensor("op_1391_axes_0"), val = tensor([-2])]; + tensor var_1390 = transpose(perm = var_1390_perm_0, x = var_1389)[name = tensor("transpose_8")]; + tensor var_1391 = squeeze(axes = var_1391_axes_0, x = var_1390)[name = tensor("op_1391")]; tensor q_19_begin_0 = const()[name = tensor("q_19_begin_0"), val = tensor([0, 0, 0, 0])]; tensor q_19_end_0 = const()[name = tensor("q_19_end_0"), val = tensor([1, 9, 3, 256])]; tensor q_19_end_mask_0 = const()[name = tensor("q_19_end_mask_0"), val = tensor([false, true, true, true])]; tensor q_19_squeeze_mask_0 = const()[name = tensor("q_19_squeeze_mask_0"), val = tensor([true, false, false, false])]; - tensor q_19 = slice_by_index(begin = q_19_begin_0, end = q_19_end_0, end_mask = q_19_end_mask_0, squeeze_mask = q_19_squeeze_mask_0, x = var_1368)[name = tensor("q_19")]; + tensor q_19 = slice_by_index(begin = q_19_begin_0, end = q_19_end_0, end_mask = q_19_end_mask_0, squeeze_mask = q_19_squeeze_mask_0, x = var_1391)[name = tensor("q_19")]; tensor k_19_begin_0 = const()[name = tensor("k_19_begin_0"), val = tensor([1, 0, 0, 0])]; tensor k_19_end_0 = const()[name = tensor("k_19_end_0"), val = tensor([2, 9, 3, 256])]; tensor k_19_end_mask_0 = const()[name = tensor("k_19_end_mask_0"), val = tensor([false, true, true, true])]; tensor k_19_squeeze_mask_0 = const()[name = tensor("k_19_squeeze_mask_0"), val = tensor([true, false, false, false])]; - tensor k_19 = slice_by_index(begin = k_19_begin_0, end = k_19_end_0, end_mask = k_19_end_mask_0, squeeze_mask = k_19_squeeze_mask_0, x = var_1368)[name = tensor("k_19")]; + tensor k_19 = slice_by_index(begin = k_19_begin_0, end = k_19_end_0, end_mask = k_19_end_mask_0, squeeze_mask = k_19_squeeze_mask_0, x = var_1391)[name = tensor("k_19")]; tensor v_19_begin_0 = const()[name = tensor("v_19_begin_0"), val = tensor([2, 0, 0, 0])]; tensor v_19_end_0 = const()[name = tensor("v_19_end_0"), val = tensor([3, 9, 3, 256])]; tensor v_19_end_mask_0 = const()[name = tensor("v_19_end_mask_0"), val = tensor([false, true, true, true])]; tensor v_19_squeeze_mask_0 = const()[name = tensor("v_19_squeeze_mask_0"), val = tensor([true, false, false, false])]; - tensor v_19 = slice_by_index(begin = v_19_begin_0, end = v_19_end_0, end_mask = v_19_end_mask_0, squeeze_mask = v_19_squeeze_mask_0, x = var_1368)[name = tensor("v_19")]; - tensor var_1376 = const()[name = tensor("op_1376"), val = tensor([9, 12, 64])]; - tensor var_1377 = reshape(shape = var_1376, x = q_19)[name = tensor("op_1377")]; + tensor v_19 = slice_by_index(begin = v_19_begin_0, end = v_19_end_0, end_mask = v_19_end_mask_0, squeeze_mask = v_19_squeeze_mask_0, x = var_1391)[name = tensor("v_19")]; + tensor var_1399 = const()[name = tensor("op_1399"), val = tensor([9, 12, 64])]; + tensor var_1400 = reshape(shape = var_1399, x = q_19)[name = tensor("op_1400")]; tensor q_21_perm_0 = const()[name = tensor("q_21_perm_0"), val = tensor([1, 0, 2])]; - tensor var_1383 = const()[name = tensor("op_1383"), val = tensor([9, 12, 64])]; - tensor var_1384 = reshape(shape = var_1383, x = k_19)[name = tensor("op_1384")]; + tensor var_1406 = const()[name = tensor("op_1406"), val = tensor([9, 12, 64])]; + tensor var_1407 = reshape(shape = var_1406, x = k_19)[name = tensor("op_1407")]; tensor k_21_perm_0 = const()[name = tensor("k_21_perm_0"), val = tensor([1, 0, 2])]; - tensor var_1390 = const()[name = tensor("op_1390"), val = tensor([9, 12, 64])]; - tensor var_1391 = reshape(shape = var_1390, x = v_19)[name = tensor("op_1391")]; + tensor var_1413 = const()[name = tensor("op_1413"), val = tensor([9, 12, 64])]; + tensor var_1414 = reshape(shape = var_1413, x = v_19)[name = tensor("op_1414")]; tensor v_21_perm_0 = const()[name = tensor("v_21_perm_0"), val = tensor([1, 0, 2])]; - tensor var_1394 = const()[name = tensor("op_1394"), val = tensor([3, 4, 9, 64])]; - tensor q_21 = transpose(perm = q_21_perm_0, x = var_1377)[name = tensor("transpose_7")]; - tensor q = reshape(shape = var_1394, x = q_21)[name = tensor("q")]; - tensor var_1396 = const()[name = tensor("op_1396"), val = tensor([3, 4, 9, 64])]; - tensor k_21 = transpose(perm = k_21_perm_0, x = var_1384)[name = tensor("transpose_6")]; - tensor k = reshape(shape = var_1396, x = k_21)[name = tensor("k")]; - tensor var_1398 = const()[name = tensor("op_1398"), val = tensor([3, 4, 9, 64])]; - tensor v_21 = transpose(perm = v_21_perm_0, x = var_1391)[name = tensor("transpose_5")]; - tensor v = reshape(shape = var_1398, x = v_21)[name = tensor("v")]; + tensor var_1417 = const()[name = tensor("op_1417"), val = tensor([3, 4, 9, 64])]; + tensor q_21 = transpose(perm = q_21_perm_0, x = var_1400)[name = tensor("transpose_7")]; + tensor q = reshape(shape = var_1417, x = q_21)[name = tensor("q")]; + tensor var_1419 = const()[name = tensor("op_1419"), val = tensor([3, 4, 9, 64])]; + tensor k_21 = transpose(perm = k_21_perm_0, x = var_1407)[name = tensor("transpose_6")]; + tensor k = reshape(shape = var_1419, x = k_21)[name = tensor("k")]; + tensor var_1421 = const()[name = tensor("op_1421"), val = tensor([3, 4, 9, 64])]; + tensor v_21 = transpose(perm = v_21_perm_0, x = var_1414)[name = tensor("transpose_5")]; + tensor v = reshape(shape = var_1421, x = v_21)[name = tensor("v")]; tensor mul_3_y_0 = const()[name = tensor("mul_3_y_0"), val = tensor(0x1p-3)]; tensor mul_3 = mul(x = q, y = mul_3_y_0)[name = tensor("mul_3")]; tensor matmul_1_transpose_y_0 = const()[name = tensor("matmul_1_transpose_y_0"), val = tensor(true)]; @@ -1237,36 +1248,36 @@ program(1.0) tensor attn_output_9_transpose_x_0 = const()[name = tensor("attn_output_9_transpose_x_0"), val = tensor(false)]; tensor attn_output_9_transpose_y_0 = const()[name = tensor("attn_output_9_transpose_y_0"), val = tensor(false)]; tensor attn_output_9 = matmul(transpose_x = attn_output_9_transpose_x_0, transpose_y = attn_output_9_transpose_y_0, x = softmax_1, y = v)[name = tensor("attn_output_9")]; - tensor var_1401 = const()[name = tensor("op_1401"), val = tensor([2, 0, 1, 3])]; - tensor var_1406 = const()[name = tensor("op_1406"), val = tensor([27, 256])]; - tensor var_1402 = transpose(perm = var_1401, x = attn_output_9)[name = tensor("transpose_4")]; - tensor attn_output_11 = reshape(shape = var_1406, x = var_1402)[name = tensor("attn_output_11")]; - tensor attn_output_13 = linear(bias = decoder_self_attn2_1_out_proj_bias, weight = decoder_self_attn2_1_out_proj_weight, x = attn_output_11)[name = tensor("linear_53")]; - tensor var_1410 = const()[name = tensor("op_1410"), val = tensor([9, 3, 256])]; - tensor attn_output = reshape(shape = var_1410, x = attn_output_13)[name = tensor("attn_output")]; + tensor var_1424 = const()[name = tensor("op_1424"), val = tensor([2, 0, 1, 3])]; + tensor var_1429 = const()[name = tensor("op_1429"), val = tensor([27, 256])]; + tensor var_1425 = transpose(perm = var_1424, x = attn_output_9)[name = tensor("transpose_4")]; + tensor attn_output_11 = reshape(shape = var_1429, x = var_1425)[name = tensor("attn_output_11")]; + tensor attn_output_13 = linear(bias = inner_decoder_self_attn2_1_out_proj_bias, weight = inner_decoder_self_attn2_1_out_proj_weight, x = attn_output_11)[name = tensor("linear_53")]; + tensor var_1433 = const()[name = tensor("op_1433"), val = tensor([9, 3, 256])]; + tensor attn_output = reshape(shape = var_1433, x = attn_output_13)[name = tensor("attn_output")]; tensor sa2_out_perm_0 = const()[name = tensor("sa2_out_perm_0"), val = tensor([1, 0, 2])]; tensor sa2_out = transpose(perm = sa2_out_perm_0, x = attn_output)[name = tensor("transpose_3")]; - tensor input_193 = add(x = query_5, y = sa2_out)[name = tensor("input_193")]; - tensor input_195_axes_0 = const()[name = tensor("input_195_axes_0"), val = tensor([-1])]; - tensor input_195 = layer_norm(axes = input_195_axes_0, beta = decoder_norm21_1_bias, epsilon = var_985, gamma = decoder_norm21_1_weight, x = input_193)[name = tensor("input_195")]; - tensor input_197 = linear(bias = decoder_linear1_1_bias, weight = decoder_linear1_1_weight, x = input_195)[name = tensor("linear_54")]; - tensor input_199 = relu(x = input_197)[name = tensor("input_199")]; - tensor ffn_out = linear(bias = decoder_linear2_1_bias, weight = decoder_linear2_1_weight, x = input_199)[name = tensor("linear_55")]; - tensor input_201 = add(x = input_195, y = ffn_out)[name = tensor("input_201")]; + tensor input_195 = add(x = query_5, y = sa2_out)[name = tensor("input_195")]; + tensor input_197_axes_0 = const()[name = tensor("input_197_axes_0"), val = tensor([-1])]; + tensor input_197 = layer_norm(axes = input_197_axes_0, beta = inner_decoder_norm21_1_bias, epsilon = var_56, gamma = inner_decoder_norm21_1_weight, x = input_195)[name = tensor("input_197")]; + tensor input_199 = linear(bias = inner_decoder_linear1_1_bias, weight = inner_decoder_linear1_1_weight, x = input_197)[name = tensor("linear_54")]; + tensor input_201 = relu(x = input_199)[name = tensor("input_201")]; + tensor ffn_out = linear(bias = inner_decoder_linear2_1_bias, weight = inner_decoder_linear2_1_weight, x = input_201)[name = tensor("linear_55")]; + tensor input_203 = add(x = input_197, y = ffn_out)[name = tensor("input_203")]; tensor xt_axes_0 = const()[name = tensor("xt_axes_0"), val = tensor([-1])]; - tensor xt = layer_norm(axes = xt_axes_0, beta = decoder_norm22_1_bias, epsilon = var_985, gamma = decoder_norm22_1_weight, x = input_201)[name = tensor("xt")]; - tensor var_1430 = const()[name = tensor("op_1430"), val = tensor([1, 3, 9, 256])]; - tensor input = reshape(shape = var_1430, x = xt)[name = tensor("input")]; - tensor var_1432 = const()[name = tensor("op_1432"), val = tensor([-1])]; - tensor var_1433 = reduce_l2_norm(axes = var_1432, keep_dims = var_988, x = input)[name = tensor("op_1433")]; + tensor xt = layer_norm(axes = xt_axes_0, beta = inner_decoder_norm22_1_bias, epsilon = var_56, gamma = inner_decoder_norm22_1_weight, x = input_203)[name = tensor("xt")]; + tensor var_1453 = const()[name = tensor("op_1453"), val = tensor([1, 3, 9, 256])]; + tensor input = reshape(shape = var_1453, x = xt)[name = tensor("input")]; + tensor var_1455 = const()[name = tensor("op_1455"), val = tensor([-1])]; + tensor var_1456 = reduce_l2_norm(axes = var_1455, keep_dims = var_55, x = input)[name = tensor("op_1456")]; tensor const_42 = const()[name = tensor("const_42"), val = tensor(0x1.fffffep+127)]; - tensor clip_5 = clip(alpha = var_980, beta = const_42, x = var_1433)[name = tensor("clip_5")]; - tensor var_1435 = real_div(x = input, y = clip_5)[name = tensor("op_1435")]; + tensor clip_5 = clip(alpha = var_69, beta = const_42, x = var_1456)[name = tensor("clip_5")]; + tensor var_1458 = real_div(x = input, y = clip_5)[name = tensor("op_1458")]; tensor concat_6 = const()[name = tensor("concat_6"), val = tensor([3, 1, 256])]; tensor reshape_0 = reshape(shape = concat_6, x = emb)[name = tensor("reshape_0")]; tensor transpose_1_perm_0 = const()[name = tensor("transpose_1_perm_0"), val = tensor([0, 1, 3, 2])]; tensor concat_7 = const()[name = tensor("concat_7"), val = tensor([3, 256, 9])]; - tensor transpose_1 = transpose(perm = transpose_1_perm_0, x = var_1435)[name = tensor("transpose_2")]; + tensor transpose_1 = transpose(perm = transpose_1_perm_0, x = var_1458)[name = tensor("transpose_2")]; tensor reshape_1 = reshape(shape = concat_7, x = transpose_1)[name = tensor("reshape_1")]; tensor matmul_2_transpose_x_0 = const()[name = tensor("matmul_2_transpose_x_0"), val = tensor(false)]; tensor matmul_2_transpose_y_0 = const()[name = tensor("matmul_2_transpose_y_0"), val = tensor(false)]; @@ -1277,10 +1288,10 @@ program(1.0) tensor output_end_0 = const()[name = tensor("output_end_0"), val = tensor([1, 3, 8])]; tensor output_end_mask_0 = const()[name = tensor("output_end_mask_0"), val = tensor([true, true, false])]; tensor output = slice_by_index(begin = output_begin_0, end = output_end_0, end_mask = output_end_mask_0, x = reshape_2)[name = tensor("output")]; - tensor probs = sigmoid(x = output)[name = tensor("op_1439")]; - tensor var_1441_axis_0 = const()[name = tensor("op_1441_axis_0"), val = tensor(0)]; - tensor dec_kv_new = stack(axis = var_1441_axis_0, values = (var_1137, nkv))[name = tensor("op_1441")]; - tensor var_1443_axis_0 = const()[name = tensor("op_1443_axis_0"), val = tensor(0)]; - tensor dec_scale_new = stack(axis = var_1443_axis_0, values = (new_scale_9, new_scale))[name = tensor("op_1443")]; + tensor probs = sigmoid(x = output)[name = tensor("op_1462")]; + tensor var_1464_axis_0 = const()[name = tensor("op_1464_axis_0"), val = tensor(0)]; + tensor dec_kv_new = stack(axis = var_1464_axis_0, values = (var_1160, nkv))[name = tensor("op_1464")]; + tensor var_1466_axis_0 = const()[name = tensor("op_1466_axis_0"), val = tensor(0)]; + tensor dec_scale_new = stack(axis = var_1466_axis_0, values = (new_scale_9, new_scale))[name = tensor("op_1466")]; } -> (probs, enc_kv_new, enc_scale_new, enc_conv_cache_new, cnn_window_new, dec_kv_new, dec_scale_new); } \ No newline at end of file diff --git a/optimized/ch/300ms/ls_eend_ch_300ms.mlpackage/Data/com.apple.CoreML/model.mlmodel b/optimized/ch/300ms/ls_eend_ch_300ms.mlpackage/Data/com.apple.CoreML/model.mlmodel index 09eaf79a16883ad1d32f03778b55635fbbffe830..8d1afc73657769e2f6f912f61c64f916553c8734 100644 --- a/optimized/ch/300ms/ls_eend_ch_300ms.mlpackage/Data/com.apple.CoreML/model.mlmodel +++ b/optimized/ch/300ms/ls_eend_ch_300ms.mlpackage/Data/com.apple.CoreML/model.mlmodel @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:6e633b39d71e29671336855550a3677233cf59cb49fe3e97a9c94677d5acb3f1 -size 185460 +oid sha256:3f5830d50c24f80450f2c08d152e16bb1da8692b3b26381056871626e0808fbc +size 191005 diff --git a/optimized/ch/300ms/ls_eend_ch_300ms.mlpackage/Manifest.json b/optimized/ch/300ms/ls_eend_ch_300ms.mlpackage/Manifest.json index 250225980694a093a91bfcf274f8ddeff9d47549..f08e9b846fafdcbf668951e72cf7d4330ceb98f1 100644 --- a/optimized/ch/300ms/ls_eend_ch_300ms.mlpackage/Manifest.json +++ b/optimized/ch/300ms/ls_eend_ch_300ms.mlpackage/Manifest.json @@ -1,18 +1,18 @@ { "fileFormatVersion": "1.0.0", "itemInfoEntries": { - "539B39A7-7E21-40EC-852C-90E40A6CCF50": { - "author": "com.apple.CoreML", - "description": "CoreML Model Specification", - "name": "model.mlmodel", - "path": "com.apple.CoreML/model.mlmodel" - }, - "743E81BC-0B9E-4702-AC8D-3B134F65C5D1": { + "2FAA383D-FADF-4A2C-B5BA-06B37FB7467E": { "author": "com.apple.CoreML", "description": "CoreML Model Weights", "name": "weights", "path": "com.apple.CoreML/weights" + }, + "60829799-5C9A-43C1-B425-7785E458D03C": { + "author": "com.apple.CoreML", + "description": "CoreML Model Specification", + "name": "model.mlmodel", + "path": "com.apple.CoreML/model.mlmodel" } }, - "rootModelIdentifier": "539B39A7-7E21-40EC-852C-90E40A6CCF50" + "rootModelIdentifier": "60829799-5C9A-43C1-B425-7785E458D03C" } diff --git a/optimized/ch/400ms/ls_eend_ch_400ms.mlmodelc/analytics/coremldata.bin b/optimized/ch/400ms/ls_eend_ch_400ms.mlmodelc/analytics/coremldata.bin index 8ef1ae6589fe15764e0c192933da3263e5f868b8..e90bf2e6774457a495929bffad5627e2beeb349f 100644 --- a/optimized/ch/400ms/ls_eend_ch_400ms.mlmodelc/analytics/coremldata.bin +++ b/optimized/ch/400ms/ls_eend_ch_400ms.mlmodelc/analytics/coremldata.bin @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:030a2bc7f91827f6efdac515150972512aa462ecb691365f25d97480cc537e7b +oid sha256:d475261f5ecfad58b0970d8de176466ec976310030a3c98ebbde85306d5a944b size 243 diff --git a/optimized/ch/400ms/ls_eend_ch_400ms.mlmodelc/coremldata.bin b/optimized/ch/400ms/ls_eend_ch_400ms.mlmodelc/coremldata.bin index a82c5d753734fd18061bb4de2c2b1b5334ef3f57..ce4dc38e34ee492c982a194732ec4cb2dae10950 100644 --- a/optimized/ch/400ms/ls_eend_ch_400ms.mlmodelc/coremldata.bin +++ b/optimized/ch/400ms/ls_eend_ch_400ms.mlmodelc/coremldata.bin @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:d8fe19dbc10aceae5862d5502e24f0c7c00f356c66d3828aae97b2c41383a71b -size 1301 +oid sha256:fe75c57185d939ba1e81b4dcd5f4f2b95480d18f25636cd453e7b4eec7f29680 +size 1404 diff --git a/optimized/ch/400ms/ls_eend_ch_400ms.mlmodelc/metadata.json b/optimized/ch/400ms/ls_eend_ch_400ms.mlmodelc/metadata.json index 1479f59b57daba4565d10a86496002152bdb5208..b0e2f6533304881151f96ea28c711c20a4457ff4 100644 --- a/optimized/ch/400ms/ls_eend_ch_400ms.mlmodelc/metadata.json +++ b/optimized/ch/400ms/ls_eend_ch_400ms.mlmodelc/metadata.json @@ -1,7 +1,7 @@ [ { "metadataOutputVersion" : "3.0", - "shortDescription" : "LS-EEND CALLHOME streaming diarizer (pipeline, T=4, max_speakers=7)", + "shortDescription" : "LS-EEND CALLHOME streaming diarizer (pipeline, T=4, max_speakers=7, layout=raw_mel, cu=all)", "outputSchema" : [ { "hasShapeFlexibility" : "0", @@ -81,7 +81,7 @@ "specificationVersion" : 8, "mlProgramOperationTypeHistogram" : { "Ios16.reduceL2Norm" : 2, - "Ios17.reshape" : 68, + "Ios17.reshape" : 69, "Ios16.softmax" : 2, "Ios17.matmul" : 29, "Ios17.transpose" : 57, @@ -89,7 +89,7 @@ "Ios17.expandDims" : 3, "Ios17.add" : 46, "Ios16.sigmoid" : 5, - "Ios17.sliceByIndex" : 64, + "Ios17.sliceByIndex" : 68, "Tile" : 1, "Ios16.reduceSum" : 1, "Ios17.squeeze" : 2, @@ -101,7 +101,7 @@ "Ios16.silu" : 18, "Ios17.realDiv" : 20, "Ios17.linear" : 56, - "Stack" : 5, + "Stack" : 6, "Ios17.concat" : 22, "Ios16.relu" : 2, "Ios16.cumsum" : 1, @@ -128,9 +128,9 @@ "hasShapeFlexibility" : "0", "isOptional" : "0", "dataType" : "Float32", - "formattedType" : "MultiArray (Float32 1 × 4 × 345)", + "formattedType" : "MultiArray (Float32 1 × 45 × 23)", "shortDescription" : "", - "shape" : "[1, 4, 345]", + "shape" : "[1, 45, 23]", "name" : "features", "type" : "MultiArray" }, @@ -206,8 +206,8 @@ } ], "userDefinedMetadata" : { - "com.github.apple.coremltools.conversion_date" : "2026-04-16", - "config" : "{\"model_name\": \"ch\", \"model_label\": \"CALLHOME\", \"variant\": \"pipeline\", \"chunk_size\": 4, \"step_duration_ms\": 400, \"frame_duration_ms\": 100, \"frame_duration_seconds\": 0.1, \"max_speakers\": 7, \"max_nspks\": 9, \"n_units\": 256, \"n_heads\": 4, \"enc_n_layers\": 4, \"dec_n_layers\": 2, \"conv_kernel_size\": 16, \"conv_delay\": 9, \"sample_rate\": 8000, \"win_length\": 200, \"hop_length\": 80, \"n_mels\": 23, \"context_size\": 7, \"subsampling\": 10, \"feat_type\": \"logmel23_cummn\", \"pure_roll\": true}", + "com.github.apple.coremltools.conversion_date" : "2026-04-18", + "config" : "{\"model_name\": \"ch\", \"model_label\": \"CALLHOME\", \"variant\": \"pipeline\", \"chunk_size\": 4, \"step_duration_ms\": 400, \"frame_duration_ms\": 100, \"frame_duration_seconds\": 0.1, \"max_speakers\": 7, \"max_nspks\": 9, \"n_units\": 256, \"n_heads\": 4, \"enc_n_layers\": 4, \"dec_n_layers\": 2, \"conv_kernel_size\": 16, \"conv_delay\": 9, \"sample_rate\": 8000, \"win_length\": 200, \"hop_length\": 80, \"n_mels\": 23, \"context_size\": 7, \"subsampling\": 10, \"feat_type\": \"logmel23_cummn\", \"pure_roll\": true, \"input_layout\": \"raw_mel\", \"compute_units_export\": \"all\", \"raw_mel_length\": 45}", "com.github.apple.coremltools.source" : "torch==2.6.0", "com.github.apple.coremltools.version" : "9.0", "com.github.apple.coremltools.source_dialect" : "TorchScript" diff --git a/optimized/ch/400ms/ls_eend_ch_400ms.mlmodelc/model.mil b/optimized/ch/400ms/ls_eend_ch_400ms.mlmodelc/model.mil index 8e884a60ea34232c188dc4cdb669760172249293..0c554f0ab421d447767fe6f10a2fde3b97d7301b 100644 --- a/optimized/ch/400ms/ls_eend_ch_400ms.mlmodelc/model.mil +++ b/optimized/ch/400ms/ls_eend_ch_400ms.mlmodelc/model.mil @@ -1,234 +1,256 @@ program(1.0) [buildInfo = dict, tensor>({{"coremlc-component-MIL", "3510.2.1"}, {"coremlc-version", "3500.32.1"}, {"coremltools-component-torch", "2.6.0"}, {"coremltools-source-dialect", "TorchScript"}, {"coremltools-version", "9.0"}})] { - func main(tensor cnn_window, tensor dec_kv, tensor dec_scale, tensor enc_conv_cache, tensor enc_kv, tensor enc_scale, tensor features, tensor valid_mask) { - tensor encoder_cnn_bias = const()[name = tensor("encoder_cnn_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(64)))]; - tensor encoder_cnn_weight = const()[name = tensor("encoder_cnn_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1152)))]; - tensor encoder__causal_mask = const()[name = tensor("encoder__causal_mask"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4981952)))]; - tensor encoder__t_index = const()[name = tensor("encoder__t_index"), val = tensor([0x1p+0, 0x1p+1, 0x1.8p+1, 0x1p+2])]; - tensor encoder_input_projection_linear_bias = const()[name = tensor("encoder_input_projection_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4982080)))]; - tensor encoder_input_projection_linear_weight = const()[name = tensor("encoder_input_projection_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4983168)))]; - tensor encoder_pre_layer_norm_bias = const()[name = tensor("encoder_pre_layer_norm_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(5336512)))]; - tensor encoder_pre_layer_norm_weight = const()[name = tensor("encoder_pre_layer_norm_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(5337600)))]; - tensor encoder_ffn1_0_module_sequential_0_bias = const()[name = tensor("encoder_ffn1_0_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(5338688)))]; - tensor encoder_ffn1_0_module_sequential_0_weight = const()[name = tensor("encoder_ffn1_0_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(5339776)))]; - tensor encoder_ffn1_0_module_sequential_1_linear_bias = const()[name = tensor("encoder_ffn1_0_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(5340864)))]; - tensor encoder_ffn1_0_module_sequential_1_linear_weight = const()[name = tensor("encoder_ffn1_0_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(5345024)))]; - tensor encoder_ffn1_0_module_sequential_4_linear_bias = const()[name = tensor("encoder_ffn1_0_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(6393664)))]; - tensor encoder_ffn1_0_module_sequential_4_linear_weight = const()[name = tensor("encoder_ffn1_0_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(6394752)))]; - tensor encoder_ret_lns_0_bias = const()[name = tensor("encoder_ret_lns_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7443392)))]; - tensor encoder_ret_lns_0_weight = const()[name = tensor("encoder_ret_lns_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7444480)))]; - tensor encoder_q_proj_0_bias = const()[name = tensor("encoder_q_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7445568)))]; - tensor encoder_q_proj_0_weight = const()[name = tensor("encoder_q_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7446656)))]; - tensor encoder_k_proj_0_bias = const()[name = tensor("encoder_k_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7708864)))]; - tensor encoder_k_proj_0_weight = const()[name = tensor("encoder_k_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7709952)))]; - tensor encoder_v_proj_0_bias = const()[name = tensor("encoder_v_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7972160)))]; - tensor encoder_v_proj_0_weight = const()[name = tensor("encoder_v_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7973248)))]; - tensor encoder_g_proj_0_bias = const()[name = tensor("encoder_g_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8235456)))]; - tensor encoder_g_proj_0_weight = const()[name = tensor("encoder_g_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8236544)))]; - tensor encoder_out_proj_0_bias = const()[name = tensor("encoder_out_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8498752)))]; - tensor encoder_out_proj_0_weight = const()[name = tensor("encoder_out_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8499840)))]; - tensor encoder_conv_module_0_sequential_0_bias = const()[name = tensor("encoder_conv_module_0_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8762048)))]; - tensor encoder_conv_module_0_sequential_0_weight = const()[name = tensor("encoder_conv_module_0_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8763136)))]; - tensor encoder_conv_module_0_sequential_2_conv_bias = const()[name = tensor("encoder_conv_module_0_sequential_2_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8764224)))]; - tensor encoder_conv_module_0_sequential_2_conv_weight = const()[name = tensor("encoder_conv_module_0_sequential_2_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8766336)))]; - tensor encoder_conv_module_0_sequential_4_conv_weight = const()[name = tensor("encoder_conv_module_0_sequential_4_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9290688)))]; - tensor encoder_conv_module_0_sequential_5_running_var = const()[name = tensor("encoder_conv_module_0_sequential_5_running_var"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9307136)))]; - tensor encoder_conv_module_0_sequential_5_running_mean = const()[name = tensor("encoder_conv_module_0_sequential_5_running_mean"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9308224)))]; - tensor encoder_conv_module_0_sequential_5_bias = const()[name = tensor("encoder_conv_module_0_sequential_5_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9309312)))]; - tensor encoder_conv_module_0_sequential_5_weight = const()[name = tensor("encoder_conv_module_0_sequential_5_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9310400)))]; - tensor encoder_conv_module_0_sequential_7_conv_bias = const()[name = tensor("encoder_conv_module_0_sequential_7_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9311488)))]; - tensor encoder_conv_module_0_sequential_7_conv_weight = const()[name = tensor("encoder_conv_module_0_sequential_7_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9312576)))]; - tensor encoder_ffn2_0_module_sequential_0_bias = const()[name = tensor("encoder_ffn2_0_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9574784)))]; - tensor encoder_ffn2_0_module_sequential_0_weight = const()[name = tensor("encoder_ffn2_0_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9575872)))]; - tensor encoder_ffn2_0_module_sequential_1_linear_bias = const()[name = tensor("encoder_ffn2_0_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9576960)))]; - tensor encoder_ffn2_0_module_sequential_1_linear_weight = const()[name = tensor("encoder_ffn2_0_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9581120)))]; - tensor encoder_ffn2_0_module_sequential_4_linear_bias = const()[name = tensor("encoder_ffn2_0_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(10629760)))]; - tensor encoder_ffn2_0_module_sequential_4_linear_weight = const()[name = tensor("encoder_ffn2_0_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(10630848)))]; - tensor encoder_layer_norm_0_bias = const()[name = tensor("encoder_layer_norm_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11679488)))]; - tensor encoder_layer_norm_0_weight = const()[name = tensor("encoder_layer_norm_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11680576)))]; - tensor encoder_ffn1_1_module_sequential_0_bias = const()[name = tensor("encoder_ffn1_1_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11681664)))]; - tensor encoder_ffn1_1_module_sequential_0_weight = const()[name = tensor("encoder_ffn1_1_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11682752)))]; - tensor encoder_ffn1_1_module_sequential_1_linear_bias = const()[name = tensor("encoder_ffn1_1_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11683840)))]; - tensor encoder_ffn1_1_module_sequential_1_linear_weight = const()[name = tensor("encoder_ffn1_1_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11688000)))]; - tensor encoder_ffn1_1_module_sequential_4_linear_bias = const()[name = tensor("encoder_ffn1_1_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(12736640)))]; - tensor encoder_ffn1_1_module_sequential_4_linear_weight = const()[name = tensor("encoder_ffn1_1_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(12737728)))]; - tensor encoder_ret_lns_1_bias = const()[name = tensor("encoder_ret_lns_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(13786368)))]; - tensor encoder_ret_lns_1_weight = const()[name = tensor("encoder_ret_lns_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(13787456)))]; - tensor encoder_q_proj_1_bias = const()[name = tensor("encoder_q_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(13788544)))]; - tensor encoder_q_proj_1_weight = const()[name = tensor("encoder_q_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(13789632)))]; - tensor encoder_k_proj_1_bias = const()[name = tensor("encoder_k_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14051840)))]; - tensor encoder_k_proj_1_weight = const()[name = tensor("encoder_k_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14052928)))]; - tensor encoder_v_proj_1_bias = const()[name = tensor("encoder_v_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14315136)))]; - tensor encoder_v_proj_1_weight = const()[name = tensor("encoder_v_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14316224)))]; - tensor encoder_g_proj_1_bias = const()[name = tensor("encoder_g_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14578432)))]; - tensor encoder_g_proj_1_weight = const()[name = tensor("encoder_g_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14579520)))]; - tensor encoder_out_proj_1_bias = const()[name = tensor("encoder_out_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14841728)))]; - tensor encoder_out_proj_1_weight = const()[name = tensor("encoder_out_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14842816)))]; - tensor encoder_conv_module_1_sequential_0_bias = const()[name = tensor("encoder_conv_module_1_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15105024)))]; - tensor encoder_conv_module_1_sequential_0_weight = const()[name = tensor("encoder_conv_module_1_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15106112)))]; - tensor encoder_conv_module_1_sequential_2_conv_bias = const()[name = tensor("encoder_conv_module_1_sequential_2_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15107200)))]; - tensor encoder_conv_module_1_sequential_2_conv_weight = const()[name = tensor("encoder_conv_module_1_sequential_2_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15109312)))]; - tensor encoder_conv_module_1_sequential_4_conv_weight = const()[name = tensor("encoder_conv_module_1_sequential_4_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15633664)))]; - tensor encoder_conv_module_1_sequential_5_running_var = const()[name = tensor("encoder_conv_module_1_sequential_5_running_var"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15650112)))]; - tensor encoder_conv_module_1_sequential_5_running_mean = const()[name = tensor("encoder_conv_module_1_sequential_5_running_mean"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15651200)))]; - tensor encoder_conv_module_1_sequential_5_bias = const()[name = tensor("encoder_conv_module_1_sequential_5_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15652288)))]; - tensor encoder_conv_module_1_sequential_5_weight = const()[name = tensor("encoder_conv_module_1_sequential_5_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15653376)))]; - tensor encoder_conv_module_1_sequential_7_conv_bias = const()[name = tensor("encoder_conv_module_1_sequential_7_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15654464)))]; - tensor encoder_conv_module_1_sequential_7_conv_weight = const()[name = tensor("encoder_conv_module_1_sequential_7_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15655552)))]; - tensor encoder_ffn2_1_module_sequential_0_bias = const()[name = tensor("encoder_ffn2_1_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15917760)))]; - tensor encoder_ffn2_1_module_sequential_0_weight = const()[name = tensor("encoder_ffn2_1_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15918848)))]; - tensor encoder_ffn2_1_module_sequential_1_linear_bias = const()[name = tensor("encoder_ffn2_1_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15919936)))]; - tensor encoder_ffn2_1_module_sequential_1_linear_weight = const()[name = tensor("encoder_ffn2_1_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15924096)))]; - tensor encoder_ffn2_1_module_sequential_4_linear_bias = const()[name = tensor("encoder_ffn2_1_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(16972736)))]; - tensor encoder_ffn2_1_module_sequential_4_linear_weight = const()[name = tensor("encoder_ffn2_1_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(16973824)))]; - tensor encoder_layer_norm_1_bias = const()[name = tensor("encoder_layer_norm_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18022464)))]; - tensor encoder_layer_norm_1_weight = const()[name = tensor("encoder_layer_norm_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18023552)))]; - tensor encoder_ffn1_2_module_sequential_0_bias = const()[name = tensor("encoder_ffn1_2_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18024640)))]; - tensor encoder_ffn1_2_module_sequential_0_weight = const()[name = tensor("encoder_ffn1_2_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18025728)))]; - tensor encoder_ffn1_2_module_sequential_1_linear_bias = const()[name = tensor("encoder_ffn1_2_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18026816)))]; - tensor encoder_ffn1_2_module_sequential_1_linear_weight = const()[name = tensor("encoder_ffn1_2_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18030976)))]; - tensor encoder_ffn1_2_module_sequential_4_linear_bias = const()[name = tensor("encoder_ffn1_2_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(19079616)))]; - tensor encoder_ffn1_2_module_sequential_4_linear_weight = const()[name = tensor("encoder_ffn1_2_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(19080704)))]; - tensor encoder_ret_lns_2_bias = const()[name = tensor("encoder_ret_lns_2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20129344)))]; - tensor encoder_ret_lns_2_weight = const()[name = tensor("encoder_ret_lns_2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20130432)))]; - tensor encoder_q_proj_2_bias = const()[name = tensor("encoder_q_proj_2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20131520)))]; - tensor encoder_q_proj_2_weight = const()[name = tensor("encoder_q_proj_2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20132608)))]; - tensor encoder_k_proj_2_bias = const()[name = tensor("encoder_k_proj_2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20394816)))]; - tensor encoder_k_proj_2_weight = const()[name = tensor("encoder_k_proj_2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20395904)))]; - tensor encoder_v_proj_2_bias = const()[name = tensor("encoder_v_proj_2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20658112)))]; - tensor encoder_v_proj_2_weight = const()[name = tensor("encoder_v_proj_2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20659200)))]; - tensor encoder_g_proj_2_bias = const()[name = tensor("encoder_g_proj_2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20921408)))]; - tensor encoder_g_proj_2_weight = const()[name = tensor("encoder_g_proj_2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20922496)))]; - tensor encoder_out_proj_2_bias = const()[name = tensor("encoder_out_proj_2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21184704)))]; - tensor encoder_out_proj_2_weight = const()[name = tensor("encoder_out_proj_2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21185792)))]; - tensor encoder_conv_module_2_sequential_0_bias = const()[name = tensor("encoder_conv_module_2_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21448000)))]; - tensor encoder_conv_module_2_sequential_0_weight = const()[name = tensor("encoder_conv_module_2_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21449088)))]; - tensor encoder_conv_module_2_sequential_2_conv_bias = const()[name = tensor("encoder_conv_module_2_sequential_2_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21450176)))]; - tensor encoder_conv_module_2_sequential_2_conv_weight = const()[name = tensor("encoder_conv_module_2_sequential_2_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21452288)))]; - tensor encoder_conv_module_2_sequential_4_conv_weight = const()[name = tensor("encoder_conv_module_2_sequential_4_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21976640)))]; - tensor encoder_conv_module_2_sequential_5_running_var = const()[name = tensor("encoder_conv_module_2_sequential_5_running_var"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21993088)))]; - tensor encoder_conv_module_2_sequential_5_running_mean = const()[name = tensor("encoder_conv_module_2_sequential_5_running_mean"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21994176)))]; - tensor encoder_conv_module_2_sequential_5_bias = const()[name = tensor("encoder_conv_module_2_sequential_5_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21995264)))]; - tensor encoder_conv_module_2_sequential_5_weight = const()[name = tensor("encoder_conv_module_2_sequential_5_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21996352)))]; - tensor encoder_conv_module_2_sequential_7_conv_bias = const()[name = tensor("encoder_conv_module_2_sequential_7_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21997440)))]; - tensor encoder_conv_module_2_sequential_7_conv_weight = const()[name = tensor("encoder_conv_module_2_sequential_7_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21998528)))]; - tensor encoder_ffn2_2_module_sequential_0_bias = const()[name = tensor("encoder_ffn2_2_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(22260736)))]; - tensor encoder_ffn2_2_module_sequential_0_weight = const()[name = tensor("encoder_ffn2_2_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(22261824)))]; - tensor encoder_ffn2_2_module_sequential_1_linear_bias = const()[name = tensor("encoder_ffn2_2_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(22262912)))]; - tensor encoder_ffn2_2_module_sequential_1_linear_weight = const()[name = tensor("encoder_ffn2_2_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(22267072)))]; - tensor encoder_ffn2_2_module_sequential_4_linear_bias = const()[name = tensor("encoder_ffn2_2_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(23315712)))]; - tensor encoder_ffn2_2_module_sequential_4_linear_weight = const()[name = tensor("encoder_ffn2_2_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(23316800)))]; - tensor encoder_layer_norm_2_bias = const()[name = tensor("encoder_layer_norm_2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(24365440)))]; - tensor encoder_layer_norm_2_weight = const()[name = tensor("encoder_layer_norm_2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(24366528)))]; - tensor encoder_ffn1_3_module_sequential_0_bias = const()[name = tensor("encoder_ffn1_3_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(24367616)))]; - tensor encoder_ffn1_3_module_sequential_0_weight = const()[name = tensor("encoder_ffn1_3_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(24368704)))]; - tensor encoder_ffn1_3_module_sequential_1_linear_bias = const()[name = tensor("encoder_ffn1_3_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(24369792)))]; - tensor encoder_ffn1_3_module_sequential_1_linear_weight = const()[name = tensor("encoder_ffn1_3_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(24373952)))]; - tensor encoder_ffn1_3_module_sequential_4_linear_bias = const()[name = tensor("encoder_ffn1_3_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(25422592)))]; - tensor encoder_ffn1_3_module_sequential_4_linear_weight = const()[name = tensor("encoder_ffn1_3_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(25423680)))]; - tensor encoder_ret_lns_3_bias = const()[name = tensor("encoder_ret_lns_3_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(26472320)))]; - tensor encoder_ret_lns_3_weight = const()[name = tensor("encoder_ret_lns_3_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(26473408)))]; - tensor encoder_q_proj_3_bias = const()[name = tensor("encoder_q_proj_3_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(26474496)))]; - tensor encoder_q_proj_3_weight = const()[name = tensor("encoder_q_proj_3_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(26475584)))]; - tensor encoder_k_proj_3_bias = const()[name = tensor("encoder_k_proj_3_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(26737792)))]; - tensor encoder_k_proj_3_weight = const()[name = tensor("encoder_k_proj_3_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(26738880)))]; - tensor encoder_v_proj_3_bias = const()[name = tensor("encoder_v_proj_3_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27001088)))]; - tensor encoder_v_proj_3_weight = const()[name = tensor("encoder_v_proj_3_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27002176)))]; - tensor encoder_g_proj_3_bias = const()[name = tensor("encoder_g_proj_3_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27264384)))]; - tensor encoder_g_proj_3_weight = const()[name = tensor("encoder_g_proj_3_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27265472)))]; - tensor encoder_out_proj_3_bias = const()[name = tensor("encoder_out_proj_3_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27527680)))]; - tensor encoder_out_proj_3_weight = const()[name = tensor("encoder_out_proj_3_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27528768)))]; - tensor encoder_conv_module_3_sequential_0_bias = const()[name = tensor("encoder_conv_module_3_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27790976)))]; - tensor encoder_conv_module_3_sequential_0_weight = const()[name = tensor("encoder_conv_module_3_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27792064)))]; - tensor encoder_conv_module_3_sequential_2_conv_bias = const()[name = tensor("encoder_conv_module_3_sequential_2_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27793152)))]; - tensor encoder_conv_module_3_sequential_2_conv_weight = const()[name = tensor("encoder_conv_module_3_sequential_2_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27795264)))]; - tensor encoder_conv_module_3_sequential_4_conv_weight = const()[name = tensor("encoder_conv_module_3_sequential_4_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28319616)))]; - tensor encoder_conv_module_3_sequential_5_running_var = const()[name = tensor("encoder_conv_module_3_sequential_5_running_var"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28336064)))]; - tensor encoder_conv_module_3_sequential_5_running_mean = const()[name = tensor("encoder_conv_module_3_sequential_5_running_mean"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28337152)))]; - tensor encoder_conv_module_3_sequential_5_bias = const()[name = tensor("encoder_conv_module_3_sequential_5_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28338240)))]; - tensor encoder_conv_module_3_sequential_5_weight = const()[name = tensor("encoder_conv_module_3_sequential_5_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28339328)))]; - tensor encoder_conv_module_3_sequential_7_conv_bias = const()[name = tensor("encoder_conv_module_3_sequential_7_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28340416)))]; - tensor encoder_conv_module_3_sequential_7_conv_weight = const()[name = tensor("encoder_conv_module_3_sequential_7_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28341504)))]; - tensor encoder_ffn2_3_module_sequential_0_bias = const()[name = tensor("encoder_ffn2_3_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28603712)))]; - tensor encoder_ffn2_3_module_sequential_0_weight = const()[name = tensor("encoder_ffn2_3_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28604800)))]; - tensor encoder_ffn2_3_module_sequential_1_linear_bias = const()[name = tensor("encoder_ffn2_3_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28605888)))]; - tensor encoder_ffn2_3_module_sequential_1_linear_weight = const()[name = tensor("encoder_ffn2_3_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28610048)))]; - tensor encoder_ffn2_3_module_sequential_4_linear_bias = const()[name = tensor("encoder_ffn2_3_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(29658688)))]; - tensor encoder_ffn2_3_module_sequential_4_linear_weight = const()[name = tensor("encoder_ffn2_3_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(29659776)))]; - tensor encoder_layer_norm_3_bias = const()[name = tensor("encoder_layer_norm_3_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(30708416)))]; - tensor encoder_layer_norm_3_weight = const()[name = tensor("encoder_layer_norm_3_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(30709504)))]; - tensor decoder__causal_mask = const()[name = tensor("decoder__causal_mask"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(30710592)))]; - tensor decoder_convert_bias = const()[name = tensor("decoder_convert_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(30710720)))]; - tensor decoder_convert_weight = const()[name = tensor("decoder_convert_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(30711808)))]; - tensor decoder_q_proj_0_bias = const()[name = tensor("decoder_q_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(31236160)))]; - tensor decoder_q_proj_0_weight = const()[name = tensor("decoder_q_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(31237248)))]; - tensor decoder_k_proj_0_bias = const()[name = tensor("decoder_k_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(31499456)))]; - tensor decoder_k_proj_0_weight = const()[name = tensor("decoder_k_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(31500544)))]; - tensor decoder_v_proj_0_bias = const()[name = tensor("decoder_v_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(31762752)))]; - tensor decoder_v_proj_0_weight = const()[name = tensor("decoder_v_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(31763840)))]; - tensor decoder_g_proj_0_bias = const()[name = tensor("decoder_g_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32026048)))]; - tensor decoder_g_proj_0_weight = const()[name = tensor("decoder_g_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32027136)))]; - tensor decoder_out_proj_0_bias = const()[name = tensor("decoder_out_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32289344)))]; - tensor decoder_out_proj_0_weight = const()[name = tensor("decoder_out_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32290432)))]; - tensor decoder_norm11_0_bias = const()[name = tensor("decoder_norm11_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32552640)))]; - tensor decoder_norm11_0_weight = const()[name = tensor("decoder_norm11_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32553728)))]; - tensor decoder_self_attn2_0_out_proj_bias = const()[name = tensor("decoder_self_attn2_0_out_proj_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32554816)))]; - tensor decoder_self_attn2_0_out_proj_weight = const()[name = tensor("decoder_self_attn2_0_out_proj_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32555904)))]; - tensor decoder_self_attn2_0_in_proj_bias = const()[name = tensor("decoder_self_attn2_0_in_proj_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32818112)))]; - tensor decoder_self_attn2_0_in_proj_weight = const()[name = tensor("decoder_self_attn2_0_in_proj_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32821248)))]; - tensor decoder_norm21_0_bias = const()[name = tensor("decoder_norm21_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(33607744)))]; - tensor decoder_norm21_0_weight = const()[name = tensor("decoder_norm21_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(33608832)))]; - tensor decoder_linear1_0_bias = const()[name = tensor("decoder_linear1_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(33609920)))]; - tensor decoder_linear1_0_weight = const()[name = tensor("decoder_linear1_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(33618176)))]; - tensor decoder_linear2_0_bias = const()[name = tensor("decoder_linear2_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(35715392)))]; - tensor decoder_linear2_0_weight = const()[name = tensor("decoder_linear2_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(35716480)))]; - tensor decoder_norm22_0_bias = const()[name = tensor("decoder_norm22_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(37813696)))]; - tensor decoder_norm22_0_weight = const()[name = tensor("decoder_norm22_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(37814784)))]; - tensor decoder_q_proj_1_bias = const()[name = tensor("decoder_q_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(37815872)))]; - tensor decoder_q_proj_1_weight = const()[name = tensor("decoder_q_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(37816960)))]; - tensor decoder_k_proj_1_bias = const()[name = tensor("decoder_k_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38079168)))]; - tensor decoder_k_proj_1_weight = const()[name = tensor("decoder_k_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38080256)))]; - tensor decoder_v_proj_1_bias = const()[name = tensor("decoder_v_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38342464)))]; - tensor decoder_v_proj_1_weight = const()[name = tensor("decoder_v_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38343552)))]; - tensor decoder_g_proj_1_bias = const()[name = tensor("decoder_g_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38605760)))]; - tensor decoder_g_proj_1_weight = const()[name = tensor("decoder_g_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38606848)))]; - tensor decoder_out_proj_1_bias = const()[name = tensor("decoder_out_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38869056)))]; - tensor decoder_out_proj_1_weight = const()[name = tensor("decoder_out_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38870144)))]; - tensor decoder_norm11_1_bias = const()[name = tensor("decoder_norm11_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(39132352)))]; - tensor decoder_norm11_1_weight = const()[name = tensor("decoder_norm11_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(39133440)))]; - tensor decoder_self_attn2_1_out_proj_bias = const()[name = tensor("decoder_self_attn2_1_out_proj_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(39134528)))]; - tensor decoder_self_attn2_1_out_proj_weight = const()[name = tensor("decoder_self_attn2_1_out_proj_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(39135616)))]; - tensor decoder_self_attn2_1_in_proj_bias = const()[name = tensor("decoder_self_attn2_1_in_proj_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(39397824)))]; - tensor decoder_self_attn2_1_in_proj_weight = const()[name = tensor("decoder_self_attn2_1_in_proj_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(39400960)))]; - tensor decoder_norm21_1_bias = const()[name = tensor("decoder_norm21_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(40187456)))]; - tensor decoder_norm21_1_weight = const()[name = tensor("decoder_norm21_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(40188544)))]; - tensor decoder_linear1_1_bias = const()[name = tensor("decoder_linear1_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(40189632)))]; - tensor decoder_linear1_1_weight = const()[name = tensor("decoder_linear1_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(40197888)))]; - tensor decoder_linear2_1_bias = const()[name = tensor("decoder_linear2_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(42295104)))]; - tensor decoder_linear2_1_weight = const()[name = tensor("decoder_linear2_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(42296192)))]; - tensor decoder_norm22_1_bias = const()[name = tensor("decoder_norm22_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(44393408)))]; - tensor decoder_norm22_1_weight = const()[name = tensor("decoder_norm22_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(44394496)))]; - tensor var_11 = const()[name = tensor("op_11"), val = tensor(0x1.5798eep-27)]; - tensor var_17 = const()[name = tensor("op_17"), val = tensor(0x1.0c6f7ap-20)]; - tensor var_20 = const()[name = tensor("op_20"), val = tensor(2)]; - tensor var_23 = const()[name = tensor("op_23"), val = tensor(0)]; - tensor var_26 = const()[name = tensor("op_26"), val = tensor(1)]; - tensor var_28 = const()[name = tensor("op_28"), val = tensor(0x1.4f8b58p-17)]; - tensor var_29 = const()[name = tensor("op_29"), val = tensor(true)]; - tensor input_1 = linear(bias = encoder_input_projection_linear_bias, weight = encoder_input_projection_linear_weight, x = features)[name = tensor("linear_0")]; - tensor input_3_axes_0 = const()[name = tensor("input_3_axes_0"), val = tensor([-1])]; - tensor input_3 = layer_norm(axes = input_3_axes_0, beta = encoder_pre_layer_norm_bias, epsilon = var_28, gamma = encoder_pre_layer_norm_weight, x = input_1)[name = tensor("input_3")]; + func main(tensor cnn_window, tensor dec_kv, tensor dec_scale, tensor enc_conv_cache, tensor enc_kv, tensor enc_scale, tensor features, tensor valid_mask) { + tensor inner_encoder_cnn_bias = const()[name = tensor("inner_encoder_cnn_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(64)))]; + tensor inner_encoder_cnn_weight = const()[name = tensor("inner_encoder_cnn_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1152)))]; + tensor inner_encoder__causal_mask = const()[name = tensor("inner_encoder__causal_mask"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4981952)))]; + tensor inner_encoder__t_index = const()[name = tensor("inner_encoder__t_index"), val = tensor([0x1p+0, 0x1p+1, 0x1.8p+1, 0x1p+2])]; + tensor inner_encoder_input_projection_linear_bias = const()[name = tensor("inner_encoder_input_projection_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4982080)))]; + tensor inner_encoder_input_projection_linear_weight = const()[name = tensor("inner_encoder_input_projection_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4983168)))]; + tensor inner_encoder_pre_layer_norm_bias = const()[name = tensor("inner_encoder_pre_layer_norm_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(5336512)))]; + tensor inner_encoder_pre_layer_norm_weight = const()[name = tensor("inner_encoder_pre_layer_norm_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(5337600)))]; + tensor inner_encoder_ffn1_0_module_sequential_0_bias = const()[name = tensor("inner_encoder_ffn1_0_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(5338688)))]; + tensor inner_encoder_ffn1_0_module_sequential_0_weight = const()[name = tensor("inner_encoder_ffn1_0_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(5339776)))]; + tensor inner_encoder_ffn1_0_module_sequential_1_linear_bias = const()[name = tensor("inner_encoder_ffn1_0_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(5340864)))]; + tensor inner_encoder_ffn1_0_module_sequential_1_linear_weight = const()[name = tensor("inner_encoder_ffn1_0_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(5345024)))]; + tensor inner_encoder_ffn1_0_module_sequential_4_linear_bias = const()[name = tensor("inner_encoder_ffn1_0_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(6393664)))]; + tensor inner_encoder_ffn1_0_module_sequential_4_linear_weight = const()[name = tensor("inner_encoder_ffn1_0_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(6394752)))]; + tensor inner_encoder_ret_lns_0_bias = const()[name = tensor("inner_encoder_ret_lns_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7443392)))]; + tensor inner_encoder_ret_lns_0_weight = const()[name = tensor("inner_encoder_ret_lns_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7444480)))]; + tensor inner_encoder_q_proj_0_bias = const()[name = tensor("inner_encoder_q_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7445568)))]; + tensor inner_encoder_q_proj_0_weight = const()[name = tensor("inner_encoder_q_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7446656)))]; + tensor inner_encoder_k_proj_0_bias = const()[name = tensor("inner_encoder_k_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7708864)))]; + tensor inner_encoder_k_proj_0_weight = const()[name = tensor("inner_encoder_k_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7709952)))]; + tensor inner_encoder_v_proj_0_bias = const()[name = tensor("inner_encoder_v_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7972160)))]; + tensor inner_encoder_v_proj_0_weight = const()[name = tensor("inner_encoder_v_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7973248)))]; + tensor inner_encoder_g_proj_0_bias = const()[name = tensor("inner_encoder_g_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8235456)))]; + tensor inner_encoder_g_proj_0_weight = const()[name = tensor("inner_encoder_g_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8236544)))]; + tensor inner_encoder_out_proj_0_bias = const()[name = tensor("inner_encoder_out_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8498752)))]; + tensor inner_encoder_out_proj_0_weight = const()[name = tensor("inner_encoder_out_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8499840)))]; + tensor inner_encoder_conv_module_0_sequential_0_bias = const()[name = tensor("inner_encoder_conv_module_0_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8762048)))]; + tensor inner_encoder_conv_module_0_sequential_0_weight = const()[name = tensor("inner_encoder_conv_module_0_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8763136)))]; + tensor inner_encoder_conv_module_0_sequential_2_conv_bias = const()[name = tensor("inner_encoder_conv_module_0_sequential_2_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8764224)))]; + tensor inner_encoder_conv_module_0_sequential_2_conv_weight = const()[name = tensor("inner_encoder_conv_module_0_sequential_2_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8766336)))]; + tensor inner_encoder_conv_module_0_sequential_4_conv_weight = const()[name = tensor("inner_encoder_conv_module_0_sequential_4_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9290688)))]; + tensor inner_encoder_conv_module_0_sequential_5_running_var = const()[name = tensor("inner_encoder_conv_module_0_sequential_5_running_var"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9307136)))]; + tensor inner_encoder_conv_module_0_sequential_5_running_mean = const()[name = tensor("inner_encoder_conv_module_0_sequential_5_running_mean"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9308224)))]; + tensor inner_encoder_conv_module_0_sequential_5_bias = const()[name = tensor("inner_encoder_conv_module_0_sequential_5_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9309312)))]; + tensor inner_encoder_conv_module_0_sequential_5_weight = const()[name = tensor("inner_encoder_conv_module_0_sequential_5_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9310400)))]; + tensor inner_encoder_conv_module_0_sequential_7_conv_bias = const()[name = tensor("inner_encoder_conv_module_0_sequential_7_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9311488)))]; + tensor inner_encoder_conv_module_0_sequential_7_conv_weight = const()[name = tensor("inner_encoder_conv_module_0_sequential_7_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9312576)))]; + tensor inner_encoder_ffn2_0_module_sequential_0_bias = const()[name = tensor("inner_encoder_ffn2_0_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9574784)))]; + tensor inner_encoder_ffn2_0_module_sequential_0_weight = const()[name = tensor("inner_encoder_ffn2_0_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9575872)))]; + tensor inner_encoder_ffn2_0_module_sequential_1_linear_bias = const()[name = tensor("inner_encoder_ffn2_0_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9576960)))]; + tensor inner_encoder_ffn2_0_module_sequential_1_linear_weight = const()[name = tensor("inner_encoder_ffn2_0_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9581120)))]; + tensor inner_encoder_ffn2_0_module_sequential_4_linear_bias = const()[name = tensor("inner_encoder_ffn2_0_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(10629760)))]; + tensor inner_encoder_ffn2_0_module_sequential_4_linear_weight = const()[name = tensor("inner_encoder_ffn2_0_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(10630848)))]; + tensor inner_encoder_layer_norm_0_bias = const()[name = tensor("inner_encoder_layer_norm_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11679488)))]; + tensor inner_encoder_layer_norm_0_weight = const()[name = tensor("inner_encoder_layer_norm_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11680576)))]; + tensor inner_encoder_ffn1_1_module_sequential_0_bias = const()[name = tensor("inner_encoder_ffn1_1_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11681664)))]; + tensor inner_encoder_ffn1_1_module_sequential_0_weight = const()[name = tensor("inner_encoder_ffn1_1_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11682752)))]; + tensor inner_encoder_ffn1_1_module_sequential_1_linear_bias = const()[name = tensor("inner_encoder_ffn1_1_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11683840)))]; + tensor inner_encoder_ffn1_1_module_sequential_1_linear_weight = const()[name = tensor("inner_encoder_ffn1_1_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11688000)))]; + tensor inner_encoder_ffn1_1_module_sequential_4_linear_bias = const()[name = tensor("inner_encoder_ffn1_1_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(12736640)))]; + tensor inner_encoder_ffn1_1_module_sequential_4_linear_weight = const()[name = tensor("inner_encoder_ffn1_1_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(12737728)))]; + tensor inner_encoder_ret_lns_1_bias = const()[name = tensor("inner_encoder_ret_lns_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(13786368)))]; + tensor inner_encoder_ret_lns_1_weight = const()[name = tensor("inner_encoder_ret_lns_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(13787456)))]; + tensor inner_encoder_q_proj_1_bias = const()[name = tensor("inner_encoder_q_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(13788544)))]; + tensor inner_encoder_q_proj_1_weight = const()[name = tensor("inner_encoder_q_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(13789632)))]; + tensor inner_encoder_k_proj_1_bias = const()[name = tensor("inner_encoder_k_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14051840)))]; + tensor inner_encoder_k_proj_1_weight = const()[name = tensor("inner_encoder_k_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14052928)))]; + tensor inner_encoder_v_proj_1_bias = const()[name = tensor("inner_encoder_v_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14315136)))]; + tensor inner_encoder_v_proj_1_weight = const()[name = tensor("inner_encoder_v_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14316224)))]; + tensor inner_encoder_g_proj_1_bias = const()[name = tensor("inner_encoder_g_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14578432)))]; + tensor inner_encoder_g_proj_1_weight = const()[name = tensor("inner_encoder_g_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14579520)))]; + tensor inner_encoder_out_proj_1_bias = const()[name = tensor("inner_encoder_out_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14841728)))]; + tensor inner_encoder_out_proj_1_weight = const()[name = tensor("inner_encoder_out_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14842816)))]; + tensor inner_encoder_conv_module_1_sequential_0_bias = const()[name = tensor("inner_encoder_conv_module_1_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15105024)))]; + tensor inner_encoder_conv_module_1_sequential_0_weight = const()[name = tensor("inner_encoder_conv_module_1_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15106112)))]; + tensor inner_encoder_conv_module_1_sequential_2_conv_bias = const()[name = tensor("inner_encoder_conv_module_1_sequential_2_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15107200)))]; + tensor inner_encoder_conv_module_1_sequential_2_conv_weight = const()[name = tensor("inner_encoder_conv_module_1_sequential_2_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15109312)))]; + tensor inner_encoder_conv_module_1_sequential_4_conv_weight = const()[name = tensor("inner_encoder_conv_module_1_sequential_4_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15633664)))]; + tensor inner_encoder_conv_module_1_sequential_5_running_var = const()[name = tensor("inner_encoder_conv_module_1_sequential_5_running_var"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15650112)))]; + tensor inner_encoder_conv_module_1_sequential_5_running_mean = const()[name = tensor("inner_encoder_conv_module_1_sequential_5_running_mean"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15651200)))]; + tensor inner_encoder_conv_module_1_sequential_5_bias = const()[name = tensor("inner_encoder_conv_module_1_sequential_5_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15652288)))]; + tensor inner_encoder_conv_module_1_sequential_5_weight = const()[name = tensor("inner_encoder_conv_module_1_sequential_5_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15653376)))]; + tensor inner_encoder_conv_module_1_sequential_7_conv_bias = const()[name = tensor("inner_encoder_conv_module_1_sequential_7_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15654464)))]; + tensor inner_encoder_conv_module_1_sequential_7_conv_weight = const()[name = tensor("inner_encoder_conv_module_1_sequential_7_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15655552)))]; + tensor inner_encoder_ffn2_1_module_sequential_0_bias = const()[name = tensor("inner_encoder_ffn2_1_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15917760)))]; + tensor inner_encoder_ffn2_1_module_sequential_0_weight = const()[name = tensor("inner_encoder_ffn2_1_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15918848)))]; + tensor inner_encoder_ffn2_1_module_sequential_1_linear_bias = const()[name = tensor("inner_encoder_ffn2_1_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15919936)))]; + tensor inner_encoder_ffn2_1_module_sequential_1_linear_weight = const()[name = tensor("inner_encoder_ffn2_1_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15924096)))]; + tensor inner_encoder_ffn2_1_module_sequential_4_linear_bias = const()[name = tensor("inner_encoder_ffn2_1_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(16972736)))]; + tensor inner_encoder_ffn2_1_module_sequential_4_linear_weight = const()[name = tensor("inner_encoder_ffn2_1_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(16973824)))]; + tensor inner_encoder_layer_norm_1_bias = const()[name = tensor("inner_encoder_layer_norm_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18022464)))]; + tensor inner_encoder_layer_norm_1_weight = const()[name = tensor("inner_encoder_layer_norm_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18023552)))]; + tensor inner_encoder_ffn1_2_module_sequential_0_bias = const()[name = tensor("inner_encoder_ffn1_2_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18024640)))]; + tensor inner_encoder_ffn1_2_module_sequential_0_weight = const()[name = tensor("inner_encoder_ffn1_2_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18025728)))]; + tensor inner_encoder_ffn1_2_module_sequential_1_linear_bias = const()[name = tensor("inner_encoder_ffn1_2_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18026816)))]; + tensor inner_encoder_ffn1_2_module_sequential_1_linear_weight = const()[name = tensor("inner_encoder_ffn1_2_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18030976)))]; + tensor inner_encoder_ffn1_2_module_sequential_4_linear_bias = const()[name = tensor("inner_encoder_ffn1_2_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(19079616)))]; + tensor inner_encoder_ffn1_2_module_sequential_4_linear_weight = const()[name = tensor("inner_encoder_ffn1_2_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(19080704)))]; + tensor inner_encoder_ret_lns_2_bias = const()[name = tensor("inner_encoder_ret_lns_2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20129344)))]; + tensor inner_encoder_ret_lns_2_weight = const()[name = tensor("inner_encoder_ret_lns_2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20130432)))]; + tensor inner_encoder_q_proj_2_bias = const()[name = tensor("inner_encoder_q_proj_2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20131520)))]; + tensor inner_encoder_q_proj_2_weight = const()[name = tensor("inner_encoder_q_proj_2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20132608)))]; + tensor inner_encoder_k_proj_2_bias = const()[name = tensor("inner_encoder_k_proj_2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20394816)))]; + tensor inner_encoder_k_proj_2_weight = const()[name = tensor("inner_encoder_k_proj_2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20395904)))]; + tensor inner_encoder_v_proj_2_bias = const()[name = tensor("inner_encoder_v_proj_2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20658112)))]; + tensor inner_encoder_v_proj_2_weight = const()[name = tensor("inner_encoder_v_proj_2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20659200)))]; + tensor inner_encoder_g_proj_2_bias = const()[name = tensor("inner_encoder_g_proj_2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20921408)))]; + tensor inner_encoder_g_proj_2_weight = const()[name = tensor("inner_encoder_g_proj_2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20922496)))]; + tensor inner_encoder_out_proj_2_bias = const()[name = tensor("inner_encoder_out_proj_2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21184704)))]; + tensor inner_encoder_out_proj_2_weight = const()[name = tensor("inner_encoder_out_proj_2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21185792)))]; + tensor inner_encoder_conv_module_2_sequential_0_bias = const()[name = tensor("inner_encoder_conv_module_2_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21448000)))]; + tensor inner_encoder_conv_module_2_sequential_0_weight = const()[name = tensor("inner_encoder_conv_module_2_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21449088)))]; + tensor inner_encoder_conv_module_2_sequential_2_conv_bias = const()[name = tensor("inner_encoder_conv_module_2_sequential_2_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21450176)))]; + tensor inner_encoder_conv_module_2_sequential_2_conv_weight = const()[name = tensor("inner_encoder_conv_module_2_sequential_2_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21452288)))]; + tensor inner_encoder_conv_module_2_sequential_4_conv_weight = const()[name = tensor("inner_encoder_conv_module_2_sequential_4_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21976640)))]; + tensor inner_encoder_conv_module_2_sequential_5_running_var = const()[name = tensor("inner_encoder_conv_module_2_sequential_5_running_var"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21993088)))]; + tensor inner_encoder_conv_module_2_sequential_5_running_mean = const()[name = tensor("inner_encoder_conv_module_2_sequential_5_running_mean"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21994176)))]; + tensor inner_encoder_conv_module_2_sequential_5_bias = const()[name = tensor("inner_encoder_conv_module_2_sequential_5_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21995264)))]; + tensor inner_encoder_conv_module_2_sequential_5_weight = const()[name = tensor("inner_encoder_conv_module_2_sequential_5_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21996352)))]; + tensor inner_encoder_conv_module_2_sequential_7_conv_bias = const()[name = tensor("inner_encoder_conv_module_2_sequential_7_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21997440)))]; + tensor inner_encoder_conv_module_2_sequential_7_conv_weight = const()[name = tensor("inner_encoder_conv_module_2_sequential_7_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21998528)))]; + tensor inner_encoder_ffn2_2_module_sequential_0_bias = const()[name = tensor("inner_encoder_ffn2_2_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(22260736)))]; + tensor inner_encoder_ffn2_2_module_sequential_0_weight = const()[name = tensor("inner_encoder_ffn2_2_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(22261824)))]; + tensor inner_encoder_ffn2_2_module_sequential_1_linear_bias = const()[name = tensor("inner_encoder_ffn2_2_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(22262912)))]; + tensor inner_encoder_ffn2_2_module_sequential_1_linear_weight = const()[name = tensor("inner_encoder_ffn2_2_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(22267072)))]; + tensor inner_encoder_ffn2_2_module_sequential_4_linear_bias = const()[name = tensor("inner_encoder_ffn2_2_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(23315712)))]; + tensor inner_encoder_ffn2_2_module_sequential_4_linear_weight = const()[name = tensor("inner_encoder_ffn2_2_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(23316800)))]; + tensor inner_encoder_layer_norm_2_bias = const()[name = tensor("inner_encoder_layer_norm_2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(24365440)))]; + tensor inner_encoder_layer_norm_2_weight = const()[name = tensor("inner_encoder_layer_norm_2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(24366528)))]; + tensor inner_encoder_ffn1_3_module_sequential_0_bias = const()[name = tensor("inner_encoder_ffn1_3_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(24367616)))]; + tensor inner_encoder_ffn1_3_module_sequential_0_weight = const()[name = tensor("inner_encoder_ffn1_3_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(24368704)))]; + tensor inner_encoder_ffn1_3_module_sequential_1_linear_bias = const()[name = tensor("inner_encoder_ffn1_3_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(24369792)))]; + tensor inner_encoder_ffn1_3_module_sequential_1_linear_weight = const()[name = tensor("inner_encoder_ffn1_3_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(24373952)))]; + tensor inner_encoder_ffn1_3_module_sequential_4_linear_bias = const()[name = tensor("inner_encoder_ffn1_3_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(25422592)))]; + tensor inner_encoder_ffn1_3_module_sequential_4_linear_weight = const()[name = tensor("inner_encoder_ffn1_3_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(25423680)))]; + tensor inner_encoder_ret_lns_3_bias = const()[name = tensor("inner_encoder_ret_lns_3_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(26472320)))]; + tensor inner_encoder_ret_lns_3_weight = const()[name = tensor("inner_encoder_ret_lns_3_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(26473408)))]; + tensor inner_encoder_q_proj_3_bias = const()[name = tensor("inner_encoder_q_proj_3_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(26474496)))]; + tensor inner_encoder_q_proj_3_weight = const()[name = tensor("inner_encoder_q_proj_3_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(26475584)))]; + tensor inner_encoder_k_proj_3_bias = const()[name = tensor("inner_encoder_k_proj_3_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(26737792)))]; + tensor inner_encoder_k_proj_3_weight = const()[name = tensor("inner_encoder_k_proj_3_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(26738880)))]; + tensor inner_encoder_v_proj_3_bias = const()[name = tensor("inner_encoder_v_proj_3_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27001088)))]; + tensor inner_encoder_v_proj_3_weight = const()[name = tensor("inner_encoder_v_proj_3_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27002176)))]; + tensor inner_encoder_g_proj_3_bias = const()[name = tensor("inner_encoder_g_proj_3_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27264384)))]; + tensor inner_encoder_g_proj_3_weight = const()[name = tensor("inner_encoder_g_proj_3_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27265472)))]; + tensor inner_encoder_out_proj_3_bias = const()[name = tensor("inner_encoder_out_proj_3_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27527680)))]; + tensor inner_encoder_out_proj_3_weight = const()[name = tensor("inner_encoder_out_proj_3_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27528768)))]; + tensor inner_encoder_conv_module_3_sequential_0_bias = const()[name = tensor("inner_encoder_conv_module_3_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27790976)))]; + tensor inner_encoder_conv_module_3_sequential_0_weight = const()[name = tensor("inner_encoder_conv_module_3_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27792064)))]; + tensor inner_encoder_conv_module_3_sequential_2_conv_bias = const()[name = tensor("inner_encoder_conv_module_3_sequential_2_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27793152)))]; + tensor inner_encoder_conv_module_3_sequential_2_conv_weight = const()[name = tensor("inner_encoder_conv_module_3_sequential_2_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27795264)))]; + tensor inner_encoder_conv_module_3_sequential_4_conv_weight = const()[name = tensor("inner_encoder_conv_module_3_sequential_4_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28319616)))]; + tensor inner_encoder_conv_module_3_sequential_5_running_var = const()[name = tensor("inner_encoder_conv_module_3_sequential_5_running_var"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28336064)))]; + tensor inner_encoder_conv_module_3_sequential_5_running_mean = const()[name = tensor("inner_encoder_conv_module_3_sequential_5_running_mean"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28337152)))]; + tensor inner_encoder_conv_module_3_sequential_5_bias = const()[name = tensor("inner_encoder_conv_module_3_sequential_5_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28338240)))]; + tensor inner_encoder_conv_module_3_sequential_5_weight = const()[name = tensor("inner_encoder_conv_module_3_sequential_5_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28339328)))]; + tensor inner_encoder_conv_module_3_sequential_7_conv_bias = const()[name = tensor("inner_encoder_conv_module_3_sequential_7_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28340416)))]; + tensor inner_encoder_conv_module_3_sequential_7_conv_weight = const()[name = tensor("inner_encoder_conv_module_3_sequential_7_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28341504)))]; + tensor inner_encoder_ffn2_3_module_sequential_0_bias = const()[name = tensor("inner_encoder_ffn2_3_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28603712)))]; + tensor inner_encoder_ffn2_3_module_sequential_0_weight = const()[name = tensor("inner_encoder_ffn2_3_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28604800)))]; + tensor inner_encoder_ffn2_3_module_sequential_1_linear_bias = const()[name = tensor("inner_encoder_ffn2_3_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28605888)))]; + tensor inner_encoder_ffn2_3_module_sequential_1_linear_weight = const()[name = tensor("inner_encoder_ffn2_3_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28610048)))]; + tensor inner_encoder_ffn2_3_module_sequential_4_linear_bias = const()[name = tensor("inner_encoder_ffn2_3_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(29658688)))]; + tensor inner_encoder_ffn2_3_module_sequential_4_linear_weight = const()[name = tensor("inner_encoder_ffn2_3_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(29659776)))]; + tensor inner_encoder_layer_norm_3_bias = const()[name = tensor("inner_encoder_layer_norm_3_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(30708416)))]; + tensor inner_encoder_layer_norm_3_weight = const()[name = tensor("inner_encoder_layer_norm_3_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(30709504)))]; + tensor inner_decoder__causal_mask = const()[name = tensor("inner_decoder__causal_mask"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(30710592)))]; + tensor inner_decoder_convert_bias = const()[name = tensor("inner_decoder_convert_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(30710720)))]; + tensor inner_decoder_convert_weight = const()[name = tensor("inner_decoder_convert_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(30711808)))]; + tensor inner_decoder_q_proj_0_bias = const()[name = tensor("inner_decoder_q_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(31236160)))]; + tensor inner_decoder_q_proj_0_weight = const()[name = tensor("inner_decoder_q_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(31237248)))]; + tensor inner_decoder_k_proj_0_bias = const()[name = tensor("inner_decoder_k_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(31499456)))]; + tensor inner_decoder_k_proj_0_weight = const()[name = tensor("inner_decoder_k_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(31500544)))]; + tensor inner_decoder_v_proj_0_bias = const()[name = tensor("inner_decoder_v_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(31762752)))]; + tensor inner_decoder_v_proj_0_weight = const()[name = tensor("inner_decoder_v_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(31763840)))]; + tensor inner_decoder_g_proj_0_bias = const()[name = tensor("inner_decoder_g_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32026048)))]; + tensor inner_decoder_g_proj_0_weight = const()[name = tensor("inner_decoder_g_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32027136)))]; + tensor inner_decoder_out_proj_0_bias = const()[name = tensor("inner_decoder_out_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32289344)))]; + tensor inner_decoder_out_proj_0_weight = const()[name = tensor("inner_decoder_out_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32290432)))]; + tensor inner_decoder_norm11_0_bias = const()[name = tensor("inner_decoder_norm11_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32552640)))]; + tensor inner_decoder_norm11_0_weight = const()[name = tensor("inner_decoder_norm11_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32553728)))]; + tensor inner_decoder_self_attn2_0_out_proj_bias = const()[name = tensor("inner_decoder_self_attn2_0_out_proj_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32554816)))]; + tensor inner_decoder_self_attn2_0_out_proj_weight = const()[name = tensor("inner_decoder_self_attn2_0_out_proj_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32555904)))]; + tensor inner_decoder_self_attn2_0_in_proj_bias = const()[name = tensor("inner_decoder_self_attn2_0_in_proj_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32818112)))]; + tensor inner_decoder_self_attn2_0_in_proj_weight = const()[name = tensor("inner_decoder_self_attn2_0_in_proj_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32821248)))]; + tensor inner_decoder_norm21_0_bias = const()[name = tensor("inner_decoder_norm21_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(33607744)))]; + tensor inner_decoder_norm21_0_weight = const()[name = tensor("inner_decoder_norm21_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(33608832)))]; + tensor inner_decoder_linear1_0_bias = const()[name = tensor("inner_decoder_linear1_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(33609920)))]; + tensor inner_decoder_linear1_0_weight = const()[name = tensor("inner_decoder_linear1_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(33618176)))]; + tensor inner_decoder_linear2_0_bias = const()[name = tensor("inner_decoder_linear2_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(35715392)))]; + tensor inner_decoder_linear2_0_weight = const()[name = tensor("inner_decoder_linear2_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(35716480)))]; + tensor inner_decoder_norm22_0_bias = const()[name = tensor("inner_decoder_norm22_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(37813696)))]; + tensor inner_decoder_norm22_0_weight = const()[name = tensor("inner_decoder_norm22_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(37814784)))]; + tensor inner_decoder_q_proj_1_bias = const()[name = tensor("inner_decoder_q_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(37815872)))]; + tensor inner_decoder_q_proj_1_weight = const()[name = tensor("inner_decoder_q_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(37816960)))]; + tensor inner_decoder_k_proj_1_bias = const()[name = tensor("inner_decoder_k_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38079168)))]; + tensor inner_decoder_k_proj_1_weight = const()[name = tensor("inner_decoder_k_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38080256)))]; + tensor inner_decoder_v_proj_1_bias = const()[name = tensor("inner_decoder_v_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38342464)))]; + tensor inner_decoder_v_proj_1_weight = const()[name = tensor("inner_decoder_v_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38343552)))]; + tensor inner_decoder_g_proj_1_bias = const()[name = tensor("inner_decoder_g_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38605760)))]; + tensor inner_decoder_g_proj_1_weight = const()[name = tensor("inner_decoder_g_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38606848)))]; + tensor inner_decoder_out_proj_1_bias = const()[name = tensor("inner_decoder_out_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38869056)))]; + tensor inner_decoder_out_proj_1_weight = const()[name = tensor("inner_decoder_out_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38870144)))]; + tensor inner_decoder_norm11_1_bias = const()[name = tensor("inner_decoder_norm11_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(39132352)))]; + tensor inner_decoder_norm11_1_weight = const()[name = tensor("inner_decoder_norm11_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(39133440)))]; + tensor inner_decoder_self_attn2_1_out_proj_bias = const()[name = tensor("inner_decoder_self_attn2_1_out_proj_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(39134528)))]; + tensor inner_decoder_self_attn2_1_out_proj_weight = const()[name = tensor("inner_decoder_self_attn2_1_out_proj_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(39135616)))]; + tensor inner_decoder_self_attn2_1_in_proj_bias = const()[name = tensor("inner_decoder_self_attn2_1_in_proj_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(39397824)))]; + tensor inner_decoder_self_attn2_1_in_proj_weight = const()[name = tensor("inner_decoder_self_attn2_1_in_proj_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(39400960)))]; + tensor inner_decoder_norm21_1_bias = const()[name = tensor("inner_decoder_norm21_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(40187456)))]; + tensor inner_decoder_norm21_1_weight = const()[name = tensor("inner_decoder_norm21_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(40188544)))]; + tensor inner_decoder_linear1_1_bias = const()[name = tensor("inner_decoder_linear1_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(40189632)))]; + tensor inner_decoder_linear1_1_weight = const()[name = tensor("inner_decoder_linear1_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(40197888)))]; + tensor inner_decoder_linear2_1_bias = const()[name = tensor("inner_decoder_linear2_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(42295104)))]; + tensor inner_decoder_linear2_1_weight = const()[name = tensor("inner_decoder_linear2_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(42296192)))]; + tensor inner_decoder_norm22_1_bias = const()[name = tensor("inner_decoder_norm22_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(44393408)))]; + tensor inner_decoder_norm22_1_weight = const()[name = tensor("inner_decoder_norm22_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(44394496)))]; + tensor var_19_begin_0 = const()[name = tensor("op_19_begin_0"), val = tensor([0, 0, 0])]; + tensor var_19_end_0 = const()[name = tensor("op_19_end_0"), val = tensor([1, 15, 23])]; + tensor var_19_end_mask_0 = const()[name = tensor("op_19_end_mask_0"), val = tensor([true, false, true])]; + tensor var_19 = slice_by_index(begin = var_19_begin_0, end = var_19_end_0, end_mask = var_19_end_mask_0, x = features)[name = tensor("op_19")]; + tensor var_29_begin_0 = const()[name = tensor("op_29_begin_0"), val = tensor([0, 10, 0])]; + tensor var_29_end_0 = const()[name = tensor("op_29_end_0"), val = tensor([1, 25, 23])]; + tensor var_29_end_mask_0 = const()[name = tensor("op_29_end_mask_0"), val = tensor([true, false, true])]; + tensor var_29 = slice_by_index(begin = var_29_begin_0, end = var_29_end_0, end_mask = var_29_end_mask_0, x = features)[name = tensor("op_29")]; + tensor var_39_begin_0 = const()[name = tensor("op_39_begin_0"), val = tensor([0, 20, 0])]; + tensor var_39_end_0 = const()[name = tensor("op_39_end_0"), val = tensor([1, 35, 23])]; + tensor var_39_end_mask_0 = const()[name = tensor("op_39_end_mask_0"), val = tensor([true, false, true])]; + tensor var_39 = slice_by_index(begin = var_39_begin_0, end = var_39_end_0, end_mask = var_39_end_mask_0, x = features)[name = tensor("op_39")]; + tensor var_49_begin_0 = const()[name = tensor("op_49_begin_0"), val = tensor([0, 30, 0])]; + tensor var_49_end_0 = const()[name = tensor("op_49_end_0"), val = tensor([1, 1, 23])]; + tensor var_49_end_mask_0 = const()[name = tensor("op_49_end_mask_0"), val = tensor([true, true, true])]; + tensor var_49 = slice_by_index(begin = var_49_begin_0, end = var_49_end_0, end_mask = var_49_end_mask_0, x = features)[name = tensor("op_49")]; + tensor stacked_axis_0 = const()[name = tensor("stacked_axis_0"), val = tensor(1)]; + tensor stacked = stack(axis = stacked_axis_0, values = (var_19, var_29, var_39, var_49))[name = tensor("stacked")]; + tensor var_56 = const()[name = tensor("op_56"), val = tensor([1, 4, 345])]; + tensor input_1 = reshape(shape = var_56, x = stacked)[name = tensor("input_1")]; + tensor var_59 = const()[name = tensor("op_59"), val = tensor(0x1p+0)]; + tensor var_65 = const()[name = tensor("op_65"), val = tensor(true)]; + tensor var_66 = const()[name = tensor("op_66"), val = tensor(0x1.4f8b58p-17)]; + tensor var_69 = const()[name = tensor("op_69"), val = tensor(0)]; + tensor var_71 = const()[name = tensor("op_71"), val = tensor(2)]; + tensor var_72 = const()[name = tensor("op_72"), val = tensor(-1)]; + tensor var_74 = const()[name = tensor("op_74"), val = tensor(0x1.0c6f7ap-20)]; + tensor var_79 = const()[name = tensor("op_79"), val = tensor(0x1.5798eep-27)]; + tensor var_82 = const()[name = tensor("op_82"), val = tensor(1)]; + tensor input_3 = linear(bias = inner_encoder_input_projection_linear_bias, weight = inner_encoder_input_projection_linear_weight, x = input_1)[name = tensor("linear_0")]; tensor input_5_axes_0 = const()[name = tensor("input_5_axes_0"), val = tensor([-1])]; - tensor input_5 = layer_norm(axes = input_5_axes_0, beta = encoder_ffn1_0_module_sequential_0_bias, epsilon = var_28, gamma = encoder_ffn1_0_module_sequential_0_weight, x = input_3)[name = tensor("input_5")]; - tensor inputs_1 = linear(bias = encoder_ffn1_0_module_sequential_1_linear_bias, weight = encoder_ffn1_0_module_sequential_1_linear_weight, x = input_5)[name = tensor("linear_1")]; - tensor input_7 = silu(x = inputs_1)[name = tensor("input_7")]; - tensor input_11 = linear(bias = encoder_ffn1_0_module_sequential_4_linear_bias, weight = encoder_ffn1_0_module_sequential_4_linear_weight, x = input_7)[name = tensor("linear_2")]; - tensor var_147 = const()[name = tensor("op_147"), val = tensor(0x1p-1)]; - tensor var_148 = mul(x = input_11, y = var_147)[name = tensor("op_148")]; - tensor input_13 = add(x = var_148, y = input_3)[name = tensor("input_13")]; + tensor input_5 = layer_norm(axes = input_5_axes_0, beta = inner_encoder_pre_layer_norm_bias, epsilon = var_66, gamma = inner_encoder_pre_layer_norm_weight, x = input_3)[name = tensor("input_5")]; + tensor input_7_axes_0 = const()[name = tensor("input_7_axes_0"), val = tensor([-1])]; + tensor input_7 = layer_norm(axes = input_7_axes_0, beta = inner_encoder_ffn1_0_module_sequential_0_bias, epsilon = var_66, gamma = inner_encoder_ffn1_0_module_sequential_0_weight, x = input_5)[name = tensor("input_7")]; + tensor inputs_1 = linear(bias = inner_encoder_ffn1_0_module_sequential_1_linear_bias, weight = inner_encoder_ffn1_0_module_sequential_1_linear_weight, x = input_7)[name = tensor("linear_1")]; + tensor input_9 = silu(x = inputs_1)[name = tensor("input_9")]; + tensor input_13 = linear(bias = inner_encoder_ffn1_0_module_sequential_4_linear_bias, weight = inner_encoder_ffn1_0_module_sequential_4_linear_weight, x = input_9)[name = tensor("linear_2")]; + tensor var_203 = const()[name = tensor("op_203"), val = tensor(0x1p-1)]; + tensor var_204 = mul(x = input_13, y = var_203)[name = tensor("op_204")]; + tensor input_15 = add(x = var_204, y = input_5)[name = tensor("input_15")]; tensor x_1_axes_0 = const()[name = tensor("x_1_axes_0"), val = tensor([-1])]; - tensor x_1 = layer_norm(axes = x_1_axes_0, beta = encoder_ret_lns_0_bias, epsilon = var_28, gamma = encoder_ret_lns_0_weight, x = input_13)[name = tensor("x_1")]; + tensor x_1 = layer_norm(axes = x_1_axes_0, beta = inner_encoder_ret_lns_0_bias, epsilon = var_66, gamma = inner_encoder_ret_lns_0_weight, x = input_15)[name = tensor("x_1")]; tensor prev_kv_1_begin_0 = const()[name = tensor("prev_kv_1_begin_0"), val = tensor([0, 0, 0, 0, 0])]; tensor prev_kv_1_end_0 = const()[name = tensor("prev_kv_1_end_0"), val = tensor([1, 1, 4, 64, 64])]; tensor prev_kv_1_end_mask_0 = const()[name = tensor("prev_kv_1_end_mask_0"), val = tensor([false, true, true, true, true])]; @@ -239,173 +261,173 @@ program(1.0) tensor prev_scale_1_end_mask_0 = const()[name = tensor("prev_scale_1_end_mask_0"), val = tensor([false, true])]; tensor prev_scale_1_squeeze_mask_0 = const()[name = tensor("prev_scale_1_squeeze_mask_0"), val = tensor([true, false])]; tensor prev_scale_1 = slice_by_index(begin = prev_scale_1_begin_0, end = prev_scale_1_end_0, end_mask = prev_scale_1_end_mask_0, squeeze_mask = prev_scale_1_squeeze_mask_0, x = enc_scale)[name = tensor("prev_scale_1")]; - tensor var_162 = linear(bias = encoder_q_proj_0_bias, weight = encoder_q_proj_0_weight, x = x_1)[name = tensor("linear_3")]; - tensor var_163 = const()[name = tensor("op_163"), val = tensor([1, 4, 4, 64])]; - tensor var_164 = reshape(shape = var_163, x = var_162)[name = tensor("op_164")]; + tensor var_218 = linear(bias = inner_encoder_q_proj_0_bias, weight = inner_encoder_q_proj_0_weight, x = x_1)[name = tensor("linear_3")]; + tensor var_219 = const()[name = tensor("op_219"), val = tensor([1, 4, 4, 64])]; + tensor var_220 = reshape(shape = var_219, x = var_218)[name = tensor("op_220")]; tensor q_1_perm_0 = const()[name = tensor("q_1_perm_0"), val = tensor([0, 2, 1, 3])]; - tensor var_168 = linear(bias = encoder_k_proj_0_bias, weight = encoder_k_proj_0_weight, x = x_1)[name = tensor("linear_4")]; - tensor var_169 = const()[name = tensor("op_169"), val = tensor(0x1p-3)]; - tensor var_170 = mul(x = var_168, y = var_169)[name = tensor("op_170")]; - tensor var_171 = const()[name = tensor("op_171"), val = tensor([1, 4, 4, 64])]; - tensor var_172 = reshape(shape = var_171, x = var_170)[name = tensor("op_172")]; + tensor var_224 = linear(bias = inner_encoder_k_proj_0_bias, weight = inner_encoder_k_proj_0_weight, x = x_1)[name = tensor("linear_4")]; + tensor var_225 = const()[name = tensor("op_225"), val = tensor(0x1p-3)]; + tensor var_226 = mul(x = var_224, y = var_225)[name = tensor("op_226")]; + tensor var_227 = const()[name = tensor("op_227"), val = tensor([1, 4, 4, 64])]; + tensor var_228 = reshape(shape = var_227, x = var_226)[name = tensor("op_228")]; tensor k_1_perm_0 = const()[name = tensor("k_1_perm_0"), val = tensor([0, 2, 1, 3])]; - tensor var_176 = linear(bias = encoder_v_proj_0_bias, weight = encoder_v_proj_0_weight, x = x_1)[name = tensor("linear_5")]; - tensor var_177 = const()[name = tensor("op_177"), val = tensor([1, 4, 4, 64])]; - tensor var_178 = reshape(shape = var_177, x = var_176)[name = tensor("op_178")]; + tensor var_232 = linear(bias = inner_encoder_v_proj_0_bias, weight = inner_encoder_v_proj_0_weight, x = x_1)[name = tensor("linear_5")]; + tensor var_233 = const()[name = tensor("op_233"), val = tensor([1, 4, 4, 64])]; + tensor var_234 = reshape(shape = var_233, x = var_232)[name = tensor("op_234")]; tensor v_1_perm_0 = const()[name = tensor("v_1_perm_0"), val = tensor([0, 2, 1, 3])]; - tensor input_17 = linear(bias = encoder_g_proj_0_bias, weight = encoder_g_proj_0_weight, x = x_1)[name = tensor("linear_6")]; + tensor input_19 = linear(bias = inner_encoder_g_proj_0_bias, weight = inner_encoder_g_proj_0_weight, x = x_1)[name = tensor("linear_6")]; tensor sqrt_s0_1 = sqrt(x = prev_scale_1)[name = tensor("sqrt_s0_1")]; - tensor s_t_1 = add(x = prev_scale_1, y = encoder__t_index)[name = tensor("s_t_1")]; + tensor s_t_1 = add(x = prev_scale_1, y = inner_encoder__t_index)[name = tensor("s_t_1")]; tensor sqrt_s_t_1 = sqrt(x = s_t_1)[name = tensor("sqrt_s_t_1")]; tensor qk_1_transpose_x_1 = const()[name = tensor("qk_1_transpose_x_1"), val = tensor(false)]; tensor qk_1_transpose_y_1 = const()[name = tensor("qk_1_transpose_y_1"), val = tensor(true)]; - tensor k_1 = transpose(perm = k_1_perm_0, x = var_172)[name = tensor("transpose_57")]; - tensor q_1 = transpose(perm = q_1_perm_0, x = var_164)[name = tensor("transpose_58")]; + tensor k_1 = transpose(perm = k_1_perm_0, x = var_228)[name = tensor("transpose_57")]; + tensor q_1 = transpose(perm = q_1_perm_0, x = var_220)[name = tensor("transpose_58")]; tensor qk_1 = matmul(transpose_x = qk_1_transpose_x_1, transpose_y = qk_1_transpose_y_1, x = q_1, y = k_1)[name = tensor("qk_1")]; - tensor var_188 = const()[name = tensor("op_188"), val = tensor([4, 1])]; - tensor var_189 = reshape(shape = var_188, x = sqrt_s_t_1)[name = tensor("op_189")]; - tensor M_1 = real_div(x = encoder__causal_mask, y = var_189)[name = tensor("M_1")]; - tensor var_191 = mul(x = qk_1, y = M_1)[name = tensor("op_191")]; + tensor var_244 = const()[name = tensor("op_244"), val = tensor([4, 1])]; + tensor var_245 = reshape(shape = var_244, x = sqrt_s_t_1)[name = tensor("op_245")]; + tensor M_1 = real_div(x = inner_encoder__causal_mask, y = var_245)[name = tensor("M_1")]; + tensor var_247 = mul(x = qk_1, y = M_1)[name = tensor("op_247")]; tensor inner_1_transpose_x_0 = const()[name = tensor("inner_1_transpose_x_0"), val = tensor(false)]; tensor inner_1_transpose_y_0 = const()[name = tensor("inner_1_transpose_y_0"), val = tensor(false)]; - tensor v_1 = transpose(perm = v_1_perm_0, x = var_178)[name = tensor("transpose_56")]; - tensor inner_1 = matmul(transpose_x = inner_1_transpose_x_0, transpose_y = inner_1_transpose_y_0, x = var_191, y = v_1)[name = tensor("inner_1")]; - tensor var_193_transpose_x_0 = const()[name = tensor("op_193_transpose_x_0"), val = tensor(false)]; - tensor var_193_transpose_y_0 = const()[name = tensor("op_193_transpose_y_0"), val = tensor(false)]; - tensor var_193 = matmul(transpose_x = var_193_transpose_x_0, transpose_y = var_193_transpose_y_0, x = q_1, y = prev_kv_1)[name = tensor("op_193")]; - tensor var_194 = real_div(x = sqrt_s0_1, y = sqrt_s_t_1)[name = tensor("op_194")]; - tensor var_195 = const()[name = tensor("op_195"), val = tensor([1, 1, 4, 1])]; - tensor var_196 = reshape(shape = var_195, x = var_194)[name = tensor("op_196")]; - tensor cross_1 = mul(x = var_193, y = var_196)[name = tensor("cross_1")]; + tensor v_1 = transpose(perm = v_1_perm_0, x = var_234)[name = tensor("transpose_56")]; + tensor inner_1 = matmul(transpose_x = inner_1_transpose_x_0, transpose_y = inner_1_transpose_y_0, x = var_247, y = v_1)[name = tensor("inner_1")]; + tensor var_249_transpose_x_0 = const()[name = tensor("op_249_transpose_x_0"), val = tensor(false)]; + tensor var_249_transpose_y_0 = const()[name = tensor("op_249_transpose_y_0"), val = tensor(false)]; + tensor var_249 = matmul(transpose_x = var_249_transpose_x_0, transpose_y = var_249_transpose_y_0, x = q_1, y = prev_kv_1)[name = tensor("op_249")]; + tensor var_250 = real_div(x = sqrt_s0_1, y = sqrt_s_t_1)[name = tensor("op_250")]; + tensor var_251 = const()[name = tensor("op_251"), val = tensor([1, 1, 4, 1])]; + tensor var_252 = reshape(shape = var_251, x = var_250)[name = tensor("op_252")]; + tensor cross_1 = mul(x = var_249, y = var_252)[name = tensor("cross_1")]; tensor out_1 = add(x = inner_1, y = cross_1)[name = tensor("out_1")]; - tensor var_199 = mul(x = prev_kv_1, y = sqrt_s0_1)[name = tensor("op_199")]; - tensor var_201_transpose_x_1 = const()[name = tensor("op_201_transpose_x_1"), val = tensor(true)]; - tensor var_201_transpose_y_1 = const()[name = tensor("op_201_transpose_y_1"), val = tensor(false)]; - tensor var_201 = matmul(transpose_x = var_201_transpose_x_1, transpose_y = var_201_transpose_y_1, x = k_1, y = v_1)[name = tensor("op_201")]; - tensor new_kv_unnorm_1 = add(x = var_199, y = var_201)[name = tensor("new_kv_unnorm_1")]; - tensor var_203 = const()[name = tensor("op_203"), val = tensor(0x1p+2)]; - tensor new_scale_1 = add(x = prev_scale_1, y = var_203)[name = tensor("new_scale_1")]; - tensor var_205 = sqrt(x = new_scale_1)[name = tensor("op_205")]; - tensor var_206 = real_div(x = new_kv_unnorm_1, y = var_205)[name = tensor("op_206")]; - tensor var_207_perm_0 = const()[name = tensor("op_207_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_255 = mul(x = prev_kv_1, y = sqrt_s0_1)[name = tensor("op_255")]; + tensor var_257_transpose_x_1 = const()[name = tensor("op_257_transpose_x_1"), val = tensor(true)]; + tensor var_257_transpose_y_1 = const()[name = tensor("op_257_transpose_y_1"), val = tensor(false)]; + tensor var_257 = matmul(transpose_x = var_257_transpose_x_1, transpose_y = var_257_transpose_y_1, x = k_1, y = v_1)[name = tensor("op_257")]; + tensor new_kv_unnorm_1 = add(x = var_255, y = var_257)[name = tensor("new_kv_unnorm_1")]; + tensor var_259 = const()[name = tensor("op_259"), val = tensor(0x1p+2)]; + tensor new_scale_1 = add(x = prev_scale_1, y = var_259)[name = tensor("new_scale_1")]; + tensor var_261 = sqrt(x = new_scale_1)[name = tensor("op_261")]; + tensor var_262 = real_div(x = new_kv_unnorm_1, y = var_261)[name = tensor("op_262")]; + tensor var_263_perm_0 = const()[name = tensor("op_263_perm_0"), val = tensor([0, 2, 1, 3])]; tensor out_3_axes_0 = const()[name = tensor("out_3_axes_0"), val = tensor([-1])]; - tensor var_207 = transpose(perm = var_207_perm_0, x = out_1)[name = tensor("transpose_55")]; - tensor out_3 = layer_norm(axes = out_3_axes_0, epsilon = var_17, x = var_207)[name = tensor("out_3")]; - tensor var_211 = const()[name = tensor("op_211"), val = tensor([1, 4, 256])]; - tensor out_5 = reshape(shape = var_211, x = out_3)[name = tensor("out_5")]; - tensor var_213 = silu(x = input_17)[name = tensor("op_213")]; - tensor input_19 = mul(x = var_213, y = out_5)[name = tensor("input_19")]; - tensor ret_out_1 = linear(bias = encoder_out_proj_0_bias, weight = encoder_out_proj_0_weight, x = input_19)[name = tensor("linear_7")]; - tensor x_3 = add(x = input_13, y = ret_out_1)[name = tensor("x_3")]; + tensor var_263 = transpose(perm = var_263_perm_0, x = out_1)[name = tensor("transpose_55")]; + tensor out_3 = layer_norm(axes = out_3_axes_0, epsilon = var_74, x = var_263)[name = tensor("out_3")]; + tensor var_267 = const()[name = tensor("op_267"), val = tensor([1, 4, 256])]; + tensor out_5 = reshape(shape = var_267, x = out_3)[name = tensor("out_5")]; + tensor var_269 = silu(x = input_19)[name = tensor("op_269")]; + tensor input_21 = mul(x = var_269, y = out_5)[name = tensor("input_21")]; + tensor ret_out_1 = linear(bias = inner_encoder_out_proj_0_bias, weight = inner_encoder_out_proj_0_weight, x = input_21)[name = tensor("linear_7")]; + tensor x_3 = add(x = input_15, y = ret_out_1)[name = tensor("x_3")]; tensor window_1_begin_0 = const()[name = tensor("window_1_begin_0"), val = tensor([0, 0, 0, 0])]; tensor window_1_end_0 = const()[name = tensor("window_1_end_0"), val = tensor([1, 1, 16, 256])]; tensor window_1_end_mask_0 = const()[name = tensor("window_1_end_mask_0"), val = tensor([false, true, true, true])]; tensor window_1_squeeze_mask_0 = const()[name = tensor("window_1_squeeze_mask_0"), val = tensor([true, false, false, false])]; tensor window_1 = slice_by_index(begin = window_1_begin_0, end = window_1_end_0, end_mask = window_1_end_mask_0, squeeze_mask = window_1_squeeze_mask_0, x = enc_conv_cache)[name = tensor("window_1")]; - tensor var_221_begin_0 = const()[name = tensor("op_221_begin_0"), val = tensor([0, 0, 0])]; - tensor var_221_end_0 = const()[name = tensor("op_221_end_0"), val = tensor([1, 1, 256])]; - tensor var_221_end_mask_0 = const()[name = tensor("op_221_end_mask_0"), val = tensor([true, false, true])]; - tensor var_221 = slice_by_index(begin = var_221_begin_0, end = var_221_end_0, end_mask = var_221_end_mask_0, x = x_3)[name = tensor("op_221")]; - tensor var_224_begin_0 = const()[name = tensor("op_224_begin_0"), val = tensor([0, 1, 0])]; - tensor var_224_end_0 = const()[name = tensor("op_224_end_0"), val = tensor([1, 16, 256])]; - tensor var_224_end_mask_0 = const()[name = tensor("op_224_end_mask_0"), val = tensor([true, true, true])]; - tensor var_224 = slice_by_index(begin = var_224_begin_0, end = var_224_end_0, end_mask = var_224_end_mask_0, x = window_1)[name = tensor("op_224")]; + tensor var_277_begin_0 = const()[name = tensor("op_277_begin_0"), val = tensor([0, 0, 0])]; + tensor var_277_end_0 = const()[name = tensor("op_277_end_0"), val = tensor([1, 1, 256])]; + tensor var_277_end_mask_0 = const()[name = tensor("op_277_end_mask_0"), val = tensor([true, false, true])]; + tensor var_277 = slice_by_index(begin = var_277_begin_0, end = var_277_end_0, end_mask = var_277_end_mask_0, x = x_3)[name = tensor("op_277")]; + tensor var_280_begin_0 = const()[name = tensor("op_280_begin_0"), val = tensor([0, 1, 0])]; + tensor var_280_end_0 = const()[name = tensor("op_280_end_0"), val = tensor([1, 16, 256])]; + tensor var_280_end_mask_0 = const()[name = tensor("op_280_end_mask_0"), val = tensor([true, true, true])]; + tensor var_280 = slice_by_index(begin = var_280_begin_0, end = var_280_end_0, end_mask = var_280_end_mask_0, x = window_1)[name = tensor("op_280")]; tensor window_3_interleave_0 = const()[name = tensor("window_3_interleave_0"), val = tensor(false)]; - tensor window_3 = concat(axis = var_26, interleave = window_3_interleave_0, values = (var_224, var_221))[name = tensor("window_3")]; - tensor var_229_begin_0 = const()[name = tensor("op_229_begin_0"), val = tensor([0, 1, 0])]; - tensor var_229_end_0 = const()[name = tensor("op_229_end_0"), val = tensor([1, 2, 256])]; - tensor var_229_end_mask_0 = const()[name = tensor("op_229_end_mask_0"), val = tensor([true, false, true])]; - tensor var_229 = slice_by_index(begin = var_229_begin_0, end = var_229_end_0, end_mask = var_229_end_mask_0, x = x_3)[name = tensor("op_229")]; - tensor var_232_begin_0 = const()[name = tensor("op_232_begin_0"), val = tensor([0, 1, 0])]; - tensor var_232_end_0 = const()[name = tensor("op_232_end_0"), val = tensor([1, 16, 256])]; - tensor var_232_end_mask_0 = const()[name = tensor("op_232_end_mask_0"), val = tensor([true, true, true])]; - tensor var_232 = slice_by_index(begin = var_232_begin_0, end = var_232_end_0, end_mask = var_232_end_mask_0, x = window_3)[name = tensor("op_232")]; + tensor window_3 = concat(axis = var_82, interleave = window_3_interleave_0, values = (var_280, var_277))[name = tensor("window_3")]; + tensor var_285_begin_0 = const()[name = tensor("op_285_begin_0"), val = tensor([0, 1, 0])]; + tensor var_285_end_0 = const()[name = tensor("op_285_end_0"), val = tensor([1, 2, 256])]; + tensor var_285_end_mask_0 = const()[name = tensor("op_285_end_mask_0"), val = tensor([true, false, true])]; + tensor var_285 = slice_by_index(begin = var_285_begin_0, end = var_285_end_0, end_mask = var_285_end_mask_0, x = x_3)[name = tensor("op_285")]; + tensor var_288_begin_0 = const()[name = tensor("op_288_begin_0"), val = tensor([0, 1, 0])]; + tensor var_288_end_0 = const()[name = tensor("op_288_end_0"), val = tensor([1, 16, 256])]; + tensor var_288_end_mask_0 = const()[name = tensor("op_288_end_mask_0"), val = tensor([true, true, true])]; + tensor var_288 = slice_by_index(begin = var_288_begin_0, end = var_288_end_0, end_mask = var_288_end_mask_0, x = window_3)[name = tensor("op_288")]; tensor window_5_interleave_0 = const()[name = tensor("window_5_interleave_0"), val = tensor(false)]; - tensor window_5 = concat(axis = var_26, interleave = window_5_interleave_0, values = (var_232, var_229))[name = tensor("window_5")]; - tensor var_237_begin_0 = const()[name = tensor("op_237_begin_0"), val = tensor([0, 2, 0])]; - tensor var_237_end_0 = const()[name = tensor("op_237_end_0"), val = tensor([1, 3, 256])]; - tensor var_237_end_mask_0 = const()[name = tensor("op_237_end_mask_0"), val = tensor([true, false, true])]; - tensor var_237 = slice_by_index(begin = var_237_begin_0, end = var_237_end_0, end_mask = var_237_end_mask_0, x = x_3)[name = tensor("op_237")]; - tensor var_240_begin_0 = const()[name = tensor("op_240_begin_0"), val = tensor([0, 1, 0])]; - tensor var_240_end_0 = const()[name = tensor("op_240_end_0"), val = tensor([1, 16, 256])]; - tensor var_240_end_mask_0 = const()[name = tensor("op_240_end_mask_0"), val = tensor([true, true, true])]; - tensor var_240 = slice_by_index(begin = var_240_begin_0, end = var_240_end_0, end_mask = var_240_end_mask_0, x = window_5)[name = tensor("op_240")]; + tensor window_5 = concat(axis = var_82, interleave = window_5_interleave_0, values = (var_288, var_285))[name = tensor("window_5")]; + tensor var_293_begin_0 = const()[name = tensor("op_293_begin_0"), val = tensor([0, 2, 0])]; + tensor var_293_end_0 = const()[name = tensor("op_293_end_0"), val = tensor([1, 3, 256])]; + tensor var_293_end_mask_0 = const()[name = tensor("op_293_end_mask_0"), val = tensor([true, false, true])]; + tensor var_293 = slice_by_index(begin = var_293_begin_0, end = var_293_end_0, end_mask = var_293_end_mask_0, x = x_3)[name = tensor("op_293")]; + tensor var_296_begin_0 = const()[name = tensor("op_296_begin_0"), val = tensor([0, 1, 0])]; + tensor var_296_end_0 = const()[name = tensor("op_296_end_0"), val = tensor([1, 16, 256])]; + tensor var_296_end_mask_0 = const()[name = tensor("op_296_end_mask_0"), val = tensor([true, true, true])]; + tensor var_296 = slice_by_index(begin = var_296_begin_0, end = var_296_end_0, end_mask = var_296_end_mask_0, x = window_5)[name = tensor("op_296")]; tensor window_7_interleave_0 = const()[name = tensor("window_7_interleave_0"), val = tensor(false)]; - tensor window_7 = concat(axis = var_26, interleave = window_7_interleave_0, values = (var_240, var_237))[name = tensor("window_7")]; - tensor var_245_begin_0 = const()[name = tensor("op_245_begin_0"), val = tensor([0, 3, 0])]; - tensor var_245_end_0 = const()[name = tensor("op_245_end_0"), val = tensor([1, 1, 256])]; - tensor var_245_end_mask_0 = const()[name = tensor("op_245_end_mask_0"), val = tensor([true, true, true])]; - tensor var_245 = slice_by_index(begin = var_245_begin_0, end = var_245_end_0, end_mask = var_245_end_mask_0, x = x_3)[name = tensor("op_245")]; - tensor var_248_begin_0 = const()[name = tensor("op_248_begin_0"), val = tensor([0, 1, 0])]; - tensor var_248_end_0 = const()[name = tensor("op_248_end_0"), val = tensor([1, 16, 256])]; - tensor var_248_end_mask_0 = const()[name = tensor("op_248_end_mask_0"), val = tensor([true, true, true])]; - tensor var_248 = slice_by_index(begin = var_248_begin_0, end = var_248_end_0, end_mask = var_248_end_mask_0, x = window_7)[name = tensor("op_248")]; + tensor window_7 = concat(axis = var_82, interleave = window_7_interleave_0, values = (var_296, var_293))[name = tensor("window_7")]; + tensor var_301_begin_0 = const()[name = tensor("op_301_begin_0"), val = tensor([0, 3, 0])]; + tensor var_301_end_0 = const()[name = tensor("op_301_end_0"), val = tensor([1, 1, 256])]; + tensor var_301_end_mask_0 = const()[name = tensor("op_301_end_mask_0"), val = tensor([true, true, true])]; + tensor var_301 = slice_by_index(begin = var_301_begin_0, end = var_301_end_0, end_mask = var_301_end_mask_0, x = x_3)[name = tensor("op_301")]; + tensor var_304_begin_0 = const()[name = tensor("op_304_begin_0"), val = tensor([0, 1, 0])]; + tensor var_304_end_0 = const()[name = tensor("op_304_end_0"), val = tensor([1, 16, 256])]; + tensor var_304_end_mask_0 = const()[name = tensor("op_304_end_mask_0"), val = tensor([true, true, true])]; + tensor var_304 = slice_by_index(begin = var_304_begin_0, end = var_304_end_0, end_mask = var_304_end_mask_0, x = window_7)[name = tensor("op_304")]; tensor window_9_interleave_0 = const()[name = tensor("window_9_interleave_0"), val = tensor(false)]; - tensor window_9 = concat(axis = var_26, interleave = window_9_interleave_0, values = (var_248, var_245))[name = tensor("window_9")]; - tensor input_21_interleave_0 = const()[name = tensor("input_21_interleave_0"), val = tensor(false)]; - tensor input_21 = concat(axis = var_23, interleave = input_21_interleave_0, values = (window_3, window_5, window_7, window_9))[name = tensor("input_21")]; + tensor window_9 = concat(axis = var_82, interleave = window_9_interleave_0, values = (var_304, var_301))[name = tensor("window_9")]; + tensor input_23_interleave_0 = const()[name = tensor("input_23_interleave_0"), val = tensor(false)]; + tensor input_23 = concat(axis = var_69, interleave = input_23_interleave_0, values = (window_3, window_5, window_7, window_9))[name = tensor("input_23")]; tensor x_5_axes_0 = const()[name = tensor("x_5_axes_0"), val = tensor([-1])]; - tensor x_5 = layer_norm(axes = x_5_axes_0, beta = encoder_conv_module_0_sequential_0_bias, epsilon = var_28, gamma = encoder_conv_module_0_sequential_0_weight, x = input_21)[name = tensor("x_5")]; - tensor input_23_perm_0 = const()[name = tensor("input_23_perm_0"), val = tensor([0, 2, 1])]; + tensor x_5 = layer_norm(axes = x_5_axes_0, beta = inner_encoder_conv_module_0_sequential_0_bias, epsilon = var_66, gamma = inner_encoder_conv_module_0_sequential_0_weight, x = input_23)[name = tensor("x_5")]; + tensor input_25_perm_0 = const()[name = tensor("input_25_perm_0"), val = tensor([0, 2, 1])]; tensor inputs_3_pad_type_0 = const()[name = tensor("inputs_3_pad_type_0"), val = tensor("valid")]; tensor inputs_3_strides_0 = const()[name = tensor("inputs_3_strides_0"), val = tensor([1])]; tensor inputs_3_pad_0 = const()[name = tensor("inputs_3_pad_0"), val = tensor([0, 0])]; tensor inputs_3_dilations_0 = const()[name = tensor("inputs_3_dilations_0"), val = tensor([1])]; tensor inputs_3_groups_0 = const()[name = tensor("inputs_3_groups_0"), val = tensor(1)]; - tensor input_23 = transpose(perm = input_23_perm_0, x = x_5)[name = tensor("transpose_54")]; - tensor inputs_3 = conv(bias = encoder_conv_module_0_sequential_2_conv_bias, dilations = inputs_3_dilations_0, groups = inputs_3_groups_0, pad = inputs_3_pad_0, pad_type = inputs_3_pad_type_0, strides = inputs_3_strides_0, weight = encoder_conv_module_0_sequential_2_conv_weight, x = input_23)[name = tensor("inputs_3")]; - tensor var_273_split_sizes_0 = const()[name = tensor("op_273_split_sizes_0"), val = tensor([256, 256])]; - tensor var_273_axis_0 = const()[name = tensor("op_273_axis_0"), val = tensor(1)]; - tensor var_273_0, tensor var_273_1 = split(axis = var_273_axis_0, split_sizes = var_273_split_sizes_0, x = inputs_3)[name = tensor("op_273")]; - tensor var_275 = sigmoid(x = var_273_1)[name = tensor("op_275")]; - tensor inputs_5 = mul(x = var_273_0, y = var_275)[name = tensor("inputs_5")]; + tensor input_25 = transpose(perm = input_25_perm_0, x = x_5)[name = tensor("transpose_54")]; + tensor inputs_3 = conv(bias = inner_encoder_conv_module_0_sequential_2_conv_bias, dilations = inputs_3_dilations_0, groups = inputs_3_groups_0, pad = inputs_3_pad_0, pad_type = inputs_3_pad_type_0, strides = inputs_3_strides_0, weight = inner_encoder_conv_module_0_sequential_2_conv_weight, x = input_25)[name = tensor("inputs_3")]; + tensor var_329_split_sizes_0 = const()[name = tensor("op_329_split_sizes_0"), val = tensor([256, 256])]; + tensor var_329_axis_0 = const()[name = tensor("op_329_axis_0"), val = tensor(1)]; + tensor var_329_0, tensor var_329_1 = split(axis = var_329_axis_0, split_sizes = var_329_split_sizes_0, x = inputs_3)[name = tensor("op_329")]; + tensor var_331 = sigmoid(x = var_329_1)[name = tensor("op_331")]; + tensor inputs_5 = mul(x = var_329_0, y = var_331)[name = tensor("inputs_5")]; tensor outputs_aug_1_pad_type_0 = const()[name = tensor("outputs_aug_1_pad_type_0"), val = tensor("custom")]; tensor outputs_aug_1_pad_0 = const()[name = tensor("outputs_aug_1_pad_0"), val = tensor([15, 15])]; tensor outputs_aug_1_groups_0 = const()[name = tensor("outputs_aug_1_groups_0"), val = tensor(256)]; tensor outputs_aug_1_strides_0 = const()[name = tensor("outputs_aug_1_strides_0"), val = tensor([1])]; tensor outputs_aug_1_dilations_0 = const()[name = tensor("outputs_aug_1_dilations_0"), val = tensor([1])]; - tensor outputs_aug_1 = conv(dilations = outputs_aug_1_dilations_0, groups = outputs_aug_1_groups_0, pad = outputs_aug_1_pad_0, pad_type = outputs_aug_1_pad_type_0, strides = outputs_aug_1_strides_0, weight = encoder_conv_module_0_sequential_4_conv_weight, x = inputs_5)[name = tensor("outputs_aug_1")]; - tensor input_25_begin_0 = const()[name = tensor("input_25_begin_0"), val = tensor([0, 0, 0])]; - tensor input_25_end_0 = const()[name = tensor("input_25_end_0"), val = tensor([4, 256, 16])]; - tensor input_25_end_mask_0 = const()[name = tensor("input_25_end_mask_0"), val = tensor([true, true, false])]; - tensor input_25 = slice_by_index(begin = input_25_begin_0, end = input_25_end_0, end_mask = input_25_end_mask_0, x = outputs_aug_1)[name = tensor("input_25")]; - tensor inputs_7 = batch_norm(beta = encoder_conv_module_0_sequential_5_bias, epsilon = var_28, gamma = encoder_conv_module_0_sequential_5_weight, mean = encoder_conv_module_0_sequential_5_running_mean, variance = encoder_conv_module_0_sequential_5_running_var, x = input_25)[name = tensor("inputs_7")]; - tensor input_27 = silu(x = inputs_7)[name = tensor("input_27")]; - tensor input_29_pad_type_0 = const()[name = tensor("input_29_pad_type_0"), val = tensor("valid")]; - tensor input_29_strides_0 = const()[name = tensor("input_29_strides_0"), val = tensor([1])]; - tensor input_29_pad_0 = const()[name = tensor("input_29_pad_0"), val = tensor([0, 0])]; - tensor input_29_dilations_0 = const()[name = tensor("input_29_dilations_0"), val = tensor([1])]; - tensor input_29_groups_0 = const()[name = tensor("input_29_groups_0"), val = tensor(1)]; - tensor input_29 = conv(bias = encoder_conv_module_0_sequential_7_conv_bias, dilations = input_29_dilations_0, groups = input_29_groups_0, pad = input_29_pad_0, pad_type = input_29_pad_type_0, strides = input_29_strides_0, weight = encoder_conv_module_0_sequential_7_conv_weight, x = input_27)[name = tensor("input_29")]; + tensor outputs_aug_1 = conv(dilations = outputs_aug_1_dilations_0, groups = outputs_aug_1_groups_0, pad = outputs_aug_1_pad_0, pad_type = outputs_aug_1_pad_type_0, strides = outputs_aug_1_strides_0, weight = inner_encoder_conv_module_0_sequential_4_conv_weight, x = inputs_5)[name = tensor("outputs_aug_1")]; + tensor input_27_begin_0 = const()[name = tensor("input_27_begin_0"), val = tensor([0, 0, 0])]; + tensor input_27_end_0 = const()[name = tensor("input_27_end_0"), val = tensor([4, 256, 16])]; + tensor input_27_end_mask_0 = const()[name = tensor("input_27_end_mask_0"), val = tensor([true, true, false])]; + tensor input_27 = slice_by_index(begin = input_27_begin_0, end = input_27_end_0, end_mask = input_27_end_mask_0, x = outputs_aug_1)[name = tensor("input_27")]; + tensor inputs_7 = batch_norm(beta = inner_encoder_conv_module_0_sequential_5_bias, epsilon = var_66, gamma = inner_encoder_conv_module_0_sequential_5_weight, mean = inner_encoder_conv_module_0_sequential_5_running_mean, variance = inner_encoder_conv_module_0_sequential_5_running_var, x = input_27)[name = tensor("inputs_7")]; + tensor input_29 = silu(x = inputs_7)[name = tensor("input_29")]; + tensor input_31_pad_type_0 = const()[name = tensor("input_31_pad_type_0"), val = tensor("valid")]; + tensor input_31_strides_0 = const()[name = tensor("input_31_strides_0"), val = tensor([1])]; + tensor input_31_pad_0 = const()[name = tensor("input_31_pad_0"), val = tensor([0, 0])]; + tensor input_31_dilations_0 = const()[name = tensor("input_31_dilations_0"), val = tensor([1])]; + tensor input_31_groups_0 = const()[name = tensor("input_31_groups_0"), val = tensor(1)]; + tensor input_31 = conv(bias = inner_encoder_conv_module_0_sequential_7_conv_bias, dilations = input_31_dilations_0, groups = input_31_groups_0, pad = input_31_pad_0, pad_type = input_31_pad_type_0, strides = input_31_strides_0, weight = inner_encoder_conv_module_0_sequential_7_conv_weight, x = input_29)[name = tensor("input_31")]; tensor conv_out_1_perm_0 = const()[name = tensor("conv_out_1_perm_0"), val = tensor([0, 2, 1])]; - tensor var_306_begin_0 = const()[name = tensor("op_306_begin_0"), val = tensor([0, -1, 0])]; - tensor var_306_end_0 = const()[name = tensor("op_306_end_0"), val = tensor([4, 16, 256])]; - tensor var_306_end_mask_0 = const()[name = tensor("op_306_end_mask_0"), val = tensor([true, true, true])]; - tensor conv_out_1 = transpose(perm = conv_out_1_perm_0, x = input_29)[name = tensor("transpose_53")]; - tensor var_306 = slice_by_index(begin = var_306_begin_0, end = var_306_end_0, end_mask = var_306_end_mask_0, x = conv_out_1)[name = tensor("op_306")]; - tensor var_308_perm_0 = const()[name = tensor("op_308_perm_0"), val = tensor([1, 0, 2])]; - tensor var_308 = transpose(perm = var_308_perm_0, x = var_306)[name = tensor("transpose_52")]; - tensor input_31 = add(x = x_3, y = var_308)[name = tensor("input_31")]; - tensor input_33_axes_0 = const()[name = tensor("input_33_axes_0"), val = tensor([-1])]; - tensor input_33 = layer_norm(axes = input_33_axes_0, beta = encoder_ffn2_0_module_sequential_0_bias, epsilon = var_28, gamma = encoder_ffn2_0_module_sequential_0_weight, x = input_31)[name = tensor("input_33")]; - tensor inputs_9 = linear(bias = encoder_ffn2_0_module_sequential_1_linear_bias, weight = encoder_ffn2_0_module_sequential_1_linear_weight, x = input_33)[name = tensor("linear_8")]; - tensor input_35 = silu(x = inputs_9)[name = tensor("input_35")]; - tensor input_39 = linear(bias = encoder_ffn2_0_module_sequential_4_linear_bias, weight = encoder_ffn2_0_module_sequential_4_linear_weight, x = input_35)[name = tensor("linear_9")]; - tensor var_331 = const()[name = tensor("op_331"), val = tensor(0x1p-1)]; - tensor var_332 = mul(x = input_39, y = var_331)[name = tensor("op_332")]; - tensor input_41 = add(x = var_332, y = input_31)[name = tensor("input_41")]; - tensor input_43_axes_0 = const()[name = tensor("input_43_axes_0"), val = tensor([-1])]; - tensor input_43 = layer_norm(axes = input_43_axes_0, beta = encoder_layer_norm_0_bias, epsilon = var_28, gamma = encoder_layer_norm_0_weight, x = input_41)[name = tensor("input_43")]; + tensor var_362_begin_0 = const()[name = tensor("op_362_begin_0"), val = tensor([0, -1, 0])]; + tensor var_362_end_0 = const()[name = tensor("op_362_end_0"), val = tensor([4, 16, 256])]; + tensor var_362_end_mask_0 = const()[name = tensor("op_362_end_mask_0"), val = tensor([true, true, true])]; + tensor conv_out_1 = transpose(perm = conv_out_1_perm_0, x = input_31)[name = tensor("transpose_53")]; + tensor var_362 = slice_by_index(begin = var_362_begin_0, end = var_362_end_0, end_mask = var_362_end_mask_0, x = conv_out_1)[name = tensor("op_362")]; + tensor var_364_perm_0 = const()[name = tensor("op_364_perm_0"), val = tensor([1, 0, 2])]; + tensor var_364 = transpose(perm = var_364_perm_0, x = var_362)[name = tensor("transpose_52")]; + tensor input_33 = add(x = x_3, y = var_364)[name = tensor("input_33")]; + tensor input_35_axes_0 = const()[name = tensor("input_35_axes_0"), val = tensor([-1])]; + tensor input_35 = layer_norm(axes = input_35_axes_0, beta = inner_encoder_ffn2_0_module_sequential_0_bias, epsilon = var_66, gamma = inner_encoder_ffn2_0_module_sequential_0_weight, x = input_33)[name = tensor("input_35")]; + tensor inputs_9 = linear(bias = inner_encoder_ffn2_0_module_sequential_1_linear_bias, weight = inner_encoder_ffn2_0_module_sequential_1_linear_weight, x = input_35)[name = tensor("linear_8")]; + tensor input_37 = silu(x = inputs_9)[name = tensor("input_37")]; + tensor input_41 = linear(bias = inner_encoder_ffn2_0_module_sequential_4_linear_bias, weight = inner_encoder_ffn2_0_module_sequential_4_linear_weight, x = input_37)[name = tensor("linear_9")]; + tensor var_387 = const()[name = tensor("op_387"), val = tensor(0x1p-1)]; + tensor var_388 = mul(x = input_41, y = var_387)[name = tensor("op_388")]; + tensor input_43 = add(x = var_388, y = input_33)[name = tensor("input_43")]; tensor input_45_axes_0 = const()[name = tensor("input_45_axes_0"), val = tensor([-1])]; - tensor input_45 = layer_norm(axes = input_45_axes_0, beta = encoder_ffn1_1_module_sequential_0_bias, epsilon = var_28, gamma = encoder_ffn1_1_module_sequential_0_weight, x = input_43)[name = tensor("input_45")]; - tensor inputs_11 = linear(bias = encoder_ffn1_1_module_sequential_1_linear_bias, weight = encoder_ffn1_1_module_sequential_1_linear_weight, x = input_45)[name = tensor("linear_10")]; - tensor input_47 = silu(x = inputs_11)[name = tensor("input_47")]; - tensor input_51 = linear(bias = encoder_ffn1_1_module_sequential_4_linear_bias, weight = encoder_ffn1_1_module_sequential_4_linear_weight, x = input_47)[name = tensor("linear_11")]; - tensor var_361 = const()[name = tensor("op_361"), val = tensor(0x1p-1)]; - tensor var_362 = mul(x = input_51, y = var_361)[name = tensor("op_362")]; - tensor input_53 = add(x = var_362, y = input_43)[name = tensor("input_53")]; + tensor input_45 = layer_norm(axes = input_45_axes_0, beta = inner_encoder_layer_norm_0_bias, epsilon = var_66, gamma = inner_encoder_layer_norm_0_weight, x = input_43)[name = tensor("input_45")]; + tensor input_47_axes_0 = const()[name = tensor("input_47_axes_0"), val = tensor([-1])]; + tensor input_47 = layer_norm(axes = input_47_axes_0, beta = inner_encoder_ffn1_1_module_sequential_0_bias, epsilon = var_66, gamma = inner_encoder_ffn1_1_module_sequential_0_weight, x = input_45)[name = tensor("input_47")]; + tensor inputs_11 = linear(bias = inner_encoder_ffn1_1_module_sequential_1_linear_bias, weight = inner_encoder_ffn1_1_module_sequential_1_linear_weight, x = input_47)[name = tensor("linear_10")]; + tensor input_49 = silu(x = inputs_11)[name = tensor("input_49")]; + tensor input_53 = linear(bias = inner_encoder_ffn1_1_module_sequential_4_linear_bias, weight = inner_encoder_ffn1_1_module_sequential_4_linear_weight, x = input_49)[name = tensor("linear_11")]; + tensor var_417 = const()[name = tensor("op_417"), val = tensor(0x1p-1)]; + tensor var_418 = mul(x = input_53, y = var_417)[name = tensor("op_418")]; + tensor input_55 = add(x = var_418, y = input_45)[name = tensor("input_55")]; tensor x_7_axes_0 = const()[name = tensor("x_7_axes_0"), val = tensor([-1])]; - tensor x_7 = layer_norm(axes = x_7_axes_0, beta = encoder_ret_lns_1_bias, epsilon = var_28, gamma = encoder_ret_lns_1_weight, x = input_53)[name = tensor("x_7")]; + tensor x_7 = layer_norm(axes = x_7_axes_0, beta = inner_encoder_ret_lns_1_bias, epsilon = var_66, gamma = inner_encoder_ret_lns_1_weight, x = input_55)[name = tensor("x_7")]; tensor prev_kv_3_begin_0 = const()[name = tensor("prev_kv_3_begin_0"), val = tensor([1, 0, 0, 0, 0])]; tensor prev_kv_3_end_0 = const()[name = tensor("prev_kv_3_end_0"), val = tensor([2, 1, 4, 64, 64])]; tensor prev_kv_3_end_mask_0 = const()[name = tensor("prev_kv_3_end_mask_0"), val = tensor([false, true, true, true, true])]; @@ -416,173 +438,173 @@ program(1.0) tensor prev_scale_3_end_mask_0 = const()[name = tensor("prev_scale_3_end_mask_0"), val = tensor([false, true])]; tensor prev_scale_3_squeeze_mask_0 = const()[name = tensor("prev_scale_3_squeeze_mask_0"), val = tensor([true, false])]; tensor prev_scale_3 = slice_by_index(begin = prev_scale_3_begin_0, end = prev_scale_3_end_0, end_mask = prev_scale_3_end_mask_0, squeeze_mask = prev_scale_3_squeeze_mask_0, x = enc_scale)[name = tensor("prev_scale_3")]; - tensor var_376 = linear(bias = encoder_q_proj_1_bias, weight = encoder_q_proj_1_weight, x = x_7)[name = tensor("linear_12")]; - tensor var_377 = const()[name = tensor("op_377"), val = tensor([1, 4, 4, 64])]; - tensor var_378 = reshape(shape = var_377, x = var_376)[name = tensor("op_378")]; + tensor var_432 = linear(bias = inner_encoder_q_proj_1_bias, weight = inner_encoder_q_proj_1_weight, x = x_7)[name = tensor("linear_12")]; + tensor var_433 = const()[name = tensor("op_433"), val = tensor([1, 4, 4, 64])]; + tensor var_434 = reshape(shape = var_433, x = var_432)[name = tensor("op_434")]; tensor q_3_perm_0 = const()[name = tensor("q_3_perm_0"), val = tensor([0, 2, 1, 3])]; - tensor var_382 = linear(bias = encoder_k_proj_1_bias, weight = encoder_k_proj_1_weight, x = x_7)[name = tensor("linear_13")]; - tensor var_383 = const()[name = tensor("op_383"), val = tensor(0x1p-3)]; - tensor var_384 = mul(x = var_382, y = var_383)[name = tensor("op_384")]; - tensor var_385 = const()[name = tensor("op_385"), val = tensor([1, 4, 4, 64])]; - tensor var_386 = reshape(shape = var_385, x = var_384)[name = tensor("op_386")]; + tensor var_438 = linear(bias = inner_encoder_k_proj_1_bias, weight = inner_encoder_k_proj_1_weight, x = x_7)[name = tensor("linear_13")]; + tensor var_439 = const()[name = tensor("op_439"), val = tensor(0x1p-3)]; + tensor var_440 = mul(x = var_438, y = var_439)[name = tensor("op_440")]; + tensor var_441 = const()[name = tensor("op_441"), val = tensor([1, 4, 4, 64])]; + tensor var_442 = reshape(shape = var_441, x = var_440)[name = tensor("op_442")]; tensor k_3_perm_0 = const()[name = tensor("k_3_perm_0"), val = tensor([0, 2, 1, 3])]; - tensor var_390 = linear(bias = encoder_v_proj_1_bias, weight = encoder_v_proj_1_weight, x = x_7)[name = tensor("linear_14")]; - tensor var_391 = const()[name = tensor("op_391"), val = tensor([1, 4, 4, 64])]; - tensor var_392 = reshape(shape = var_391, x = var_390)[name = tensor("op_392")]; + tensor var_446 = linear(bias = inner_encoder_v_proj_1_bias, weight = inner_encoder_v_proj_1_weight, x = x_7)[name = tensor("linear_14")]; + tensor var_447 = const()[name = tensor("op_447"), val = tensor([1, 4, 4, 64])]; + tensor var_448 = reshape(shape = var_447, x = var_446)[name = tensor("op_448")]; tensor v_3_perm_0 = const()[name = tensor("v_3_perm_0"), val = tensor([0, 2, 1, 3])]; - tensor input_57 = linear(bias = encoder_g_proj_1_bias, weight = encoder_g_proj_1_weight, x = x_7)[name = tensor("linear_15")]; + tensor input_59 = linear(bias = inner_encoder_g_proj_1_bias, weight = inner_encoder_g_proj_1_weight, x = x_7)[name = tensor("linear_15")]; tensor sqrt_s0_3 = sqrt(x = prev_scale_3)[name = tensor("sqrt_s0_3")]; - tensor s_t_3 = add(x = prev_scale_3, y = encoder__t_index)[name = tensor("s_t_3")]; + tensor s_t_3 = add(x = prev_scale_3, y = inner_encoder__t_index)[name = tensor("s_t_3")]; tensor sqrt_s_t_3 = sqrt(x = s_t_3)[name = tensor("sqrt_s_t_3")]; tensor qk_3_transpose_x_1 = const()[name = tensor("qk_3_transpose_x_1"), val = tensor(false)]; tensor qk_3_transpose_y_1 = const()[name = tensor("qk_3_transpose_y_1"), val = tensor(true)]; - tensor k_3 = transpose(perm = k_3_perm_0, x = var_386)[name = tensor("transpose_50")]; - tensor q_3 = transpose(perm = q_3_perm_0, x = var_378)[name = tensor("transpose_51")]; + tensor k_3 = transpose(perm = k_3_perm_0, x = var_442)[name = tensor("transpose_50")]; + tensor q_3 = transpose(perm = q_3_perm_0, x = var_434)[name = tensor("transpose_51")]; tensor qk_3 = matmul(transpose_x = qk_3_transpose_x_1, transpose_y = qk_3_transpose_y_1, x = q_3, y = k_3)[name = tensor("qk_3")]; - tensor var_402 = const()[name = tensor("op_402"), val = tensor([4, 1])]; - tensor var_403 = reshape(shape = var_402, x = sqrt_s_t_3)[name = tensor("op_403")]; - tensor M_3 = real_div(x = encoder__causal_mask, y = var_403)[name = tensor("M_3")]; - tensor var_405 = mul(x = qk_3, y = M_3)[name = tensor("op_405")]; + tensor var_458 = const()[name = tensor("op_458"), val = tensor([4, 1])]; + tensor var_459 = reshape(shape = var_458, x = sqrt_s_t_3)[name = tensor("op_459")]; + tensor M_3 = real_div(x = inner_encoder__causal_mask, y = var_459)[name = tensor("M_3")]; + tensor var_461 = mul(x = qk_3, y = M_3)[name = tensor("op_461")]; tensor inner_3_transpose_x_0 = const()[name = tensor("inner_3_transpose_x_0"), val = tensor(false)]; tensor inner_3_transpose_y_0 = const()[name = tensor("inner_3_transpose_y_0"), val = tensor(false)]; - tensor v_3 = transpose(perm = v_3_perm_0, x = var_392)[name = tensor("transpose_49")]; - tensor inner_3 = matmul(transpose_x = inner_3_transpose_x_0, transpose_y = inner_3_transpose_y_0, x = var_405, y = v_3)[name = tensor("inner_3")]; - tensor var_407_transpose_x_0 = const()[name = tensor("op_407_transpose_x_0"), val = tensor(false)]; - tensor var_407_transpose_y_0 = const()[name = tensor("op_407_transpose_y_0"), val = tensor(false)]; - tensor var_407 = matmul(transpose_x = var_407_transpose_x_0, transpose_y = var_407_transpose_y_0, x = q_3, y = prev_kv_3)[name = tensor("op_407")]; - tensor var_408 = real_div(x = sqrt_s0_3, y = sqrt_s_t_3)[name = tensor("op_408")]; - tensor var_409 = const()[name = tensor("op_409"), val = tensor([1, 1, 4, 1])]; - tensor var_410 = reshape(shape = var_409, x = var_408)[name = tensor("op_410")]; - tensor cross_3 = mul(x = var_407, y = var_410)[name = tensor("cross_3")]; + tensor v_3 = transpose(perm = v_3_perm_0, x = var_448)[name = tensor("transpose_49")]; + tensor inner_3 = matmul(transpose_x = inner_3_transpose_x_0, transpose_y = inner_3_transpose_y_0, x = var_461, y = v_3)[name = tensor("inner_3")]; + tensor var_463_transpose_x_0 = const()[name = tensor("op_463_transpose_x_0"), val = tensor(false)]; + tensor var_463_transpose_y_0 = const()[name = tensor("op_463_transpose_y_0"), val = tensor(false)]; + tensor var_463 = matmul(transpose_x = var_463_transpose_x_0, transpose_y = var_463_transpose_y_0, x = q_3, y = prev_kv_3)[name = tensor("op_463")]; + tensor var_464 = real_div(x = sqrt_s0_3, y = sqrt_s_t_3)[name = tensor("op_464")]; + tensor var_465 = const()[name = tensor("op_465"), val = tensor([1, 1, 4, 1])]; + tensor var_466 = reshape(shape = var_465, x = var_464)[name = tensor("op_466")]; + tensor cross_3 = mul(x = var_463, y = var_466)[name = tensor("cross_3")]; tensor out_7 = add(x = inner_3, y = cross_3)[name = tensor("out_7")]; - tensor var_413 = mul(x = prev_kv_3, y = sqrt_s0_3)[name = tensor("op_413")]; - tensor var_415_transpose_x_1 = const()[name = tensor("op_415_transpose_x_1"), val = tensor(true)]; - tensor var_415_transpose_y_1 = const()[name = tensor("op_415_transpose_y_1"), val = tensor(false)]; - tensor var_415 = matmul(transpose_x = var_415_transpose_x_1, transpose_y = var_415_transpose_y_1, x = k_3, y = v_3)[name = tensor("op_415")]; - tensor new_kv_unnorm_3 = add(x = var_413, y = var_415)[name = tensor("new_kv_unnorm_3")]; - tensor var_417 = const()[name = tensor("op_417"), val = tensor(0x1p+2)]; - tensor new_scale_3 = add(x = prev_scale_3, y = var_417)[name = tensor("new_scale_3")]; - tensor var_419 = sqrt(x = new_scale_3)[name = tensor("op_419")]; - tensor var_420 = real_div(x = new_kv_unnorm_3, y = var_419)[name = tensor("op_420")]; - tensor var_421_perm_0 = const()[name = tensor("op_421_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_469 = mul(x = prev_kv_3, y = sqrt_s0_3)[name = tensor("op_469")]; + tensor var_471_transpose_x_1 = const()[name = tensor("op_471_transpose_x_1"), val = tensor(true)]; + tensor var_471_transpose_y_1 = const()[name = tensor("op_471_transpose_y_1"), val = tensor(false)]; + tensor var_471 = matmul(transpose_x = var_471_transpose_x_1, transpose_y = var_471_transpose_y_1, x = k_3, y = v_3)[name = tensor("op_471")]; + tensor new_kv_unnorm_3 = add(x = var_469, y = var_471)[name = tensor("new_kv_unnorm_3")]; + tensor var_473 = const()[name = tensor("op_473"), val = tensor(0x1p+2)]; + tensor new_scale_3 = add(x = prev_scale_3, y = var_473)[name = tensor("new_scale_3")]; + tensor var_475 = sqrt(x = new_scale_3)[name = tensor("op_475")]; + tensor var_476 = real_div(x = new_kv_unnorm_3, y = var_475)[name = tensor("op_476")]; + tensor var_477_perm_0 = const()[name = tensor("op_477_perm_0"), val = tensor([0, 2, 1, 3])]; tensor out_9_axes_0 = const()[name = tensor("out_9_axes_0"), val = tensor([-1])]; - tensor var_421 = transpose(perm = var_421_perm_0, x = out_7)[name = tensor("transpose_48")]; - tensor out_9 = layer_norm(axes = out_9_axes_0, epsilon = var_17, x = var_421)[name = tensor("out_9")]; - tensor var_425 = const()[name = tensor("op_425"), val = tensor([1, 4, 256])]; - tensor out_11 = reshape(shape = var_425, x = out_9)[name = tensor("out_11")]; - tensor var_427 = silu(x = input_57)[name = tensor("op_427")]; - tensor input_59 = mul(x = var_427, y = out_11)[name = tensor("input_59")]; - tensor ret_out_3 = linear(bias = encoder_out_proj_1_bias, weight = encoder_out_proj_1_weight, x = input_59)[name = tensor("linear_16")]; - tensor x_9 = add(x = input_53, y = ret_out_3)[name = tensor("x_9")]; + tensor var_477 = transpose(perm = var_477_perm_0, x = out_7)[name = tensor("transpose_48")]; + tensor out_9 = layer_norm(axes = out_9_axes_0, epsilon = var_74, x = var_477)[name = tensor("out_9")]; + tensor var_481 = const()[name = tensor("op_481"), val = tensor([1, 4, 256])]; + tensor out_11 = reshape(shape = var_481, x = out_9)[name = tensor("out_11")]; + tensor var_483 = silu(x = input_59)[name = tensor("op_483")]; + tensor input_61 = mul(x = var_483, y = out_11)[name = tensor("input_61")]; + tensor ret_out_3 = linear(bias = inner_encoder_out_proj_1_bias, weight = inner_encoder_out_proj_1_weight, x = input_61)[name = tensor("linear_16")]; + tensor x_9 = add(x = input_55, y = ret_out_3)[name = tensor("x_9")]; tensor window_11_begin_0 = const()[name = tensor("window_11_begin_0"), val = tensor([1, 0, 0, 0])]; tensor window_11_end_0 = const()[name = tensor("window_11_end_0"), val = tensor([2, 1, 16, 256])]; tensor window_11_end_mask_0 = const()[name = tensor("window_11_end_mask_0"), val = tensor([false, true, true, true])]; tensor window_11_squeeze_mask_0 = const()[name = tensor("window_11_squeeze_mask_0"), val = tensor([true, false, false, false])]; tensor window_11 = slice_by_index(begin = window_11_begin_0, end = window_11_end_0, end_mask = window_11_end_mask_0, squeeze_mask = window_11_squeeze_mask_0, x = enc_conv_cache)[name = tensor("window_11")]; - tensor var_435_begin_0 = const()[name = tensor("op_435_begin_0"), val = tensor([0, 0, 0])]; - tensor var_435_end_0 = const()[name = tensor("op_435_end_0"), val = tensor([1, 1, 256])]; - tensor var_435_end_mask_0 = const()[name = tensor("op_435_end_mask_0"), val = tensor([true, false, true])]; - tensor var_435 = slice_by_index(begin = var_435_begin_0, end = var_435_end_0, end_mask = var_435_end_mask_0, x = x_9)[name = tensor("op_435")]; - tensor var_438_begin_0 = const()[name = tensor("op_438_begin_0"), val = tensor([0, 1, 0])]; - tensor var_438_end_0 = const()[name = tensor("op_438_end_0"), val = tensor([1, 16, 256])]; - tensor var_438_end_mask_0 = const()[name = tensor("op_438_end_mask_0"), val = tensor([true, true, true])]; - tensor var_438 = slice_by_index(begin = var_438_begin_0, end = var_438_end_0, end_mask = var_438_end_mask_0, x = window_11)[name = tensor("op_438")]; + tensor var_491_begin_0 = const()[name = tensor("op_491_begin_0"), val = tensor([0, 0, 0])]; + tensor var_491_end_0 = const()[name = tensor("op_491_end_0"), val = tensor([1, 1, 256])]; + tensor var_491_end_mask_0 = const()[name = tensor("op_491_end_mask_0"), val = tensor([true, false, true])]; + tensor var_491 = slice_by_index(begin = var_491_begin_0, end = var_491_end_0, end_mask = var_491_end_mask_0, x = x_9)[name = tensor("op_491")]; + tensor var_494_begin_0 = const()[name = tensor("op_494_begin_0"), val = tensor([0, 1, 0])]; + tensor var_494_end_0 = const()[name = tensor("op_494_end_0"), val = tensor([1, 16, 256])]; + tensor var_494_end_mask_0 = const()[name = tensor("op_494_end_mask_0"), val = tensor([true, true, true])]; + tensor var_494 = slice_by_index(begin = var_494_begin_0, end = var_494_end_0, end_mask = var_494_end_mask_0, x = window_11)[name = tensor("op_494")]; tensor window_13_interleave_0 = const()[name = tensor("window_13_interleave_0"), val = tensor(false)]; - tensor window_13 = concat(axis = var_26, interleave = window_13_interleave_0, values = (var_438, var_435))[name = tensor("window_13")]; - tensor var_443_begin_0 = const()[name = tensor("op_443_begin_0"), val = tensor([0, 1, 0])]; - tensor var_443_end_0 = const()[name = tensor("op_443_end_0"), val = tensor([1, 2, 256])]; - tensor var_443_end_mask_0 = const()[name = tensor("op_443_end_mask_0"), val = tensor([true, false, true])]; - tensor var_443 = slice_by_index(begin = var_443_begin_0, end = var_443_end_0, end_mask = var_443_end_mask_0, x = x_9)[name = tensor("op_443")]; - tensor var_446_begin_0 = const()[name = tensor("op_446_begin_0"), val = tensor([0, 1, 0])]; - tensor var_446_end_0 = const()[name = tensor("op_446_end_0"), val = tensor([1, 16, 256])]; - tensor var_446_end_mask_0 = const()[name = tensor("op_446_end_mask_0"), val = tensor([true, true, true])]; - tensor var_446 = slice_by_index(begin = var_446_begin_0, end = var_446_end_0, end_mask = var_446_end_mask_0, x = window_13)[name = tensor("op_446")]; + tensor window_13 = concat(axis = var_82, interleave = window_13_interleave_0, values = (var_494, var_491))[name = tensor("window_13")]; + tensor var_499_begin_0 = const()[name = tensor("op_499_begin_0"), val = tensor([0, 1, 0])]; + tensor var_499_end_0 = const()[name = tensor("op_499_end_0"), val = tensor([1, 2, 256])]; + tensor var_499_end_mask_0 = const()[name = tensor("op_499_end_mask_0"), val = tensor([true, false, true])]; + tensor var_499 = slice_by_index(begin = var_499_begin_0, end = var_499_end_0, end_mask = var_499_end_mask_0, x = x_9)[name = tensor("op_499")]; + tensor var_502_begin_0 = const()[name = tensor("op_502_begin_0"), val = tensor([0, 1, 0])]; + tensor var_502_end_0 = const()[name = tensor("op_502_end_0"), val = tensor([1, 16, 256])]; + tensor var_502_end_mask_0 = const()[name = tensor("op_502_end_mask_0"), val = tensor([true, true, true])]; + tensor var_502 = slice_by_index(begin = var_502_begin_0, end = var_502_end_0, end_mask = var_502_end_mask_0, x = window_13)[name = tensor("op_502")]; tensor window_15_interleave_0 = const()[name = tensor("window_15_interleave_0"), val = tensor(false)]; - tensor window_15 = concat(axis = var_26, interleave = window_15_interleave_0, values = (var_446, var_443))[name = tensor("window_15")]; - tensor var_451_begin_0 = const()[name = tensor("op_451_begin_0"), val = tensor([0, 2, 0])]; - tensor var_451_end_0 = const()[name = tensor("op_451_end_0"), val = tensor([1, 3, 256])]; - tensor var_451_end_mask_0 = const()[name = tensor("op_451_end_mask_0"), val = tensor([true, false, true])]; - tensor var_451 = slice_by_index(begin = var_451_begin_0, end = var_451_end_0, end_mask = var_451_end_mask_0, x = x_9)[name = tensor("op_451")]; - tensor var_454_begin_0 = const()[name = tensor("op_454_begin_0"), val = tensor([0, 1, 0])]; - tensor var_454_end_0 = const()[name = tensor("op_454_end_0"), val = tensor([1, 16, 256])]; - tensor var_454_end_mask_0 = const()[name = tensor("op_454_end_mask_0"), val = tensor([true, true, true])]; - tensor var_454 = slice_by_index(begin = var_454_begin_0, end = var_454_end_0, end_mask = var_454_end_mask_0, x = window_15)[name = tensor("op_454")]; + tensor window_15 = concat(axis = var_82, interleave = window_15_interleave_0, values = (var_502, var_499))[name = tensor("window_15")]; + tensor var_507_begin_0 = const()[name = tensor("op_507_begin_0"), val = tensor([0, 2, 0])]; + tensor var_507_end_0 = const()[name = tensor("op_507_end_0"), val = tensor([1, 3, 256])]; + tensor var_507_end_mask_0 = const()[name = tensor("op_507_end_mask_0"), val = tensor([true, false, true])]; + tensor var_507 = slice_by_index(begin = var_507_begin_0, end = var_507_end_0, end_mask = var_507_end_mask_0, x = x_9)[name = tensor("op_507")]; + tensor var_510_begin_0 = const()[name = tensor("op_510_begin_0"), val = tensor([0, 1, 0])]; + tensor var_510_end_0 = const()[name = tensor("op_510_end_0"), val = tensor([1, 16, 256])]; + tensor var_510_end_mask_0 = const()[name = tensor("op_510_end_mask_0"), val = tensor([true, true, true])]; + tensor var_510 = slice_by_index(begin = var_510_begin_0, end = var_510_end_0, end_mask = var_510_end_mask_0, x = window_15)[name = tensor("op_510")]; tensor window_17_interleave_0 = const()[name = tensor("window_17_interleave_0"), val = tensor(false)]; - tensor window_17 = concat(axis = var_26, interleave = window_17_interleave_0, values = (var_454, var_451))[name = tensor("window_17")]; - tensor var_459_begin_0 = const()[name = tensor("op_459_begin_0"), val = tensor([0, 3, 0])]; - tensor var_459_end_0 = const()[name = tensor("op_459_end_0"), val = tensor([1, 1, 256])]; - tensor var_459_end_mask_0 = const()[name = tensor("op_459_end_mask_0"), val = tensor([true, true, true])]; - tensor var_459 = slice_by_index(begin = var_459_begin_0, end = var_459_end_0, end_mask = var_459_end_mask_0, x = x_9)[name = tensor("op_459")]; - tensor var_462_begin_0 = const()[name = tensor("op_462_begin_0"), val = tensor([0, 1, 0])]; - tensor var_462_end_0 = const()[name = tensor("op_462_end_0"), val = tensor([1, 16, 256])]; - tensor var_462_end_mask_0 = const()[name = tensor("op_462_end_mask_0"), val = tensor([true, true, true])]; - tensor var_462 = slice_by_index(begin = var_462_begin_0, end = var_462_end_0, end_mask = var_462_end_mask_0, x = window_17)[name = tensor("op_462")]; + tensor window_17 = concat(axis = var_82, interleave = window_17_interleave_0, values = (var_510, var_507))[name = tensor("window_17")]; + tensor var_515_begin_0 = const()[name = tensor("op_515_begin_0"), val = tensor([0, 3, 0])]; + tensor var_515_end_0 = const()[name = tensor("op_515_end_0"), val = tensor([1, 1, 256])]; + tensor var_515_end_mask_0 = const()[name = tensor("op_515_end_mask_0"), val = tensor([true, true, true])]; + tensor var_515 = slice_by_index(begin = var_515_begin_0, end = var_515_end_0, end_mask = var_515_end_mask_0, x = x_9)[name = tensor("op_515")]; + tensor var_518_begin_0 = const()[name = tensor("op_518_begin_0"), val = tensor([0, 1, 0])]; + tensor var_518_end_0 = const()[name = tensor("op_518_end_0"), val = tensor([1, 16, 256])]; + tensor var_518_end_mask_0 = const()[name = tensor("op_518_end_mask_0"), val = tensor([true, true, true])]; + tensor var_518 = slice_by_index(begin = var_518_begin_0, end = var_518_end_0, end_mask = var_518_end_mask_0, x = window_17)[name = tensor("op_518")]; tensor window_19_interleave_0 = const()[name = tensor("window_19_interleave_0"), val = tensor(false)]; - tensor window_19 = concat(axis = var_26, interleave = window_19_interleave_0, values = (var_462, var_459))[name = tensor("window_19")]; - tensor input_61_interleave_0 = const()[name = tensor("input_61_interleave_0"), val = tensor(false)]; - tensor input_61 = concat(axis = var_23, interleave = input_61_interleave_0, values = (window_13, window_15, window_17, window_19))[name = tensor("input_61")]; + tensor window_19 = concat(axis = var_82, interleave = window_19_interleave_0, values = (var_518, var_515))[name = tensor("window_19")]; + tensor input_63_interleave_0 = const()[name = tensor("input_63_interleave_0"), val = tensor(false)]; + tensor input_63 = concat(axis = var_69, interleave = input_63_interleave_0, values = (window_13, window_15, window_17, window_19))[name = tensor("input_63")]; tensor x_11_axes_0 = const()[name = tensor("x_11_axes_0"), val = tensor([-1])]; - tensor x_11 = layer_norm(axes = x_11_axes_0, beta = encoder_conv_module_1_sequential_0_bias, epsilon = var_28, gamma = encoder_conv_module_1_sequential_0_weight, x = input_61)[name = tensor("x_11")]; - tensor input_63_perm_0 = const()[name = tensor("input_63_perm_0"), val = tensor([0, 2, 1])]; + tensor x_11 = layer_norm(axes = x_11_axes_0, beta = inner_encoder_conv_module_1_sequential_0_bias, epsilon = var_66, gamma = inner_encoder_conv_module_1_sequential_0_weight, x = input_63)[name = tensor("x_11")]; + tensor input_65_perm_0 = const()[name = tensor("input_65_perm_0"), val = tensor([0, 2, 1])]; tensor inputs_13_pad_type_0 = const()[name = tensor("inputs_13_pad_type_0"), val = tensor("valid")]; tensor inputs_13_strides_0 = const()[name = tensor("inputs_13_strides_0"), val = tensor([1])]; tensor inputs_13_pad_0 = const()[name = tensor("inputs_13_pad_0"), val = tensor([0, 0])]; tensor inputs_13_dilations_0 = const()[name = tensor("inputs_13_dilations_0"), val = tensor([1])]; tensor inputs_13_groups_0 = const()[name = tensor("inputs_13_groups_0"), val = tensor(1)]; - tensor input_63 = transpose(perm = input_63_perm_0, x = x_11)[name = tensor("transpose_47")]; - tensor inputs_13 = conv(bias = encoder_conv_module_1_sequential_2_conv_bias, dilations = inputs_13_dilations_0, groups = inputs_13_groups_0, pad = inputs_13_pad_0, pad_type = inputs_13_pad_type_0, strides = inputs_13_strides_0, weight = encoder_conv_module_1_sequential_2_conv_weight, x = input_63)[name = tensor("inputs_13")]; - tensor var_487_split_sizes_0 = const()[name = tensor("op_487_split_sizes_0"), val = tensor([256, 256])]; - tensor var_487_axis_0 = const()[name = tensor("op_487_axis_0"), val = tensor(1)]; - tensor var_487_0, tensor var_487_1 = split(axis = var_487_axis_0, split_sizes = var_487_split_sizes_0, x = inputs_13)[name = tensor("op_487")]; - tensor var_489 = sigmoid(x = var_487_1)[name = tensor("op_489")]; - tensor inputs_15 = mul(x = var_487_0, y = var_489)[name = tensor("inputs_15")]; + tensor input_65 = transpose(perm = input_65_perm_0, x = x_11)[name = tensor("transpose_47")]; + tensor inputs_13 = conv(bias = inner_encoder_conv_module_1_sequential_2_conv_bias, dilations = inputs_13_dilations_0, groups = inputs_13_groups_0, pad = inputs_13_pad_0, pad_type = inputs_13_pad_type_0, strides = inputs_13_strides_0, weight = inner_encoder_conv_module_1_sequential_2_conv_weight, x = input_65)[name = tensor("inputs_13")]; + tensor var_543_split_sizes_0 = const()[name = tensor("op_543_split_sizes_0"), val = tensor([256, 256])]; + tensor var_543_axis_0 = const()[name = tensor("op_543_axis_0"), val = tensor(1)]; + tensor var_543_0, tensor var_543_1 = split(axis = var_543_axis_0, split_sizes = var_543_split_sizes_0, x = inputs_13)[name = tensor("op_543")]; + tensor var_545 = sigmoid(x = var_543_1)[name = tensor("op_545")]; + tensor inputs_15 = mul(x = var_543_0, y = var_545)[name = tensor("inputs_15")]; tensor outputs_aug_3_pad_type_0 = const()[name = tensor("outputs_aug_3_pad_type_0"), val = tensor("custom")]; tensor outputs_aug_3_pad_0 = const()[name = tensor("outputs_aug_3_pad_0"), val = tensor([15, 15])]; tensor outputs_aug_3_groups_0 = const()[name = tensor("outputs_aug_3_groups_0"), val = tensor(256)]; tensor outputs_aug_3_strides_0 = const()[name = tensor("outputs_aug_3_strides_0"), val = tensor([1])]; tensor outputs_aug_3_dilations_0 = const()[name = tensor("outputs_aug_3_dilations_0"), val = tensor([1])]; - tensor outputs_aug_3 = conv(dilations = outputs_aug_3_dilations_0, groups = outputs_aug_3_groups_0, pad = outputs_aug_3_pad_0, pad_type = outputs_aug_3_pad_type_0, strides = outputs_aug_3_strides_0, weight = encoder_conv_module_1_sequential_4_conv_weight, x = inputs_15)[name = tensor("outputs_aug_3")]; - tensor input_65_begin_0 = const()[name = tensor("input_65_begin_0"), val = tensor([0, 0, 0])]; - tensor input_65_end_0 = const()[name = tensor("input_65_end_0"), val = tensor([4, 256, 16])]; - tensor input_65_end_mask_0 = const()[name = tensor("input_65_end_mask_0"), val = tensor([true, true, false])]; - tensor input_65 = slice_by_index(begin = input_65_begin_0, end = input_65_end_0, end_mask = input_65_end_mask_0, x = outputs_aug_3)[name = tensor("input_65")]; - tensor inputs_17 = batch_norm(beta = encoder_conv_module_1_sequential_5_bias, epsilon = var_28, gamma = encoder_conv_module_1_sequential_5_weight, mean = encoder_conv_module_1_sequential_5_running_mean, variance = encoder_conv_module_1_sequential_5_running_var, x = input_65)[name = tensor("inputs_17")]; - tensor input_67 = silu(x = inputs_17)[name = tensor("input_67")]; - tensor input_69_pad_type_0 = const()[name = tensor("input_69_pad_type_0"), val = tensor("valid")]; - tensor input_69_strides_0 = const()[name = tensor("input_69_strides_0"), val = tensor([1])]; - tensor input_69_pad_0 = const()[name = tensor("input_69_pad_0"), val = tensor([0, 0])]; - tensor input_69_dilations_0 = const()[name = tensor("input_69_dilations_0"), val = tensor([1])]; - tensor input_69_groups_0 = const()[name = tensor("input_69_groups_0"), val = tensor(1)]; - tensor input_69 = conv(bias = encoder_conv_module_1_sequential_7_conv_bias, dilations = input_69_dilations_0, groups = input_69_groups_0, pad = input_69_pad_0, pad_type = input_69_pad_type_0, strides = input_69_strides_0, weight = encoder_conv_module_1_sequential_7_conv_weight, x = input_67)[name = tensor("input_69")]; + tensor outputs_aug_3 = conv(dilations = outputs_aug_3_dilations_0, groups = outputs_aug_3_groups_0, pad = outputs_aug_3_pad_0, pad_type = outputs_aug_3_pad_type_0, strides = outputs_aug_3_strides_0, weight = inner_encoder_conv_module_1_sequential_4_conv_weight, x = inputs_15)[name = tensor("outputs_aug_3")]; + tensor input_67_begin_0 = const()[name = tensor("input_67_begin_0"), val = tensor([0, 0, 0])]; + tensor input_67_end_0 = const()[name = tensor("input_67_end_0"), val = tensor([4, 256, 16])]; + tensor input_67_end_mask_0 = const()[name = tensor("input_67_end_mask_0"), val = tensor([true, true, false])]; + tensor input_67 = slice_by_index(begin = input_67_begin_0, end = input_67_end_0, end_mask = input_67_end_mask_0, x = outputs_aug_3)[name = tensor("input_67")]; + tensor inputs_17 = batch_norm(beta = inner_encoder_conv_module_1_sequential_5_bias, epsilon = var_66, gamma = inner_encoder_conv_module_1_sequential_5_weight, mean = inner_encoder_conv_module_1_sequential_5_running_mean, variance = inner_encoder_conv_module_1_sequential_5_running_var, x = input_67)[name = tensor("inputs_17")]; + tensor input_69 = silu(x = inputs_17)[name = tensor("input_69")]; + tensor input_71_pad_type_0 = const()[name = tensor("input_71_pad_type_0"), val = tensor("valid")]; + tensor input_71_strides_0 = const()[name = tensor("input_71_strides_0"), val = tensor([1])]; + tensor input_71_pad_0 = const()[name = tensor("input_71_pad_0"), val = tensor([0, 0])]; + tensor input_71_dilations_0 = const()[name = tensor("input_71_dilations_0"), val = tensor([1])]; + tensor input_71_groups_0 = const()[name = tensor("input_71_groups_0"), val = tensor(1)]; + tensor input_71 = conv(bias = inner_encoder_conv_module_1_sequential_7_conv_bias, dilations = input_71_dilations_0, groups = input_71_groups_0, pad = input_71_pad_0, pad_type = input_71_pad_type_0, strides = input_71_strides_0, weight = inner_encoder_conv_module_1_sequential_7_conv_weight, x = input_69)[name = tensor("input_71")]; tensor conv_out_3_perm_0 = const()[name = tensor("conv_out_3_perm_0"), val = tensor([0, 2, 1])]; - tensor var_520_begin_0 = const()[name = tensor("op_520_begin_0"), val = tensor([0, -1, 0])]; - tensor var_520_end_0 = const()[name = tensor("op_520_end_0"), val = tensor([4, 16, 256])]; - tensor var_520_end_mask_0 = const()[name = tensor("op_520_end_mask_0"), val = tensor([true, true, true])]; - tensor conv_out_3 = transpose(perm = conv_out_3_perm_0, x = input_69)[name = tensor("transpose_46")]; - tensor var_520 = slice_by_index(begin = var_520_begin_0, end = var_520_end_0, end_mask = var_520_end_mask_0, x = conv_out_3)[name = tensor("op_520")]; - tensor var_522_perm_0 = const()[name = tensor("op_522_perm_0"), val = tensor([1, 0, 2])]; - tensor var_522 = transpose(perm = var_522_perm_0, x = var_520)[name = tensor("transpose_45")]; - tensor input_71 = add(x = x_9, y = var_522)[name = tensor("input_71")]; - tensor input_73_axes_0 = const()[name = tensor("input_73_axes_0"), val = tensor([-1])]; - tensor input_73 = layer_norm(axes = input_73_axes_0, beta = encoder_ffn2_1_module_sequential_0_bias, epsilon = var_28, gamma = encoder_ffn2_1_module_sequential_0_weight, x = input_71)[name = tensor("input_73")]; - tensor inputs_19 = linear(bias = encoder_ffn2_1_module_sequential_1_linear_bias, weight = encoder_ffn2_1_module_sequential_1_linear_weight, x = input_73)[name = tensor("linear_17")]; - tensor input_75 = silu(x = inputs_19)[name = tensor("input_75")]; - tensor input_79 = linear(bias = encoder_ffn2_1_module_sequential_4_linear_bias, weight = encoder_ffn2_1_module_sequential_4_linear_weight, x = input_75)[name = tensor("linear_18")]; - tensor var_545 = const()[name = tensor("op_545"), val = tensor(0x1p-1)]; - tensor var_546 = mul(x = input_79, y = var_545)[name = tensor("op_546")]; - tensor input_81 = add(x = var_546, y = input_71)[name = tensor("input_81")]; - tensor input_83_axes_0 = const()[name = tensor("input_83_axes_0"), val = tensor([-1])]; - tensor input_83 = layer_norm(axes = input_83_axes_0, beta = encoder_layer_norm_1_bias, epsilon = var_28, gamma = encoder_layer_norm_1_weight, x = input_81)[name = tensor("input_83")]; + tensor var_576_begin_0 = const()[name = tensor("op_576_begin_0"), val = tensor([0, -1, 0])]; + tensor var_576_end_0 = const()[name = tensor("op_576_end_0"), val = tensor([4, 16, 256])]; + tensor var_576_end_mask_0 = const()[name = tensor("op_576_end_mask_0"), val = tensor([true, true, true])]; + tensor conv_out_3 = transpose(perm = conv_out_3_perm_0, x = input_71)[name = tensor("transpose_46")]; + tensor var_576 = slice_by_index(begin = var_576_begin_0, end = var_576_end_0, end_mask = var_576_end_mask_0, x = conv_out_3)[name = tensor("op_576")]; + tensor var_578_perm_0 = const()[name = tensor("op_578_perm_0"), val = tensor([1, 0, 2])]; + tensor var_578 = transpose(perm = var_578_perm_0, x = var_576)[name = tensor("transpose_45")]; + tensor input_73 = add(x = x_9, y = var_578)[name = tensor("input_73")]; + tensor input_75_axes_0 = const()[name = tensor("input_75_axes_0"), val = tensor([-1])]; + tensor input_75 = layer_norm(axes = input_75_axes_0, beta = inner_encoder_ffn2_1_module_sequential_0_bias, epsilon = var_66, gamma = inner_encoder_ffn2_1_module_sequential_0_weight, x = input_73)[name = tensor("input_75")]; + tensor inputs_19 = linear(bias = inner_encoder_ffn2_1_module_sequential_1_linear_bias, weight = inner_encoder_ffn2_1_module_sequential_1_linear_weight, x = input_75)[name = tensor("linear_17")]; + tensor input_77 = silu(x = inputs_19)[name = tensor("input_77")]; + tensor input_81 = linear(bias = inner_encoder_ffn2_1_module_sequential_4_linear_bias, weight = inner_encoder_ffn2_1_module_sequential_4_linear_weight, x = input_77)[name = tensor("linear_18")]; + tensor var_601 = const()[name = tensor("op_601"), val = tensor(0x1p-1)]; + tensor var_602 = mul(x = input_81, y = var_601)[name = tensor("op_602")]; + tensor input_83 = add(x = var_602, y = input_73)[name = tensor("input_83")]; tensor input_85_axes_0 = const()[name = tensor("input_85_axes_0"), val = tensor([-1])]; - tensor input_85 = layer_norm(axes = input_85_axes_0, beta = encoder_ffn1_2_module_sequential_0_bias, epsilon = var_28, gamma = encoder_ffn1_2_module_sequential_0_weight, x = input_83)[name = tensor("input_85")]; - tensor inputs_21 = linear(bias = encoder_ffn1_2_module_sequential_1_linear_bias, weight = encoder_ffn1_2_module_sequential_1_linear_weight, x = input_85)[name = tensor("linear_19")]; - tensor input_87 = silu(x = inputs_21)[name = tensor("input_87")]; - tensor input_91 = linear(bias = encoder_ffn1_2_module_sequential_4_linear_bias, weight = encoder_ffn1_2_module_sequential_4_linear_weight, x = input_87)[name = tensor("linear_20")]; - tensor var_575 = const()[name = tensor("op_575"), val = tensor(0x1p-1)]; - tensor var_576 = mul(x = input_91, y = var_575)[name = tensor("op_576")]; - tensor input_93 = add(x = var_576, y = input_83)[name = tensor("input_93")]; + tensor input_85 = layer_norm(axes = input_85_axes_0, beta = inner_encoder_layer_norm_1_bias, epsilon = var_66, gamma = inner_encoder_layer_norm_1_weight, x = input_83)[name = tensor("input_85")]; + tensor input_87_axes_0 = const()[name = tensor("input_87_axes_0"), val = tensor([-1])]; + tensor input_87 = layer_norm(axes = input_87_axes_0, beta = inner_encoder_ffn1_2_module_sequential_0_bias, epsilon = var_66, gamma = inner_encoder_ffn1_2_module_sequential_0_weight, x = input_85)[name = tensor("input_87")]; + tensor inputs_21 = linear(bias = inner_encoder_ffn1_2_module_sequential_1_linear_bias, weight = inner_encoder_ffn1_2_module_sequential_1_linear_weight, x = input_87)[name = tensor("linear_19")]; + tensor input_89 = silu(x = inputs_21)[name = tensor("input_89")]; + tensor input_93 = linear(bias = inner_encoder_ffn1_2_module_sequential_4_linear_bias, weight = inner_encoder_ffn1_2_module_sequential_4_linear_weight, x = input_89)[name = tensor("linear_20")]; + tensor var_631 = const()[name = tensor("op_631"), val = tensor(0x1p-1)]; + tensor var_632 = mul(x = input_93, y = var_631)[name = tensor("op_632")]; + tensor input_95 = add(x = var_632, y = input_85)[name = tensor("input_95")]; tensor x_13_axes_0 = const()[name = tensor("x_13_axes_0"), val = tensor([-1])]; - tensor x_13 = layer_norm(axes = x_13_axes_0, beta = encoder_ret_lns_2_bias, epsilon = var_28, gamma = encoder_ret_lns_2_weight, x = input_93)[name = tensor("x_13")]; + tensor x_13 = layer_norm(axes = x_13_axes_0, beta = inner_encoder_ret_lns_2_bias, epsilon = var_66, gamma = inner_encoder_ret_lns_2_weight, x = input_95)[name = tensor("x_13")]; tensor prev_kv_5_begin_0 = const()[name = tensor("prev_kv_5_begin_0"), val = tensor([2, 0, 0, 0, 0])]; tensor prev_kv_5_end_0 = const()[name = tensor("prev_kv_5_end_0"), val = tensor([3, 1, 4, 64, 64])]; tensor prev_kv_5_end_mask_0 = const()[name = tensor("prev_kv_5_end_mask_0"), val = tensor([false, true, true, true, true])]; @@ -593,173 +615,173 @@ program(1.0) tensor prev_scale_5_end_mask_0 = const()[name = tensor("prev_scale_5_end_mask_0"), val = tensor([false, true])]; tensor prev_scale_5_squeeze_mask_0 = const()[name = tensor("prev_scale_5_squeeze_mask_0"), val = tensor([true, false])]; tensor prev_scale_5 = slice_by_index(begin = prev_scale_5_begin_0, end = prev_scale_5_end_0, end_mask = prev_scale_5_end_mask_0, squeeze_mask = prev_scale_5_squeeze_mask_0, x = enc_scale)[name = tensor("prev_scale_5")]; - tensor var_590 = linear(bias = encoder_q_proj_2_bias, weight = encoder_q_proj_2_weight, x = x_13)[name = tensor("linear_21")]; - tensor var_591 = const()[name = tensor("op_591"), val = tensor([1, 4, 4, 64])]; - tensor var_592 = reshape(shape = var_591, x = var_590)[name = tensor("op_592")]; + tensor var_646 = linear(bias = inner_encoder_q_proj_2_bias, weight = inner_encoder_q_proj_2_weight, x = x_13)[name = tensor("linear_21")]; + tensor var_647 = const()[name = tensor("op_647"), val = tensor([1, 4, 4, 64])]; + tensor var_648 = reshape(shape = var_647, x = var_646)[name = tensor("op_648")]; tensor q_5_perm_0 = const()[name = tensor("q_5_perm_0"), val = tensor([0, 2, 1, 3])]; - tensor var_596 = linear(bias = encoder_k_proj_2_bias, weight = encoder_k_proj_2_weight, x = x_13)[name = tensor("linear_22")]; - tensor var_597 = const()[name = tensor("op_597"), val = tensor(0x1p-3)]; - tensor var_598 = mul(x = var_596, y = var_597)[name = tensor("op_598")]; - tensor var_599 = const()[name = tensor("op_599"), val = tensor([1, 4, 4, 64])]; - tensor var_600 = reshape(shape = var_599, x = var_598)[name = tensor("op_600")]; + tensor var_652 = linear(bias = inner_encoder_k_proj_2_bias, weight = inner_encoder_k_proj_2_weight, x = x_13)[name = tensor("linear_22")]; + tensor var_653 = const()[name = tensor("op_653"), val = tensor(0x1p-3)]; + tensor var_654 = mul(x = var_652, y = var_653)[name = tensor("op_654")]; + tensor var_655 = const()[name = tensor("op_655"), val = tensor([1, 4, 4, 64])]; + tensor var_656 = reshape(shape = var_655, x = var_654)[name = tensor("op_656")]; tensor k_5_perm_0 = const()[name = tensor("k_5_perm_0"), val = tensor([0, 2, 1, 3])]; - tensor var_604 = linear(bias = encoder_v_proj_2_bias, weight = encoder_v_proj_2_weight, x = x_13)[name = tensor("linear_23")]; - tensor var_605 = const()[name = tensor("op_605"), val = tensor([1, 4, 4, 64])]; - tensor var_606 = reshape(shape = var_605, x = var_604)[name = tensor("op_606")]; + tensor var_660 = linear(bias = inner_encoder_v_proj_2_bias, weight = inner_encoder_v_proj_2_weight, x = x_13)[name = tensor("linear_23")]; + tensor var_661 = const()[name = tensor("op_661"), val = tensor([1, 4, 4, 64])]; + tensor var_662 = reshape(shape = var_661, x = var_660)[name = tensor("op_662")]; tensor v_5_perm_0 = const()[name = tensor("v_5_perm_0"), val = tensor([0, 2, 1, 3])]; - tensor input_97 = linear(bias = encoder_g_proj_2_bias, weight = encoder_g_proj_2_weight, x = x_13)[name = tensor("linear_24")]; + tensor input_99 = linear(bias = inner_encoder_g_proj_2_bias, weight = inner_encoder_g_proj_2_weight, x = x_13)[name = tensor("linear_24")]; tensor sqrt_s0_5 = sqrt(x = prev_scale_5)[name = tensor("sqrt_s0_5")]; - tensor s_t_5 = add(x = prev_scale_5, y = encoder__t_index)[name = tensor("s_t_5")]; + tensor s_t_5 = add(x = prev_scale_5, y = inner_encoder__t_index)[name = tensor("s_t_5")]; tensor sqrt_s_t_5 = sqrt(x = s_t_5)[name = tensor("sqrt_s_t_5")]; tensor qk_5_transpose_x_1 = const()[name = tensor("qk_5_transpose_x_1"), val = tensor(false)]; tensor qk_5_transpose_y_1 = const()[name = tensor("qk_5_transpose_y_1"), val = tensor(true)]; - tensor k_5 = transpose(perm = k_5_perm_0, x = var_600)[name = tensor("transpose_43")]; - tensor q_5 = transpose(perm = q_5_perm_0, x = var_592)[name = tensor("transpose_44")]; + tensor k_5 = transpose(perm = k_5_perm_0, x = var_656)[name = tensor("transpose_43")]; + tensor q_5 = transpose(perm = q_5_perm_0, x = var_648)[name = tensor("transpose_44")]; tensor qk_5 = matmul(transpose_x = qk_5_transpose_x_1, transpose_y = qk_5_transpose_y_1, x = q_5, y = k_5)[name = tensor("qk_5")]; - tensor var_616 = const()[name = tensor("op_616"), val = tensor([4, 1])]; - tensor var_617 = reshape(shape = var_616, x = sqrt_s_t_5)[name = tensor("op_617")]; - tensor M_5 = real_div(x = encoder__causal_mask, y = var_617)[name = tensor("M_5")]; - tensor var_619 = mul(x = qk_5, y = M_5)[name = tensor("op_619")]; + tensor var_672 = const()[name = tensor("op_672"), val = tensor([4, 1])]; + tensor var_673 = reshape(shape = var_672, x = sqrt_s_t_5)[name = tensor("op_673")]; + tensor M_5 = real_div(x = inner_encoder__causal_mask, y = var_673)[name = tensor("M_5")]; + tensor var_675 = mul(x = qk_5, y = M_5)[name = tensor("op_675")]; tensor inner_5_transpose_x_0 = const()[name = tensor("inner_5_transpose_x_0"), val = tensor(false)]; tensor inner_5_transpose_y_0 = const()[name = tensor("inner_5_transpose_y_0"), val = tensor(false)]; - tensor v_5 = transpose(perm = v_5_perm_0, x = var_606)[name = tensor("transpose_42")]; - tensor inner_5 = matmul(transpose_x = inner_5_transpose_x_0, transpose_y = inner_5_transpose_y_0, x = var_619, y = v_5)[name = tensor("inner_5")]; - tensor var_621_transpose_x_0 = const()[name = tensor("op_621_transpose_x_0"), val = tensor(false)]; - tensor var_621_transpose_y_0 = const()[name = tensor("op_621_transpose_y_0"), val = tensor(false)]; - tensor var_621 = matmul(transpose_x = var_621_transpose_x_0, transpose_y = var_621_transpose_y_0, x = q_5, y = prev_kv_5)[name = tensor("op_621")]; - tensor var_622 = real_div(x = sqrt_s0_5, y = sqrt_s_t_5)[name = tensor("op_622")]; - tensor var_623 = const()[name = tensor("op_623"), val = tensor([1, 1, 4, 1])]; - tensor var_624 = reshape(shape = var_623, x = var_622)[name = tensor("op_624")]; - tensor cross_5 = mul(x = var_621, y = var_624)[name = tensor("cross_5")]; + tensor v_5 = transpose(perm = v_5_perm_0, x = var_662)[name = tensor("transpose_42")]; + tensor inner_5 = matmul(transpose_x = inner_5_transpose_x_0, transpose_y = inner_5_transpose_y_0, x = var_675, y = v_5)[name = tensor("inner_5")]; + tensor var_677_transpose_x_0 = const()[name = tensor("op_677_transpose_x_0"), val = tensor(false)]; + tensor var_677_transpose_y_0 = const()[name = tensor("op_677_transpose_y_0"), val = tensor(false)]; + tensor var_677 = matmul(transpose_x = var_677_transpose_x_0, transpose_y = var_677_transpose_y_0, x = q_5, y = prev_kv_5)[name = tensor("op_677")]; + tensor var_678 = real_div(x = sqrt_s0_5, y = sqrt_s_t_5)[name = tensor("op_678")]; + tensor var_679 = const()[name = tensor("op_679"), val = tensor([1, 1, 4, 1])]; + tensor var_680 = reshape(shape = var_679, x = var_678)[name = tensor("op_680")]; + tensor cross_5 = mul(x = var_677, y = var_680)[name = tensor("cross_5")]; tensor out_13 = add(x = inner_5, y = cross_5)[name = tensor("out_13")]; - tensor var_627 = mul(x = prev_kv_5, y = sqrt_s0_5)[name = tensor("op_627")]; - tensor var_629_transpose_x_1 = const()[name = tensor("op_629_transpose_x_1"), val = tensor(true)]; - tensor var_629_transpose_y_1 = const()[name = tensor("op_629_transpose_y_1"), val = tensor(false)]; - tensor var_629 = matmul(transpose_x = var_629_transpose_x_1, transpose_y = var_629_transpose_y_1, x = k_5, y = v_5)[name = tensor("op_629")]; - tensor new_kv_unnorm_5 = add(x = var_627, y = var_629)[name = tensor("new_kv_unnorm_5")]; - tensor var_631 = const()[name = tensor("op_631"), val = tensor(0x1p+2)]; - tensor new_scale_5 = add(x = prev_scale_5, y = var_631)[name = tensor("new_scale_5")]; - tensor var_633 = sqrt(x = new_scale_5)[name = tensor("op_633")]; - tensor var_634 = real_div(x = new_kv_unnorm_5, y = var_633)[name = tensor("op_634")]; - tensor var_635_perm_0 = const()[name = tensor("op_635_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_683 = mul(x = prev_kv_5, y = sqrt_s0_5)[name = tensor("op_683")]; + tensor var_685_transpose_x_1 = const()[name = tensor("op_685_transpose_x_1"), val = tensor(true)]; + tensor var_685_transpose_y_1 = const()[name = tensor("op_685_transpose_y_1"), val = tensor(false)]; + tensor var_685 = matmul(transpose_x = var_685_transpose_x_1, transpose_y = var_685_transpose_y_1, x = k_5, y = v_5)[name = tensor("op_685")]; + tensor new_kv_unnorm_5 = add(x = var_683, y = var_685)[name = tensor("new_kv_unnorm_5")]; + tensor var_687 = const()[name = tensor("op_687"), val = tensor(0x1p+2)]; + tensor new_scale_5 = add(x = prev_scale_5, y = var_687)[name = tensor("new_scale_5")]; + tensor var_689 = sqrt(x = new_scale_5)[name = tensor("op_689")]; + tensor var_690 = real_div(x = new_kv_unnorm_5, y = var_689)[name = tensor("op_690")]; + tensor var_691_perm_0 = const()[name = tensor("op_691_perm_0"), val = tensor([0, 2, 1, 3])]; tensor out_15_axes_0 = const()[name = tensor("out_15_axes_0"), val = tensor([-1])]; - tensor var_635 = transpose(perm = var_635_perm_0, x = out_13)[name = tensor("transpose_41")]; - tensor out_15 = layer_norm(axes = out_15_axes_0, epsilon = var_17, x = var_635)[name = tensor("out_15")]; - tensor var_639 = const()[name = tensor("op_639"), val = tensor([1, 4, 256])]; - tensor out_17 = reshape(shape = var_639, x = out_15)[name = tensor("out_17")]; - tensor var_641 = silu(x = input_97)[name = tensor("op_641")]; - tensor input_99 = mul(x = var_641, y = out_17)[name = tensor("input_99")]; - tensor ret_out_5 = linear(bias = encoder_out_proj_2_bias, weight = encoder_out_proj_2_weight, x = input_99)[name = tensor("linear_25")]; - tensor x_15 = add(x = input_93, y = ret_out_5)[name = tensor("x_15")]; + tensor var_691 = transpose(perm = var_691_perm_0, x = out_13)[name = tensor("transpose_41")]; + tensor out_15 = layer_norm(axes = out_15_axes_0, epsilon = var_74, x = var_691)[name = tensor("out_15")]; + tensor var_695 = const()[name = tensor("op_695"), val = tensor([1, 4, 256])]; + tensor out_17 = reshape(shape = var_695, x = out_15)[name = tensor("out_17")]; + tensor var_697 = silu(x = input_99)[name = tensor("op_697")]; + tensor input_101 = mul(x = var_697, y = out_17)[name = tensor("input_101")]; + tensor ret_out_5 = linear(bias = inner_encoder_out_proj_2_bias, weight = inner_encoder_out_proj_2_weight, x = input_101)[name = tensor("linear_25")]; + tensor x_15 = add(x = input_95, y = ret_out_5)[name = tensor("x_15")]; tensor window_21_begin_0 = const()[name = tensor("window_21_begin_0"), val = tensor([2, 0, 0, 0])]; tensor window_21_end_0 = const()[name = tensor("window_21_end_0"), val = tensor([3, 1, 16, 256])]; tensor window_21_end_mask_0 = const()[name = tensor("window_21_end_mask_0"), val = tensor([false, true, true, true])]; tensor window_21_squeeze_mask_0 = const()[name = tensor("window_21_squeeze_mask_0"), val = tensor([true, false, false, false])]; tensor window_21 = slice_by_index(begin = window_21_begin_0, end = window_21_end_0, end_mask = window_21_end_mask_0, squeeze_mask = window_21_squeeze_mask_0, x = enc_conv_cache)[name = tensor("window_21")]; - tensor var_649_begin_0 = const()[name = tensor("op_649_begin_0"), val = tensor([0, 0, 0])]; - tensor var_649_end_0 = const()[name = tensor("op_649_end_0"), val = tensor([1, 1, 256])]; - tensor var_649_end_mask_0 = const()[name = tensor("op_649_end_mask_0"), val = tensor([true, false, true])]; - tensor var_649 = slice_by_index(begin = var_649_begin_0, end = var_649_end_0, end_mask = var_649_end_mask_0, x = x_15)[name = tensor("op_649")]; - tensor var_652_begin_0 = const()[name = tensor("op_652_begin_0"), val = tensor([0, 1, 0])]; - tensor var_652_end_0 = const()[name = tensor("op_652_end_0"), val = tensor([1, 16, 256])]; - tensor var_652_end_mask_0 = const()[name = tensor("op_652_end_mask_0"), val = tensor([true, true, true])]; - tensor var_652 = slice_by_index(begin = var_652_begin_0, end = var_652_end_0, end_mask = var_652_end_mask_0, x = window_21)[name = tensor("op_652")]; + tensor var_705_begin_0 = const()[name = tensor("op_705_begin_0"), val = tensor([0, 0, 0])]; + tensor var_705_end_0 = const()[name = tensor("op_705_end_0"), val = tensor([1, 1, 256])]; + tensor var_705_end_mask_0 = const()[name = tensor("op_705_end_mask_0"), val = tensor([true, false, true])]; + tensor var_705 = slice_by_index(begin = var_705_begin_0, end = var_705_end_0, end_mask = var_705_end_mask_0, x = x_15)[name = tensor("op_705")]; + tensor var_708_begin_0 = const()[name = tensor("op_708_begin_0"), val = tensor([0, 1, 0])]; + tensor var_708_end_0 = const()[name = tensor("op_708_end_0"), val = tensor([1, 16, 256])]; + tensor var_708_end_mask_0 = const()[name = tensor("op_708_end_mask_0"), val = tensor([true, true, true])]; + tensor var_708 = slice_by_index(begin = var_708_begin_0, end = var_708_end_0, end_mask = var_708_end_mask_0, x = window_21)[name = tensor("op_708")]; tensor window_23_interleave_0 = const()[name = tensor("window_23_interleave_0"), val = tensor(false)]; - tensor window_23 = concat(axis = var_26, interleave = window_23_interleave_0, values = (var_652, var_649))[name = tensor("window_23")]; - tensor var_657_begin_0 = const()[name = tensor("op_657_begin_0"), val = tensor([0, 1, 0])]; - tensor var_657_end_0 = const()[name = tensor("op_657_end_0"), val = tensor([1, 2, 256])]; - tensor var_657_end_mask_0 = const()[name = tensor("op_657_end_mask_0"), val = tensor([true, false, true])]; - tensor var_657 = slice_by_index(begin = var_657_begin_0, end = var_657_end_0, end_mask = var_657_end_mask_0, x = x_15)[name = tensor("op_657")]; - tensor var_660_begin_0 = const()[name = tensor("op_660_begin_0"), val = tensor([0, 1, 0])]; - tensor var_660_end_0 = const()[name = tensor("op_660_end_0"), val = tensor([1, 16, 256])]; - tensor var_660_end_mask_0 = const()[name = tensor("op_660_end_mask_0"), val = tensor([true, true, true])]; - tensor var_660 = slice_by_index(begin = var_660_begin_0, end = var_660_end_0, end_mask = var_660_end_mask_0, x = window_23)[name = tensor("op_660")]; + tensor window_23 = concat(axis = var_82, interleave = window_23_interleave_0, values = (var_708, var_705))[name = tensor("window_23")]; + tensor var_713_begin_0 = const()[name = tensor("op_713_begin_0"), val = tensor([0, 1, 0])]; + tensor var_713_end_0 = const()[name = tensor("op_713_end_0"), val = tensor([1, 2, 256])]; + tensor var_713_end_mask_0 = const()[name = tensor("op_713_end_mask_0"), val = tensor([true, false, true])]; + tensor var_713 = slice_by_index(begin = var_713_begin_0, end = var_713_end_0, end_mask = var_713_end_mask_0, x = x_15)[name = tensor("op_713")]; + tensor var_716_begin_0 = const()[name = tensor("op_716_begin_0"), val = tensor([0, 1, 0])]; + tensor var_716_end_0 = const()[name = tensor("op_716_end_0"), val = tensor([1, 16, 256])]; + tensor var_716_end_mask_0 = const()[name = tensor("op_716_end_mask_0"), val = tensor([true, true, true])]; + tensor var_716 = slice_by_index(begin = var_716_begin_0, end = var_716_end_0, end_mask = var_716_end_mask_0, x = window_23)[name = tensor("op_716")]; tensor window_25_interleave_0 = const()[name = tensor("window_25_interleave_0"), val = tensor(false)]; - tensor window_25 = concat(axis = var_26, interleave = window_25_interleave_0, values = (var_660, var_657))[name = tensor("window_25")]; - tensor var_665_begin_0 = const()[name = tensor("op_665_begin_0"), val = tensor([0, 2, 0])]; - tensor var_665_end_0 = const()[name = tensor("op_665_end_0"), val = tensor([1, 3, 256])]; - tensor var_665_end_mask_0 = const()[name = tensor("op_665_end_mask_0"), val = tensor([true, false, true])]; - tensor var_665 = slice_by_index(begin = var_665_begin_0, end = var_665_end_0, end_mask = var_665_end_mask_0, x = x_15)[name = tensor("op_665")]; - tensor var_668_begin_0 = const()[name = tensor("op_668_begin_0"), val = tensor([0, 1, 0])]; - tensor var_668_end_0 = const()[name = tensor("op_668_end_0"), val = tensor([1, 16, 256])]; - tensor var_668_end_mask_0 = const()[name = tensor("op_668_end_mask_0"), val = tensor([true, true, true])]; - tensor var_668 = slice_by_index(begin = var_668_begin_0, end = var_668_end_0, end_mask = var_668_end_mask_0, x = window_25)[name = tensor("op_668")]; + tensor window_25 = concat(axis = var_82, interleave = window_25_interleave_0, values = (var_716, var_713))[name = tensor("window_25")]; + tensor var_721_begin_0 = const()[name = tensor("op_721_begin_0"), val = tensor([0, 2, 0])]; + tensor var_721_end_0 = const()[name = tensor("op_721_end_0"), val = tensor([1, 3, 256])]; + tensor var_721_end_mask_0 = const()[name = tensor("op_721_end_mask_0"), val = tensor([true, false, true])]; + tensor var_721 = slice_by_index(begin = var_721_begin_0, end = var_721_end_0, end_mask = var_721_end_mask_0, x = x_15)[name = tensor("op_721")]; + tensor var_724_begin_0 = const()[name = tensor("op_724_begin_0"), val = tensor([0, 1, 0])]; + tensor var_724_end_0 = const()[name = tensor("op_724_end_0"), val = tensor([1, 16, 256])]; + tensor var_724_end_mask_0 = const()[name = tensor("op_724_end_mask_0"), val = tensor([true, true, true])]; + tensor var_724 = slice_by_index(begin = var_724_begin_0, end = var_724_end_0, end_mask = var_724_end_mask_0, x = window_25)[name = tensor("op_724")]; tensor window_27_interleave_0 = const()[name = tensor("window_27_interleave_0"), val = tensor(false)]; - tensor window_27 = concat(axis = var_26, interleave = window_27_interleave_0, values = (var_668, var_665))[name = tensor("window_27")]; - tensor var_673_begin_0 = const()[name = tensor("op_673_begin_0"), val = tensor([0, 3, 0])]; - tensor var_673_end_0 = const()[name = tensor("op_673_end_0"), val = tensor([1, 1, 256])]; - tensor var_673_end_mask_0 = const()[name = tensor("op_673_end_mask_0"), val = tensor([true, true, true])]; - tensor var_673 = slice_by_index(begin = var_673_begin_0, end = var_673_end_0, end_mask = var_673_end_mask_0, x = x_15)[name = tensor("op_673")]; - tensor var_676_begin_0 = const()[name = tensor("op_676_begin_0"), val = tensor([0, 1, 0])]; - tensor var_676_end_0 = const()[name = tensor("op_676_end_0"), val = tensor([1, 16, 256])]; - tensor var_676_end_mask_0 = const()[name = tensor("op_676_end_mask_0"), val = tensor([true, true, true])]; - tensor var_676 = slice_by_index(begin = var_676_begin_0, end = var_676_end_0, end_mask = var_676_end_mask_0, x = window_27)[name = tensor("op_676")]; + tensor window_27 = concat(axis = var_82, interleave = window_27_interleave_0, values = (var_724, var_721))[name = tensor("window_27")]; + tensor var_729_begin_0 = const()[name = tensor("op_729_begin_0"), val = tensor([0, 3, 0])]; + tensor var_729_end_0 = const()[name = tensor("op_729_end_0"), val = tensor([1, 1, 256])]; + tensor var_729_end_mask_0 = const()[name = tensor("op_729_end_mask_0"), val = tensor([true, true, true])]; + tensor var_729 = slice_by_index(begin = var_729_begin_0, end = var_729_end_0, end_mask = var_729_end_mask_0, x = x_15)[name = tensor("op_729")]; + tensor var_732_begin_0 = const()[name = tensor("op_732_begin_0"), val = tensor([0, 1, 0])]; + tensor var_732_end_0 = const()[name = tensor("op_732_end_0"), val = tensor([1, 16, 256])]; + tensor var_732_end_mask_0 = const()[name = tensor("op_732_end_mask_0"), val = tensor([true, true, true])]; + tensor var_732 = slice_by_index(begin = var_732_begin_0, end = var_732_end_0, end_mask = var_732_end_mask_0, x = window_27)[name = tensor("op_732")]; tensor window_29_interleave_0 = const()[name = tensor("window_29_interleave_0"), val = tensor(false)]; - tensor window_29 = concat(axis = var_26, interleave = window_29_interleave_0, values = (var_676, var_673))[name = tensor("window_29")]; - tensor input_101_interleave_0 = const()[name = tensor("input_101_interleave_0"), val = tensor(false)]; - tensor input_101 = concat(axis = var_23, interleave = input_101_interleave_0, values = (window_23, window_25, window_27, window_29))[name = tensor("input_101")]; + tensor window_29 = concat(axis = var_82, interleave = window_29_interleave_0, values = (var_732, var_729))[name = tensor("window_29")]; + tensor input_103_interleave_0 = const()[name = tensor("input_103_interleave_0"), val = tensor(false)]; + tensor input_103 = concat(axis = var_69, interleave = input_103_interleave_0, values = (window_23, window_25, window_27, window_29))[name = tensor("input_103")]; tensor x_17_axes_0 = const()[name = tensor("x_17_axes_0"), val = tensor([-1])]; - tensor x_17 = layer_norm(axes = x_17_axes_0, beta = encoder_conv_module_2_sequential_0_bias, epsilon = var_28, gamma = encoder_conv_module_2_sequential_0_weight, x = input_101)[name = tensor("x_17")]; - tensor input_103_perm_0 = const()[name = tensor("input_103_perm_0"), val = tensor([0, 2, 1])]; + tensor x_17 = layer_norm(axes = x_17_axes_0, beta = inner_encoder_conv_module_2_sequential_0_bias, epsilon = var_66, gamma = inner_encoder_conv_module_2_sequential_0_weight, x = input_103)[name = tensor("x_17")]; + tensor input_105_perm_0 = const()[name = tensor("input_105_perm_0"), val = tensor([0, 2, 1])]; tensor inputs_23_pad_type_0 = const()[name = tensor("inputs_23_pad_type_0"), val = tensor("valid")]; tensor inputs_23_strides_0 = const()[name = tensor("inputs_23_strides_0"), val = tensor([1])]; tensor inputs_23_pad_0 = const()[name = tensor("inputs_23_pad_0"), val = tensor([0, 0])]; tensor inputs_23_dilations_0 = const()[name = tensor("inputs_23_dilations_0"), val = tensor([1])]; tensor inputs_23_groups_0 = const()[name = tensor("inputs_23_groups_0"), val = tensor(1)]; - tensor input_103 = transpose(perm = input_103_perm_0, x = x_17)[name = tensor("transpose_40")]; - tensor inputs_23 = conv(bias = encoder_conv_module_2_sequential_2_conv_bias, dilations = inputs_23_dilations_0, groups = inputs_23_groups_0, pad = inputs_23_pad_0, pad_type = inputs_23_pad_type_0, strides = inputs_23_strides_0, weight = encoder_conv_module_2_sequential_2_conv_weight, x = input_103)[name = tensor("inputs_23")]; - tensor var_701_split_sizes_0 = const()[name = tensor("op_701_split_sizes_0"), val = tensor([256, 256])]; - tensor var_701_axis_0 = const()[name = tensor("op_701_axis_0"), val = tensor(1)]; - tensor var_701_0, tensor var_701_1 = split(axis = var_701_axis_0, split_sizes = var_701_split_sizes_0, x = inputs_23)[name = tensor("op_701")]; - tensor var_703 = sigmoid(x = var_701_1)[name = tensor("op_703")]; - tensor inputs_25 = mul(x = var_701_0, y = var_703)[name = tensor("inputs_25")]; + tensor input_105 = transpose(perm = input_105_perm_0, x = x_17)[name = tensor("transpose_40")]; + tensor inputs_23 = conv(bias = inner_encoder_conv_module_2_sequential_2_conv_bias, dilations = inputs_23_dilations_0, groups = inputs_23_groups_0, pad = inputs_23_pad_0, pad_type = inputs_23_pad_type_0, strides = inputs_23_strides_0, weight = inner_encoder_conv_module_2_sequential_2_conv_weight, x = input_105)[name = tensor("inputs_23")]; + tensor var_757_split_sizes_0 = const()[name = tensor("op_757_split_sizes_0"), val = tensor([256, 256])]; + tensor var_757_axis_0 = const()[name = tensor("op_757_axis_0"), val = tensor(1)]; + tensor var_757_0, tensor var_757_1 = split(axis = var_757_axis_0, split_sizes = var_757_split_sizes_0, x = inputs_23)[name = tensor("op_757")]; + tensor var_759 = sigmoid(x = var_757_1)[name = tensor("op_759")]; + tensor inputs_25 = mul(x = var_757_0, y = var_759)[name = tensor("inputs_25")]; tensor outputs_aug_5_pad_type_0 = const()[name = tensor("outputs_aug_5_pad_type_0"), val = tensor("custom")]; tensor outputs_aug_5_pad_0 = const()[name = tensor("outputs_aug_5_pad_0"), val = tensor([15, 15])]; tensor outputs_aug_5_groups_0 = const()[name = tensor("outputs_aug_5_groups_0"), val = tensor(256)]; tensor outputs_aug_5_strides_0 = const()[name = tensor("outputs_aug_5_strides_0"), val = tensor([1])]; tensor outputs_aug_5_dilations_0 = const()[name = tensor("outputs_aug_5_dilations_0"), val = tensor([1])]; - tensor outputs_aug_5 = conv(dilations = outputs_aug_5_dilations_0, groups = outputs_aug_5_groups_0, pad = outputs_aug_5_pad_0, pad_type = outputs_aug_5_pad_type_0, strides = outputs_aug_5_strides_0, weight = encoder_conv_module_2_sequential_4_conv_weight, x = inputs_25)[name = tensor("outputs_aug_5")]; - tensor input_105_begin_0 = const()[name = tensor("input_105_begin_0"), val = tensor([0, 0, 0])]; - tensor input_105_end_0 = const()[name = tensor("input_105_end_0"), val = tensor([4, 256, 16])]; - tensor input_105_end_mask_0 = const()[name = tensor("input_105_end_mask_0"), val = tensor([true, true, false])]; - tensor input_105 = slice_by_index(begin = input_105_begin_0, end = input_105_end_0, end_mask = input_105_end_mask_0, x = outputs_aug_5)[name = tensor("input_105")]; - tensor inputs_27 = batch_norm(beta = encoder_conv_module_2_sequential_5_bias, epsilon = var_28, gamma = encoder_conv_module_2_sequential_5_weight, mean = encoder_conv_module_2_sequential_5_running_mean, variance = encoder_conv_module_2_sequential_5_running_var, x = input_105)[name = tensor("inputs_27")]; - tensor input_107 = silu(x = inputs_27)[name = tensor("input_107")]; - tensor input_109_pad_type_0 = const()[name = tensor("input_109_pad_type_0"), val = tensor("valid")]; - tensor input_109_strides_0 = const()[name = tensor("input_109_strides_0"), val = tensor([1])]; - tensor input_109_pad_0 = const()[name = tensor("input_109_pad_0"), val = tensor([0, 0])]; - tensor input_109_dilations_0 = const()[name = tensor("input_109_dilations_0"), val = tensor([1])]; - tensor input_109_groups_0 = const()[name = tensor("input_109_groups_0"), val = tensor(1)]; - tensor input_109 = conv(bias = encoder_conv_module_2_sequential_7_conv_bias, dilations = input_109_dilations_0, groups = input_109_groups_0, pad = input_109_pad_0, pad_type = input_109_pad_type_0, strides = input_109_strides_0, weight = encoder_conv_module_2_sequential_7_conv_weight, x = input_107)[name = tensor("input_109")]; + tensor outputs_aug_5 = conv(dilations = outputs_aug_5_dilations_0, groups = outputs_aug_5_groups_0, pad = outputs_aug_5_pad_0, pad_type = outputs_aug_5_pad_type_0, strides = outputs_aug_5_strides_0, weight = inner_encoder_conv_module_2_sequential_4_conv_weight, x = inputs_25)[name = tensor("outputs_aug_5")]; + tensor input_107_begin_0 = const()[name = tensor("input_107_begin_0"), val = tensor([0, 0, 0])]; + tensor input_107_end_0 = const()[name = tensor("input_107_end_0"), val = tensor([4, 256, 16])]; + tensor input_107_end_mask_0 = const()[name = tensor("input_107_end_mask_0"), val = tensor([true, true, false])]; + tensor input_107 = slice_by_index(begin = input_107_begin_0, end = input_107_end_0, end_mask = input_107_end_mask_0, x = outputs_aug_5)[name = tensor("input_107")]; + tensor inputs_27 = batch_norm(beta = inner_encoder_conv_module_2_sequential_5_bias, epsilon = var_66, gamma = inner_encoder_conv_module_2_sequential_5_weight, mean = inner_encoder_conv_module_2_sequential_5_running_mean, variance = inner_encoder_conv_module_2_sequential_5_running_var, x = input_107)[name = tensor("inputs_27")]; + tensor input_109 = silu(x = inputs_27)[name = tensor("input_109")]; + tensor input_111_pad_type_0 = const()[name = tensor("input_111_pad_type_0"), val = tensor("valid")]; + tensor input_111_strides_0 = const()[name = tensor("input_111_strides_0"), val = tensor([1])]; + tensor input_111_pad_0 = const()[name = tensor("input_111_pad_0"), val = tensor([0, 0])]; + tensor input_111_dilations_0 = const()[name = tensor("input_111_dilations_0"), val = tensor([1])]; + tensor input_111_groups_0 = const()[name = tensor("input_111_groups_0"), val = tensor(1)]; + tensor input_111 = conv(bias = inner_encoder_conv_module_2_sequential_7_conv_bias, dilations = input_111_dilations_0, groups = input_111_groups_0, pad = input_111_pad_0, pad_type = input_111_pad_type_0, strides = input_111_strides_0, weight = inner_encoder_conv_module_2_sequential_7_conv_weight, x = input_109)[name = tensor("input_111")]; tensor conv_out_5_perm_0 = const()[name = tensor("conv_out_5_perm_0"), val = tensor([0, 2, 1])]; - tensor var_734_begin_0 = const()[name = tensor("op_734_begin_0"), val = tensor([0, -1, 0])]; - tensor var_734_end_0 = const()[name = tensor("op_734_end_0"), val = tensor([4, 16, 256])]; - tensor var_734_end_mask_0 = const()[name = tensor("op_734_end_mask_0"), val = tensor([true, true, true])]; - tensor conv_out_5 = transpose(perm = conv_out_5_perm_0, x = input_109)[name = tensor("transpose_39")]; - tensor var_734 = slice_by_index(begin = var_734_begin_0, end = var_734_end_0, end_mask = var_734_end_mask_0, x = conv_out_5)[name = tensor("op_734")]; - tensor var_736_perm_0 = const()[name = tensor("op_736_perm_0"), val = tensor([1, 0, 2])]; - tensor var_736 = transpose(perm = var_736_perm_0, x = var_734)[name = tensor("transpose_38")]; - tensor input_111 = add(x = x_15, y = var_736)[name = tensor("input_111")]; - tensor input_113_axes_0 = const()[name = tensor("input_113_axes_0"), val = tensor([-1])]; - tensor input_113 = layer_norm(axes = input_113_axes_0, beta = encoder_ffn2_2_module_sequential_0_bias, epsilon = var_28, gamma = encoder_ffn2_2_module_sequential_0_weight, x = input_111)[name = tensor("input_113")]; - tensor inputs_29 = linear(bias = encoder_ffn2_2_module_sequential_1_linear_bias, weight = encoder_ffn2_2_module_sequential_1_linear_weight, x = input_113)[name = tensor("linear_26")]; - tensor input_115 = silu(x = inputs_29)[name = tensor("input_115")]; - tensor input_119 = linear(bias = encoder_ffn2_2_module_sequential_4_linear_bias, weight = encoder_ffn2_2_module_sequential_4_linear_weight, x = input_115)[name = tensor("linear_27")]; - tensor var_759 = const()[name = tensor("op_759"), val = tensor(0x1p-1)]; - tensor var_760 = mul(x = input_119, y = var_759)[name = tensor("op_760")]; - tensor input_121 = add(x = var_760, y = input_111)[name = tensor("input_121")]; - tensor input_123_axes_0 = const()[name = tensor("input_123_axes_0"), val = tensor([-1])]; - tensor input_123 = layer_norm(axes = input_123_axes_0, beta = encoder_layer_norm_2_bias, epsilon = var_28, gamma = encoder_layer_norm_2_weight, x = input_121)[name = tensor("input_123")]; + tensor var_790_begin_0 = const()[name = tensor("op_790_begin_0"), val = tensor([0, -1, 0])]; + tensor var_790_end_0 = const()[name = tensor("op_790_end_0"), val = tensor([4, 16, 256])]; + tensor var_790_end_mask_0 = const()[name = tensor("op_790_end_mask_0"), val = tensor([true, true, true])]; + tensor conv_out_5 = transpose(perm = conv_out_5_perm_0, x = input_111)[name = tensor("transpose_39")]; + tensor var_790 = slice_by_index(begin = var_790_begin_0, end = var_790_end_0, end_mask = var_790_end_mask_0, x = conv_out_5)[name = tensor("op_790")]; + tensor var_792_perm_0 = const()[name = tensor("op_792_perm_0"), val = tensor([1, 0, 2])]; + tensor var_792 = transpose(perm = var_792_perm_0, x = var_790)[name = tensor("transpose_38")]; + tensor input_113 = add(x = x_15, y = var_792)[name = tensor("input_113")]; + tensor input_115_axes_0 = const()[name = tensor("input_115_axes_0"), val = tensor([-1])]; + tensor input_115 = layer_norm(axes = input_115_axes_0, beta = inner_encoder_ffn2_2_module_sequential_0_bias, epsilon = var_66, gamma = inner_encoder_ffn2_2_module_sequential_0_weight, x = input_113)[name = tensor("input_115")]; + tensor inputs_29 = linear(bias = inner_encoder_ffn2_2_module_sequential_1_linear_bias, weight = inner_encoder_ffn2_2_module_sequential_1_linear_weight, x = input_115)[name = tensor("linear_26")]; + tensor input_117 = silu(x = inputs_29)[name = tensor("input_117")]; + tensor input_121 = linear(bias = inner_encoder_ffn2_2_module_sequential_4_linear_bias, weight = inner_encoder_ffn2_2_module_sequential_4_linear_weight, x = input_117)[name = tensor("linear_27")]; + tensor var_815 = const()[name = tensor("op_815"), val = tensor(0x1p-1)]; + tensor var_816 = mul(x = input_121, y = var_815)[name = tensor("op_816")]; + tensor input_123 = add(x = var_816, y = input_113)[name = tensor("input_123")]; tensor input_125_axes_0 = const()[name = tensor("input_125_axes_0"), val = tensor([-1])]; - tensor input_125 = layer_norm(axes = input_125_axes_0, beta = encoder_ffn1_3_module_sequential_0_bias, epsilon = var_28, gamma = encoder_ffn1_3_module_sequential_0_weight, x = input_123)[name = tensor("input_125")]; - tensor inputs_31 = linear(bias = encoder_ffn1_3_module_sequential_1_linear_bias, weight = encoder_ffn1_3_module_sequential_1_linear_weight, x = input_125)[name = tensor("linear_28")]; - tensor input_127 = silu(x = inputs_31)[name = tensor("input_127")]; - tensor input_131 = linear(bias = encoder_ffn1_3_module_sequential_4_linear_bias, weight = encoder_ffn1_3_module_sequential_4_linear_weight, x = input_127)[name = tensor("linear_29")]; - tensor var_789 = const()[name = tensor("op_789"), val = tensor(0x1p-1)]; - tensor var_790 = mul(x = input_131, y = var_789)[name = tensor("op_790")]; - tensor input_133 = add(x = var_790, y = input_123)[name = tensor("input_133")]; + tensor input_125 = layer_norm(axes = input_125_axes_0, beta = inner_encoder_layer_norm_2_bias, epsilon = var_66, gamma = inner_encoder_layer_norm_2_weight, x = input_123)[name = tensor("input_125")]; + tensor input_127_axes_0 = const()[name = tensor("input_127_axes_0"), val = tensor([-1])]; + tensor input_127 = layer_norm(axes = input_127_axes_0, beta = inner_encoder_ffn1_3_module_sequential_0_bias, epsilon = var_66, gamma = inner_encoder_ffn1_3_module_sequential_0_weight, x = input_125)[name = tensor("input_127")]; + tensor inputs_31 = linear(bias = inner_encoder_ffn1_3_module_sequential_1_linear_bias, weight = inner_encoder_ffn1_3_module_sequential_1_linear_weight, x = input_127)[name = tensor("linear_28")]; + tensor input_129 = silu(x = inputs_31)[name = tensor("input_129")]; + tensor input_133 = linear(bias = inner_encoder_ffn1_3_module_sequential_4_linear_bias, weight = inner_encoder_ffn1_3_module_sequential_4_linear_weight, x = input_129)[name = tensor("linear_29")]; + tensor var_845 = const()[name = tensor("op_845"), val = tensor(0x1p-1)]; + tensor var_846 = mul(x = input_133, y = var_845)[name = tensor("op_846")]; + tensor input_135 = add(x = var_846, y = input_125)[name = tensor("input_135")]; tensor x_19_axes_0 = const()[name = tensor("x_19_axes_0"), val = tensor([-1])]; - tensor x_19 = layer_norm(axes = x_19_axes_0, beta = encoder_ret_lns_3_bias, epsilon = var_28, gamma = encoder_ret_lns_3_weight, x = input_133)[name = tensor("x_19")]; + tensor x_19 = layer_norm(axes = x_19_axes_0, beta = inner_encoder_ret_lns_3_bias, epsilon = var_66, gamma = inner_encoder_ret_lns_3_weight, x = input_135)[name = tensor("x_19")]; tensor prev_kv_7_begin_0 = const()[name = tensor("prev_kv_7_begin_0"), val = tensor([3, 0, 0, 0, 0])]; tensor prev_kv_7_end_0 = const()[name = tensor("prev_kv_7_end_0"), val = tensor([4, 1, 4, 64, 64])]; tensor prev_kv_7_end_mask_0 = const()[name = tensor("prev_kv_7_end_mask_0"), val = tensor([false, true, true, true, true])]; @@ -770,209 +792,202 @@ program(1.0) tensor prev_scale_7_end_mask_0 = const()[name = tensor("prev_scale_7_end_mask_0"), val = tensor([false, true])]; tensor prev_scale_7_squeeze_mask_0 = const()[name = tensor("prev_scale_7_squeeze_mask_0"), val = tensor([true, false])]; tensor prev_scale_7 = slice_by_index(begin = prev_scale_7_begin_0, end = prev_scale_7_end_0, end_mask = prev_scale_7_end_mask_0, squeeze_mask = prev_scale_7_squeeze_mask_0, x = enc_scale)[name = tensor("prev_scale_7")]; - tensor var_804 = linear(bias = encoder_q_proj_3_bias, weight = encoder_q_proj_3_weight, x = x_19)[name = tensor("linear_30")]; - tensor var_805 = const()[name = tensor("op_805"), val = tensor([1, 4, 4, 64])]; - tensor var_806 = reshape(shape = var_805, x = var_804)[name = tensor("op_806")]; + tensor var_860 = linear(bias = inner_encoder_q_proj_3_bias, weight = inner_encoder_q_proj_3_weight, x = x_19)[name = tensor("linear_30")]; + tensor var_861 = const()[name = tensor("op_861"), val = tensor([1, 4, 4, 64])]; + tensor var_862 = reshape(shape = var_861, x = var_860)[name = tensor("op_862")]; tensor q_7_perm_0 = const()[name = tensor("q_7_perm_0"), val = tensor([0, 2, 1, 3])]; - tensor var_810 = linear(bias = encoder_k_proj_3_bias, weight = encoder_k_proj_3_weight, x = x_19)[name = tensor("linear_31")]; - tensor var_811 = const()[name = tensor("op_811"), val = tensor(0x1p-3)]; - tensor var_812 = mul(x = var_810, y = var_811)[name = tensor("op_812")]; - tensor var_813 = const()[name = tensor("op_813"), val = tensor([1, 4, 4, 64])]; - tensor var_814 = reshape(shape = var_813, x = var_812)[name = tensor("op_814")]; + tensor var_866 = linear(bias = inner_encoder_k_proj_3_bias, weight = inner_encoder_k_proj_3_weight, x = x_19)[name = tensor("linear_31")]; + tensor var_867 = const()[name = tensor("op_867"), val = tensor(0x1p-3)]; + tensor var_868 = mul(x = var_866, y = var_867)[name = tensor("op_868")]; + tensor var_869 = const()[name = tensor("op_869"), val = tensor([1, 4, 4, 64])]; + tensor var_870 = reshape(shape = var_869, x = var_868)[name = tensor("op_870")]; tensor k_7_perm_0 = const()[name = tensor("k_7_perm_0"), val = tensor([0, 2, 1, 3])]; - tensor var_818 = linear(bias = encoder_v_proj_3_bias, weight = encoder_v_proj_3_weight, x = x_19)[name = tensor("linear_32")]; - tensor var_819 = const()[name = tensor("op_819"), val = tensor([1, 4, 4, 64])]; - tensor var_820 = reshape(shape = var_819, x = var_818)[name = tensor("op_820")]; + tensor var_874 = linear(bias = inner_encoder_v_proj_3_bias, weight = inner_encoder_v_proj_3_weight, x = x_19)[name = tensor("linear_32")]; + tensor var_875 = const()[name = tensor("op_875"), val = tensor([1, 4, 4, 64])]; + tensor var_876 = reshape(shape = var_875, x = var_874)[name = tensor("op_876")]; tensor v_7_perm_0 = const()[name = tensor("v_7_perm_0"), val = tensor([0, 2, 1, 3])]; - tensor input_137 = linear(bias = encoder_g_proj_3_bias, weight = encoder_g_proj_3_weight, x = x_19)[name = tensor("linear_33")]; + tensor input_139 = linear(bias = inner_encoder_g_proj_3_bias, weight = inner_encoder_g_proj_3_weight, x = x_19)[name = tensor("linear_33")]; tensor sqrt_s0_7 = sqrt(x = prev_scale_7)[name = tensor("sqrt_s0_7")]; - tensor s_t_7 = add(x = prev_scale_7, y = encoder__t_index)[name = tensor("s_t_7")]; + tensor s_t_7 = add(x = prev_scale_7, y = inner_encoder__t_index)[name = tensor("s_t_7")]; tensor sqrt_s_t_7 = sqrt(x = s_t_7)[name = tensor("sqrt_s_t_7")]; tensor qk_7_transpose_x_1 = const()[name = tensor("qk_7_transpose_x_1"), val = tensor(false)]; tensor qk_7_transpose_y_1 = const()[name = tensor("qk_7_transpose_y_1"), val = tensor(true)]; - tensor k_7 = transpose(perm = k_7_perm_0, x = var_814)[name = tensor("transpose_36")]; - tensor q_7 = transpose(perm = q_7_perm_0, x = var_806)[name = tensor("transpose_37")]; + tensor k_7 = transpose(perm = k_7_perm_0, x = var_870)[name = tensor("transpose_36")]; + tensor q_7 = transpose(perm = q_7_perm_0, x = var_862)[name = tensor("transpose_37")]; tensor qk_7 = matmul(transpose_x = qk_7_transpose_x_1, transpose_y = qk_7_transpose_y_1, x = q_7, y = k_7)[name = tensor("qk_7")]; - tensor var_830 = const()[name = tensor("op_830"), val = tensor([4, 1])]; - tensor var_831 = reshape(shape = var_830, x = sqrt_s_t_7)[name = tensor("op_831")]; - tensor M_7 = real_div(x = encoder__causal_mask, y = var_831)[name = tensor("M_7")]; - tensor var_833 = mul(x = qk_7, y = M_7)[name = tensor("op_833")]; + tensor var_886 = const()[name = tensor("op_886"), val = tensor([4, 1])]; + tensor var_887 = reshape(shape = var_886, x = sqrt_s_t_7)[name = tensor("op_887")]; + tensor M_7 = real_div(x = inner_encoder__causal_mask, y = var_887)[name = tensor("M_7")]; + tensor var_889 = mul(x = qk_7, y = M_7)[name = tensor("op_889")]; tensor inner_7_transpose_x_0 = const()[name = tensor("inner_7_transpose_x_0"), val = tensor(false)]; tensor inner_7_transpose_y_0 = const()[name = tensor("inner_7_transpose_y_0"), val = tensor(false)]; - tensor v_7 = transpose(perm = v_7_perm_0, x = var_820)[name = tensor("transpose_35")]; - tensor inner_7 = matmul(transpose_x = inner_7_transpose_x_0, transpose_y = inner_7_transpose_y_0, x = var_833, y = v_7)[name = tensor("inner_7")]; - tensor var_835_transpose_x_0 = const()[name = tensor("op_835_transpose_x_0"), val = tensor(false)]; - tensor var_835_transpose_y_0 = const()[name = tensor("op_835_transpose_y_0"), val = tensor(false)]; - tensor var_835 = matmul(transpose_x = var_835_transpose_x_0, transpose_y = var_835_transpose_y_0, x = q_7, y = prev_kv_7)[name = tensor("op_835")]; - tensor var_836 = real_div(x = sqrt_s0_7, y = sqrt_s_t_7)[name = tensor("op_836")]; - tensor var_837 = const()[name = tensor("op_837"), val = tensor([1, 1, 4, 1])]; - tensor var_838 = reshape(shape = var_837, x = var_836)[name = tensor("op_838")]; - tensor cross_7 = mul(x = var_835, y = var_838)[name = tensor("cross_7")]; + tensor v_7 = transpose(perm = v_7_perm_0, x = var_876)[name = tensor("transpose_35")]; + tensor inner_7 = matmul(transpose_x = inner_7_transpose_x_0, transpose_y = inner_7_transpose_y_0, x = var_889, y = v_7)[name = tensor("inner_7")]; + tensor var_891_transpose_x_0 = const()[name = tensor("op_891_transpose_x_0"), val = tensor(false)]; + tensor var_891_transpose_y_0 = const()[name = tensor("op_891_transpose_y_0"), val = tensor(false)]; + tensor var_891 = matmul(transpose_x = var_891_transpose_x_0, transpose_y = var_891_transpose_y_0, x = q_7, y = prev_kv_7)[name = tensor("op_891")]; + tensor var_892 = real_div(x = sqrt_s0_7, y = sqrt_s_t_7)[name = tensor("op_892")]; + tensor var_893 = const()[name = tensor("op_893"), val = tensor([1, 1, 4, 1])]; + tensor var_894 = reshape(shape = var_893, x = var_892)[name = tensor("op_894")]; + tensor cross_7 = mul(x = var_891, y = var_894)[name = tensor("cross_7")]; tensor out_19 = add(x = inner_7, y = cross_7)[name = tensor("out_19")]; - tensor var_841 = mul(x = prev_kv_7, y = sqrt_s0_7)[name = tensor("op_841")]; - tensor var_843_transpose_x_1 = const()[name = tensor("op_843_transpose_x_1"), val = tensor(true)]; - tensor var_843_transpose_y_1 = const()[name = tensor("op_843_transpose_y_1"), val = tensor(false)]; - tensor var_843 = matmul(transpose_x = var_843_transpose_x_1, transpose_y = var_843_transpose_y_1, x = k_7, y = v_7)[name = tensor("op_843")]; - tensor new_kv_unnorm_7 = add(x = var_841, y = var_843)[name = tensor("new_kv_unnorm_7")]; - tensor var_845 = const()[name = tensor("op_845"), val = tensor(0x1p+2)]; - tensor new_scale_7 = add(x = prev_scale_7, y = var_845)[name = tensor("new_scale_7")]; - tensor var_847 = sqrt(x = new_scale_7)[name = tensor("op_847")]; - tensor nkv_1 = real_div(x = new_kv_unnorm_7, y = var_847)[name = tensor("nkv_1")]; - tensor var_849_perm_0 = const()[name = tensor("op_849_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_897 = mul(x = prev_kv_7, y = sqrt_s0_7)[name = tensor("op_897")]; + tensor var_899_transpose_x_1 = const()[name = tensor("op_899_transpose_x_1"), val = tensor(true)]; + tensor var_899_transpose_y_1 = const()[name = tensor("op_899_transpose_y_1"), val = tensor(false)]; + tensor var_899 = matmul(transpose_x = var_899_transpose_x_1, transpose_y = var_899_transpose_y_1, x = k_7, y = v_7)[name = tensor("op_899")]; + tensor new_kv_unnorm_7 = add(x = var_897, y = var_899)[name = tensor("new_kv_unnorm_7")]; + tensor var_901 = const()[name = tensor("op_901"), val = tensor(0x1p+2)]; + tensor new_scale_7 = add(x = prev_scale_7, y = var_901)[name = tensor("new_scale_7")]; + tensor var_903 = sqrt(x = new_scale_7)[name = tensor("op_903")]; + tensor nkv_1 = real_div(x = new_kv_unnorm_7, y = var_903)[name = tensor("nkv_1")]; + tensor var_905_perm_0 = const()[name = tensor("op_905_perm_0"), val = tensor([0, 2, 1, 3])]; tensor out_21_axes_0 = const()[name = tensor("out_21_axes_0"), val = tensor([-1])]; - tensor var_849 = transpose(perm = var_849_perm_0, x = out_19)[name = tensor("transpose_34")]; - tensor out_21 = layer_norm(axes = out_21_axes_0, epsilon = var_17, x = var_849)[name = tensor("out_21")]; - tensor var_853 = const()[name = tensor("op_853"), val = tensor([1, 4, 256])]; - tensor out_23 = reshape(shape = var_853, x = out_21)[name = tensor("out_23")]; - tensor var_855 = silu(x = input_137)[name = tensor("op_855")]; - tensor input_139 = mul(x = var_855, y = out_23)[name = tensor("input_139")]; - tensor ret_out_7 = linear(bias = encoder_out_proj_3_bias, weight = encoder_out_proj_3_weight, x = input_139)[name = tensor("linear_34")]; - tensor x_21 = add(x = input_133, y = ret_out_7)[name = tensor("x_21")]; + tensor var_905 = transpose(perm = var_905_perm_0, x = out_19)[name = tensor("transpose_34")]; + tensor out_21 = layer_norm(axes = out_21_axes_0, epsilon = var_74, x = var_905)[name = tensor("out_21")]; + tensor var_909 = const()[name = tensor("op_909"), val = tensor([1, 4, 256])]; + tensor out_23 = reshape(shape = var_909, x = out_21)[name = tensor("out_23")]; + tensor var_911 = silu(x = input_139)[name = tensor("op_911")]; + tensor input_141 = mul(x = var_911, y = out_23)[name = tensor("input_141")]; + tensor ret_out_7 = linear(bias = inner_encoder_out_proj_3_bias, weight = inner_encoder_out_proj_3_weight, x = input_141)[name = tensor("linear_34")]; + tensor x_21 = add(x = input_135, y = ret_out_7)[name = tensor("x_21")]; tensor window_31_begin_0 = const()[name = tensor("window_31_begin_0"), val = tensor([3, 0, 0, 0])]; tensor window_31_end_0 = const()[name = tensor("window_31_end_0"), val = tensor([4, 1, 16, 256])]; tensor window_31_end_mask_0 = const()[name = tensor("window_31_end_mask_0"), val = tensor([false, true, true, true])]; tensor window_31_squeeze_mask_0 = const()[name = tensor("window_31_squeeze_mask_0"), val = tensor([true, false, false, false])]; tensor window_31 = slice_by_index(begin = window_31_begin_0, end = window_31_end_0, end_mask = window_31_end_mask_0, squeeze_mask = window_31_squeeze_mask_0, x = enc_conv_cache)[name = tensor("window_31")]; - tensor var_863_begin_0 = const()[name = tensor("op_863_begin_0"), val = tensor([0, 0, 0])]; - tensor var_863_end_0 = const()[name = tensor("op_863_end_0"), val = tensor([1, 1, 256])]; - tensor var_863_end_mask_0 = const()[name = tensor("op_863_end_mask_0"), val = tensor([true, false, true])]; - tensor var_863 = slice_by_index(begin = var_863_begin_0, end = var_863_end_0, end_mask = var_863_end_mask_0, x = x_21)[name = tensor("op_863")]; - tensor var_866_begin_0 = const()[name = tensor("op_866_begin_0"), val = tensor([0, 1, 0])]; - tensor var_866_end_0 = const()[name = tensor("op_866_end_0"), val = tensor([1, 16, 256])]; - tensor var_866_end_mask_0 = const()[name = tensor("op_866_end_mask_0"), val = tensor([true, true, true])]; - tensor var_866 = slice_by_index(begin = var_866_begin_0, end = var_866_end_0, end_mask = var_866_end_mask_0, x = window_31)[name = tensor("op_866")]; + tensor var_919_begin_0 = const()[name = tensor("op_919_begin_0"), val = tensor([0, 0, 0])]; + tensor var_919_end_0 = const()[name = tensor("op_919_end_0"), val = tensor([1, 1, 256])]; + tensor var_919_end_mask_0 = const()[name = tensor("op_919_end_mask_0"), val = tensor([true, false, true])]; + tensor var_919 = slice_by_index(begin = var_919_begin_0, end = var_919_end_0, end_mask = var_919_end_mask_0, x = x_21)[name = tensor("op_919")]; + tensor var_922_begin_0 = const()[name = tensor("op_922_begin_0"), val = tensor([0, 1, 0])]; + tensor var_922_end_0 = const()[name = tensor("op_922_end_0"), val = tensor([1, 16, 256])]; + tensor var_922_end_mask_0 = const()[name = tensor("op_922_end_mask_0"), val = tensor([true, true, true])]; + tensor var_922 = slice_by_index(begin = var_922_begin_0, end = var_922_end_0, end_mask = var_922_end_mask_0, x = window_31)[name = tensor("op_922")]; tensor window_33_interleave_0 = const()[name = tensor("window_33_interleave_0"), val = tensor(false)]; - tensor window_33 = concat(axis = var_26, interleave = window_33_interleave_0, values = (var_866, var_863))[name = tensor("window_33")]; - tensor var_871_begin_0 = const()[name = tensor("op_871_begin_0"), val = tensor([0, 1, 0])]; - tensor var_871_end_0 = const()[name = tensor("op_871_end_0"), val = tensor([1, 2, 256])]; - tensor var_871_end_mask_0 = const()[name = tensor("op_871_end_mask_0"), val = tensor([true, false, true])]; - tensor var_871 = slice_by_index(begin = var_871_begin_0, end = var_871_end_0, end_mask = var_871_end_mask_0, x = x_21)[name = tensor("op_871")]; - tensor var_874_begin_0 = const()[name = tensor("op_874_begin_0"), val = tensor([0, 1, 0])]; - tensor var_874_end_0 = const()[name = tensor("op_874_end_0"), val = tensor([1, 16, 256])]; - tensor var_874_end_mask_0 = const()[name = tensor("op_874_end_mask_0"), val = tensor([true, true, true])]; - tensor var_874 = slice_by_index(begin = var_874_begin_0, end = var_874_end_0, end_mask = var_874_end_mask_0, x = window_33)[name = tensor("op_874")]; + tensor window_33 = concat(axis = var_82, interleave = window_33_interleave_0, values = (var_922, var_919))[name = tensor("window_33")]; + tensor var_927_begin_0 = const()[name = tensor("op_927_begin_0"), val = tensor([0, 1, 0])]; + tensor var_927_end_0 = const()[name = tensor("op_927_end_0"), val = tensor([1, 2, 256])]; + tensor var_927_end_mask_0 = const()[name = tensor("op_927_end_mask_0"), val = tensor([true, false, true])]; + tensor var_927 = slice_by_index(begin = var_927_begin_0, end = var_927_end_0, end_mask = var_927_end_mask_0, x = x_21)[name = tensor("op_927")]; + tensor var_930_begin_0 = const()[name = tensor("op_930_begin_0"), val = tensor([0, 1, 0])]; + tensor var_930_end_0 = const()[name = tensor("op_930_end_0"), val = tensor([1, 16, 256])]; + tensor var_930_end_mask_0 = const()[name = tensor("op_930_end_mask_0"), val = tensor([true, true, true])]; + tensor var_930 = slice_by_index(begin = var_930_begin_0, end = var_930_end_0, end_mask = var_930_end_mask_0, x = window_33)[name = tensor("op_930")]; tensor window_35_interleave_0 = const()[name = tensor("window_35_interleave_0"), val = tensor(false)]; - tensor window_35 = concat(axis = var_26, interleave = window_35_interleave_0, values = (var_874, var_871))[name = tensor("window_35")]; - tensor var_879_begin_0 = const()[name = tensor("op_879_begin_0"), val = tensor([0, 2, 0])]; - tensor var_879_end_0 = const()[name = tensor("op_879_end_0"), val = tensor([1, 3, 256])]; - tensor var_879_end_mask_0 = const()[name = tensor("op_879_end_mask_0"), val = tensor([true, false, true])]; - tensor var_879 = slice_by_index(begin = var_879_begin_0, end = var_879_end_0, end_mask = var_879_end_mask_0, x = x_21)[name = tensor("op_879")]; - tensor var_882_begin_0 = const()[name = tensor("op_882_begin_0"), val = tensor([0, 1, 0])]; - tensor var_882_end_0 = const()[name = tensor("op_882_end_0"), val = tensor([1, 16, 256])]; - tensor var_882_end_mask_0 = const()[name = tensor("op_882_end_mask_0"), val = tensor([true, true, true])]; - tensor var_882 = slice_by_index(begin = var_882_begin_0, end = var_882_end_0, end_mask = var_882_end_mask_0, x = window_35)[name = tensor("op_882")]; + tensor window_35 = concat(axis = var_82, interleave = window_35_interleave_0, values = (var_930, var_927))[name = tensor("window_35")]; + tensor var_935_begin_0 = const()[name = tensor("op_935_begin_0"), val = tensor([0, 2, 0])]; + tensor var_935_end_0 = const()[name = tensor("op_935_end_0"), val = tensor([1, 3, 256])]; + tensor var_935_end_mask_0 = const()[name = tensor("op_935_end_mask_0"), val = tensor([true, false, true])]; + tensor var_935 = slice_by_index(begin = var_935_begin_0, end = var_935_end_0, end_mask = var_935_end_mask_0, x = x_21)[name = tensor("op_935")]; + tensor var_938_begin_0 = const()[name = tensor("op_938_begin_0"), val = tensor([0, 1, 0])]; + tensor var_938_end_0 = const()[name = tensor("op_938_end_0"), val = tensor([1, 16, 256])]; + tensor var_938_end_mask_0 = const()[name = tensor("op_938_end_mask_0"), val = tensor([true, true, true])]; + tensor var_938 = slice_by_index(begin = var_938_begin_0, end = var_938_end_0, end_mask = var_938_end_mask_0, x = window_35)[name = tensor("op_938")]; tensor window_37_interleave_0 = const()[name = tensor("window_37_interleave_0"), val = tensor(false)]; - tensor window_37 = concat(axis = var_26, interleave = window_37_interleave_0, values = (var_882, var_879))[name = tensor("window_37")]; - tensor var_887_begin_0 = const()[name = tensor("op_887_begin_0"), val = tensor([0, 3, 0])]; - tensor var_887_end_0 = const()[name = tensor("op_887_end_0"), val = tensor([1, 1, 256])]; - tensor var_887_end_mask_0 = const()[name = tensor("op_887_end_mask_0"), val = tensor([true, true, true])]; - tensor var_887 = slice_by_index(begin = var_887_begin_0, end = var_887_end_0, end_mask = var_887_end_mask_0, x = x_21)[name = tensor("op_887")]; - tensor var_890_begin_0 = const()[name = tensor("op_890_begin_0"), val = tensor([0, 1, 0])]; - tensor var_890_end_0 = const()[name = tensor("op_890_end_0"), val = tensor([1, 16, 256])]; - tensor var_890_end_mask_0 = const()[name = tensor("op_890_end_mask_0"), val = tensor([true, true, true])]; - tensor var_890 = slice_by_index(begin = var_890_begin_0, end = var_890_end_0, end_mask = var_890_end_mask_0, x = window_37)[name = tensor("op_890")]; + tensor window_37 = concat(axis = var_82, interleave = window_37_interleave_0, values = (var_938, var_935))[name = tensor("window_37")]; + tensor var_943_begin_0 = const()[name = tensor("op_943_begin_0"), val = tensor([0, 3, 0])]; + tensor var_943_end_0 = const()[name = tensor("op_943_end_0"), val = tensor([1, 1, 256])]; + tensor var_943_end_mask_0 = const()[name = tensor("op_943_end_mask_0"), val = tensor([true, true, true])]; + tensor var_943 = slice_by_index(begin = var_943_begin_0, end = var_943_end_0, end_mask = var_943_end_mask_0, x = x_21)[name = tensor("op_943")]; + tensor var_946_begin_0 = const()[name = tensor("op_946_begin_0"), val = tensor([0, 1, 0])]; + tensor var_946_end_0 = const()[name = tensor("op_946_end_0"), val = tensor([1, 16, 256])]; + tensor var_946_end_mask_0 = const()[name = tensor("op_946_end_mask_0"), val = tensor([true, true, true])]; + tensor var_946 = slice_by_index(begin = var_946_begin_0, end = var_946_end_0, end_mask = var_946_end_mask_0, x = window_37)[name = tensor("op_946")]; tensor window_interleave_0 = const()[name = tensor("window_interleave_0"), val = tensor(false)]; - tensor window = concat(axis = var_26, interleave = window_interleave_0, values = (var_890, var_887))[name = tensor("window")]; - tensor input_141_interleave_0 = const()[name = tensor("input_141_interleave_0"), val = tensor(false)]; - tensor input_141 = concat(axis = var_23, interleave = input_141_interleave_0, values = (window_33, window_35, window_37, window))[name = tensor("input_141")]; + tensor window = concat(axis = var_82, interleave = window_interleave_0, values = (var_946, var_943))[name = tensor("window")]; + tensor input_143_interleave_0 = const()[name = tensor("input_143_interleave_0"), val = tensor(false)]; + tensor input_143 = concat(axis = var_69, interleave = input_143_interleave_0, values = (window_33, window_35, window_37, window))[name = tensor("input_143")]; tensor x_23_axes_0 = const()[name = tensor("x_23_axes_0"), val = tensor([-1])]; - tensor x_23 = layer_norm(axes = x_23_axes_0, beta = encoder_conv_module_3_sequential_0_bias, epsilon = var_28, gamma = encoder_conv_module_3_sequential_0_weight, x = input_141)[name = tensor("x_23")]; - tensor input_143_perm_0 = const()[name = tensor("input_143_perm_0"), val = tensor([0, 2, 1])]; + tensor x_23 = layer_norm(axes = x_23_axes_0, beta = inner_encoder_conv_module_3_sequential_0_bias, epsilon = var_66, gamma = inner_encoder_conv_module_3_sequential_0_weight, x = input_143)[name = tensor("x_23")]; + tensor input_145_perm_0 = const()[name = tensor("input_145_perm_0"), val = tensor([0, 2, 1])]; tensor inputs_33_pad_type_0 = const()[name = tensor("inputs_33_pad_type_0"), val = tensor("valid")]; tensor inputs_33_strides_0 = const()[name = tensor("inputs_33_strides_0"), val = tensor([1])]; tensor inputs_33_pad_0 = const()[name = tensor("inputs_33_pad_0"), val = tensor([0, 0])]; tensor inputs_33_dilations_0 = const()[name = tensor("inputs_33_dilations_0"), val = tensor([1])]; tensor inputs_33_groups_0 = const()[name = tensor("inputs_33_groups_0"), val = tensor(1)]; - tensor input_143 = transpose(perm = input_143_perm_0, x = x_23)[name = tensor("transpose_33")]; - tensor inputs_33 = conv(bias = encoder_conv_module_3_sequential_2_conv_bias, dilations = inputs_33_dilations_0, groups = inputs_33_groups_0, pad = inputs_33_pad_0, pad_type = inputs_33_pad_type_0, strides = inputs_33_strides_0, weight = encoder_conv_module_3_sequential_2_conv_weight, x = input_143)[name = tensor("inputs_33")]; - tensor var_915_split_sizes_0 = const()[name = tensor("op_915_split_sizes_0"), val = tensor([256, 256])]; - tensor var_915_axis_0 = const()[name = tensor("op_915_axis_0"), val = tensor(1)]; - tensor var_915_0, tensor var_915_1 = split(axis = var_915_axis_0, split_sizes = var_915_split_sizes_0, x = inputs_33)[name = tensor("op_915")]; - tensor var_917 = sigmoid(x = var_915_1)[name = tensor("op_917")]; - tensor inputs_35 = mul(x = var_915_0, y = var_917)[name = tensor("inputs_35")]; + tensor input_145 = transpose(perm = input_145_perm_0, x = x_23)[name = tensor("transpose_33")]; + tensor inputs_33 = conv(bias = inner_encoder_conv_module_3_sequential_2_conv_bias, dilations = inputs_33_dilations_0, groups = inputs_33_groups_0, pad = inputs_33_pad_0, pad_type = inputs_33_pad_type_0, strides = inputs_33_strides_0, weight = inner_encoder_conv_module_3_sequential_2_conv_weight, x = input_145)[name = tensor("inputs_33")]; + tensor var_971_split_sizes_0 = const()[name = tensor("op_971_split_sizes_0"), val = tensor([256, 256])]; + tensor var_971_axis_0 = const()[name = tensor("op_971_axis_0"), val = tensor(1)]; + tensor var_971_0, tensor var_971_1 = split(axis = var_971_axis_0, split_sizes = var_971_split_sizes_0, x = inputs_33)[name = tensor("op_971")]; + tensor var_973 = sigmoid(x = var_971_1)[name = tensor("op_973")]; + tensor inputs_35 = mul(x = var_971_0, y = var_973)[name = tensor("inputs_35")]; tensor outputs_aug_pad_type_0 = const()[name = tensor("outputs_aug_pad_type_0"), val = tensor("custom")]; tensor outputs_aug_pad_0 = const()[name = tensor("outputs_aug_pad_0"), val = tensor([15, 15])]; tensor outputs_aug_groups_0 = const()[name = tensor("outputs_aug_groups_0"), val = tensor(256)]; tensor outputs_aug_strides_0 = const()[name = tensor("outputs_aug_strides_0"), val = tensor([1])]; tensor outputs_aug_dilations_0 = const()[name = tensor("outputs_aug_dilations_0"), val = tensor([1])]; - tensor outputs_aug = conv(dilations = outputs_aug_dilations_0, groups = outputs_aug_groups_0, pad = outputs_aug_pad_0, pad_type = outputs_aug_pad_type_0, strides = outputs_aug_strides_0, weight = encoder_conv_module_3_sequential_4_conv_weight, x = inputs_35)[name = tensor("outputs_aug")]; - tensor input_145_begin_0 = const()[name = tensor("input_145_begin_0"), val = tensor([0, 0, 0])]; - tensor input_145_end_0 = const()[name = tensor("input_145_end_0"), val = tensor([4, 256, 16])]; - tensor input_145_end_mask_0 = const()[name = tensor("input_145_end_mask_0"), val = tensor([true, true, false])]; - tensor input_145 = slice_by_index(begin = input_145_begin_0, end = input_145_end_0, end_mask = input_145_end_mask_0, x = outputs_aug)[name = tensor("input_145")]; - tensor inputs_37 = batch_norm(beta = encoder_conv_module_3_sequential_5_bias, epsilon = var_28, gamma = encoder_conv_module_3_sequential_5_weight, mean = encoder_conv_module_3_sequential_5_running_mean, variance = encoder_conv_module_3_sequential_5_running_var, x = input_145)[name = tensor("inputs_37")]; - tensor input_147 = silu(x = inputs_37)[name = tensor("input_147")]; - tensor input_149_pad_type_0 = const()[name = tensor("input_149_pad_type_0"), val = tensor("valid")]; - tensor input_149_strides_0 = const()[name = tensor("input_149_strides_0"), val = tensor([1])]; - tensor input_149_pad_0 = const()[name = tensor("input_149_pad_0"), val = tensor([0, 0])]; - tensor input_149_dilations_0 = const()[name = tensor("input_149_dilations_0"), val = tensor([1])]; - tensor input_149_groups_0 = const()[name = tensor("input_149_groups_0"), val = tensor(1)]; - tensor input_149 = conv(bias = encoder_conv_module_3_sequential_7_conv_bias, dilations = input_149_dilations_0, groups = input_149_groups_0, pad = input_149_pad_0, pad_type = input_149_pad_type_0, strides = input_149_strides_0, weight = encoder_conv_module_3_sequential_7_conv_weight, x = input_147)[name = tensor("input_149")]; + tensor outputs_aug = conv(dilations = outputs_aug_dilations_0, groups = outputs_aug_groups_0, pad = outputs_aug_pad_0, pad_type = outputs_aug_pad_type_0, strides = outputs_aug_strides_0, weight = inner_encoder_conv_module_3_sequential_4_conv_weight, x = inputs_35)[name = tensor("outputs_aug")]; + tensor input_147_begin_0 = const()[name = tensor("input_147_begin_0"), val = tensor([0, 0, 0])]; + tensor input_147_end_0 = const()[name = tensor("input_147_end_0"), val = tensor([4, 256, 16])]; + tensor input_147_end_mask_0 = const()[name = tensor("input_147_end_mask_0"), val = tensor([true, true, false])]; + tensor input_147 = slice_by_index(begin = input_147_begin_0, end = input_147_end_0, end_mask = input_147_end_mask_0, x = outputs_aug)[name = tensor("input_147")]; + tensor inputs_37 = batch_norm(beta = inner_encoder_conv_module_3_sequential_5_bias, epsilon = var_66, gamma = inner_encoder_conv_module_3_sequential_5_weight, mean = inner_encoder_conv_module_3_sequential_5_running_mean, variance = inner_encoder_conv_module_3_sequential_5_running_var, x = input_147)[name = tensor("inputs_37")]; + tensor input_149 = silu(x = inputs_37)[name = tensor("input_149")]; + tensor input_151_pad_type_0 = const()[name = tensor("input_151_pad_type_0"), val = tensor("valid")]; + tensor input_151_strides_0 = const()[name = tensor("input_151_strides_0"), val = tensor([1])]; + tensor input_151_pad_0 = const()[name = tensor("input_151_pad_0"), val = tensor([0, 0])]; + tensor input_151_dilations_0 = const()[name = tensor("input_151_dilations_0"), val = tensor([1])]; + tensor input_151_groups_0 = const()[name = tensor("input_151_groups_0"), val = tensor(1)]; + tensor input_151 = conv(bias = inner_encoder_conv_module_3_sequential_7_conv_bias, dilations = input_151_dilations_0, groups = input_151_groups_0, pad = input_151_pad_0, pad_type = input_151_pad_type_0, strides = input_151_strides_0, weight = inner_encoder_conv_module_3_sequential_7_conv_weight, x = input_149)[name = tensor("input_151")]; tensor conv_out_7_perm_0 = const()[name = tensor("conv_out_7_perm_0"), val = tensor([0, 2, 1])]; - tensor var_948_begin_0 = const()[name = tensor("op_948_begin_0"), val = tensor([0, -1, 0])]; - tensor var_948_end_0 = const()[name = tensor("op_948_end_0"), val = tensor([4, 16, 256])]; - tensor var_948_end_mask_0 = const()[name = tensor("op_948_end_mask_0"), val = tensor([true, true, true])]; - tensor conv_out_7 = transpose(perm = conv_out_7_perm_0, x = input_149)[name = tensor("transpose_32")]; - tensor var_948 = slice_by_index(begin = var_948_begin_0, end = var_948_end_0, end_mask = var_948_end_mask_0, x = conv_out_7)[name = tensor("op_948")]; - tensor var_950_perm_0 = const()[name = tensor("op_950_perm_0"), val = tensor([1, 0, 2])]; - tensor var_950 = transpose(perm = var_950_perm_0, x = var_948)[name = tensor("transpose_31")]; - tensor input_151 = add(x = x_21, y = var_950)[name = tensor("input_151")]; - tensor input_153_axes_0 = const()[name = tensor("input_153_axes_0"), val = tensor([-1])]; - tensor input_153 = layer_norm(axes = input_153_axes_0, beta = encoder_ffn2_3_module_sequential_0_bias, epsilon = var_28, gamma = encoder_ffn2_3_module_sequential_0_weight, x = input_151)[name = tensor("input_153")]; - tensor inputs = linear(bias = encoder_ffn2_3_module_sequential_1_linear_bias, weight = encoder_ffn2_3_module_sequential_1_linear_weight, x = input_153)[name = tensor("linear_35")]; - tensor input_155 = silu(x = inputs)[name = tensor("input_155")]; - tensor input_159 = linear(bias = encoder_ffn2_3_module_sequential_4_linear_bias, weight = encoder_ffn2_3_module_sequential_4_linear_weight, x = input_155)[name = tensor("linear_36")]; - tensor var_973 = const()[name = tensor("op_973"), val = tensor(0x1p-1)]; - tensor var_974 = mul(x = input_159, y = var_973)[name = tensor("op_974")]; - tensor input_161 = add(x = var_974, y = input_151)[name = tensor("input_161")]; + tensor var_1004_begin_0 = const()[name = tensor("op_1004_begin_0"), val = tensor([0, -1, 0])]; + tensor var_1004_end_0 = const()[name = tensor("op_1004_end_0"), val = tensor([4, 16, 256])]; + tensor var_1004_end_mask_0 = const()[name = tensor("op_1004_end_mask_0"), val = tensor([true, true, true])]; + tensor conv_out_7 = transpose(perm = conv_out_7_perm_0, x = input_151)[name = tensor("transpose_32")]; + tensor var_1004 = slice_by_index(begin = var_1004_begin_0, end = var_1004_end_0, end_mask = var_1004_end_mask_0, x = conv_out_7)[name = tensor("op_1004")]; + tensor var_1006_perm_0 = const()[name = tensor("op_1006_perm_0"), val = tensor([1, 0, 2])]; + tensor var_1006 = transpose(perm = var_1006_perm_0, x = var_1004)[name = tensor("transpose_31")]; + tensor input_153 = add(x = x_21, y = var_1006)[name = tensor("input_153")]; + tensor input_155_axes_0 = const()[name = tensor("input_155_axes_0"), val = tensor([-1])]; + tensor input_155 = layer_norm(axes = input_155_axes_0, beta = inner_encoder_ffn2_3_module_sequential_0_bias, epsilon = var_66, gamma = inner_encoder_ffn2_3_module_sequential_0_weight, x = input_153)[name = tensor("input_155")]; + tensor inputs = linear(bias = inner_encoder_ffn2_3_module_sequential_1_linear_bias, weight = inner_encoder_ffn2_3_module_sequential_1_linear_weight, x = input_155)[name = tensor("linear_35")]; + tensor input_157 = silu(x = inputs)[name = tensor("input_157")]; + tensor input_161 = linear(bias = inner_encoder_ffn2_3_module_sequential_4_linear_bias, weight = inner_encoder_ffn2_3_module_sequential_4_linear_weight, x = input_157)[name = tensor("linear_36")]; + tensor var_1029 = const()[name = tensor("op_1029"), val = tensor(0x1p-1)]; + tensor var_1030 = mul(x = input_161, y = var_1029)[name = tensor("op_1030")]; + tensor input_163 = add(x = var_1030, y = input_153)[name = tensor("input_163")]; tensor x_25_axes_0 = const()[name = tensor("x_25_axes_0"), val = tensor([-1])]; - tensor x_25 = layer_norm(axes = x_25_axes_0, beta = encoder_layer_norm_3_bias, epsilon = var_28, gamma = encoder_layer_norm_3_weight, x = input_161)[name = tensor("x_25")]; + tensor x_25 = layer_norm(axes = x_25_axes_0, beta = inner_encoder_layer_norm_3_bias, epsilon = var_66, gamma = inner_encoder_layer_norm_3_weight, x = input_163)[name = tensor("x_25")]; tensor x_ct_perm_0 = const()[name = tensor("x_ct_perm_0"), val = tensor([0, 2, 1])]; tensor cat_interleave_0 = const()[name = tensor("cat_interleave_0"), val = tensor(false)]; tensor x_ct = transpose(perm = x_ct_perm_0, x = x_25)[name = tensor("transpose_30")]; - tensor cat = concat(axis = var_20, interleave = cat_interleave_0, values = (cnn_window, x_ct))[name = tensor("cat")]; + tensor cat = concat(axis = var_71, interleave = cat_interleave_0, values = (cnn_window, x_ct))[name = tensor("cat")]; tensor conv_out_pad_type_0 = const()[name = tensor("conv_out_pad_type_0"), val = tensor("valid")]; tensor conv_out_strides_0 = const()[name = tensor("conv_out_strides_0"), val = tensor([1])]; tensor conv_out_pad_0 = const()[name = tensor("conv_out_pad_0"), val = tensor([0, 0])]; tensor conv_out_dilations_0 = const()[name = tensor("conv_out_dilations_0"), val = tensor([1])]; tensor conv_out_groups_0 = const()[name = tensor("conv_out_groups_0"), val = tensor(1)]; - tensor conv_out = conv(bias = encoder_cnn_bias, dilations = conv_out_dilations_0, groups = conv_out_groups_0, pad = conv_out_pad_0, pad_type = conv_out_pad_type_0, strides = conv_out_strides_0, weight = encoder_cnn_weight, x = cat)[name = tensor("conv_out")]; - tensor var_992_begin_0 = const()[name = tensor("op_992_begin_0"), val = tensor([0, 0, 4])]; - tensor var_992_end_0 = const()[name = tensor("op_992_end_0"), val = tensor([1, 256, 22])]; - tensor var_992_end_mask_0 = const()[name = tensor("op_992_end_mask_0"), val = tensor([true, true, true])]; - tensor cnn_window_new = slice_by_index(begin = var_992_begin_0, end = var_992_end_0, end_mask = var_992_end_mask_0, x = cat)[name = tensor("op_992")]; - tensor input_163_perm_0 = const()[name = tensor("input_163_perm_0"), val = tensor([0, 2, 1])]; - tensor var_994 = const()[name = tensor("op_994"), val = tensor([-1])]; - tensor input_163 = transpose(perm = input_163_perm_0, x = conv_out)[name = tensor("transpose_29")]; - tensor var_995 = reduce_l2_norm(axes = var_994, keep_dims = var_29, x = input_163)[name = tensor("op_995")]; + tensor conv_out = conv(bias = inner_encoder_cnn_bias, dilations = conv_out_dilations_0, groups = conv_out_groups_0, pad = conv_out_pad_0, pad_type = conv_out_pad_type_0, strides = conv_out_strides_0, weight = inner_encoder_cnn_weight, x = cat)[name = tensor("conv_out")]; + tensor var_1048_begin_0 = const()[name = tensor("op_1048_begin_0"), val = tensor([0, 0, 4])]; + tensor var_1048_end_0 = const()[name = tensor("op_1048_end_0"), val = tensor([1, 256, 22])]; + tensor var_1048_end_mask_0 = const()[name = tensor("op_1048_end_mask_0"), val = tensor([true, true, true])]; + tensor cnn_window_new = slice_by_index(begin = var_1048_begin_0, end = var_1048_end_0, end_mask = var_1048_end_mask_0, x = cat)[name = tensor("op_1048")]; + tensor input_165_perm_0 = const()[name = tensor("input_165_perm_0"), val = tensor([0, 2, 1])]; + tensor var_1050 = const()[name = tensor("op_1050"), val = tensor([-1])]; + tensor input_165 = transpose(perm = input_165_perm_0, x = conv_out)[name = tensor("transpose_29")]; + tensor var_1051 = reduce_l2_norm(axes = var_1050, keep_dims = var_65, x = input_165)[name = tensor("op_1051")]; tensor const_12 = const()[name = tensor("const_12"), val = tensor(0x1.fffffep+127)]; - tensor clip_0 = clip(alpha = var_11, beta = const_12, x = var_995)[name = tensor("clip_0")]; - tensor emb = real_div(x = input_163, y = clip_0)[name = tensor("emb")]; - tensor var_999_axis_0 = const()[name = tensor("op_999_axis_0"), val = tensor(0)]; - tensor enc_kv_new = stack(axis = var_999_axis_0, values = (var_206, var_420, var_634, nkv_1))[name = tensor("op_999")]; - tensor var_1001_axis_0 = const()[name = tensor("op_1001_axis_0"), val = tensor(0)]; - tensor enc_scale_new = stack(axis = var_1001_axis_0, values = (new_scale_1, new_scale_3, new_scale_5, new_scale_7))[name = tensor("op_1001")]; - tensor var_1003_axis_0 = const()[name = tensor("op_1003_axis_0"), val = tensor(0)]; - tensor enc_conv_cache_new = stack(axis = var_1003_axis_0, values = (window_9, window_19, window_29, window))[name = tensor("op_1003")]; - tensor var_1012 = const()[name = tensor("op_1012"), val = tensor(0x1.5798eep-27)]; - tensor var_1017 = const()[name = tensor("op_1017"), val = tensor(0x1.4f8b58p-17)]; - tensor var_1019 = const()[name = tensor("op_1019"), val = tensor(0x1.0c6f7ap-20)]; - tensor var_1020 = const()[name = tensor("op_1020"), val = tensor(true)]; - tensor var_1022 = const()[name = tensor("op_1022"), val = tensor(0x1p+0)]; - tensor var_1026 = const()[name = tensor("op_1026"), val = tensor(-1)]; - tensor var_1032 = const()[name = tensor("op_1032"), val = tensor(0)]; + tensor clip_0 = clip(alpha = var_79, beta = const_12, x = var_1051)[name = tensor("clip_0")]; + tensor emb = real_div(x = input_165, y = clip_0)[name = tensor("emb")]; + tensor var_1055_axis_0 = const()[name = tensor("op_1055_axis_0"), val = tensor(0)]; + tensor enc_kv_new = stack(axis = var_1055_axis_0, values = (var_262, var_476, var_690, nkv_1))[name = tensor("op_1055")]; + tensor var_1057_axis_0 = const()[name = tensor("op_1057_axis_0"), val = tensor(0)]; + tensor enc_scale_new = stack(axis = var_1057_axis_0, values = (new_scale_1, new_scale_3, new_scale_5, new_scale_7))[name = tensor("op_1057")]; + tensor var_1059_axis_0 = const()[name = tensor("op_1059_axis_0"), val = tensor(0)]; + tensor enc_conv_cache_new = stack(axis = var_1059_axis_0, values = (window_9, window_19, window_29, window))[name = tensor("op_1059")]; tensor pos = const()[name = tensor("pos"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(44395584)))]; - tensor var_1094_axes_0 = const()[name = tensor("op_1094_axes_0"), val = tensor([2])]; - tensor var_1094 = expand_dims(axes = var_1094_axes_0, x = emb)[name = tensor("op_1094")]; + tensor var_1127_axes_0 = const()[name = tensor("op_1127_axes_0"), val = tensor([2])]; + tensor var_1127 = expand_dims(axes = var_1127_axes_0, x = emb)[name = tensor("op_1127")]; tensor emb_exp_reps_0 = const()[name = tensor("emb_exp_reps_0"), val = tensor([1, 1, 9, 1])]; - tensor emb_exp = tile(reps = emb_exp_reps_0, x = var_1094)[name = tensor("emb_exp")]; - tensor input_165_interleave_0 = const()[name = tensor("input_165_interleave_0"), val = tensor(false)]; - tensor input_165 = concat(axis = var_1026, interleave = input_165_interleave_0, values = (emb_exp, pos))[name = tensor("input_165")]; - tensor x_27 = linear(bias = decoder_convert_bias, weight = decoder_convert_weight, x = input_165)[name = tensor("linear_37")]; - tensor var_1102_perm_0 = const()[name = tensor("op_1102_perm_0"), val = tensor([0, 2, 1, 3])]; - tensor var_1106 = const()[name = tensor("op_1106"), val = tensor([9, 4, 256])]; - tensor var_1102 = transpose(perm = var_1102_perm_0, x = x_27)[name = tensor("transpose_28")]; - tensor x_29 = reshape(shape = var_1106, x = var_1102)[name = tensor("x_29")]; + tensor emb_exp = tile(reps = emb_exp_reps_0, x = var_1127)[name = tensor("emb_exp")]; + tensor input_167_interleave_0 = const()[name = tensor("input_167_interleave_0"), val = tensor(false)]; + tensor input_167 = concat(axis = var_72, interleave = input_167_interleave_0, values = (emb_exp, pos))[name = tensor("input_167")]; + tensor x_27 = linear(bias = inner_decoder_convert_bias, weight = inner_decoder_convert_weight, x = input_167)[name = tensor("linear_37")]; + tensor var_1135_perm_0 = const()[name = tensor("op_1135_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_1139 = const()[name = tensor("op_1139"), val = tensor([9, 4, 256])]; + tensor var_1135 = transpose(perm = var_1135_perm_0, x = x_27)[name = tensor("transpose_28")]; + tensor x_29 = reshape(shape = var_1139, x = var_1135)[name = tensor("x_29")]; tensor prev_kv_9_begin_0 = const()[name = tensor("prev_kv_9_begin_0"), val = tensor([0, 0, 0, 0, 0])]; tensor prev_kv_9_end_0 = const()[name = tensor("prev_kv_9_end_0"), val = tensor([1, 9, 4, 64, 64])]; tensor prev_kv_9_end_mask_0 = const()[name = tensor("prev_kv_9_end_mask_0"), val = tensor([false, true, true, true, true])]; @@ -983,132 +998,132 @@ program(1.0) tensor prev_scale_9_end_mask_0 = const()[name = tensor("prev_scale_9_end_mask_0"), val = tensor([false, true])]; tensor prev_scale_9_squeeze_mask_0 = const()[name = tensor("prev_scale_9_squeeze_mask_0"), val = tensor([true, false])]; tensor prev_scale_9 = slice_by_index(begin = prev_scale_9_begin_0, end = prev_scale_9_end_0, end_mask = prev_scale_9_end_mask_0, squeeze_mask = prev_scale_9_squeeze_mask_0, x = dec_scale)[name = tensor("prev_scale_9")]; - tensor var_1114 = linear(bias = decoder_q_proj_0_bias, weight = decoder_q_proj_0_weight, x = x_29)[name = tensor("linear_38")]; - tensor var_1115 = const()[name = tensor("op_1115"), val = tensor([9, 4, 4, 64])]; - tensor var_1116 = reshape(shape = var_1115, x = var_1114)[name = tensor("op_1116")]; + tensor var_1147 = linear(bias = inner_decoder_q_proj_0_bias, weight = inner_decoder_q_proj_0_weight, x = x_29)[name = tensor("linear_38")]; + tensor var_1148 = const()[name = tensor("op_1148"), val = tensor([9, 4, 4, 64])]; + tensor var_1149 = reshape(shape = var_1148, x = var_1147)[name = tensor("op_1149")]; tensor q_9_perm_0 = const()[name = tensor("q_9_perm_0"), val = tensor([0, 2, 1, 3])]; - tensor var_1120 = linear(bias = decoder_k_proj_0_bias, weight = decoder_k_proj_0_weight, x = x_29)[name = tensor("linear_39")]; - tensor var_1121 = const()[name = tensor("op_1121"), val = tensor(0x1p-3)]; - tensor var_1122 = mul(x = var_1120, y = var_1121)[name = tensor("op_1122")]; - tensor var_1123 = const()[name = tensor("op_1123"), val = tensor([9, 4, 4, 64])]; - tensor var_1124 = reshape(shape = var_1123, x = var_1122)[name = tensor("op_1124")]; + tensor var_1153 = linear(bias = inner_decoder_k_proj_0_bias, weight = inner_decoder_k_proj_0_weight, x = x_29)[name = tensor("linear_39")]; + tensor var_1154 = const()[name = tensor("op_1154"), val = tensor(0x1p-3)]; + tensor var_1155 = mul(x = var_1153, y = var_1154)[name = tensor("op_1155")]; + tensor var_1156 = const()[name = tensor("op_1156"), val = tensor([9, 4, 4, 64])]; + tensor var_1157 = reshape(shape = var_1156, x = var_1155)[name = tensor("op_1157")]; tensor k_9_perm_0 = const()[name = tensor("k_9_perm_0"), val = tensor([0, 2, 1, 3])]; - tensor var_1128 = linear(bias = decoder_v_proj_0_bias, weight = decoder_v_proj_0_weight, x = x_29)[name = tensor("linear_40")]; - tensor var_1129 = const()[name = tensor("op_1129"), val = tensor([9, 4, 4, 64])]; - tensor var_1130 = reshape(shape = var_1129, x = var_1128)[name = tensor("op_1130")]; + tensor var_1161 = linear(bias = inner_decoder_v_proj_0_bias, weight = inner_decoder_v_proj_0_weight, x = x_29)[name = tensor("linear_40")]; + tensor var_1162 = const()[name = tensor("op_1162"), val = tensor([9, 4, 4, 64])]; + tensor var_1163 = reshape(shape = var_1162, x = var_1161)[name = tensor("op_1163")]; tensor v_9_perm_0 = const()[name = tensor("v_9_perm_0"), val = tensor([0, 2, 1, 3])]; - tensor input_169 = linear(bias = decoder_g_proj_0_bias, weight = decoder_g_proj_0_weight, x = x_29)[name = tensor("linear_41")]; + tensor input_171 = linear(bias = inner_decoder_g_proj_0_bias, weight = inner_decoder_g_proj_0_weight, x = x_29)[name = tensor("linear_41")]; tensor cumsum_mask_1_exclusive_0 = const()[name = tensor("cumsum_mask_1_exclusive_0"), val = tensor(false)]; tensor cumsum_mask_1_reverse_0 = const()[name = tensor("cumsum_mask_1_reverse_0"), val = tensor(false)]; - tensor cumsum_mask_1 = cumsum(axis = var_1032, exclusive = cumsum_mask_1_exclusive_0, reverse = cumsum_mask_1_reverse_0, x = valid_mask)[name = tensor("cumsum_mask_1")]; + tensor cumsum_mask_1 = cumsum(axis = var_69, exclusive = cumsum_mask_1_exclusive_0, reverse = cumsum_mask_1_reverse_0, x = valid_mask)[name = tensor("cumsum_mask_1")]; tensor sqrt_s0_9 = sqrt(x = prev_scale_9)[name = tensor("sqrt_s0_9")]; tensor s_t_9 = add(x = prev_scale_9, y = cumsum_mask_1)[name = tensor("s_t_9")]; tensor const_20 = const()[name = tensor("const_20"), val = tensor(0x1.fffffep+127)]; - tensor clip_1 = clip(alpha = var_1022, beta = const_20, x = s_t_9)[name = tensor("clip_1")]; + tensor clip_1 = clip(alpha = var_59, beta = const_20, x = s_t_9)[name = tensor("clip_1")]; tensor sqrt_s_t_9 = sqrt(x = clip_1)[name = tensor("sqrt_s_t_9")]; tensor qk_9_transpose_x_1 = const()[name = tensor("qk_9_transpose_x_1"), val = tensor(false)]; tensor qk_9_transpose_y_1 = const()[name = tensor("qk_9_transpose_y_1"), val = tensor(true)]; - tensor k_9 = transpose(perm = k_9_perm_0, x = var_1124)[name = tensor("transpose_26")]; - tensor q_9 = transpose(perm = q_9_perm_0, x = var_1116)[name = tensor("transpose_27")]; + tensor k_9 = transpose(perm = k_9_perm_0, x = var_1157)[name = tensor("transpose_26")]; + tensor q_9 = transpose(perm = q_9_perm_0, x = var_1149)[name = tensor("transpose_27")]; tensor qk_9 = matmul(transpose_x = qk_9_transpose_x_1, transpose_y = qk_9_transpose_y_1, x = q_9, y = k_9)[name = tensor("qk_9")]; - tensor var_1142 = const()[name = tensor("op_1142"), val = tensor([1, 4])]; - tensor var_1143 = reshape(shape = var_1142, x = valid_mask)[name = tensor("op_1143")]; - tensor causal_with_valid_1 = mul(x = decoder__causal_mask, y = var_1143)[name = tensor("causal_with_valid_1")]; - tensor var_1145 = const()[name = tensor("op_1145"), val = tensor([4, 1])]; - tensor var_1146 = reshape(shape = var_1145, x = sqrt_s_t_9)[name = tensor("op_1146")]; - tensor M_9 = real_div(x = causal_with_valid_1, y = var_1146)[name = tensor("M_9")]; - tensor var_1148 = mul(x = qk_9, y = M_9)[name = tensor("op_1148")]; + tensor var_1175 = const()[name = tensor("op_1175"), val = tensor([1, 4])]; + tensor var_1176 = reshape(shape = var_1175, x = valid_mask)[name = tensor("op_1176")]; + tensor causal_with_valid_1 = mul(x = inner_decoder__causal_mask, y = var_1176)[name = tensor("causal_with_valid_1")]; + tensor var_1178 = const()[name = tensor("op_1178"), val = tensor([4, 1])]; + tensor var_1179 = reshape(shape = var_1178, x = sqrt_s_t_9)[name = tensor("op_1179")]; + tensor M_9 = real_div(x = causal_with_valid_1, y = var_1179)[name = tensor("M_9")]; + tensor var_1181 = mul(x = qk_9, y = M_9)[name = tensor("op_1181")]; tensor inner_9_transpose_x_0 = const()[name = tensor("inner_9_transpose_x_0"), val = tensor(false)]; tensor inner_9_transpose_y_0 = const()[name = tensor("inner_9_transpose_y_0"), val = tensor(false)]; - tensor v_9 = transpose(perm = v_9_perm_0, x = var_1130)[name = tensor("transpose_25")]; - tensor inner_9 = matmul(transpose_x = inner_9_transpose_x_0, transpose_y = inner_9_transpose_y_0, x = var_1148, y = v_9)[name = tensor("inner_9")]; - tensor var_1150_transpose_x_0 = const()[name = tensor("op_1150_transpose_x_0"), val = tensor(false)]; - tensor var_1150_transpose_y_0 = const()[name = tensor("op_1150_transpose_y_0"), val = tensor(false)]; - tensor var_1150 = matmul(transpose_x = var_1150_transpose_x_0, transpose_y = var_1150_transpose_y_0, x = q_9, y = prev_kv_9)[name = tensor("op_1150")]; - tensor var_1151 = real_div(x = sqrt_s0_9, y = sqrt_s_t_9)[name = tensor("op_1151")]; - tensor var_1152 = const()[name = tensor("op_1152"), val = tensor([1, 1, 4, 1])]; - tensor var_1153 = reshape(shape = var_1152, x = var_1151)[name = tensor("op_1153")]; - tensor cross_9 = mul(x = var_1150, y = var_1153)[name = tensor("cross_9")]; + tensor v_9 = transpose(perm = v_9_perm_0, x = var_1163)[name = tensor("transpose_25")]; + tensor inner_9 = matmul(transpose_x = inner_9_transpose_x_0, transpose_y = inner_9_transpose_y_0, x = var_1181, y = v_9)[name = tensor("inner_9")]; + tensor var_1183_transpose_x_0 = const()[name = tensor("op_1183_transpose_x_0"), val = tensor(false)]; + tensor var_1183_transpose_y_0 = const()[name = tensor("op_1183_transpose_y_0"), val = tensor(false)]; + tensor var_1183 = matmul(transpose_x = var_1183_transpose_x_0, transpose_y = var_1183_transpose_y_0, x = q_9, y = prev_kv_9)[name = tensor("op_1183")]; + tensor var_1184 = real_div(x = sqrt_s0_9, y = sqrt_s_t_9)[name = tensor("op_1184")]; + tensor var_1185 = const()[name = tensor("op_1185"), val = tensor([1, 1, 4, 1])]; + tensor var_1186 = reshape(shape = var_1185, x = var_1184)[name = tensor("op_1186")]; + tensor cross_9 = mul(x = var_1183, y = var_1186)[name = tensor("cross_9")]; tensor out_25 = add(x = inner_9, y = cross_9)[name = tensor("out_25")]; - tensor var_1156 = const()[name = tensor("op_1156"), val = tensor([1, 1, 4, 1])]; - tensor var_1157 = reshape(shape = var_1156, x = valid_mask)[name = tensor("op_1157")]; - tensor v_masked_1 = mul(x = v_9, y = var_1157)[name = tensor("v_masked_1")]; - tensor var_1159 = mul(x = prev_kv_9, y = sqrt_s0_9)[name = tensor("op_1159")]; - tensor var_1161_transpose_x_1 = const()[name = tensor("op_1161_transpose_x_1"), val = tensor(true)]; - tensor var_1161_transpose_y_1 = const()[name = tensor("op_1161_transpose_y_1"), val = tensor(false)]; - tensor var_1161 = matmul(transpose_x = var_1161_transpose_x_1, transpose_y = var_1161_transpose_y_1, x = k_9, y = v_masked_1)[name = tensor("op_1161")]; - tensor new_kv_unnorm_9 = add(x = var_1159, y = var_1161)[name = tensor("new_kv_unnorm_9")]; - tensor var_1163_keep_dims_0 = const()[name = tensor("op_1163_keep_dims_0"), val = tensor(false)]; - tensor var_1163 = reduce_sum(keep_dims = var_1163_keep_dims_0, x = valid_mask)[name = tensor("op_1163")]; - tensor var_1164 = const()[name = tensor("op_1164"), val = tensor([1])]; - tensor var_1165 = reshape(shape = var_1164, x = var_1163)[name = tensor("op_1165")]; - tensor new_scale_9 = add(x = prev_scale_9, y = var_1165)[name = tensor("new_scale_9")]; + tensor var_1189 = const()[name = tensor("op_1189"), val = tensor([1, 1, 4, 1])]; + tensor var_1190 = reshape(shape = var_1189, x = valid_mask)[name = tensor("op_1190")]; + tensor v_masked_1 = mul(x = v_9, y = var_1190)[name = tensor("v_masked_1")]; + tensor var_1192 = mul(x = prev_kv_9, y = sqrt_s0_9)[name = tensor("op_1192")]; + tensor var_1194_transpose_x_1 = const()[name = tensor("op_1194_transpose_x_1"), val = tensor(true)]; + tensor var_1194_transpose_y_1 = const()[name = tensor("op_1194_transpose_y_1"), val = tensor(false)]; + tensor var_1194 = matmul(transpose_x = var_1194_transpose_x_1, transpose_y = var_1194_transpose_y_1, x = k_9, y = v_masked_1)[name = tensor("op_1194")]; + tensor new_kv_unnorm_9 = add(x = var_1192, y = var_1194)[name = tensor("new_kv_unnorm_9")]; + tensor var_1196_keep_dims_0 = const()[name = tensor("op_1196_keep_dims_0"), val = tensor(false)]; + tensor var_1196 = reduce_sum(keep_dims = var_1196_keep_dims_0, x = valid_mask)[name = tensor("op_1196")]; + tensor var_1197 = const()[name = tensor("op_1197"), val = tensor([1])]; + tensor var_1198 = reshape(shape = var_1197, x = var_1196)[name = tensor("op_1198")]; + tensor new_scale_9 = add(x = prev_scale_9, y = var_1198)[name = tensor("new_scale_9")]; tensor const_21 = const()[name = tensor("const_21"), val = tensor(0x1.fffffep+127)]; - tensor clip_2 = clip(alpha = var_1022, beta = const_21, x = new_scale_9)[name = tensor("clip_2")]; + tensor clip_2 = clip(alpha = var_59, beta = const_21, x = new_scale_9)[name = tensor("clip_2")]; tensor sqrt_new_scale_1 = sqrt(x = clip_2)[name = tensor("sqrt_new_scale_1")]; - tensor var_1169 = real_div(x = new_kv_unnorm_9, y = sqrt_new_scale_1)[name = tensor("op_1169")]; - tensor var_1170_perm_0 = const()[name = tensor("op_1170_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_1202 = real_div(x = new_kv_unnorm_9, y = sqrt_new_scale_1)[name = tensor("op_1202")]; + tensor var_1203_perm_0 = const()[name = tensor("op_1203_perm_0"), val = tensor([0, 2, 1, 3])]; tensor out_27_axes_0 = const()[name = tensor("out_27_axes_0"), val = tensor([-1])]; - tensor var_1170 = transpose(perm = var_1170_perm_0, x = out_25)[name = tensor("transpose_24")]; - tensor out_27 = layer_norm(axes = out_27_axes_0, epsilon = var_1019, x = var_1170)[name = tensor("out_27")]; - tensor var_1174 = const()[name = tensor("op_1174"), val = tensor([9, 4, 256])]; - tensor out_29 = reshape(shape = var_1174, x = out_27)[name = tensor("out_29")]; - tensor var_1176 = silu(x = input_169)[name = tensor("op_1176")]; - tensor input_171 = mul(x = var_1176, y = out_29)[name = tensor("input_171")]; - tensor ret_out_9 = linear(bias = decoder_out_proj_0_bias, weight = decoder_out_proj_0_weight, x = input_171)[name = tensor("linear_42")]; - tensor input_173 = add(x = x_29, y = ret_out_9)[name = tensor("input_173")]; + tensor var_1203 = transpose(perm = var_1203_perm_0, x = out_25)[name = tensor("transpose_24")]; + tensor out_27 = layer_norm(axes = out_27_axes_0, epsilon = var_74, x = var_1203)[name = tensor("out_27")]; + tensor var_1207 = const()[name = tensor("op_1207"), val = tensor([9, 4, 256])]; + tensor out_29 = reshape(shape = var_1207, x = out_27)[name = tensor("out_29")]; + tensor var_1209 = silu(x = input_171)[name = tensor("op_1209")]; + tensor input_173 = mul(x = var_1209, y = out_29)[name = tensor("input_173")]; + tensor ret_out_9 = linear(bias = inner_decoder_out_proj_0_bias, weight = inner_decoder_out_proj_0_weight, x = input_173)[name = tensor("linear_42")]; + tensor input_175 = add(x = x_29, y = ret_out_9)[name = tensor("input_175")]; tensor xt_1_axes_0 = const()[name = tensor("xt_1_axes_0"), val = tensor([-1])]; - tensor xt_1 = layer_norm(axes = xt_1_axes_0, beta = decoder_norm11_0_bias, epsilon = var_1017, gamma = decoder_norm11_0_weight, x = input_173)[name = tensor("xt_1")]; - tensor var_1186 = const()[name = tensor("op_1186"), val = tensor([1, 9, 4, 256])]; - tensor var_1187 = reshape(shape = var_1186, x = xt_1)[name = tensor("op_1187")]; - tensor var_1188_perm_0 = const()[name = tensor("op_1188_perm_0"), val = tensor([0, 2, 1, 3])]; - tensor var_1191 = const()[name = tensor("op_1191"), val = tensor([4, 9, 256])]; - tensor var_1188 = transpose(perm = var_1188_perm_0, x = var_1187)[name = tensor("transpose_23")]; - tensor query_1 = reshape(shape = var_1191, x = var_1188)[name = tensor("query_1")]; + tensor xt_1 = layer_norm(axes = xt_1_axes_0, beta = inner_decoder_norm11_0_bias, epsilon = var_66, gamma = inner_decoder_norm11_0_weight, x = input_175)[name = tensor("xt_1")]; + tensor var_1219 = const()[name = tensor("op_1219"), val = tensor([1, 9, 4, 256])]; + tensor var_1220 = reshape(shape = var_1219, x = xt_1)[name = tensor("op_1220")]; + tensor var_1221_perm_0 = const()[name = tensor("op_1221_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_1224 = const()[name = tensor("op_1224"), val = tensor([4, 9, 256])]; + tensor var_1221 = transpose(perm = var_1221_perm_0, x = var_1220)[name = tensor("transpose_23")]; + tensor query_1 = reshape(shape = var_1224, x = var_1221)[name = tensor("query_1")]; tensor query_3_perm_0 = const()[name = tensor("query_3_perm_0"), val = tensor([1, 0, 2])]; tensor query_3 = transpose(perm = query_3_perm_0, x = query_1)[name = tensor("transpose_22")]; - tensor var_1214 = linear(bias = decoder_self_attn2_0_in_proj_bias, weight = decoder_self_attn2_0_in_proj_weight, x = query_3)[name = tensor("linear_43")]; + tensor var_1247 = linear(bias = inner_decoder_self_attn2_0_in_proj_bias, weight = inner_decoder_self_attn2_0_in_proj_weight, x = query_3)[name = tensor("linear_43")]; tensor concat_1 = const()[name = tensor("concat_1"), val = tensor([9, 4, 3, 256])]; - tensor var_1216 = reshape(shape = concat_1, x = var_1214)[name = tensor("op_1216")]; - tensor var_1217_axes_0 = const()[name = tensor("op_1217_axes_0"), val = tensor([0])]; - tensor var_1217 = expand_dims(axes = var_1217_axes_0, x = var_1216)[name = tensor("op_1217")]; - tensor var_1218_perm_0 = const()[name = tensor("op_1218_perm_0"), val = tensor([-2, 1, 2, 0, 4])]; - tensor var_1219_axes_0 = const()[name = tensor("op_1219_axes_0"), val = tensor([-2])]; - tensor var_1218 = transpose(perm = var_1218_perm_0, x = var_1217)[name = tensor("transpose_21")]; - tensor var_1219 = squeeze(axes = var_1219_axes_0, x = var_1218)[name = tensor("op_1219")]; + tensor var_1249 = reshape(shape = concat_1, x = var_1247)[name = tensor("op_1249")]; + tensor var_1250_axes_0 = const()[name = tensor("op_1250_axes_0"), val = tensor([0])]; + tensor var_1250 = expand_dims(axes = var_1250_axes_0, x = var_1249)[name = tensor("op_1250")]; + tensor var_1251_perm_0 = const()[name = tensor("op_1251_perm_0"), val = tensor([-2, 1, 2, 0, 4])]; + tensor var_1252_axes_0 = const()[name = tensor("op_1252_axes_0"), val = tensor([-2])]; + tensor var_1251 = transpose(perm = var_1251_perm_0, x = var_1250)[name = tensor("transpose_21")]; + tensor var_1252 = squeeze(axes = var_1252_axes_0, x = var_1251)[name = tensor("op_1252")]; tensor q_11_begin_0 = const()[name = tensor("q_11_begin_0"), val = tensor([0, 0, 0, 0])]; tensor q_11_end_0 = const()[name = tensor("q_11_end_0"), val = tensor([1, 9, 4, 256])]; tensor q_11_end_mask_0 = const()[name = tensor("q_11_end_mask_0"), val = tensor([false, true, true, true])]; tensor q_11_squeeze_mask_0 = const()[name = tensor("q_11_squeeze_mask_0"), val = tensor([true, false, false, false])]; - tensor q_11 = slice_by_index(begin = q_11_begin_0, end = q_11_end_0, end_mask = q_11_end_mask_0, squeeze_mask = q_11_squeeze_mask_0, x = var_1219)[name = tensor("q_11")]; + tensor q_11 = slice_by_index(begin = q_11_begin_0, end = q_11_end_0, end_mask = q_11_end_mask_0, squeeze_mask = q_11_squeeze_mask_0, x = var_1252)[name = tensor("q_11")]; tensor k_11_begin_0 = const()[name = tensor("k_11_begin_0"), val = tensor([1, 0, 0, 0])]; tensor k_11_end_0 = const()[name = tensor("k_11_end_0"), val = tensor([2, 9, 4, 256])]; tensor k_11_end_mask_0 = const()[name = tensor("k_11_end_mask_0"), val = tensor([false, true, true, true])]; tensor k_11_squeeze_mask_0 = const()[name = tensor("k_11_squeeze_mask_0"), val = tensor([true, false, false, false])]; - tensor k_11 = slice_by_index(begin = k_11_begin_0, end = k_11_end_0, end_mask = k_11_end_mask_0, squeeze_mask = k_11_squeeze_mask_0, x = var_1219)[name = tensor("k_11")]; + tensor k_11 = slice_by_index(begin = k_11_begin_0, end = k_11_end_0, end_mask = k_11_end_mask_0, squeeze_mask = k_11_squeeze_mask_0, x = var_1252)[name = tensor("k_11")]; tensor v_11_begin_0 = const()[name = tensor("v_11_begin_0"), val = tensor([2, 0, 0, 0])]; tensor v_11_end_0 = const()[name = tensor("v_11_end_0"), val = tensor([3, 9, 4, 256])]; tensor v_11_end_mask_0 = const()[name = tensor("v_11_end_mask_0"), val = tensor([false, true, true, true])]; tensor v_11_squeeze_mask_0 = const()[name = tensor("v_11_squeeze_mask_0"), val = tensor([true, false, false, false])]; - tensor v_11 = slice_by_index(begin = v_11_begin_0, end = v_11_end_0, end_mask = v_11_end_mask_0, squeeze_mask = v_11_squeeze_mask_0, x = var_1219)[name = tensor("v_11")]; - tensor var_1227 = const()[name = tensor("op_1227"), val = tensor([9, 16, 64])]; - tensor var_1228 = reshape(shape = var_1227, x = q_11)[name = tensor("op_1228")]; + tensor v_11 = slice_by_index(begin = v_11_begin_0, end = v_11_end_0, end_mask = v_11_end_mask_0, squeeze_mask = v_11_squeeze_mask_0, x = var_1252)[name = tensor("v_11")]; + tensor var_1260 = const()[name = tensor("op_1260"), val = tensor([9, 16, 64])]; + tensor var_1261 = reshape(shape = var_1260, x = q_11)[name = tensor("op_1261")]; tensor q_13_perm_0 = const()[name = tensor("q_13_perm_0"), val = tensor([1, 0, 2])]; - tensor var_1234 = const()[name = tensor("op_1234"), val = tensor([9, 16, 64])]; - tensor var_1235 = reshape(shape = var_1234, x = k_11)[name = tensor("op_1235")]; + tensor var_1267 = const()[name = tensor("op_1267"), val = tensor([9, 16, 64])]; + tensor var_1268 = reshape(shape = var_1267, x = k_11)[name = tensor("op_1268")]; tensor k_13_perm_0 = const()[name = tensor("k_13_perm_0"), val = tensor([1, 0, 2])]; - tensor var_1241 = const()[name = tensor("op_1241"), val = tensor([9, 16, 64])]; - tensor var_1242 = reshape(shape = var_1241, x = v_11)[name = tensor("op_1242")]; + tensor var_1274 = const()[name = tensor("op_1274"), val = tensor([9, 16, 64])]; + tensor var_1275 = reshape(shape = var_1274, x = v_11)[name = tensor("op_1275")]; tensor v_13_perm_0 = const()[name = tensor("v_13_perm_0"), val = tensor([1, 0, 2])]; - tensor var_1245 = const()[name = tensor("op_1245"), val = tensor([4, 4, 9, 64])]; - tensor q_13 = transpose(perm = q_13_perm_0, x = var_1228)[name = tensor("transpose_20")]; - tensor q_15 = reshape(shape = var_1245, x = q_13)[name = tensor("q_15")]; - tensor var_1247 = const()[name = tensor("op_1247"), val = tensor([4, 4, 9, 64])]; - tensor k_13 = transpose(perm = k_13_perm_0, x = var_1235)[name = tensor("transpose_19")]; - tensor k_15 = reshape(shape = var_1247, x = k_13)[name = tensor("k_15")]; - tensor var_1249 = const()[name = tensor("op_1249"), val = tensor([4, 4, 9, 64])]; - tensor v_13 = transpose(perm = v_13_perm_0, x = var_1242)[name = tensor("transpose_18")]; - tensor v_15 = reshape(shape = var_1249, x = v_13)[name = tensor("v_15")]; + tensor var_1278 = const()[name = tensor("op_1278"), val = tensor([4, 4, 9, 64])]; + tensor q_13 = transpose(perm = q_13_perm_0, x = var_1261)[name = tensor("transpose_20")]; + tensor q_15 = reshape(shape = var_1278, x = q_13)[name = tensor("q_15")]; + tensor var_1280 = const()[name = tensor("op_1280"), val = tensor([4, 4, 9, 64])]; + tensor k_13 = transpose(perm = k_13_perm_0, x = var_1268)[name = tensor("transpose_19")]; + tensor k_15 = reshape(shape = var_1280, x = k_13)[name = tensor("k_15")]; + tensor var_1282 = const()[name = tensor("op_1282"), val = tensor([4, 4, 9, 64])]; + tensor v_13 = transpose(perm = v_13_perm_0, x = var_1275)[name = tensor("transpose_18")]; + tensor v_15 = reshape(shape = var_1282, x = v_13)[name = tensor("v_15")]; tensor mul_1_y_0 = const()[name = tensor("mul_1_y_0"), val = tensor(0x1p-3)]; tensor mul_1 = mul(x = q_15, y = mul_1_y_0)[name = tensor("mul_1")]; tensor matmul_0_transpose_y_0 = const()[name = tensor("matmul_0_transpose_y_0"), val = tensor(true)]; @@ -1119,30 +1134,30 @@ program(1.0) tensor attn_output_1_transpose_x_0 = const()[name = tensor("attn_output_1_transpose_x_0"), val = tensor(false)]; tensor attn_output_1_transpose_y_0 = const()[name = tensor("attn_output_1_transpose_y_0"), val = tensor(false)]; tensor attn_output_1 = matmul(transpose_x = attn_output_1_transpose_x_0, transpose_y = attn_output_1_transpose_y_0, x = softmax_0, y = v_15)[name = tensor("attn_output_1")]; - tensor var_1252 = const()[name = tensor("op_1252"), val = tensor([2, 0, 1, 3])]; - tensor var_1257 = const()[name = tensor("op_1257"), val = tensor([36, 256])]; - tensor var_1253 = transpose(perm = var_1252, x = attn_output_1)[name = tensor("transpose_17")]; - tensor attn_output_3 = reshape(shape = var_1257, x = var_1253)[name = tensor("attn_output_3")]; - tensor attn_output_5 = linear(bias = decoder_self_attn2_0_out_proj_bias, weight = decoder_self_attn2_0_out_proj_weight, x = attn_output_3)[name = tensor("linear_44")]; - tensor var_1261 = const()[name = tensor("op_1261"), val = tensor([9, 4, 256])]; - tensor attn_output_7 = reshape(shape = var_1261, x = attn_output_5)[name = tensor("attn_output_7")]; + tensor var_1285 = const()[name = tensor("op_1285"), val = tensor([2, 0, 1, 3])]; + tensor var_1290 = const()[name = tensor("op_1290"), val = tensor([36, 256])]; + tensor var_1286 = transpose(perm = var_1285, x = attn_output_1)[name = tensor("transpose_17")]; + tensor attn_output_3 = reshape(shape = var_1290, x = var_1286)[name = tensor("attn_output_3")]; + tensor attn_output_5 = linear(bias = inner_decoder_self_attn2_0_out_proj_bias, weight = inner_decoder_self_attn2_0_out_proj_weight, x = attn_output_3)[name = tensor("linear_44")]; + tensor var_1294 = const()[name = tensor("op_1294"), val = tensor([9, 4, 256])]; + tensor attn_output_7 = reshape(shape = var_1294, x = attn_output_5)[name = tensor("attn_output_7")]; tensor sa2_out_1_perm_0 = const()[name = tensor("sa2_out_1_perm_0"), val = tensor([1, 0, 2])]; tensor sa2_out_1 = transpose(perm = sa2_out_1_perm_0, x = attn_output_7)[name = tensor("transpose_16")]; - tensor input_175 = add(x = query_1, y = sa2_out_1)[name = tensor("input_175")]; - tensor input_177_axes_0 = const()[name = tensor("input_177_axes_0"), val = tensor([-1])]; - tensor input_177 = layer_norm(axes = input_177_axes_0, beta = decoder_norm21_0_bias, epsilon = var_1017, gamma = decoder_norm21_0_weight, x = input_175)[name = tensor("input_177")]; - tensor input_179 = linear(bias = decoder_linear1_0_bias, weight = decoder_linear1_0_weight, x = input_177)[name = tensor("linear_45")]; - tensor input_181 = relu(x = input_179)[name = tensor("input_181")]; - tensor ffn_out_1 = linear(bias = decoder_linear2_0_bias, weight = decoder_linear2_0_weight, x = input_181)[name = tensor("linear_46")]; - tensor input_183 = add(x = input_177, y = ffn_out_1)[name = tensor("input_183")]; + tensor input_177 = add(x = query_1, y = sa2_out_1)[name = tensor("input_177")]; + tensor input_179_axes_0 = const()[name = tensor("input_179_axes_0"), val = tensor([-1])]; + tensor input_179 = layer_norm(axes = input_179_axes_0, beta = inner_decoder_norm21_0_bias, epsilon = var_66, gamma = inner_decoder_norm21_0_weight, x = input_177)[name = tensor("input_179")]; + tensor input_181 = linear(bias = inner_decoder_linear1_0_bias, weight = inner_decoder_linear1_0_weight, x = input_179)[name = tensor("linear_45")]; + tensor input_183 = relu(x = input_181)[name = tensor("input_183")]; + tensor ffn_out_1 = linear(bias = inner_decoder_linear2_0_bias, weight = inner_decoder_linear2_0_weight, x = input_183)[name = tensor("linear_46")]; + tensor input_185 = add(x = input_179, y = ffn_out_1)[name = tensor("input_185")]; tensor xt_3_axes_0 = const()[name = tensor("xt_3_axes_0"), val = tensor([-1])]; - tensor xt_3 = layer_norm(axes = xt_3_axes_0, beta = decoder_norm22_0_bias, epsilon = var_1017, gamma = decoder_norm22_0_weight, x = input_183)[name = tensor("xt_3")]; - tensor var_1281 = const()[name = tensor("op_1281"), val = tensor([1, 4, 9, 256])]; - tensor x_31 = reshape(shape = var_1281, x = xt_3)[name = tensor("x_31")]; - tensor var_1283_perm_0 = const()[name = tensor("op_1283_perm_0"), val = tensor([0, 2, 1, 3])]; - tensor var_1287 = const()[name = tensor("op_1287"), val = tensor([9, 4, 256])]; - tensor var_1283 = transpose(perm = var_1283_perm_0, x = x_31)[name = tensor("transpose_15")]; - tensor x = reshape(shape = var_1287, x = var_1283)[name = tensor("x")]; + tensor xt_3 = layer_norm(axes = xt_3_axes_0, beta = inner_decoder_norm22_0_bias, epsilon = var_66, gamma = inner_decoder_norm22_0_weight, x = input_185)[name = tensor("xt_3")]; + tensor var_1314 = const()[name = tensor("op_1314"), val = tensor([1, 4, 9, 256])]; + tensor x_31 = reshape(shape = var_1314, x = xt_3)[name = tensor("x_31")]; + tensor var_1316_perm_0 = const()[name = tensor("op_1316_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_1320 = const()[name = tensor("op_1320"), val = tensor([9, 4, 256])]; + tensor var_1316 = transpose(perm = var_1316_perm_0, x = x_31)[name = tensor("transpose_15")]; + tensor x = reshape(shape = var_1320, x = var_1316)[name = tensor("x")]; tensor prev_kv_begin_0 = const()[name = tensor("prev_kv_begin_0"), val = tensor([1, 0, 0, 0, 0])]; tensor prev_kv_end_0 = const()[name = tensor("prev_kv_end_0"), val = tensor([2, 9, 4, 64, 64])]; tensor prev_kv_end_mask_0 = const()[name = tensor("prev_kv_end_mask_0"), val = tensor([false, true, true, true, true])]; @@ -1153,120 +1168,120 @@ program(1.0) tensor prev_scale_end_mask_0 = const()[name = tensor("prev_scale_end_mask_0"), val = tensor([false, true])]; tensor prev_scale_squeeze_mask_0 = const()[name = tensor("prev_scale_squeeze_mask_0"), val = tensor([true, false])]; tensor prev_scale = slice_by_index(begin = prev_scale_begin_0, end = prev_scale_end_0, end_mask = prev_scale_end_mask_0, squeeze_mask = prev_scale_squeeze_mask_0, x = dec_scale)[name = tensor("prev_scale")]; - tensor var_1295 = linear(bias = decoder_q_proj_1_bias, weight = decoder_q_proj_1_weight, x = x)[name = tensor("linear_47")]; - tensor var_1296 = const()[name = tensor("op_1296"), val = tensor([9, 4, 4, 64])]; - tensor var_1297 = reshape(shape = var_1296, x = var_1295)[name = tensor("op_1297")]; + tensor var_1328 = linear(bias = inner_decoder_q_proj_1_bias, weight = inner_decoder_q_proj_1_weight, x = x)[name = tensor("linear_47")]; + tensor var_1329 = const()[name = tensor("op_1329"), val = tensor([9, 4, 4, 64])]; + tensor var_1330 = reshape(shape = var_1329, x = var_1328)[name = tensor("op_1330")]; tensor q_17_perm_0 = const()[name = tensor("q_17_perm_0"), val = tensor([0, 2, 1, 3])]; - tensor var_1301 = linear(bias = decoder_k_proj_1_bias, weight = decoder_k_proj_1_weight, x = x)[name = tensor("linear_48")]; - tensor var_1302 = const()[name = tensor("op_1302"), val = tensor(0x1p-3)]; - tensor var_1303 = mul(x = var_1301, y = var_1302)[name = tensor("op_1303")]; - tensor var_1304 = const()[name = tensor("op_1304"), val = tensor([9, 4, 4, 64])]; - tensor var_1305 = reshape(shape = var_1304, x = var_1303)[name = tensor("op_1305")]; + tensor var_1334 = linear(bias = inner_decoder_k_proj_1_bias, weight = inner_decoder_k_proj_1_weight, x = x)[name = tensor("linear_48")]; + tensor var_1335 = const()[name = tensor("op_1335"), val = tensor(0x1p-3)]; + tensor var_1336 = mul(x = var_1334, y = var_1335)[name = tensor("op_1336")]; + tensor var_1337 = const()[name = tensor("op_1337"), val = tensor([9, 4, 4, 64])]; + tensor var_1338 = reshape(shape = var_1337, x = var_1336)[name = tensor("op_1338")]; tensor k_17_perm_0 = const()[name = tensor("k_17_perm_0"), val = tensor([0, 2, 1, 3])]; - tensor var_1309 = linear(bias = decoder_v_proj_1_bias, weight = decoder_v_proj_1_weight, x = x)[name = tensor("linear_49")]; - tensor var_1310 = const()[name = tensor("op_1310"), val = tensor([9, 4, 4, 64])]; - tensor var_1311 = reshape(shape = var_1310, x = var_1309)[name = tensor("op_1311")]; + tensor var_1342 = linear(bias = inner_decoder_v_proj_1_bias, weight = inner_decoder_v_proj_1_weight, x = x)[name = tensor("linear_49")]; + tensor var_1343 = const()[name = tensor("op_1343"), val = tensor([9, 4, 4, 64])]; + tensor var_1344 = reshape(shape = var_1343, x = var_1342)[name = tensor("op_1344")]; tensor v_17_perm_0 = const()[name = tensor("v_17_perm_0"), val = tensor([0, 2, 1, 3])]; - tensor input_187 = linear(bias = decoder_g_proj_1_bias, weight = decoder_g_proj_1_weight, x = x)[name = tensor("linear_50")]; + tensor input_189 = linear(bias = inner_decoder_g_proj_1_bias, weight = inner_decoder_g_proj_1_weight, x = x)[name = tensor("linear_50")]; tensor sqrt_s0 = sqrt(x = prev_scale)[name = tensor("sqrt_s0")]; tensor s_t = add(x = prev_scale, y = cumsum_mask_1)[name = tensor("s_t")]; tensor const_32 = const()[name = tensor("const_32"), val = tensor(0x1.fffffep+127)]; - tensor clip_3 = clip(alpha = var_1022, beta = const_32, x = s_t)[name = tensor("clip_3")]; + tensor clip_3 = clip(alpha = var_59, beta = const_32, x = s_t)[name = tensor("clip_3")]; tensor sqrt_s_t = sqrt(x = clip_3)[name = tensor("sqrt_s_t")]; tensor qk_transpose_x_1 = const()[name = tensor("qk_transpose_x_1"), val = tensor(false)]; tensor qk_transpose_y_1 = const()[name = tensor("qk_transpose_y_1"), val = tensor(true)]; - tensor k_17 = transpose(perm = k_17_perm_0, x = var_1305)[name = tensor("transpose_13")]; - tensor q_17 = transpose(perm = q_17_perm_0, x = var_1297)[name = tensor("transpose_14")]; + tensor k_17 = transpose(perm = k_17_perm_0, x = var_1338)[name = tensor("transpose_13")]; + tensor q_17 = transpose(perm = q_17_perm_0, x = var_1330)[name = tensor("transpose_14")]; tensor qk = matmul(transpose_x = qk_transpose_x_1, transpose_y = qk_transpose_y_1, x = q_17, y = k_17)[name = tensor("qk")]; - tensor var_1326 = const()[name = tensor("op_1326"), val = tensor([4, 1])]; - tensor var_1327 = reshape(shape = var_1326, x = sqrt_s_t)[name = tensor("op_1327")]; - tensor M = real_div(x = causal_with_valid_1, y = var_1327)[name = tensor("M")]; - tensor var_1329 = mul(x = qk, y = M)[name = tensor("op_1329")]; - tensor inner_transpose_x_0 = const()[name = tensor("inner_transpose_x_0"), val = tensor(false)]; - tensor inner_transpose_y_0 = const()[name = tensor("inner_transpose_y_0"), val = tensor(false)]; - tensor v_17 = transpose(perm = v_17_perm_0, x = var_1311)[name = tensor("transpose_12")]; - tensor inner = matmul(transpose_x = inner_transpose_x_0, transpose_y = inner_transpose_y_0, x = var_1329, y = v_17)[name = tensor("inner")]; - tensor var_1331_transpose_x_0 = const()[name = tensor("op_1331_transpose_x_0"), val = tensor(false)]; - tensor var_1331_transpose_y_0 = const()[name = tensor("op_1331_transpose_y_0"), val = tensor(false)]; - tensor var_1331 = matmul(transpose_x = var_1331_transpose_x_0, transpose_y = var_1331_transpose_y_0, x = q_17, y = prev_kv)[name = tensor("op_1331")]; - tensor var_1332 = real_div(x = sqrt_s0, y = sqrt_s_t)[name = tensor("op_1332")]; - tensor var_1333 = const()[name = tensor("op_1333"), val = tensor([1, 1, 4, 1])]; - tensor var_1334 = reshape(shape = var_1333, x = var_1332)[name = tensor("op_1334")]; - tensor cross = mul(x = var_1331, y = var_1334)[name = tensor("cross")]; - tensor out_31 = add(x = inner, y = cross)[name = tensor("out_31")]; - tensor v_masked = mul(x = v_17, y = var_1157)[name = tensor("v_masked")]; - tensor var_1340 = mul(x = prev_kv, y = sqrt_s0)[name = tensor("op_1340")]; - tensor var_1342_transpose_x_1 = const()[name = tensor("op_1342_transpose_x_1"), val = tensor(true)]; - tensor var_1342_transpose_y_1 = const()[name = tensor("op_1342_transpose_y_1"), val = tensor(false)]; - tensor var_1342 = matmul(transpose_x = var_1342_transpose_x_1, transpose_y = var_1342_transpose_y_1, x = k_17, y = v_masked)[name = tensor("op_1342")]; - tensor new_kv_unnorm = add(x = var_1340, y = var_1342)[name = tensor("new_kv_unnorm")]; - tensor new_scale = add(x = prev_scale, y = var_1165)[name = tensor("new_scale")]; + tensor var_1359 = const()[name = tensor("op_1359"), val = tensor([4, 1])]; + tensor var_1360 = reshape(shape = var_1359, x = sqrt_s_t)[name = tensor("op_1360")]; + tensor M = real_div(x = causal_with_valid_1, y = var_1360)[name = tensor("M")]; + tensor var_1362 = mul(x = qk, y = M)[name = tensor("op_1362")]; + tensor inner_11_transpose_x_0 = const()[name = tensor("inner_11_transpose_x_0"), val = tensor(false)]; + tensor inner_11_transpose_y_0 = const()[name = tensor("inner_11_transpose_y_0"), val = tensor(false)]; + tensor v_17 = transpose(perm = v_17_perm_0, x = var_1344)[name = tensor("transpose_12")]; + tensor inner_11 = matmul(transpose_x = inner_11_transpose_x_0, transpose_y = inner_11_transpose_y_0, x = var_1362, y = v_17)[name = tensor("inner_11")]; + tensor var_1364_transpose_x_0 = const()[name = tensor("op_1364_transpose_x_0"), val = tensor(false)]; + tensor var_1364_transpose_y_0 = const()[name = tensor("op_1364_transpose_y_0"), val = tensor(false)]; + tensor var_1364 = matmul(transpose_x = var_1364_transpose_x_0, transpose_y = var_1364_transpose_y_0, x = q_17, y = prev_kv)[name = tensor("op_1364")]; + tensor var_1365 = real_div(x = sqrt_s0, y = sqrt_s_t)[name = tensor("op_1365")]; + tensor var_1366 = const()[name = tensor("op_1366"), val = tensor([1, 1, 4, 1])]; + tensor var_1367 = reshape(shape = var_1366, x = var_1365)[name = tensor("op_1367")]; + tensor cross = mul(x = var_1364, y = var_1367)[name = tensor("cross")]; + tensor out_31 = add(x = inner_11, y = cross)[name = tensor("out_31")]; + tensor v_masked = mul(x = v_17, y = var_1190)[name = tensor("v_masked")]; + tensor var_1373 = mul(x = prev_kv, y = sqrt_s0)[name = tensor("op_1373")]; + tensor var_1375_transpose_x_1 = const()[name = tensor("op_1375_transpose_x_1"), val = tensor(true)]; + tensor var_1375_transpose_y_1 = const()[name = tensor("op_1375_transpose_y_1"), val = tensor(false)]; + tensor var_1375 = matmul(transpose_x = var_1375_transpose_x_1, transpose_y = var_1375_transpose_y_1, x = k_17, y = v_masked)[name = tensor("op_1375")]; + tensor new_kv_unnorm = add(x = var_1373, y = var_1375)[name = tensor("new_kv_unnorm")]; + tensor new_scale = add(x = prev_scale, y = var_1198)[name = tensor("new_scale")]; tensor const_33 = const()[name = tensor("const_33"), val = tensor(0x1.fffffep+127)]; - tensor clip_4 = clip(alpha = var_1022, beta = const_33, x = new_scale)[name = tensor("clip_4")]; + tensor clip_4 = clip(alpha = var_59, beta = const_33, x = new_scale)[name = tensor("clip_4")]; tensor sqrt_new_scale = sqrt(x = clip_4)[name = tensor("sqrt_new_scale")]; tensor nkv = real_div(x = new_kv_unnorm, y = sqrt_new_scale)[name = tensor("nkv")]; - tensor var_1351_perm_0 = const()[name = tensor("op_1351_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_1384_perm_0 = const()[name = tensor("op_1384_perm_0"), val = tensor([0, 2, 1, 3])]; tensor out_33_axes_0 = const()[name = tensor("out_33_axes_0"), val = tensor([-1])]; - tensor var_1351 = transpose(perm = var_1351_perm_0, x = out_31)[name = tensor("transpose_11")]; - tensor out_33 = layer_norm(axes = out_33_axes_0, epsilon = var_1019, x = var_1351)[name = tensor("out_33")]; - tensor var_1355 = const()[name = tensor("op_1355"), val = tensor([9, 4, 256])]; - tensor out = reshape(shape = var_1355, x = out_33)[name = tensor("out")]; - tensor var_1357 = silu(x = input_187)[name = tensor("op_1357")]; - tensor input_189 = mul(x = var_1357, y = out)[name = tensor("input_189")]; - tensor ret_out = linear(bias = decoder_out_proj_1_bias, weight = decoder_out_proj_1_weight, x = input_189)[name = tensor("linear_51")]; - tensor input_191 = add(x = x, y = ret_out)[name = tensor("input_191")]; + tensor var_1384 = transpose(perm = var_1384_perm_0, x = out_31)[name = tensor("transpose_11")]; + tensor out_33 = layer_norm(axes = out_33_axes_0, epsilon = var_74, x = var_1384)[name = tensor("out_33")]; + tensor var_1388 = const()[name = tensor("op_1388"), val = tensor([9, 4, 256])]; + tensor out = reshape(shape = var_1388, x = out_33)[name = tensor("out")]; + tensor var_1390 = silu(x = input_189)[name = tensor("op_1390")]; + tensor input_191 = mul(x = var_1390, y = out)[name = tensor("input_191")]; + tensor ret_out = linear(bias = inner_decoder_out_proj_1_bias, weight = inner_decoder_out_proj_1_weight, x = input_191)[name = tensor("linear_51")]; + tensor input_193 = add(x = x, y = ret_out)[name = tensor("input_193")]; tensor xt_5_axes_0 = const()[name = tensor("xt_5_axes_0"), val = tensor([-1])]; - tensor xt_5 = layer_norm(axes = xt_5_axes_0, beta = decoder_norm11_1_bias, epsilon = var_1017, gamma = decoder_norm11_1_weight, x = input_191)[name = tensor("xt_5")]; - tensor var_1367 = const()[name = tensor("op_1367"), val = tensor([1, 9, 4, 256])]; - tensor var_1368 = reshape(shape = var_1367, x = xt_5)[name = tensor("op_1368")]; - tensor var_1369_perm_0 = const()[name = tensor("op_1369_perm_0"), val = tensor([0, 2, 1, 3])]; - tensor var_1372 = const()[name = tensor("op_1372"), val = tensor([4, 9, 256])]; - tensor var_1369 = transpose(perm = var_1369_perm_0, x = var_1368)[name = tensor("transpose_10")]; - tensor query_5 = reshape(shape = var_1372, x = var_1369)[name = tensor("query_5")]; + tensor xt_5 = layer_norm(axes = xt_5_axes_0, beta = inner_decoder_norm11_1_bias, epsilon = var_66, gamma = inner_decoder_norm11_1_weight, x = input_193)[name = tensor("xt_5")]; + tensor var_1400 = const()[name = tensor("op_1400"), val = tensor([1, 9, 4, 256])]; + tensor var_1401 = reshape(shape = var_1400, x = xt_5)[name = tensor("op_1401")]; + tensor var_1402_perm_0 = const()[name = tensor("op_1402_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_1405 = const()[name = tensor("op_1405"), val = tensor([4, 9, 256])]; + tensor var_1402 = transpose(perm = var_1402_perm_0, x = var_1401)[name = tensor("transpose_10")]; + tensor query_5 = reshape(shape = var_1405, x = var_1402)[name = tensor("query_5")]; tensor query_perm_0 = const()[name = tensor("query_perm_0"), val = tensor([1, 0, 2])]; tensor query = transpose(perm = query_perm_0, x = query_5)[name = tensor("transpose_9")]; - tensor var_1395 = linear(bias = decoder_self_attn2_1_in_proj_bias, weight = decoder_self_attn2_1_in_proj_weight, x = query)[name = tensor("linear_52")]; + tensor var_1428 = linear(bias = inner_decoder_self_attn2_1_in_proj_bias, weight = inner_decoder_self_attn2_1_in_proj_weight, x = query)[name = tensor("linear_52")]; tensor concat_2 = const()[name = tensor("concat_2"), val = tensor([9, 4, 3, 256])]; - tensor var_1397 = reshape(shape = concat_2, x = var_1395)[name = tensor("op_1397")]; - tensor var_1398_axes_0 = const()[name = tensor("op_1398_axes_0"), val = tensor([0])]; - tensor var_1398 = expand_dims(axes = var_1398_axes_0, x = var_1397)[name = tensor("op_1398")]; - tensor var_1399_perm_0 = const()[name = tensor("op_1399_perm_0"), val = tensor([-2, 1, 2, 0, 4])]; - tensor var_1400_axes_0 = const()[name = tensor("op_1400_axes_0"), val = tensor([-2])]; - tensor var_1399 = transpose(perm = var_1399_perm_0, x = var_1398)[name = tensor("transpose_8")]; - tensor var_1400 = squeeze(axes = var_1400_axes_0, x = var_1399)[name = tensor("op_1400")]; + tensor var_1430 = reshape(shape = concat_2, x = var_1428)[name = tensor("op_1430")]; + tensor var_1431_axes_0 = const()[name = tensor("op_1431_axes_0"), val = tensor([0])]; + tensor var_1431 = expand_dims(axes = var_1431_axes_0, x = var_1430)[name = tensor("op_1431")]; + tensor var_1432_perm_0 = const()[name = tensor("op_1432_perm_0"), val = tensor([-2, 1, 2, 0, 4])]; + tensor var_1433_axes_0 = const()[name = tensor("op_1433_axes_0"), val = tensor([-2])]; + tensor var_1432 = transpose(perm = var_1432_perm_0, x = var_1431)[name = tensor("transpose_8")]; + tensor var_1433 = squeeze(axes = var_1433_axes_0, x = var_1432)[name = tensor("op_1433")]; tensor q_19_begin_0 = const()[name = tensor("q_19_begin_0"), val = tensor([0, 0, 0, 0])]; tensor q_19_end_0 = const()[name = tensor("q_19_end_0"), val = tensor([1, 9, 4, 256])]; tensor q_19_end_mask_0 = const()[name = tensor("q_19_end_mask_0"), val = tensor([false, true, true, true])]; tensor q_19_squeeze_mask_0 = const()[name = tensor("q_19_squeeze_mask_0"), val = tensor([true, false, false, false])]; - tensor q_19 = slice_by_index(begin = q_19_begin_0, end = q_19_end_0, end_mask = q_19_end_mask_0, squeeze_mask = q_19_squeeze_mask_0, x = var_1400)[name = tensor("q_19")]; + tensor q_19 = slice_by_index(begin = q_19_begin_0, end = q_19_end_0, end_mask = q_19_end_mask_0, squeeze_mask = q_19_squeeze_mask_0, x = var_1433)[name = tensor("q_19")]; tensor k_19_begin_0 = const()[name = tensor("k_19_begin_0"), val = tensor([1, 0, 0, 0])]; tensor k_19_end_0 = const()[name = tensor("k_19_end_0"), val = tensor([2, 9, 4, 256])]; tensor k_19_end_mask_0 = const()[name = tensor("k_19_end_mask_0"), val = tensor([false, true, true, true])]; tensor k_19_squeeze_mask_0 = const()[name = tensor("k_19_squeeze_mask_0"), val = tensor([true, false, false, false])]; - tensor k_19 = slice_by_index(begin = k_19_begin_0, end = k_19_end_0, end_mask = k_19_end_mask_0, squeeze_mask = k_19_squeeze_mask_0, x = var_1400)[name = tensor("k_19")]; + tensor k_19 = slice_by_index(begin = k_19_begin_0, end = k_19_end_0, end_mask = k_19_end_mask_0, squeeze_mask = k_19_squeeze_mask_0, x = var_1433)[name = tensor("k_19")]; tensor v_19_begin_0 = const()[name = tensor("v_19_begin_0"), val = tensor([2, 0, 0, 0])]; tensor v_19_end_0 = const()[name = tensor("v_19_end_0"), val = tensor([3, 9, 4, 256])]; tensor v_19_end_mask_0 = const()[name = tensor("v_19_end_mask_0"), val = tensor([false, true, true, true])]; tensor v_19_squeeze_mask_0 = const()[name = tensor("v_19_squeeze_mask_0"), val = tensor([true, false, false, false])]; - tensor v_19 = slice_by_index(begin = v_19_begin_0, end = v_19_end_0, end_mask = v_19_end_mask_0, squeeze_mask = v_19_squeeze_mask_0, x = var_1400)[name = tensor("v_19")]; - tensor var_1408 = const()[name = tensor("op_1408"), val = tensor([9, 16, 64])]; - tensor var_1409 = reshape(shape = var_1408, x = q_19)[name = tensor("op_1409")]; + tensor v_19 = slice_by_index(begin = v_19_begin_0, end = v_19_end_0, end_mask = v_19_end_mask_0, squeeze_mask = v_19_squeeze_mask_0, x = var_1433)[name = tensor("v_19")]; + tensor var_1441 = const()[name = tensor("op_1441"), val = tensor([9, 16, 64])]; + tensor var_1442 = reshape(shape = var_1441, x = q_19)[name = tensor("op_1442")]; tensor q_21_perm_0 = const()[name = tensor("q_21_perm_0"), val = tensor([1, 0, 2])]; - tensor var_1415 = const()[name = tensor("op_1415"), val = tensor([9, 16, 64])]; - tensor var_1416 = reshape(shape = var_1415, x = k_19)[name = tensor("op_1416")]; + tensor var_1448 = const()[name = tensor("op_1448"), val = tensor([9, 16, 64])]; + tensor var_1449 = reshape(shape = var_1448, x = k_19)[name = tensor("op_1449")]; tensor k_21_perm_0 = const()[name = tensor("k_21_perm_0"), val = tensor([1, 0, 2])]; - tensor var_1422 = const()[name = tensor("op_1422"), val = tensor([9, 16, 64])]; - tensor var_1423 = reshape(shape = var_1422, x = v_19)[name = tensor("op_1423")]; + tensor var_1455 = const()[name = tensor("op_1455"), val = tensor([9, 16, 64])]; + tensor var_1456 = reshape(shape = var_1455, x = v_19)[name = tensor("op_1456")]; tensor v_21_perm_0 = const()[name = tensor("v_21_perm_0"), val = tensor([1, 0, 2])]; - tensor var_1426 = const()[name = tensor("op_1426"), val = tensor([4, 4, 9, 64])]; - tensor q_21 = transpose(perm = q_21_perm_0, x = var_1409)[name = tensor("transpose_7")]; - tensor q = reshape(shape = var_1426, x = q_21)[name = tensor("q")]; - tensor var_1428 = const()[name = tensor("op_1428"), val = tensor([4, 4, 9, 64])]; - tensor k_21 = transpose(perm = k_21_perm_0, x = var_1416)[name = tensor("transpose_6")]; - tensor k = reshape(shape = var_1428, x = k_21)[name = tensor("k")]; - tensor var_1430 = const()[name = tensor("op_1430"), val = tensor([4, 4, 9, 64])]; - tensor v_21 = transpose(perm = v_21_perm_0, x = var_1423)[name = tensor("transpose_5")]; - tensor v = reshape(shape = var_1430, x = v_21)[name = tensor("v")]; + tensor var_1459 = const()[name = tensor("op_1459"), val = tensor([4, 4, 9, 64])]; + tensor q_21 = transpose(perm = q_21_perm_0, x = var_1442)[name = tensor("transpose_7")]; + tensor q = reshape(shape = var_1459, x = q_21)[name = tensor("q")]; + tensor var_1461 = const()[name = tensor("op_1461"), val = tensor([4, 4, 9, 64])]; + tensor k_21 = transpose(perm = k_21_perm_0, x = var_1449)[name = tensor("transpose_6")]; + tensor k = reshape(shape = var_1461, x = k_21)[name = tensor("k")]; + tensor var_1463 = const()[name = tensor("op_1463"), val = tensor([4, 4, 9, 64])]; + tensor v_21 = transpose(perm = v_21_perm_0, x = var_1456)[name = tensor("transpose_5")]; + tensor v = reshape(shape = var_1463, x = v_21)[name = tensor("v")]; tensor mul_3_y_0 = const()[name = tensor("mul_3_y_0"), val = tensor(0x1p-3)]; tensor mul_3 = mul(x = q, y = mul_3_y_0)[name = tensor("mul_3")]; tensor matmul_1_transpose_y_0 = const()[name = tensor("matmul_1_transpose_y_0"), val = tensor(true)]; @@ -1277,36 +1292,36 @@ program(1.0) tensor attn_output_9_transpose_x_0 = const()[name = tensor("attn_output_9_transpose_x_0"), val = tensor(false)]; tensor attn_output_9_transpose_y_0 = const()[name = tensor("attn_output_9_transpose_y_0"), val = tensor(false)]; tensor attn_output_9 = matmul(transpose_x = attn_output_9_transpose_x_0, transpose_y = attn_output_9_transpose_y_0, x = softmax_1, y = v)[name = tensor("attn_output_9")]; - tensor var_1433 = const()[name = tensor("op_1433"), val = tensor([2, 0, 1, 3])]; - tensor var_1438 = const()[name = tensor("op_1438"), val = tensor([36, 256])]; - tensor var_1434 = transpose(perm = var_1433, x = attn_output_9)[name = tensor("transpose_4")]; - tensor attn_output_11 = reshape(shape = var_1438, x = var_1434)[name = tensor("attn_output_11")]; - tensor attn_output_13 = linear(bias = decoder_self_attn2_1_out_proj_bias, weight = decoder_self_attn2_1_out_proj_weight, x = attn_output_11)[name = tensor("linear_53")]; - tensor var_1442 = const()[name = tensor("op_1442"), val = tensor([9, 4, 256])]; - tensor attn_output = reshape(shape = var_1442, x = attn_output_13)[name = tensor("attn_output")]; + tensor var_1466 = const()[name = tensor("op_1466"), val = tensor([2, 0, 1, 3])]; + tensor var_1471 = const()[name = tensor("op_1471"), val = tensor([36, 256])]; + tensor var_1467 = transpose(perm = var_1466, x = attn_output_9)[name = tensor("transpose_4")]; + tensor attn_output_11 = reshape(shape = var_1471, x = var_1467)[name = tensor("attn_output_11")]; + tensor attn_output_13 = linear(bias = inner_decoder_self_attn2_1_out_proj_bias, weight = inner_decoder_self_attn2_1_out_proj_weight, x = attn_output_11)[name = tensor("linear_53")]; + tensor var_1475 = const()[name = tensor("op_1475"), val = tensor([9, 4, 256])]; + tensor attn_output = reshape(shape = var_1475, x = attn_output_13)[name = tensor("attn_output")]; tensor sa2_out_perm_0 = const()[name = tensor("sa2_out_perm_0"), val = tensor([1, 0, 2])]; tensor sa2_out = transpose(perm = sa2_out_perm_0, x = attn_output)[name = tensor("transpose_3")]; - tensor input_193 = add(x = query_5, y = sa2_out)[name = tensor("input_193")]; - tensor input_195_axes_0 = const()[name = tensor("input_195_axes_0"), val = tensor([-1])]; - tensor input_195 = layer_norm(axes = input_195_axes_0, beta = decoder_norm21_1_bias, epsilon = var_1017, gamma = decoder_norm21_1_weight, x = input_193)[name = tensor("input_195")]; - tensor input_197 = linear(bias = decoder_linear1_1_bias, weight = decoder_linear1_1_weight, x = input_195)[name = tensor("linear_54")]; - tensor input_199 = relu(x = input_197)[name = tensor("input_199")]; - tensor ffn_out = linear(bias = decoder_linear2_1_bias, weight = decoder_linear2_1_weight, x = input_199)[name = tensor("linear_55")]; - tensor input_201 = add(x = input_195, y = ffn_out)[name = tensor("input_201")]; + tensor input_195 = add(x = query_5, y = sa2_out)[name = tensor("input_195")]; + tensor input_197_axes_0 = const()[name = tensor("input_197_axes_0"), val = tensor([-1])]; + tensor input_197 = layer_norm(axes = input_197_axes_0, beta = inner_decoder_norm21_1_bias, epsilon = var_66, gamma = inner_decoder_norm21_1_weight, x = input_195)[name = tensor("input_197")]; + tensor input_199 = linear(bias = inner_decoder_linear1_1_bias, weight = inner_decoder_linear1_1_weight, x = input_197)[name = tensor("linear_54")]; + tensor input_201 = relu(x = input_199)[name = tensor("input_201")]; + tensor ffn_out = linear(bias = inner_decoder_linear2_1_bias, weight = inner_decoder_linear2_1_weight, x = input_201)[name = tensor("linear_55")]; + tensor input_203 = add(x = input_197, y = ffn_out)[name = tensor("input_203")]; tensor xt_axes_0 = const()[name = tensor("xt_axes_0"), val = tensor([-1])]; - tensor xt = layer_norm(axes = xt_axes_0, beta = decoder_norm22_1_bias, epsilon = var_1017, gamma = decoder_norm22_1_weight, x = input_201)[name = tensor("xt")]; - tensor var_1462 = const()[name = tensor("op_1462"), val = tensor([1, 4, 9, 256])]; - tensor input = reshape(shape = var_1462, x = xt)[name = tensor("input")]; - tensor var_1464 = const()[name = tensor("op_1464"), val = tensor([-1])]; - tensor var_1465 = reduce_l2_norm(axes = var_1464, keep_dims = var_1020, x = input)[name = tensor("op_1465")]; + tensor xt = layer_norm(axes = xt_axes_0, beta = inner_decoder_norm22_1_bias, epsilon = var_66, gamma = inner_decoder_norm22_1_weight, x = input_203)[name = tensor("xt")]; + tensor var_1495 = const()[name = tensor("op_1495"), val = tensor([1, 4, 9, 256])]; + tensor input = reshape(shape = var_1495, x = xt)[name = tensor("input")]; + tensor var_1497 = const()[name = tensor("op_1497"), val = tensor([-1])]; + tensor var_1498 = reduce_l2_norm(axes = var_1497, keep_dims = var_65, x = input)[name = tensor("op_1498")]; tensor const_42 = const()[name = tensor("const_42"), val = tensor(0x1.fffffep+127)]; - tensor clip_5 = clip(alpha = var_1012, beta = const_42, x = var_1465)[name = tensor("clip_5")]; - tensor var_1467 = real_div(x = input, y = clip_5)[name = tensor("op_1467")]; + tensor clip_5 = clip(alpha = var_79, beta = const_42, x = var_1498)[name = tensor("clip_5")]; + tensor var_1500 = real_div(x = input, y = clip_5)[name = tensor("op_1500")]; tensor concat_6 = const()[name = tensor("concat_6"), val = tensor([4, 1, 256])]; tensor reshape_0 = reshape(shape = concat_6, x = emb)[name = tensor("reshape_0")]; tensor transpose_1_perm_0 = const()[name = tensor("transpose_1_perm_0"), val = tensor([0, 1, 3, 2])]; tensor concat_7 = const()[name = tensor("concat_7"), val = tensor([4, 256, 9])]; - tensor transpose_1 = transpose(perm = transpose_1_perm_0, x = var_1467)[name = tensor("transpose_2")]; + tensor transpose_1 = transpose(perm = transpose_1_perm_0, x = var_1500)[name = tensor("transpose_2")]; tensor reshape_1 = reshape(shape = concat_7, x = transpose_1)[name = tensor("reshape_1")]; tensor matmul_2_transpose_x_0 = const()[name = tensor("matmul_2_transpose_x_0"), val = tensor(false)]; tensor matmul_2_transpose_y_0 = const()[name = tensor("matmul_2_transpose_y_0"), val = tensor(false)]; @@ -1317,10 +1332,10 @@ program(1.0) tensor output_end_0 = const()[name = tensor("output_end_0"), val = tensor([1, 4, 8])]; tensor output_end_mask_0 = const()[name = tensor("output_end_mask_0"), val = tensor([true, true, false])]; tensor output = slice_by_index(begin = output_begin_0, end = output_end_0, end_mask = output_end_mask_0, x = reshape_2)[name = tensor("output")]; - tensor probs = sigmoid(x = output)[name = tensor("op_1471")]; - tensor var_1473_axis_0 = const()[name = tensor("op_1473_axis_0"), val = tensor(0)]; - tensor dec_kv_new = stack(axis = var_1473_axis_0, values = (var_1169, nkv))[name = tensor("op_1473")]; - tensor var_1475_axis_0 = const()[name = tensor("op_1475_axis_0"), val = tensor(0)]; - tensor dec_scale_new = stack(axis = var_1475_axis_0, values = (new_scale_9, new_scale))[name = tensor("op_1475")]; + tensor probs = sigmoid(x = output)[name = tensor("op_1504")]; + tensor var_1506_axis_0 = const()[name = tensor("op_1506_axis_0"), val = tensor(0)]; + tensor dec_kv_new = stack(axis = var_1506_axis_0, values = (var_1202, nkv))[name = tensor("op_1506")]; + tensor var_1508_axis_0 = const()[name = tensor("op_1508_axis_0"), val = tensor(0)]; + tensor dec_scale_new = stack(axis = var_1508_axis_0, values = (new_scale_9, new_scale))[name = tensor("op_1508")]; } -> (probs, enc_kv_new, enc_scale_new, enc_conv_cache_new, cnn_window_new, dec_kv_new, dec_scale_new); } \ No newline at end of file diff --git a/optimized/ch/400ms/ls_eend_ch_400ms.mlpackage/Data/com.apple.CoreML/model.mlmodel b/optimized/ch/400ms/ls_eend_ch_400ms.mlpackage/Data/com.apple.CoreML/model.mlmodel index adac95464c7b8d573c30be0062638e5c2c0320d8..ed72efa3240836143eb2ec2a8f2cbc986306ded9 100644 --- a/optimized/ch/400ms/ls_eend_ch_400ms.mlpackage/Data/com.apple.CoreML/model.mlmodel +++ b/optimized/ch/400ms/ls_eend_ch_400ms.mlpackage/Data/com.apple.CoreML/model.mlmodel @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:182619e9c96f802ccae152cbda7d15ae2b22f480c04c689709f70aafdc13e145 -size 191044 +oid sha256:ea37aa7ff0809165f05c3e642105047e5ab12f90ec70be3299df1257bc98c557 +size 197116 diff --git a/optimized/ch/400ms/ls_eend_ch_400ms.mlpackage/Manifest.json b/optimized/ch/400ms/ls_eend_ch_400ms.mlpackage/Manifest.json index 36202097fcb62b8f48444015affe7358861ad8e4..95621140d601caa12673af33e82bbf0318b2be4e 100644 --- a/optimized/ch/400ms/ls_eend_ch_400ms.mlpackage/Manifest.json +++ b/optimized/ch/400ms/ls_eend_ch_400ms.mlpackage/Manifest.json @@ -1,18 +1,18 @@ { "fileFormatVersion": "1.0.0", "itemInfoEntries": { - "1C991CEA-BA44-4416-8759-2F7BBCD14BF6": { + "6C11BC4B-3170-4769-81B4-E5118EDB1B86": { "author": "com.apple.CoreML", "description": "CoreML Model Weights", "name": "weights", "path": "com.apple.CoreML/weights" }, - "676FBDCA-C6C7-4101-9A73-E68F17B5F2F1": { + "CF204ECD-C3A4-43DD-9F68-85A7874B4336": { "author": "com.apple.CoreML", "description": "CoreML Model Specification", "name": "model.mlmodel", "path": "com.apple.CoreML/model.mlmodel" } }, - "rootModelIdentifier": "676FBDCA-C6C7-4101-9A73-E68F17B5F2F1" + "rootModelIdentifier": "CF204ECD-C3A4-43DD-9F68-85A7874B4336" } diff --git a/optimized/ch/500ms/ls_eend_ch_500ms.mlmodelc/analytics/coremldata.bin b/optimized/ch/500ms/ls_eend_ch_500ms.mlmodelc/analytics/coremldata.bin index 3126a415e64304c0c88002baacbed295e38ad760..2c56ac9511ab6867e4dcc2bc80303dacc61cdec9 100644 --- a/optimized/ch/500ms/ls_eend_ch_500ms.mlmodelc/analytics/coremldata.bin +++ b/optimized/ch/500ms/ls_eend_ch_500ms.mlmodelc/analytics/coremldata.bin @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:631a472bbce1c64c900ce12bd152385af02044081a60d2b7634e29c403b9bc69 +oid sha256:8959c901ace7b3872ee7f214ad15c6993022f8cff276a7f35f918465bdd6cfe3 size 243 diff --git a/optimized/ch/500ms/ls_eend_ch_500ms.mlmodelc/coremldata.bin b/optimized/ch/500ms/ls_eend_ch_500ms.mlmodelc/coremldata.bin index 54dd550a42c104727ee2b06639b7b4ba3bc7bf83..f63b0a6d0db3ceb7d6600b3e6617798efb173804 100644 --- a/optimized/ch/500ms/ls_eend_ch_500ms.mlmodelc/coremldata.bin +++ b/optimized/ch/500ms/ls_eend_ch_500ms.mlmodelc/coremldata.bin @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:3cce9489e7b9d36fdbfd8fa6395dc2943f598f22d4ce2b28f349fff244c15703 -size 1301 +oid sha256:758464a340f78679e799f784a793d9900805b7b05ca4480cbbce6bdf1e0eea42 +size 1404 diff --git a/optimized/ch/500ms/ls_eend_ch_500ms.mlmodelc/metadata.json b/optimized/ch/500ms/ls_eend_ch_500ms.mlmodelc/metadata.json index 6c6b1d83c6b133bc9631ee243751876ec6e39cef..d9da2c68b4a00990419d3882ff7963f01c13d62f 100644 --- a/optimized/ch/500ms/ls_eend_ch_500ms.mlmodelc/metadata.json +++ b/optimized/ch/500ms/ls_eend_ch_500ms.mlmodelc/metadata.json @@ -1,7 +1,7 @@ [ { "metadataOutputVersion" : "3.0", - "shortDescription" : "LS-EEND CALLHOME streaming diarizer (pipeline, T=5, max_speakers=7)", + "shortDescription" : "LS-EEND CALLHOME streaming diarizer (pipeline, T=5, max_speakers=7, layout=raw_mel, cu=all)", "outputSchema" : [ { "hasShapeFlexibility" : "0", @@ -81,7 +81,7 @@ "specificationVersion" : 8, "mlProgramOperationTypeHistogram" : { "Ios16.reduceL2Norm" : 2, - "Ios17.reshape" : 68, + "Ios17.reshape" : 69, "Ios16.softmax" : 2, "Ios17.matmul" : 29, "Ios17.transpose" : 57, @@ -89,7 +89,7 @@ "Ios17.expandDims" : 3, "Ios17.add" : 46, "Ios16.sigmoid" : 5, - "Ios17.sliceByIndex" : 72, + "Ios17.sliceByIndex" : 77, "Tile" : 1, "Ios16.reduceSum" : 1, "Ios17.squeeze" : 2, @@ -101,7 +101,7 @@ "Ios16.silu" : 18, "Ios17.realDiv" : 20, "Ios17.linear" : 56, - "Stack" : 5, + "Stack" : 6, "Ios17.concat" : 26, "Ios16.relu" : 2, "Ios16.cumsum" : 1, @@ -128,9 +128,9 @@ "hasShapeFlexibility" : "0", "isOptional" : "0", "dataType" : "Float32", - "formattedType" : "MultiArray (Float32 1 × 5 × 345)", + "formattedType" : "MultiArray (Float32 1 × 55 × 23)", "shortDescription" : "", - "shape" : "[1, 5, 345]", + "shape" : "[1, 55, 23]", "name" : "features", "type" : "MultiArray" }, @@ -206,8 +206,8 @@ } ], "userDefinedMetadata" : { - "com.github.apple.coremltools.conversion_date" : "2026-04-16", - "config" : "{\"model_name\": \"ch\", \"model_label\": \"CALLHOME\", \"variant\": \"pipeline\", \"chunk_size\": 5, \"step_duration_ms\": 500, \"frame_duration_ms\": 100, \"frame_duration_seconds\": 0.1, \"max_speakers\": 7, \"max_nspks\": 9, \"n_units\": 256, \"n_heads\": 4, \"enc_n_layers\": 4, \"dec_n_layers\": 2, \"conv_kernel_size\": 16, \"conv_delay\": 9, \"sample_rate\": 8000, \"win_length\": 200, \"hop_length\": 80, \"n_mels\": 23, \"context_size\": 7, \"subsampling\": 10, \"feat_type\": \"logmel23_cummn\", \"pure_roll\": true}", + "com.github.apple.coremltools.conversion_date" : "2026-04-18", + "config" : "{\"model_name\": \"ch\", \"model_label\": \"CALLHOME\", \"variant\": \"pipeline\", \"chunk_size\": 5, \"step_duration_ms\": 500, \"frame_duration_ms\": 100, \"frame_duration_seconds\": 0.1, \"max_speakers\": 7, \"max_nspks\": 9, \"n_units\": 256, \"n_heads\": 4, \"enc_n_layers\": 4, \"dec_n_layers\": 2, \"conv_kernel_size\": 16, \"conv_delay\": 9, \"sample_rate\": 8000, \"win_length\": 200, \"hop_length\": 80, \"n_mels\": 23, \"context_size\": 7, \"subsampling\": 10, \"feat_type\": \"logmel23_cummn\", \"pure_roll\": true, \"input_layout\": \"raw_mel\", \"compute_units_export\": \"all\", \"raw_mel_length\": 55}", "com.github.apple.coremltools.source" : "torch==2.6.0", "com.github.apple.coremltools.version" : "9.0", "com.github.apple.coremltools.source_dialect" : "TorchScript" diff --git a/optimized/ch/500ms/ls_eend_ch_500ms.mlmodelc/model.mil b/optimized/ch/500ms/ls_eend_ch_500ms.mlmodelc/model.mil index fce20b41f60d6b106fa6a87a31e58eb7e064dc8b..23d359348d66fd1d2d9e467d139aa1b83f5a183a 100644 --- a/optimized/ch/500ms/ls_eend_ch_500ms.mlmodelc/model.mil +++ b/optimized/ch/500ms/ls_eend_ch_500ms.mlmodelc/model.mil @@ -1,234 +1,260 @@ program(1.0) [buildInfo = dict, tensor>({{"coremlc-component-MIL", "3510.2.1"}, {"coremlc-version", "3500.32.1"}, {"coremltools-component-torch", "2.6.0"}, {"coremltools-source-dialect", "TorchScript"}, {"coremltools-version", "9.0"}})] { - func main(tensor cnn_window, tensor dec_kv, tensor dec_scale, tensor enc_conv_cache, tensor enc_kv, tensor enc_scale, tensor features, tensor valid_mask) { - tensor encoder_cnn_bias = const()[name = tensor("encoder_cnn_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(64)))]; - tensor encoder_cnn_weight = const()[name = tensor("encoder_cnn_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1152)))]; - tensor encoder__causal_mask = const()[name = tensor("encoder__causal_mask"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4981952)))]; - tensor encoder__t_index = const()[name = tensor("encoder__t_index"), val = tensor([0x1p+0, 0x1p+1, 0x1.8p+1, 0x1p+2, 0x1.4p+2])]; - tensor encoder_input_projection_linear_bias = const()[name = tensor("encoder_input_projection_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4982144)))]; - tensor encoder_input_projection_linear_weight = const()[name = tensor("encoder_input_projection_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4983232)))]; - tensor encoder_pre_layer_norm_bias = const()[name = tensor("encoder_pre_layer_norm_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(5336576)))]; - tensor encoder_pre_layer_norm_weight = const()[name = tensor("encoder_pre_layer_norm_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(5337664)))]; - tensor encoder_ffn1_0_module_sequential_0_bias = const()[name = tensor("encoder_ffn1_0_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(5338752)))]; - tensor encoder_ffn1_0_module_sequential_0_weight = const()[name = tensor("encoder_ffn1_0_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(5339840)))]; - tensor encoder_ffn1_0_module_sequential_1_linear_bias = const()[name = tensor("encoder_ffn1_0_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(5340928)))]; - tensor encoder_ffn1_0_module_sequential_1_linear_weight = const()[name = tensor("encoder_ffn1_0_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(5345088)))]; - tensor encoder_ffn1_0_module_sequential_4_linear_bias = const()[name = tensor("encoder_ffn1_0_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(6393728)))]; - tensor encoder_ffn1_0_module_sequential_4_linear_weight = const()[name = tensor("encoder_ffn1_0_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(6394816)))]; - tensor encoder_ret_lns_0_bias = const()[name = tensor("encoder_ret_lns_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7443456)))]; - tensor encoder_ret_lns_0_weight = const()[name = tensor("encoder_ret_lns_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7444544)))]; - tensor encoder_q_proj_0_bias = const()[name = tensor("encoder_q_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7445632)))]; - tensor encoder_q_proj_0_weight = const()[name = tensor("encoder_q_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7446720)))]; - tensor encoder_k_proj_0_bias = const()[name = tensor("encoder_k_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7708928)))]; - tensor encoder_k_proj_0_weight = const()[name = tensor("encoder_k_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7710016)))]; - tensor encoder_v_proj_0_bias = const()[name = tensor("encoder_v_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7972224)))]; - tensor encoder_v_proj_0_weight = const()[name = tensor("encoder_v_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7973312)))]; - tensor encoder_g_proj_0_bias = const()[name = tensor("encoder_g_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8235520)))]; - tensor encoder_g_proj_0_weight = const()[name = tensor("encoder_g_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8236608)))]; - tensor encoder_out_proj_0_bias = const()[name = tensor("encoder_out_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8498816)))]; - tensor encoder_out_proj_0_weight = const()[name = tensor("encoder_out_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8499904)))]; - tensor encoder_conv_module_0_sequential_0_bias = const()[name = tensor("encoder_conv_module_0_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8762112)))]; - tensor encoder_conv_module_0_sequential_0_weight = const()[name = tensor("encoder_conv_module_0_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8763200)))]; - tensor encoder_conv_module_0_sequential_2_conv_bias = const()[name = tensor("encoder_conv_module_0_sequential_2_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8764288)))]; - tensor encoder_conv_module_0_sequential_2_conv_weight = const()[name = tensor("encoder_conv_module_0_sequential_2_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8766400)))]; - tensor encoder_conv_module_0_sequential_4_conv_weight = const()[name = tensor("encoder_conv_module_0_sequential_4_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9290752)))]; - tensor encoder_conv_module_0_sequential_5_running_var = const()[name = tensor("encoder_conv_module_0_sequential_5_running_var"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9307200)))]; - tensor encoder_conv_module_0_sequential_5_running_mean = const()[name = tensor("encoder_conv_module_0_sequential_5_running_mean"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9308288)))]; - tensor encoder_conv_module_0_sequential_5_bias = const()[name = tensor("encoder_conv_module_0_sequential_5_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9309376)))]; - tensor encoder_conv_module_0_sequential_5_weight = const()[name = tensor("encoder_conv_module_0_sequential_5_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9310464)))]; - tensor encoder_conv_module_0_sequential_7_conv_bias = const()[name = tensor("encoder_conv_module_0_sequential_7_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9311552)))]; - tensor encoder_conv_module_0_sequential_7_conv_weight = const()[name = tensor("encoder_conv_module_0_sequential_7_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9312640)))]; - tensor encoder_ffn2_0_module_sequential_0_bias = const()[name = tensor("encoder_ffn2_0_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9574848)))]; - tensor encoder_ffn2_0_module_sequential_0_weight = const()[name = tensor("encoder_ffn2_0_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9575936)))]; - tensor encoder_ffn2_0_module_sequential_1_linear_bias = const()[name = tensor("encoder_ffn2_0_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9577024)))]; - tensor encoder_ffn2_0_module_sequential_1_linear_weight = const()[name = tensor("encoder_ffn2_0_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9581184)))]; - tensor encoder_ffn2_0_module_sequential_4_linear_bias = const()[name = tensor("encoder_ffn2_0_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(10629824)))]; - tensor encoder_ffn2_0_module_sequential_4_linear_weight = const()[name = tensor("encoder_ffn2_0_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(10630912)))]; - tensor encoder_layer_norm_0_bias = const()[name = tensor("encoder_layer_norm_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11679552)))]; - tensor encoder_layer_norm_0_weight = const()[name = tensor("encoder_layer_norm_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11680640)))]; - tensor encoder_ffn1_1_module_sequential_0_bias = const()[name = tensor("encoder_ffn1_1_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11681728)))]; - tensor encoder_ffn1_1_module_sequential_0_weight = const()[name = tensor("encoder_ffn1_1_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11682816)))]; - tensor encoder_ffn1_1_module_sequential_1_linear_bias = const()[name = tensor("encoder_ffn1_1_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11683904)))]; - tensor encoder_ffn1_1_module_sequential_1_linear_weight = const()[name = tensor("encoder_ffn1_1_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11688064)))]; - tensor encoder_ffn1_1_module_sequential_4_linear_bias = const()[name = tensor("encoder_ffn1_1_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(12736704)))]; - tensor encoder_ffn1_1_module_sequential_4_linear_weight = const()[name = tensor("encoder_ffn1_1_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(12737792)))]; - tensor encoder_ret_lns_1_bias = const()[name = tensor("encoder_ret_lns_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(13786432)))]; - tensor encoder_ret_lns_1_weight = const()[name = tensor("encoder_ret_lns_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(13787520)))]; - tensor encoder_q_proj_1_bias = const()[name = tensor("encoder_q_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(13788608)))]; - tensor encoder_q_proj_1_weight = const()[name = tensor("encoder_q_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(13789696)))]; - tensor encoder_k_proj_1_bias = const()[name = tensor("encoder_k_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14051904)))]; - tensor encoder_k_proj_1_weight = const()[name = tensor("encoder_k_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14052992)))]; - tensor encoder_v_proj_1_bias = const()[name = tensor("encoder_v_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14315200)))]; - tensor encoder_v_proj_1_weight = const()[name = tensor("encoder_v_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14316288)))]; - tensor encoder_g_proj_1_bias = const()[name = tensor("encoder_g_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14578496)))]; - tensor encoder_g_proj_1_weight = const()[name = tensor("encoder_g_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14579584)))]; - tensor encoder_out_proj_1_bias = const()[name = tensor("encoder_out_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14841792)))]; - tensor encoder_out_proj_1_weight = const()[name = tensor("encoder_out_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14842880)))]; - tensor encoder_conv_module_1_sequential_0_bias = const()[name = tensor("encoder_conv_module_1_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15105088)))]; - tensor encoder_conv_module_1_sequential_0_weight = const()[name = tensor("encoder_conv_module_1_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15106176)))]; - tensor encoder_conv_module_1_sequential_2_conv_bias = const()[name = tensor("encoder_conv_module_1_sequential_2_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15107264)))]; - tensor encoder_conv_module_1_sequential_2_conv_weight = const()[name = tensor("encoder_conv_module_1_sequential_2_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15109376)))]; - tensor encoder_conv_module_1_sequential_4_conv_weight = const()[name = tensor("encoder_conv_module_1_sequential_4_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15633728)))]; - tensor encoder_conv_module_1_sequential_5_running_var = const()[name = tensor("encoder_conv_module_1_sequential_5_running_var"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15650176)))]; - tensor encoder_conv_module_1_sequential_5_running_mean = const()[name = tensor("encoder_conv_module_1_sequential_5_running_mean"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15651264)))]; - tensor encoder_conv_module_1_sequential_5_bias = const()[name = tensor("encoder_conv_module_1_sequential_5_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15652352)))]; - tensor encoder_conv_module_1_sequential_5_weight = const()[name = tensor("encoder_conv_module_1_sequential_5_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15653440)))]; - tensor encoder_conv_module_1_sequential_7_conv_bias = const()[name = tensor("encoder_conv_module_1_sequential_7_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15654528)))]; - tensor encoder_conv_module_1_sequential_7_conv_weight = const()[name = tensor("encoder_conv_module_1_sequential_7_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15655616)))]; - tensor encoder_ffn2_1_module_sequential_0_bias = const()[name = tensor("encoder_ffn2_1_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15917824)))]; - tensor encoder_ffn2_1_module_sequential_0_weight = const()[name = tensor("encoder_ffn2_1_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15918912)))]; - tensor encoder_ffn2_1_module_sequential_1_linear_bias = const()[name = tensor("encoder_ffn2_1_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15920000)))]; - tensor encoder_ffn2_1_module_sequential_1_linear_weight = const()[name = tensor("encoder_ffn2_1_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15924160)))]; - tensor encoder_ffn2_1_module_sequential_4_linear_bias = const()[name = tensor("encoder_ffn2_1_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(16972800)))]; - tensor encoder_ffn2_1_module_sequential_4_linear_weight = const()[name = tensor("encoder_ffn2_1_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(16973888)))]; - tensor encoder_layer_norm_1_bias = const()[name = tensor("encoder_layer_norm_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18022528)))]; - tensor encoder_layer_norm_1_weight = const()[name = tensor("encoder_layer_norm_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18023616)))]; - tensor encoder_ffn1_2_module_sequential_0_bias = const()[name = tensor("encoder_ffn1_2_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18024704)))]; - tensor encoder_ffn1_2_module_sequential_0_weight = const()[name = tensor("encoder_ffn1_2_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18025792)))]; - tensor encoder_ffn1_2_module_sequential_1_linear_bias = const()[name = tensor("encoder_ffn1_2_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18026880)))]; - tensor encoder_ffn1_2_module_sequential_1_linear_weight = const()[name = tensor("encoder_ffn1_2_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18031040)))]; - tensor encoder_ffn1_2_module_sequential_4_linear_bias = const()[name = tensor("encoder_ffn1_2_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(19079680)))]; - tensor encoder_ffn1_2_module_sequential_4_linear_weight = const()[name = tensor("encoder_ffn1_2_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(19080768)))]; - tensor encoder_ret_lns_2_bias = const()[name = tensor("encoder_ret_lns_2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20129408)))]; - tensor encoder_ret_lns_2_weight = const()[name = tensor("encoder_ret_lns_2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20130496)))]; - tensor encoder_q_proj_2_bias = const()[name = tensor("encoder_q_proj_2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20131584)))]; - tensor encoder_q_proj_2_weight = const()[name = tensor("encoder_q_proj_2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20132672)))]; - tensor encoder_k_proj_2_bias = const()[name = tensor("encoder_k_proj_2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20394880)))]; - tensor encoder_k_proj_2_weight = const()[name = tensor("encoder_k_proj_2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20395968)))]; - tensor encoder_v_proj_2_bias = const()[name = tensor("encoder_v_proj_2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20658176)))]; - tensor encoder_v_proj_2_weight = const()[name = tensor("encoder_v_proj_2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20659264)))]; - tensor encoder_g_proj_2_bias = const()[name = tensor("encoder_g_proj_2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20921472)))]; - tensor encoder_g_proj_2_weight = const()[name = tensor("encoder_g_proj_2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20922560)))]; - tensor encoder_out_proj_2_bias = const()[name = tensor("encoder_out_proj_2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21184768)))]; - tensor encoder_out_proj_2_weight = const()[name = tensor("encoder_out_proj_2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21185856)))]; - tensor encoder_conv_module_2_sequential_0_bias = const()[name = tensor("encoder_conv_module_2_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21448064)))]; - tensor encoder_conv_module_2_sequential_0_weight = const()[name = tensor("encoder_conv_module_2_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21449152)))]; - tensor encoder_conv_module_2_sequential_2_conv_bias = const()[name = tensor("encoder_conv_module_2_sequential_2_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21450240)))]; - tensor encoder_conv_module_2_sequential_2_conv_weight = const()[name = tensor("encoder_conv_module_2_sequential_2_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21452352)))]; - tensor encoder_conv_module_2_sequential_4_conv_weight = const()[name = tensor("encoder_conv_module_2_sequential_4_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21976704)))]; - tensor encoder_conv_module_2_sequential_5_running_var = const()[name = tensor("encoder_conv_module_2_sequential_5_running_var"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21993152)))]; - tensor encoder_conv_module_2_sequential_5_running_mean = const()[name = tensor("encoder_conv_module_2_sequential_5_running_mean"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21994240)))]; - tensor encoder_conv_module_2_sequential_5_bias = const()[name = tensor("encoder_conv_module_2_sequential_5_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21995328)))]; - tensor encoder_conv_module_2_sequential_5_weight = const()[name = tensor("encoder_conv_module_2_sequential_5_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21996416)))]; - tensor encoder_conv_module_2_sequential_7_conv_bias = const()[name = tensor("encoder_conv_module_2_sequential_7_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21997504)))]; - tensor encoder_conv_module_2_sequential_7_conv_weight = const()[name = tensor("encoder_conv_module_2_sequential_7_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21998592)))]; - tensor encoder_ffn2_2_module_sequential_0_bias = const()[name = tensor("encoder_ffn2_2_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(22260800)))]; - tensor encoder_ffn2_2_module_sequential_0_weight = const()[name = tensor("encoder_ffn2_2_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(22261888)))]; - tensor encoder_ffn2_2_module_sequential_1_linear_bias = const()[name = tensor("encoder_ffn2_2_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(22262976)))]; - tensor encoder_ffn2_2_module_sequential_1_linear_weight = const()[name = tensor("encoder_ffn2_2_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(22267136)))]; - tensor encoder_ffn2_2_module_sequential_4_linear_bias = const()[name = tensor("encoder_ffn2_2_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(23315776)))]; - tensor encoder_ffn2_2_module_sequential_4_linear_weight = const()[name = tensor("encoder_ffn2_2_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(23316864)))]; - tensor encoder_layer_norm_2_bias = const()[name = tensor("encoder_layer_norm_2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(24365504)))]; - tensor encoder_layer_norm_2_weight = const()[name = tensor("encoder_layer_norm_2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(24366592)))]; - tensor encoder_ffn1_3_module_sequential_0_bias = const()[name = tensor("encoder_ffn1_3_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(24367680)))]; - tensor encoder_ffn1_3_module_sequential_0_weight = const()[name = tensor("encoder_ffn1_3_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(24368768)))]; - tensor encoder_ffn1_3_module_sequential_1_linear_bias = const()[name = tensor("encoder_ffn1_3_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(24369856)))]; - tensor encoder_ffn1_3_module_sequential_1_linear_weight = const()[name = tensor("encoder_ffn1_3_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(24374016)))]; - tensor encoder_ffn1_3_module_sequential_4_linear_bias = const()[name = tensor("encoder_ffn1_3_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(25422656)))]; - tensor encoder_ffn1_3_module_sequential_4_linear_weight = const()[name = tensor("encoder_ffn1_3_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(25423744)))]; - tensor encoder_ret_lns_3_bias = const()[name = tensor("encoder_ret_lns_3_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(26472384)))]; - tensor encoder_ret_lns_3_weight = const()[name = tensor("encoder_ret_lns_3_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(26473472)))]; - tensor encoder_q_proj_3_bias = const()[name = tensor("encoder_q_proj_3_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(26474560)))]; - tensor encoder_q_proj_3_weight = const()[name = tensor("encoder_q_proj_3_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(26475648)))]; - tensor encoder_k_proj_3_bias = const()[name = tensor("encoder_k_proj_3_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(26737856)))]; - tensor encoder_k_proj_3_weight = const()[name = tensor("encoder_k_proj_3_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(26738944)))]; - tensor encoder_v_proj_3_bias = const()[name = tensor("encoder_v_proj_3_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27001152)))]; - tensor encoder_v_proj_3_weight = const()[name = tensor("encoder_v_proj_3_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27002240)))]; - tensor encoder_g_proj_3_bias = const()[name = tensor("encoder_g_proj_3_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27264448)))]; - tensor encoder_g_proj_3_weight = const()[name = tensor("encoder_g_proj_3_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27265536)))]; - tensor encoder_out_proj_3_bias = const()[name = tensor("encoder_out_proj_3_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27527744)))]; - tensor encoder_out_proj_3_weight = const()[name = tensor("encoder_out_proj_3_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27528832)))]; - tensor encoder_conv_module_3_sequential_0_bias = const()[name = tensor("encoder_conv_module_3_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27791040)))]; - tensor encoder_conv_module_3_sequential_0_weight = const()[name = tensor("encoder_conv_module_3_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27792128)))]; - tensor encoder_conv_module_3_sequential_2_conv_bias = const()[name = tensor("encoder_conv_module_3_sequential_2_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27793216)))]; - tensor encoder_conv_module_3_sequential_2_conv_weight = const()[name = tensor("encoder_conv_module_3_sequential_2_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27795328)))]; - tensor encoder_conv_module_3_sequential_4_conv_weight = const()[name = tensor("encoder_conv_module_3_sequential_4_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28319680)))]; - tensor encoder_conv_module_3_sequential_5_running_var = const()[name = tensor("encoder_conv_module_3_sequential_5_running_var"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28336128)))]; - tensor encoder_conv_module_3_sequential_5_running_mean = const()[name = tensor("encoder_conv_module_3_sequential_5_running_mean"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28337216)))]; - tensor encoder_conv_module_3_sequential_5_bias = const()[name = tensor("encoder_conv_module_3_sequential_5_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28338304)))]; - tensor encoder_conv_module_3_sequential_5_weight = const()[name = tensor("encoder_conv_module_3_sequential_5_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28339392)))]; - tensor encoder_conv_module_3_sequential_7_conv_bias = const()[name = tensor("encoder_conv_module_3_sequential_7_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28340480)))]; - tensor encoder_conv_module_3_sequential_7_conv_weight = const()[name = tensor("encoder_conv_module_3_sequential_7_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28341568)))]; - tensor encoder_ffn2_3_module_sequential_0_bias = const()[name = tensor("encoder_ffn2_3_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28603776)))]; - tensor encoder_ffn2_3_module_sequential_0_weight = const()[name = tensor("encoder_ffn2_3_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28604864)))]; - tensor encoder_ffn2_3_module_sequential_1_linear_bias = const()[name = tensor("encoder_ffn2_3_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28605952)))]; - tensor encoder_ffn2_3_module_sequential_1_linear_weight = const()[name = tensor("encoder_ffn2_3_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28610112)))]; - tensor encoder_ffn2_3_module_sequential_4_linear_bias = const()[name = tensor("encoder_ffn2_3_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(29658752)))]; - tensor encoder_ffn2_3_module_sequential_4_linear_weight = const()[name = tensor("encoder_ffn2_3_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(29659840)))]; - tensor encoder_layer_norm_3_bias = const()[name = tensor("encoder_layer_norm_3_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(30708480)))]; - tensor encoder_layer_norm_3_weight = const()[name = tensor("encoder_layer_norm_3_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(30709568)))]; - tensor decoder__causal_mask = const()[name = tensor("decoder__causal_mask"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(30710656)))]; - tensor decoder_convert_bias = const()[name = tensor("decoder_convert_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(30710848)))]; - tensor decoder_convert_weight = const()[name = tensor("decoder_convert_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(30711936)))]; - tensor decoder_q_proj_0_bias = const()[name = tensor("decoder_q_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(31236288)))]; - tensor decoder_q_proj_0_weight = const()[name = tensor("decoder_q_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(31237376)))]; - tensor decoder_k_proj_0_bias = const()[name = tensor("decoder_k_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(31499584)))]; - tensor decoder_k_proj_0_weight = const()[name = tensor("decoder_k_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(31500672)))]; - tensor decoder_v_proj_0_bias = const()[name = tensor("decoder_v_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(31762880)))]; - tensor decoder_v_proj_0_weight = const()[name = tensor("decoder_v_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(31763968)))]; - tensor decoder_g_proj_0_bias = const()[name = tensor("decoder_g_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32026176)))]; - tensor decoder_g_proj_0_weight = const()[name = tensor("decoder_g_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32027264)))]; - tensor decoder_out_proj_0_bias = const()[name = tensor("decoder_out_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32289472)))]; - tensor decoder_out_proj_0_weight = const()[name = tensor("decoder_out_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32290560)))]; - tensor decoder_norm11_0_bias = const()[name = tensor("decoder_norm11_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32552768)))]; - tensor decoder_norm11_0_weight = const()[name = tensor("decoder_norm11_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32553856)))]; - tensor decoder_self_attn2_0_out_proj_bias = const()[name = tensor("decoder_self_attn2_0_out_proj_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32554944)))]; - tensor decoder_self_attn2_0_out_proj_weight = const()[name = tensor("decoder_self_attn2_0_out_proj_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32556032)))]; - tensor decoder_self_attn2_0_in_proj_bias = const()[name = tensor("decoder_self_attn2_0_in_proj_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32818240)))]; - tensor decoder_self_attn2_0_in_proj_weight = const()[name = tensor("decoder_self_attn2_0_in_proj_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32821376)))]; - tensor decoder_norm21_0_bias = const()[name = tensor("decoder_norm21_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(33607872)))]; - tensor decoder_norm21_0_weight = const()[name = tensor("decoder_norm21_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(33608960)))]; - tensor decoder_linear1_0_bias = const()[name = tensor("decoder_linear1_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(33610048)))]; - tensor decoder_linear1_0_weight = const()[name = tensor("decoder_linear1_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(33618304)))]; - tensor decoder_linear2_0_bias = const()[name = tensor("decoder_linear2_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(35715520)))]; - tensor decoder_linear2_0_weight = const()[name = tensor("decoder_linear2_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(35716608)))]; - tensor decoder_norm22_0_bias = const()[name = tensor("decoder_norm22_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(37813824)))]; - tensor decoder_norm22_0_weight = const()[name = tensor("decoder_norm22_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(37814912)))]; - tensor decoder_q_proj_1_bias = const()[name = tensor("decoder_q_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(37816000)))]; - tensor decoder_q_proj_1_weight = const()[name = tensor("decoder_q_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(37817088)))]; - tensor decoder_k_proj_1_bias = const()[name = tensor("decoder_k_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38079296)))]; - tensor decoder_k_proj_1_weight = const()[name = tensor("decoder_k_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38080384)))]; - tensor decoder_v_proj_1_bias = const()[name = tensor("decoder_v_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38342592)))]; - tensor decoder_v_proj_1_weight = const()[name = tensor("decoder_v_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38343680)))]; - tensor decoder_g_proj_1_bias = const()[name = tensor("decoder_g_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38605888)))]; - tensor decoder_g_proj_1_weight = const()[name = tensor("decoder_g_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38606976)))]; - tensor decoder_out_proj_1_bias = const()[name = tensor("decoder_out_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38869184)))]; - tensor decoder_out_proj_1_weight = const()[name = tensor("decoder_out_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38870272)))]; - tensor decoder_norm11_1_bias = const()[name = tensor("decoder_norm11_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(39132480)))]; - tensor decoder_norm11_1_weight = const()[name = tensor("decoder_norm11_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(39133568)))]; - tensor decoder_self_attn2_1_out_proj_bias = const()[name = tensor("decoder_self_attn2_1_out_proj_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(39134656)))]; - tensor decoder_self_attn2_1_out_proj_weight = const()[name = tensor("decoder_self_attn2_1_out_proj_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(39135744)))]; - tensor decoder_self_attn2_1_in_proj_bias = const()[name = tensor("decoder_self_attn2_1_in_proj_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(39397952)))]; - tensor decoder_self_attn2_1_in_proj_weight = const()[name = tensor("decoder_self_attn2_1_in_proj_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(39401088)))]; - tensor decoder_norm21_1_bias = const()[name = tensor("decoder_norm21_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(40187584)))]; - tensor decoder_norm21_1_weight = const()[name = tensor("decoder_norm21_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(40188672)))]; - tensor decoder_linear1_1_bias = const()[name = tensor("decoder_linear1_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(40189760)))]; - tensor decoder_linear1_1_weight = const()[name = tensor("decoder_linear1_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(40198016)))]; - tensor decoder_linear2_1_bias = const()[name = tensor("decoder_linear2_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(42295232)))]; - tensor decoder_linear2_1_weight = const()[name = tensor("decoder_linear2_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(42296320)))]; - tensor decoder_norm22_1_bias = const()[name = tensor("decoder_norm22_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(44393536)))]; - tensor decoder_norm22_1_weight = const()[name = tensor("decoder_norm22_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(44394624)))]; - tensor var_11 = const()[name = tensor("op_11"), val = tensor(0x1.5798eep-27)]; - tensor var_18 = const()[name = tensor("op_18"), val = tensor(0x1.0c6f7ap-20)]; - tensor var_21 = const()[name = tensor("op_21"), val = tensor(2)]; - tensor var_24 = const()[name = tensor("op_24"), val = tensor(0)]; - tensor var_27 = const()[name = tensor("op_27"), val = tensor(1)]; - tensor var_29 = const()[name = tensor("op_29"), val = tensor(0x1.4f8b58p-17)]; - tensor var_30 = const()[name = tensor("op_30"), val = tensor(true)]; - tensor input_1 = linear(bias = encoder_input_projection_linear_bias, weight = encoder_input_projection_linear_weight, x = features)[name = tensor("linear_0")]; - tensor input_3_axes_0 = const()[name = tensor("input_3_axes_0"), val = tensor([-1])]; - tensor input_3 = layer_norm(axes = input_3_axes_0, beta = encoder_pre_layer_norm_bias, epsilon = var_29, gamma = encoder_pre_layer_norm_weight, x = input_1)[name = tensor("input_3")]; + func main(tensor cnn_window, tensor dec_kv, tensor dec_scale, tensor enc_conv_cache, tensor enc_kv, tensor enc_scale, tensor features, tensor valid_mask) { + tensor inner_encoder_cnn_bias = const()[name = tensor("inner_encoder_cnn_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(64)))]; + tensor inner_encoder_cnn_weight = const()[name = tensor("inner_encoder_cnn_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1152)))]; + tensor inner_encoder__causal_mask = const()[name = tensor("inner_encoder__causal_mask"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4981952)))]; + tensor inner_encoder__t_index = const()[name = tensor("inner_encoder__t_index"), val = tensor([0x1p+0, 0x1p+1, 0x1.8p+1, 0x1p+2, 0x1.4p+2])]; + tensor inner_encoder_input_projection_linear_bias = const()[name = tensor("inner_encoder_input_projection_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4982144)))]; + tensor inner_encoder_input_projection_linear_weight = const()[name = tensor("inner_encoder_input_projection_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4983232)))]; + tensor inner_encoder_pre_layer_norm_bias = const()[name = tensor("inner_encoder_pre_layer_norm_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(5336576)))]; + tensor inner_encoder_pre_layer_norm_weight = const()[name = tensor("inner_encoder_pre_layer_norm_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(5337664)))]; + tensor inner_encoder_ffn1_0_module_sequential_0_bias = const()[name = tensor("inner_encoder_ffn1_0_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(5338752)))]; + tensor inner_encoder_ffn1_0_module_sequential_0_weight = const()[name = tensor("inner_encoder_ffn1_0_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(5339840)))]; + tensor inner_encoder_ffn1_0_module_sequential_1_linear_bias = const()[name = tensor("inner_encoder_ffn1_0_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(5340928)))]; + tensor inner_encoder_ffn1_0_module_sequential_1_linear_weight = const()[name = tensor("inner_encoder_ffn1_0_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(5345088)))]; + tensor inner_encoder_ffn1_0_module_sequential_4_linear_bias = const()[name = tensor("inner_encoder_ffn1_0_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(6393728)))]; + tensor inner_encoder_ffn1_0_module_sequential_4_linear_weight = const()[name = tensor("inner_encoder_ffn1_0_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(6394816)))]; + tensor inner_encoder_ret_lns_0_bias = const()[name = tensor("inner_encoder_ret_lns_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7443456)))]; + tensor inner_encoder_ret_lns_0_weight = const()[name = tensor("inner_encoder_ret_lns_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7444544)))]; + tensor inner_encoder_q_proj_0_bias = const()[name = tensor("inner_encoder_q_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7445632)))]; + tensor inner_encoder_q_proj_0_weight = const()[name = tensor("inner_encoder_q_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7446720)))]; + tensor inner_encoder_k_proj_0_bias = const()[name = tensor("inner_encoder_k_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7708928)))]; + tensor inner_encoder_k_proj_0_weight = const()[name = tensor("inner_encoder_k_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7710016)))]; + tensor inner_encoder_v_proj_0_bias = const()[name = tensor("inner_encoder_v_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7972224)))]; + tensor inner_encoder_v_proj_0_weight = const()[name = tensor("inner_encoder_v_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7973312)))]; + tensor inner_encoder_g_proj_0_bias = const()[name = tensor("inner_encoder_g_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8235520)))]; + tensor inner_encoder_g_proj_0_weight = const()[name = tensor("inner_encoder_g_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8236608)))]; + tensor inner_encoder_out_proj_0_bias = const()[name = tensor("inner_encoder_out_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8498816)))]; + tensor inner_encoder_out_proj_0_weight = const()[name = tensor("inner_encoder_out_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8499904)))]; + tensor inner_encoder_conv_module_0_sequential_0_bias = const()[name = tensor("inner_encoder_conv_module_0_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8762112)))]; + tensor inner_encoder_conv_module_0_sequential_0_weight = const()[name = tensor("inner_encoder_conv_module_0_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8763200)))]; + tensor inner_encoder_conv_module_0_sequential_2_conv_bias = const()[name = tensor("inner_encoder_conv_module_0_sequential_2_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8764288)))]; + tensor inner_encoder_conv_module_0_sequential_2_conv_weight = const()[name = tensor("inner_encoder_conv_module_0_sequential_2_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8766400)))]; + tensor inner_encoder_conv_module_0_sequential_4_conv_weight = const()[name = tensor("inner_encoder_conv_module_0_sequential_4_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9290752)))]; + tensor inner_encoder_conv_module_0_sequential_5_running_var = const()[name = tensor("inner_encoder_conv_module_0_sequential_5_running_var"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9307200)))]; + tensor inner_encoder_conv_module_0_sequential_5_running_mean = const()[name = tensor("inner_encoder_conv_module_0_sequential_5_running_mean"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9308288)))]; + tensor inner_encoder_conv_module_0_sequential_5_bias = const()[name = tensor("inner_encoder_conv_module_0_sequential_5_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9309376)))]; + tensor inner_encoder_conv_module_0_sequential_5_weight = const()[name = tensor("inner_encoder_conv_module_0_sequential_5_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9310464)))]; + tensor inner_encoder_conv_module_0_sequential_7_conv_bias = const()[name = tensor("inner_encoder_conv_module_0_sequential_7_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9311552)))]; + tensor inner_encoder_conv_module_0_sequential_7_conv_weight = const()[name = tensor("inner_encoder_conv_module_0_sequential_7_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9312640)))]; + tensor inner_encoder_ffn2_0_module_sequential_0_bias = const()[name = tensor("inner_encoder_ffn2_0_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9574848)))]; + tensor inner_encoder_ffn2_0_module_sequential_0_weight = const()[name = tensor("inner_encoder_ffn2_0_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9575936)))]; + tensor inner_encoder_ffn2_0_module_sequential_1_linear_bias = const()[name = tensor("inner_encoder_ffn2_0_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9577024)))]; + tensor inner_encoder_ffn2_0_module_sequential_1_linear_weight = const()[name = tensor("inner_encoder_ffn2_0_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9581184)))]; + tensor inner_encoder_ffn2_0_module_sequential_4_linear_bias = const()[name = tensor("inner_encoder_ffn2_0_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(10629824)))]; + tensor inner_encoder_ffn2_0_module_sequential_4_linear_weight = const()[name = tensor("inner_encoder_ffn2_0_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(10630912)))]; + tensor inner_encoder_layer_norm_0_bias = const()[name = tensor("inner_encoder_layer_norm_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11679552)))]; + tensor inner_encoder_layer_norm_0_weight = const()[name = tensor("inner_encoder_layer_norm_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11680640)))]; + tensor inner_encoder_ffn1_1_module_sequential_0_bias = const()[name = tensor("inner_encoder_ffn1_1_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11681728)))]; + tensor inner_encoder_ffn1_1_module_sequential_0_weight = const()[name = tensor("inner_encoder_ffn1_1_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11682816)))]; + tensor inner_encoder_ffn1_1_module_sequential_1_linear_bias = const()[name = tensor("inner_encoder_ffn1_1_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11683904)))]; + tensor inner_encoder_ffn1_1_module_sequential_1_linear_weight = const()[name = tensor("inner_encoder_ffn1_1_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11688064)))]; + tensor inner_encoder_ffn1_1_module_sequential_4_linear_bias = const()[name = tensor("inner_encoder_ffn1_1_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(12736704)))]; + tensor inner_encoder_ffn1_1_module_sequential_4_linear_weight = const()[name = tensor("inner_encoder_ffn1_1_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(12737792)))]; + tensor inner_encoder_ret_lns_1_bias = const()[name = tensor("inner_encoder_ret_lns_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(13786432)))]; + tensor inner_encoder_ret_lns_1_weight = const()[name = tensor("inner_encoder_ret_lns_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(13787520)))]; + tensor inner_encoder_q_proj_1_bias = const()[name = tensor("inner_encoder_q_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(13788608)))]; + tensor inner_encoder_q_proj_1_weight = const()[name = tensor("inner_encoder_q_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(13789696)))]; + tensor inner_encoder_k_proj_1_bias = const()[name = tensor("inner_encoder_k_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14051904)))]; + tensor inner_encoder_k_proj_1_weight = const()[name = tensor("inner_encoder_k_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14052992)))]; + tensor inner_encoder_v_proj_1_bias = const()[name = tensor("inner_encoder_v_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14315200)))]; + tensor inner_encoder_v_proj_1_weight = const()[name = tensor("inner_encoder_v_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14316288)))]; + tensor inner_encoder_g_proj_1_bias = const()[name = tensor("inner_encoder_g_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14578496)))]; + tensor inner_encoder_g_proj_1_weight = const()[name = tensor("inner_encoder_g_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14579584)))]; + tensor inner_encoder_out_proj_1_bias = const()[name = tensor("inner_encoder_out_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14841792)))]; + tensor inner_encoder_out_proj_1_weight = const()[name = tensor("inner_encoder_out_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14842880)))]; + tensor inner_encoder_conv_module_1_sequential_0_bias = const()[name = tensor("inner_encoder_conv_module_1_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15105088)))]; + tensor inner_encoder_conv_module_1_sequential_0_weight = const()[name = tensor("inner_encoder_conv_module_1_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15106176)))]; + tensor inner_encoder_conv_module_1_sequential_2_conv_bias = const()[name = tensor("inner_encoder_conv_module_1_sequential_2_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15107264)))]; + tensor inner_encoder_conv_module_1_sequential_2_conv_weight = const()[name = tensor("inner_encoder_conv_module_1_sequential_2_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15109376)))]; + tensor inner_encoder_conv_module_1_sequential_4_conv_weight = const()[name = tensor("inner_encoder_conv_module_1_sequential_4_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15633728)))]; + tensor inner_encoder_conv_module_1_sequential_5_running_var = const()[name = tensor("inner_encoder_conv_module_1_sequential_5_running_var"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15650176)))]; + tensor inner_encoder_conv_module_1_sequential_5_running_mean = const()[name = tensor("inner_encoder_conv_module_1_sequential_5_running_mean"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15651264)))]; + tensor inner_encoder_conv_module_1_sequential_5_bias = const()[name = tensor("inner_encoder_conv_module_1_sequential_5_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15652352)))]; + tensor inner_encoder_conv_module_1_sequential_5_weight = const()[name = tensor("inner_encoder_conv_module_1_sequential_5_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15653440)))]; + tensor inner_encoder_conv_module_1_sequential_7_conv_bias = const()[name = tensor("inner_encoder_conv_module_1_sequential_7_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15654528)))]; + tensor inner_encoder_conv_module_1_sequential_7_conv_weight = const()[name = tensor("inner_encoder_conv_module_1_sequential_7_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15655616)))]; + tensor inner_encoder_ffn2_1_module_sequential_0_bias = const()[name = tensor("inner_encoder_ffn2_1_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15917824)))]; + tensor inner_encoder_ffn2_1_module_sequential_0_weight = const()[name = tensor("inner_encoder_ffn2_1_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15918912)))]; + tensor inner_encoder_ffn2_1_module_sequential_1_linear_bias = const()[name = tensor("inner_encoder_ffn2_1_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15920000)))]; + tensor inner_encoder_ffn2_1_module_sequential_1_linear_weight = const()[name = tensor("inner_encoder_ffn2_1_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15924160)))]; + tensor inner_encoder_ffn2_1_module_sequential_4_linear_bias = const()[name = tensor("inner_encoder_ffn2_1_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(16972800)))]; + tensor inner_encoder_ffn2_1_module_sequential_4_linear_weight = const()[name = tensor("inner_encoder_ffn2_1_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(16973888)))]; + tensor inner_encoder_layer_norm_1_bias = const()[name = tensor("inner_encoder_layer_norm_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18022528)))]; + tensor inner_encoder_layer_norm_1_weight = const()[name = tensor("inner_encoder_layer_norm_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18023616)))]; + tensor inner_encoder_ffn1_2_module_sequential_0_bias = const()[name = tensor("inner_encoder_ffn1_2_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18024704)))]; + tensor inner_encoder_ffn1_2_module_sequential_0_weight = const()[name = tensor("inner_encoder_ffn1_2_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18025792)))]; + tensor inner_encoder_ffn1_2_module_sequential_1_linear_bias = const()[name = tensor("inner_encoder_ffn1_2_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18026880)))]; + tensor inner_encoder_ffn1_2_module_sequential_1_linear_weight = const()[name = tensor("inner_encoder_ffn1_2_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18031040)))]; + tensor inner_encoder_ffn1_2_module_sequential_4_linear_bias = const()[name = tensor("inner_encoder_ffn1_2_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(19079680)))]; + tensor inner_encoder_ffn1_2_module_sequential_4_linear_weight = const()[name = tensor("inner_encoder_ffn1_2_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(19080768)))]; + tensor inner_encoder_ret_lns_2_bias = const()[name = tensor("inner_encoder_ret_lns_2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20129408)))]; + tensor inner_encoder_ret_lns_2_weight = const()[name = tensor("inner_encoder_ret_lns_2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20130496)))]; + tensor inner_encoder_q_proj_2_bias = const()[name = tensor("inner_encoder_q_proj_2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20131584)))]; + tensor inner_encoder_q_proj_2_weight = const()[name = tensor("inner_encoder_q_proj_2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20132672)))]; + tensor inner_encoder_k_proj_2_bias = const()[name = tensor("inner_encoder_k_proj_2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20394880)))]; + tensor inner_encoder_k_proj_2_weight = const()[name = tensor("inner_encoder_k_proj_2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20395968)))]; + tensor inner_encoder_v_proj_2_bias = const()[name = tensor("inner_encoder_v_proj_2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20658176)))]; + tensor inner_encoder_v_proj_2_weight = const()[name = tensor("inner_encoder_v_proj_2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20659264)))]; + tensor inner_encoder_g_proj_2_bias = const()[name = tensor("inner_encoder_g_proj_2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20921472)))]; + tensor inner_encoder_g_proj_2_weight = const()[name = tensor("inner_encoder_g_proj_2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20922560)))]; + tensor inner_encoder_out_proj_2_bias = const()[name = tensor("inner_encoder_out_proj_2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21184768)))]; + tensor inner_encoder_out_proj_2_weight = const()[name = tensor("inner_encoder_out_proj_2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21185856)))]; + tensor inner_encoder_conv_module_2_sequential_0_bias = const()[name = tensor("inner_encoder_conv_module_2_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21448064)))]; + tensor inner_encoder_conv_module_2_sequential_0_weight = const()[name = tensor("inner_encoder_conv_module_2_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21449152)))]; + tensor inner_encoder_conv_module_2_sequential_2_conv_bias = const()[name = tensor("inner_encoder_conv_module_2_sequential_2_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21450240)))]; + tensor inner_encoder_conv_module_2_sequential_2_conv_weight = const()[name = tensor("inner_encoder_conv_module_2_sequential_2_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21452352)))]; + tensor inner_encoder_conv_module_2_sequential_4_conv_weight = const()[name = tensor("inner_encoder_conv_module_2_sequential_4_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21976704)))]; + tensor inner_encoder_conv_module_2_sequential_5_running_var = const()[name = tensor("inner_encoder_conv_module_2_sequential_5_running_var"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21993152)))]; + tensor inner_encoder_conv_module_2_sequential_5_running_mean = const()[name = tensor("inner_encoder_conv_module_2_sequential_5_running_mean"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21994240)))]; + tensor inner_encoder_conv_module_2_sequential_5_bias = const()[name = tensor("inner_encoder_conv_module_2_sequential_5_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21995328)))]; + tensor inner_encoder_conv_module_2_sequential_5_weight = const()[name = tensor("inner_encoder_conv_module_2_sequential_5_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21996416)))]; + tensor inner_encoder_conv_module_2_sequential_7_conv_bias = const()[name = tensor("inner_encoder_conv_module_2_sequential_7_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21997504)))]; + tensor inner_encoder_conv_module_2_sequential_7_conv_weight = const()[name = tensor("inner_encoder_conv_module_2_sequential_7_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21998592)))]; + tensor inner_encoder_ffn2_2_module_sequential_0_bias = const()[name = tensor("inner_encoder_ffn2_2_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(22260800)))]; + tensor inner_encoder_ffn2_2_module_sequential_0_weight = const()[name = tensor("inner_encoder_ffn2_2_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(22261888)))]; + tensor inner_encoder_ffn2_2_module_sequential_1_linear_bias = const()[name = tensor("inner_encoder_ffn2_2_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(22262976)))]; + tensor inner_encoder_ffn2_2_module_sequential_1_linear_weight = const()[name = tensor("inner_encoder_ffn2_2_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(22267136)))]; + tensor inner_encoder_ffn2_2_module_sequential_4_linear_bias = const()[name = tensor("inner_encoder_ffn2_2_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(23315776)))]; + tensor inner_encoder_ffn2_2_module_sequential_4_linear_weight = const()[name = tensor("inner_encoder_ffn2_2_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(23316864)))]; + tensor inner_encoder_layer_norm_2_bias = const()[name = tensor("inner_encoder_layer_norm_2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(24365504)))]; + tensor inner_encoder_layer_norm_2_weight = const()[name = tensor("inner_encoder_layer_norm_2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(24366592)))]; + tensor inner_encoder_ffn1_3_module_sequential_0_bias = const()[name = tensor("inner_encoder_ffn1_3_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(24367680)))]; + tensor inner_encoder_ffn1_3_module_sequential_0_weight = const()[name = tensor("inner_encoder_ffn1_3_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(24368768)))]; + tensor inner_encoder_ffn1_3_module_sequential_1_linear_bias = const()[name = tensor("inner_encoder_ffn1_3_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(24369856)))]; + tensor inner_encoder_ffn1_3_module_sequential_1_linear_weight = const()[name = tensor("inner_encoder_ffn1_3_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(24374016)))]; + tensor inner_encoder_ffn1_3_module_sequential_4_linear_bias = const()[name = tensor("inner_encoder_ffn1_3_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(25422656)))]; + tensor inner_encoder_ffn1_3_module_sequential_4_linear_weight = const()[name = tensor("inner_encoder_ffn1_3_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(25423744)))]; + tensor inner_encoder_ret_lns_3_bias = const()[name = tensor("inner_encoder_ret_lns_3_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(26472384)))]; + tensor inner_encoder_ret_lns_3_weight = const()[name = tensor("inner_encoder_ret_lns_3_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(26473472)))]; + tensor inner_encoder_q_proj_3_bias = const()[name = tensor("inner_encoder_q_proj_3_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(26474560)))]; + tensor inner_encoder_q_proj_3_weight = const()[name = tensor("inner_encoder_q_proj_3_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(26475648)))]; + tensor inner_encoder_k_proj_3_bias = const()[name = tensor("inner_encoder_k_proj_3_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(26737856)))]; + tensor inner_encoder_k_proj_3_weight = const()[name = tensor("inner_encoder_k_proj_3_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(26738944)))]; + tensor inner_encoder_v_proj_3_bias = const()[name = tensor("inner_encoder_v_proj_3_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27001152)))]; + tensor inner_encoder_v_proj_3_weight = const()[name = tensor("inner_encoder_v_proj_3_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27002240)))]; + tensor inner_encoder_g_proj_3_bias = const()[name = tensor("inner_encoder_g_proj_3_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27264448)))]; + tensor inner_encoder_g_proj_3_weight = const()[name = tensor("inner_encoder_g_proj_3_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27265536)))]; + tensor inner_encoder_out_proj_3_bias = const()[name = tensor("inner_encoder_out_proj_3_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27527744)))]; + tensor inner_encoder_out_proj_3_weight = const()[name = tensor("inner_encoder_out_proj_3_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27528832)))]; + tensor inner_encoder_conv_module_3_sequential_0_bias = const()[name = tensor("inner_encoder_conv_module_3_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27791040)))]; + tensor inner_encoder_conv_module_3_sequential_0_weight = const()[name = tensor("inner_encoder_conv_module_3_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27792128)))]; + tensor inner_encoder_conv_module_3_sequential_2_conv_bias = const()[name = tensor("inner_encoder_conv_module_3_sequential_2_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27793216)))]; + tensor inner_encoder_conv_module_3_sequential_2_conv_weight = const()[name = tensor("inner_encoder_conv_module_3_sequential_2_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27795328)))]; + tensor inner_encoder_conv_module_3_sequential_4_conv_weight = const()[name = tensor("inner_encoder_conv_module_3_sequential_4_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28319680)))]; + tensor inner_encoder_conv_module_3_sequential_5_running_var = const()[name = tensor("inner_encoder_conv_module_3_sequential_5_running_var"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28336128)))]; + tensor inner_encoder_conv_module_3_sequential_5_running_mean = const()[name = tensor("inner_encoder_conv_module_3_sequential_5_running_mean"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28337216)))]; + tensor inner_encoder_conv_module_3_sequential_5_bias = const()[name = tensor("inner_encoder_conv_module_3_sequential_5_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28338304)))]; + tensor inner_encoder_conv_module_3_sequential_5_weight = const()[name = tensor("inner_encoder_conv_module_3_sequential_5_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28339392)))]; + tensor inner_encoder_conv_module_3_sequential_7_conv_bias = const()[name = tensor("inner_encoder_conv_module_3_sequential_7_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28340480)))]; + tensor inner_encoder_conv_module_3_sequential_7_conv_weight = const()[name = tensor("inner_encoder_conv_module_3_sequential_7_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28341568)))]; + tensor inner_encoder_ffn2_3_module_sequential_0_bias = const()[name = tensor("inner_encoder_ffn2_3_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28603776)))]; + tensor inner_encoder_ffn2_3_module_sequential_0_weight = const()[name = tensor("inner_encoder_ffn2_3_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28604864)))]; + tensor inner_encoder_ffn2_3_module_sequential_1_linear_bias = const()[name = tensor("inner_encoder_ffn2_3_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28605952)))]; + tensor inner_encoder_ffn2_3_module_sequential_1_linear_weight = const()[name = tensor("inner_encoder_ffn2_3_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28610112)))]; + tensor inner_encoder_ffn2_3_module_sequential_4_linear_bias = const()[name = tensor("inner_encoder_ffn2_3_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(29658752)))]; + tensor inner_encoder_ffn2_3_module_sequential_4_linear_weight = const()[name = tensor("inner_encoder_ffn2_3_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(29659840)))]; + tensor inner_encoder_layer_norm_3_bias = const()[name = tensor("inner_encoder_layer_norm_3_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(30708480)))]; + tensor inner_encoder_layer_norm_3_weight = const()[name = tensor("inner_encoder_layer_norm_3_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(30709568)))]; + tensor inner_decoder__causal_mask = const()[name = tensor("inner_decoder__causal_mask"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(30710656)))]; + tensor inner_decoder_convert_bias = const()[name = tensor("inner_decoder_convert_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(30710848)))]; + tensor inner_decoder_convert_weight = const()[name = tensor("inner_decoder_convert_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(30711936)))]; + tensor inner_decoder_q_proj_0_bias = const()[name = tensor("inner_decoder_q_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(31236288)))]; + tensor inner_decoder_q_proj_0_weight = const()[name = tensor("inner_decoder_q_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(31237376)))]; + tensor inner_decoder_k_proj_0_bias = const()[name = tensor("inner_decoder_k_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(31499584)))]; + tensor inner_decoder_k_proj_0_weight = const()[name = tensor("inner_decoder_k_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(31500672)))]; + tensor inner_decoder_v_proj_0_bias = const()[name = tensor("inner_decoder_v_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(31762880)))]; + tensor inner_decoder_v_proj_0_weight = const()[name = tensor("inner_decoder_v_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(31763968)))]; + tensor inner_decoder_g_proj_0_bias = const()[name = tensor("inner_decoder_g_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32026176)))]; + tensor inner_decoder_g_proj_0_weight = const()[name = tensor("inner_decoder_g_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32027264)))]; + tensor inner_decoder_out_proj_0_bias = const()[name = tensor("inner_decoder_out_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32289472)))]; + tensor inner_decoder_out_proj_0_weight = const()[name = tensor("inner_decoder_out_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32290560)))]; + tensor inner_decoder_norm11_0_bias = const()[name = tensor("inner_decoder_norm11_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32552768)))]; + tensor inner_decoder_norm11_0_weight = const()[name = tensor("inner_decoder_norm11_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32553856)))]; + tensor inner_decoder_self_attn2_0_out_proj_bias = const()[name = tensor("inner_decoder_self_attn2_0_out_proj_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32554944)))]; + tensor inner_decoder_self_attn2_0_out_proj_weight = const()[name = tensor("inner_decoder_self_attn2_0_out_proj_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32556032)))]; + tensor inner_decoder_self_attn2_0_in_proj_bias = const()[name = tensor("inner_decoder_self_attn2_0_in_proj_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32818240)))]; + tensor inner_decoder_self_attn2_0_in_proj_weight = const()[name = tensor("inner_decoder_self_attn2_0_in_proj_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32821376)))]; + tensor inner_decoder_norm21_0_bias = const()[name = tensor("inner_decoder_norm21_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(33607872)))]; + tensor inner_decoder_norm21_0_weight = const()[name = tensor("inner_decoder_norm21_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(33608960)))]; + tensor inner_decoder_linear1_0_bias = const()[name = tensor("inner_decoder_linear1_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(33610048)))]; + tensor inner_decoder_linear1_0_weight = const()[name = tensor("inner_decoder_linear1_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(33618304)))]; + tensor inner_decoder_linear2_0_bias = const()[name = tensor("inner_decoder_linear2_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(35715520)))]; + tensor inner_decoder_linear2_0_weight = const()[name = tensor("inner_decoder_linear2_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(35716608)))]; + tensor inner_decoder_norm22_0_bias = const()[name = tensor("inner_decoder_norm22_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(37813824)))]; + tensor inner_decoder_norm22_0_weight = const()[name = tensor("inner_decoder_norm22_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(37814912)))]; + tensor inner_decoder_q_proj_1_bias = const()[name = tensor("inner_decoder_q_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(37816000)))]; + tensor inner_decoder_q_proj_1_weight = const()[name = tensor("inner_decoder_q_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(37817088)))]; + tensor inner_decoder_k_proj_1_bias = const()[name = tensor("inner_decoder_k_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38079296)))]; + tensor inner_decoder_k_proj_1_weight = const()[name = tensor("inner_decoder_k_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38080384)))]; + tensor inner_decoder_v_proj_1_bias = const()[name = tensor("inner_decoder_v_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38342592)))]; + tensor inner_decoder_v_proj_1_weight = const()[name = tensor("inner_decoder_v_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38343680)))]; + tensor inner_decoder_g_proj_1_bias = const()[name = tensor("inner_decoder_g_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38605888)))]; + tensor inner_decoder_g_proj_1_weight = const()[name = tensor("inner_decoder_g_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38606976)))]; + tensor inner_decoder_out_proj_1_bias = const()[name = tensor("inner_decoder_out_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38869184)))]; + tensor inner_decoder_out_proj_1_weight = const()[name = tensor("inner_decoder_out_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38870272)))]; + tensor inner_decoder_norm11_1_bias = const()[name = tensor("inner_decoder_norm11_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(39132480)))]; + tensor inner_decoder_norm11_1_weight = const()[name = tensor("inner_decoder_norm11_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(39133568)))]; + tensor inner_decoder_self_attn2_1_out_proj_bias = const()[name = tensor("inner_decoder_self_attn2_1_out_proj_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(39134656)))]; + tensor inner_decoder_self_attn2_1_out_proj_weight = const()[name = tensor("inner_decoder_self_attn2_1_out_proj_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(39135744)))]; + tensor inner_decoder_self_attn2_1_in_proj_bias = const()[name = tensor("inner_decoder_self_attn2_1_in_proj_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(39397952)))]; + tensor inner_decoder_self_attn2_1_in_proj_weight = const()[name = tensor("inner_decoder_self_attn2_1_in_proj_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(39401088)))]; + tensor inner_decoder_norm21_1_bias = const()[name = tensor("inner_decoder_norm21_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(40187584)))]; + tensor inner_decoder_norm21_1_weight = const()[name = tensor("inner_decoder_norm21_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(40188672)))]; + tensor inner_decoder_linear1_1_bias = const()[name = tensor("inner_decoder_linear1_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(40189760)))]; + tensor inner_decoder_linear1_1_weight = const()[name = tensor("inner_decoder_linear1_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(40198016)))]; + tensor inner_decoder_linear2_1_bias = const()[name = tensor("inner_decoder_linear2_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(42295232)))]; + tensor inner_decoder_linear2_1_weight = const()[name = tensor("inner_decoder_linear2_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(42296320)))]; + tensor inner_decoder_norm22_1_bias = const()[name = tensor("inner_decoder_norm22_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(44393536)))]; + tensor inner_decoder_norm22_1_weight = const()[name = tensor("inner_decoder_norm22_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(44394624)))]; + tensor var_19_begin_0 = const()[name = tensor("op_19_begin_0"), val = tensor([0, 0, 0])]; + tensor var_19_end_0 = const()[name = tensor("op_19_end_0"), val = tensor([1, 15, 23])]; + tensor var_19_end_mask_0 = const()[name = tensor("op_19_end_mask_0"), val = tensor([true, false, true])]; + tensor var_19 = slice_by_index(begin = var_19_begin_0, end = var_19_end_0, end_mask = var_19_end_mask_0, x = features)[name = tensor("op_19")]; + tensor var_29_begin_0 = const()[name = tensor("op_29_begin_0"), val = tensor([0, 10, 0])]; + tensor var_29_end_0 = const()[name = tensor("op_29_end_0"), val = tensor([1, 25, 23])]; + tensor var_29_end_mask_0 = const()[name = tensor("op_29_end_mask_0"), val = tensor([true, false, true])]; + tensor var_29 = slice_by_index(begin = var_29_begin_0, end = var_29_end_0, end_mask = var_29_end_mask_0, x = features)[name = tensor("op_29")]; + tensor var_39_begin_0 = const()[name = tensor("op_39_begin_0"), val = tensor([0, 20, 0])]; + tensor var_39_end_0 = const()[name = tensor("op_39_end_0"), val = tensor([1, 35, 23])]; + tensor var_39_end_mask_0 = const()[name = tensor("op_39_end_mask_0"), val = tensor([true, false, true])]; + tensor var_39 = slice_by_index(begin = var_39_begin_0, end = var_39_end_0, end_mask = var_39_end_mask_0, x = features)[name = tensor("op_39")]; + tensor var_49_begin_0 = const()[name = tensor("op_49_begin_0"), val = tensor([0, 30, 0])]; + tensor var_49_end_0 = const()[name = tensor("op_49_end_0"), val = tensor([1, 45, 23])]; + tensor var_49_end_mask_0 = const()[name = tensor("op_49_end_mask_0"), val = tensor([true, false, true])]; + tensor var_49 = slice_by_index(begin = var_49_begin_0, end = var_49_end_0, end_mask = var_49_end_mask_0, x = features)[name = tensor("op_49")]; + tensor var_59_begin_0 = const()[name = tensor("op_59_begin_0"), val = tensor([0, 40, 0])]; + tensor var_59_end_0 = const()[name = tensor("op_59_end_0"), val = tensor([1, 1, 23])]; + tensor var_59_end_mask_0 = const()[name = tensor("op_59_end_mask_0"), val = tensor([true, true, true])]; + tensor var_59 = slice_by_index(begin = var_59_begin_0, end = var_59_end_0, end_mask = var_59_end_mask_0, x = features)[name = tensor("op_59")]; + tensor stacked_axis_0 = const()[name = tensor("stacked_axis_0"), val = tensor(1)]; + tensor stacked = stack(axis = stacked_axis_0, values = (var_19, var_29, var_39, var_49, var_59))[name = tensor("stacked")]; + tensor var_66 = const()[name = tensor("op_66"), val = tensor([1, 5, 345])]; + tensor input_1 = reshape(shape = var_66, x = stacked)[name = tensor("input_1")]; + tensor var_69 = const()[name = tensor("op_69"), val = tensor(0x1p+0)]; + tensor var_75 = const()[name = tensor("op_75"), val = tensor(true)]; + tensor var_76 = const()[name = tensor("op_76"), val = tensor(0x1.4f8b58p-17)]; + tensor var_79 = const()[name = tensor("op_79"), val = tensor(0)]; + tensor var_81 = const()[name = tensor("op_81"), val = tensor(2)]; + tensor var_82 = const()[name = tensor("op_82"), val = tensor(-1)]; + tensor var_84 = const()[name = tensor("op_84"), val = tensor(0x1.0c6f7ap-20)]; + tensor var_90 = const()[name = tensor("op_90"), val = tensor(0x1.5798eep-27)]; + tensor var_93 = const()[name = tensor("op_93"), val = tensor(1)]; + tensor input_3 = linear(bias = inner_encoder_input_projection_linear_bias, weight = inner_encoder_input_projection_linear_weight, x = input_1)[name = tensor("linear_0")]; tensor input_5_axes_0 = const()[name = tensor("input_5_axes_0"), val = tensor([-1])]; - tensor input_5 = layer_norm(axes = input_5_axes_0, beta = encoder_ffn1_0_module_sequential_0_bias, epsilon = var_29, gamma = encoder_ffn1_0_module_sequential_0_weight, x = input_3)[name = tensor("input_5")]; - tensor inputs_1 = linear(bias = encoder_ffn1_0_module_sequential_1_linear_bias, weight = encoder_ffn1_0_module_sequential_1_linear_weight, x = input_5)[name = tensor("linear_1")]; - tensor input_7 = silu(x = inputs_1)[name = tensor("input_7")]; - tensor input_11 = linear(bias = encoder_ffn1_0_module_sequential_4_linear_bias, weight = encoder_ffn1_0_module_sequential_4_linear_weight, x = input_7)[name = tensor("linear_2")]; - tensor var_148 = const()[name = tensor("op_148"), val = tensor(0x1p-1)]; - tensor var_149 = mul(x = input_11, y = var_148)[name = tensor("op_149")]; - tensor input_13 = add(x = var_149, y = input_3)[name = tensor("input_13")]; + tensor input_5 = layer_norm(axes = input_5_axes_0, beta = inner_encoder_pre_layer_norm_bias, epsilon = var_76, gamma = inner_encoder_pre_layer_norm_weight, x = input_3)[name = tensor("input_5")]; + tensor input_7_axes_0 = const()[name = tensor("input_7_axes_0"), val = tensor([-1])]; + tensor input_7 = layer_norm(axes = input_7_axes_0, beta = inner_encoder_ffn1_0_module_sequential_0_bias, epsilon = var_76, gamma = inner_encoder_ffn1_0_module_sequential_0_weight, x = input_5)[name = tensor("input_7")]; + tensor inputs_1 = linear(bias = inner_encoder_ffn1_0_module_sequential_1_linear_bias, weight = inner_encoder_ffn1_0_module_sequential_1_linear_weight, x = input_7)[name = tensor("linear_1")]; + tensor input_9 = silu(x = inputs_1)[name = tensor("input_9")]; + tensor input_13 = linear(bias = inner_encoder_ffn1_0_module_sequential_4_linear_bias, weight = inner_encoder_ffn1_0_module_sequential_4_linear_weight, x = input_9)[name = tensor("linear_2")]; + tensor var_214 = const()[name = tensor("op_214"), val = tensor(0x1p-1)]; + tensor var_215 = mul(x = input_13, y = var_214)[name = tensor("op_215")]; + tensor input_15 = add(x = var_215, y = input_5)[name = tensor("input_15")]; tensor x_1_axes_0 = const()[name = tensor("x_1_axes_0"), val = tensor([-1])]; - tensor x_1 = layer_norm(axes = x_1_axes_0, beta = encoder_ret_lns_0_bias, epsilon = var_29, gamma = encoder_ret_lns_0_weight, x = input_13)[name = tensor("x_1")]; + tensor x_1 = layer_norm(axes = x_1_axes_0, beta = inner_encoder_ret_lns_0_bias, epsilon = var_76, gamma = inner_encoder_ret_lns_0_weight, x = input_15)[name = tensor("x_1")]; tensor prev_kv_1_begin_0 = const()[name = tensor("prev_kv_1_begin_0"), val = tensor([0, 0, 0, 0, 0])]; tensor prev_kv_1_end_0 = const()[name = tensor("prev_kv_1_end_0"), val = tensor([1, 1, 4, 64, 64])]; tensor prev_kv_1_end_mask_0 = const()[name = tensor("prev_kv_1_end_mask_0"), val = tensor([false, true, true, true, true])]; @@ -239,183 +265,183 @@ program(1.0) tensor prev_scale_1_end_mask_0 = const()[name = tensor("prev_scale_1_end_mask_0"), val = tensor([false, true])]; tensor prev_scale_1_squeeze_mask_0 = const()[name = tensor("prev_scale_1_squeeze_mask_0"), val = tensor([true, false])]; tensor prev_scale_1 = slice_by_index(begin = prev_scale_1_begin_0, end = prev_scale_1_end_0, end_mask = prev_scale_1_end_mask_0, squeeze_mask = prev_scale_1_squeeze_mask_0, x = enc_scale)[name = tensor("prev_scale_1")]; - tensor var_163 = linear(bias = encoder_q_proj_0_bias, weight = encoder_q_proj_0_weight, x = x_1)[name = tensor("linear_3")]; - tensor var_164 = const()[name = tensor("op_164"), val = tensor([1, 5, 4, 64])]; - tensor var_165 = reshape(shape = var_164, x = var_163)[name = tensor("op_165")]; + tensor var_229 = linear(bias = inner_encoder_q_proj_0_bias, weight = inner_encoder_q_proj_0_weight, x = x_1)[name = tensor("linear_3")]; + tensor var_230 = const()[name = tensor("op_230"), val = tensor([1, 5, 4, 64])]; + tensor var_231 = reshape(shape = var_230, x = var_229)[name = tensor("op_231")]; tensor q_1_perm_0 = const()[name = tensor("q_1_perm_0"), val = tensor([0, 2, 1, 3])]; - tensor var_169 = linear(bias = encoder_k_proj_0_bias, weight = encoder_k_proj_0_weight, x = x_1)[name = tensor("linear_4")]; - tensor var_170 = const()[name = tensor("op_170"), val = tensor(0x1p-3)]; - tensor var_171 = mul(x = var_169, y = var_170)[name = tensor("op_171")]; - tensor var_172 = const()[name = tensor("op_172"), val = tensor([1, 5, 4, 64])]; - tensor var_173 = reshape(shape = var_172, x = var_171)[name = tensor("op_173")]; + tensor var_235 = linear(bias = inner_encoder_k_proj_0_bias, weight = inner_encoder_k_proj_0_weight, x = x_1)[name = tensor("linear_4")]; + tensor var_236 = const()[name = tensor("op_236"), val = tensor(0x1p-3)]; + tensor var_237 = mul(x = var_235, y = var_236)[name = tensor("op_237")]; + tensor var_238 = const()[name = tensor("op_238"), val = tensor([1, 5, 4, 64])]; + tensor var_239 = reshape(shape = var_238, x = var_237)[name = tensor("op_239")]; tensor k_1_perm_0 = const()[name = tensor("k_1_perm_0"), val = tensor([0, 2, 1, 3])]; - tensor var_177 = linear(bias = encoder_v_proj_0_bias, weight = encoder_v_proj_0_weight, x = x_1)[name = tensor("linear_5")]; - tensor var_178 = const()[name = tensor("op_178"), val = tensor([1, 5, 4, 64])]; - tensor var_179 = reshape(shape = var_178, x = var_177)[name = tensor("op_179")]; + tensor var_243 = linear(bias = inner_encoder_v_proj_0_bias, weight = inner_encoder_v_proj_0_weight, x = x_1)[name = tensor("linear_5")]; + tensor var_244 = const()[name = tensor("op_244"), val = tensor([1, 5, 4, 64])]; + tensor var_245 = reshape(shape = var_244, x = var_243)[name = tensor("op_245")]; tensor v_1_perm_0 = const()[name = tensor("v_1_perm_0"), val = tensor([0, 2, 1, 3])]; - tensor input_17 = linear(bias = encoder_g_proj_0_bias, weight = encoder_g_proj_0_weight, x = x_1)[name = tensor("linear_6")]; + tensor input_19 = linear(bias = inner_encoder_g_proj_0_bias, weight = inner_encoder_g_proj_0_weight, x = x_1)[name = tensor("linear_6")]; tensor sqrt_s0_1 = sqrt(x = prev_scale_1)[name = tensor("sqrt_s0_1")]; - tensor s_t_1 = add(x = prev_scale_1, y = encoder__t_index)[name = tensor("s_t_1")]; + tensor s_t_1 = add(x = prev_scale_1, y = inner_encoder__t_index)[name = tensor("s_t_1")]; tensor sqrt_s_t_1 = sqrt(x = s_t_1)[name = tensor("sqrt_s_t_1")]; tensor qk_1_transpose_x_1 = const()[name = tensor("qk_1_transpose_x_1"), val = tensor(false)]; tensor qk_1_transpose_y_1 = const()[name = tensor("qk_1_transpose_y_1"), val = tensor(true)]; - tensor k_1 = transpose(perm = k_1_perm_0, x = var_173)[name = tensor("transpose_57")]; - tensor q_1 = transpose(perm = q_1_perm_0, x = var_165)[name = tensor("transpose_58")]; + tensor k_1 = transpose(perm = k_1_perm_0, x = var_239)[name = tensor("transpose_57")]; + tensor q_1 = transpose(perm = q_1_perm_0, x = var_231)[name = tensor("transpose_58")]; tensor qk_1 = matmul(transpose_x = qk_1_transpose_x_1, transpose_y = qk_1_transpose_y_1, x = q_1, y = k_1)[name = tensor("qk_1")]; - tensor var_189 = const()[name = tensor("op_189"), val = tensor([5, 1])]; - tensor var_190 = reshape(shape = var_189, x = sqrt_s_t_1)[name = tensor("op_190")]; - tensor M_1 = real_div(x = encoder__causal_mask, y = var_190)[name = tensor("M_1")]; - tensor var_192 = mul(x = qk_1, y = M_1)[name = tensor("op_192")]; + tensor var_255 = const()[name = tensor("op_255"), val = tensor([5, 1])]; + tensor var_256 = reshape(shape = var_255, x = sqrt_s_t_1)[name = tensor("op_256")]; + tensor M_1 = real_div(x = inner_encoder__causal_mask, y = var_256)[name = tensor("M_1")]; + tensor var_258 = mul(x = qk_1, y = M_1)[name = tensor("op_258")]; tensor inner_1_transpose_x_0 = const()[name = tensor("inner_1_transpose_x_0"), val = tensor(false)]; tensor inner_1_transpose_y_0 = const()[name = tensor("inner_1_transpose_y_0"), val = tensor(false)]; - tensor v_1 = transpose(perm = v_1_perm_0, x = var_179)[name = tensor("transpose_56")]; - tensor inner_1 = matmul(transpose_x = inner_1_transpose_x_0, transpose_y = inner_1_transpose_y_0, x = var_192, y = v_1)[name = tensor("inner_1")]; - tensor var_194_transpose_x_0 = const()[name = tensor("op_194_transpose_x_0"), val = tensor(false)]; - tensor var_194_transpose_y_0 = const()[name = tensor("op_194_transpose_y_0"), val = tensor(false)]; - tensor var_194 = matmul(transpose_x = var_194_transpose_x_0, transpose_y = var_194_transpose_y_0, x = q_1, y = prev_kv_1)[name = tensor("op_194")]; - tensor var_195 = real_div(x = sqrt_s0_1, y = sqrt_s_t_1)[name = tensor("op_195")]; - tensor var_196 = const()[name = tensor("op_196"), val = tensor([1, 1, 5, 1])]; - tensor var_197 = reshape(shape = var_196, x = var_195)[name = tensor("op_197")]; - tensor cross_1 = mul(x = var_194, y = var_197)[name = tensor("cross_1")]; + tensor v_1 = transpose(perm = v_1_perm_0, x = var_245)[name = tensor("transpose_56")]; + tensor inner_1 = matmul(transpose_x = inner_1_transpose_x_0, transpose_y = inner_1_transpose_y_0, x = var_258, y = v_1)[name = tensor("inner_1")]; + tensor var_260_transpose_x_0 = const()[name = tensor("op_260_transpose_x_0"), val = tensor(false)]; + tensor var_260_transpose_y_0 = const()[name = tensor("op_260_transpose_y_0"), val = tensor(false)]; + tensor var_260 = matmul(transpose_x = var_260_transpose_x_0, transpose_y = var_260_transpose_y_0, x = q_1, y = prev_kv_1)[name = tensor("op_260")]; + tensor var_261 = real_div(x = sqrt_s0_1, y = sqrt_s_t_1)[name = tensor("op_261")]; + tensor var_262 = const()[name = tensor("op_262"), val = tensor([1, 1, 5, 1])]; + tensor var_263 = reshape(shape = var_262, x = var_261)[name = tensor("op_263")]; + tensor cross_1 = mul(x = var_260, y = var_263)[name = tensor("cross_1")]; tensor out_1 = add(x = inner_1, y = cross_1)[name = tensor("out_1")]; - tensor var_200 = mul(x = prev_kv_1, y = sqrt_s0_1)[name = tensor("op_200")]; - tensor var_202_transpose_x_1 = const()[name = tensor("op_202_transpose_x_1"), val = tensor(true)]; - tensor var_202_transpose_y_1 = const()[name = tensor("op_202_transpose_y_1"), val = tensor(false)]; - tensor var_202 = matmul(transpose_x = var_202_transpose_x_1, transpose_y = var_202_transpose_y_1, x = k_1, y = v_1)[name = tensor("op_202")]; - tensor new_kv_unnorm_1 = add(x = var_200, y = var_202)[name = tensor("new_kv_unnorm_1")]; - tensor var_204 = const()[name = tensor("op_204"), val = tensor(0x1.4p+2)]; - tensor new_scale_1 = add(x = prev_scale_1, y = var_204)[name = tensor("new_scale_1")]; - tensor var_206 = sqrt(x = new_scale_1)[name = tensor("op_206")]; - tensor var_207 = real_div(x = new_kv_unnorm_1, y = var_206)[name = tensor("op_207")]; - tensor var_208_perm_0 = const()[name = tensor("op_208_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_266 = mul(x = prev_kv_1, y = sqrt_s0_1)[name = tensor("op_266")]; + tensor var_268_transpose_x_1 = const()[name = tensor("op_268_transpose_x_1"), val = tensor(true)]; + tensor var_268_transpose_y_1 = const()[name = tensor("op_268_transpose_y_1"), val = tensor(false)]; + tensor var_268 = matmul(transpose_x = var_268_transpose_x_1, transpose_y = var_268_transpose_y_1, x = k_1, y = v_1)[name = tensor("op_268")]; + tensor new_kv_unnorm_1 = add(x = var_266, y = var_268)[name = tensor("new_kv_unnorm_1")]; + tensor var_270 = const()[name = tensor("op_270"), val = tensor(0x1.4p+2)]; + tensor new_scale_1 = add(x = prev_scale_1, y = var_270)[name = tensor("new_scale_1")]; + tensor var_272 = sqrt(x = new_scale_1)[name = tensor("op_272")]; + tensor var_273 = real_div(x = new_kv_unnorm_1, y = var_272)[name = tensor("op_273")]; + tensor var_274_perm_0 = const()[name = tensor("op_274_perm_0"), val = tensor([0, 2, 1, 3])]; tensor out_3_axes_0 = const()[name = tensor("out_3_axes_0"), val = tensor([-1])]; - tensor var_208 = transpose(perm = var_208_perm_0, x = out_1)[name = tensor("transpose_55")]; - tensor out_3 = layer_norm(axes = out_3_axes_0, epsilon = var_18, x = var_208)[name = tensor("out_3")]; - tensor var_212 = const()[name = tensor("op_212"), val = tensor([1, 5, 256])]; - tensor out_5 = reshape(shape = var_212, x = out_3)[name = tensor("out_5")]; - tensor var_214 = silu(x = input_17)[name = tensor("op_214")]; - tensor input_19 = mul(x = var_214, y = out_5)[name = tensor("input_19")]; - tensor ret_out_1 = linear(bias = encoder_out_proj_0_bias, weight = encoder_out_proj_0_weight, x = input_19)[name = tensor("linear_7")]; - tensor x_3 = add(x = input_13, y = ret_out_1)[name = tensor("x_3")]; + tensor var_274 = transpose(perm = var_274_perm_0, x = out_1)[name = tensor("transpose_55")]; + tensor out_3 = layer_norm(axes = out_3_axes_0, epsilon = var_84, x = var_274)[name = tensor("out_3")]; + tensor var_278 = const()[name = tensor("op_278"), val = tensor([1, 5, 256])]; + tensor out_5 = reshape(shape = var_278, x = out_3)[name = tensor("out_5")]; + tensor var_280 = silu(x = input_19)[name = tensor("op_280")]; + tensor input_21 = mul(x = var_280, y = out_5)[name = tensor("input_21")]; + tensor ret_out_1 = linear(bias = inner_encoder_out_proj_0_bias, weight = inner_encoder_out_proj_0_weight, x = input_21)[name = tensor("linear_7")]; + tensor x_3 = add(x = input_15, y = ret_out_1)[name = tensor("x_3")]; tensor window_1_begin_0 = const()[name = tensor("window_1_begin_0"), val = tensor([0, 0, 0, 0])]; tensor window_1_end_0 = const()[name = tensor("window_1_end_0"), val = tensor([1, 1, 16, 256])]; tensor window_1_end_mask_0 = const()[name = tensor("window_1_end_mask_0"), val = tensor([false, true, true, true])]; tensor window_1_squeeze_mask_0 = const()[name = tensor("window_1_squeeze_mask_0"), val = tensor([true, false, false, false])]; tensor window_1 = slice_by_index(begin = window_1_begin_0, end = window_1_end_0, end_mask = window_1_end_mask_0, squeeze_mask = window_1_squeeze_mask_0, x = enc_conv_cache)[name = tensor("window_1")]; - tensor var_222_begin_0 = const()[name = tensor("op_222_begin_0"), val = tensor([0, 0, 0])]; - tensor var_222_end_0 = const()[name = tensor("op_222_end_0"), val = tensor([1, 1, 256])]; - tensor var_222_end_mask_0 = const()[name = tensor("op_222_end_mask_0"), val = tensor([true, false, true])]; - tensor var_222 = slice_by_index(begin = var_222_begin_0, end = var_222_end_0, end_mask = var_222_end_mask_0, x = x_3)[name = tensor("op_222")]; - tensor var_225_begin_0 = const()[name = tensor("op_225_begin_0"), val = tensor([0, 1, 0])]; - tensor var_225_end_0 = const()[name = tensor("op_225_end_0"), val = tensor([1, 16, 256])]; - tensor var_225_end_mask_0 = const()[name = tensor("op_225_end_mask_0"), val = tensor([true, true, true])]; - tensor var_225 = slice_by_index(begin = var_225_begin_0, end = var_225_end_0, end_mask = var_225_end_mask_0, x = window_1)[name = tensor("op_225")]; + tensor var_288_begin_0 = const()[name = tensor("op_288_begin_0"), val = tensor([0, 0, 0])]; + tensor var_288_end_0 = const()[name = tensor("op_288_end_0"), val = tensor([1, 1, 256])]; + tensor var_288_end_mask_0 = const()[name = tensor("op_288_end_mask_0"), val = tensor([true, false, true])]; + tensor var_288 = slice_by_index(begin = var_288_begin_0, end = var_288_end_0, end_mask = var_288_end_mask_0, x = x_3)[name = tensor("op_288")]; + tensor var_291_begin_0 = const()[name = tensor("op_291_begin_0"), val = tensor([0, 1, 0])]; + tensor var_291_end_0 = const()[name = tensor("op_291_end_0"), val = tensor([1, 16, 256])]; + tensor var_291_end_mask_0 = const()[name = tensor("op_291_end_mask_0"), val = tensor([true, true, true])]; + tensor var_291 = slice_by_index(begin = var_291_begin_0, end = var_291_end_0, end_mask = var_291_end_mask_0, x = window_1)[name = tensor("op_291")]; tensor window_3_interleave_0 = const()[name = tensor("window_3_interleave_0"), val = tensor(false)]; - tensor window_3 = concat(axis = var_27, interleave = window_3_interleave_0, values = (var_225, var_222))[name = tensor("window_3")]; - tensor var_230_begin_0 = const()[name = tensor("op_230_begin_0"), val = tensor([0, 1, 0])]; - tensor var_230_end_0 = const()[name = tensor("op_230_end_0"), val = tensor([1, 2, 256])]; - tensor var_230_end_mask_0 = const()[name = tensor("op_230_end_mask_0"), val = tensor([true, false, true])]; - tensor var_230 = slice_by_index(begin = var_230_begin_0, end = var_230_end_0, end_mask = var_230_end_mask_0, x = x_3)[name = tensor("op_230")]; - tensor var_233_begin_0 = const()[name = tensor("op_233_begin_0"), val = tensor([0, 1, 0])]; - tensor var_233_end_0 = const()[name = tensor("op_233_end_0"), val = tensor([1, 16, 256])]; - tensor var_233_end_mask_0 = const()[name = tensor("op_233_end_mask_0"), val = tensor([true, true, true])]; - tensor var_233 = slice_by_index(begin = var_233_begin_0, end = var_233_end_0, end_mask = var_233_end_mask_0, x = window_3)[name = tensor("op_233")]; + tensor window_3 = concat(axis = var_93, interleave = window_3_interleave_0, values = (var_291, var_288))[name = tensor("window_3")]; + tensor var_296_begin_0 = const()[name = tensor("op_296_begin_0"), val = tensor([0, 1, 0])]; + tensor var_296_end_0 = const()[name = tensor("op_296_end_0"), val = tensor([1, 2, 256])]; + tensor var_296_end_mask_0 = const()[name = tensor("op_296_end_mask_0"), val = tensor([true, false, true])]; + tensor var_296 = slice_by_index(begin = var_296_begin_0, end = var_296_end_0, end_mask = var_296_end_mask_0, x = x_3)[name = tensor("op_296")]; + tensor var_299_begin_0 = const()[name = tensor("op_299_begin_0"), val = tensor([0, 1, 0])]; + tensor var_299_end_0 = const()[name = tensor("op_299_end_0"), val = tensor([1, 16, 256])]; + tensor var_299_end_mask_0 = const()[name = tensor("op_299_end_mask_0"), val = tensor([true, true, true])]; + tensor var_299 = slice_by_index(begin = var_299_begin_0, end = var_299_end_0, end_mask = var_299_end_mask_0, x = window_3)[name = tensor("op_299")]; tensor window_5_interleave_0 = const()[name = tensor("window_5_interleave_0"), val = tensor(false)]; - tensor window_5 = concat(axis = var_27, interleave = window_5_interleave_0, values = (var_233, var_230))[name = tensor("window_5")]; - tensor var_238_begin_0 = const()[name = tensor("op_238_begin_0"), val = tensor([0, 2, 0])]; - tensor var_238_end_0 = const()[name = tensor("op_238_end_0"), val = tensor([1, 3, 256])]; - tensor var_238_end_mask_0 = const()[name = tensor("op_238_end_mask_0"), val = tensor([true, false, true])]; - tensor var_238 = slice_by_index(begin = var_238_begin_0, end = var_238_end_0, end_mask = var_238_end_mask_0, x = x_3)[name = tensor("op_238")]; - tensor var_241_begin_0 = const()[name = tensor("op_241_begin_0"), val = tensor([0, 1, 0])]; - tensor var_241_end_0 = const()[name = tensor("op_241_end_0"), val = tensor([1, 16, 256])]; - tensor var_241_end_mask_0 = const()[name = tensor("op_241_end_mask_0"), val = tensor([true, true, true])]; - tensor var_241 = slice_by_index(begin = var_241_begin_0, end = var_241_end_0, end_mask = var_241_end_mask_0, x = window_5)[name = tensor("op_241")]; + tensor window_5 = concat(axis = var_93, interleave = window_5_interleave_0, values = (var_299, var_296))[name = tensor("window_5")]; + tensor var_304_begin_0 = const()[name = tensor("op_304_begin_0"), val = tensor([0, 2, 0])]; + tensor var_304_end_0 = const()[name = tensor("op_304_end_0"), val = tensor([1, 3, 256])]; + tensor var_304_end_mask_0 = const()[name = tensor("op_304_end_mask_0"), val = tensor([true, false, true])]; + tensor var_304 = slice_by_index(begin = var_304_begin_0, end = var_304_end_0, end_mask = var_304_end_mask_0, x = x_3)[name = tensor("op_304")]; + tensor var_307_begin_0 = const()[name = tensor("op_307_begin_0"), val = tensor([0, 1, 0])]; + tensor var_307_end_0 = const()[name = tensor("op_307_end_0"), val = tensor([1, 16, 256])]; + tensor var_307_end_mask_0 = const()[name = tensor("op_307_end_mask_0"), val = tensor([true, true, true])]; + tensor var_307 = slice_by_index(begin = var_307_begin_0, end = var_307_end_0, end_mask = var_307_end_mask_0, x = window_5)[name = tensor("op_307")]; tensor window_7_interleave_0 = const()[name = tensor("window_7_interleave_0"), val = tensor(false)]; - tensor window_7 = concat(axis = var_27, interleave = window_7_interleave_0, values = (var_241, var_238))[name = tensor("window_7")]; - tensor var_246_begin_0 = const()[name = tensor("op_246_begin_0"), val = tensor([0, 3, 0])]; - tensor var_246_end_0 = const()[name = tensor("op_246_end_0"), val = tensor([1, 4, 256])]; - tensor var_246_end_mask_0 = const()[name = tensor("op_246_end_mask_0"), val = tensor([true, false, true])]; - tensor var_246 = slice_by_index(begin = var_246_begin_0, end = var_246_end_0, end_mask = var_246_end_mask_0, x = x_3)[name = tensor("op_246")]; - tensor var_249_begin_0 = const()[name = tensor("op_249_begin_0"), val = tensor([0, 1, 0])]; - tensor var_249_end_0 = const()[name = tensor("op_249_end_0"), val = tensor([1, 16, 256])]; - tensor var_249_end_mask_0 = const()[name = tensor("op_249_end_mask_0"), val = tensor([true, true, true])]; - tensor var_249 = slice_by_index(begin = var_249_begin_0, end = var_249_end_0, end_mask = var_249_end_mask_0, x = window_7)[name = tensor("op_249")]; + tensor window_7 = concat(axis = var_93, interleave = window_7_interleave_0, values = (var_307, var_304))[name = tensor("window_7")]; + tensor var_312_begin_0 = const()[name = tensor("op_312_begin_0"), val = tensor([0, 3, 0])]; + tensor var_312_end_0 = const()[name = tensor("op_312_end_0"), val = tensor([1, 4, 256])]; + tensor var_312_end_mask_0 = const()[name = tensor("op_312_end_mask_0"), val = tensor([true, false, true])]; + tensor var_312 = slice_by_index(begin = var_312_begin_0, end = var_312_end_0, end_mask = var_312_end_mask_0, x = x_3)[name = tensor("op_312")]; + tensor var_315_begin_0 = const()[name = tensor("op_315_begin_0"), val = tensor([0, 1, 0])]; + tensor var_315_end_0 = const()[name = tensor("op_315_end_0"), val = tensor([1, 16, 256])]; + tensor var_315_end_mask_0 = const()[name = tensor("op_315_end_mask_0"), val = tensor([true, true, true])]; + tensor var_315 = slice_by_index(begin = var_315_begin_0, end = var_315_end_0, end_mask = var_315_end_mask_0, x = window_7)[name = tensor("op_315")]; tensor window_9_interleave_0 = const()[name = tensor("window_9_interleave_0"), val = tensor(false)]; - tensor window_9 = concat(axis = var_27, interleave = window_9_interleave_0, values = (var_249, var_246))[name = tensor("window_9")]; - tensor var_254_begin_0 = const()[name = tensor("op_254_begin_0"), val = tensor([0, 4, 0])]; - tensor var_254_end_0 = const()[name = tensor("op_254_end_0"), val = tensor([1, 1, 256])]; - tensor var_254_end_mask_0 = const()[name = tensor("op_254_end_mask_0"), val = tensor([true, true, true])]; - tensor var_254 = slice_by_index(begin = var_254_begin_0, end = var_254_end_0, end_mask = var_254_end_mask_0, x = x_3)[name = tensor("op_254")]; - tensor var_257_begin_0 = const()[name = tensor("op_257_begin_0"), val = tensor([0, 1, 0])]; - tensor var_257_end_0 = const()[name = tensor("op_257_end_0"), val = tensor([1, 16, 256])]; - tensor var_257_end_mask_0 = const()[name = tensor("op_257_end_mask_0"), val = tensor([true, true, true])]; - tensor var_257 = slice_by_index(begin = var_257_begin_0, end = var_257_end_0, end_mask = var_257_end_mask_0, x = window_9)[name = tensor("op_257")]; + tensor window_9 = concat(axis = var_93, interleave = window_9_interleave_0, values = (var_315, var_312))[name = tensor("window_9")]; + tensor var_320_begin_0 = const()[name = tensor("op_320_begin_0"), val = tensor([0, 4, 0])]; + tensor var_320_end_0 = const()[name = tensor("op_320_end_0"), val = tensor([1, 1, 256])]; + tensor var_320_end_mask_0 = const()[name = tensor("op_320_end_mask_0"), val = tensor([true, true, true])]; + tensor var_320 = slice_by_index(begin = var_320_begin_0, end = var_320_end_0, end_mask = var_320_end_mask_0, x = x_3)[name = tensor("op_320")]; + tensor var_323_begin_0 = const()[name = tensor("op_323_begin_0"), val = tensor([0, 1, 0])]; + tensor var_323_end_0 = const()[name = tensor("op_323_end_0"), val = tensor([1, 16, 256])]; + tensor var_323_end_mask_0 = const()[name = tensor("op_323_end_mask_0"), val = tensor([true, true, true])]; + tensor var_323 = slice_by_index(begin = var_323_begin_0, end = var_323_end_0, end_mask = var_323_end_mask_0, x = window_9)[name = tensor("op_323")]; tensor window_11_interleave_0 = const()[name = tensor("window_11_interleave_0"), val = tensor(false)]; - tensor window_11 = concat(axis = var_27, interleave = window_11_interleave_0, values = (var_257, var_254))[name = tensor("window_11")]; - tensor input_21_interleave_0 = const()[name = tensor("input_21_interleave_0"), val = tensor(false)]; - tensor input_21 = concat(axis = var_24, interleave = input_21_interleave_0, values = (window_3, window_5, window_7, window_9, window_11))[name = tensor("input_21")]; + tensor window_11 = concat(axis = var_93, interleave = window_11_interleave_0, values = (var_323, var_320))[name = tensor("window_11")]; + tensor input_23_interleave_0 = const()[name = tensor("input_23_interleave_0"), val = tensor(false)]; + tensor input_23 = concat(axis = var_79, interleave = input_23_interleave_0, values = (window_3, window_5, window_7, window_9, window_11))[name = tensor("input_23")]; tensor x_5_axes_0 = const()[name = tensor("x_5_axes_0"), val = tensor([-1])]; - tensor x_5 = layer_norm(axes = x_5_axes_0, beta = encoder_conv_module_0_sequential_0_bias, epsilon = var_29, gamma = encoder_conv_module_0_sequential_0_weight, x = input_21)[name = tensor("x_5")]; - tensor input_23_perm_0 = const()[name = tensor("input_23_perm_0"), val = tensor([0, 2, 1])]; + tensor x_5 = layer_norm(axes = x_5_axes_0, beta = inner_encoder_conv_module_0_sequential_0_bias, epsilon = var_76, gamma = inner_encoder_conv_module_0_sequential_0_weight, x = input_23)[name = tensor("x_5")]; + tensor input_25_perm_0 = const()[name = tensor("input_25_perm_0"), val = tensor([0, 2, 1])]; tensor inputs_3_pad_type_0 = const()[name = tensor("inputs_3_pad_type_0"), val = tensor("valid")]; tensor inputs_3_strides_0 = const()[name = tensor("inputs_3_strides_0"), val = tensor([1])]; tensor inputs_3_pad_0 = const()[name = tensor("inputs_3_pad_0"), val = tensor([0, 0])]; tensor inputs_3_dilations_0 = const()[name = tensor("inputs_3_dilations_0"), val = tensor([1])]; tensor inputs_3_groups_0 = const()[name = tensor("inputs_3_groups_0"), val = tensor(1)]; - tensor input_23 = transpose(perm = input_23_perm_0, x = x_5)[name = tensor("transpose_54")]; - tensor inputs_3 = conv(bias = encoder_conv_module_0_sequential_2_conv_bias, dilations = inputs_3_dilations_0, groups = inputs_3_groups_0, pad = inputs_3_pad_0, pad_type = inputs_3_pad_type_0, strides = inputs_3_strides_0, weight = encoder_conv_module_0_sequential_2_conv_weight, x = input_23)[name = tensor("inputs_3")]; - tensor var_282_split_sizes_0 = const()[name = tensor("op_282_split_sizes_0"), val = tensor([256, 256])]; - tensor var_282_axis_0 = const()[name = tensor("op_282_axis_0"), val = tensor(1)]; - tensor var_282_0, tensor var_282_1 = split(axis = var_282_axis_0, split_sizes = var_282_split_sizes_0, x = inputs_3)[name = tensor("op_282")]; - tensor var_284 = sigmoid(x = var_282_1)[name = tensor("op_284")]; - tensor inputs_5 = mul(x = var_282_0, y = var_284)[name = tensor("inputs_5")]; + tensor input_25 = transpose(perm = input_25_perm_0, x = x_5)[name = tensor("transpose_54")]; + tensor inputs_3 = conv(bias = inner_encoder_conv_module_0_sequential_2_conv_bias, dilations = inputs_3_dilations_0, groups = inputs_3_groups_0, pad = inputs_3_pad_0, pad_type = inputs_3_pad_type_0, strides = inputs_3_strides_0, weight = inner_encoder_conv_module_0_sequential_2_conv_weight, x = input_25)[name = tensor("inputs_3")]; + tensor var_348_split_sizes_0 = const()[name = tensor("op_348_split_sizes_0"), val = tensor([256, 256])]; + tensor var_348_axis_0 = const()[name = tensor("op_348_axis_0"), val = tensor(1)]; + tensor var_348_0, tensor var_348_1 = split(axis = var_348_axis_0, split_sizes = var_348_split_sizes_0, x = inputs_3)[name = tensor("op_348")]; + tensor var_350 = sigmoid(x = var_348_1)[name = tensor("op_350")]; + tensor inputs_5 = mul(x = var_348_0, y = var_350)[name = tensor("inputs_5")]; tensor outputs_aug_1_pad_type_0 = const()[name = tensor("outputs_aug_1_pad_type_0"), val = tensor("custom")]; tensor outputs_aug_1_pad_0 = const()[name = tensor("outputs_aug_1_pad_0"), val = tensor([15, 15])]; tensor outputs_aug_1_groups_0 = const()[name = tensor("outputs_aug_1_groups_0"), val = tensor(256)]; tensor outputs_aug_1_strides_0 = const()[name = tensor("outputs_aug_1_strides_0"), val = tensor([1])]; tensor outputs_aug_1_dilations_0 = const()[name = tensor("outputs_aug_1_dilations_0"), val = tensor([1])]; - tensor outputs_aug_1 = conv(dilations = outputs_aug_1_dilations_0, groups = outputs_aug_1_groups_0, pad = outputs_aug_1_pad_0, pad_type = outputs_aug_1_pad_type_0, strides = outputs_aug_1_strides_0, weight = encoder_conv_module_0_sequential_4_conv_weight, x = inputs_5)[name = tensor("outputs_aug_1")]; - tensor input_25_begin_0 = const()[name = tensor("input_25_begin_0"), val = tensor([0, 0, 0])]; - tensor input_25_end_0 = const()[name = tensor("input_25_end_0"), val = tensor([5, 256, 16])]; - tensor input_25_end_mask_0 = const()[name = tensor("input_25_end_mask_0"), val = tensor([true, true, false])]; - tensor input_25 = slice_by_index(begin = input_25_begin_0, end = input_25_end_0, end_mask = input_25_end_mask_0, x = outputs_aug_1)[name = tensor("input_25")]; - tensor inputs_7 = batch_norm(beta = encoder_conv_module_0_sequential_5_bias, epsilon = var_29, gamma = encoder_conv_module_0_sequential_5_weight, mean = encoder_conv_module_0_sequential_5_running_mean, variance = encoder_conv_module_0_sequential_5_running_var, x = input_25)[name = tensor("inputs_7")]; - tensor input_27 = silu(x = inputs_7)[name = tensor("input_27")]; - tensor input_29_pad_type_0 = const()[name = tensor("input_29_pad_type_0"), val = tensor("valid")]; - tensor input_29_strides_0 = const()[name = tensor("input_29_strides_0"), val = tensor([1])]; - tensor input_29_pad_0 = const()[name = tensor("input_29_pad_0"), val = tensor([0, 0])]; - tensor input_29_dilations_0 = const()[name = tensor("input_29_dilations_0"), val = tensor([1])]; - tensor input_29_groups_0 = const()[name = tensor("input_29_groups_0"), val = tensor(1)]; - tensor input_29 = conv(bias = encoder_conv_module_0_sequential_7_conv_bias, dilations = input_29_dilations_0, groups = input_29_groups_0, pad = input_29_pad_0, pad_type = input_29_pad_type_0, strides = input_29_strides_0, weight = encoder_conv_module_0_sequential_7_conv_weight, x = input_27)[name = tensor("input_29")]; + tensor outputs_aug_1 = conv(dilations = outputs_aug_1_dilations_0, groups = outputs_aug_1_groups_0, pad = outputs_aug_1_pad_0, pad_type = outputs_aug_1_pad_type_0, strides = outputs_aug_1_strides_0, weight = inner_encoder_conv_module_0_sequential_4_conv_weight, x = inputs_5)[name = tensor("outputs_aug_1")]; + tensor input_27_begin_0 = const()[name = tensor("input_27_begin_0"), val = tensor([0, 0, 0])]; + tensor input_27_end_0 = const()[name = tensor("input_27_end_0"), val = tensor([5, 256, 16])]; + tensor input_27_end_mask_0 = const()[name = tensor("input_27_end_mask_0"), val = tensor([true, true, false])]; + tensor input_27 = slice_by_index(begin = input_27_begin_0, end = input_27_end_0, end_mask = input_27_end_mask_0, x = outputs_aug_1)[name = tensor("input_27")]; + tensor inputs_7 = batch_norm(beta = inner_encoder_conv_module_0_sequential_5_bias, epsilon = var_76, gamma = inner_encoder_conv_module_0_sequential_5_weight, mean = inner_encoder_conv_module_0_sequential_5_running_mean, variance = inner_encoder_conv_module_0_sequential_5_running_var, x = input_27)[name = tensor("inputs_7")]; + tensor input_29 = silu(x = inputs_7)[name = tensor("input_29")]; + tensor input_31_pad_type_0 = const()[name = tensor("input_31_pad_type_0"), val = tensor("valid")]; + tensor input_31_strides_0 = const()[name = tensor("input_31_strides_0"), val = tensor([1])]; + tensor input_31_pad_0 = const()[name = tensor("input_31_pad_0"), val = tensor([0, 0])]; + tensor input_31_dilations_0 = const()[name = tensor("input_31_dilations_0"), val = tensor([1])]; + tensor input_31_groups_0 = const()[name = tensor("input_31_groups_0"), val = tensor(1)]; + tensor input_31 = conv(bias = inner_encoder_conv_module_0_sequential_7_conv_bias, dilations = input_31_dilations_0, groups = input_31_groups_0, pad = input_31_pad_0, pad_type = input_31_pad_type_0, strides = input_31_strides_0, weight = inner_encoder_conv_module_0_sequential_7_conv_weight, x = input_29)[name = tensor("input_31")]; tensor conv_out_1_perm_0 = const()[name = tensor("conv_out_1_perm_0"), val = tensor([0, 2, 1])]; - tensor var_315_begin_0 = const()[name = tensor("op_315_begin_0"), val = tensor([0, -1, 0])]; - tensor var_315_end_0 = const()[name = tensor("op_315_end_0"), val = tensor([5, 16, 256])]; - tensor var_315_end_mask_0 = const()[name = tensor("op_315_end_mask_0"), val = tensor([true, true, true])]; - tensor conv_out_1 = transpose(perm = conv_out_1_perm_0, x = input_29)[name = tensor("transpose_53")]; - tensor var_315 = slice_by_index(begin = var_315_begin_0, end = var_315_end_0, end_mask = var_315_end_mask_0, x = conv_out_1)[name = tensor("op_315")]; - tensor var_317_perm_0 = const()[name = tensor("op_317_perm_0"), val = tensor([1, 0, 2])]; - tensor var_317 = transpose(perm = var_317_perm_0, x = var_315)[name = tensor("transpose_52")]; - tensor input_31 = add(x = x_3, y = var_317)[name = tensor("input_31")]; - tensor input_33_axes_0 = const()[name = tensor("input_33_axes_0"), val = tensor([-1])]; - tensor input_33 = layer_norm(axes = input_33_axes_0, beta = encoder_ffn2_0_module_sequential_0_bias, epsilon = var_29, gamma = encoder_ffn2_0_module_sequential_0_weight, x = input_31)[name = tensor("input_33")]; - tensor inputs_9 = linear(bias = encoder_ffn2_0_module_sequential_1_linear_bias, weight = encoder_ffn2_0_module_sequential_1_linear_weight, x = input_33)[name = tensor("linear_8")]; - tensor input_35 = silu(x = inputs_9)[name = tensor("input_35")]; - tensor input_39 = linear(bias = encoder_ffn2_0_module_sequential_4_linear_bias, weight = encoder_ffn2_0_module_sequential_4_linear_weight, x = input_35)[name = tensor("linear_9")]; - tensor var_340 = const()[name = tensor("op_340"), val = tensor(0x1p-1)]; - tensor var_341 = mul(x = input_39, y = var_340)[name = tensor("op_341")]; - tensor input_41 = add(x = var_341, y = input_31)[name = tensor("input_41")]; - tensor input_43_axes_0 = const()[name = tensor("input_43_axes_0"), val = tensor([-1])]; - tensor input_43 = layer_norm(axes = input_43_axes_0, beta = encoder_layer_norm_0_bias, epsilon = var_29, gamma = encoder_layer_norm_0_weight, x = input_41)[name = tensor("input_43")]; + tensor var_381_begin_0 = const()[name = tensor("op_381_begin_0"), val = tensor([0, -1, 0])]; + tensor var_381_end_0 = const()[name = tensor("op_381_end_0"), val = tensor([5, 16, 256])]; + tensor var_381_end_mask_0 = const()[name = tensor("op_381_end_mask_0"), val = tensor([true, true, true])]; + tensor conv_out_1 = transpose(perm = conv_out_1_perm_0, x = input_31)[name = tensor("transpose_53")]; + tensor var_381 = slice_by_index(begin = var_381_begin_0, end = var_381_end_0, end_mask = var_381_end_mask_0, x = conv_out_1)[name = tensor("op_381")]; + tensor var_383_perm_0 = const()[name = tensor("op_383_perm_0"), val = tensor([1, 0, 2])]; + tensor var_383 = transpose(perm = var_383_perm_0, x = var_381)[name = tensor("transpose_52")]; + tensor input_33 = add(x = x_3, y = var_383)[name = tensor("input_33")]; + tensor input_35_axes_0 = const()[name = tensor("input_35_axes_0"), val = tensor([-1])]; + tensor input_35 = layer_norm(axes = input_35_axes_0, beta = inner_encoder_ffn2_0_module_sequential_0_bias, epsilon = var_76, gamma = inner_encoder_ffn2_0_module_sequential_0_weight, x = input_33)[name = tensor("input_35")]; + tensor inputs_9 = linear(bias = inner_encoder_ffn2_0_module_sequential_1_linear_bias, weight = inner_encoder_ffn2_0_module_sequential_1_linear_weight, x = input_35)[name = tensor("linear_8")]; + tensor input_37 = silu(x = inputs_9)[name = tensor("input_37")]; + tensor input_41 = linear(bias = inner_encoder_ffn2_0_module_sequential_4_linear_bias, weight = inner_encoder_ffn2_0_module_sequential_4_linear_weight, x = input_37)[name = tensor("linear_9")]; + tensor var_406 = const()[name = tensor("op_406"), val = tensor(0x1p-1)]; + tensor var_407 = mul(x = input_41, y = var_406)[name = tensor("op_407")]; + tensor input_43 = add(x = var_407, y = input_33)[name = tensor("input_43")]; tensor input_45_axes_0 = const()[name = tensor("input_45_axes_0"), val = tensor([-1])]; - tensor input_45 = layer_norm(axes = input_45_axes_0, beta = encoder_ffn1_1_module_sequential_0_bias, epsilon = var_29, gamma = encoder_ffn1_1_module_sequential_0_weight, x = input_43)[name = tensor("input_45")]; - tensor inputs_11 = linear(bias = encoder_ffn1_1_module_sequential_1_linear_bias, weight = encoder_ffn1_1_module_sequential_1_linear_weight, x = input_45)[name = tensor("linear_10")]; - tensor input_47 = silu(x = inputs_11)[name = tensor("input_47")]; - tensor input_51 = linear(bias = encoder_ffn1_1_module_sequential_4_linear_bias, weight = encoder_ffn1_1_module_sequential_4_linear_weight, x = input_47)[name = tensor("linear_11")]; - tensor var_370 = const()[name = tensor("op_370"), val = tensor(0x1p-1)]; - tensor var_371 = mul(x = input_51, y = var_370)[name = tensor("op_371")]; - tensor input_53 = add(x = var_371, y = input_43)[name = tensor("input_53")]; + tensor input_45 = layer_norm(axes = input_45_axes_0, beta = inner_encoder_layer_norm_0_bias, epsilon = var_76, gamma = inner_encoder_layer_norm_0_weight, x = input_43)[name = tensor("input_45")]; + tensor input_47_axes_0 = const()[name = tensor("input_47_axes_0"), val = tensor([-1])]; + tensor input_47 = layer_norm(axes = input_47_axes_0, beta = inner_encoder_ffn1_1_module_sequential_0_bias, epsilon = var_76, gamma = inner_encoder_ffn1_1_module_sequential_0_weight, x = input_45)[name = tensor("input_47")]; + tensor inputs_11 = linear(bias = inner_encoder_ffn1_1_module_sequential_1_linear_bias, weight = inner_encoder_ffn1_1_module_sequential_1_linear_weight, x = input_47)[name = tensor("linear_10")]; + tensor input_49 = silu(x = inputs_11)[name = tensor("input_49")]; + tensor input_53 = linear(bias = inner_encoder_ffn1_1_module_sequential_4_linear_bias, weight = inner_encoder_ffn1_1_module_sequential_4_linear_weight, x = input_49)[name = tensor("linear_11")]; + tensor var_436 = const()[name = tensor("op_436"), val = tensor(0x1p-1)]; + tensor var_437 = mul(x = input_53, y = var_436)[name = tensor("op_437")]; + tensor input_55 = add(x = var_437, y = input_45)[name = tensor("input_55")]; tensor x_7_axes_0 = const()[name = tensor("x_7_axes_0"), val = tensor([-1])]; - tensor x_7 = layer_norm(axes = x_7_axes_0, beta = encoder_ret_lns_1_bias, epsilon = var_29, gamma = encoder_ret_lns_1_weight, x = input_53)[name = tensor("x_7")]; + tensor x_7 = layer_norm(axes = x_7_axes_0, beta = inner_encoder_ret_lns_1_bias, epsilon = var_76, gamma = inner_encoder_ret_lns_1_weight, x = input_55)[name = tensor("x_7")]; tensor prev_kv_3_begin_0 = const()[name = tensor("prev_kv_3_begin_0"), val = tensor([1, 0, 0, 0, 0])]; tensor prev_kv_3_end_0 = const()[name = tensor("prev_kv_3_end_0"), val = tensor([2, 1, 4, 64, 64])]; tensor prev_kv_3_end_mask_0 = const()[name = tensor("prev_kv_3_end_mask_0"), val = tensor([false, true, true, true, true])]; @@ -426,183 +452,183 @@ program(1.0) tensor prev_scale_3_end_mask_0 = const()[name = tensor("prev_scale_3_end_mask_0"), val = tensor([false, true])]; tensor prev_scale_3_squeeze_mask_0 = const()[name = tensor("prev_scale_3_squeeze_mask_0"), val = tensor([true, false])]; tensor prev_scale_3 = slice_by_index(begin = prev_scale_3_begin_0, end = prev_scale_3_end_0, end_mask = prev_scale_3_end_mask_0, squeeze_mask = prev_scale_3_squeeze_mask_0, x = enc_scale)[name = tensor("prev_scale_3")]; - tensor var_385 = linear(bias = encoder_q_proj_1_bias, weight = encoder_q_proj_1_weight, x = x_7)[name = tensor("linear_12")]; - tensor var_386 = const()[name = tensor("op_386"), val = tensor([1, 5, 4, 64])]; - tensor var_387 = reshape(shape = var_386, x = var_385)[name = tensor("op_387")]; + tensor var_451 = linear(bias = inner_encoder_q_proj_1_bias, weight = inner_encoder_q_proj_1_weight, x = x_7)[name = tensor("linear_12")]; + tensor var_452 = const()[name = tensor("op_452"), val = tensor([1, 5, 4, 64])]; + tensor var_453 = reshape(shape = var_452, x = var_451)[name = tensor("op_453")]; tensor q_3_perm_0 = const()[name = tensor("q_3_perm_0"), val = tensor([0, 2, 1, 3])]; - tensor var_391 = linear(bias = encoder_k_proj_1_bias, weight = encoder_k_proj_1_weight, x = x_7)[name = tensor("linear_13")]; - tensor var_392 = const()[name = tensor("op_392"), val = tensor(0x1p-3)]; - tensor var_393 = mul(x = var_391, y = var_392)[name = tensor("op_393")]; - tensor var_394 = const()[name = tensor("op_394"), val = tensor([1, 5, 4, 64])]; - tensor var_395 = reshape(shape = var_394, x = var_393)[name = tensor("op_395")]; + tensor var_457 = linear(bias = inner_encoder_k_proj_1_bias, weight = inner_encoder_k_proj_1_weight, x = x_7)[name = tensor("linear_13")]; + tensor var_458 = const()[name = tensor("op_458"), val = tensor(0x1p-3)]; + tensor var_459 = mul(x = var_457, y = var_458)[name = tensor("op_459")]; + tensor var_460 = const()[name = tensor("op_460"), val = tensor([1, 5, 4, 64])]; + tensor var_461 = reshape(shape = var_460, x = var_459)[name = tensor("op_461")]; tensor k_3_perm_0 = const()[name = tensor("k_3_perm_0"), val = tensor([0, 2, 1, 3])]; - tensor var_399 = linear(bias = encoder_v_proj_1_bias, weight = encoder_v_proj_1_weight, x = x_7)[name = tensor("linear_14")]; - tensor var_400 = const()[name = tensor("op_400"), val = tensor([1, 5, 4, 64])]; - tensor var_401 = reshape(shape = var_400, x = var_399)[name = tensor("op_401")]; + tensor var_465 = linear(bias = inner_encoder_v_proj_1_bias, weight = inner_encoder_v_proj_1_weight, x = x_7)[name = tensor("linear_14")]; + tensor var_466 = const()[name = tensor("op_466"), val = tensor([1, 5, 4, 64])]; + tensor var_467 = reshape(shape = var_466, x = var_465)[name = tensor("op_467")]; tensor v_3_perm_0 = const()[name = tensor("v_3_perm_0"), val = tensor([0, 2, 1, 3])]; - tensor input_57 = linear(bias = encoder_g_proj_1_bias, weight = encoder_g_proj_1_weight, x = x_7)[name = tensor("linear_15")]; + tensor input_59 = linear(bias = inner_encoder_g_proj_1_bias, weight = inner_encoder_g_proj_1_weight, x = x_7)[name = tensor("linear_15")]; tensor sqrt_s0_3 = sqrt(x = prev_scale_3)[name = tensor("sqrt_s0_3")]; - tensor s_t_3 = add(x = prev_scale_3, y = encoder__t_index)[name = tensor("s_t_3")]; + tensor s_t_3 = add(x = prev_scale_3, y = inner_encoder__t_index)[name = tensor("s_t_3")]; tensor sqrt_s_t_3 = sqrt(x = s_t_3)[name = tensor("sqrt_s_t_3")]; tensor qk_3_transpose_x_1 = const()[name = tensor("qk_3_transpose_x_1"), val = tensor(false)]; tensor qk_3_transpose_y_1 = const()[name = tensor("qk_3_transpose_y_1"), val = tensor(true)]; - tensor k_3 = transpose(perm = k_3_perm_0, x = var_395)[name = tensor("transpose_50")]; - tensor q_3 = transpose(perm = q_3_perm_0, x = var_387)[name = tensor("transpose_51")]; + tensor k_3 = transpose(perm = k_3_perm_0, x = var_461)[name = tensor("transpose_50")]; + tensor q_3 = transpose(perm = q_3_perm_0, x = var_453)[name = tensor("transpose_51")]; tensor qk_3 = matmul(transpose_x = qk_3_transpose_x_1, transpose_y = qk_3_transpose_y_1, x = q_3, y = k_3)[name = tensor("qk_3")]; - tensor var_411 = const()[name = tensor("op_411"), val = tensor([5, 1])]; - tensor var_412 = reshape(shape = var_411, x = sqrt_s_t_3)[name = tensor("op_412")]; - tensor M_3 = real_div(x = encoder__causal_mask, y = var_412)[name = tensor("M_3")]; - tensor var_414 = mul(x = qk_3, y = M_3)[name = tensor("op_414")]; + tensor var_477 = const()[name = tensor("op_477"), val = tensor([5, 1])]; + tensor var_478 = reshape(shape = var_477, x = sqrt_s_t_3)[name = tensor("op_478")]; + tensor M_3 = real_div(x = inner_encoder__causal_mask, y = var_478)[name = tensor("M_3")]; + tensor var_480 = mul(x = qk_3, y = M_3)[name = tensor("op_480")]; tensor inner_3_transpose_x_0 = const()[name = tensor("inner_3_transpose_x_0"), val = tensor(false)]; tensor inner_3_transpose_y_0 = const()[name = tensor("inner_3_transpose_y_0"), val = tensor(false)]; - tensor v_3 = transpose(perm = v_3_perm_0, x = var_401)[name = tensor("transpose_49")]; - tensor inner_3 = matmul(transpose_x = inner_3_transpose_x_0, transpose_y = inner_3_transpose_y_0, x = var_414, y = v_3)[name = tensor("inner_3")]; - tensor var_416_transpose_x_0 = const()[name = tensor("op_416_transpose_x_0"), val = tensor(false)]; - tensor var_416_transpose_y_0 = const()[name = tensor("op_416_transpose_y_0"), val = tensor(false)]; - tensor var_416 = matmul(transpose_x = var_416_transpose_x_0, transpose_y = var_416_transpose_y_0, x = q_3, y = prev_kv_3)[name = tensor("op_416")]; - tensor var_417 = real_div(x = sqrt_s0_3, y = sqrt_s_t_3)[name = tensor("op_417")]; - tensor var_418 = const()[name = tensor("op_418"), val = tensor([1, 1, 5, 1])]; - tensor var_419 = reshape(shape = var_418, x = var_417)[name = tensor("op_419")]; - tensor cross_3 = mul(x = var_416, y = var_419)[name = tensor("cross_3")]; + tensor v_3 = transpose(perm = v_3_perm_0, x = var_467)[name = tensor("transpose_49")]; + tensor inner_3 = matmul(transpose_x = inner_3_transpose_x_0, transpose_y = inner_3_transpose_y_0, x = var_480, y = v_3)[name = tensor("inner_3")]; + tensor var_482_transpose_x_0 = const()[name = tensor("op_482_transpose_x_0"), val = tensor(false)]; + tensor var_482_transpose_y_0 = const()[name = tensor("op_482_transpose_y_0"), val = tensor(false)]; + tensor var_482 = matmul(transpose_x = var_482_transpose_x_0, transpose_y = var_482_transpose_y_0, x = q_3, y = prev_kv_3)[name = tensor("op_482")]; + tensor var_483 = real_div(x = sqrt_s0_3, y = sqrt_s_t_3)[name = tensor("op_483")]; + tensor var_484 = const()[name = tensor("op_484"), val = tensor([1, 1, 5, 1])]; + tensor var_485 = reshape(shape = var_484, x = var_483)[name = tensor("op_485")]; + tensor cross_3 = mul(x = var_482, y = var_485)[name = tensor("cross_3")]; tensor out_7 = add(x = inner_3, y = cross_3)[name = tensor("out_7")]; - tensor var_422 = mul(x = prev_kv_3, y = sqrt_s0_3)[name = tensor("op_422")]; - tensor var_424_transpose_x_1 = const()[name = tensor("op_424_transpose_x_1"), val = tensor(true)]; - tensor var_424_transpose_y_1 = const()[name = tensor("op_424_transpose_y_1"), val = tensor(false)]; - tensor var_424 = matmul(transpose_x = var_424_transpose_x_1, transpose_y = var_424_transpose_y_1, x = k_3, y = v_3)[name = tensor("op_424")]; - tensor new_kv_unnorm_3 = add(x = var_422, y = var_424)[name = tensor("new_kv_unnorm_3")]; - tensor var_426 = const()[name = tensor("op_426"), val = tensor(0x1.4p+2)]; - tensor new_scale_3 = add(x = prev_scale_3, y = var_426)[name = tensor("new_scale_3")]; - tensor var_428 = sqrt(x = new_scale_3)[name = tensor("op_428")]; - tensor var_429 = real_div(x = new_kv_unnorm_3, y = var_428)[name = tensor("op_429")]; - tensor var_430_perm_0 = const()[name = tensor("op_430_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_488 = mul(x = prev_kv_3, y = sqrt_s0_3)[name = tensor("op_488")]; + tensor var_490_transpose_x_1 = const()[name = tensor("op_490_transpose_x_1"), val = tensor(true)]; + tensor var_490_transpose_y_1 = const()[name = tensor("op_490_transpose_y_1"), val = tensor(false)]; + tensor var_490 = matmul(transpose_x = var_490_transpose_x_1, transpose_y = var_490_transpose_y_1, x = k_3, y = v_3)[name = tensor("op_490")]; + tensor new_kv_unnorm_3 = add(x = var_488, y = var_490)[name = tensor("new_kv_unnorm_3")]; + tensor var_492 = const()[name = tensor("op_492"), val = tensor(0x1.4p+2)]; + tensor new_scale_3 = add(x = prev_scale_3, y = var_492)[name = tensor("new_scale_3")]; + tensor var_494 = sqrt(x = new_scale_3)[name = tensor("op_494")]; + tensor var_495 = real_div(x = new_kv_unnorm_3, y = var_494)[name = tensor("op_495")]; + tensor var_496_perm_0 = const()[name = tensor("op_496_perm_0"), val = tensor([0, 2, 1, 3])]; tensor out_9_axes_0 = const()[name = tensor("out_9_axes_0"), val = tensor([-1])]; - tensor var_430 = transpose(perm = var_430_perm_0, x = out_7)[name = tensor("transpose_48")]; - tensor out_9 = layer_norm(axes = out_9_axes_0, epsilon = var_18, x = var_430)[name = tensor("out_9")]; - tensor var_434 = const()[name = tensor("op_434"), val = tensor([1, 5, 256])]; - tensor out_11 = reshape(shape = var_434, x = out_9)[name = tensor("out_11")]; - tensor var_436 = silu(x = input_57)[name = tensor("op_436")]; - tensor input_59 = mul(x = var_436, y = out_11)[name = tensor("input_59")]; - tensor ret_out_3 = linear(bias = encoder_out_proj_1_bias, weight = encoder_out_proj_1_weight, x = input_59)[name = tensor("linear_16")]; - tensor x_9 = add(x = input_53, y = ret_out_3)[name = tensor("x_9")]; + tensor var_496 = transpose(perm = var_496_perm_0, x = out_7)[name = tensor("transpose_48")]; + tensor out_9 = layer_norm(axes = out_9_axes_0, epsilon = var_84, x = var_496)[name = tensor("out_9")]; + tensor var_500 = const()[name = tensor("op_500"), val = tensor([1, 5, 256])]; + tensor out_11 = reshape(shape = var_500, x = out_9)[name = tensor("out_11")]; + tensor var_502 = silu(x = input_59)[name = tensor("op_502")]; + tensor input_61 = mul(x = var_502, y = out_11)[name = tensor("input_61")]; + tensor ret_out_3 = linear(bias = inner_encoder_out_proj_1_bias, weight = inner_encoder_out_proj_1_weight, x = input_61)[name = tensor("linear_16")]; + tensor x_9 = add(x = input_55, y = ret_out_3)[name = tensor("x_9")]; tensor window_13_begin_0 = const()[name = tensor("window_13_begin_0"), val = tensor([1, 0, 0, 0])]; tensor window_13_end_0 = const()[name = tensor("window_13_end_0"), val = tensor([2, 1, 16, 256])]; tensor window_13_end_mask_0 = const()[name = tensor("window_13_end_mask_0"), val = tensor([false, true, true, true])]; tensor window_13_squeeze_mask_0 = const()[name = tensor("window_13_squeeze_mask_0"), val = tensor([true, false, false, false])]; tensor window_13 = slice_by_index(begin = window_13_begin_0, end = window_13_end_0, end_mask = window_13_end_mask_0, squeeze_mask = window_13_squeeze_mask_0, x = enc_conv_cache)[name = tensor("window_13")]; - tensor var_444_begin_0 = const()[name = tensor("op_444_begin_0"), val = tensor([0, 0, 0])]; - tensor var_444_end_0 = const()[name = tensor("op_444_end_0"), val = tensor([1, 1, 256])]; - tensor var_444_end_mask_0 = const()[name = tensor("op_444_end_mask_0"), val = tensor([true, false, true])]; - tensor var_444 = slice_by_index(begin = var_444_begin_0, end = var_444_end_0, end_mask = var_444_end_mask_0, x = x_9)[name = tensor("op_444")]; - tensor var_447_begin_0 = const()[name = tensor("op_447_begin_0"), val = tensor([0, 1, 0])]; - tensor var_447_end_0 = const()[name = tensor("op_447_end_0"), val = tensor([1, 16, 256])]; - tensor var_447_end_mask_0 = const()[name = tensor("op_447_end_mask_0"), val = tensor([true, true, true])]; - tensor var_447 = slice_by_index(begin = var_447_begin_0, end = var_447_end_0, end_mask = var_447_end_mask_0, x = window_13)[name = tensor("op_447")]; + tensor var_510_begin_0 = const()[name = tensor("op_510_begin_0"), val = tensor([0, 0, 0])]; + tensor var_510_end_0 = const()[name = tensor("op_510_end_0"), val = tensor([1, 1, 256])]; + tensor var_510_end_mask_0 = const()[name = tensor("op_510_end_mask_0"), val = tensor([true, false, true])]; + tensor var_510 = slice_by_index(begin = var_510_begin_0, end = var_510_end_0, end_mask = var_510_end_mask_0, x = x_9)[name = tensor("op_510")]; + tensor var_513_begin_0 = const()[name = tensor("op_513_begin_0"), val = tensor([0, 1, 0])]; + tensor var_513_end_0 = const()[name = tensor("op_513_end_0"), val = tensor([1, 16, 256])]; + tensor var_513_end_mask_0 = const()[name = tensor("op_513_end_mask_0"), val = tensor([true, true, true])]; + tensor var_513 = slice_by_index(begin = var_513_begin_0, end = var_513_end_0, end_mask = var_513_end_mask_0, x = window_13)[name = tensor("op_513")]; tensor window_15_interleave_0 = const()[name = tensor("window_15_interleave_0"), val = tensor(false)]; - tensor window_15 = concat(axis = var_27, interleave = window_15_interleave_0, values = (var_447, var_444))[name = tensor("window_15")]; - tensor var_452_begin_0 = const()[name = tensor("op_452_begin_0"), val = tensor([0, 1, 0])]; - tensor var_452_end_0 = const()[name = tensor("op_452_end_0"), val = tensor([1, 2, 256])]; - tensor var_452_end_mask_0 = const()[name = tensor("op_452_end_mask_0"), val = tensor([true, false, true])]; - tensor var_452 = slice_by_index(begin = var_452_begin_0, end = var_452_end_0, end_mask = var_452_end_mask_0, x = x_9)[name = tensor("op_452")]; - tensor var_455_begin_0 = const()[name = tensor("op_455_begin_0"), val = tensor([0, 1, 0])]; - tensor var_455_end_0 = const()[name = tensor("op_455_end_0"), val = tensor([1, 16, 256])]; - tensor var_455_end_mask_0 = const()[name = tensor("op_455_end_mask_0"), val = tensor([true, true, true])]; - tensor var_455 = slice_by_index(begin = var_455_begin_0, end = var_455_end_0, end_mask = var_455_end_mask_0, x = window_15)[name = tensor("op_455")]; + tensor window_15 = concat(axis = var_93, interleave = window_15_interleave_0, values = (var_513, var_510))[name = tensor("window_15")]; + tensor var_518_begin_0 = const()[name = tensor("op_518_begin_0"), val = tensor([0, 1, 0])]; + tensor var_518_end_0 = const()[name = tensor("op_518_end_0"), val = tensor([1, 2, 256])]; + tensor var_518_end_mask_0 = const()[name = tensor("op_518_end_mask_0"), val = tensor([true, false, true])]; + tensor var_518 = slice_by_index(begin = var_518_begin_0, end = var_518_end_0, end_mask = var_518_end_mask_0, x = x_9)[name = tensor("op_518")]; + tensor var_521_begin_0 = const()[name = tensor("op_521_begin_0"), val = tensor([0, 1, 0])]; + tensor var_521_end_0 = const()[name = tensor("op_521_end_0"), val = tensor([1, 16, 256])]; + tensor var_521_end_mask_0 = const()[name = tensor("op_521_end_mask_0"), val = tensor([true, true, true])]; + tensor var_521 = slice_by_index(begin = var_521_begin_0, end = var_521_end_0, end_mask = var_521_end_mask_0, x = window_15)[name = tensor("op_521")]; tensor window_17_interleave_0 = const()[name = tensor("window_17_interleave_0"), val = tensor(false)]; - tensor window_17 = concat(axis = var_27, interleave = window_17_interleave_0, values = (var_455, var_452))[name = tensor("window_17")]; - tensor var_460_begin_0 = const()[name = tensor("op_460_begin_0"), val = tensor([0, 2, 0])]; - tensor var_460_end_0 = const()[name = tensor("op_460_end_0"), val = tensor([1, 3, 256])]; - tensor var_460_end_mask_0 = const()[name = tensor("op_460_end_mask_0"), val = tensor([true, false, true])]; - tensor var_460 = slice_by_index(begin = var_460_begin_0, end = var_460_end_0, end_mask = var_460_end_mask_0, x = x_9)[name = tensor("op_460")]; - tensor var_463_begin_0 = const()[name = tensor("op_463_begin_0"), val = tensor([0, 1, 0])]; - tensor var_463_end_0 = const()[name = tensor("op_463_end_0"), val = tensor([1, 16, 256])]; - tensor var_463_end_mask_0 = const()[name = tensor("op_463_end_mask_0"), val = tensor([true, true, true])]; - tensor var_463 = slice_by_index(begin = var_463_begin_0, end = var_463_end_0, end_mask = var_463_end_mask_0, x = window_17)[name = tensor("op_463")]; + tensor window_17 = concat(axis = var_93, interleave = window_17_interleave_0, values = (var_521, var_518))[name = tensor("window_17")]; + tensor var_526_begin_0 = const()[name = tensor("op_526_begin_0"), val = tensor([0, 2, 0])]; + tensor var_526_end_0 = const()[name = tensor("op_526_end_0"), val = tensor([1, 3, 256])]; + tensor var_526_end_mask_0 = const()[name = tensor("op_526_end_mask_0"), val = tensor([true, false, true])]; + tensor var_526 = slice_by_index(begin = var_526_begin_0, end = var_526_end_0, end_mask = var_526_end_mask_0, x = x_9)[name = tensor("op_526")]; + tensor var_529_begin_0 = const()[name = tensor("op_529_begin_0"), val = tensor([0, 1, 0])]; + tensor var_529_end_0 = const()[name = tensor("op_529_end_0"), val = tensor([1, 16, 256])]; + tensor var_529_end_mask_0 = const()[name = tensor("op_529_end_mask_0"), val = tensor([true, true, true])]; + tensor var_529 = slice_by_index(begin = var_529_begin_0, end = var_529_end_0, end_mask = var_529_end_mask_0, x = window_17)[name = tensor("op_529")]; tensor window_19_interleave_0 = const()[name = tensor("window_19_interleave_0"), val = tensor(false)]; - tensor window_19 = concat(axis = var_27, interleave = window_19_interleave_0, values = (var_463, var_460))[name = tensor("window_19")]; - tensor var_468_begin_0 = const()[name = tensor("op_468_begin_0"), val = tensor([0, 3, 0])]; - tensor var_468_end_0 = const()[name = tensor("op_468_end_0"), val = tensor([1, 4, 256])]; - tensor var_468_end_mask_0 = const()[name = tensor("op_468_end_mask_0"), val = tensor([true, false, true])]; - tensor var_468 = slice_by_index(begin = var_468_begin_0, end = var_468_end_0, end_mask = var_468_end_mask_0, x = x_9)[name = tensor("op_468")]; - tensor var_471_begin_0 = const()[name = tensor("op_471_begin_0"), val = tensor([0, 1, 0])]; - tensor var_471_end_0 = const()[name = tensor("op_471_end_0"), val = tensor([1, 16, 256])]; - tensor var_471_end_mask_0 = const()[name = tensor("op_471_end_mask_0"), val = tensor([true, true, true])]; - tensor var_471 = slice_by_index(begin = var_471_begin_0, end = var_471_end_0, end_mask = var_471_end_mask_0, x = window_19)[name = tensor("op_471")]; + tensor window_19 = concat(axis = var_93, interleave = window_19_interleave_0, values = (var_529, var_526))[name = tensor("window_19")]; + tensor var_534_begin_0 = const()[name = tensor("op_534_begin_0"), val = tensor([0, 3, 0])]; + tensor var_534_end_0 = const()[name = tensor("op_534_end_0"), val = tensor([1, 4, 256])]; + tensor var_534_end_mask_0 = const()[name = tensor("op_534_end_mask_0"), val = tensor([true, false, true])]; + tensor var_534 = slice_by_index(begin = var_534_begin_0, end = var_534_end_0, end_mask = var_534_end_mask_0, x = x_9)[name = tensor("op_534")]; + tensor var_537_begin_0 = const()[name = tensor("op_537_begin_0"), val = tensor([0, 1, 0])]; + tensor var_537_end_0 = const()[name = tensor("op_537_end_0"), val = tensor([1, 16, 256])]; + tensor var_537_end_mask_0 = const()[name = tensor("op_537_end_mask_0"), val = tensor([true, true, true])]; + tensor var_537 = slice_by_index(begin = var_537_begin_0, end = var_537_end_0, end_mask = var_537_end_mask_0, x = window_19)[name = tensor("op_537")]; tensor window_21_interleave_0 = const()[name = tensor("window_21_interleave_0"), val = tensor(false)]; - tensor window_21 = concat(axis = var_27, interleave = window_21_interleave_0, values = (var_471, var_468))[name = tensor("window_21")]; - tensor var_476_begin_0 = const()[name = tensor("op_476_begin_0"), val = tensor([0, 4, 0])]; - tensor var_476_end_0 = const()[name = tensor("op_476_end_0"), val = tensor([1, 1, 256])]; - tensor var_476_end_mask_0 = const()[name = tensor("op_476_end_mask_0"), val = tensor([true, true, true])]; - tensor var_476 = slice_by_index(begin = var_476_begin_0, end = var_476_end_0, end_mask = var_476_end_mask_0, x = x_9)[name = tensor("op_476")]; - tensor var_479_begin_0 = const()[name = tensor("op_479_begin_0"), val = tensor([0, 1, 0])]; - tensor var_479_end_0 = const()[name = tensor("op_479_end_0"), val = tensor([1, 16, 256])]; - tensor var_479_end_mask_0 = const()[name = tensor("op_479_end_mask_0"), val = tensor([true, true, true])]; - tensor var_479 = slice_by_index(begin = var_479_begin_0, end = var_479_end_0, end_mask = var_479_end_mask_0, x = window_21)[name = tensor("op_479")]; + tensor window_21 = concat(axis = var_93, interleave = window_21_interleave_0, values = (var_537, var_534))[name = tensor("window_21")]; + tensor var_542_begin_0 = const()[name = tensor("op_542_begin_0"), val = tensor([0, 4, 0])]; + tensor var_542_end_0 = const()[name = tensor("op_542_end_0"), val = tensor([1, 1, 256])]; + tensor var_542_end_mask_0 = const()[name = tensor("op_542_end_mask_0"), val = tensor([true, true, true])]; + tensor var_542 = slice_by_index(begin = var_542_begin_0, end = var_542_end_0, end_mask = var_542_end_mask_0, x = x_9)[name = tensor("op_542")]; + tensor var_545_begin_0 = const()[name = tensor("op_545_begin_0"), val = tensor([0, 1, 0])]; + tensor var_545_end_0 = const()[name = tensor("op_545_end_0"), val = tensor([1, 16, 256])]; + tensor var_545_end_mask_0 = const()[name = tensor("op_545_end_mask_0"), val = tensor([true, true, true])]; + tensor var_545 = slice_by_index(begin = var_545_begin_0, end = var_545_end_0, end_mask = var_545_end_mask_0, x = window_21)[name = tensor("op_545")]; tensor window_23_interleave_0 = const()[name = tensor("window_23_interleave_0"), val = tensor(false)]; - tensor window_23 = concat(axis = var_27, interleave = window_23_interleave_0, values = (var_479, var_476))[name = tensor("window_23")]; - tensor input_61_interleave_0 = const()[name = tensor("input_61_interleave_0"), val = tensor(false)]; - tensor input_61 = concat(axis = var_24, interleave = input_61_interleave_0, values = (window_15, window_17, window_19, window_21, window_23))[name = tensor("input_61")]; + tensor window_23 = concat(axis = var_93, interleave = window_23_interleave_0, values = (var_545, var_542))[name = tensor("window_23")]; + tensor input_63_interleave_0 = const()[name = tensor("input_63_interleave_0"), val = tensor(false)]; + tensor input_63 = concat(axis = var_79, interleave = input_63_interleave_0, values = (window_15, window_17, window_19, window_21, window_23))[name = tensor("input_63")]; tensor x_11_axes_0 = const()[name = tensor("x_11_axes_0"), val = tensor([-1])]; - tensor x_11 = layer_norm(axes = x_11_axes_0, beta = encoder_conv_module_1_sequential_0_bias, epsilon = var_29, gamma = encoder_conv_module_1_sequential_0_weight, x = input_61)[name = tensor("x_11")]; - tensor input_63_perm_0 = const()[name = tensor("input_63_perm_0"), val = tensor([0, 2, 1])]; + tensor x_11 = layer_norm(axes = x_11_axes_0, beta = inner_encoder_conv_module_1_sequential_0_bias, epsilon = var_76, gamma = inner_encoder_conv_module_1_sequential_0_weight, x = input_63)[name = tensor("x_11")]; + tensor input_65_perm_0 = const()[name = tensor("input_65_perm_0"), val = tensor([0, 2, 1])]; tensor inputs_13_pad_type_0 = const()[name = tensor("inputs_13_pad_type_0"), val = tensor("valid")]; tensor inputs_13_strides_0 = const()[name = tensor("inputs_13_strides_0"), val = tensor([1])]; tensor inputs_13_pad_0 = const()[name = tensor("inputs_13_pad_0"), val = tensor([0, 0])]; tensor inputs_13_dilations_0 = const()[name = tensor("inputs_13_dilations_0"), val = tensor([1])]; tensor inputs_13_groups_0 = const()[name = tensor("inputs_13_groups_0"), val = tensor(1)]; - tensor input_63 = transpose(perm = input_63_perm_0, x = x_11)[name = tensor("transpose_47")]; - tensor inputs_13 = conv(bias = encoder_conv_module_1_sequential_2_conv_bias, dilations = inputs_13_dilations_0, groups = inputs_13_groups_0, pad = inputs_13_pad_0, pad_type = inputs_13_pad_type_0, strides = inputs_13_strides_0, weight = encoder_conv_module_1_sequential_2_conv_weight, x = input_63)[name = tensor("inputs_13")]; - tensor var_504_split_sizes_0 = const()[name = tensor("op_504_split_sizes_0"), val = tensor([256, 256])]; - tensor var_504_axis_0 = const()[name = tensor("op_504_axis_0"), val = tensor(1)]; - tensor var_504_0, tensor var_504_1 = split(axis = var_504_axis_0, split_sizes = var_504_split_sizes_0, x = inputs_13)[name = tensor("op_504")]; - tensor var_506 = sigmoid(x = var_504_1)[name = tensor("op_506")]; - tensor inputs_15 = mul(x = var_504_0, y = var_506)[name = tensor("inputs_15")]; + tensor input_65 = transpose(perm = input_65_perm_0, x = x_11)[name = tensor("transpose_47")]; + tensor inputs_13 = conv(bias = inner_encoder_conv_module_1_sequential_2_conv_bias, dilations = inputs_13_dilations_0, groups = inputs_13_groups_0, pad = inputs_13_pad_0, pad_type = inputs_13_pad_type_0, strides = inputs_13_strides_0, weight = inner_encoder_conv_module_1_sequential_2_conv_weight, x = input_65)[name = tensor("inputs_13")]; + tensor var_570_split_sizes_0 = const()[name = tensor("op_570_split_sizes_0"), val = tensor([256, 256])]; + tensor var_570_axis_0 = const()[name = tensor("op_570_axis_0"), val = tensor(1)]; + tensor var_570_0, tensor var_570_1 = split(axis = var_570_axis_0, split_sizes = var_570_split_sizes_0, x = inputs_13)[name = tensor("op_570")]; + tensor var_572 = sigmoid(x = var_570_1)[name = tensor("op_572")]; + tensor inputs_15 = mul(x = var_570_0, y = var_572)[name = tensor("inputs_15")]; tensor outputs_aug_3_pad_type_0 = const()[name = tensor("outputs_aug_3_pad_type_0"), val = tensor("custom")]; tensor outputs_aug_3_pad_0 = const()[name = tensor("outputs_aug_3_pad_0"), val = tensor([15, 15])]; tensor outputs_aug_3_groups_0 = const()[name = tensor("outputs_aug_3_groups_0"), val = tensor(256)]; tensor outputs_aug_3_strides_0 = const()[name = tensor("outputs_aug_3_strides_0"), val = tensor([1])]; tensor outputs_aug_3_dilations_0 = const()[name = tensor("outputs_aug_3_dilations_0"), val = tensor([1])]; - tensor outputs_aug_3 = conv(dilations = outputs_aug_3_dilations_0, groups = outputs_aug_3_groups_0, pad = outputs_aug_3_pad_0, pad_type = outputs_aug_3_pad_type_0, strides = outputs_aug_3_strides_0, weight = encoder_conv_module_1_sequential_4_conv_weight, x = inputs_15)[name = tensor("outputs_aug_3")]; - tensor input_65_begin_0 = const()[name = tensor("input_65_begin_0"), val = tensor([0, 0, 0])]; - tensor input_65_end_0 = const()[name = tensor("input_65_end_0"), val = tensor([5, 256, 16])]; - tensor input_65_end_mask_0 = const()[name = tensor("input_65_end_mask_0"), val = tensor([true, true, false])]; - tensor input_65 = slice_by_index(begin = input_65_begin_0, end = input_65_end_0, end_mask = input_65_end_mask_0, x = outputs_aug_3)[name = tensor("input_65")]; - tensor inputs_17 = batch_norm(beta = encoder_conv_module_1_sequential_5_bias, epsilon = var_29, gamma = encoder_conv_module_1_sequential_5_weight, mean = encoder_conv_module_1_sequential_5_running_mean, variance = encoder_conv_module_1_sequential_5_running_var, x = input_65)[name = tensor("inputs_17")]; - tensor input_67 = silu(x = inputs_17)[name = tensor("input_67")]; - tensor input_69_pad_type_0 = const()[name = tensor("input_69_pad_type_0"), val = tensor("valid")]; - tensor input_69_strides_0 = const()[name = tensor("input_69_strides_0"), val = tensor([1])]; - tensor input_69_pad_0 = const()[name = tensor("input_69_pad_0"), val = tensor([0, 0])]; - tensor input_69_dilations_0 = const()[name = tensor("input_69_dilations_0"), val = tensor([1])]; - tensor input_69_groups_0 = const()[name = tensor("input_69_groups_0"), val = tensor(1)]; - tensor input_69 = conv(bias = encoder_conv_module_1_sequential_7_conv_bias, dilations = input_69_dilations_0, groups = input_69_groups_0, pad = input_69_pad_0, pad_type = input_69_pad_type_0, strides = input_69_strides_0, weight = encoder_conv_module_1_sequential_7_conv_weight, x = input_67)[name = tensor("input_69")]; + tensor outputs_aug_3 = conv(dilations = outputs_aug_3_dilations_0, groups = outputs_aug_3_groups_0, pad = outputs_aug_3_pad_0, pad_type = outputs_aug_3_pad_type_0, strides = outputs_aug_3_strides_0, weight = inner_encoder_conv_module_1_sequential_4_conv_weight, x = inputs_15)[name = tensor("outputs_aug_3")]; + tensor input_67_begin_0 = const()[name = tensor("input_67_begin_0"), val = tensor([0, 0, 0])]; + tensor input_67_end_0 = const()[name = tensor("input_67_end_0"), val = tensor([5, 256, 16])]; + tensor input_67_end_mask_0 = const()[name = tensor("input_67_end_mask_0"), val = tensor([true, true, false])]; + tensor input_67 = slice_by_index(begin = input_67_begin_0, end = input_67_end_0, end_mask = input_67_end_mask_0, x = outputs_aug_3)[name = tensor("input_67")]; + tensor inputs_17 = batch_norm(beta = inner_encoder_conv_module_1_sequential_5_bias, epsilon = var_76, gamma = inner_encoder_conv_module_1_sequential_5_weight, mean = inner_encoder_conv_module_1_sequential_5_running_mean, variance = inner_encoder_conv_module_1_sequential_5_running_var, x = input_67)[name = tensor("inputs_17")]; + tensor input_69 = silu(x = inputs_17)[name = tensor("input_69")]; + tensor input_71_pad_type_0 = const()[name = tensor("input_71_pad_type_0"), val = tensor("valid")]; + tensor input_71_strides_0 = const()[name = tensor("input_71_strides_0"), val = tensor([1])]; + tensor input_71_pad_0 = const()[name = tensor("input_71_pad_0"), val = tensor([0, 0])]; + tensor input_71_dilations_0 = const()[name = tensor("input_71_dilations_0"), val = tensor([1])]; + tensor input_71_groups_0 = const()[name = tensor("input_71_groups_0"), val = tensor(1)]; + tensor input_71 = conv(bias = inner_encoder_conv_module_1_sequential_7_conv_bias, dilations = input_71_dilations_0, groups = input_71_groups_0, pad = input_71_pad_0, pad_type = input_71_pad_type_0, strides = input_71_strides_0, weight = inner_encoder_conv_module_1_sequential_7_conv_weight, x = input_69)[name = tensor("input_71")]; tensor conv_out_3_perm_0 = const()[name = tensor("conv_out_3_perm_0"), val = tensor([0, 2, 1])]; - tensor var_537_begin_0 = const()[name = tensor("op_537_begin_0"), val = tensor([0, -1, 0])]; - tensor var_537_end_0 = const()[name = tensor("op_537_end_0"), val = tensor([5, 16, 256])]; - tensor var_537_end_mask_0 = const()[name = tensor("op_537_end_mask_0"), val = tensor([true, true, true])]; - tensor conv_out_3 = transpose(perm = conv_out_3_perm_0, x = input_69)[name = tensor("transpose_46")]; - tensor var_537 = slice_by_index(begin = var_537_begin_0, end = var_537_end_0, end_mask = var_537_end_mask_0, x = conv_out_3)[name = tensor("op_537")]; - tensor var_539_perm_0 = const()[name = tensor("op_539_perm_0"), val = tensor([1, 0, 2])]; - tensor var_539 = transpose(perm = var_539_perm_0, x = var_537)[name = tensor("transpose_45")]; - tensor input_71 = add(x = x_9, y = var_539)[name = tensor("input_71")]; - tensor input_73_axes_0 = const()[name = tensor("input_73_axes_0"), val = tensor([-1])]; - tensor input_73 = layer_norm(axes = input_73_axes_0, beta = encoder_ffn2_1_module_sequential_0_bias, epsilon = var_29, gamma = encoder_ffn2_1_module_sequential_0_weight, x = input_71)[name = tensor("input_73")]; - tensor inputs_19 = linear(bias = encoder_ffn2_1_module_sequential_1_linear_bias, weight = encoder_ffn2_1_module_sequential_1_linear_weight, x = input_73)[name = tensor("linear_17")]; - tensor input_75 = silu(x = inputs_19)[name = tensor("input_75")]; - tensor input_79 = linear(bias = encoder_ffn2_1_module_sequential_4_linear_bias, weight = encoder_ffn2_1_module_sequential_4_linear_weight, x = input_75)[name = tensor("linear_18")]; - tensor var_562 = const()[name = tensor("op_562"), val = tensor(0x1p-1)]; - tensor var_563 = mul(x = input_79, y = var_562)[name = tensor("op_563")]; - tensor input_81 = add(x = var_563, y = input_71)[name = tensor("input_81")]; - tensor input_83_axes_0 = const()[name = tensor("input_83_axes_0"), val = tensor([-1])]; - tensor input_83 = layer_norm(axes = input_83_axes_0, beta = encoder_layer_norm_1_bias, epsilon = var_29, gamma = encoder_layer_norm_1_weight, x = input_81)[name = tensor("input_83")]; + tensor var_603_begin_0 = const()[name = tensor("op_603_begin_0"), val = tensor([0, -1, 0])]; + tensor var_603_end_0 = const()[name = tensor("op_603_end_0"), val = tensor([5, 16, 256])]; + tensor var_603_end_mask_0 = const()[name = tensor("op_603_end_mask_0"), val = tensor([true, true, true])]; + tensor conv_out_3 = transpose(perm = conv_out_3_perm_0, x = input_71)[name = tensor("transpose_46")]; + tensor var_603 = slice_by_index(begin = var_603_begin_0, end = var_603_end_0, end_mask = var_603_end_mask_0, x = conv_out_3)[name = tensor("op_603")]; + tensor var_605_perm_0 = const()[name = tensor("op_605_perm_0"), val = tensor([1, 0, 2])]; + tensor var_605 = transpose(perm = var_605_perm_0, x = var_603)[name = tensor("transpose_45")]; + tensor input_73 = add(x = x_9, y = var_605)[name = tensor("input_73")]; + tensor input_75_axes_0 = const()[name = tensor("input_75_axes_0"), val = tensor([-1])]; + tensor input_75 = layer_norm(axes = input_75_axes_0, beta = inner_encoder_ffn2_1_module_sequential_0_bias, epsilon = var_76, gamma = inner_encoder_ffn2_1_module_sequential_0_weight, x = input_73)[name = tensor("input_75")]; + tensor inputs_19 = linear(bias = inner_encoder_ffn2_1_module_sequential_1_linear_bias, weight = inner_encoder_ffn2_1_module_sequential_1_linear_weight, x = input_75)[name = tensor("linear_17")]; + tensor input_77 = silu(x = inputs_19)[name = tensor("input_77")]; + tensor input_81 = linear(bias = inner_encoder_ffn2_1_module_sequential_4_linear_bias, weight = inner_encoder_ffn2_1_module_sequential_4_linear_weight, x = input_77)[name = tensor("linear_18")]; + tensor var_628 = const()[name = tensor("op_628"), val = tensor(0x1p-1)]; + tensor var_629 = mul(x = input_81, y = var_628)[name = tensor("op_629")]; + tensor input_83 = add(x = var_629, y = input_73)[name = tensor("input_83")]; tensor input_85_axes_0 = const()[name = tensor("input_85_axes_0"), val = tensor([-1])]; - tensor input_85 = layer_norm(axes = input_85_axes_0, beta = encoder_ffn1_2_module_sequential_0_bias, epsilon = var_29, gamma = encoder_ffn1_2_module_sequential_0_weight, x = input_83)[name = tensor("input_85")]; - tensor inputs_21 = linear(bias = encoder_ffn1_2_module_sequential_1_linear_bias, weight = encoder_ffn1_2_module_sequential_1_linear_weight, x = input_85)[name = tensor("linear_19")]; - tensor input_87 = silu(x = inputs_21)[name = tensor("input_87")]; - tensor input_91 = linear(bias = encoder_ffn1_2_module_sequential_4_linear_bias, weight = encoder_ffn1_2_module_sequential_4_linear_weight, x = input_87)[name = tensor("linear_20")]; - tensor var_592 = const()[name = tensor("op_592"), val = tensor(0x1p-1)]; - tensor var_593 = mul(x = input_91, y = var_592)[name = tensor("op_593")]; - tensor input_93 = add(x = var_593, y = input_83)[name = tensor("input_93")]; + tensor input_85 = layer_norm(axes = input_85_axes_0, beta = inner_encoder_layer_norm_1_bias, epsilon = var_76, gamma = inner_encoder_layer_norm_1_weight, x = input_83)[name = tensor("input_85")]; + tensor input_87_axes_0 = const()[name = tensor("input_87_axes_0"), val = tensor([-1])]; + tensor input_87 = layer_norm(axes = input_87_axes_0, beta = inner_encoder_ffn1_2_module_sequential_0_bias, epsilon = var_76, gamma = inner_encoder_ffn1_2_module_sequential_0_weight, x = input_85)[name = tensor("input_87")]; + tensor inputs_21 = linear(bias = inner_encoder_ffn1_2_module_sequential_1_linear_bias, weight = inner_encoder_ffn1_2_module_sequential_1_linear_weight, x = input_87)[name = tensor("linear_19")]; + tensor input_89 = silu(x = inputs_21)[name = tensor("input_89")]; + tensor input_93 = linear(bias = inner_encoder_ffn1_2_module_sequential_4_linear_bias, weight = inner_encoder_ffn1_2_module_sequential_4_linear_weight, x = input_89)[name = tensor("linear_20")]; + tensor var_658 = const()[name = tensor("op_658"), val = tensor(0x1p-1)]; + tensor var_659 = mul(x = input_93, y = var_658)[name = tensor("op_659")]; + tensor input_95 = add(x = var_659, y = input_85)[name = tensor("input_95")]; tensor x_13_axes_0 = const()[name = tensor("x_13_axes_0"), val = tensor([-1])]; - tensor x_13 = layer_norm(axes = x_13_axes_0, beta = encoder_ret_lns_2_bias, epsilon = var_29, gamma = encoder_ret_lns_2_weight, x = input_93)[name = tensor("x_13")]; + tensor x_13 = layer_norm(axes = x_13_axes_0, beta = inner_encoder_ret_lns_2_bias, epsilon = var_76, gamma = inner_encoder_ret_lns_2_weight, x = input_95)[name = tensor("x_13")]; tensor prev_kv_5_begin_0 = const()[name = tensor("prev_kv_5_begin_0"), val = tensor([2, 0, 0, 0, 0])]; tensor prev_kv_5_end_0 = const()[name = tensor("prev_kv_5_end_0"), val = tensor([3, 1, 4, 64, 64])]; tensor prev_kv_5_end_mask_0 = const()[name = tensor("prev_kv_5_end_mask_0"), val = tensor([false, true, true, true, true])]; @@ -613,183 +639,183 @@ program(1.0) tensor prev_scale_5_end_mask_0 = const()[name = tensor("prev_scale_5_end_mask_0"), val = tensor([false, true])]; tensor prev_scale_5_squeeze_mask_0 = const()[name = tensor("prev_scale_5_squeeze_mask_0"), val = tensor([true, false])]; tensor prev_scale_5 = slice_by_index(begin = prev_scale_5_begin_0, end = prev_scale_5_end_0, end_mask = prev_scale_5_end_mask_0, squeeze_mask = prev_scale_5_squeeze_mask_0, x = enc_scale)[name = tensor("prev_scale_5")]; - tensor var_607 = linear(bias = encoder_q_proj_2_bias, weight = encoder_q_proj_2_weight, x = x_13)[name = tensor("linear_21")]; - tensor var_608 = const()[name = tensor("op_608"), val = tensor([1, 5, 4, 64])]; - tensor var_609 = reshape(shape = var_608, x = var_607)[name = tensor("op_609")]; + tensor var_673 = linear(bias = inner_encoder_q_proj_2_bias, weight = inner_encoder_q_proj_2_weight, x = x_13)[name = tensor("linear_21")]; + tensor var_674 = const()[name = tensor("op_674"), val = tensor([1, 5, 4, 64])]; + tensor var_675 = reshape(shape = var_674, x = var_673)[name = tensor("op_675")]; tensor q_5_perm_0 = const()[name = tensor("q_5_perm_0"), val = tensor([0, 2, 1, 3])]; - tensor var_613 = linear(bias = encoder_k_proj_2_bias, weight = encoder_k_proj_2_weight, x = x_13)[name = tensor("linear_22")]; - tensor var_614 = const()[name = tensor("op_614"), val = tensor(0x1p-3)]; - tensor var_615 = mul(x = var_613, y = var_614)[name = tensor("op_615")]; - tensor var_616 = const()[name = tensor("op_616"), val = tensor([1, 5, 4, 64])]; - tensor var_617 = reshape(shape = var_616, x = var_615)[name = tensor("op_617")]; + tensor var_679 = linear(bias = inner_encoder_k_proj_2_bias, weight = inner_encoder_k_proj_2_weight, x = x_13)[name = tensor("linear_22")]; + tensor var_680 = const()[name = tensor("op_680"), val = tensor(0x1p-3)]; + tensor var_681 = mul(x = var_679, y = var_680)[name = tensor("op_681")]; + tensor var_682 = const()[name = tensor("op_682"), val = tensor([1, 5, 4, 64])]; + tensor var_683 = reshape(shape = var_682, x = var_681)[name = tensor("op_683")]; tensor k_5_perm_0 = const()[name = tensor("k_5_perm_0"), val = tensor([0, 2, 1, 3])]; - tensor var_621 = linear(bias = encoder_v_proj_2_bias, weight = encoder_v_proj_2_weight, x = x_13)[name = tensor("linear_23")]; - tensor var_622 = const()[name = tensor("op_622"), val = tensor([1, 5, 4, 64])]; - tensor var_623 = reshape(shape = var_622, x = var_621)[name = tensor("op_623")]; + tensor var_687 = linear(bias = inner_encoder_v_proj_2_bias, weight = inner_encoder_v_proj_2_weight, x = x_13)[name = tensor("linear_23")]; + tensor var_688 = const()[name = tensor("op_688"), val = tensor([1, 5, 4, 64])]; + tensor var_689 = reshape(shape = var_688, x = var_687)[name = tensor("op_689")]; tensor v_5_perm_0 = const()[name = tensor("v_5_perm_0"), val = tensor([0, 2, 1, 3])]; - tensor input_97 = linear(bias = encoder_g_proj_2_bias, weight = encoder_g_proj_2_weight, x = x_13)[name = tensor("linear_24")]; + tensor input_99 = linear(bias = inner_encoder_g_proj_2_bias, weight = inner_encoder_g_proj_2_weight, x = x_13)[name = tensor("linear_24")]; tensor sqrt_s0_5 = sqrt(x = prev_scale_5)[name = tensor("sqrt_s0_5")]; - tensor s_t_5 = add(x = prev_scale_5, y = encoder__t_index)[name = tensor("s_t_5")]; + tensor s_t_5 = add(x = prev_scale_5, y = inner_encoder__t_index)[name = tensor("s_t_5")]; tensor sqrt_s_t_5 = sqrt(x = s_t_5)[name = tensor("sqrt_s_t_5")]; tensor qk_5_transpose_x_1 = const()[name = tensor("qk_5_transpose_x_1"), val = tensor(false)]; tensor qk_5_transpose_y_1 = const()[name = tensor("qk_5_transpose_y_1"), val = tensor(true)]; - tensor k_5 = transpose(perm = k_5_perm_0, x = var_617)[name = tensor("transpose_43")]; - tensor q_5 = transpose(perm = q_5_perm_0, x = var_609)[name = tensor("transpose_44")]; + tensor k_5 = transpose(perm = k_5_perm_0, x = var_683)[name = tensor("transpose_43")]; + tensor q_5 = transpose(perm = q_5_perm_0, x = var_675)[name = tensor("transpose_44")]; tensor qk_5 = matmul(transpose_x = qk_5_transpose_x_1, transpose_y = qk_5_transpose_y_1, x = q_5, y = k_5)[name = tensor("qk_5")]; - tensor var_633 = const()[name = tensor("op_633"), val = tensor([5, 1])]; - tensor var_634 = reshape(shape = var_633, x = sqrt_s_t_5)[name = tensor("op_634")]; - tensor M_5 = real_div(x = encoder__causal_mask, y = var_634)[name = tensor("M_5")]; - tensor var_636 = mul(x = qk_5, y = M_5)[name = tensor("op_636")]; + tensor var_699 = const()[name = tensor("op_699"), val = tensor([5, 1])]; + tensor var_700 = reshape(shape = var_699, x = sqrt_s_t_5)[name = tensor("op_700")]; + tensor M_5 = real_div(x = inner_encoder__causal_mask, y = var_700)[name = tensor("M_5")]; + tensor var_702 = mul(x = qk_5, y = M_5)[name = tensor("op_702")]; tensor inner_5_transpose_x_0 = const()[name = tensor("inner_5_transpose_x_0"), val = tensor(false)]; tensor inner_5_transpose_y_0 = const()[name = tensor("inner_5_transpose_y_0"), val = tensor(false)]; - tensor v_5 = transpose(perm = v_5_perm_0, x = var_623)[name = tensor("transpose_42")]; - tensor inner_5 = matmul(transpose_x = inner_5_transpose_x_0, transpose_y = inner_5_transpose_y_0, x = var_636, y = v_5)[name = tensor("inner_5")]; - tensor var_638_transpose_x_0 = const()[name = tensor("op_638_transpose_x_0"), val = tensor(false)]; - tensor var_638_transpose_y_0 = const()[name = tensor("op_638_transpose_y_0"), val = tensor(false)]; - tensor var_638 = matmul(transpose_x = var_638_transpose_x_0, transpose_y = var_638_transpose_y_0, x = q_5, y = prev_kv_5)[name = tensor("op_638")]; - tensor var_639 = real_div(x = sqrt_s0_5, y = sqrt_s_t_5)[name = tensor("op_639")]; - tensor var_640 = const()[name = tensor("op_640"), val = tensor([1, 1, 5, 1])]; - tensor var_641 = reshape(shape = var_640, x = var_639)[name = tensor("op_641")]; - tensor cross_5 = mul(x = var_638, y = var_641)[name = tensor("cross_5")]; + tensor v_5 = transpose(perm = v_5_perm_0, x = var_689)[name = tensor("transpose_42")]; + tensor inner_5 = matmul(transpose_x = inner_5_transpose_x_0, transpose_y = inner_5_transpose_y_0, x = var_702, y = v_5)[name = tensor("inner_5")]; + tensor var_704_transpose_x_0 = const()[name = tensor("op_704_transpose_x_0"), val = tensor(false)]; + tensor var_704_transpose_y_0 = const()[name = tensor("op_704_transpose_y_0"), val = tensor(false)]; + tensor var_704 = matmul(transpose_x = var_704_transpose_x_0, transpose_y = var_704_transpose_y_0, x = q_5, y = prev_kv_5)[name = tensor("op_704")]; + tensor var_705 = real_div(x = sqrt_s0_5, y = sqrt_s_t_5)[name = tensor("op_705")]; + tensor var_706 = const()[name = tensor("op_706"), val = tensor([1, 1, 5, 1])]; + tensor var_707 = reshape(shape = var_706, x = var_705)[name = tensor("op_707")]; + tensor cross_5 = mul(x = var_704, y = var_707)[name = tensor("cross_5")]; tensor out_13 = add(x = inner_5, y = cross_5)[name = tensor("out_13")]; - tensor var_644 = mul(x = prev_kv_5, y = sqrt_s0_5)[name = tensor("op_644")]; - tensor var_646_transpose_x_1 = const()[name = tensor("op_646_transpose_x_1"), val = tensor(true)]; - tensor var_646_transpose_y_1 = const()[name = tensor("op_646_transpose_y_1"), val = tensor(false)]; - tensor var_646 = matmul(transpose_x = var_646_transpose_x_1, transpose_y = var_646_transpose_y_1, x = k_5, y = v_5)[name = tensor("op_646")]; - tensor new_kv_unnorm_5 = add(x = var_644, y = var_646)[name = tensor("new_kv_unnorm_5")]; - tensor var_648 = const()[name = tensor("op_648"), val = tensor(0x1.4p+2)]; - tensor new_scale_5 = add(x = prev_scale_5, y = var_648)[name = tensor("new_scale_5")]; - tensor var_650 = sqrt(x = new_scale_5)[name = tensor("op_650")]; - tensor var_651 = real_div(x = new_kv_unnorm_5, y = var_650)[name = tensor("op_651")]; - tensor var_652_perm_0 = const()[name = tensor("op_652_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_710 = mul(x = prev_kv_5, y = sqrt_s0_5)[name = tensor("op_710")]; + tensor var_712_transpose_x_1 = const()[name = tensor("op_712_transpose_x_1"), val = tensor(true)]; + tensor var_712_transpose_y_1 = const()[name = tensor("op_712_transpose_y_1"), val = tensor(false)]; + tensor var_712 = matmul(transpose_x = var_712_transpose_x_1, transpose_y = var_712_transpose_y_1, x = k_5, y = v_5)[name = tensor("op_712")]; + tensor new_kv_unnorm_5 = add(x = var_710, y = var_712)[name = tensor("new_kv_unnorm_5")]; + tensor var_714 = const()[name = tensor("op_714"), val = tensor(0x1.4p+2)]; + tensor new_scale_5 = add(x = prev_scale_5, y = var_714)[name = tensor("new_scale_5")]; + tensor var_716 = sqrt(x = new_scale_5)[name = tensor("op_716")]; + tensor var_717 = real_div(x = new_kv_unnorm_5, y = var_716)[name = tensor("op_717")]; + tensor var_718_perm_0 = const()[name = tensor("op_718_perm_0"), val = tensor([0, 2, 1, 3])]; tensor out_15_axes_0 = const()[name = tensor("out_15_axes_0"), val = tensor([-1])]; - tensor var_652 = transpose(perm = var_652_perm_0, x = out_13)[name = tensor("transpose_41")]; - tensor out_15 = layer_norm(axes = out_15_axes_0, epsilon = var_18, x = var_652)[name = tensor("out_15")]; - tensor var_656 = const()[name = tensor("op_656"), val = tensor([1, 5, 256])]; - tensor out_17 = reshape(shape = var_656, x = out_15)[name = tensor("out_17")]; - tensor var_658 = silu(x = input_97)[name = tensor("op_658")]; - tensor input_99 = mul(x = var_658, y = out_17)[name = tensor("input_99")]; - tensor ret_out_5 = linear(bias = encoder_out_proj_2_bias, weight = encoder_out_proj_2_weight, x = input_99)[name = tensor("linear_25")]; - tensor x_15 = add(x = input_93, y = ret_out_5)[name = tensor("x_15")]; + tensor var_718 = transpose(perm = var_718_perm_0, x = out_13)[name = tensor("transpose_41")]; + tensor out_15 = layer_norm(axes = out_15_axes_0, epsilon = var_84, x = var_718)[name = tensor("out_15")]; + tensor var_722 = const()[name = tensor("op_722"), val = tensor([1, 5, 256])]; + tensor out_17 = reshape(shape = var_722, x = out_15)[name = tensor("out_17")]; + tensor var_724 = silu(x = input_99)[name = tensor("op_724")]; + tensor input_101 = mul(x = var_724, y = out_17)[name = tensor("input_101")]; + tensor ret_out_5 = linear(bias = inner_encoder_out_proj_2_bias, weight = inner_encoder_out_proj_2_weight, x = input_101)[name = tensor("linear_25")]; + tensor x_15 = add(x = input_95, y = ret_out_5)[name = tensor("x_15")]; tensor window_25_begin_0 = const()[name = tensor("window_25_begin_0"), val = tensor([2, 0, 0, 0])]; tensor window_25_end_0 = const()[name = tensor("window_25_end_0"), val = tensor([3, 1, 16, 256])]; tensor window_25_end_mask_0 = const()[name = tensor("window_25_end_mask_0"), val = tensor([false, true, true, true])]; tensor window_25_squeeze_mask_0 = const()[name = tensor("window_25_squeeze_mask_0"), val = tensor([true, false, false, false])]; tensor window_25 = slice_by_index(begin = window_25_begin_0, end = window_25_end_0, end_mask = window_25_end_mask_0, squeeze_mask = window_25_squeeze_mask_0, x = enc_conv_cache)[name = tensor("window_25")]; - tensor var_666_begin_0 = const()[name = tensor("op_666_begin_0"), val = tensor([0, 0, 0])]; - tensor var_666_end_0 = const()[name = tensor("op_666_end_0"), val = tensor([1, 1, 256])]; - tensor var_666_end_mask_0 = const()[name = tensor("op_666_end_mask_0"), val = tensor([true, false, true])]; - tensor var_666 = slice_by_index(begin = var_666_begin_0, end = var_666_end_0, end_mask = var_666_end_mask_0, x = x_15)[name = tensor("op_666")]; - tensor var_669_begin_0 = const()[name = tensor("op_669_begin_0"), val = tensor([0, 1, 0])]; - tensor var_669_end_0 = const()[name = tensor("op_669_end_0"), val = tensor([1, 16, 256])]; - tensor var_669_end_mask_0 = const()[name = tensor("op_669_end_mask_0"), val = tensor([true, true, true])]; - tensor var_669 = slice_by_index(begin = var_669_begin_0, end = var_669_end_0, end_mask = var_669_end_mask_0, x = window_25)[name = tensor("op_669")]; + tensor var_732_begin_0 = const()[name = tensor("op_732_begin_0"), val = tensor([0, 0, 0])]; + tensor var_732_end_0 = const()[name = tensor("op_732_end_0"), val = tensor([1, 1, 256])]; + tensor var_732_end_mask_0 = const()[name = tensor("op_732_end_mask_0"), val = tensor([true, false, true])]; + tensor var_732 = slice_by_index(begin = var_732_begin_0, end = var_732_end_0, end_mask = var_732_end_mask_0, x = x_15)[name = tensor("op_732")]; + tensor var_735_begin_0 = const()[name = tensor("op_735_begin_0"), val = tensor([0, 1, 0])]; + tensor var_735_end_0 = const()[name = tensor("op_735_end_0"), val = tensor([1, 16, 256])]; + tensor var_735_end_mask_0 = const()[name = tensor("op_735_end_mask_0"), val = tensor([true, true, true])]; + tensor var_735 = slice_by_index(begin = var_735_begin_0, end = var_735_end_0, end_mask = var_735_end_mask_0, x = window_25)[name = tensor("op_735")]; tensor window_27_interleave_0 = const()[name = tensor("window_27_interleave_0"), val = tensor(false)]; - tensor window_27 = concat(axis = var_27, interleave = window_27_interleave_0, values = (var_669, var_666))[name = tensor("window_27")]; - tensor var_674_begin_0 = const()[name = tensor("op_674_begin_0"), val = tensor([0, 1, 0])]; - tensor var_674_end_0 = const()[name = tensor("op_674_end_0"), val = tensor([1, 2, 256])]; - tensor var_674_end_mask_0 = const()[name = tensor("op_674_end_mask_0"), val = tensor([true, false, true])]; - tensor var_674 = slice_by_index(begin = var_674_begin_0, end = var_674_end_0, end_mask = var_674_end_mask_0, x = x_15)[name = tensor("op_674")]; - tensor var_677_begin_0 = const()[name = tensor("op_677_begin_0"), val = tensor([0, 1, 0])]; - tensor var_677_end_0 = const()[name = tensor("op_677_end_0"), val = tensor([1, 16, 256])]; - tensor var_677_end_mask_0 = const()[name = tensor("op_677_end_mask_0"), val = tensor([true, true, true])]; - tensor var_677 = slice_by_index(begin = var_677_begin_0, end = var_677_end_0, end_mask = var_677_end_mask_0, x = window_27)[name = tensor("op_677")]; + tensor window_27 = concat(axis = var_93, interleave = window_27_interleave_0, values = (var_735, var_732))[name = tensor("window_27")]; + tensor var_740_begin_0 = const()[name = tensor("op_740_begin_0"), val = tensor([0, 1, 0])]; + tensor var_740_end_0 = const()[name = tensor("op_740_end_0"), val = tensor([1, 2, 256])]; + tensor var_740_end_mask_0 = const()[name = tensor("op_740_end_mask_0"), val = tensor([true, false, true])]; + tensor var_740 = slice_by_index(begin = var_740_begin_0, end = var_740_end_0, end_mask = var_740_end_mask_0, x = x_15)[name = tensor("op_740")]; + tensor var_743_begin_0 = const()[name = tensor("op_743_begin_0"), val = tensor([0, 1, 0])]; + tensor var_743_end_0 = const()[name = tensor("op_743_end_0"), val = tensor([1, 16, 256])]; + tensor var_743_end_mask_0 = const()[name = tensor("op_743_end_mask_0"), val = tensor([true, true, true])]; + tensor var_743 = slice_by_index(begin = var_743_begin_0, end = var_743_end_0, end_mask = var_743_end_mask_0, x = window_27)[name = tensor("op_743")]; tensor window_29_interleave_0 = const()[name = tensor("window_29_interleave_0"), val = tensor(false)]; - tensor window_29 = concat(axis = var_27, interleave = window_29_interleave_0, values = (var_677, var_674))[name = tensor("window_29")]; - tensor var_682_begin_0 = const()[name = tensor("op_682_begin_0"), val = tensor([0, 2, 0])]; - tensor var_682_end_0 = const()[name = tensor("op_682_end_0"), val = tensor([1, 3, 256])]; - tensor var_682_end_mask_0 = const()[name = tensor("op_682_end_mask_0"), val = tensor([true, false, true])]; - tensor var_682 = slice_by_index(begin = var_682_begin_0, end = var_682_end_0, end_mask = var_682_end_mask_0, x = x_15)[name = tensor("op_682")]; - tensor var_685_begin_0 = const()[name = tensor("op_685_begin_0"), val = tensor([0, 1, 0])]; - tensor var_685_end_0 = const()[name = tensor("op_685_end_0"), val = tensor([1, 16, 256])]; - tensor var_685_end_mask_0 = const()[name = tensor("op_685_end_mask_0"), val = tensor([true, true, true])]; - tensor var_685 = slice_by_index(begin = var_685_begin_0, end = var_685_end_0, end_mask = var_685_end_mask_0, x = window_29)[name = tensor("op_685")]; + tensor window_29 = concat(axis = var_93, interleave = window_29_interleave_0, values = (var_743, var_740))[name = tensor("window_29")]; + tensor var_748_begin_0 = const()[name = tensor("op_748_begin_0"), val = tensor([0, 2, 0])]; + tensor var_748_end_0 = const()[name = tensor("op_748_end_0"), val = tensor([1, 3, 256])]; + tensor var_748_end_mask_0 = const()[name = tensor("op_748_end_mask_0"), val = tensor([true, false, true])]; + tensor var_748 = slice_by_index(begin = var_748_begin_0, end = var_748_end_0, end_mask = var_748_end_mask_0, x = x_15)[name = tensor("op_748")]; + tensor var_751_begin_0 = const()[name = tensor("op_751_begin_0"), val = tensor([0, 1, 0])]; + tensor var_751_end_0 = const()[name = tensor("op_751_end_0"), val = tensor([1, 16, 256])]; + tensor var_751_end_mask_0 = const()[name = tensor("op_751_end_mask_0"), val = tensor([true, true, true])]; + tensor var_751 = slice_by_index(begin = var_751_begin_0, end = var_751_end_0, end_mask = var_751_end_mask_0, x = window_29)[name = tensor("op_751")]; tensor window_31_interleave_0 = const()[name = tensor("window_31_interleave_0"), val = tensor(false)]; - tensor window_31 = concat(axis = var_27, interleave = window_31_interleave_0, values = (var_685, var_682))[name = tensor("window_31")]; - tensor var_690_begin_0 = const()[name = tensor("op_690_begin_0"), val = tensor([0, 3, 0])]; - tensor var_690_end_0 = const()[name = tensor("op_690_end_0"), val = tensor([1, 4, 256])]; - tensor var_690_end_mask_0 = const()[name = tensor("op_690_end_mask_0"), val = tensor([true, false, true])]; - tensor var_690 = slice_by_index(begin = var_690_begin_0, end = var_690_end_0, end_mask = var_690_end_mask_0, x = x_15)[name = tensor("op_690")]; - tensor var_693_begin_0 = const()[name = tensor("op_693_begin_0"), val = tensor([0, 1, 0])]; - tensor var_693_end_0 = const()[name = tensor("op_693_end_0"), val = tensor([1, 16, 256])]; - tensor var_693_end_mask_0 = const()[name = tensor("op_693_end_mask_0"), val = tensor([true, true, true])]; - tensor var_693 = slice_by_index(begin = var_693_begin_0, end = var_693_end_0, end_mask = var_693_end_mask_0, x = window_31)[name = tensor("op_693")]; + tensor window_31 = concat(axis = var_93, interleave = window_31_interleave_0, values = (var_751, var_748))[name = tensor("window_31")]; + tensor var_756_begin_0 = const()[name = tensor("op_756_begin_0"), val = tensor([0, 3, 0])]; + tensor var_756_end_0 = const()[name = tensor("op_756_end_0"), val = tensor([1, 4, 256])]; + tensor var_756_end_mask_0 = const()[name = tensor("op_756_end_mask_0"), val = tensor([true, false, true])]; + tensor var_756 = slice_by_index(begin = var_756_begin_0, end = var_756_end_0, end_mask = var_756_end_mask_0, x = x_15)[name = tensor("op_756")]; + tensor var_759_begin_0 = const()[name = tensor("op_759_begin_0"), val = tensor([0, 1, 0])]; + tensor var_759_end_0 = const()[name = tensor("op_759_end_0"), val = tensor([1, 16, 256])]; + tensor var_759_end_mask_0 = const()[name = tensor("op_759_end_mask_0"), val = tensor([true, true, true])]; + tensor var_759 = slice_by_index(begin = var_759_begin_0, end = var_759_end_0, end_mask = var_759_end_mask_0, x = window_31)[name = tensor("op_759")]; tensor window_33_interleave_0 = const()[name = tensor("window_33_interleave_0"), val = tensor(false)]; - tensor window_33 = concat(axis = var_27, interleave = window_33_interleave_0, values = (var_693, var_690))[name = tensor("window_33")]; - tensor var_698_begin_0 = const()[name = tensor("op_698_begin_0"), val = tensor([0, 4, 0])]; - tensor var_698_end_0 = const()[name = tensor("op_698_end_0"), val = tensor([1, 1, 256])]; - tensor var_698_end_mask_0 = const()[name = tensor("op_698_end_mask_0"), val = tensor([true, true, true])]; - tensor var_698 = slice_by_index(begin = var_698_begin_0, end = var_698_end_0, end_mask = var_698_end_mask_0, x = x_15)[name = tensor("op_698")]; - tensor var_701_begin_0 = const()[name = tensor("op_701_begin_0"), val = tensor([0, 1, 0])]; - tensor var_701_end_0 = const()[name = tensor("op_701_end_0"), val = tensor([1, 16, 256])]; - tensor var_701_end_mask_0 = const()[name = tensor("op_701_end_mask_0"), val = tensor([true, true, true])]; - tensor var_701 = slice_by_index(begin = var_701_begin_0, end = var_701_end_0, end_mask = var_701_end_mask_0, x = window_33)[name = tensor("op_701")]; + tensor window_33 = concat(axis = var_93, interleave = window_33_interleave_0, values = (var_759, var_756))[name = tensor("window_33")]; + tensor var_764_begin_0 = const()[name = tensor("op_764_begin_0"), val = tensor([0, 4, 0])]; + tensor var_764_end_0 = const()[name = tensor("op_764_end_0"), val = tensor([1, 1, 256])]; + tensor var_764_end_mask_0 = const()[name = tensor("op_764_end_mask_0"), val = tensor([true, true, true])]; + tensor var_764 = slice_by_index(begin = var_764_begin_0, end = var_764_end_0, end_mask = var_764_end_mask_0, x = x_15)[name = tensor("op_764")]; + tensor var_767_begin_0 = const()[name = tensor("op_767_begin_0"), val = tensor([0, 1, 0])]; + tensor var_767_end_0 = const()[name = tensor("op_767_end_0"), val = tensor([1, 16, 256])]; + tensor var_767_end_mask_0 = const()[name = tensor("op_767_end_mask_0"), val = tensor([true, true, true])]; + tensor var_767 = slice_by_index(begin = var_767_begin_0, end = var_767_end_0, end_mask = var_767_end_mask_0, x = window_33)[name = tensor("op_767")]; tensor window_35_interleave_0 = const()[name = tensor("window_35_interleave_0"), val = tensor(false)]; - tensor window_35 = concat(axis = var_27, interleave = window_35_interleave_0, values = (var_701, var_698))[name = tensor("window_35")]; - tensor input_101_interleave_0 = const()[name = tensor("input_101_interleave_0"), val = tensor(false)]; - tensor input_101 = concat(axis = var_24, interleave = input_101_interleave_0, values = (window_27, window_29, window_31, window_33, window_35))[name = tensor("input_101")]; + tensor window_35 = concat(axis = var_93, interleave = window_35_interleave_0, values = (var_767, var_764))[name = tensor("window_35")]; + tensor input_103_interleave_0 = const()[name = tensor("input_103_interleave_0"), val = tensor(false)]; + tensor input_103 = concat(axis = var_79, interleave = input_103_interleave_0, values = (window_27, window_29, window_31, window_33, window_35))[name = tensor("input_103")]; tensor x_17_axes_0 = const()[name = tensor("x_17_axes_0"), val = tensor([-1])]; - tensor x_17 = layer_norm(axes = x_17_axes_0, beta = encoder_conv_module_2_sequential_0_bias, epsilon = var_29, gamma = encoder_conv_module_2_sequential_0_weight, x = input_101)[name = tensor("x_17")]; - tensor input_103_perm_0 = const()[name = tensor("input_103_perm_0"), val = tensor([0, 2, 1])]; + tensor x_17 = layer_norm(axes = x_17_axes_0, beta = inner_encoder_conv_module_2_sequential_0_bias, epsilon = var_76, gamma = inner_encoder_conv_module_2_sequential_0_weight, x = input_103)[name = tensor("x_17")]; + tensor input_105_perm_0 = const()[name = tensor("input_105_perm_0"), val = tensor([0, 2, 1])]; tensor inputs_23_pad_type_0 = const()[name = tensor("inputs_23_pad_type_0"), val = tensor("valid")]; tensor inputs_23_strides_0 = const()[name = tensor("inputs_23_strides_0"), val = tensor([1])]; tensor inputs_23_pad_0 = const()[name = tensor("inputs_23_pad_0"), val = tensor([0, 0])]; tensor inputs_23_dilations_0 = const()[name = tensor("inputs_23_dilations_0"), val = tensor([1])]; tensor inputs_23_groups_0 = const()[name = tensor("inputs_23_groups_0"), val = tensor(1)]; - tensor input_103 = transpose(perm = input_103_perm_0, x = x_17)[name = tensor("transpose_40")]; - tensor inputs_23 = conv(bias = encoder_conv_module_2_sequential_2_conv_bias, dilations = inputs_23_dilations_0, groups = inputs_23_groups_0, pad = inputs_23_pad_0, pad_type = inputs_23_pad_type_0, strides = inputs_23_strides_0, weight = encoder_conv_module_2_sequential_2_conv_weight, x = input_103)[name = tensor("inputs_23")]; - tensor var_726_split_sizes_0 = const()[name = tensor("op_726_split_sizes_0"), val = tensor([256, 256])]; - tensor var_726_axis_0 = const()[name = tensor("op_726_axis_0"), val = tensor(1)]; - tensor var_726_0, tensor var_726_1 = split(axis = var_726_axis_0, split_sizes = var_726_split_sizes_0, x = inputs_23)[name = tensor("op_726")]; - tensor var_728 = sigmoid(x = var_726_1)[name = tensor("op_728")]; - tensor inputs_25 = mul(x = var_726_0, y = var_728)[name = tensor("inputs_25")]; + tensor input_105 = transpose(perm = input_105_perm_0, x = x_17)[name = tensor("transpose_40")]; + tensor inputs_23 = conv(bias = inner_encoder_conv_module_2_sequential_2_conv_bias, dilations = inputs_23_dilations_0, groups = inputs_23_groups_0, pad = inputs_23_pad_0, pad_type = inputs_23_pad_type_0, strides = inputs_23_strides_0, weight = inner_encoder_conv_module_2_sequential_2_conv_weight, x = input_105)[name = tensor("inputs_23")]; + tensor var_792_split_sizes_0 = const()[name = tensor("op_792_split_sizes_0"), val = tensor([256, 256])]; + tensor var_792_axis_0 = const()[name = tensor("op_792_axis_0"), val = tensor(1)]; + tensor var_792_0, tensor var_792_1 = split(axis = var_792_axis_0, split_sizes = var_792_split_sizes_0, x = inputs_23)[name = tensor("op_792")]; + tensor var_794 = sigmoid(x = var_792_1)[name = tensor("op_794")]; + tensor inputs_25 = mul(x = var_792_0, y = var_794)[name = tensor("inputs_25")]; tensor outputs_aug_5_pad_type_0 = const()[name = tensor("outputs_aug_5_pad_type_0"), val = tensor("custom")]; tensor outputs_aug_5_pad_0 = const()[name = tensor("outputs_aug_5_pad_0"), val = tensor([15, 15])]; tensor outputs_aug_5_groups_0 = const()[name = tensor("outputs_aug_5_groups_0"), val = tensor(256)]; tensor outputs_aug_5_strides_0 = const()[name = tensor("outputs_aug_5_strides_0"), val = tensor([1])]; tensor outputs_aug_5_dilations_0 = const()[name = tensor("outputs_aug_5_dilations_0"), val = tensor([1])]; - tensor outputs_aug_5 = conv(dilations = outputs_aug_5_dilations_0, groups = outputs_aug_5_groups_0, pad = outputs_aug_5_pad_0, pad_type = outputs_aug_5_pad_type_0, strides = outputs_aug_5_strides_0, weight = encoder_conv_module_2_sequential_4_conv_weight, x = inputs_25)[name = tensor("outputs_aug_5")]; - tensor input_105_begin_0 = const()[name = tensor("input_105_begin_0"), val = tensor([0, 0, 0])]; - tensor input_105_end_0 = const()[name = tensor("input_105_end_0"), val = tensor([5, 256, 16])]; - tensor input_105_end_mask_0 = const()[name = tensor("input_105_end_mask_0"), val = tensor([true, true, false])]; - tensor input_105 = slice_by_index(begin = input_105_begin_0, end = input_105_end_0, end_mask = input_105_end_mask_0, x = outputs_aug_5)[name = tensor("input_105")]; - tensor inputs_27 = batch_norm(beta = encoder_conv_module_2_sequential_5_bias, epsilon = var_29, gamma = encoder_conv_module_2_sequential_5_weight, mean = encoder_conv_module_2_sequential_5_running_mean, variance = encoder_conv_module_2_sequential_5_running_var, x = input_105)[name = tensor("inputs_27")]; - tensor input_107 = silu(x = inputs_27)[name = tensor("input_107")]; - tensor input_109_pad_type_0 = const()[name = tensor("input_109_pad_type_0"), val = tensor("valid")]; - tensor input_109_strides_0 = const()[name = tensor("input_109_strides_0"), val = tensor([1])]; - tensor input_109_pad_0 = const()[name = tensor("input_109_pad_0"), val = tensor([0, 0])]; - tensor input_109_dilations_0 = const()[name = tensor("input_109_dilations_0"), val = tensor([1])]; - tensor input_109_groups_0 = const()[name = tensor("input_109_groups_0"), val = tensor(1)]; - tensor input_109 = conv(bias = encoder_conv_module_2_sequential_7_conv_bias, dilations = input_109_dilations_0, groups = input_109_groups_0, pad = input_109_pad_0, pad_type = input_109_pad_type_0, strides = input_109_strides_0, weight = encoder_conv_module_2_sequential_7_conv_weight, x = input_107)[name = tensor("input_109")]; + tensor outputs_aug_5 = conv(dilations = outputs_aug_5_dilations_0, groups = outputs_aug_5_groups_0, pad = outputs_aug_5_pad_0, pad_type = outputs_aug_5_pad_type_0, strides = outputs_aug_5_strides_0, weight = inner_encoder_conv_module_2_sequential_4_conv_weight, x = inputs_25)[name = tensor("outputs_aug_5")]; + tensor input_107_begin_0 = const()[name = tensor("input_107_begin_0"), val = tensor([0, 0, 0])]; + tensor input_107_end_0 = const()[name = tensor("input_107_end_0"), val = tensor([5, 256, 16])]; + tensor input_107_end_mask_0 = const()[name = tensor("input_107_end_mask_0"), val = tensor([true, true, false])]; + tensor input_107 = slice_by_index(begin = input_107_begin_0, end = input_107_end_0, end_mask = input_107_end_mask_0, x = outputs_aug_5)[name = tensor("input_107")]; + tensor inputs_27 = batch_norm(beta = inner_encoder_conv_module_2_sequential_5_bias, epsilon = var_76, gamma = inner_encoder_conv_module_2_sequential_5_weight, mean = inner_encoder_conv_module_2_sequential_5_running_mean, variance = inner_encoder_conv_module_2_sequential_5_running_var, x = input_107)[name = tensor("inputs_27")]; + tensor input_109 = silu(x = inputs_27)[name = tensor("input_109")]; + tensor input_111_pad_type_0 = const()[name = tensor("input_111_pad_type_0"), val = tensor("valid")]; + tensor input_111_strides_0 = const()[name = tensor("input_111_strides_0"), val = tensor([1])]; + tensor input_111_pad_0 = const()[name = tensor("input_111_pad_0"), val = tensor([0, 0])]; + tensor input_111_dilations_0 = const()[name = tensor("input_111_dilations_0"), val = tensor([1])]; + tensor input_111_groups_0 = const()[name = tensor("input_111_groups_0"), val = tensor(1)]; + tensor input_111 = conv(bias = inner_encoder_conv_module_2_sequential_7_conv_bias, dilations = input_111_dilations_0, groups = input_111_groups_0, pad = input_111_pad_0, pad_type = input_111_pad_type_0, strides = input_111_strides_0, weight = inner_encoder_conv_module_2_sequential_7_conv_weight, x = input_109)[name = tensor("input_111")]; tensor conv_out_5_perm_0 = const()[name = tensor("conv_out_5_perm_0"), val = tensor([0, 2, 1])]; - tensor var_759_begin_0 = const()[name = tensor("op_759_begin_0"), val = tensor([0, -1, 0])]; - tensor var_759_end_0 = const()[name = tensor("op_759_end_0"), val = tensor([5, 16, 256])]; - tensor var_759_end_mask_0 = const()[name = tensor("op_759_end_mask_0"), val = tensor([true, true, true])]; - tensor conv_out_5 = transpose(perm = conv_out_5_perm_0, x = input_109)[name = tensor("transpose_39")]; - tensor var_759 = slice_by_index(begin = var_759_begin_0, end = var_759_end_0, end_mask = var_759_end_mask_0, x = conv_out_5)[name = tensor("op_759")]; - tensor var_761_perm_0 = const()[name = tensor("op_761_perm_0"), val = tensor([1, 0, 2])]; - tensor var_761 = transpose(perm = var_761_perm_0, x = var_759)[name = tensor("transpose_38")]; - tensor input_111 = add(x = x_15, y = var_761)[name = tensor("input_111")]; - tensor input_113_axes_0 = const()[name = tensor("input_113_axes_0"), val = tensor([-1])]; - tensor input_113 = layer_norm(axes = input_113_axes_0, beta = encoder_ffn2_2_module_sequential_0_bias, epsilon = var_29, gamma = encoder_ffn2_2_module_sequential_0_weight, x = input_111)[name = tensor("input_113")]; - tensor inputs_29 = linear(bias = encoder_ffn2_2_module_sequential_1_linear_bias, weight = encoder_ffn2_2_module_sequential_1_linear_weight, x = input_113)[name = tensor("linear_26")]; - tensor input_115 = silu(x = inputs_29)[name = tensor("input_115")]; - tensor input_119 = linear(bias = encoder_ffn2_2_module_sequential_4_linear_bias, weight = encoder_ffn2_2_module_sequential_4_linear_weight, x = input_115)[name = tensor("linear_27")]; - tensor var_784 = const()[name = tensor("op_784"), val = tensor(0x1p-1)]; - tensor var_785 = mul(x = input_119, y = var_784)[name = tensor("op_785")]; - tensor input_121 = add(x = var_785, y = input_111)[name = tensor("input_121")]; - tensor input_123_axes_0 = const()[name = tensor("input_123_axes_0"), val = tensor([-1])]; - tensor input_123 = layer_norm(axes = input_123_axes_0, beta = encoder_layer_norm_2_bias, epsilon = var_29, gamma = encoder_layer_norm_2_weight, x = input_121)[name = tensor("input_123")]; + tensor var_825_begin_0 = const()[name = tensor("op_825_begin_0"), val = tensor([0, -1, 0])]; + tensor var_825_end_0 = const()[name = tensor("op_825_end_0"), val = tensor([5, 16, 256])]; + tensor var_825_end_mask_0 = const()[name = tensor("op_825_end_mask_0"), val = tensor([true, true, true])]; + tensor conv_out_5 = transpose(perm = conv_out_5_perm_0, x = input_111)[name = tensor("transpose_39")]; + tensor var_825 = slice_by_index(begin = var_825_begin_0, end = var_825_end_0, end_mask = var_825_end_mask_0, x = conv_out_5)[name = tensor("op_825")]; + tensor var_827_perm_0 = const()[name = tensor("op_827_perm_0"), val = tensor([1, 0, 2])]; + tensor var_827 = transpose(perm = var_827_perm_0, x = var_825)[name = tensor("transpose_38")]; + tensor input_113 = add(x = x_15, y = var_827)[name = tensor("input_113")]; + tensor input_115_axes_0 = const()[name = tensor("input_115_axes_0"), val = tensor([-1])]; + tensor input_115 = layer_norm(axes = input_115_axes_0, beta = inner_encoder_ffn2_2_module_sequential_0_bias, epsilon = var_76, gamma = inner_encoder_ffn2_2_module_sequential_0_weight, x = input_113)[name = tensor("input_115")]; + tensor inputs_29 = linear(bias = inner_encoder_ffn2_2_module_sequential_1_linear_bias, weight = inner_encoder_ffn2_2_module_sequential_1_linear_weight, x = input_115)[name = tensor("linear_26")]; + tensor input_117 = silu(x = inputs_29)[name = tensor("input_117")]; + tensor input_121 = linear(bias = inner_encoder_ffn2_2_module_sequential_4_linear_bias, weight = inner_encoder_ffn2_2_module_sequential_4_linear_weight, x = input_117)[name = tensor("linear_27")]; + tensor var_850 = const()[name = tensor("op_850"), val = tensor(0x1p-1)]; + tensor var_851 = mul(x = input_121, y = var_850)[name = tensor("op_851")]; + tensor input_123 = add(x = var_851, y = input_113)[name = tensor("input_123")]; tensor input_125_axes_0 = const()[name = tensor("input_125_axes_0"), val = tensor([-1])]; - tensor input_125 = layer_norm(axes = input_125_axes_0, beta = encoder_ffn1_3_module_sequential_0_bias, epsilon = var_29, gamma = encoder_ffn1_3_module_sequential_0_weight, x = input_123)[name = tensor("input_125")]; - tensor inputs_31 = linear(bias = encoder_ffn1_3_module_sequential_1_linear_bias, weight = encoder_ffn1_3_module_sequential_1_linear_weight, x = input_125)[name = tensor("linear_28")]; - tensor input_127 = silu(x = inputs_31)[name = tensor("input_127")]; - tensor input_131 = linear(bias = encoder_ffn1_3_module_sequential_4_linear_bias, weight = encoder_ffn1_3_module_sequential_4_linear_weight, x = input_127)[name = tensor("linear_29")]; - tensor var_814 = const()[name = tensor("op_814"), val = tensor(0x1p-1)]; - tensor var_815 = mul(x = input_131, y = var_814)[name = tensor("op_815")]; - tensor input_133 = add(x = var_815, y = input_123)[name = tensor("input_133")]; + tensor input_125 = layer_norm(axes = input_125_axes_0, beta = inner_encoder_layer_norm_2_bias, epsilon = var_76, gamma = inner_encoder_layer_norm_2_weight, x = input_123)[name = tensor("input_125")]; + tensor input_127_axes_0 = const()[name = tensor("input_127_axes_0"), val = tensor([-1])]; + tensor input_127 = layer_norm(axes = input_127_axes_0, beta = inner_encoder_ffn1_3_module_sequential_0_bias, epsilon = var_76, gamma = inner_encoder_ffn1_3_module_sequential_0_weight, x = input_125)[name = tensor("input_127")]; + tensor inputs_31 = linear(bias = inner_encoder_ffn1_3_module_sequential_1_linear_bias, weight = inner_encoder_ffn1_3_module_sequential_1_linear_weight, x = input_127)[name = tensor("linear_28")]; + tensor input_129 = silu(x = inputs_31)[name = tensor("input_129")]; + tensor input_133 = linear(bias = inner_encoder_ffn1_3_module_sequential_4_linear_bias, weight = inner_encoder_ffn1_3_module_sequential_4_linear_weight, x = input_129)[name = tensor("linear_29")]; + tensor var_880 = const()[name = tensor("op_880"), val = tensor(0x1p-1)]; + tensor var_881 = mul(x = input_133, y = var_880)[name = tensor("op_881")]; + tensor input_135 = add(x = var_881, y = input_125)[name = tensor("input_135")]; tensor x_19_axes_0 = const()[name = tensor("x_19_axes_0"), val = tensor([-1])]; - tensor x_19 = layer_norm(axes = x_19_axes_0, beta = encoder_ret_lns_3_bias, epsilon = var_29, gamma = encoder_ret_lns_3_weight, x = input_133)[name = tensor("x_19")]; + tensor x_19 = layer_norm(axes = x_19_axes_0, beta = inner_encoder_ret_lns_3_bias, epsilon = var_76, gamma = inner_encoder_ret_lns_3_weight, x = input_135)[name = tensor("x_19")]; tensor prev_kv_7_begin_0 = const()[name = tensor("prev_kv_7_begin_0"), val = tensor([3, 0, 0, 0, 0])]; tensor prev_kv_7_end_0 = const()[name = tensor("prev_kv_7_end_0"), val = tensor([4, 1, 4, 64, 64])]; tensor prev_kv_7_end_mask_0 = const()[name = tensor("prev_kv_7_end_mask_0"), val = tensor([false, true, true, true, true])]; @@ -800,219 +826,212 @@ program(1.0) tensor prev_scale_7_end_mask_0 = const()[name = tensor("prev_scale_7_end_mask_0"), val = tensor([false, true])]; tensor prev_scale_7_squeeze_mask_0 = const()[name = tensor("prev_scale_7_squeeze_mask_0"), val = tensor([true, false])]; tensor prev_scale_7 = slice_by_index(begin = prev_scale_7_begin_0, end = prev_scale_7_end_0, end_mask = prev_scale_7_end_mask_0, squeeze_mask = prev_scale_7_squeeze_mask_0, x = enc_scale)[name = tensor("prev_scale_7")]; - tensor var_829 = linear(bias = encoder_q_proj_3_bias, weight = encoder_q_proj_3_weight, x = x_19)[name = tensor("linear_30")]; - tensor var_830 = const()[name = tensor("op_830"), val = tensor([1, 5, 4, 64])]; - tensor var_831 = reshape(shape = var_830, x = var_829)[name = tensor("op_831")]; + tensor var_895 = linear(bias = inner_encoder_q_proj_3_bias, weight = inner_encoder_q_proj_3_weight, x = x_19)[name = tensor("linear_30")]; + tensor var_896 = const()[name = tensor("op_896"), val = tensor([1, 5, 4, 64])]; + tensor var_897 = reshape(shape = var_896, x = var_895)[name = tensor("op_897")]; tensor q_7_perm_0 = const()[name = tensor("q_7_perm_0"), val = tensor([0, 2, 1, 3])]; - tensor var_835 = linear(bias = encoder_k_proj_3_bias, weight = encoder_k_proj_3_weight, x = x_19)[name = tensor("linear_31")]; - tensor var_836 = const()[name = tensor("op_836"), val = tensor(0x1p-3)]; - tensor var_837 = mul(x = var_835, y = var_836)[name = tensor("op_837")]; - tensor var_838 = const()[name = tensor("op_838"), val = tensor([1, 5, 4, 64])]; - tensor var_839 = reshape(shape = var_838, x = var_837)[name = tensor("op_839")]; + tensor var_901 = linear(bias = inner_encoder_k_proj_3_bias, weight = inner_encoder_k_proj_3_weight, x = x_19)[name = tensor("linear_31")]; + tensor var_902 = const()[name = tensor("op_902"), val = tensor(0x1p-3)]; + tensor var_903 = mul(x = var_901, y = var_902)[name = tensor("op_903")]; + tensor var_904 = const()[name = tensor("op_904"), val = tensor([1, 5, 4, 64])]; + tensor var_905 = reshape(shape = var_904, x = var_903)[name = tensor("op_905")]; tensor k_7_perm_0 = const()[name = tensor("k_7_perm_0"), val = tensor([0, 2, 1, 3])]; - tensor var_843 = linear(bias = encoder_v_proj_3_bias, weight = encoder_v_proj_3_weight, x = x_19)[name = tensor("linear_32")]; - tensor var_844 = const()[name = tensor("op_844"), val = tensor([1, 5, 4, 64])]; - tensor var_845 = reshape(shape = var_844, x = var_843)[name = tensor("op_845")]; + tensor var_909 = linear(bias = inner_encoder_v_proj_3_bias, weight = inner_encoder_v_proj_3_weight, x = x_19)[name = tensor("linear_32")]; + tensor var_910 = const()[name = tensor("op_910"), val = tensor([1, 5, 4, 64])]; + tensor var_911 = reshape(shape = var_910, x = var_909)[name = tensor("op_911")]; tensor v_7_perm_0 = const()[name = tensor("v_7_perm_0"), val = tensor([0, 2, 1, 3])]; - tensor input_137 = linear(bias = encoder_g_proj_3_bias, weight = encoder_g_proj_3_weight, x = x_19)[name = tensor("linear_33")]; + tensor input_139 = linear(bias = inner_encoder_g_proj_3_bias, weight = inner_encoder_g_proj_3_weight, x = x_19)[name = tensor("linear_33")]; tensor sqrt_s0_7 = sqrt(x = prev_scale_7)[name = tensor("sqrt_s0_7")]; - tensor s_t_7 = add(x = prev_scale_7, y = encoder__t_index)[name = tensor("s_t_7")]; + tensor s_t_7 = add(x = prev_scale_7, y = inner_encoder__t_index)[name = tensor("s_t_7")]; tensor sqrt_s_t_7 = sqrt(x = s_t_7)[name = tensor("sqrt_s_t_7")]; tensor qk_7_transpose_x_1 = const()[name = tensor("qk_7_transpose_x_1"), val = tensor(false)]; tensor qk_7_transpose_y_1 = const()[name = tensor("qk_7_transpose_y_1"), val = tensor(true)]; - tensor k_7 = transpose(perm = k_7_perm_0, x = var_839)[name = tensor("transpose_36")]; - tensor q_7 = transpose(perm = q_7_perm_0, x = var_831)[name = tensor("transpose_37")]; + tensor k_7 = transpose(perm = k_7_perm_0, x = var_905)[name = tensor("transpose_36")]; + tensor q_7 = transpose(perm = q_7_perm_0, x = var_897)[name = tensor("transpose_37")]; tensor qk_7 = matmul(transpose_x = qk_7_transpose_x_1, transpose_y = qk_7_transpose_y_1, x = q_7, y = k_7)[name = tensor("qk_7")]; - tensor var_855 = const()[name = tensor("op_855"), val = tensor([5, 1])]; - tensor var_856 = reshape(shape = var_855, x = sqrt_s_t_7)[name = tensor("op_856")]; - tensor M_7 = real_div(x = encoder__causal_mask, y = var_856)[name = tensor("M_7")]; - tensor var_858 = mul(x = qk_7, y = M_7)[name = tensor("op_858")]; + tensor var_921 = const()[name = tensor("op_921"), val = tensor([5, 1])]; + tensor var_922 = reshape(shape = var_921, x = sqrt_s_t_7)[name = tensor("op_922")]; + tensor M_7 = real_div(x = inner_encoder__causal_mask, y = var_922)[name = tensor("M_7")]; + tensor var_924 = mul(x = qk_7, y = M_7)[name = tensor("op_924")]; tensor inner_7_transpose_x_0 = const()[name = tensor("inner_7_transpose_x_0"), val = tensor(false)]; tensor inner_7_transpose_y_0 = const()[name = tensor("inner_7_transpose_y_0"), val = tensor(false)]; - tensor v_7 = transpose(perm = v_7_perm_0, x = var_845)[name = tensor("transpose_35")]; - tensor inner_7 = matmul(transpose_x = inner_7_transpose_x_0, transpose_y = inner_7_transpose_y_0, x = var_858, y = v_7)[name = tensor("inner_7")]; - tensor var_860_transpose_x_0 = const()[name = tensor("op_860_transpose_x_0"), val = tensor(false)]; - tensor var_860_transpose_y_0 = const()[name = tensor("op_860_transpose_y_0"), val = tensor(false)]; - tensor var_860 = matmul(transpose_x = var_860_transpose_x_0, transpose_y = var_860_transpose_y_0, x = q_7, y = prev_kv_7)[name = tensor("op_860")]; - tensor var_861 = real_div(x = sqrt_s0_7, y = sqrt_s_t_7)[name = tensor("op_861")]; - tensor var_862 = const()[name = tensor("op_862"), val = tensor([1, 1, 5, 1])]; - tensor var_863 = reshape(shape = var_862, x = var_861)[name = tensor("op_863")]; - tensor cross_7 = mul(x = var_860, y = var_863)[name = tensor("cross_7")]; + tensor v_7 = transpose(perm = v_7_perm_0, x = var_911)[name = tensor("transpose_35")]; + tensor inner_7 = matmul(transpose_x = inner_7_transpose_x_0, transpose_y = inner_7_transpose_y_0, x = var_924, y = v_7)[name = tensor("inner_7")]; + tensor var_926_transpose_x_0 = const()[name = tensor("op_926_transpose_x_0"), val = tensor(false)]; + tensor var_926_transpose_y_0 = const()[name = tensor("op_926_transpose_y_0"), val = tensor(false)]; + tensor var_926 = matmul(transpose_x = var_926_transpose_x_0, transpose_y = var_926_transpose_y_0, x = q_7, y = prev_kv_7)[name = tensor("op_926")]; + tensor var_927 = real_div(x = sqrt_s0_7, y = sqrt_s_t_7)[name = tensor("op_927")]; + tensor var_928 = const()[name = tensor("op_928"), val = tensor([1, 1, 5, 1])]; + tensor var_929 = reshape(shape = var_928, x = var_927)[name = tensor("op_929")]; + tensor cross_7 = mul(x = var_926, y = var_929)[name = tensor("cross_7")]; tensor out_19 = add(x = inner_7, y = cross_7)[name = tensor("out_19")]; - tensor var_866 = mul(x = prev_kv_7, y = sqrt_s0_7)[name = tensor("op_866")]; - tensor var_868_transpose_x_1 = const()[name = tensor("op_868_transpose_x_1"), val = tensor(true)]; - tensor var_868_transpose_y_1 = const()[name = tensor("op_868_transpose_y_1"), val = tensor(false)]; - tensor var_868 = matmul(transpose_x = var_868_transpose_x_1, transpose_y = var_868_transpose_y_1, x = k_7, y = v_7)[name = tensor("op_868")]; - tensor new_kv_unnorm_7 = add(x = var_866, y = var_868)[name = tensor("new_kv_unnorm_7")]; - tensor var_870 = const()[name = tensor("op_870"), val = tensor(0x1.4p+2)]; - tensor new_scale_7 = add(x = prev_scale_7, y = var_870)[name = tensor("new_scale_7")]; - tensor var_872 = sqrt(x = new_scale_7)[name = tensor("op_872")]; - tensor nkv_1 = real_div(x = new_kv_unnorm_7, y = var_872)[name = tensor("nkv_1")]; - tensor var_874_perm_0 = const()[name = tensor("op_874_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_932 = mul(x = prev_kv_7, y = sqrt_s0_7)[name = tensor("op_932")]; + tensor var_934_transpose_x_1 = const()[name = tensor("op_934_transpose_x_1"), val = tensor(true)]; + tensor var_934_transpose_y_1 = const()[name = tensor("op_934_transpose_y_1"), val = tensor(false)]; + tensor var_934 = matmul(transpose_x = var_934_transpose_x_1, transpose_y = var_934_transpose_y_1, x = k_7, y = v_7)[name = tensor("op_934")]; + tensor new_kv_unnorm_7 = add(x = var_932, y = var_934)[name = tensor("new_kv_unnorm_7")]; + tensor var_936 = const()[name = tensor("op_936"), val = tensor(0x1.4p+2)]; + tensor new_scale_7 = add(x = prev_scale_7, y = var_936)[name = tensor("new_scale_7")]; + tensor var_938 = sqrt(x = new_scale_7)[name = tensor("op_938")]; + tensor nkv_1 = real_div(x = new_kv_unnorm_7, y = var_938)[name = tensor("nkv_1")]; + tensor var_940_perm_0 = const()[name = tensor("op_940_perm_0"), val = tensor([0, 2, 1, 3])]; tensor out_21_axes_0 = const()[name = tensor("out_21_axes_0"), val = tensor([-1])]; - tensor var_874 = transpose(perm = var_874_perm_0, x = out_19)[name = tensor("transpose_34")]; - tensor out_21 = layer_norm(axes = out_21_axes_0, epsilon = var_18, x = var_874)[name = tensor("out_21")]; - tensor var_878 = const()[name = tensor("op_878"), val = tensor([1, 5, 256])]; - tensor out_23 = reshape(shape = var_878, x = out_21)[name = tensor("out_23")]; - tensor var_880 = silu(x = input_137)[name = tensor("op_880")]; - tensor input_139 = mul(x = var_880, y = out_23)[name = tensor("input_139")]; - tensor ret_out_7 = linear(bias = encoder_out_proj_3_bias, weight = encoder_out_proj_3_weight, x = input_139)[name = tensor("linear_34")]; - tensor x_21 = add(x = input_133, y = ret_out_7)[name = tensor("x_21")]; + tensor var_940 = transpose(perm = var_940_perm_0, x = out_19)[name = tensor("transpose_34")]; + tensor out_21 = layer_norm(axes = out_21_axes_0, epsilon = var_84, x = var_940)[name = tensor("out_21")]; + tensor var_944 = const()[name = tensor("op_944"), val = tensor([1, 5, 256])]; + tensor out_23 = reshape(shape = var_944, x = out_21)[name = tensor("out_23")]; + tensor var_946 = silu(x = input_139)[name = tensor("op_946")]; + tensor input_141 = mul(x = var_946, y = out_23)[name = tensor("input_141")]; + tensor ret_out_7 = linear(bias = inner_encoder_out_proj_3_bias, weight = inner_encoder_out_proj_3_weight, x = input_141)[name = tensor("linear_34")]; + tensor x_21 = add(x = input_135, y = ret_out_7)[name = tensor("x_21")]; tensor window_37_begin_0 = const()[name = tensor("window_37_begin_0"), val = tensor([3, 0, 0, 0])]; tensor window_37_end_0 = const()[name = tensor("window_37_end_0"), val = tensor([4, 1, 16, 256])]; tensor window_37_end_mask_0 = const()[name = tensor("window_37_end_mask_0"), val = tensor([false, true, true, true])]; tensor window_37_squeeze_mask_0 = const()[name = tensor("window_37_squeeze_mask_0"), val = tensor([true, false, false, false])]; tensor window_37 = slice_by_index(begin = window_37_begin_0, end = window_37_end_0, end_mask = window_37_end_mask_0, squeeze_mask = window_37_squeeze_mask_0, x = enc_conv_cache)[name = tensor("window_37")]; - tensor var_888_begin_0 = const()[name = tensor("op_888_begin_0"), val = tensor([0, 0, 0])]; - tensor var_888_end_0 = const()[name = tensor("op_888_end_0"), val = tensor([1, 1, 256])]; - tensor var_888_end_mask_0 = const()[name = tensor("op_888_end_mask_0"), val = tensor([true, false, true])]; - tensor var_888 = slice_by_index(begin = var_888_begin_0, end = var_888_end_0, end_mask = var_888_end_mask_0, x = x_21)[name = tensor("op_888")]; - tensor var_891_begin_0 = const()[name = tensor("op_891_begin_0"), val = tensor([0, 1, 0])]; - tensor var_891_end_0 = const()[name = tensor("op_891_end_0"), val = tensor([1, 16, 256])]; - tensor var_891_end_mask_0 = const()[name = tensor("op_891_end_mask_0"), val = tensor([true, true, true])]; - tensor var_891 = slice_by_index(begin = var_891_begin_0, end = var_891_end_0, end_mask = var_891_end_mask_0, x = window_37)[name = tensor("op_891")]; + tensor var_954_begin_0 = const()[name = tensor("op_954_begin_0"), val = tensor([0, 0, 0])]; + tensor var_954_end_0 = const()[name = tensor("op_954_end_0"), val = tensor([1, 1, 256])]; + tensor var_954_end_mask_0 = const()[name = tensor("op_954_end_mask_0"), val = tensor([true, false, true])]; + tensor var_954 = slice_by_index(begin = var_954_begin_0, end = var_954_end_0, end_mask = var_954_end_mask_0, x = x_21)[name = tensor("op_954")]; + tensor var_957_begin_0 = const()[name = tensor("op_957_begin_0"), val = tensor([0, 1, 0])]; + tensor var_957_end_0 = const()[name = tensor("op_957_end_0"), val = tensor([1, 16, 256])]; + tensor var_957_end_mask_0 = const()[name = tensor("op_957_end_mask_0"), val = tensor([true, true, true])]; + tensor var_957 = slice_by_index(begin = var_957_begin_0, end = var_957_end_0, end_mask = var_957_end_mask_0, x = window_37)[name = tensor("op_957")]; tensor window_39_interleave_0 = const()[name = tensor("window_39_interleave_0"), val = tensor(false)]; - tensor window_39 = concat(axis = var_27, interleave = window_39_interleave_0, values = (var_891, var_888))[name = tensor("window_39")]; - tensor var_896_begin_0 = const()[name = tensor("op_896_begin_0"), val = tensor([0, 1, 0])]; - tensor var_896_end_0 = const()[name = tensor("op_896_end_0"), val = tensor([1, 2, 256])]; - tensor var_896_end_mask_0 = const()[name = tensor("op_896_end_mask_0"), val = tensor([true, false, true])]; - tensor var_896 = slice_by_index(begin = var_896_begin_0, end = var_896_end_0, end_mask = var_896_end_mask_0, x = x_21)[name = tensor("op_896")]; - tensor var_899_begin_0 = const()[name = tensor("op_899_begin_0"), val = tensor([0, 1, 0])]; - tensor var_899_end_0 = const()[name = tensor("op_899_end_0"), val = tensor([1, 16, 256])]; - tensor var_899_end_mask_0 = const()[name = tensor("op_899_end_mask_0"), val = tensor([true, true, true])]; - tensor var_899 = slice_by_index(begin = var_899_begin_0, end = var_899_end_0, end_mask = var_899_end_mask_0, x = window_39)[name = tensor("op_899")]; + tensor window_39 = concat(axis = var_93, interleave = window_39_interleave_0, values = (var_957, var_954))[name = tensor("window_39")]; + tensor var_962_begin_0 = const()[name = tensor("op_962_begin_0"), val = tensor([0, 1, 0])]; + tensor var_962_end_0 = const()[name = tensor("op_962_end_0"), val = tensor([1, 2, 256])]; + tensor var_962_end_mask_0 = const()[name = tensor("op_962_end_mask_0"), val = tensor([true, false, true])]; + tensor var_962 = slice_by_index(begin = var_962_begin_0, end = var_962_end_0, end_mask = var_962_end_mask_0, x = x_21)[name = tensor("op_962")]; + tensor var_965_begin_0 = const()[name = tensor("op_965_begin_0"), val = tensor([0, 1, 0])]; + tensor var_965_end_0 = const()[name = tensor("op_965_end_0"), val = tensor([1, 16, 256])]; + tensor var_965_end_mask_0 = const()[name = tensor("op_965_end_mask_0"), val = tensor([true, true, true])]; + tensor var_965 = slice_by_index(begin = var_965_begin_0, end = var_965_end_0, end_mask = var_965_end_mask_0, x = window_39)[name = tensor("op_965")]; tensor window_41_interleave_0 = const()[name = tensor("window_41_interleave_0"), val = tensor(false)]; - tensor window_41 = concat(axis = var_27, interleave = window_41_interleave_0, values = (var_899, var_896))[name = tensor("window_41")]; - tensor var_904_begin_0 = const()[name = tensor("op_904_begin_0"), val = tensor([0, 2, 0])]; - tensor var_904_end_0 = const()[name = tensor("op_904_end_0"), val = tensor([1, 3, 256])]; - tensor var_904_end_mask_0 = const()[name = tensor("op_904_end_mask_0"), val = tensor([true, false, true])]; - tensor var_904 = slice_by_index(begin = var_904_begin_0, end = var_904_end_0, end_mask = var_904_end_mask_0, x = x_21)[name = tensor("op_904")]; - tensor var_907_begin_0 = const()[name = tensor("op_907_begin_0"), val = tensor([0, 1, 0])]; - tensor var_907_end_0 = const()[name = tensor("op_907_end_0"), val = tensor([1, 16, 256])]; - tensor var_907_end_mask_0 = const()[name = tensor("op_907_end_mask_0"), val = tensor([true, true, true])]; - tensor var_907 = slice_by_index(begin = var_907_begin_0, end = var_907_end_0, end_mask = var_907_end_mask_0, x = window_41)[name = tensor("op_907")]; + tensor window_41 = concat(axis = var_93, interleave = window_41_interleave_0, values = (var_965, var_962))[name = tensor("window_41")]; + tensor var_970_begin_0 = const()[name = tensor("op_970_begin_0"), val = tensor([0, 2, 0])]; + tensor var_970_end_0 = const()[name = tensor("op_970_end_0"), val = tensor([1, 3, 256])]; + tensor var_970_end_mask_0 = const()[name = tensor("op_970_end_mask_0"), val = tensor([true, false, true])]; + tensor var_970 = slice_by_index(begin = var_970_begin_0, end = var_970_end_0, end_mask = var_970_end_mask_0, x = x_21)[name = tensor("op_970")]; + tensor var_973_begin_0 = const()[name = tensor("op_973_begin_0"), val = tensor([0, 1, 0])]; + tensor var_973_end_0 = const()[name = tensor("op_973_end_0"), val = tensor([1, 16, 256])]; + tensor var_973_end_mask_0 = const()[name = tensor("op_973_end_mask_0"), val = tensor([true, true, true])]; + tensor var_973 = slice_by_index(begin = var_973_begin_0, end = var_973_end_0, end_mask = var_973_end_mask_0, x = window_41)[name = tensor("op_973")]; tensor window_43_interleave_0 = const()[name = tensor("window_43_interleave_0"), val = tensor(false)]; - tensor window_43 = concat(axis = var_27, interleave = window_43_interleave_0, values = (var_907, var_904))[name = tensor("window_43")]; - tensor var_912_begin_0 = const()[name = tensor("op_912_begin_0"), val = tensor([0, 3, 0])]; - tensor var_912_end_0 = const()[name = tensor("op_912_end_0"), val = tensor([1, 4, 256])]; - tensor var_912_end_mask_0 = const()[name = tensor("op_912_end_mask_0"), val = tensor([true, false, true])]; - tensor var_912 = slice_by_index(begin = var_912_begin_0, end = var_912_end_0, end_mask = var_912_end_mask_0, x = x_21)[name = tensor("op_912")]; - tensor var_915_begin_0 = const()[name = tensor("op_915_begin_0"), val = tensor([0, 1, 0])]; - tensor var_915_end_0 = const()[name = tensor("op_915_end_0"), val = tensor([1, 16, 256])]; - tensor var_915_end_mask_0 = const()[name = tensor("op_915_end_mask_0"), val = tensor([true, true, true])]; - tensor var_915 = slice_by_index(begin = var_915_begin_0, end = var_915_end_0, end_mask = var_915_end_mask_0, x = window_43)[name = tensor("op_915")]; + tensor window_43 = concat(axis = var_93, interleave = window_43_interleave_0, values = (var_973, var_970))[name = tensor("window_43")]; + tensor var_978_begin_0 = const()[name = tensor("op_978_begin_0"), val = tensor([0, 3, 0])]; + tensor var_978_end_0 = const()[name = tensor("op_978_end_0"), val = tensor([1, 4, 256])]; + tensor var_978_end_mask_0 = const()[name = tensor("op_978_end_mask_0"), val = tensor([true, false, true])]; + tensor var_978 = slice_by_index(begin = var_978_begin_0, end = var_978_end_0, end_mask = var_978_end_mask_0, x = x_21)[name = tensor("op_978")]; + tensor var_981_begin_0 = const()[name = tensor("op_981_begin_0"), val = tensor([0, 1, 0])]; + tensor var_981_end_0 = const()[name = tensor("op_981_end_0"), val = tensor([1, 16, 256])]; + tensor var_981_end_mask_0 = const()[name = tensor("op_981_end_mask_0"), val = tensor([true, true, true])]; + tensor var_981 = slice_by_index(begin = var_981_begin_0, end = var_981_end_0, end_mask = var_981_end_mask_0, x = window_43)[name = tensor("op_981")]; tensor window_45_interleave_0 = const()[name = tensor("window_45_interleave_0"), val = tensor(false)]; - tensor window_45 = concat(axis = var_27, interleave = window_45_interleave_0, values = (var_915, var_912))[name = tensor("window_45")]; - tensor var_920_begin_0 = const()[name = tensor("op_920_begin_0"), val = tensor([0, 4, 0])]; - tensor var_920_end_0 = const()[name = tensor("op_920_end_0"), val = tensor([1, 1, 256])]; - tensor var_920_end_mask_0 = const()[name = tensor("op_920_end_mask_0"), val = tensor([true, true, true])]; - tensor var_920 = slice_by_index(begin = var_920_begin_0, end = var_920_end_0, end_mask = var_920_end_mask_0, x = x_21)[name = tensor("op_920")]; - tensor var_923_begin_0 = const()[name = tensor("op_923_begin_0"), val = tensor([0, 1, 0])]; - tensor var_923_end_0 = const()[name = tensor("op_923_end_0"), val = tensor([1, 16, 256])]; - tensor var_923_end_mask_0 = const()[name = tensor("op_923_end_mask_0"), val = tensor([true, true, true])]; - tensor var_923 = slice_by_index(begin = var_923_begin_0, end = var_923_end_0, end_mask = var_923_end_mask_0, x = window_45)[name = tensor("op_923")]; + tensor window_45 = concat(axis = var_93, interleave = window_45_interleave_0, values = (var_981, var_978))[name = tensor("window_45")]; + tensor var_986_begin_0 = const()[name = tensor("op_986_begin_0"), val = tensor([0, 4, 0])]; + tensor var_986_end_0 = const()[name = tensor("op_986_end_0"), val = tensor([1, 1, 256])]; + tensor var_986_end_mask_0 = const()[name = tensor("op_986_end_mask_0"), val = tensor([true, true, true])]; + tensor var_986 = slice_by_index(begin = var_986_begin_0, end = var_986_end_0, end_mask = var_986_end_mask_0, x = x_21)[name = tensor("op_986")]; + tensor var_989_begin_0 = const()[name = tensor("op_989_begin_0"), val = tensor([0, 1, 0])]; + tensor var_989_end_0 = const()[name = tensor("op_989_end_0"), val = tensor([1, 16, 256])]; + tensor var_989_end_mask_0 = const()[name = tensor("op_989_end_mask_0"), val = tensor([true, true, true])]; + tensor var_989 = slice_by_index(begin = var_989_begin_0, end = var_989_end_0, end_mask = var_989_end_mask_0, x = window_45)[name = tensor("op_989")]; tensor window_interleave_0 = const()[name = tensor("window_interleave_0"), val = tensor(false)]; - tensor window = concat(axis = var_27, interleave = window_interleave_0, values = (var_923, var_920))[name = tensor("window")]; - tensor input_141_interleave_0 = const()[name = tensor("input_141_interleave_0"), val = tensor(false)]; - tensor input_141 = concat(axis = var_24, interleave = input_141_interleave_0, values = (window_39, window_41, window_43, window_45, window))[name = tensor("input_141")]; + tensor window = concat(axis = var_93, interleave = window_interleave_0, values = (var_989, var_986))[name = tensor("window")]; + tensor input_143_interleave_0 = const()[name = tensor("input_143_interleave_0"), val = tensor(false)]; + tensor input_143 = concat(axis = var_79, interleave = input_143_interleave_0, values = (window_39, window_41, window_43, window_45, window))[name = tensor("input_143")]; tensor x_23_axes_0 = const()[name = tensor("x_23_axes_0"), val = tensor([-1])]; - tensor x_23 = layer_norm(axes = x_23_axes_0, beta = encoder_conv_module_3_sequential_0_bias, epsilon = var_29, gamma = encoder_conv_module_3_sequential_0_weight, x = input_141)[name = tensor("x_23")]; - tensor input_143_perm_0 = const()[name = tensor("input_143_perm_0"), val = tensor([0, 2, 1])]; + tensor x_23 = layer_norm(axes = x_23_axes_0, beta = inner_encoder_conv_module_3_sequential_0_bias, epsilon = var_76, gamma = inner_encoder_conv_module_3_sequential_0_weight, x = input_143)[name = tensor("x_23")]; + tensor input_145_perm_0 = const()[name = tensor("input_145_perm_0"), val = tensor([0, 2, 1])]; tensor inputs_33_pad_type_0 = const()[name = tensor("inputs_33_pad_type_0"), val = tensor("valid")]; tensor inputs_33_strides_0 = const()[name = tensor("inputs_33_strides_0"), val = tensor([1])]; tensor inputs_33_pad_0 = const()[name = tensor("inputs_33_pad_0"), val = tensor([0, 0])]; tensor inputs_33_dilations_0 = const()[name = tensor("inputs_33_dilations_0"), val = tensor([1])]; tensor inputs_33_groups_0 = const()[name = tensor("inputs_33_groups_0"), val = tensor(1)]; - tensor input_143 = transpose(perm = input_143_perm_0, x = x_23)[name = tensor("transpose_33")]; - tensor inputs_33 = conv(bias = encoder_conv_module_3_sequential_2_conv_bias, dilations = inputs_33_dilations_0, groups = inputs_33_groups_0, pad = inputs_33_pad_0, pad_type = inputs_33_pad_type_0, strides = inputs_33_strides_0, weight = encoder_conv_module_3_sequential_2_conv_weight, x = input_143)[name = tensor("inputs_33")]; - tensor var_948_split_sizes_0 = const()[name = tensor("op_948_split_sizes_0"), val = tensor([256, 256])]; - tensor var_948_axis_0 = const()[name = tensor("op_948_axis_0"), val = tensor(1)]; - tensor var_948_0, tensor var_948_1 = split(axis = var_948_axis_0, split_sizes = var_948_split_sizes_0, x = inputs_33)[name = tensor("op_948")]; - tensor var_950 = sigmoid(x = var_948_1)[name = tensor("op_950")]; - tensor inputs_35 = mul(x = var_948_0, y = var_950)[name = tensor("inputs_35")]; + tensor input_145 = transpose(perm = input_145_perm_0, x = x_23)[name = tensor("transpose_33")]; + tensor inputs_33 = conv(bias = inner_encoder_conv_module_3_sequential_2_conv_bias, dilations = inputs_33_dilations_0, groups = inputs_33_groups_0, pad = inputs_33_pad_0, pad_type = inputs_33_pad_type_0, strides = inputs_33_strides_0, weight = inner_encoder_conv_module_3_sequential_2_conv_weight, x = input_145)[name = tensor("inputs_33")]; + tensor var_1014_split_sizes_0 = const()[name = tensor("op_1014_split_sizes_0"), val = tensor([256, 256])]; + tensor var_1014_axis_0 = const()[name = tensor("op_1014_axis_0"), val = tensor(1)]; + tensor var_1014_0, tensor var_1014_1 = split(axis = var_1014_axis_0, split_sizes = var_1014_split_sizes_0, x = inputs_33)[name = tensor("op_1014")]; + tensor var_1016 = sigmoid(x = var_1014_1)[name = tensor("op_1016")]; + tensor inputs_35 = mul(x = var_1014_0, y = var_1016)[name = tensor("inputs_35")]; tensor outputs_aug_pad_type_0 = const()[name = tensor("outputs_aug_pad_type_0"), val = tensor("custom")]; tensor outputs_aug_pad_0 = const()[name = tensor("outputs_aug_pad_0"), val = tensor([15, 15])]; tensor outputs_aug_groups_0 = const()[name = tensor("outputs_aug_groups_0"), val = tensor(256)]; tensor outputs_aug_strides_0 = const()[name = tensor("outputs_aug_strides_0"), val = tensor([1])]; tensor outputs_aug_dilations_0 = const()[name = tensor("outputs_aug_dilations_0"), val = tensor([1])]; - tensor outputs_aug = conv(dilations = outputs_aug_dilations_0, groups = outputs_aug_groups_0, pad = outputs_aug_pad_0, pad_type = outputs_aug_pad_type_0, strides = outputs_aug_strides_0, weight = encoder_conv_module_3_sequential_4_conv_weight, x = inputs_35)[name = tensor("outputs_aug")]; - tensor input_145_begin_0 = const()[name = tensor("input_145_begin_0"), val = tensor([0, 0, 0])]; - tensor input_145_end_0 = const()[name = tensor("input_145_end_0"), val = tensor([5, 256, 16])]; - tensor input_145_end_mask_0 = const()[name = tensor("input_145_end_mask_0"), val = tensor([true, true, false])]; - tensor input_145 = slice_by_index(begin = input_145_begin_0, end = input_145_end_0, end_mask = input_145_end_mask_0, x = outputs_aug)[name = tensor("input_145")]; - tensor inputs_37 = batch_norm(beta = encoder_conv_module_3_sequential_5_bias, epsilon = var_29, gamma = encoder_conv_module_3_sequential_5_weight, mean = encoder_conv_module_3_sequential_5_running_mean, variance = encoder_conv_module_3_sequential_5_running_var, x = input_145)[name = tensor("inputs_37")]; - tensor input_147 = silu(x = inputs_37)[name = tensor("input_147")]; - tensor input_149_pad_type_0 = const()[name = tensor("input_149_pad_type_0"), val = tensor("valid")]; - tensor input_149_strides_0 = const()[name = tensor("input_149_strides_0"), val = tensor([1])]; - tensor input_149_pad_0 = const()[name = tensor("input_149_pad_0"), val = tensor([0, 0])]; - tensor input_149_dilations_0 = const()[name = tensor("input_149_dilations_0"), val = tensor([1])]; - tensor input_149_groups_0 = const()[name = tensor("input_149_groups_0"), val = tensor(1)]; - tensor input_149 = conv(bias = encoder_conv_module_3_sequential_7_conv_bias, dilations = input_149_dilations_0, groups = input_149_groups_0, pad = input_149_pad_0, pad_type = input_149_pad_type_0, strides = input_149_strides_0, weight = encoder_conv_module_3_sequential_7_conv_weight, x = input_147)[name = tensor("input_149")]; + tensor outputs_aug = conv(dilations = outputs_aug_dilations_0, groups = outputs_aug_groups_0, pad = outputs_aug_pad_0, pad_type = outputs_aug_pad_type_0, strides = outputs_aug_strides_0, weight = inner_encoder_conv_module_3_sequential_4_conv_weight, x = inputs_35)[name = tensor("outputs_aug")]; + tensor input_147_begin_0 = const()[name = tensor("input_147_begin_0"), val = tensor([0, 0, 0])]; + tensor input_147_end_0 = const()[name = tensor("input_147_end_0"), val = tensor([5, 256, 16])]; + tensor input_147_end_mask_0 = const()[name = tensor("input_147_end_mask_0"), val = tensor([true, true, false])]; + tensor input_147 = slice_by_index(begin = input_147_begin_0, end = input_147_end_0, end_mask = input_147_end_mask_0, x = outputs_aug)[name = tensor("input_147")]; + tensor inputs_37 = batch_norm(beta = inner_encoder_conv_module_3_sequential_5_bias, epsilon = var_76, gamma = inner_encoder_conv_module_3_sequential_5_weight, mean = inner_encoder_conv_module_3_sequential_5_running_mean, variance = inner_encoder_conv_module_3_sequential_5_running_var, x = input_147)[name = tensor("inputs_37")]; + tensor input_149 = silu(x = inputs_37)[name = tensor("input_149")]; + tensor input_151_pad_type_0 = const()[name = tensor("input_151_pad_type_0"), val = tensor("valid")]; + tensor input_151_strides_0 = const()[name = tensor("input_151_strides_0"), val = tensor([1])]; + tensor input_151_pad_0 = const()[name = tensor("input_151_pad_0"), val = tensor([0, 0])]; + tensor input_151_dilations_0 = const()[name = tensor("input_151_dilations_0"), val = tensor([1])]; + tensor input_151_groups_0 = const()[name = tensor("input_151_groups_0"), val = tensor(1)]; + tensor input_151 = conv(bias = inner_encoder_conv_module_3_sequential_7_conv_bias, dilations = input_151_dilations_0, groups = input_151_groups_0, pad = input_151_pad_0, pad_type = input_151_pad_type_0, strides = input_151_strides_0, weight = inner_encoder_conv_module_3_sequential_7_conv_weight, x = input_149)[name = tensor("input_151")]; tensor conv_out_7_perm_0 = const()[name = tensor("conv_out_7_perm_0"), val = tensor([0, 2, 1])]; - tensor var_981_begin_0 = const()[name = tensor("op_981_begin_0"), val = tensor([0, -1, 0])]; - tensor var_981_end_0 = const()[name = tensor("op_981_end_0"), val = tensor([5, 16, 256])]; - tensor var_981_end_mask_0 = const()[name = tensor("op_981_end_mask_0"), val = tensor([true, true, true])]; - tensor conv_out_7 = transpose(perm = conv_out_7_perm_0, x = input_149)[name = tensor("transpose_32")]; - tensor var_981 = slice_by_index(begin = var_981_begin_0, end = var_981_end_0, end_mask = var_981_end_mask_0, x = conv_out_7)[name = tensor("op_981")]; - tensor var_983_perm_0 = const()[name = tensor("op_983_perm_0"), val = tensor([1, 0, 2])]; - tensor var_983 = transpose(perm = var_983_perm_0, x = var_981)[name = tensor("transpose_31")]; - tensor input_151 = add(x = x_21, y = var_983)[name = tensor("input_151")]; - tensor input_153_axes_0 = const()[name = tensor("input_153_axes_0"), val = tensor([-1])]; - tensor input_153 = layer_norm(axes = input_153_axes_0, beta = encoder_ffn2_3_module_sequential_0_bias, epsilon = var_29, gamma = encoder_ffn2_3_module_sequential_0_weight, x = input_151)[name = tensor("input_153")]; - tensor inputs = linear(bias = encoder_ffn2_3_module_sequential_1_linear_bias, weight = encoder_ffn2_3_module_sequential_1_linear_weight, x = input_153)[name = tensor("linear_35")]; - tensor input_155 = silu(x = inputs)[name = tensor("input_155")]; - tensor input_159 = linear(bias = encoder_ffn2_3_module_sequential_4_linear_bias, weight = encoder_ffn2_3_module_sequential_4_linear_weight, x = input_155)[name = tensor("linear_36")]; - tensor var_1006 = const()[name = tensor("op_1006"), val = tensor(0x1p-1)]; - tensor var_1007 = mul(x = input_159, y = var_1006)[name = tensor("op_1007")]; - tensor input_161 = add(x = var_1007, y = input_151)[name = tensor("input_161")]; + tensor var_1047_begin_0 = const()[name = tensor("op_1047_begin_0"), val = tensor([0, -1, 0])]; + tensor var_1047_end_0 = const()[name = tensor("op_1047_end_0"), val = tensor([5, 16, 256])]; + tensor var_1047_end_mask_0 = const()[name = tensor("op_1047_end_mask_0"), val = tensor([true, true, true])]; + tensor conv_out_7 = transpose(perm = conv_out_7_perm_0, x = input_151)[name = tensor("transpose_32")]; + tensor var_1047 = slice_by_index(begin = var_1047_begin_0, end = var_1047_end_0, end_mask = var_1047_end_mask_0, x = conv_out_7)[name = tensor("op_1047")]; + tensor var_1049_perm_0 = const()[name = tensor("op_1049_perm_0"), val = tensor([1, 0, 2])]; + tensor var_1049 = transpose(perm = var_1049_perm_0, x = var_1047)[name = tensor("transpose_31")]; + tensor input_153 = add(x = x_21, y = var_1049)[name = tensor("input_153")]; + tensor input_155_axes_0 = const()[name = tensor("input_155_axes_0"), val = tensor([-1])]; + tensor input_155 = layer_norm(axes = input_155_axes_0, beta = inner_encoder_ffn2_3_module_sequential_0_bias, epsilon = var_76, gamma = inner_encoder_ffn2_3_module_sequential_0_weight, x = input_153)[name = tensor("input_155")]; + tensor inputs = linear(bias = inner_encoder_ffn2_3_module_sequential_1_linear_bias, weight = inner_encoder_ffn2_3_module_sequential_1_linear_weight, x = input_155)[name = tensor("linear_35")]; + tensor input_157 = silu(x = inputs)[name = tensor("input_157")]; + tensor input_161 = linear(bias = inner_encoder_ffn2_3_module_sequential_4_linear_bias, weight = inner_encoder_ffn2_3_module_sequential_4_linear_weight, x = input_157)[name = tensor("linear_36")]; + tensor var_1072 = const()[name = tensor("op_1072"), val = tensor(0x1p-1)]; + tensor var_1073 = mul(x = input_161, y = var_1072)[name = tensor("op_1073")]; + tensor input_163 = add(x = var_1073, y = input_153)[name = tensor("input_163")]; tensor x_25_axes_0 = const()[name = tensor("x_25_axes_0"), val = tensor([-1])]; - tensor x_25 = layer_norm(axes = x_25_axes_0, beta = encoder_layer_norm_3_bias, epsilon = var_29, gamma = encoder_layer_norm_3_weight, x = input_161)[name = tensor("x_25")]; + tensor x_25 = layer_norm(axes = x_25_axes_0, beta = inner_encoder_layer_norm_3_bias, epsilon = var_76, gamma = inner_encoder_layer_norm_3_weight, x = input_163)[name = tensor("x_25")]; tensor x_ct_perm_0 = const()[name = tensor("x_ct_perm_0"), val = tensor([0, 2, 1])]; tensor cat_interleave_0 = const()[name = tensor("cat_interleave_0"), val = tensor(false)]; tensor x_ct = transpose(perm = x_ct_perm_0, x = x_25)[name = tensor("transpose_30")]; - tensor cat = concat(axis = var_21, interleave = cat_interleave_0, values = (cnn_window, x_ct))[name = tensor("cat")]; + tensor cat = concat(axis = var_81, interleave = cat_interleave_0, values = (cnn_window, x_ct))[name = tensor("cat")]; tensor conv_out_pad_type_0 = const()[name = tensor("conv_out_pad_type_0"), val = tensor("valid")]; tensor conv_out_strides_0 = const()[name = tensor("conv_out_strides_0"), val = tensor([1])]; tensor conv_out_pad_0 = const()[name = tensor("conv_out_pad_0"), val = tensor([0, 0])]; tensor conv_out_dilations_0 = const()[name = tensor("conv_out_dilations_0"), val = tensor([1])]; tensor conv_out_groups_0 = const()[name = tensor("conv_out_groups_0"), val = tensor(1)]; - tensor conv_out = conv(bias = encoder_cnn_bias, dilations = conv_out_dilations_0, groups = conv_out_groups_0, pad = conv_out_pad_0, pad_type = conv_out_pad_type_0, strides = conv_out_strides_0, weight = encoder_cnn_weight, x = cat)[name = tensor("conv_out")]; - tensor var_1025_begin_0 = const()[name = tensor("op_1025_begin_0"), val = tensor([0, 0, 5])]; - tensor var_1025_end_0 = const()[name = tensor("op_1025_end_0"), val = tensor([1, 256, 23])]; - tensor var_1025_end_mask_0 = const()[name = tensor("op_1025_end_mask_0"), val = tensor([true, true, true])]; - tensor cnn_window_new = slice_by_index(begin = var_1025_begin_0, end = var_1025_end_0, end_mask = var_1025_end_mask_0, x = cat)[name = tensor("op_1025")]; - tensor input_163_perm_0 = const()[name = tensor("input_163_perm_0"), val = tensor([0, 2, 1])]; - tensor var_1027 = const()[name = tensor("op_1027"), val = tensor([-1])]; - tensor input_163 = transpose(perm = input_163_perm_0, x = conv_out)[name = tensor("transpose_29")]; - tensor var_1028 = reduce_l2_norm(axes = var_1027, keep_dims = var_30, x = input_163)[name = tensor("op_1028")]; + tensor conv_out = conv(bias = inner_encoder_cnn_bias, dilations = conv_out_dilations_0, groups = conv_out_groups_0, pad = conv_out_pad_0, pad_type = conv_out_pad_type_0, strides = conv_out_strides_0, weight = inner_encoder_cnn_weight, x = cat)[name = tensor("conv_out")]; + tensor var_1091_begin_0 = const()[name = tensor("op_1091_begin_0"), val = tensor([0, 0, 5])]; + tensor var_1091_end_0 = const()[name = tensor("op_1091_end_0"), val = tensor([1, 256, 23])]; + tensor var_1091_end_mask_0 = const()[name = tensor("op_1091_end_mask_0"), val = tensor([true, true, true])]; + tensor cnn_window_new = slice_by_index(begin = var_1091_begin_0, end = var_1091_end_0, end_mask = var_1091_end_mask_0, x = cat)[name = tensor("op_1091")]; + tensor input_165_perm_0 = const()[name = tensor("input_165_perm_0"), val = tensor([0, 2, 1])]; + tensor var_1093 = const()[name = tensor("op_1093"), val = tensor([-1])]; + tensor input_165 = transpose(perm = input_165_perm_0, x = conv_out)[name = tensor("transpose_29")]; + tensor var_1094 = reduce_l2_norm(axes = var_1093, keep_dims = var_75, x = input_165)[name = tensor("op_1094")]; tensor const_12 = const()[name = tensor("const_12"), val = tensor(0x1.fffffep+127)]; - tensor clip_0 = clip(alpha = var_11, beta = const_12, x = var_1028)[name = tensor("clip_0")]; - tensor emb = real_div(x = input_163, y = clip_0)[name = tensor("emb")]; - tensor var_1032_axis_0 = const()[name = tensor("op_1032_axis_0"), val = tensor(0)]; - tensor enc_kv_new = stack(axis = var_1032_axis_0, values = (var_207, var_429, var_651, nkv_1))[name = tensor("op_1032")]; - tensor var_1034_axis_0 = const()[name = tensor("op_1034_axis_0"), val = tensor(0)]; - tensor enc_scale_new = stack(axis = var_1034_axis_0, values = (new_scale_1, new_scale_3, new_scale_5, new_scale_7))[name = tensor("op_1034")]; - tensor var_1036_axis_0 = const()[name = tensor("op_1036_axis_0"), val = tensor(0)]; - tensor enc_conv_cache_new = stack(axis = var_1036_axis_0, values = (window_11, window_23, window_35, window))[name = tensor("op_1036")]; - tensor var_1045 = const()[name = tensor("op_1045"), val = tensor(0x1.5798eep-27)]; - tensor var_1050 = const()[name = tensor("op_1050"), val = tensor(0x1.4f8b58p-17)]; - tensor var_1052 = const()[name = tensor("op_1052"), val = tensor(0x1.0c6f7ap-20)]; - tensor var_1053 = const()[name = tensor("op_1053"), val = tensor(true)]; - tensor var_1055 = const()[name = tensor("op_1055"), val = tensor(0x1p+0)]; - tensor var_1059 = const()[name = tensor("op_1059"), val = tensor(-1)]; - tensor var_1065 = const()[name = tensor("op_1065"), val = tensor(0)]; + tensor clip_0 = clip(alpha = var_90, beta = const_12, x = var_1094)[name = tensor("clip_0")]; + tensor emb = real_div(x = input_165, y = clip_0)[name = tensor("emb")]; + tensor var_1098_axis_0 = const()[name = tensor("op_1098_axis_0"), val = tensor(0)]; + tensor enc_kv_new = stack(axis = var_1098_axis_0, values = (var_273, var_495, var_717, nkv_1))[name = tensor("op_1098")]; + tensor var_1100_axis_0 = const()[name = tensor("op_1100_axis_0"), val = tensor(0)]; + tensor enc_scale_new = stack(axis = var_1100_axis_0, values = (new_scale_1, new_scale_3, new_scale_5, new_scale_7))[name = tensor("op_1100")]; + tensor var_1102_axis_0 = const()[name = tensor("op_1102_axis_0"), val = tensor(0)]; + tensor enc_conv_cache_new = stack(axis = var_1102_axis_0, values = (window_11, window_23, window_35, window))[name = tensor("op_1102")]; tensor pos = const()[name = tensor("pos"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(44395712)))]; - tensor var_1127_axes_0 = const()[name = tensor("op_1127_axes_0"), val = tensor([2])]; - tensor var_1127 = expand_dims(axes = var_1127_axes_0, x = emb)[name = tensor("op_1127")]; + tensor var_1170_axes_0 = const()[name = tensor("op_1170_axes_0"), val = tensor([2])]; + tensor var_1170 = expand_dims(axes = var_1170_axes_0, x = emb)[name = tensor("op_1170")]; tensor emb_exp_reps_0 = const()[name = tensor("emb_exp_reps_0"), val = tensor([1, 1, 9, 1])]; - tensor emb_exp = tile(reps = emb_exp_reps_0, x = var_1127)[name = tensor("emb_exp")]; - tensor input_165_interleave_0 = const()[name = tensor("input_165_interleave_0"), val = tensor(false)]; - tensor input_165 = concat(axis = var_1059, interleave = input_165_interleave_0, values = (emb_exp, pos))[name = tensor("input_165")]; - tensor x_27 = linear(bias = decoder_convert_bias, weight = decoder_convert_weight, x = input_165)[name = tensor("linear_37")]; - tensor var_1135_perm_0 = const()[name = tensor("op_1135_perm_0"), val = tensor([0, 2, 1, 3])]; - tensor var_1139 = const()[name = tensor("op_1139"), val = tensor([9, 5, 256])]; - tensor var_1135 = transpose(perm = var_1135_perm_0, x = x_27)[name = tensor("transpose_28")]; - tensor x_29 = reshape(shape = var_1139, x = var_1135)[name = tensor("x_29")]; + tensor emb_exp = tile(reps = emb_exp_reps_0, x = var_1170)[name = tensor("emb_exp")]; + tensor input_167_interleave_0 = const()[name = tensor("input_167_interleave_0"), val = tensor(false)]; + tensor input_167 = concat(axis = var_82, interleave = input_167_interleave_0, values = (emb_exp, pos))[name = tensor("input_167")]; + tensor x_27 = linear(bias = inner_decoder_convert_bias, weight = inner_decoder_convert_weight, x = input_167)[name = tensor("linear_37")]; + tensor var_1178_perm_0 = const()[name = tensor("op_1178_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_1182 = const()[name = tensor("op_1182"), val = tensor([9, 5, 256])]; + tensor var_1178 = transpose(perm = var_1178_perm_0, x = x_27)[name = tensor("transpose_28")]; + tensor x_29 = reshape(shape = var_1182, x = var_1178)[name = tensor("x_29")]; tensor prev_kv_9_begin_0 = const()[name = tensor("prev_kv_9_begin_0"), val = tensor([0, 0, 0, 0, 0])]; tensor prev_kv_9_end_0 = const()[name = tensor("prev_kv_9_end_0"), val = tensor([1, 9, 4, 64, 64])]; tensor prev_kv_9_end_mask_0 = const()[name = tensor("prev_kv_9_end_mask_0"), val = tensor([false, true, true, true, true])]; @@ -1023,132 +1042,132 @@ program(1.0) tensor prev_scale_9_end_mask_0 = const()[name = tensor("prev_scale_9_end_mask_0"), val = tensor([false, true])]; tensor prev_scale_9_squeeze_mask_0 = const()[name = tensor("prev_scale_9_squeeze_mask_0"), val = tensor([true, false])]; tensor prev_scale_9 = slice_by_index(begin = prev_scale_9_begin_0, end = prev_scale_9_end_0, end_mask = prev_scale_9_end_mask_0, squeeze_mask = prev_scale_9_squeeze_mask_0, x = dec_scale)[name = tensor("prev_scale_9")]; - tensor var_1147 = linear(bias = decoder_q_proj_0_bias, weight = decoder_q_proj_0_weight, x = x_29)[name = tensor("linear_38")]; - tensor var_1148 = const()[name = tensor("op_1148"), val = tensor([9, 5, 4, 64])]; - tensor var_1149 = reshape(shape = var_1148, x = var_1147)[name = tensor("op_1149")]; + tensor var_1190 = linear(bias = inner_decoder_q_proj_0_bias, weight = inner_decoder_q_proj_0_weight, x = x_29)[name = tensor("linear_38")]; + tensor var_1191 = const()[name = tensor("op_1191"), val = tensor([9, 5, 4, 64])]; + tensor var_1192 = reshape(shape = var_1191, x = var_1190)[name = tensor("op_1192")]; tensor q_9_perm_0 = const()[name = tensor("q_9_perm_0"), val = tensor([0, 2, 1, 3])]; - tensor var_1153 = linear(bias = decoder_k_proj_0_bias, weight = decoder_k_proj_0_weight, x = x_29)[name = tensor("linear_39")]; - tensor var_1154 = const()[name = tensor("op_1154"), val = tensor(0x1p-3)]; - tensor var_1155 = mul(x = var_1153, y = var_1154)[name = tensor("op_1155")]; - tensor var_1156 = const()[name = tensor("op_1156"), val = tensor([9, 5, 4, 64])]; - tensor var_1157 = reshape(shape = var_1156, x = var_1155)[name = tensor("op_1157")]; + tensor var_1196 = linear(bias = inner_decoder_k_proj_0_bias, weight = inner_decoder_k_proj_0_weight, x = x_29)[name = tensor("linear_39")]; + tensor var_1197 = const()[name = tensor("op_1197"), val = tensor(0x1p-3)]; + tensor var_1198 = mul(x = var_1196, y = var_1197)[name = tensor("op_1198")]; + tensor var_1199 = const()[name = tensor("op_1199"), val = tensor([9, 5, 4, 64])]; + tensor var_1200 = reshape(shape = var_1199, x = var_1198)[name = tensor("op_1200")]; tensor k_9_perm_0 = const()[name = tensor("k_9_perm_0"), val = tensor([0, 2, 1, 3])]; - tensor var_1161 = linear(bias = decoder_v_proj_0_bias, weight = decoder_v_proj_0_weight, x = x_29)[name = tensor("linear_40")]; - tensor var_1162 = const()[name = tensor("op_1162"), val = tensor([9, 5, 4, 64])]; - tensor var_1163 = reshape(shape = var_1162, x = var_1161)[name = tensor("op_1163")]; + tensor var_1204 = linear(bias = inner_decoder_v_proj_0_bias, weight = inner_decoder_v_proj_0_weight, x = x_29)[name = tensor("linear_40")]; + tensor var_1205 = const()[name = tensor("op_1205"), val = tensor([9, 5, 4, 64])]; + tensor var_1206 = reshape(shape = var_1205, x = var_1204)[name = tensor("op_1206")]; tensor v_9_perm_0 = const()[name = tensor("v_9_perm_0"), val = tensor([0, 2, 1, 3])]; - tensor input_169 = linear(bias = decoder_g_proj_0_bias, weight = decoder_g_proj_0_weight, x = x_29)[name = tensor("linear_41")]; + tensor input_171 = linear(bias = inner_decoder_g_proj_0_bias, weight = inner_decoder_g_proj_0_weight, x = x_29)[name = tensor("linear_41")]; tensor cumsum_mask_1_exclusive_0 = const()[name = tensor("cumsum_mask_1_exclusive_0"), val = tensor(false)]; tensor cumsum_mask_1_reverse_0 = const()[name = tensor("cumsum_mask_1_reverse_0"), val = tensor(false)]; - tensor cumsum_mask_1 = cumsum(axis = var_1065, exclusive = cumsum_mask_1_exclusive_0, reverse = cumsum_mask_1_reverse_0, x = valid_mask)[name = tensor("cumsum_mask_1")]; + tensor cumsum_mask_1 = cumsum(axis = var_79, exclusive = cumsum_mask_1_exclusive_0, reverse = cumsum_mask_1_reverse_0, x = valid_mask)[name = tensor("cumsum_mask_1")]; tensor sqrt_s0_9 = sqrt(x = prev_scale_9)[name = tensor("sqrt_s0_9")]; tensor s_t_9 = add(x = prev_scale_9, y = cumsum_mask_1)[name = tensor("s_t_9")]; tensor const_20 = const()[name = tensor("const_20"), val = tensor(0x1.fffffep+127)]; - tensor clip_1 = clip(alpha = var_1055, beta = const_20, x = s_t_9)[name = tensor("clip_1")]; + tensor clip_1 = clip(alpha = var_69, beta = const_20, x = s_t_9)[name = tensor("clip_1")]; tensor sqrt_s_t_9 = sqrt(x = clip_1)[name = tensor("sqrt_s_t_9")]; tensor qk_9_transpose_x_1 = const()[name = tensor("qk_9_transpose_x_1"), val = tensor(false)]; tensor qk_9_transpose_y_1 = const()[name = tensor("qk_9_transpose_y_1"), val = tensor(true)]; - tensor k_9 = transpose(perm = k_9_perm_0, x = var_1157)[name = tensor("transpose_26")]; - tensor q_9 = transpose(perm = q_9_perm_0, x = var_1149)[name = tensor("transpose_27")]; + tensor k_9 = transpose(perm = k_9_perm_0, x = var_1200)[name = tensor("transpose_26")]; + tensor q_9 = transpose(perm = q_9_perm_0, x = var_1192)[name = tensor("transpose_27")]; tensor qk_9 = matmul(transpose_x = qk_9_transpose_x_1, transpose_y = qk_9_transpose_y_1, x = q_9, y = k_9)[name = tensor("qk_9")]; - tensor var_1175 = const()[name = tensor("op_1175"), val = tensor([1, 5])]; - tensor var_1176 = reshape(shape = var_1175, x = valid_mask)[name = tensor("op_1176")]; - tensor causal_with_valid_1 = mul(x = decoder__causal_mask, y = var_1176)[name = tensor("causal_with_valid_1")]; - tensor var_1178 = const()[name = tensor("op_1178"), val = tensor([5, 1])]; - tensor var_1179 = reshape(shape = var_1178, x = sqrt_s_t_9)[name = tensor("op_1179")]; - tensor M_9 = real_div(x = causal_with_valid_1, y = var_1179)[name = tensor("M_9")]; - tensor var_1181 = mul(x = qk_9, y = M_9)[name = tensor("op_1181")]; + tensor var_1218 = const()[name = tensor("op_1218"), val = tensor([1, 5])]; + tensor var_1219 = reshape(shape = var_1218, x = valid_mask)[name = tensor("op_1219")]; + tensor causal_with_valid_1 = mul(x = inner_decoder__causal_mask, y = var_1219)[name = tensor("causal_with_valid_1")]; + tensor var_1221 = const()[name = tensor("op_1221"), val = tensor([5, 1])]; + tensor var_1222 = reshape(shape = var_1221, x = sqrt_s_t_9)[name = tensor("op_1222")]; + tensor M_9 = real_div(x = causal_with_valid_1, y = var_1222)[name = tensor("M_9")]; + tensor var_1224 = mul(x = qk_9, y = M_9)[name = tensor("op_1224")]; tensor inner_9_transpose_x_0 = const()[name = tensor("inner_9_transpose_x_0"), val = tensor(false)]; tensor inner_9_transpose_y_0 = const()[name = tensor("inner_9_transpose_y_0"), val = tensor(false)]; - tensor v_9 = transpose(perm = v_9_perm_0, x = var_1163)[name = tensor("transpose_25")]; - tensor inner_9 = matmul(transpose_x = inner_9_transpose_x_0, transpose_y = inner_9_transpose_y_0, x = var_1181, y = v_9)[name = tensor("inner_9")]; - tensor var_1183_transpose_x_0 = const()[name = tensor("op_1183_transpose_x_0"), val = tensor(false)]; - tensor var_1183_transpose_y_0 = const()[name = tensor("op_1183_transpose_y_0"), val = tensor(false)]; - tensor var_1183 = matmul(transpose_x = var_1183_transpose_x_0, transpose_y = var_1183_transpose_y_0, x = q_9, y = prev_kv_9)[name = tensor("op_1183")]; - tensor var_1184 = real_div(x = sqrt_s0_9, y = sqrt_s_t_9)[name = tensor("op_1184")]; - tensor var_1185 = const()[name = tensor("op_1185"), val = tensor([1, 1, 5, 1])]; - tensor var_1186 = reshape(shape = var_1185, x = var_1184)[name = tensor("op_1186")]; - tensor cross_9 = mul(x = var_1183, y = var_1186)[name = tensor("cross_9")]; + tensor v_9 = transpose(perm = v_9_perm_0, x = var_1206)[name = tensor("transpose_25")]; + tensor inner_9 = matmul(transpose_x = inner_9_transpose_x_0, transpose_y = inner_9_transpose_y_0, x = var_1224, y = v_9)[name = tensor("inner_9")]; + tensor var_1226_transpose_x_0 = const()[name = tensor("op_1226_transpose_x_0"), val = tensor(false)]; + tensor var_1226_transpose_y_0 = const()[name = tensor("op_1226_transpose_y_0"), val = tensor(false)]; + tensor var_1226 = matmul(transpose_x = var_1226_transpose_x_0, transpose_y = var_1226_transpose_y_0, x = q_9, y = prev_kv_9)[name = tensor("op_1226")]; + tensor var_1227 = real_div(x = sqrt_s0_9, y = sqrt_s_t_9)[name = tensor("op_1227")]; + tensor var_1228 = const()[name = tensor("op_1228"), val = tensor([1, 1, 5, 1])]; + tensor var_1229 = reshape(shape = var_1228, x = var_1227)[name = tensor("op_1229")]; + tensor cross_9 = mul(x = var_1226, y = var_1229)[name = tensor("cross_9")]; tensor out_25 = add(x = inner_9, y = cross_9)[name = tensor("out_25")]; - tensor var_1189 = const()[name = tensor("op_1189"), val = tensor([1, 1, 5, 1])]; - tensor var_1190 = reshape(shape = var_1189, x = valid_mask)[name = tensor("op_1190")]; - tensor v_masked_1 = mul(x = v_9, y = var_1190)[name = tensor("v_masked_1")]; - tensor var_1192 = mul(x = prev_kv_9, y = sqrt_s0_9)[name = tensor("op_1192")]; - tensor var_1194_transpose_x_1 = const()[name = tensor("op_1194_transpose_x_1"), val = tensor(true)]; - tensor var_1194_transpose_y_1 = const()[name = tensor("op_1194_transpose_y_1"), val = tensor(false)]; - tensor var_1194 = matmul(transpose_x = var_1194_transpose_x_1, transpose_y = var_1194_transpose_y_1, x = k_9, y = v_masked_1)[name = tensor("op_1194")]; - tensor new_kv_unnorm_9 = add(x = var_1192, y = var_1194)[name = tensor("new_kv_unnorm_9")]; - tensor var_1196_keep_dims_0 = const()[name = tensor("op_1196_keep_dims_0"), val = tensor(false)]; - tensor var_1196 = reduce_sum(keep_dims = var_1196_keep_dims_0, x = valid_mask)[name = tensor("op_1196")]; - tensor var_1197 = const()[name = tensor("op_1197"), val = tensor([1])]; - tensor var_1198 = reshape(shape = var_1197, x = var_1196)[name = tensor("op_1198")]; - tensor new_scale_9 = add(x = prev_scale_9, y = var_1198)[name = tensor("new_scale_9")]; + tensor var_1232 = const()[name = tensor("op_1232"), val = tensor([1, 1, 5, 1])]; + tensor var_1233 = reshape(shape = var_1232, x = valid_mask)[name = tensor("op_1233")]; + tensor v_masked_1 = mul(x = v_9, y = var_1233)[name = tensor("v_masked_1")]; + tensor var_1235 = mul(x = prev_kv_9, y = sqrt_s0_9)[name = tensor("op_1235")]; + tensor var_1237_transpose_x_1 = const()[name = tensor("op_1237_transpose_x_1"), val = tensor(true)]; + tensor var_1237_transpose_y_1 = const()[name = tensor("op_1237_transpose_y_1"), val = tensor(false)]; + tensor var_1237 = matmul(transpose_x = var_1237_transpose_x_1, transpose_y = var_1237_transpose_y_1, x = k_9, y = v_masked_1)[name = tensor("op_1237")]; + tensor new_kv_unnorm_9 = add(x = var_1235, y = var_1237)[name = tensor("new_kv_unnorm_9")]; + tensor var_1239_keep_dims_0 = const()[name = tensor("op_1239_keep_dims_0"), val = tensor(false)]; + tensor var_1239 = reduce_sum(keep_dims = var_1239_keep_dims_0, x = valid_mask)[name = tensor("op_1239")]; + tensor var_1240 = const()[name = tensor("op_1240"), val = tensor([1])]; + tensor var_1241 = reshape(shape = var_1240, x = var_1239)[name = tensor("op_1241")]; + tensor new_scale_9 = add(x = prev_scale_9, y = var_1241)[name = tensor("new_scale_9")]; tensor const_21 = const()[name = tensor("const_21"), val = tensor(0x1.fffffep+127)]; - tensor clip_2 = clip(alpha = var_1055, beta = const_21, x = new_scale_9)[name = tensor("clip_2")]; + tensor clip_2 = clip(alpha = var_69, beta = const_21, x = new_scale_9)[name = tensor("clip_2")]; tensor sqrt_new_scale_1 = sqrt(x = clip_2)[name = tensor("sqrt_new_scale_1")]; - tensor var_1202 = real_div(x = new_kv_unnorm_9, y = sqrt_new_scale_1)[name = tensor("op_1202")]; - tensor var_1203_perm_0 = const()[name = tensor("op_1203_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_1245 = real_div(x = new_kv_unnorm_9, y = sqrt_new_scale_1)[name = tensor("op_1245")]; + tensor var_1246_perm_0 = const()[name = tensor("op_1246_perm_0"), val = tensor([0, 2, 1, 3])]; tensor out_27_axes_0 = const()[name = tensor("out_27_axes_0"), val = tensor([-1])]; - tensor var_1203 = transpose(perm = var_1203_perm_0, x = out_25)[name = tensor("transpose_24")]; - tensor out_27 = layer_norm(axes = out_27_axes_0, epsilon = var_1052, x = var_1203)[name = tensor("out_27")]; - tensor var_1207 = const()[name = tensor("op_1207"), val = tensor([9, 5, 256])]; - tensor out_29 = reshape(shape = var_1207, x = out_27)[name = tensor("out_29")]; - tensor var_1209 = silu(x = input_169)[name = tensor("op_1209")]; - tensor input_171 = mul(x = var_1209, y = out_29)[name = tensor("input_171")]; - tensor ret_out_9 = linear(bias = decoder_out_proj_0_bias, weight = decoder_out_proj_0_weight, x = input_171)[name = tensor("linear_42")]; - tensor input_173 = add(x = x_29, y = ret_out_9)[name = tensor("input_173")]; + tensor var_1246 = transpose(perm = var_1246_perm_0, x = out_25)[name = tensor("transpose_24")]; + tensor out_27 = layer_norm(axes = out_27_axes_0, epsilon = var_84, x = var_1246)[name = tensor("out_27")]; + tensor var_1250 = const()[name = tensor("op_1250"), val = tensor([9, 5, 256])]; + tensor out_29 = reshape(shape = var_1250, x = out_27)[name = tensor("out_29")]; + tensor var_1252 = silu(x = input_171)[name = tensor("op_1252")]; + tensor input_173 = mul(x = var_1252, y = out_29)[name = tensor("input_173")]; + tensor ret_out_9 = linear(bias = inner_decoder_out_proj_0_bias, weight = inner_decoder_out_proj_0_weight, x = input_173)[name = tensor("linear_42")]; + tensor input_175 = add(x = x_29, y = ret_out_9)[name = tensor("input_175")]; tensor xt_1_axes_0 = const()[name = tensor("xt_1_axes_0"), val = tensor([-1])]; - tensor xt_1 = layer_norm(axes = xt_1_axes_0, beta = decoder_norm11_0_bias, epsilon = var_1050, gamma = decoder_norm11_0_weight, x = input_173)[name = tensor("xt_1")]; - tensor var_1219 = const()[name = tensor("op_1219"), val = tensor([1, 9, 5, 256])]; - tensor var_1220 = reshape(shape = var_1219, x = xt_1)[name = tensor("op_1220")]; - tensor var_1221_perm_0 = const()[name = tensor("op_1221_perm_0"), val = tensor([0, 2, 1, 3])]; - tensor var_1224 = const()[name = tensor("op_1224"), val = tensor([5, 9, 256])]; - tensor var_1221 = transpose(perm = var_1221_perm_0, x = var_1220)[name = tensor("transpose_23")]; - tensor query_1 = reshape(shape = var_1224, x = var_1221)[name = tensor("query_1")]; + tensor xt_1 = layer_norm(axes = xt_1_axes_0, beta = inner_decoder_norm11_0_bias, epsilon = var_76, gamma = inner_decoder_norm11_0_weight, x = input_175)[name = tensor("xt_1")]; + tensor var_1262 = const()[name = tensor("op_1262"), val = tensor([1, 9, 5, 256])]; + tensor var_1263 = reshape(shape = var_1262, x = xt_1)[name = tensor("op_1263")]; + tensor var_1264_perm_0 = const()[name = tensor("op_1264_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_1267 = const()[name = tensor("op_1267"), val = tensor([5, 9, 256])]; + tensor var_1264 = transpose(perm = var_1264_perm_0, x = var_1263)[name = tensor("transpose_23")]; + tensor query_1 = reshape(shape = var_1267, x = var_1264)[name = tensor("query_1")]; tensor query_3_perm_0 = const()[name = tensor("query_3_perm_0"), val = tensor([1, 0, 2])]; tensor query_3 = transpose(perm = query_3_perm_0, x = query_1)[name = tensor("transpose_22")]; - tensor var_1247 = linear(bias = decoder_self_attn2_0_in_proj_bias, weight = decoder_self_attn2_0_in_proj_weight, x = query_3)[name = tensor("linear_43")]; + tensor var_1290 = linear(bias = inner_decoder_self_attn2_0_in_proj_bias, weight = inner_decoder_self_attn2_0_in_proj_weight, x = query_3)[name = tensor("linear_43")]; tensor concat_1 = const()[name = tensor("concat_1"), val = tensor([9, 5, 3, 256])]; - tensor var_1249 = reshape(shape = concat_1, x = var_1247)[name = tensor("op_1249")]; - tensor var_1250_axes_0 = const()[name = tensor("op_1250_axes_0"), val = tensor([0])]; - tensor var_1250 = expand_dims(axes = var_1250_axes_0, x = var_1249)[name = tensor("op_1250")]; - tensor var_1251_perm_0 = const()[name = tensor("op_1251_perm_0"), val = tensor([-2, 1, 2, 0, 4])]; - tensor var_1252_axes_0 = const()[name = tensor("op_1252_axes_0"), val = tensor([-2])]; - tensor var_1251 = transpose(perm = var_1251_perm_0, x = var_1250)[name = tensor("transpose_21")]; - tensor var_1252 = squeeze(axes = var_1252_axes_0, x = var_1251)[name = tensor("op_1252")]; + tensor var_1292 = reshape(shape = concat_1, x = var_1290)[name = tensor("op_1292")]; + tensor var_1293_axes_0 = const()[name = tensor("op_1293_axes_0"), val = tensor([0])]; + tensor var_1293 = expand_dims(axes = var_1293_axes_0, x = var_1292)[name = tensor("op_1293")]; + tensor var_1294_perm_0 = const()[name = tensor("op_1294_perm_0"), val = tensor([-2, 1, 2, 0, 4])]; + tensor var_1295_axes_0 = const()[name = tensor("op_1295_axes_0"), val = tensor([-2])]; + tensor var_1294 = transpose(perm = var_1294_perm_0, x = var_1293)[name = tensor("transpose_21")]; + tensor var_1295 = squeeze(axes = var_1295_axes_0, x = var_1294)[name = tensor("op_1295")]; tensor q_11_begin_0 = const()[name = tensor("q_11_begin_0"), val = tensor([0, 0, 0, 0])]; tensor q_11_end_0 = const()[name = tensor("q_11_end_0"), val = tensor([1, 9, 5, 256])]; tensor q_11_end_mask_0 = const()[name = tensor("q_11_end_mask_0"), val = tensor([false, true, true, true])]; tensor q_11_squeeze_mask_0 = const()[name = tensor("q_11_squeeze_mask_0"), val = tensor([true, false, false, false])]; - tensor q_11 = slice_by_index(begin = q_11_begin_0, end = q_11_end_0, end_mask = q_11_end_mask_0, squeeze_mask = q_11_squeeze_mask_0, x = var_1252)[name = tensor("q_11")]; + tensor q_11 = slice_by_index(begin = q_11_begin_0, end = q_11_end_0, end_mask = q_11_end_mask_0, squeeze_mask = q_11_squeeze_mask_0, x = var_1295)[name = tensor("q_11")]; tensor k_11_begin_0 = const()[name = tensor("k_11_begin_0"), val = tensor([1, 0, 0, 0])]; tensor k_11_end_0 = const()[name = tensor("k_11_end_0"), val = tensor([2, 9, 5, 256])]; tensor k_11_end_mask_0 = const()[name = tensor("k_11_end_mask_0"), val = tensor([false, true, true, true])]; tensor k_11_squeeze_mask_0 = const()[name = tensor("k_11_squeeze_mask_0"), val = tensor([true, false, false, false])]; - tensor k_11 = slice_by_index(begin = k_11_begin_0, end = k_11_end_0, end_mask = k_11_end_mask_0, squeeze_mask = k_11_squeeze_mask_0, x = var_1252)[name = tensor("k_11")]; + tensor k_11 = slice_by_index(begin = k_11_begin_0, end = k_11_end_0, end_mask = k_11_end_mask_0, squeeze_mask = k_11_squeeze_mask_0, x = var_1295)[name = tensor("k_11")]; tensor v_11_begin_0 = const()[name = tensor("v_11_begin_0"), val = tensor([2, 0, 0, 0])]; tensor v_11_end_0 = const()[name = tensor("v_11_end_0"), val = tensor([3, 9, 5, 256])]; tensor v_11_end_mask_0 = const()[name = tensor("v_11_end_mask_0"), val = tensor([false, true, true, true])]; tensor v_11_squeeze_mask_0 = const()[name = tensor("v_11_squeeze_mask_0"), val = tensor([true, false, false, false])]; - tensor v_11 = slice_by_index(begin = v_11_begin_0, end = v_11_end_0, end_mask = v_11_end_mask_0, squeeze_mask = v_11_squeeze_mask_0, x = var_1252)[name = tensor("v_11")]; - tensor var_1260 = const()[name = tensor("op_1260"), val = tensor([9, 20, 64])]; - tensor var_1261 = reshape(shape = var_1260, x = q_11)[name = tensor("op_1261")]; + tensor v_11 = slice_by_index(begin = v_11_begin_0, end = v_11_end_0, end_mask = v_11_end_mask_0, squeeze_mask = v_11_squeeze_mask_0, x = var_1295)[name = tensor("v_11")]; + tensor var_1303 = const()[name = tensor("op_1303"), val = tensor([9, 20, 64])]; + tensor var_1304 = reshape(shape = var_1303, x = q_11)[name = tensor("op_1304")]; tensor q_13_perm_0 = const()[name = tensor("q_13_perm_0"), val = tensor([1, 0, 2])]; - tensor var_1267 = const()[name = tensor("op_1267"), val = tensor([9, 20, 64])]; - tensor var_1268 = reshape(shape = var_1267, x = k_11)[name = tensor("op_1268")]; + tensor var_1310 = const()[name = tensor("op_1310"), val = tensor([9, 20, 64])]; + tensor var_1311 = reshape(shape = var_1310, x = k_11)[name = tensor("op_1311")]; tensor k_13_perm_0 = const()[name = tensor("k_13_perm_0"), val = tensor([1, 0, 2])]; - tensor var_1274 = const()[name = tensor("op_1274"), val = tensor([9, 20, 64])]; - tensor var_1275 = reshape(shape = var_1274, x = v_11)[name = tensor("op_1275")]; + tensor var_1317 = const()[name = tensor("op_1317"), val = tensor([9, 20, 64])]; + tensor var_1318 = reshape(shape = var_1317, x = v_11)[name = tensor("op_1318")]; tensor v_13_perm_0 = const()[name = tensor("v_13_perm_0"), val = tensor([1, 0, 2])]; - tensor var_1278 = const()[name = tensor("op_1278"), val = tensor([5, 4, 9, 64])]; - tensor q_13 = transpose(perm = q_13_perm_0, x = var_1261)[name = tensor("transpose_20")]; - tensor q_15 = reshape(shape = var_1278, x = q_13)[name = tensor("q_15")]; - tensor var_1280 = const()[name = tensor("op_1280"), val = tensor([5, 4, 9, 64])]; - tensor k_13 = transpose(perm = k_13_perm_0, x = var_1268)[name = tensor("transpose_19")]; - tensor k_15 = reshape(shape = var_1280, x = k_13)[name = tensor("k_15")]; - tensor var_1282 = const()[name = tensor("op_1282"), val = tensor([5, 4, 9, 64])]; - tensor v_13 = transpose(perm = v_13_perm_0, x = var_1275)[name = tensor("transpose_18")]; - tensor v_15 = reshape(shape = var_1282, x = v_13)[name = tensor("v_15")]; + tensor var_1321 = const()[name = tensor("op_1321"), val = tensor([5, 4, 9, 64])]; + tensor q_13 = transpose(perm = q_13_perm_0, x = var_1304)[name = tensor("transpose_20")]; + tensor q_15 = reshape(shape = var_1321, x = q_13)[name = tensor("q_15")]; + tensor var_1323 = const()[name = tensor("op_1323"), val = tensor([5, 4, 9, 64])]; + tensor k_13 = transpose(perm = k_13_perm_0, x = var_1311)[name = tensor("transpose_19")]; + tensor k_15 = reshape(shape = var_1323, x = k_13)[name = tensor("k_15")]; + tensor var_1325 = const()[name = tensor("op_1325"), val = tensor([5, 4, 9, 64])]; + tensor v_13 = transpose(perm = v_13_perm_0, x = var_1318)[name = tensor("transpose_18")]; + tensor v_15 = reshape(shape = var_1325, x = v_13)[name = tensor("v_15")]; tensor mul_1_y_0 = const()[name = tensor("mul_1_y_0"), val = tensor(0x1p-3)]; tensor mul_1 = mul(x = q_15, y = mul_1_y_0)[name = tensor("mul_1")]; tensor matmul_0_transpose_y_0 = const()[name = tensor("matmul_0_transpose_y_0"), val = tensor(true)]; @@ -1159,30 +1178,30 @@ program(1.0) tensor attn_output_1_transpose_x_0 = const()[name = tensor("attn_output_1_transpose_x_0"), val = tensor(false)]; tensor attn_output_1_transpose_y_0 = const()[name = tensor("attn_output_1_transpose_y_0"), val = tensor(false)]; tensor attn_output_1 = matmul(transpose_x = attn_output_1_transpose_x_0, transpose_y = attn_output_1_transpose_y_0, x = softmax_0, y = v_15)[name = tensor("attn_output_1")]; - tensor var_1285 = const()[name = tensor("op_1285"), val = tensor([2, 0, 1, 3])]; - tensor var_1290 = const()[name = tensor("op_1290"), val = tensor([45, 256])]; - tensor var_1286 = transpose(perm = var_1285, x = attn_output_1)[name = tensor("transpose_17")]; - tensor attn_output_3 = reshape(shape = var_1290, x = var_1286)[name = tensor("attn_output_3")]; - tensor attn_output_5 = linear(bias = decoder_self_attn2_0_out_proj_bias, weight = decoder_self_attn2_0_out_proj_weight, x = attn_output_3)[name = tensor("linear_44")]; - tensor var_1294 = const()[name = tensor("op_1294"), val = tensor([9, 5, 256])]; - tensor attn_output_7 = reshape(shape = var_1294, x = attn_output_5)[name = tensor("attn_output_7")]; + tensor var_1328 = const()[name = tensor("op_1328"), val = tensor([2, 0, 1, 3])]; + tensor var_1333 = const()[name = tensor("op_1333"), val = tensor([45, 256])]; + tensor var_1329 = transpose(perm = var_1328, x = attn_output_1)[name = tensor("transpose_17")]; + tensor attn_output_3 = reshape(shape = var_1333, x = var_1329)[name = tensor("attn_output_3")]; + tensor attn_output_5 = linear(bias = inner_decoder_self_attn2_0_out_proj_bias, weight = inner_decoder_self_attn2_0_out_proj_weight, x = attn_output_3)[name = tensor("linear_44")]; + tensor var_1337 = const()[name = tensor("op_1337"), val = tensor([9, 5, 256])]; + tensor attn_output_7 = reshape(shape = var_1337, x = attn_output_5)[name = tensor("attn_output_7")]; tensor sa2_out_1_perm_0 = const()[name = tensor("sa2_out_1_perm_0"), val = tensor([1, 0, 2])]; tensor sa2_out_1 = transpose(perm = sa2_out_1_perm_0, x = attn_output_7)[name = tensor("transpose_16")]; - tensor input_175 = add(x = query_1, y = sa2_out_1)[name = tensor("input_175")]; - tensor input_177_axes_0 = const()[name = tensor("input_177_axes_0"), val = tensor([-1])]; - tensor input_177 = layer_norm(axes = input_177_axes_0, beta = decoder_norm21_0_bias, epsilon = var_1050, gamma = decoder_norm21_0_weight, x = input_175)[name = tensor("input_177")]; - tensor input_179 = linear(bias = decoder_linear1_0_bias, weight = decoder_linear1_0_weight, x = input_177)[name = tensor("linear_45")]; - tensor input_181 = relu(x = input_179)[name = tensor("input_181")]; - tensor ffn_out_1 = linear(bias = decoder_linear2_0_bias, weight = decoder_linear2_0_weight, x = input_181)[name = tensor("linear_46")]; - tensor input_183 = add(x = input_177, y = ffn_out_1)[name = tensor("input_183")]; + tensor input_177 = add(x = query_1, y = sa2_out_1)[name = tensor("input_177")]; + tensor input_179_axes_0 = const()[name = tensor("input_179_axes_0"), val = tensor([-1])]; + tensor input_179 = layer_norm(axes = input_179_axes_0, beta = inner_decoder_norm21_0_bias, epsilon = var_76, gamma = inner_decoder_norm21_0_weight, x = input_177)[name = tensor("input_179")]; + tensor input_181 = linear(bias = inner_decoder_linear1_0_bias, weight = inner_decoder_linear1_0_weight, x = input_179)[name = tensor("linear_45")]; + tensor input_183 = relu(x = input_181)[name = tensor("input_183")]; + tensor ffn_out_1 = linear(bias = inner_decoder_linear2_0_bias, weight = inner_decoder_linear2_0_weight, x = input_183)[name = tensor("linear_46")]; + tensor input_185 = add(x = input_179, y = ffn_out_1)[name = tensor("input_185")]; tensor xt_3_axes_0 = const()[name = tensor("xt_3_axes_0"), val = tensor([-1])]; - tensor xt_3 = layer_norm(axes = xt_3_axes_0, beta = decoder_norm22_0_bias, epsilon = var_1050, gamma = decoder_norm22_0_weight, x = input_183)[name = tensor("xt_3")]; - tensor var_1314 = const()[name = tensor("op_1314"), val = tensor([1, 5, 9, 256])]; - tensor x_31 = reshape(shape = var_1314, x = xt_3)[name = tensor("x_31")]; - tensor var_1316_perm_0 = const()[name = tensor("op_1316_perm_0"), val = tensor([0, 2, 1, 3])]; - tensor var_1320 = const()[name = tensor("op_1320"), val = tensor([9, 5, 256])]; - tensor var_1316 = transpose(perm = var_1316_perm_0, x = x_31)[name = tensor("transpose_15")]; - tensor x = reshape(shape = var_1320, x = var_1316)[name = tensor("x")]; + tensor xt_3 = layer_norm(axes = xt_3_axes_0, beta = inner_decoder_norm22_0_bias, epsilon = var_76, gamma = inner_decoder_norm22_0_weight, x = input_185)[name = tensor("xt_3")]; + tensor var_1357 = const()[name = tensor("op_1357"), val = tensor([1, 5, 9, 256])]; + tensor x_31 = reshape(shape = var_1357, x = xt_3)[name = tensor("x_31")]; + tensor var_1359_perm_0 = const()[name = tensor("op_1359_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_1363 = const()[name = tensor("op_1363"), val = tensor([9, 5, 256])]; + tensor var_1359 = transpose(perm = var_1359_perm_0, x = x_31)[name = tensor("transpose_15")]; + tensor x = reshape(shape = var_1363, x = var_1359)[name = tensor("x")]; tensor prev_kv_begin_0 = const()[name = tensor("prev_kv_begin_0"), val = tensor([1, 0, 0, 0, 0])]; tensor prev_kv_end_0 = const()[name = tensor("prev_kv_end_0"), val = tensor([2, 9, 4, 64, 64])]; tensor prev_kv_end_mask_0 = const()[name = tensor("prev_kv_end_mask_0"), val = tensor([false, true, true, true, true])]; @@ -1193,120 +1212,120 @@ program(1.0) tensor prev_scale_end_mask_0 = const()[name = tensor("prev_scale_end_mask_0"), val = tensor([false, true])]; tensor prev_scale_squeeze_mask_0 = const()[name = tensor("prev_scale_squeeze_mask_0"), val = tensor([true, false])]; tensor prev_scale = slice_by_index(begin = prev_scale_begin_0, end = prev_scale_end_0, end_mask = prev_scale_end_mask_0, squeeze_mask = prev_scale_squeeze_mask_0, x = dec_scale)[name = tensor("prev_scale")]; - tensor var_1328 = linear(bias = decoder_q_proj_1_bias, weight = decoder_q_proj_1_weight, x = x)[name = tensor("linear_47")]; - tensor var_1329 = const()[name = tensor("op_1329"), val = tensor([9, 5, 4, 64])]; - tensor var_1330 = reshape(shape = var_1329, x = var_1328)[name = tensor("op_1330")]; + tensor var_1371 = linear(bias = inner_decoder_q_proj_1_bias, weight = inner_decoder_q_proj_1_weight, x = x)[name = tensor("linear_47")]; + tensor var_1372 = const()[name = tensor("op_1372"), val = tensor([9, 5, 4, 64])]; + tensor var_1373 = reshape(shape = var_1372, x = var_1371)[name = tensor("op_1373")]; tensor q_17_perm_0 = const()[name = tensor("q_17_perm_0"), val = tensor([0, 2, 1, 3])]; - tensor var_1334 = linear(bias = decoder_k_proj_1_bias, weight = decoder_k_proj_1_weight, x = x)[name = tensor("linear_48")]; - tensor var_1335 = const()[name = tensor("op_1335"), val = tensor(0x1p-3)]; - tensor var_1336 = mul(x = var_1334, y = var_1335)[name = tensor("op_1336")]; - tensor var_1337 = const()[name = tensor("op_1337"), val = tensor([9, 5, 4, 64])]; - tensor var_1338 = reshape(shape = var_1337, x = var_1336)[name = tensor("op_1338")]; + tensor var_1377 = linear(bias = inner_decoder_k_proj_1_bias, weight = inner_decoder_k_proj_1_weight, x = x)[name = tensor("linear_48")]; + tensor var_1378 = const()[name = tensor("op_1378"), val = tensor(0x1p-3)]; + tensor var_1379 = mul(x = var_1377, y = var_1378)[name = tensor("op_1379")]; + tensor var_1380 = const()[name = tensor("op_1380"), val = tensor([9, 5, 4, 64])]; + tensor var_1381 = reshape(shape = var_1380, x = var_1379)[name = tensor("op_1381")]; tensor k_17_perm_0 = const()[name = tensor("k_17_perm_0"), val = tensor([0, 2, 1, 3])]; - tensor var_1342 = linear(bias = decoder_v_proj_1_bias, weight = decoder_v_proj_1_weight, x = x)[name = tensor("linear_49")]; - tensor var_1343 = const()[name = tensor("op_1343"), val = tensor([9, 5, 4, 64])]; - tensor var_1344 = reshape(shape = var_1343, x = var_1342)[name = tensor("op_1344")]; + tensor var_1385 = linear(bias = inner_decoder_v_proj_1_bias, weight = inner_decoder_v_proj_1_weight, x = x)[name = tensor("linear_49")]; + tensor var_1386 = const()[name = tensor("op_1386"), val = tensor([9, 5, 4, 64])]; + tensor var_1387 = reshape(shape = var_1386, x = var_1385)[name = tensor("op_1387")]; tensor v_17_perm_0 = const()[name = tensor("v_17_perm_0"), val = tensor([0, 2, 1, 3])]; - tensor input_187 = linear(bias = decoder_g_proj_1_bias, weight = decoder_g_proj_1_weight, x = x)[name = tensor("linear_50")]; + tensor input_189 = linear(bias = inner_decoder_g_proj_1_bias, weight = inner_decoder_g_proj_1_weight, x = x)[name = tensor("linear_50")]; tensor sqrt_s0 = sqrt(x = prev_scale)[name = tensor("sqrt_s0")]; tensor s_t = add(x = prev_scale, y = cumsum_mask_1)[name = tensor("s_t")]; tensor const_32 = const()[name = tensor("const_32"), val = tensor(0x1.fffffep+127)]; - tensor clip_3 = clip(alpha = var_1055, beta = const_32, x = s_t)[name = tensor("clip_3")]; + tensor clip_3 = clip(alpha = var_69, beta = const_32, x = s_t)[name = tensor("clip_3")]; tensor sqrt_s_t = sqrt(x = clip_3)[name = tensor("sqrt_s_t")]; tensor qk_transpose_x_1 = const()[name = tensor("qk_transpose_x_1"), val = tensor(false)]; tensor qk_transpose_y_1 = const()[name = tensor("qk_transpose_y_1"), val = tensor(true)]; - tensor k_17 = transpose(perm = k_17_perm_0, x = var_1338)[name = tensor("transpose_13")]; - tensor q_17 = transpose(perm = q_17_perm_0, x = var_1330)[name = tensor("transpose_14")]; + tensor k_17 = transpose(perm = k_17_perm_0, x = var_1381)[name = tensor("transpose_13")]; + tensor q_17 = transpose(perm = q_17_perm_0, x = var_1373)[name = tensor("transpose_14")]; tensor qk = matmul(transpose_x = qk_transpose_x_1, transpose_y = qk_transpose_y_1, x = q_17, y = k_17)[name = tensor("qk")]; - tensor var_1359 = const()[name = tensor("op_1359"), val = tensor([5, 1])]; - tensor var_1360 = reshape(shape = var_1359, x = sqrt_s_t)[name = tensor("op_1360")]; - tensor M = real_div(x = causal_with_valid_1, y = var_1360)[name = tensor("M")]; - tensor var_1362 = mul(x = qk, y = M)[name = tensor("op_1362")]; - tensor inner_transpose_x_0 = const()[name = tensor("inner_transpose_x_0"), val = tensor(false)]; - tensor inner_transpose_y_0 = const()[name = tensor("inner_transpose_y_0"), val = tensor(false)]; - tensor v_17 = transpose(perm = v_17_perm_0, x = var_1344)[name = tensor("transpose_12")]; - tensor inner = matmul(transpose_x = inner_transpose_x_0, transpose_y = inner_transpose_y_0, x = var_1362, y = v_17)[name = tensor("inner")]; - tensor var_1364_transpose_x_0 = const()[name = tensor("op_1364_transpose_x_0"), val = tensor(false)]; - tensor var_1364_transpose_y_0 = const()[name = tensor("op_1364_transpose_y_0"), val = tensor(false)]; - tensor var_1364 = matmul(transpose_x = var_1364_transpose_x_0, transpose_y = var_1364_transpose_y_0, x = q_17, y = prev_kv)[name = tensor("op_1364")]; - tensor var_1365 = real_div(x = sqrt_s0, y = sqrt_s_t)[name = tensor("op_1365")]; - tensor var_1366 = const()[name = tensor("op_1366"), val = tensor([1, 1, 5, 1])]; - tensor var_1367 = reshape(shape = var_1366, x = var_1365)[name = tensor("op_1367")]; - tensor cross = mul(x = var_1364, y = var_1367)[name = tensor("cross")]; - tensor out_31 = add(x = inner, y = cross)[name = tensor("out_31")]; - tensor v_masked = mul(x = v_17, y = var_1190)[name = tensor("v_masked")]; - tensor var_1373 = mul(x = prev_kv, y = sqrt_s0)[name = tensor("op_1373")]; - tensor var_1375_transpose_x_1 = const()[name = tensor("op_1375_transpose_x_1"), val = tensor(true)]; - tensor var_1375_transpose_y_1 = const()[name = tensor("op_1375_transpose_y_1"), val = tensor(false)]; - tensor var_1375 = matmul(transpose_x = var_1375_transpose_x_1, transpose_y = var_1375_transpose_y_1, x = k_17, y = v_masked)[name = tensor("op_1375")]; - tensor new_kv_unnorm = add(x = var_1373, y = var_1375)[name = tensor("new_kv_unnorm")]; - tensor new_scale = add(x = prev_scale, y = var_1198)[name = tensor("new_scale")]; + tensor var_1402 = const()[name = tensor("op_1402"), val = tensor([5, 1])]; + tensor var_1403 = reshape(shape = var_1402, x = sqrt_s_t)[name = tensor("op_1403")]; + tensor M = real_div(x = causal_with_valid_1, y = var_1403)[name = tensor("M")]; + tensor var_1405 = mul(x = qk, y = M)[name = tensor("op_1405")]; + tensor inner_11_transpose_x_0 = const()[name = tensor("inner_11_transpose_x_0"), val = tensor(false)]; + tensor inner_11_transpose_y_0 = const()[name = tensor("inner_11_transpose_y_0"), val = tensor(false)]; + tensor v_17 = transpose(perm = v_17_perm_0, x = var_1387)[name = tensor("transpose_12")]; + tensor inner_11 = matmul(transpose_x = inner_11_transpose_x_0, transpose_y = inner_11_transpose_y_0, x = var_1405, y = v_17)[name = tensor("inner_11")]; + tensor var_1407_transpose_x_0 = const()[name = tensor("op_1407_transpose_x_0"), val = tensor(false)]; + tensor var_1407_transpose_y_0 = const()[name = tensor("op_1407_transpose_y_0"), val = tensor(false)]; + tensor var_1407 = matmul(transpose_x = var_1407_transpose_x_0, transpose_y = var_1407_transpose_y_0, x = q_17, y = prev_kv)[name = tensor("op_1407")]; + tensor var_1408 = real_div(x = sqrt_s0, y = sqrt_s_t)[name = tensor("op_1408")]; + tensor var_1409 = const()[name = tensor("op_1409"), val = tensor([1, 1, 5, 1])]; + tensor var_1410 = reshape(shape = var_1409, x = var_1408)[name = tensor("op_1410")]; + tensor cross = mul(x = var_1407, y = var_1410)[name = tensor("cross")]; + tensor out_31 = add(x = inner_11, y = cross)[name = tensor("out_31")]; + tensor v_masked = mul(x = v_17, y = var_1233)[name = tensor("v_masked")]; + tensor var_1416 = mul(x = prev_kv, y = sqrt_s0)[name = tensor("op_1416")]; + tensor var_1418_transpose_x_1 = const()[name = tensor("op_1418_transpose_x_1"), val = tensor(true)]; + tensor var_1418_transpose_y_1 = const()[name = tensor("op_1418_transpose_y_1"), val = tensor(false)]; + tensor var_1418 = matmul(transpose_x = var_1418_transpose_x_1, transpose_y = var_1418_transpose_y_1, x = k_17, y = v_masked)[name = tensor("op_1418")]; + tensor new_kv_unnorm = add(x = var_1416, y = var_1418)[name = tensor("new_kv_unnorm")]; + tensor new_scale = add(x = prev_scale, y = var_1241)[name = tensor("new_scale")]; tensor const_33 = const()[name = tensor("const_33"), val = tensor(0x1.fffffep+127)]; - tensor clip_4 = clip(alpha = var_1055, beta = const_33, x = new_scale)[name = tensor("clip_4")]; + tensor clip_4 = clip(alpha = var_69, beta = const_33, x = new_scale)[name = tensor("clip_4")]; tensor sqrt_new_scale = sqrt(x = clip_4)[name = tensor("sqrt_new_scale")]; tensor nkv = real_div(x = new_kv_unnorm, y = sqrt_new_scale)[name = tensor("nkv")]; - tensor var_1384_perm_0 = const()[name = tensor("op_1384_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_1427_perm_0 = const()[name = tensor("op_1427_perm_0"), val = tensor([0, 2, 1, 3])]; tensor out_33_axes_0 = const()[name = tensor("out_33_axes_0"), val = tensor([-1])]; - tensor var_1384 = transpose(perm = var_1384_perm_0, x = out_31)[name = tensor("transpose_11")]; - tensor out_33 = layer_norm(axes = out_33_axes_0, epsilon = var_1052, x = var_1384)[name = tensor("out_33")]; - tensor var_1388 = const()[name = tensor("op_1388"), val = tensor([9, 5, 256])]; - tensor out = reshape(shape = var_1388, x = out_33)[name = tensor("out")]; - tensor var_1390 = silu(x = input_187)[name = tensor("op_1390")]; - tensor input_189 = mul(x = var_1390, y = out)[name = tensor("input_189")]; - tensor ret_out = linear(bias = decoder_out_proj_1_bias, weight = decoder_out_proj_1_weight, x = input_189)[name = tensor("linear_51")]; - tensor input_191 = add(x = x, y = ret_out)[name = tensor("input_191")]; + tensor var_1427 = transpose(perm = var_1427_perm_0, x = out_31)[name = tensor("transpose_11")]; + tensor out_33 = layer_norm(axes = out_33_axes_0, epsilon = var_84, x = var_1427)[name = tensor("out_33")]; + tensor var_1431 = const()[name = tensor("op_1431"), val = tensor([9, 5, 256])]; + tensor out = reshape(shape = var_1431, x = out_33)[name = tensor("out")]; + tensor var_1433 = silu(x = input_189)[name = tensor("op_1433")]; + tensor input_191 = mul(x = var_1433, y = out)[name = tensor("input_191")]; + tensor ret_out = linear(bias = inner_decoder_out_proj_1_bias, weight = inner_decoder_out_proj_1_weight, x = input_191)[name = tensor("linear_51")]; + tensor input_193 = add(x = x, y = ret_out)[name = tensor("input_193")]; tensor xt_5_axes_0 = const()[name = tensor("xt_5_axes_0"), val = tensor([-1])]; - tensor xt_5 = layer_norm(axes = xt_5_axes_0, beta = decoder_norm11_1_bias, epsilon = var_1050, gamma = decoder_norm11_1_weight, x = input_191)[name = tensor("xt_5")]; - tensor var_1400 = const()[name = tensor("op_1400"), val = tensor([1, 9, 5, 256])]; - tensor var_1401 = reshape(shape = var_1400, x = xt_5)[name = tensor("op_1401")]; - tensor var_1402_perm_0 = const()[name = tensor("op_1402_perm_0"), val = tensor([0, 2, 1, 3])]; - tensor var_1405 = const()[name = tensor("op_1405"), val = tensor([5, 9, 256])]; - tensor var_1402 = transpose(perm = var_1402_perm_0, x = var_1401)[name = tensor("transpose_10")]; - tensor query_5 = reshape(shape = var_1405, x = var_1402)[name = tensor("query_5")]; + tensor xt_5 = layer_norm(axes = xt_5_axes_0, beta = inner_decoder_norm11_1_bias, epsilon = var_76, gamma = inner_decoder_norm11_1_weight, x = input_193)[name = tensor("xt_5")]; + tensor var_1443 = const()[name = tensor("op_1443"), val = tensor([1, 9, 5, 256])]; + tensor var_1444 = reshape(shape = var_1443, x = xt_5)[name = tensor("op_1444")]; + tensor var_1445_perm_0 = const()[name = tensor("op_1445_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_1448 = const()[name = tensor("op_1448"), val = tensor([5, 9, 256])]; + tensor var_1445 = transpose(perm = var_1445_perm_0, x = var_1444)[name = tensor("transpose_10")]; + tensor query_5 = reshape(shape = var_1448, x = var_1445)[name = tensor("query_5")]; tensor query_perm_0 = const()[name = tensor("query_perm_0"), val = tensor([1, 0, 2])]; tensor query = transpose(perm = query_perm_0, x = query_5)[name = tensor("transpose_9")]; - tensor var_1428 = linear(bias = decoder_self_attn2_1_in_proj_bias, weight = decoder_self_attn2_1_in_proj_weight, x = query)[name = tensor("linear_52")]; + tensor var_1471 = linear(bias = inner_decoder_self_attn2_1_in_proj_bias, weight = inner_decoder_self_attn2_1_in_proj_weight, x = query)[name = tensor("linear_52")]; tensor concat_2 = const()[name = tensor("concat_2"), val = tensor([9, 5, 3, 256])]; - tensor var_1430 = reshape(shape = concat_2, x = var_1428)[name = tensor("op_1430")]; - tensor var_1431_axes_0 = const()[name = tensor("op_1431_axes_0"), val = tensor([0])]; - tensor var_1431 = expand_dims(axes = var_1431_axes_0, x = var_1430)[name = tensor("op_1431")]; - tensor var_1432_perm_0 = const()[name = tensor("op_1432_perm_0"), val = tensor([-2, 1, 2, 0, 4])]; - tensor var_1433_axes_0 = const()[name = tensor("op_1433_axes_0"), val = tensor([-2])]; - tensor var_1432 = transpose(perm = var_1432_perm_0, x = var_1431)[name = tensor("transpose_8")]; - tensor var_1433 = squeeze(axes = var_1433_axes_0, x = var_1432)[name = tensor("op_1433")]; + tensor var_1473 = reshape(shape = concat_2, x = var_1471)[name = tensor("op_1473")]; + tensor var_1474_axes_0 = const()[name = tensor("op_1474_axes_0"), val = tensor([0])]; + tensor var_1474 = expand_dims(axes = var_1474_axes_0, x = var_1473)[name = tensor("op_1474")]; + tensor var_1475_perm_0 = const()[name = tensor("op_1475_perm_0"), val = tensor([-2, 1, 2, 0, 4])]; + tensor var_1476_axes_0 = const()[name = tensor("op_1476_axes_0"), val = tensor([-2])]; + tensor var_1475 = transpose(perm = var_1475_perm_0, x = var_1474)[name = tensor("transpose_8")]; + tensor var_1476 = squeeze(axes = var_1476_axes_0, x = var_1475)[name = tensor("op_1476")]; tensor q_19_begin_0 = const()[name = tensor("q_19_begin_0"), val = tensor([0, 0, 0, 0])]; tensor q_19_end_0 = const()[name = tensor("q_19_end_0"), val = tensor([1, 9, 5, 256])]; tensor q_19_end_mask_0 = const()[name = tensor("q_19_end_mask_0"), val = tensor([false, true, true, true])]; tensor q_19_squeeze_mask_0 = const()[name = tensor("q_19_squeeze_mask_0"), val = tensor([true, false, false, false])]; - tensor q_19 = slice_by_index(begin = q_19_begin_0, end = q_19_end_0, end_mask = q_19_end_mask_0, squeeze_mask = q_19_squeeze_mask_0, x = var_1433)[name = tensor("q_19")]; + tensor q_19 = slice_by_index(begin = q_19_begin_0, end = q_19_end_0, end_mask = q_19_end_mask_0, squeeze_mask = q_19_squeeze_mask_0, x = var_1476)[name = tensor("q_19")]; tensor k_19_begin_0 = const()[name = tensor("k_19_begin_0"), val = tensor([1, 0, 0, 0])]; tensor k_19_end_0 = const()[name = tensor("k_19_end_0"), val = tensor([2, 9, 5, 256])]; tensor k_19_end_mask_0 = const()[name = tensor("k_19_end_mask_0"), val = tensor([false, true, true, true])]; tensor k_19_squeeze_mask_0 = const()[name = tensor("k_19_squeeze_mask_0"), val = tensor([true, false, false, false])]; - tensor k_19 = slice_by_index(begin = k_19_begin_0, end = k_19_end_0, end_mask = k_19_end_mask_0, squeeze_mask = k_19_squeeze_mask_0, x = var_1433)[name = tensor("k_19")]; + tensor k_19 = slice_by_index(begin = k_19_begin_0, end = k_19_end_0, end_mask = k_19_end_mask_0, squeeze_mask = k_19_squeeze_mask_0, x = var_1476)[name = tensor("k_19")]; tensor v_19_begin_0 = const()[name = tensor("v_19_begin_0"), val = tensor([2, 0, 0, 0])]; tensor v_19_end_0 = const()[name = tensor("v_19_end_0"), val = tensor([3, 9, 5, 256])]; tensor v_19_end_mask_0 = const()[name = tensor("v_19_end_mask_0"), val = tensor([false, true, true, true])]; tensor v_19_squeeze_mask_0 = const()[name = tensor("v_19_squeeze_mask_0"), val = tensor([true, false, false, false])]; - tensor v_19 = slice_by_index(begin = v_19_begin_0, end = v_19_end_0, end_mask = v_19_end_mask_0, squeeze_mask = v_19_squeeze_mask_0, x = var_1433)[name = tensor("v_19")]; - tensor var_1441 = const()[name = tensor("op_1441"), val = tensor([9, 20, 64])]; - tensor var_1442 = reshape(shape = var_1441, x = q_19)[name = tensor("op_1442")]; + tensor v_19 = slice_by_index(begin = v_19_begin_0, end = v_19_end_0, end_mask = v_19_end_mask_0, squeeze_mask = v_19_squeeze_mask_0, x = var_1476)[name = tensor("v_19")]; + tensor var_1484 = const()[name = tensor("op_1484"), val = tensor([9, 20, 64])]; + tensor var_1485 = reshape(shape = var_1484, x = q_19)[name = tensor("op_1485")]; tensor q_21_perm_0 = const()[name = tensor("q_21_perm_0"), val = tensor([1, 0, 2])]; - tensor var_1448 = const()[name = tensor("op_1448"), val = tensor([9, 20, 64])]; - tensor var_1449 = reshape(shape = var_1448, x = k_19)[name = tensor("op_1449")]; + tensor var_1491 = const()[name = tensor("op_1491"), val = tensor([9, 20, 64])]; + tensor var_1492 = reshape(shape = var_1491, x = k_19)[name = tensor("op_1492")]; tensor k_21_perm_0 = const()[name = tensor("k_21_perm_0"), val = tensor([1, 0, 2])]; - tensor var_1455 = const()[name = tensor("op_1455"), val = tensor([9, 20, 64])]; - tensor var_1456 = reshape(shape = var_1455, x = v_19)[name = tensor("op_1456")]; + tensor var_1498 = const()[name = tensor("op_1498"), val = tensor([9, 20, 64])]; + tensor var_1499 = reshape(shape = var_1498, x = v_19)[name = tensor("op_1499")]; tensor v_21_perm_0 = const()[name = tensor("v_21_perm_0"), val = tensor([1, 0, 2])]; - tensor var_1459 = const()[name = tensor("op_1459"), val = tensor([5, 4, 9, 64])]; - tensor q_21 = transpose(perm = q_21_perm_0, x = var_1442)[name = tensor("transpose_7")]; - tensor q = reshape(shape = var_1459, x = q_21)[name = tensor("q")]; - tensor var_1461 = const()[name = tensor("op_1461"), val = tensor([5, 4, 9, 64])]; - tensor k_21 = transpose(perm = k_21_perm_0, x = var_1449)[name = tensor("transpose_6")]; - tensor k = reshape(shape = var_1461, x = k_21)[name = tensor("k")]; - tensor var_1463 = const()[name = tensor("op_1463"), val = tensor([5, 4, 9, 64])]; - tensor v_21 = transpose(perm = v_21_perm_0, x = var_1456)[name = tensor("transpose_5")]; - tensor v = reshape(shape = var_1463, x = v_21)[name = tensor("v")]; + tensor var_1502 = const()[name = tensor("op_1502"), val = tensor([5, 4, 9, 64])]; + tensor q_21 = transpose(perm = q_21_perm_0, x = var_1485)[name = tensor("transpose_7")]; + tensor q = reshape(shape = var_1502, x = q_21)[name = tensor("q")]; + tensor var_1504 = const()[name = tensor("op_1504"), val = tensor([5, 4, 9, 64])]; + tensor k_21 = transpose(perm = k_21_perm_0, x = var_1492)[name = tensor("transpose_6")]; + tensor k = reshape(shape = var_1504, x = k_21)[name = tensor("k")]; + tensor var_1506 = const()[name = tensor("op_1506"), val = tensor([5, 4, 9, 64])]; + tensor v_21 = transpose(perm = v_21_perm_0, x = var_1499)[name = tensor("transpose_5")]; + tensor v = reshape(shape = var_1506, x = v_21)[name = tensor("v")]; tensor mul_3_y_0 = const()[name = tensor("mul_3_y_0"), val = tensor(0x1p-3)]; tensor mul_3 = mul(x = q, y = mul_3_y_0)[name = tensor("mul_3")]; tensor matmul_1_transpose_y_0 = const()[name = tensor("matmul_1_transpose_y_0"), val = tensor(true)]; @@ -1317,36 +1336,36 @@ program(1.0) tensor attn_output_9_transpose_x_0 = const()[name = tensor("attn_output_9_transpose_x_0"), val = tensor(false)]; tensor attn_output_9_transpose_y_0 = const()[name = tensor("attn_output_9_transpose_y_0"), val = tensor(false)]; tensor attn_output_9 = matmul(transpose_x = attn_output_9_transpose_x_0, transpose_y = attn_output_9_transpose_y_0, x = softmax_1, y = v)[name = tensor("attn_output_9")]; - tensor var_1466 = const()[name = tensor("op_1466"), val = tensor([2, 0, 1, 3])]; - tensor var_1471 = const()[name = tensor("op_1471"), val = tensor([45, 256])]; - tensor var_1467 = transpose(perm = var_1466, x = attn_output_9)[name = tensor("transpose_4")]; - tensor attn_output_11 = reshape(shape = var_1471, x = var_1467)[name = tensor("attn_output_11")]; - tensor attn_output_13 = linear(bias = decoder_self_attn2_1_out_proj_bias, weight = decoder_self_attn2_1_out_proj_weight, x = attn_output_11)[name = tensor("linear_53")]; - tensor var_1475 = const()[name = tensor("op_1475"), val = tensor([9, 5, 256])]; - tensor attn_output = reshape(shape = var_1475, x = attn_output_13)[name = tensor("attn_output")]; + tensor var_1509 = const()[name = tensor("op_1509"), val = tensor([2, 0, 1, 3])]; + tensor var_1514 = const()[name = tensor("op_1514"), val = tensor([45, 256])]; + tensor var_1510 = transpose(perm = var_1509, x = attn_output_9)[name = tensor("transpose_4")]; + tensor attn_output_11 = reshape(shape = var_1514, x = var_1510)[name = tensor("attn_output_11")]; + tensor attn_output_13 = linear(bias = inner_decoder_self_attn2_1_out_proj_bias, weight = inner_decoder_self_attn2_1_out_proj_weight, x = attn_output_11)[name = tensor("linear_53")]; + tensor var_1518 = const()[name = tensor("op_1518"), val = tensor([9, 5, 256])]; + tensor attn_output = reshape(shape = var_1518, x = attn_output_13)[name = tensor("attn_output")]; tensor sa2_out_perm_0 = const()[name = tensor("sa2_out_perm_0"), val = tensor([1, 0, 2])]; tensor sa2_out = transpose(perm = sa2_out_perm_0, x = attn_output)[name = tensor("transpose_3")]; - tensor input_193 = add(x = query_5, y = sa2_out)[name = tensor("input_193")]; - tensor input_195_axes_0 = const()[name = tensor("input_195_axes_0"), val = tensor([-1])]; - tensor input_195 = layer_norm(axes = input_195_axes_0, beta = decoder_norm21_1_bias, epsilon = var_1050, gamma = decoder_norm21_1_weight, x = input_193)[name = tensor("input_195")]; - tensor input_197 = linear(bias = decoder_linear1_1_bias, weight = decoder_linear1_1_weight, x = input_195)[name = tensor("linear_54")]; - tensor input_199 = relu(x = input_197)[name = tensor("input_199")]; - tensor ffn_out = linear(bias = decoder_linear2_1_bias, weight = decoder_linear2_1_weight, x = input_199)[name = tensor("linear_55")]; - tensor input_201 = add(x = input_195, y = ffn_out)[name = tensor("input_201")]; + tensor input_195 = add(x = query_5, y = sa2_out)[name = tensor("input_195")]; + tensor input_197_axes_0 = const()[name = tensor("input_197_axes_0"), val = tensor([-1])]; + tensor input_197 = layer_norm(axes = input_197_axes_0, beta = inner_decoder_norm21_1_bias, epsilon = var_76, gamma = inner_decoder_norm21_1_weight, x = input_195)[name = tensor("input_197")]; + tensor input_199 = linear(bias = inner_decoder_linear1_1_bias, weight = inner_decoder_linear1_1_weight, x = input_197)[name = tensor("linear_54")]; + tensor input_201 = relu(x = input_199)[name = tensor("input_201")]; + tensor ffn_out = linear(bias = inner_decoder_linear2_1_bias, weight = inner_decoder_linear2_1_weight, x = input_201)[name = tensor("linear_55")]; + tensor input_203 = add(x = input_197, y = ffn_out)[name = tensor("input_203")]; tensor xt_axes_0 = const()[name = tensor("xt_axes_0"), val = tensor([-1])]; - tensor xt = layer_norm(axes = xt_axes_0, beta = decoder_norm22_1_bias, epsilon = var_1050, gamma = decoder_norm22_1_weight, x = input_201)[name = tensor("xt")]; - tensor var_1495 = const()[name = tensor("op_1495"), val = tensor([1, 5, 9, 256])]; - tensor input = reshape(shape = var_1495, x = xt)[name = tensor("input")]; - tensor var_1497 = const()[name = tensor("op_1497"), val = tensor([-1])]; - tensor var_1498 = reduce_l2_norm(axes = var_1497, keep_dims = var_1053, x = input)[name = tensor("op_1498")]; + tensor xt = layer_norm(axes = xt_axes_0, beta = inner_decoder_norm22_1_bias, epsilon = var_76, gamma = inner_decoder_norm22_1_weight, x = input_203)[name = tensor("xt")]; + tensor var_1538 = const()[name = tensor("op_1538"), val = tensor([1, 5, 9, 256])]; + tensor input = reshape(shape = var_1538, x = xt)[name = tensor("input")]; + tensor var_1540 = const()[name = tensor("op_1540"), val = tensor([-1])]; + tensor var_1541 = reduce_l2_norm(axes = var_1540, keep_dims = var_75, x = input)[name = tensor("op_1541")]; tensor const_42 = const()[name = tensor("const_42"), val = tensor(0x1.fffffep+127)]; - tensor clip_5 = clip(alpha = var_1045, beta = const_42, x = var_1498)[name = tensor("clip_5")]; - tensor var_1500 = real_div(x = input, y = clip_5)[name = tensor("op_1500")]; + tensor clip_5 = clip(alpha = var_90, beta = const_42, x = var_1541)[name = tensor("clip_5")]; + tensor var_1543 = real_div(x = input, y = clip_5)[name = tensor("op_1543")]; tensor concat_6 = const()[name = tensor("concat_6"), val = tensor([5, 1, 256])]; tensor reshape_0 = reshape(shape = concat_6, x = emb)[name = tensor("reshape_0")]; tensor transpose_1_perm_0 = const()[name = tensor("transpose_1_perm_0"), val = tensor([0, 1, 3, 2])]; tensor concat_7 = const()[name = tensor("concat_7"), val = tensor([5, 256, 9])]; - tensor transpose_1 = transpose(perm = transpose_1_perm_0, x = var_1500)[name = tensor("transpose_2")]; + tensor transpose_1 = transpose(perm = transpose_1_perm_0, x = var_1543)[name = tensor("transpose_2")]; tensor reshape_1 = reshape(shape = concat_7, x = transpose_1)[name = tensor("reshape_1")]; tensor matmul_2_transpose_x_0 = const()[name = tensor("matmul_2_transpose_x_0"), val = tensor(false)]; tensor matmul_2_transpose_y_0 = const()[name = tensor("matmul_2_transpose_y_0"), val = tensor(false)]; @@ -1357,10 +1376,10 @@ program(1.0) tensor output_end_0 = const()[name = tensor("output_end_0"), val = tensor([1, 5, 8])]; tensor output_end_mask_0 = const()[name = tensor("output_end_mask_0"), val = tensor([true, true, false])]; tensor output = slice_by_index(begin = output_begin_0, end = output_end_0, end_mask = output_end_mask_0, x = reshape_2)[name = tensor("output")]; - tensor probs = sigmoid(x = output)[name = tensor("op_1504")]; - tensor var_1506_axis_0 = const()[name = tensor("op_1506_axis_0"), val = tensor(0)]; - tensor dec_kv_new = stack(axis = var_1506_axis_0, values = (var_1202, nkv))[name = tensor("op_1506")]; - tensor var_1508_axis_0 = const()[name = tensor("op_1508_axis_0"), val = tensor(0)]; - tensor dec_scale_new = stack(axis = var_1508_axis_0, values = (new_scale_9, new_scale))[name = tensor("op_1508")]; + tensor probs = sigmoid(x = output)[name = tensor("op_1547")]; + tensor var_1549_axis_0 = const()[name = tensor("op_1549_axis_0"), val = tensor(0)]; + tensor dec_kv_new = stack(axis = var_1549_axis_0, values = (var_1245, nkv))[name = tensor("op_1549")]; + tensor var_1551_axis_0 = const()[name = tensor("op_1551_axis_0"), val = tensor(0)]; + tensor dec_scale_new = stack(axis = var_1551_axis_0, values = (new_scale_9, new_scale))[name = tensor("op_1551")]; } -> (probs, enc_kv_new, enc_scale_new, enc_conv_cache_new, cnn_window_new, dec_kv_new, dec_scale_new); } \ No newline at end of file diff --git a/optimized/ch/500ms/ls_eend_ch_500ms.mlpackage/Data/com.apple.CoreML/model.mlmodel b/optimized/ch/500ms/ls_eend_ch_500ms.mlpackage/Data/com.apple.CoreML/model.mlmodel index e15cb45b1297e6fc6556ae5ba31a2b881b5eb03b..1b93c691eb7ab0a9b93c9f62a746822f7ce0e698 100644 --- a/optimized/ch/500ms/ls_eend_ch_500ms.mlpackage/Data/com.apple.CoreML/model.mlmodel +++ b/optimized/ch/500ms/ls_eend_ch_500ms.mlpackage/Data/com.apple.CoreML/model.mlmodel @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:32b8d373605b82b210571b934838ef9e0ede35dc97c852d616acea69d2d1c22c -size 196620 +oid sha256:9faeb4f3c5526ac70955a9cbc1da548053b2fdeeda49edad279bca5b0543d453 +size 203220 diff --git a/optimized/ch/500ms/ls_eend_ch_500ms.mlpackage/Manifest.json b/optimized/ch/500ms/ls_eend_ch_500ms.mlpackage/Manifest.json index 7b1a12bb9811d9e20f8485718da28926f0fe3cd1..880c068fe5e42be2c295514041d528d410ce8114 100644 --- a/optimized/ch/500ms/ls_eend_ch_500ms.mlpackage/Manifest.json +++ b/optimized/ch/500ms/ls_eend_ch_500ms.mlpackage/Manifest.json @@ -1,18 +1,18 @@ { "fileFormatVersion": "1.0.0", "itemInfoEntries": { - "B3E5EFE2-8BD2-41C3-B202-D931C2CCA18E": { + "22DDF204-0B49-4CB0-A43C-85A28EC40D19": { "author": "com.apple.CoreML", "description": "CoreML Model Weights", "name": "weights", "path": "com.apple.CoreML/weights" }, - "E2319A52-06A1-4D52-8B9C-D3854ED1D555": { + "8D5435E8-36A0-406C-8F2A-A64DE6AB7ADA": { "author": "com.apple.CoreML", "description": "CoreML Model Specification", "name": "model.mlmodel", "path": "com.apple.CoreML/model.mlmodel" } }, - "rootModelIdentifier": "E2319A52-06A1-4D52-8B9C-D3854ED1D555" + "rootModelIdentifier": "8D5435E8-36A0-406C-8F2A-A64DE6AB7ADA" } diff --git a/optimized/dih2/100ms/ls_eend_dih2_100ms.mlmodelc/analytics/coremldata.bin b/optimized/dih2/100ms/ls_eend_dih2_100ms.mlmodelc/analytics/coremldata.bin index e6a49061db71bd6609e70b767cc7b7d63298f84f..08b08077481b93da48cc59029ff2a55c5e492230 100644 --- a/optimized/dih2/100ms/ls_eend_dih2_100ms.mlmodelc/analytics/coremldata.bin +++ b/optimized/dih2/100ms/ls_eend_dih2_100ms.mlmodelc/analytics/coremldata.bin @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:a4b9f84d7ea9b63709ef0b529a4154d7c460335b9b0ff8de63b48ab92f0e2e2c +oid sha256:bbc1211c80f5d2d6a66d9a9a8588f469eb038399ffa9625b2bd5e7675d88e026 size 243 diff --git a/optimized/dih2/100ms/ls_eend_dih2_100ms.mlmodelc/coremldata.bin b/optimized/dih2/100ms/ls_eend_dih2_100ms.mlmodelc/coremldata.bin index 2f3d2e94407edd8a732a883b6a075a76fb034d65..187a1081ee38d9af1585849e48ffcc03dce7a55e 100644 --- a/optimized/dih2/100ms/ls_eend_dih2_100ms.mlmodelc/coremldata.bin +++ b/optimized/dih2/100ms/ls_eend_dih2_100ms.mlmodelc/coremldata.bin @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:7b761b596de7a3abcc362d422df9b4f2626ec351b373ef5652bc478cea8e6568 -size 1308 +oid sha256:c08015c27fd3eb10b72b9dcf0cc5426bbc510677b6c471695746401352d58f94 +size 1411 diff --git a/optimized/dih2/100ms/ls_eend_dih2_100ms.mlmodelc/metadata.json b/optimized/dih2/100ms/ls_eend_dih2_100ms.mlmodelc/metadata.json index 36071321ca497cfeb4a01fcebd77a489dc4f5a14..1f8297ee0f85446fd7dff2b19281eaa7f36166e9 100644 --- a/optimized/dih2/100ms/ls_eend_dih2_100ms.mlmodelc/metadata.json +++ b/optimized/dih2/100ms/ls_eend_dih2_100ms.mlmodelc/metadata.json @@ -1,7 +1,7 @@ [ { "metadataOutputVersion" : "3.0", - "shortDescription" : "LS-EEND DIHARD II streaming diarizer (pipeline, T=1, max_speakers=10)", + "shortDescription" : "LS-EEND DIHARD II streaming diarizer (pipeline, T=1, max_speakers=10, layout=raw_mel, cu=all)", "outputSchema" : [ { "hasShapeFlexibility" : "0", @@ -81,12 +81,12 @@ "specificationVersion" : 8, "mlProgramOperationTypeHistogram" : { "Ios16.reduceL2Norm" : 2, - "Ios17.reshape" : 66, + "Ios17.reshape" : 67, "Ios16.softmax" : 2, "Ios17.matmul" : 29, "Ios17.transpose" : 57, "Split" : 4, - "Ios17.expandDims" : 3, + "Ios17.expandDims" : 4, "Ios17.add" : 46, "Ios16.sigmoid" : 5, "Ios17.sliceByIndex" : 36, @@ -128,9 +128,9 @@ "hasShapeFlexibility" : "0", "isOptional" : "0", "dataType" : "Float32", - "formattedType" : "MultiArray (Float32 1 × 1 × 345)", + "formattedType" : "MultiArray (Float32 1 × 15 × 23)", "shortDescription" : "", - "shape" : "[1, 1, 345]", + "shape" : "[1, 15, 23]", "name" : "features", "type" : "MultiArray" }, @@ -206,8 +206,8 @@ } ], "userDefinedMetadata" : { - "com.github.apple.coremltools.conversion_date" : "2026-04-16", - "config" : "{\"model_name\": \"dih2\", \"model_label\": \"DIHARD II\", \"variant\": \"pipeline\", \"chunk_size\": 1, \"step_duration_ms\": 100, \"frame_duration_ms\": 100, \"frame_duration_seconds\": 0.1, \"max_speakers\": 10, \"max_nspks\": 12, \"n_units\": 256, \"n_heads\": 4, \"enc_n_layers\": 4, \"dec_n_layers\": 2, \"conv_kernel_size\": 16, \"conv_delay\": 9, \"sample_rate\": 8000, \"win_length\": 200, \"hop_length\": 80, \"n_mels\": 23, \"context_size\": 7, \"subsampling\": 10, \"feat_type\": \"logmel23_cummn\", \"pure_roll\": true}", + "com.github.apple.coremltools.conversion_date" : "2026-04-18", + "config" : "{\"model_name\": \"dih2\", \"model_label\": \"DIHARD II\", \"variant\": \"pipeline\", \"chunk_size\": 1, \"step_duration_ms\": 100, \"frame_duration_ms\": 100, \"frame_duration_seconds\": 0.1, \"max_speakers\": 10, \"max_nspks\": 12, \"n_units\": 256, \"n_heads\": 4, \"enc_n_layers\": 4, \"dec_n_layers\": 2, \"conv_kernel_size\": 16, \"conv_delay\": 9, \"sample_rate\": 8000, \"win_length\": 200, \"hop_length\": 80, \"n_mels\": 23, \"context_size\": 7, \"subsampling\": 10, \"feat_type\": \"logmel23_cummn\", \"pure_roll\": true, \"input_layout\": \"raw_mel\", \"compute_units_export\": \"all\", \"raw_mel_length\": 15}", "com.github.apple.coremltools.source" : "torch==2.6.0", "com.github.apple.coremltools.version" : "9.0", "com.github.apple.coremltools.source_dialect" : "TorchScript" diff --git a/optimized/dih2/100ms/ls_eend_dih2_100ms.mlmodelc/model.mil b/optimized/dih2/100ms/ls_eend_dih2_100ms.mlmodelc/model.mil index 182f88767b6a7156fba8d5e17cc71543704b1b0a..4bc2f36465c8a6b77c6ec5021d7ddf40dec1643b 100644 --- a/optimized/dih2/100ms/ls_eend_dih2_100ms.mlmodelc/model.mil +++ b/optimized/dih2/100ms/ls_eend_dih2_100ms.mlmodelc/model.mil @@ -1,233 +1,239 @@ program(1.0) [buildInfo = dict, tensor>({{"coremlc-component-MIL", "3510.2.1"}, {"coremlc-version", "3500.32.1"}, {"coremltools-component-torch", "2.6.0"}, {"coremltools-source-dialect", "TorchScript"}, {"coremltools-version", "9.0"}})] { - func main(tensor cnn_window, tensor dec_kv, tensor dec_scale, tensor enc_conv_cache, tensor enc_kv, tensor enc_scale, tensor features, tensor valid_mask) { - tensor encoder_cnn_bias = const()[name = tensor("encoder_cnn_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(64)))]; - tensor encoder_cnn_weight = const()[name = tensor("encoder_cnn_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1152)))]; - tensor encoder__causal_mask = const()[name = tensor("encoder__causal_mask"), val = tensor([[0x1p+0]])]; - tensor encoder__t_index = const()[name = tensor("encoder__t_index"), val = tensor([0x1p+0])]; - tensor encoder_input_projection_linear_bias = const()[name = tensor("encoder_input_projection_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4981952)))]; - tensor encoder_input_projection_linear_weight = const()[name = tensor("encoder_input_projection_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4983040)))]; - tensor encoder_pre_layer_norm_bias = const()[name = tensor("encoder_pre_layer_norm_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(5336384)))]; - tensor encoder_pre_layer_norm_weight = const()[name = tensor("encoder_pre_layer_norm_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(5337472)))]; - tensor encoder_ffn1_0_module_sequential_0_bias = const()[name = tensor("encoder_ffn1_0_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(5338560)))]; - tensor encoder_ffn1_0_module_sequential_0_weight = const()[name = tensor("encoder_ffn1_0_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(5339648)))]; - tensor encoder_ffn1_0_module_sequential_1_linear_bias = const()[name = tensor("encoder_ffn1_0_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(5340736)))]; - tensor encoder_ffn1_0_module_sequential_1_linear_weight = const()[name = tensor("encoder_ffn1_0_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(5344896)))]; - tensor encoder_ffn1_0_module_sequential_4_linear_bias = const()[name = tensor("encoder_ffn1_0_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(6393536)))]; - tensor encoder_ffn1_0_module_sequential_4_linear_weight = const()[name = tensor("encoder_ffn1_0_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(6394624)))]; - tensor encoder_ret_lns_0_bias = const()[name = tensor("encoder_ret_lns_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7443264)))]; - tensor encoder_ret_lns_0_weight = const()[name = tensor("encoder_ret_lns_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7444352)))]; - tensor encoder_q_proj_0_bias = const()[name = tensor("encoder_q_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7445440)))]; - tensor encoder_q_proj_0_weight = const()[name = tensor("encoder_q_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7446528)))]; - tensor encoder_k_proj_0_bias = const()[name = tensor("encoder_k_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7708736)))]; - tensor encoder_k_proj_0_weight = const()[name = tensor("encoder_k_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7709824)))]; - tensor encoder_v_proj_0_bias = const()[name = tensor("encoder_v_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7972032)))]; - tensor encoder_v_proj_0_weight = const()[name = tensor("encoder_v_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7973120)))]; - tensor encoder_g_proj_0_bias = const()[name = tensor("encoder_g_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8235328)))]; - tensor encoder_g_proj_0_weight = const()[name = tensor("encoder_g_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8236416)))]; - tensor encoder_out_proj_0_bias = const()[name = tensor("encoder_out_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8498624)))]; - tensor encoder_out_proj_0_weight = const()[name = tensor("encoder_out_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8499712)))]; - tensor encoder_conv_module_0_sequential_0_bias = const()[name = tensor("encoder_conv_module_0_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8761920)))]; - tensor encoder_conv_module_0_sequential_0_weight = const()[name = tensor("encoder_conv_module_0_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8763008)))]; - tensor encoder_conv_module_0_sequential_2_conv_bias = const()[name = tensor("encoder_conv_module_0_sequential_2_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8764096)))]; - tensor encoder_conv_module_0_sequential_2_conv_weight = const()[name = tensor("encoder_conv_module_0_sequential_2_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8766208)))]; - tensor encoder_conv_module_0_sequential_4_conv_weight = const()[name = tensor("encoder_conv_module_0_sequential_4_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9290560)))]; - tensor encoder_conv_module_0_sequential_5_running_var = const()[name = tensor("encoder_conv_module_0_sequential_5_running_var"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9307008)))]; - tensor encoder_conv_module_0_sequential_5_running_mean = const()[name = tensor("encoder_conv_module_0_sequential_5_running_mean"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9308096)))]; - tensor encoder_conv_module_0_sequential_5_bias = const()[name = tensor("encoder_conv_module_0_sequential_5_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9309184)))]; - tensor encoder_conv_module_0_sequential_5_weight = const()[name = tensor("encoder_conv_module_0_sequential_5_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9310272)))]; - tensor encoder_conv_module_0_sequential_7_conv_bias = const()[name = tensor("encoder_conv_module_0_sequential_7_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9311360)))]; - tensor encoder_conv_module_0_sequential_7_conv_weight = const()[name = tensor("encoder_conv_module_0_sequential_7_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9312448)))]; - tensor encoder_ffn2_0_module_sequential_0_bias = const()[name = tensor("encoder_ffn2_0_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9574656)))]; - tensor encoder_ffn2_0_module_sequential_0_weight = const()[name = tensor("encoder_ffn2_0_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9575744)))]; - tensor encoder_ffn2_0_module_sequential_1_linear_bias = const()[name = tensor("encoder_ffn2_0_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9576832)))]; - tensor encoder_ffn2_0_module_sequential_1_linear_weight = const()[name = tensor("encoder_ffn2_0_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9580992)))]; - tensor encoder_ffn2_0_module_sequential_4_linear_bias = const()[name = tensor("encoder_ffn2_0_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(10629632)))]; - tensor encoder_ffn2_0_module_sequential_4_linear_weight = const()[name = tensor("encoder_ffn2_0_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(10630720)))]; - tensor encoder_layer_norm_0_bias = const()[name = tensor("encoder_layer_norm_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11679360)))]; - tensor encoder_layer_norm_0_weight = const()[name = tensor("encoder_layer_norm_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11680448)))]; - tensor encoder_ffn1_1_module_sequential_0_bias = const()[name = tensor("encoder_ffn1_1_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11681536)))]; - tensor encoder_ffn1_1_module_sequential_0_weight = const()[name = tensor("encoder_ffn1_1_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11682624)))]; - tensor encoder_ffn1_1_module_sequential_1_linear_bias = const()[name = tensor("encoder_ffn1_1_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11683712)))]; - tensor encoder_ffn1_1_module_sequential_1_linear_weight = const()[name = tensor("encoder_ffn1_1_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11687872)))]; - tensor encoder_ffn1_1_module_sequential_4_linear_bias = const()[name = tensor("encoder_ffn1_1_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(12736512)))]; - tensor encoder_ffn1_1_module_sequential_4_linear_weight = const()[name = tensor("encoder_ffn1_1_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(12737600)))]; - tensor encoder_ret_lns_1_bias = const()[name = tensor("encoder_ret_lns_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(13786240)))]; - tensor encoder_ret_lns_1_weight = const()[name = tensor("encoder_ret_lns_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(13787328)))]; - tensor encoder_q_proj_1_bias = const()[name = tensor("encoder_q_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(13788416)))]; - tensor encoder_q_proj_1_weight = const()[name = tensor("encoder_q_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(13789504)))]; - tensor encoder_k_proj_1_bias = const()[name = tensor("encoder_k_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14051712)))]; - tensor encoder_k_proj_1_weight = const()[name = tensor("encoder_k_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14052800)))]; - tensor encoder_v_proj_1_bias = const()[name = tensor("encoder_v_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14315008)))]; - tensor encoder_v_proj_1_weight = const()[name = tensor("encoder_v_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14316096)))]; - tensor encoder_g_proj_1_bias = const()[name = tensor("encoder_g_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14578304)))]; - tensor encoder_g_proj_1_weight = const()[name = tensor("encoder_g_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14579392)))]; - tensor encoder_out_proj_1_bias = const()[name = tensor("encoder_out_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14841600)))]; - tensor encoder_out_proj_1_weight = const()[name = tensor("encoder_out_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14842688)))]; - tensor encoder_conv_module_1_sequential_0_bias = const()[name = tensor("encoder_conv_module_1_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15104896)))]; - tensor encoder_conv_module_1_sequential_0_weight = const()[name = tensor("encoder_conv_module_1_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15105984)))]; - tensor encoder_conv_module_1_sequential_2_conv_bias = const()[name = tensor("encoder_conv_module_1_sequential_2_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15107072)))]; - tensor encoder_conv_module_1_sequential_2_conv_weight = const()[name = tensor("encoder_conv_module_1_sequential_2_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15109184)))]; - tensor encoder_conv_module_1_sequential_4_conv_weight = const()[name = tensor("encoder_conv_module_1_sequential_4_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15633536)))]; - tensor encoder_conv_module_1_sequential_5_running_var = const()[name = tensor("encoder_conv_module_1_sequential_5_running_var"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15649984)))]; - tensor encoder_conv_module_1_sequential_5_running_mean = const()[name = tensor("encoder_conv_module_1_sequential_5_running_mean"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15651072)))]; - tensor encoder_conv_module_1_sequential_5_bias = const()[name = tensor("encoder_conv_module_1_sequential_5_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15652160)))]; - tensor encoder_conv_module_1_sequential_5_weight = const()[name = tensor("encoder_conv_module_1_sequential_5_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15653248)))]; - tensor encoder_conv_module_1_sequential_7_conv_bias = const()[name = tensor("encoder_conv_module_1_sequential_7_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15654336)))]; - tensor encoder_conv_module_1_sequential_7_conv_weight = const()[name = tensor("encoder_conv_module_1_sequential_7_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15655424)))]; - tensor encoder_ffn2_1_module_sequential_0_bias = const()[name = tensor("encoder_ffn2_1_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15917632)))]; - tensor encoder_ffn2_1_module_sequential_0_weight = const()[name = tensor("encoder_ffn2_1_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15918720)))]; - tensor encoder_ffn2_1_module_sequential_1_linear_bias = const()[name = tensor("encoder_ffn2_1_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15919808)))]; - tensor encoder_ffn2_1_module_sequential_1_linear_weight = const()[name = tensor("encoder_ffn2_1_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15923968)))]; - tensor encoder_ffn2_1_module_sequential_4_linear_bias = const()[name = tensor("encoder_ffn2_1_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(16972608)))]; - tensor encoder_ffn2_1_module_sequential_4_linear_weight = const()[name = tensor("encoder_ffn2_1_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(16973696)))]; - tensor encoder_layer_norm_1_bias = const()[name = tensor("encoder_layer_norm_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18022336)))]; - tensor encoder_layer_norm_1_weight = const()[name = tensor("encoder_layer_norm_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18023424)))]; - tensor encoder_ffn1_2_module_sequential_0_bias = const()[name = tensor("encoder_ffn1_2_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18024512)))]; - tensor encoder_ffn1_2_module_sequential_0_weight = const()[name = tensor("encoder_ffn1_2_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18025600)))]; - tensor encoder_ffn1_2_module_sequential_1_linear_bias = const()[name = tensor("encoder_ffn1_2_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18026688)))]; - tensor encoder_ffn1_2_module_sequential_1_linear_weight = const()[name = tensor("encoder_ffn1_2_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18030848)))]; - tensor encoder_ffn1_2_module_sequential_4_linear_bias = const()[name = tensor("encoder_ffn1_2_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(19079488)))]; - tensor encoder_ffn1_2_module_sequential_4_linear_weight = const()[name = tensor("encoder_ffn1_2_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(19080576)))]; - tensor encoder_ret_lns_2_bias = const()[name = tensor("encoder_ret_lns_2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20129216)))]; - tensor encoder_ret_lns_2_weight = const()[name = tensor("encoder_ret_lns_2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20130304)))]; - tensor encoder_q_proj_2_bias = const()[name = tensor("encoder_q_proj_2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20131392)))]; - tensor encoder_q_proj_2_weight = const()[name = tensor("encoder_q_proj_2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20132480)))]; - tensor encoder_k_proj_2_bias = const()[name = tensor("encoder_k_proj_2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20394688)))]; - tensor encoder_k_proj_2_weight = const()[name = tensor("encoder_k_proj_2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20395776)))]; - tensor encoder_v_proj_2_bias = const()[name = tensor("encoder_v_proj_2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20657984)))]; - tensor encoder_v_proj_2_weight = const()[name = tensor("encoder_v_proj_2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20659072)))]; - tensor encoder_g_proj_2_bias = const()[name = tensor("encoder_g_proj_2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20921280)))]; - tensor encoder_g_proj_2_weight = const()[name = tensor("encoder_g_proj_2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20922368)))]; - tensor encoder_out_proj_2_bias = const()[name = tensor("encoder_out_proj_2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21184576)))]; - tensor encoder_out_proj_2_weight = const()[name = tensor("encoder_out_proj_2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21185664)))]; - tensor encoder_conv_module_2_sequential_0_bias = const()[name = tensor("encoder_conv_module_2_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21447872)))]; - tensor encoder_conv_module_2_sequential_0_weight = const()[name = tensor("encoder_conv_module_2_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21448960)))]; - tensor encoder_conv_module_2_sequential_2_conv_bias = const()[name = tensor("encoder_conv_module_2_sequential_2_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21450048)))]; - tensor encoder_conv_module_2_sequential_2_conv_weight = const()[name = tensor("encoder_conv_module_2_sequential_2_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21452160)))]; - tensor encoder_conv_module_2_sequential_4_conv_weight = const()[name = tensor("encoder_conv_module_2_sequential_4_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21976512)))]; - tensor encoder_conv_module_2_sequential_5_running_var = const()[name = tensor("encoder_conv_module_2_sequential_5_running_var"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21992960)))]; - tensor encoder_conv_module_2_sequential_5_running_mean = const()[name = tensor("encoder_conv_module_2_sequential_5_running_mean"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21994048)))]; - tensor encoder_conv_module_2_sequential_5_bias = const()[name = tensor("encoder_conv_module_2_sequential_5_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21995136)))]; - tensor encoder_conv_module_2_sequential_5_weight = const()[name = tensor("encoder_conv_module_2_sequential_5_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21996224)))]; - tensor encoder_conv_module_2_sequential_7_conv_bias = const()[name = tensor("encoder_conv_module_2_sequential_7_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21997312)))]; - tensor encoder_conv_module_2_sequential_7_conv_weight = const()[name = tensor("encoder_conv_module_2_sequential_7_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21998400)))]; - tensor encoder_ffn2_2_module_sequential_0_bias = const()[name = tensor("encoder_ffn2_2_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(22260608)))]; - tensor encoder_ffn2_2_module_sequential_0_weight = const()[name = tensor("encoder_ffn2_2_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(22261696)))]; - tensor encoder_ffn2_2_module_sequential_1_linear_bias = const()[name = tensor("encoder_ffn2_2_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(22262784)))]; - tensor encoder_ffn2_2_module_sequential_1_linear_weight = const()[name = tensor("encoder_ffn2_2_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(22266944)))]; - tensor encoder_ffn2_2_module_sequential_4_linear_bias = const()[name = tensor("encoder_ffn2_2_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(23315584)))]; - tensor encoder_ffn2_2_module_sequential_4_linear_weight = const()[name = tensor("encoder_ffn2_2_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(23316672)))]; - tensor encoder_layer_norm_2_bias = const()[name = tensor("encoder_layer_norm_2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(24365312)))]; - tensor encoder_layer_norm_2_weight = const()[name = tensor("encoder_layer_norm_2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(24366400)))]; - tensor encoder_ffn1_3_module_sequential_0_bias = const()[name = tensor("encoder_ffn1_3_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(24367488)))]; - tensor encoder_ffn1_3_module_sequential_0_weight = const()[name = tensor("encoder_ffn1_3_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(24368576)))]; - tensor encoder_ffn1_3_module_sequential_1_linear_bias = const()[name = tensor("encoder_ffn1_3_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(24369664)))]; - tensor encoder_ffn1_3_module_sequential_1_linear_weight = const()[name = tensor("encoder_ffn1_3_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(24373824)))]; - tensor encoder_ffn1_3_module_sequential_4_linear_bias = const()[name = tensor("encoder_ffn1_3_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(25422464)))]; - tensor encoder_ffn1_3_module_sequential_4_linear_weight = const()[name = tensor("encoder_ffn1_3_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(25423552)))]; - tensor encoder_ret_lns_3_bias = const()[name = tensor("encoder_ret_lns_3_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(26472192)))]; - tensor encoder_ret_lns_3_weight = const()[name = tensor("encoder_ret_lns_3_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(26473280)))]; - tensor encoder_q_proj_3_bias = const()[name = tensor("encoder_q_proj_3_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(26474368)))]; - tensor encoder_q_proj_3_weight = const()[name = tensor("encoder_q_proj_3_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(26475456)))]; - tensor encoder_k_proj_3_bias = const()[name = tensor("encoder_k_proj_3_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(26737664)))]; - tensor encoder_k_proj_3_weight = const()[name = tensor("encoder_k_proj_3_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(26738752)))]; - tensor encoder_v_proj_3_bias = const()[name = tensor("encoder_v_proj_3_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27000960)))]; - tensor encoder_v_proj_3_weight = const()[name = tensor("encoder_v_proj_3_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27002048)))]; - tensor encoder_g_proj_3_bias = const()[name = tensor("encoder_g_proj_3_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27264256)))]; - tensor encoder_g_proj_3_weight = const()[name = tensor("encoder_g_proj_3_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27265344)))]; - tensor encoder_out_proj_3_bias = const()[name = tensor("encoder_out_proj_3_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27527552)))]; - tensor encoder_out_proj_3_weight = const()[name = tensor("encoder_out_proj_3_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27528640)))]; - tensor encoder_conv_module_3_sequential_0_bias = const()[name = tensor("encoder_conv_module_3_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27790848)))]; - tensor encoder_conv_module_3_sequential_0_weight = const()[name = tensor("encoder_conv_module_3_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27791936)))]; - tensor encoder_conv_module_3_sequential_2_conv_bias = const()[name = tensor("encoder_conv_module_3_sequential_2_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27793024)))]; - tensor encoder_conv_module_3_sequential_2_conv_weight = const()[name = tensor("encoder_conv_module_3_sequential_2_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27795136)))]; - tensor encoder_conv_module_3_sequential_4_conv_weight = const()[name = tensor("encoder_conv_module_3_sequential_4_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28319488)))]; - tensor encoder_conv_module_3_sequential_5_running_var = const()[name = tensor("encoder_conv_module_3_sequential_5_running_var"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28335936)))]; - tensor encoder_conv_module_3_sequential_5_running_mean = const()[name = tensor("encoder_conv_module_3_sequential_5_running_mean"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28337024)))]; - tensor encoder_conv_module_3_sequential_5_bias = const()[name = tensor("encoder_conv_module_3_sequential_5_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28338112)))]; - tensor encoder_conv_module_3_sequential_5_weight = const()[name = tensor("encoder_conv_module_3_sequential_5_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28339200)))]; - tensor encoder_conv_module_3_sequential_7_conv_bias = const()[name = tensor("encoder_conv_module_3_sequential_7_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28340288)))]; - tensor encoder_conv_module_3_sequential_7_conv_weight = const()[name = tensor("encoder_conv_module_3_sequential_7_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28341376)))]; - tensor encoder_ffn2_3_module_sequential_0_bias = const()[name = tensor("encoder_ffn2_3_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28603584)))]; - tensor encoder_ffn2_3_module_sequential_0_weight = const()[name = tensor("encoder_ffn2_3_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28604672)))]; - tensor encoder_ffn2_3_module_sequential_1_linear_bias = const()[name = tensor("encoder_ffn2_3_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28605760)))]; - tensor encoder_ffn2_3_module_sequential_1_linear_weight = const()[name = tensor("encoder_ffn2_3_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28609920)))]; - tensor encoder_ffn2_3_module_sequential_4_linear_bias = const()[name = tensor("encoder_ffn2_3_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(29658560)))]; - tensor encoder_ffn2_3_module_sequential_4_linear_weight = const()[name = tensor("encoder_ffn2_3_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(29659648)))]; - tensor encoder_layer_norm_3_bias = const()[name = tensor("encoder_layer_norm_3_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(30708288)))]; - tensor encoder_layer_norm_3_weight = const()[name = tensor("encoder_layer_norm_3_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(30709376)))]; - tensor decoder_convert_bias = const()[name = tensor("decoder_convert_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(30710464)))]; - tensor decoder_convert_weight = const()[name = tensor("decoder_convert_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(30711552)))]; - tensor decoder_q_proj_0_bias = const()[name = tensor("decoder_q_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(31235904)))]; - tensor decoder_q_proj_0_weight = const()[name = tensor("decoder_q_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(31236992)))]; - tensor decoder_k_proj_0_bias = const()[name = tensor("decoder_k_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(31499200)))]; - tensor decoder_k_proj_0_weight = const()[name = tensor("decoder_k_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(31500288)))]; - tensor decoder_v_proj_0_bias = const()[name = tensor("decoder_v_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(31762496)))]; - tensor decoder_v_proj_0_weight = const()[name = tensor("decoder_v_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(31763584)))]; - tensor decoder_g_proj_0_bias = const()[name = tensor("decoder_g_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32025792)))]; - tensor decoder_g_proj_0_weight = const()[name = tensor("decoder_g_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32026880)))]; - tensor decoder_out_proj_0_bias = const()[name = tensor("decoder_out_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32289088)))]; - tensor decoder_out_proj_0_weight = const()[name = tensor("decoder_out_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32290176)))]; - tensor decoder_norm11_0_bias = const()[name = tensor("decoder_norm11_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32552384)))]; - tensor decoder_norm11_0_weight = const()[name = tensor("decoder_norm11_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32553472)))]; - tensor decoder_self_attn2_0_out_proj_bias = const()[name = tensor("decoder_self_attn2_0_out_proj_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32554560)))]; - tensor decoder_self_attn2_0_out_proj_weight = const()[name = tensor("decoder_self_attn2_0_out_proj_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32555648)))]; - tensor decoder_self_attn2_0_in_proj_bias = const()[name = tensor("decoder_self_attn2_0_in_proj_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32817856)))]; - tensor decoder_self_attn2_0_in_proj_weight = const()[name = tensor("decoder_self_attn2_0_in_proj_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32820992)))]; - tensor decoder_norm21_0_bias = const()[name = tensor("decoder_norm21_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(33607488)))]; - tensor decoder_norm21_0_weight = const()[name = tensor("decoder_norm21_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(33608576)))]; - tensor decoder_linear1_0_bias = const()[name = tensor("decoder_linear1_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(33609664)))]; - tensor decoder_linear1_0_weight = const()[name = tensor("decoder_linear1_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(33617920)))]; - tensor decoder_linear2_0_bias = const()[name = tensor("decoder_linear2_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(35715136)))]; - tensor decoder_linear2_0_weight = const()[name = tensor("decoder_linear2_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(35716224)))]; - tensor decoder_norm22_0_bias = const()[name = tensor("decoder_norm22_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(37813440)))]; - tensor decoder_norm22_0_weight = const()[name = tensor("decoder_norm22_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(37814528)))]; - tensor decoder_q_proj_1_bias = const()[name = tensor("decoder_q_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(37815616)))]; - tensor decoder_q_proj_1_weight = const()[name = tensor("decoder_q_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(37816704)))]; - tensor decoder_k_proj_1_bias = const()[name = tensor("decoder_k_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38078912)))]; - tensor decoder_k_proj_1_weight = const()[name = tensor("decoder_k_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38080000)))]; - tensor decoder_v_proj_1_bias = const()[name = tensor("decoder_v_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38342208)))]; - tensor decoder_v_proj_1_weight = const()[name = tensor("decoder_v_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38343296)))]; - tensor decoder_g_proj_1_bias = const()[name = tensor("decoder_g_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38605504)))]; - tensor decoder_g_proj_1_weight = const()[name = tensor("decoder_g_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38606592)))]; - tensor decoder_out_proj_1_bias = const()[name = tensor("decoder_out_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38868800)))]; - tensor decoder_out_proj_1_weight = const()[name = tensor("decoder_out_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38869888)))]; - tensor decoder_norm11_1_bias = const()[name = tensor("decoder_norm11_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(39132096)))]; - tensor decoder_norm11_1_weight = const()[name = tensor("decoder_norm11_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(39133184)))]; - tensor decoder_self_attn2_1_out_proj_bias = const()[name = tensor("decoder_self_attn2_1_out_proj_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(39134272)))]; - tensor decoder_self_attn2_1_out_proj_weight = const()[name = tensor("decoder_self_attn2_1_out_proj_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(39135360)))]; - tensor decoder_self_attn2_1_in_proj_bias = const()[name = tensor("decoder_self_attn2_1_in_proj_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(39397568)))]; - tensor decoder_self_attn2_1_in_proj_weight = const()[name = tensor("decoder_self_attn2_1_in_proj_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(39400704)))]; - tensor decoder_norm21_1_bias = const()[name = tensor("decoder_norm21_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(40187200)))]; - tensor decoder_norm21_1_weight = const()[name = tensor("decoder_norm21_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(40188288)))]; - tensor decoder_linear1_1_bias = const()[name = tensor("decoder_linear1_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(40189376)))]; - tensor decoder_linear1_1_weight = const()[name = tensor("decoder_linear1_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(40197632)))]; - tensor decoder_linear2_1_bias = const()[name = tensor("decoder_linear2_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(42294848)))]; - tensor decoder_linear2_1_weight = const()[name = tensor("decoder_linear2_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(42295936)))]; - tensor decoder_norm22_1_bias = const()[name = tensor("decoder_norm22_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(44393152)))]; - tensor decoder_norm22_1_weight = const()[name = tensor("decoder_norm22_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(44394240)))]; - tensor var_11 = const()[name = tensor("op_11"), val = tensor(0x1.5798eep-27)]; - tensor var_17 = const()[name = tensor("op_17"), val = tensor(0x1.0c6f7ap-20)]; - tensor var_20 = const()[name = tensor("op_20"), val = tensor(2)]; - tensor var_23 = const()[name = tensor("op_23"), val = tensor(0)]; - tensor var_26 = const()[name = tensor("op_26"), val = tensor(1)]; - tensor var_28 = const()[name = tensor("op_28"), val = tensor(0x1.4f8b58p-17)]; - tensor var_29 = const()[name = tensor("op_29"), val = tensor(true)]; - tensor input_1 = linear(bias = encoder_input_projection_linear_bias, weight = encoder_input_projection_linear_weight, x = features)[name = tensor("linear_0")]; - tensor input_3_axes_0 = const()[name = tensor("input_3_axes_0"), val = tensor([-1])]; - tensor input_3 = layer_norm(axes = input_3_axes_0, beta = encoder_pre_layer_norm_bias, epsilon = var_28, gamma = encoder_pre_layer_norm_weight, x = input_1)[name = tensor("input_3")]; + func main(tensor cnn_window, tensor dec_kv, tensor dec_scale, tensor enc_conv_cache, tensor enc_kv, tensor enc_scale, tensor features, tensor valid_mask) { + tensor inner_encoder_cnn_bias = const()[name = tensor("inner_encoder_cnn_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(64)))]; + tensor inner_encoder_cnn_weight = const()[name = tensor("inner_encoder_cnn_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1152)))]; + tensor inner_encoder__causal_mask = const()[name = tensor("inner_encoder__causal_mask"), val = tensor([[0x1p+0]])]; + tensor inner_encoder__t_index = const()[name = tensor("inner_encoder__t_index"), val = tensor([0x1p+0])]; + tensor inner_encoder_input_projection_linear_bias = const()[name = tensor("inner_encoder_input_projection_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4981952)))]; + tensor inner_encoder_input_projection_linear_weight = const()[name = tensor("inner_encoder_input_projection_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4983040)))]; + tensor inner_encoder_pre_layer_norm_bias = const()[name = tensor("inner_encoder_pre_layer_norm_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(5336384)))]; + tensor inner_encoder_pre_layer_norm_weight = const()[name = tensor("inner_encoder_pre_layer_norm_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(5337472)))]; + tensor inner_encoder_ffn1_0_module_sequential_0_bias = const()[name = tensor("inner_encoder_ffn1_0_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(5338560)))]; + tensor inner_encoder_ffn1_0_module_sequential_0_weight = const()[name = tensor("inner_encoder_ffn1_0_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(5339648)))]; + tensor inner_encoder_ffn1_0_module_sequential_1_linear_bias = const()[name = tensor("inner_encoder_ffn1_0_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(5340736)))]; + tensor inner_encoder_ffn1_0_module_sequential_1_linear_weight = const()[name = tensor("inner_encoder_ffn1_0_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(5344896)))]; + tensor inner_encoder_ffn1_0_module_sequential_4_linear_bias = const()[name = tensor("inner_encoder_ffn1_0_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(6393536)))]; + tensor inner_encoder_ffn1_0_module_sequential_4_linear_weight = const()[name = tensor("inner_encoder_ffn1_0_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(6394624)))]; + tensor inner_encoder_ret_lns_0_bias = const()[name = tensor("inner_encoder_ret_lns_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7443264)))]; + tensor inner_encoder_ret_lns_0_weight = const()[name = tensor("inner_encoder_ret_lns_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7444352)))]; + tensor inner_encoder_q_proj_0_bias = const()[name = tensor("inner_encoder_q_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7445440)))]; + tensor inner_encoder_q_proj_0_weight = const()[name = tensor("inner_encoder_q_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7446528)))]; + tensor inner_encoder_k_proj_0_bias = const()[name = tensor("inner_encoder_k_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7708736)))]; + tensor inner_encoder_k_proj_0_weight = const()[name = tensor("inner_encoder_k_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7709824)))]; + tensor inner_encoder_v_proj_0_bias = const()[name = tensor("inner_encoder_v_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7972032)))]; + tensor inner_encoder_v_proj_0_weight = const()[name = tensor("inner_encoder_v_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7973120)))]; + tensor inner_encoder_g_proj_0_bias = const()[name = tensor("inner_encoder_g_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8235328)))]; + tensor inner_encoder_g_proj_0_weight = const()[name = tensor("inner_encoder_g_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8236416)))]; + tensor inner_encoder_out_proj_0_bias = const()[name = tensor("inner_encoder_out_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8498624)))]; + tensor inner_encoder_out_proj_0_weight = const()[name = tensor("inner_encoder_out_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8499712)))]; + tensor inner_encoder_conv_module_0_sequential_0_bias = const()[name = tensor("inner_encoder_conv_module_0_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8761920)))]; + tensor inner_encoder_conv_module_0_sequential_0_weight = const()[name = tensor("inner_encoder_conv_module_0_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8763008)))]; + tensor inner_encoder_conv_module_0_sequential_2_conv_bias = const()[name = tensor("inner_encoder_conv_module_0_sequential_2_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8764096)))]; + tensor inner_encoder_conv_module_0_sequential_2_conv_weight = const()[name = tensor("inner_encoder_conv_module_0_sequential_2_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8766208)))]; + tensor inner_encoder_conv_module_0_sequential_4_conv_weight = const()[name = tensor("inner_encoder_conv_module_0_sequential_4_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9290560)))]; + tensor inner_encoder_conv_module_0_sequential_5_running_var = const()[name = tensor("inner_encoder_conv_module_0_sequential_5_running_var"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9307008)))]; + tensor inner_encoder_conv_module_0_sequential_5_running_mean = const()[name = tensor("inner_encoder_conv_module_0_sequential_5_running_mean"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9308096)))]; + tensor inner_encoder_conv_module_0_sequential_5_bias = const()[name = tensor("inner_encoder_conv_module_0_sequential_5_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9309184)))]; + tensor inner_encoder_conv_module_0_sequential_5_weight = const()[name = tensor("inner_encoder_conv_module_0_sequential_5_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9310272)))]; + tensor inner_encoder_conv_module_0_sequential_7_conv_bias = const()[name = tensor("inner_encoder_conv_module_0_sequential_7_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9311360)))]; + tensor inner_encoder_conv_module_0_sequential_7_conv_weight = const()[name = tensor("inner_encoder_conv_module_0_sequential_7_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9312448)))]; + tensor inner_encoder_ffn2_0_module_sequential_0_bias = const()[name = tensor("inner_encoder_ffn2_0_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9574656)))]; + tensor inner_encoder_ffn2_0_module_sequential_0_weight = const()[name = tensor("inner_encoder_ffn2_0_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9575744)))]; + tensor inner_encoder_ffn2_0_module_sequential_1_linear_bias = const()[name = tensor("inner_encoder_ffn2_0_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9576832)))]; + tensor inner_encoder_ffn2_0_module_sequential_1_linear_weight = const()[name = tensor("inner_encoder_ffn2_0_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9580992)))]; + tensor inner_encoder_ffn2_0_module_sequential_4_linear_bias = const()[name = tensor("inner_encoder_ffn2_0_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(10629632)))]; + tensor inner_encoder_ffn2_0_module_sequential_4_linear_weight = const()[name = tensor("inner_encoder_ffn2_0_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(10630720)))]; + tensor inner_encoder_layer_norm_0_bias = const()[name = tensor("inner_encoder_layer_norm_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11679360)))]; + tensor inner_encoder_layer_norm_0_weight = const()[name = tensor("inner_encoder_layer_norm_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11680448)))]; + tensor inner_encoder_ffn1_1_module_sequential_0_bias = const()[name = tensor("inner_encoder_ffn1_1_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11681536)))]; + tensor inner_encoder_ffn1_1_module_sequential_0_weight = const()[name = tensor("inner_encoder_ffn1_1_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11682624)))]; + tensor inner_encoder_ffn1_1_module_sequential_1_linear_bias = const()[name = tensor("inner_encoder_ffn1_1_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11683712)))]; + tensor inner_encoder_ffn1_1_module_sequential_1_linear_weight = const()[name = tensor("inner_encoder_ffn1_1_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11687872)))]; + tensor inner_encoder_ffn1_1_module_sequential_4_linear_bias = const()[name = tensor("inner_encoder_ffn1_1_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(12736512)))]; + tensor inner_encoder_ffn1_1_module_sequential_4_linear_weight = const()[name = tensor("inner_encoder_ffn1_1_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(12737600)))]; + tensor inner_encoder_ret_lns_1_bias = const()[name = tensor("inner_encoder_ret_lns_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(13786240)))]; + tensor inner_encoder_ret_lns_1_weight = const()[name = tensor("inner_encoder_ret_lns_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(13787328)))]; + tensor inner_encoder_q_proj_1_bias = const()[name = tensor("inner_encoder_q_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(13788416)))]; + tensor inner_encoder_q_proj_1_weight = const()[name = tensor("inner_encoder_q_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(13789504)))]; + tensor inner_encoder_k_proj_1_bias = const()[name = tensor("inner_encoder_k_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14051712)))]; + tensor inner_encoder_k_proj_1_weight = const()[name = tensor("inner_encoder_k_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14052800)))]; + tensor inner_encoder_v_proj_1_bias = const()[name = tensor("inner_encoder_v_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14315008)))]; + tensor inner_encoder_v_proj_1_weight = const()[name = tensor("inner_encoder_v_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14316096)))]; + tensor inner_encoder_g_proj_1_bias = const()[name = tensor("inner_encoder_g_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14578304)))]; + tensor inner_encoder_g_proj_1_weight = const()[name = tensor("inner_encoder_g_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14579392)))]; + tensor inner_encoder_out_proj_1_bias = const()[name = tensor("inner_encoder_out_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14841600)))]; + tensor inner_encoder_out_proj_1_weight = const()[name = tensor("inner_encoder_out_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14842688)))]; + tensor inner_encoder_conv_module_1_sequential_0_bias = const()[name = tensor("inner_encoder_conv_module_1_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15104896)))]; + tensor inner_encoder_conv_module_1_sequential_0_weight = const()[name = tensor("inner_encoder_conv_module_1_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15105984)))]; + tensor inner_encoder_conv_module_1_sequential_2_conv_bias = const()[name = tensor("inner_encoder_conv_module_1_sequential_2_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15107072)))]; + tensor inner_encoder_conv_module_1_sequential_2_conv_weight = const()[name = tensor("inner_encoder_conv_module_1_sequential_2_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15109184)))]; + tensor inner_encoder_conv_module_1_sequential_4_conv_weight = const()[name = tensor("inner_encoder_conv_module_1_sequential_4_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15633536)))]; + tensor inner_encoder_conv_module_1_sequential_5_running_var = const()[name = tensor("inner_encoder_conv_module_1_sequential_5_running_var"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15649984)))]; + tensor inner_encoder_conv_module_1_sequential_5_running_mean = const()[name = tensor("inner_encoder_conv_module_1_sequential_5_running_mean"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15651072)))]; + tensor inner_encoder_conv_module_1_sequential_5_bias = const()[name = tensor("inner_encoder_conv_module_1_sequential_5_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15652160)))]; + tensor inner_encoder_conv_module_1_sequential_5_weight = const()[name = tensor("inner_encoder_conv_module_1_sequential_5_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15653248)))]; + tensor inner_encoder_conv_module_1_sequential_7_conv_bias = const()[name = tensor("inner_encoder_conv_module_1_sequential_7_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15654336)))]; + tensor inner_encoder_conv_module_1_sequential_7_conv_weight = const()[name = tensor("inner_encoder_conv_module_1_sequential_7_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15655424)))]; + tensor inner_encoder_ffn2_1_module_sequential_0_bias = const()[name = tensor("inner_encoder_ffn2_1_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15917632)))]; + tensor inner_encoder_ffn2_1_module_sequential_0_weight = const()[name = tensor("inner_encoder_ffn2_1_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15918720)))]; + tensor inner_encoder_ffn2_1_module_sequential_1_linear_bias = const()[name = tensor("inner_encoder_ffn2_1_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15919808)))]; + tensor inner_encoder_ffn2_1_module_sequential_1_linear_weight = const()[name = tensor("inner_encoder_ffn2_1_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15923968)))]; + tensor inner_encoder_ffn2_1_module_sequential_4_linear_bias = const()[name = tensor("inner_encoder_ffn2_1_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(16972608)))]; + tensor inner_encoder_ffn2_1_module_sequential_4_linear_weight = const()[name = tensor("inner_encoder_ffn2_1_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(16973696)))]; + tensor inner_encoder_layer_norm_1_bias = const()[name = tensor("inner_encoder_layer_norm_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18022336)))]; + tensor inner_encoder_layer_norm_1_weight = const()[name = tensor("inner_encoder_layer_norm_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18023424)))]; + tensor inner_encoder_ffn1_2_module_sequential_0_bias = const()[name = tensor("inner_encoder_ffn1_2_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18024512)))]; + tensor inner_encoder_ffn1_2_module_sequential_0_weight = const()[name = tensor("inner_encoder_ffn1_2_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18025600)))]; + tensor inner_encoder_ffn1_2_module_sequential_1_linear_bias = const()[name = tensor("inner_encoder_ffn1_2_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18026688)))]; + tensor inner_encoder_ffn1_2_module_sequential_1_linear_weight = const()[name = tensor("inner_encoder_ffn1_2_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18030848)))]; + tensor inner_encoder_ffn1_2_module_sequential_4_linear_bias = const()[name = tensor("inner_encoder_ffn1_2_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(19079488)))]; + tensor inner_encoder_ffn1_2_module_sequential_4_linear_weight = const()[name = tensor("inner_encoder_ffn1_2_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(19080576)))]; + tensor inner_encoder_ret_lns_2_bias = const()[name = tensor("inner_encoder_ret_lns_2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20129216)))]; + tensor inner_encoder_ret_lns_2_weight = const()[name = tensor("inner_encoder_ret_lns_2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20130304)))]; + tensor inner_encoder_q_proj_2_bias = const()[name = tensor("inner_encoder_q_proj_2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20131392)))]; + tensor inner_encoder_q_proj_2_weight = const()[name = tensor("inner_encoder_q_proj_2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20132480)))]; + tensor inner_encoder_k_proj_2_bias = const()[name = tensor("inner_encoder_k_proj_2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20394688)))]; + tensor inner_encoder_k_proj_2_weight = const()[name = tensor("inner_encoder_k_proj_2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20395776)))]; + tensor inner_encoder_v_proj_2_bias = const()[name = tensor("inner_encoder_v_proj_2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20657984)))]; + tensor inner_encoder_v_proj_2_weight = const()[name = tensor("inner_encoder_v_proj_2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20659072)))]; + tensor inner_encoder_g_proj_2_bias = const()[name = tensor("inner_encoder_g_proj_2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20921280)))]; + tensor inner_encoder_g_proj_2_weight = const()[name = tensor("inner_encoder_g_proj_2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20922368)))]; + tensor inner_encoder_out_proj_2_bias = const()[name = tensor("inner_encoder_out_proj_2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21184576)))]; + tensor inner_encoder_out_proj_2_weight = const()[name = tensor("inner_encoder_out_proj_2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21185664)))]; + tensor inner_encoder_conv_module_2_sequential_0_bias = const()[name = tensor("inner_encoder_conv_module_2_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21447872)))]; + tensor inner_encoder_conv_module_2_sequential_0_weight = const()[name = tensor("inner_encoder_conv_module_2_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21448960)))]; + tensor inner_encoder_conv_module_2_sequential_2_conv_bias = const()[name = tensor("inner_encoder_conv_module_2_sequential_2_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21450048)))]; + tensor inner_encoder_conv_module_2_sequential_2_conv_weight = const()[name = tensor("inner_encoder_conv_module_2_sequential_2_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21452160)))]; + tensor inner_encoder_conv_module_2_sequential_4_conv_weight = const()[name = tensor("inner_encoder_conv_module_2_sequential_4_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21976512)))]; + tensor inner_encoder_conv_module_2_sequential_5_running_var = const()[name = tensor("inner_encoder_conv_module_2_sequential_5_running_var"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21992960)))]; + tensor inner_encoder_conv_module_2_sequential_5_running_mean = const()[name = tensor("inner_encoder_conv_module_2_sequential_5_running_mean"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21994048)))]; + tensor inner_encoder_conv_module_2_sequential_5_bias = const()[name = tensor("inner_encoder_conv_module_2_sequential_5_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21995136)))]; + tensor inner_encoder_conv_module_2_sequential_5_weight = const()[name = tensor("inner_encoder_conv_module_2_sequential_5_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21996224)))]; + tensor inner_encoder_conv_module_2_sequential_7_conv_bias = const()[name = tensor("inner_encoder_conv_module_2_sequential_7_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21997312)))]; + tensor inner_encoder_conv_module_2_sequential_7_conv_weight = const()[name = tensor("inner_encoder_conv_module_2_sequential_7_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21998400)))]; + tensor inner_encoder_ffn2_2_module_sequential_0_bias = const()[name = tensor("inner_encoder_ffn2_2_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(22260608)))]; + tensor inner_encoder_ffn2_2_module_sequential_0_weight = const()[name = tensor("inner_encoder_ffn2_2_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(22261696)))]; + tensor inner_encoder_ffn2_2_module_sequential_1_linear_bias = const()[name = tensor("inner_encoder_ffn2_2_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(22262784)))]; + tensor inner_encoder_ffn2_2_module_sequential_1_linear_weight = const()[name = tensor("inner_encoder_ffn2_2_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(22266944)))]; + tensor inner_encoder_ffn2_2_module_sequential_4_linear_bias = const()[name = tensor("inner_encoder_ffn2_2_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(23315584)))]; + tensor inner_encoder_ffn2_2_module_sequential_4_linear_weight = const()[name = tensor("inner_encoder_ffn2_2_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(23316672)))]; + tensor inner_encoder_layer_norm_2_bias = const()[name = tensor("inner_encoder_layer_norm_2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(24365312)))]; + tensor inner_encoder_layer_norm_2_weight = const()[name = tensor("inner_encoder_layer_norm_2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(24366400)))]; + tensor inner_encoder_ffn1_3_module_sequential_0_bias = const()[name = tensor("inner_encoder_ffn1_3_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(24367488)))]; + tensor inner_encoder_ffn1_3_module_sequential_0_weight = const()[name = tensor("inner_encoder_ffn1_3_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(24368576)))]; + tensor inner_encoder_ffn1_3_module_sequential_1_linear_bias = const()[name = tensor("inner_encoder_ffn1_3_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(24369664)))]; + tensor inner_encoder_ffn1_3_module_sequential_1_linear_weight = const()[name = tensor("inner_encoder_ffn1_3_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(24373824)))]; + tensor inner_encoder_ffn1_3_module_sequential_4_linear_bias = const()[name = tensor("inner_encoder_ffn1_3_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(25422464)))]; + tensor inner_encoder_ffn1_3_module_sequential_4_linear_weight = const()[name = tensor("inner_encoder_ffn1_3_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(25423552)))]; + tensor inner_encoder_ret_lns_3_bias = const()[name = tensor("inner_encoder_ret_lns_3_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(26472192)))]; + tensor inner_encoder_ret_lns_3_weight = const()[name = tensor("inner_encoder_ret_lns_3_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(26473280)))]; + tensor inner_encoder_q_proj_3_bias = const()[name = tensor("inner_encoder_q_proj_3_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(26474368)))]; + tensor inner_encoder_q_proj_3_weight = const()[name = tensor("inner_encoder_q_proj_3_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(26475456)))]; + tensor inner_encoder_k_proj_3_bias = const()[name = tensor("inner_encoder_k_proj_3_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(26737664)))]; + tensor inner_encoder_k_proj_3_weight = const()[name = tensor("inner_encoder_k_proj_3_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(26738752)))]; + tensor inner_encoder_v_proj_3_bias = const()[name = tensor("inner_encoder_v_proj_3_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27000960)))]; + tensor inner_encoder_v_proj_3_weight = const()[name = tensor("inner_encoder_v_proj_3_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27002048)))]; + tensor inner_encoder_g_proj_3_bias = const()[name = tensor("inner_encoder_g_proj_3_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27264256)))]; + tensor inner_encoder_g_proj_3_weight = const()[name = tensor("inner_encoder_g_proj_3_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27265344)))]; + tensor inner_encoder_out_proj_3_bias = const()[name = tensor("inner_encoder_out_proj_3_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27527552)))]; + tensor inner_encoder_out_proj_3_weight = const()[name = tensor("inner_encoder_out_proj_3_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27528640)))]; + tensor inner_encoder_conv_module_3_sequential_0_bias = const()[name = tensor("inner_encoder_conv_module_3_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27790848)))]; + tensor inner_encoder_conv_module_3_sequential_0_weight = const()[name = tensor("inner_encoder_conv_module_3_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27791936)))]; + tensor inner_encoder_conv_module_3_sequential_2_conv_bias = const()[name = tensor("inner_encoder_conv_module_3_sequential_2_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27793024)))]; + tensor inner_encoder_conv_module_3_sequential_2_conv_weight = const()[name = tensor("inner_encoder_conv_module_3_sequential_2_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27795136)))]; + tensor inner_encoder_conv_module_3_sequential_4_conv_weight = const()[name = tensor("inner_encoder_conv_module_3_sequential_4_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28319488)))]; + tensor inner_encoder_conv_module_3_sequential_5_running_var = const()[name = tensor("inner_encoder_conv_module_3_sequential_5_running_var"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28335936)))]; + tensor inner_encoder_conv_module_3_sequential_5_running_mean = const()[name = tensor("inner_encoder_conv_module_3_sequential_5_running_mean"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28337024)))]; + tensor inner_encoder_conv_module_3_sequential_5_bias = const()[name = tensor("inner_encoder_conv_module_3_sequential_5_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28338112)))]; + tensor inner_encoder_conv_module_3_sequential_5_weight = const()[name = tensor("inner_encoder_conv_module_3_sequential_5_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28339200)))]; + tensor inner_encoder_conv_module_3_sequential_7_conv_bias = const()[name = tensor("inner_encoder_conv_module_3_sequential_7_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28340288)))]; + tensor inner_encoder_conv_module_3_sequential_7_conv_weight = const()[name = tensor("inner_encoder_conv_module_3_sequential_7_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28341376)))]; + tensor inner_encoder_ffn2_3_module_sequential_0_bias = const()[name = tensor("inner_encoder_ffn2_3_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28603584)))]; + tensor inner_encoder_ffn2_3_module_sequential_0_weight = const()[name = tensor("inner_encoder_ffn2_3_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28604672)))]; + tensor inner_encoder_ffn2_3_module_sequential_1_linear_bias = const()[name = tensor("inner_encoder_ffn2_3_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28605760)))]; + tensor inner_encoder_ffn2_3_module_sequential_1_linear_weight = const()[name = tensor("inner_encoder_ffn2_3_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28609920)))]; + tensor inner_encoder_ffn2_3_module_sequential_4_linear_bias = const()[name = tensor("inner_encoder_ffn2_3_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(29658560)))]; + tensor inner_encoder_ffn2_3_module_sequential_4_linear_weight = const()[name = tensor("inner_encoder_ffn2_3_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(29659648)))]; + tensor inner_encoder_layer_norm_3_bias = const()[name = tensor("inner_encoder_layer_norm_3_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(30708288)))]; + tensor inner_encoder_layer_norm_3_weight = const()[name = tensor("inner_encoder_layer_norm_3_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(30709376)))]; + tensor inner_decoder_convert_bias = const()[name = tensor("inner_decoder_convert_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(30710464)))]; + tensor inner_decoder_convert_weight = const()[name = tensor("inner_decoder_convert_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(30711552)))]; + tensor inner_decoder_q_proj_0_bias = const()[name = tensor("inner_decoder_q_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(31235904)))]; + tensor inner_decoder_q_proj_0_weight = const()[name = tensor("inner_decoder_q_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(31236992)))]; + tensor inner_decoder_k_proj_0_bias = const()[name = tensor("inner_decoder_k_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(31499200)))]; + tensor inner_decoder_k_proj_0_weight = const()[name = tensor("inner_decoder_k_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(31500288)))]; + tensor inner_decoder_v_proj_0_bias = const()[name = tensor("inner_decoder_v_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(31762496)))]; + tensor inner_decoder_v_proj_0_weight = const()[name = tensor("inner_decoder_v_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(31763584)))]; + tensor inner_decoder_g_proj_0_bias = const()[name = tensor("inner_decoder_g_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32025792)))]; + tensor inner_decoder_g_proj_0_weight = const()[name = tensor("inner_decoder_g_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32026880)))]; + tensor inner_decoder_out_proj_0_bias = const()[name = tensor("inner_decoder_out_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32289088)))]; + tensor inner_decoder_out_proj_0_weight = const()[name = tensor("inner_decoder_out_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32290176)))]; + tensor inner_decoder_norm11_0_bias = const()[name = tensor("inner_decoder_norm11_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32552384)))]; + tensor inner_decoder_norm11_0_weight = const()[name = tensor("inner_decoder_norm11_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32553472)))]; + tensor inner_decoder_self_attn2_0_out_proj_bias = const()[name = tensor("inner_decoder_self_attn2_0_out_proj_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32554560)))]; + tensor inner_decoder_self_attn2_0_out_proj_weight = const()[name = tensor("inner_decoder_self_attn2_0_out_proj_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32555648)))]; + tensor inner_decoder_self_attn2_0_in_proj_bias = const()[name = tensor("inner_decoder_self_attn2_0_in_proj_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32817856)))]; + tensor inner_decoder_self_attn2_0_in_proj_weight = const()[name = tensor("inner_decoder_self_attn2_0_in_proj_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32820992)))]; + tensor inner_decoder_norm21_0_bias = const()[name = tensor("inner_decoder_norm21_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(33607488)))]; + tensor inner_decoder_norm21_0_weight = const()[name = tensor("inner_decoder_norm21_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(33608576)))]; + tensor inner_decoder_linear1_0_bias = const()[name = tensor("inner_decoder_linear1_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(33609664)))]; + tensor inner_decoder_linear1_0_weight = const()[name = tensor("inner_decoder_linear1_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(33617920)))]; + tensor inner_decoder_linear2_0_bias = const()[name = tensor("inner_decoder_linear2_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(35715136)))]; + tensor inner_decoder_linear2_0_weight = const()[name = tensor("inner_decoder_linear2_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(35716224)))]; + tensor inner_decoder_norm22_0_bias = const()[name = tensor("inner_decoder_norm22_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(37813440)))]; + tensor inner_decoder_norm22_0_weight = const()[name = tensor("inner_decoder_norm22_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(37814528)))]; + tensor inner_decoder_q_proj_1_bias = const()[name = tensor("inner_decoder_q_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(37815616)))]; + tensor inner_decoder_q_proj_1_weight = const()[name = tensor("inner_decoder_q_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(37816704)))]; + tensor inner_decoder_k_proj_1_bias = const()[name = tensor("inner_decoder_k_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38078912)))]; + tensor inner_decoder_k_proj_1_weight = const()[name = tensor("inner_decoder_k_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38080000)))]; + tensor inner_decoder_v_proj_1_bias = const()[name = tensor("inner_decoder_v_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38342208)))]; + tensor inner_decoder_v_proj_1_weight = const()[name = tensor("inner_decoder_v_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38343296)))]; + tensor inner_decoder_g_proj_1_bias = const()[name = tensor("inner_decoder_g_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38605504)))]; + tensor inner_decoder_g_proj_1_weight = const()[name = tensor("inner_decoder_g_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38606592)))]; + tensor inner_decoder_out_proj_1_bias = const()[name = tensor("inner_decoder_out_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38868800)))]; + tensor inner_decoder_out_proj_1_weight = const()[name = tensor("inner_decoder_out_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38869888)))]; + tensor inner_decoder_norm11_1_bias = const()[name = tensor("inner_decoder_norm11_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(39132096)))]; + tensor inner_decoder_norm11_1_weight = const()[name = tensor("inner_decoder_norm11_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(39133184)))]; + tensor inner_decoder_self_attn2_1_out_proj_bias = const()[name = tensor("inner_decoder_self_attn2_1_out_proj_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(39134272)))]; + tensor inner_decoder_self_attn2_1_out_proj_weight = const()[name = tensor("inner_decoder_self_attn2_1_out_proj_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(39135360)))]; + tensor inner_decoder_self_attn2_1_in_proj_bias = const()[name = tensor("inner_decoder_self_attn2_1_in_proj_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(39397568)))]; + tensor inner_decoder_self_attn2_1_in_proj_weight = const()[name = tensor("inner_decoder_self_attn2_1_in_proj_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(39400704)))]; + tensor inner_decoder_norm21_1_bias = const()[name = tensor("inner_decoder_norm21_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(40187200)))]; + tensor inner_decoder_norm21_1_weight = const()[name = tensor("inner_decoder_norm21_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(40188288)))]; + tensor inner_decoder_linear1_1_bias = const()[name = tensor("inner_decoder_linear1_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(40189376)))]; + tensor inner_decoder_linear1_1_weight = const()[name = tensor("inner_decoder_linear1_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(40197632)))]; + tensor inner_decoder_linear2_1_bias = const()[name = tensor("inner_decoder_linear2_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(42294848)))]; + tensor inner_decoder_linear2_1_weight = const()[name = tensor("inner_decoder_linear2_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(42295936)))]; + tensor inner_decoder_norm22_1_bias = const()[name = tensor("inner_decoder_norm22_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(44393152)))]; + tensor inner_decoder_norm22_1_weight = const()[name = tensor("inner_decoder_norm22_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(44394240)))]; + tensor stacked_axes_0 = const()[name = tensor("stacked_axes_0"), val = tensor([1])]; + tensor stacked = expand_dims(axes = stacked_axes_0, x = features)[name = tensor("stacked")]; + tensor var_26 = const()[name = tensor("op_26"), val = tensor([1, 1, 345])]; + tensor input_1 = reshape(shape = var_26, x = stacked)[name = tensor("input_1")]; + tensor var_29 = const()[name = tensor("op_29"), val = tensor(0x1p+0)]; + tensor var_35 = const()[name = tensor("op_35"), val = tensor(true)]; + tensor var_36 = const()[name = tensor("op_36"), val = tensor(0x1.4f8b58p-17)]; + tensor var_39 = const()[name = tensor("op_39"), val = tensor(0)]; + tensor var_41 = const()[name = tensor("op_41"), val = tensor(2)]; + tensor var_42 = const()[name = tensor("op_42"), val = tensor(-1)]; + tensor var_44 = const()[name = tensor("op_44"), val = tensor(0x1.0c6f7ap-20)]; + tensor var_49 = const()[name = tensor("op_49"), val = tensor(0x1.5798eep-27)]; + tensor var_52 = const()[name = tensor("op_52"), val = tensor(1)]; + tensor input_3 = linear(bias = inner_encoder_input_projection_linear_bias, weight = inner_encoder_input_projection_linear_weight, x = input_1)[name = tensor("linear_0")]; tensor input_5_axes_0 = const()[name = tensor("input_5_axes_0"), val = tensor([-1])]; - tensor input_5 = layer_norm(axes = input_5_axes_0, beta = encoder_ffn1_0_module_sequential_0_bias, epsilon = var_28, gamma = encoder_ffn1_0_module_sequential_0_weight, x = input_3)[name = tensor("input_5")]; - tensor inputs_1 = linear(bias = encoder_ffn1_0_module_sequential_1_linear_bias, weight = encoder_ffn1_0_module_sequential_1_linear_weight, x = input_5)[name = tensor("linear_1")]; - tensor input_7 = silu(x = inputs_1)[name = tensor("input_7")]; - tensor input_11 = linear(bias = encoder_ffn1_0_module_sequential_4_linear_bias, weight = encoder_ffn1_0_module_sequential_4_linear_weight, x = input_7)[name = tensor("linear_2")]; - tensor var_147 = const()[name = tensor("op_147"), val = tensor(0x1p-1)]; - tensor var_148 = mul(x = input_11, y = var_147)[name = tensor("op_148")]; - tensor input_13 = add(x = var_148, y = input_3)[name = tensor("input_13")]; + tensor input_5 = layer_norm(axes = input_5_axes_0, beta = inner_encoder_pre_layer_norm_bias, epsilon = var_36, gamma = inner_encoder_pre_layer_norm_weight, x = input_3)[name = tensor("input_5")]; + tensor input_7_axes_0 = const()[name = tensor("input_7_axes_0"), val = tensor([-1])]; + tensor input_7 = layer_norm(axes = input_7_axes_0, beta = inner_encoder_ffn1_0_module_sequential_0_bias, epsilon = var_36, gamma = inner_encoder_ffn1_0_module_sequential_0_weight, x = input_5)[name = tensor("input_7")]; + tensor inputs_1 = linear(bias = inner_encoder_ffn1_0_module_sequential_1_linear_bias, weight = inner_encoder_ffn1_0_module_sequential_1_linear_weight, x = input_7)[name = tensor("linear_1")]; + tensor input_9 = silu(x = inputs_1)[name = tensor("input_9")]; + tensor input_13 = linear(bias = inner_encoder_ffn1_0_module_sequential_4_linear_bias, weight = inner_encoder_ffn1_0_module_sequential_4_linear_weight, x = input_9)[name = tensor("linear_2")]; + tensor var_173 = const()[name = tensor("op_173"), val = tensor(0x1p-1)]; + tensor var_174 = mul(x = input_13, y = var_173)[name = tensor("op_174")]; + tensor input_15 = add(x = var_174, y = input_5)[name = tensor("input_15")]; tensor x_1_axes_0 = const()[name = tensor("x_1_axes_0"), val = tensor([-1])]; - tensor x_1 = layer_norm(axes = x_1_axes_0, beta = encoder_ret_lns_0_bias, epsilon = var_28, gamma = encoder_ret_lns_0_weight, x = input_13)[name = tensor("x_1")]; + tensor x_1 = layer_norm(axes = x_1_axes_0, beta = inner_encoder_ret_lns_0_bias, epsilon = var_36, gamma = inner_encoder_ret_lns_0_weight, x = input_15)[name = tensor("x_1")]; tensor prev_kv_1_begin_0 = const()[name = tensor("prev_kv_1_begin_0"), val = tensor([0, 0, 0, 0, 0])]; tensor prev_kv_1_end_0 = const()[name = tensor("prev_kv_1_end_0"), val = tensor([1, 1, 4, 64, 64])]; tensor prev_kv_1_end_mask_0 = const()[name = tensor("prev_kv_1_end_mask_0"), val = tensor([false, true, true, true, true])]; @@ -238,139 +244,139 @@ program(1.0) tensor prev_scale_1_end_mask_0 = const()[name = tensor("prev_scale_1_end_mask_0"), val = tensor([false, true])]; tensor prev_scale_1_squeeze_mask_0 = const()[name = tensor("prev_scale_1_squeeze_mask_0"), val = tensor([true, false])]; tensor prev_scale_1 = slice_by_index(begin = prev_scale_1_begin_0, end = prev_scale_1_end_0, end_mask = prev_scale_1_end_mask_0, squeeze_mask = prev_scale_1_squeeze_mask_0, x = enc_scale)[name = tensor("prev_scale_1")]; - tensor var_162 = linear(bias = encoder_q_proj_0_bias, weight = encoder_q_proj_0_weight, x = x_1)[name = tensor("linear_3")]; - tensor var_163 = const()[name = tensor("op_163"), val = tensor([1, 1, 4, 64])]; - tensor var_164 = reshape(shape = var_163, x = var_162)[name = tensor("op_164")]; + tensor var_188 = linear(bias = inner_encoder_q_proj_0_bias, weight = inner_encoder_q_proj_0_weight, x = x_1)[name = tensor("linear_3")]; + tensor var_189 = const()[name = tensor("op_189"), val = tensor([1, 1, 4, 64])]; + tensor var_190 = reshape(shape = var_189, x = var_188)[name = tensor("op_190")]; tensor q_1_perm_0 = const()[name = tensor("q_1_perm_0"), val = tensor([0, 2, 1, 3])]; - tensor var_168 = linear(bias = encoder_k_proj_0_bias, weight = encoder_k_proj_0_weight, x = x_1)[name = tensor("linear_4")]; - tensor var_169 = const()[name = tensor("op_169"), val = tensor(0x1p-3)]; - tensor var_170 = mul(x = var_168, y = var_169)[name = tensor("op_170")]; - tensor var_171 = const()[name = tensor("op_171"), val = tensor([1, 1, 4, 64])]; - tensor var_172 = reshape(shape = var_171, x = var_170)[name = tensor("op_172")]; + tensor var_194 = linear(bias = inner_encoder_k_proj_0_bias, weight = inner_encoder_k_proj_0_weight, x = x_1)[name = tensor("linear_4")]; + tensor var_195 = const()[name = tensor("op_195"), val = tensor(0x1p-3)]; + tensor var_196 = mul(x = var_194, y = var_195)[name = tensor("op_196")]; + tensor var_197 = const()[name = tensor("op_197"), val = tensor([1, 1, 4, 64])]; + tensor var_198 = reshape(shape = var_197, x = var_196)[name = tensor("op_198")]; tensor k_1_perm_0 = const()[name = tensor("k_1_perm_0"), val = tensor([0, 2, 1, 3])]; - tensor var_176 = linear(bias = encoder_v_proj_0_bias, weight = encoder_v_proj_0_weight, x = x_1)[name = tensor("linear_5")]; - tensor var_177 = const()[name = tensor("op_177"), val = tensor([1, 1, 4, 64])]; - tensor var_178 = reshape(shape = var_177, x = var_176)[name = tensor("op_178")]; + tensor var_202 = linear(bias = inner_encoder_v_proj_0_bias, weight = inner_encoder_v_proj_0_weight, x = x_1)[name = tensor("linear_5")]; + tensor var_203 = const()[name = tensor("op_203"), val = tensor([1, 1, 4, 64])]; + tensor var_204 = reshape(shape = var_203, x = var_202)[name = tensor("op_204")]; tensor v_1_perm_0 = const()[name = tensor("v_1_perm_0"), val = tensor([0, 2, 1, 3])]; - tensor input_17 = linear(bias = encoder_g_proj_0_bias, weight = encoder_g_proj_0_weight, x = x_1)[name = tensor("linear_6")]; + tensor input_19 = linear(bias = inner_encoder_g_proj_0_bias, weight = inner_encoder_g_proj_0_weight, x = x_1)[name = tensor("linear_6")]; tensor sqrt_s0_1 = sqrt(x = prev_scale_1)[name = tensor("sqrt_s0_1")]; - tensor s_t_1 = add(x = prev_scale_1, y = encoder__t_index)[name = tensor("s_t_1")]; + tensor s_t_1 = add(x = prev_scale_1, y = inner_encoder__t_index)[name = tensor("s_t_1")]; tensor sqrt_s_t_1 = sqrt(x = s_t_1)[name = tensor("sqrt_s_t_1")]; tensor qk_1_transpose_x_1 = const()[name = tensor("qk_1_transpose_x_1"), val = tensor(false)]; tensor qk_1_transpose_y_1 = const()[name = tensor("qk_1_transpose_y_1"), val = tensor(true)]; - tensor k_1 = transpose(perm = k_1_perm_0, x = var_172)[name = tensor("transpose_57")]; - tensor q_1 = transpose(perm = q_1_perm_0, x = var_164)[name = tensor("transpose_58")]; + tensor k_1 = transpose(perm = k_1_perm_0, x = var_198)[name = tensor("transpose_57")]; + tensor q_1 = transpose(perm = q_1_perm_0, x = var_190)[name = tensor("transpose_58")]; tensor qk_1 = matmul(transpose_x = qk_1_transpose_x_1, transpose_y = qk_1_transpose_y_1, x = q_1, y = k_1)[name = tensor("qk_1")]; - tensor var_188 = const()[name = tensor("op_188"), val = tensor([1, 1])]; - tensor var_189 = reshape(shape = var_188, x = sqrt_s_t_1)[name = tensor("op_189")]; - tensor M_1 = real_div(x = encoder__causal_mask, y = var_189)[name = tensor("M_1")]; - tensor var_191 = mul(x = qk_1, y = M_1)[name = tensor("op_191")]; + tensor var_214 = const()[name = tensor("op_214"), val = tensor([1, 1])]; + tensor var_215 = reshape(shape = var_214, x = sqrt_s_t_1)[name = tensor("op_215")]; + tensor M_1 = real_div(x = inner_encoder__causal_mask, y = var_215)[name = tensor("M_1")]; + tensor var_217 = mul(x = qk_1, y = M_1)[name = tensor("op_217")]; tensor inner_1_transpose_x_0 = const()[name = tensor("inner_1_transpose_x_0"), val = tensor(false)]; tensor inner_1_transpose_y_0 = const()[name = tensor("inner_1_transpose_y_0"), val = tensor(false)]; - tensor v_1 = transpose(perm = v_1_perm_0, x = var_178)[name = tensor("transpose_56")]; - tensor inner_1 = matmul(transpose_x = inner_1_transpose_x_0, transpose_y = inner_1_transpose_y_0, x = var_191, y = v_1)[name = tensor("inner_1")]; - tensor var_193_transpose_x_0 = const()[name = tensor("op_193_transpose_x_0"), val = tensor(false)]; - tensor var_193_transpose_y_0 = const()[name = tensor("op_193_transpose_y_0"), val = tensor(false)]; - tensor var_193 = matmul(transpose_x = var_193_transpose_x_0, transpose_y = var_193_transpose_y_0, x = q_1, y = prev_kv_1)[name = tensor("op_193")]; - tensor var_194 = real_div(x = sqrt_s0_1, y = sqrt_s_t_1)[name = tensor("op_194")]; - tensor var_195 = const()[name = tensor("op_195"), val = tensor([1, 1, 1, 1])]; - tensor var_196 = reshape(shape = var_195, x = var_194)[name = tensor("op_196")]; - tensor cross_1 = mul(x = var_193, y = var_196)[name = tensor("cross_1")]; + tensor v_1 = transpose(perm = v_1_perm_0, x = var_204)[name = tensor("transpose_56")]; + tensor inner_1 = matmul(transpose_x = inner_1_transpose_x_0, transpose_y = inner_1_transpose_y_0, x = var_217, y = v_1)[name = tensor("inner_1")]; + tensor var_219_transpose_x_0 = const()[name = tensor("op_219_transpose_x_0"), val = tensor(false)]; + tensor var_219_transpose_y_0 = const()[name = tensor("op_219_transpose_y_0"), val = tensor(false)]; + tensor var_219 = matmul(transpose_x = var_219_transpose_x_0, transpose_y = var_219_transpose_y_0, x = q_1, y = prev_kv_1)[name = tensor("op_219")]; + tensor var_220 = real_div(x = sqrt_s0_1, y = sqrt_s_t_1)[name = tensor("op_220")]; + tensor var_221 = const()[name = tensor("op_221"), val = tensor([1, 1, 1, 1])]; + tensor var_222 = reshape(shape = var_221, x = var_220)[name = tensor("op_222")]; + tensor cross_1 = mul(x = var_219, y = var_222)[name = tensor("cross_1")]; tensor out_1 = add(x = inner_1, y = cross_1)[name = tensor("out_1")]; - tensor var_199 = mul(x = prev_kv_1, y = sqrt_s0_1)[name = tensor("op_199")]; - tensor var_201_transpose_x_1 = const()[name = tensor("op_201_transpose_x_1"), val = tensor(true)]; - tensor var_201_transpose_y_1 = const()[name = tensor("op_201_transpose_y_1"), val = tensor(false)]; - tensor var_201 = matmul(transpose_x = var_201_transpose_x_1, transpose_y = var_201_transpose_y_1, x = k_1, y = v_1)[name = tensor("op_201")]; - tensor new_kv_unnorm_1 = add(x = var_199, y = var_201)[name = tensor("new_kv_unnorm_1")]; - tensor var_203 = const()[name = tensor("op_203"), val = tensor(0x1p+0)]; - tensor new_scale_1 = add(x = prev_scale_1, y = var_203)[name = tensor("new_scale_1")]; - tensor var_205 = sqrt(x = new_scale_1)[name = tensor("op_205")]; - tensor var_206 = real_div(x = new_kv_unnorm_1, y = var_205)[name = tensor("op_206")]; - tensor var_207_perm_0 = const()[name = tensor("op_207_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_225 = mul(x = prev_kv_1, y = sqrt_s0_1)[name = tensor("op_225")]; + tensor var_227_transpose_x_1 = const()[name = tensor("op_227_transpose_x_1"), val = tensor(true)]; + tensor var_227_transpose_y_1 = const()[name = tensor("op_227_transpose_y_1"), val = tensor(false)]; + tensor var_227 = matmul(transpose_x = var_227_transpose_x_1, transpose_y = var_227_transpose_y_1, x = k_1, y = v_1)[name = tensor("op_227")]; + tensor new_kv_unnorm_1 = add(x = var_225, y = var_227)[name = tensor("new_kv_unnorm_1")]; + tensor var_229 = const()[name = tensor("op_229"), val = tensor(0x1p+0)]; + tensor new_scale_1 = add(x = prev_scale_1, y = var_229)[name = tensor("new_scale_1")]; + tensor var_231 = sqrt(x = new_scale_1)[name = tensor("op_231")]; + tensor var_232 = real_div(x = new_kv_unnorm_1, y = var_231)[name = tensor("op_232")]; + tensor var_233_perm_0 = const()[name = tensor("op_233_perm_0"), val = tensor([0, 2, 1, 3])]; tensor out_3_axes_0 = const()[name = tensor("out_3_axes_0"), val = tensor([-1])]; - tensor var_207 = transpose(perm = var_207_perm_0, x = out_1)[name = tensor("transpose_55")]; - tensor out_3 = layer_norm(axes = out_3_axes_0, epsilon = var_17, x = var_207)[name = tensor("out_3")]; - tensor var_211 = const()[name = tensor("op_211"), val = tensor([1, 1, 256])]; - tensor out_5 = reshape(shape = var_211, x = out_3)[name = tensor("out_5")]; - tensor var_213 = silu(x = input_17)[name = tensor("op_213")]; - tensor input_19 = mul(x = var_213, y = out_5)[name = tensor("input_19")]; - tensor ret_out_1 = linear(bias = encoder_out_proj_0_bias, weight = encoder_out_proj_0_weight, x = input_19)[name = tensor("linear_7")]; - tensor x_3 = add(x = input_13, y = ret_out_1)[name = tensor("x_3")]; + tensor var_233 = transpose(perm = var_233_perm_0, x = out_1)[name = tensor("transpose_55")]; + tensor out_3 = layer_norm(axes = out_3_axes_0, epsilon = var_44, x = var_233)[name = tensor("out_3")]; + tensor var_237 = const()[name = tensor("op_237"), val = tensor([1, 1, 256])]; + tensor out_5 = reshape(shape = var_237, x = out_3)[name = tensor("out_5")]; + tensor var_239 = silu(x = input_19)[name = tensor("op_239")]; + tensor input_21 = mul(x = var_239, y = out_5)[name = tensor("input_21")]; + tensor ret_out_1 = linear(bias = inner_encoder_out_proj_0_bias, weight = inner_encoder_out_proj_0_weight, x = input_21)[name = tensor("linear_7")]; + tensor x_3 = add(x = input_15, y = ret_out_1)[name = tensor("x_3")]; tensor window_1_begin_0 = const()[name = tensor("window_1_begin_0"), val = tensor([0, 0, 0, 0])]; tensor window_1_end_0 = const()[name = tensor("window_1_end_0"), val = tensor([1, 1, 16, 256])]; tensor window_1_end_mask_0 = const()[name = tensor("window_1_end_mask_0"), val = tensor([false, true, true, true])]; tensor window_1_squeeze_mask_0 = const()[name = tensor("window_1_squeeze_mask_0"), val = tensor([true, false, false, false])]; tensor window_1 = slice_by_index(begin = window_1_begin_0, end = window_1_end_0, end_mask = window_1_end_mask_0, squeeze_mask = window_1_squeeze_mask_0, x = enc_conv_cache)[name = tensor("window_1")]; - tensor var_224_begin_0 = const()[name = tensor("op_224_begin_0"), val = tensor([0, 1, 0])]; - tensor var_224_end_0 = const()[name = tensor("op_224_end_0"), val = tensor([1, 16, 256])]; - tensor var_224_end_mask_0 = const()[name = tensor("op_224_end_mask_0"), val = tensor([true, true, true])]; - tensor var_224 = slice_by_index(begin = var_224_begin_0, end = var_224_end_0, end_mask = var_224_end_mask_0, x = window_1)[name = tensor("op_224")]; + tensor var_250_begin_0 = const()[name = tensor("op_250_begin_0"), val = tensor([0, 1, 0])]; + tensor var_250_end_0 = const()[name = tensor("op_250_end_0"), val = tensor([1, 16, 256])]; + tensor var_250_end_mask_0 = const()[name = tensor("op_250_end_mask_0"), val = tensor([true, true, true])]; + tensor var_250 = slice_by_index(begin = var_250_begin_0, end = var_250_end_0, end_mask = var_250_end_mask_0, x = window_1)[name = tensor("op_250")]; tensor window_3_interleave_0 = const()[name = tensor("window_3_interleave_0"), val = tensor(false)]; - tensor window_3 = concat(axis = var_26, interleave = window_3_interleave_0, values = (var_224, x_3))[name = tensor("window_3")]; - tensor input_21_interleave_0 = const()[name = tensor("input_21_interleave_0"), val = tensor(false)]; - tensor input_21 = concat(axis = var_23, interleave = input_21_interleave_0, values = window_3)[name = tensor("input_21")]; + tensor window_3 = concat(axis = var_52, interleave = window_3_interleave_0, values = (var_250, x_3))[name = tensor("window_3")]; + tensor input_23_interleave_0 = const()[name = tensor("input_23_interleave_0"), val = tensor(false)]; + tensor input_23 = concat(axis = var_39, interleave = input_23_interleave_0, values = window_3)[name = tensor("input_23")]; tensor x_5_axes_0 = const()[name = tensor("x_5_axes_0"), val = tensor([-1])]; - tensor x_5 = layer_norm(axes = x_5_axes_0, beta = encoder_conv_module_0_sequential_0_bias, epsilon = var_28, gamma = encoder_conv_module_0_sequential_0_weight, x = input_21)[name = tensor("x_5")]; - tensor input_23_perm_0 = const()[name = tensor("input_23_perm_0"), val = tensor([0, 2, 1])]; + tensor x_5 = layer_norm(axes = x_5_axes_0, beta = inner_encoder_conv_module_0_sequential_0_bias, epsilon = var_36, gamma = inner_encoder_conv_module_0_sequential_0_weight, x = input_23)[name = tensor("x_5")]; + tensor input_25_perm_0 = const()[name = tensor("input_25_perm_0"), val = tensor([0, 2, 1])]; tensor inputs_3_pad_type_0 = const()[name = tensor("inputs_3_pad_type_0"), val = tensor("valid")]; tensor inputs_3_strides_0 = const()[name = tensor("inputs_3_strides_0"), val = tensor([1])]; tensor inputs_3_pad_0 = const()[name = tensor("inputs_3_pad_0"), val = tensor([0, 0])]; tensor inputs_3_dilations_0 = const()[name = tensor("inputs_3_dilations_0"), val = tensor([1])]; tensor inputs_3_groups_0 = const()[name = tensor("inputs_3_groups_0"), val = tensor(1)]; - tensor input_23 = transpose(perm = input_23_perm_0, x = x_5)[name = tensor("transpose_54")]; - tensor inputs_3 = conv(bias = encoder_conv_module_0_sequential_2_conv_bias, dilations = inputs_3_dilations_0, groups = inputs_3_groups_0, pad = inputs_3_pad_0, pad_type = inputs_3_pad_type_0, strides = inputs_3_strides_0, weight = encoder_conv_module_0_sequential_2_conv_weight, x = input_23)[name = tensor("inputs_3")]; - tensor var_249_split_sizes_0 = const()[name = tensor("op_249_split_sizes_0"), val = tensor([256, 256])]; - tensor var_249_axis_0 = const()[name = tensor("op_249_axis_0"), val = tensor(1)]; - tensor var_249_0, tensor var_249_1 = split(axis = var_249_axis_0, split_sizes = var_249_split_sizes_0, x = inputs_3)[name = tensor("op_249")]; - tensor var_251 = sigmoid(x = var_249_1)[name = tensor("op_251")]; - tensor inputs_5 = mul(x = var_249_0, y = var_251)[name = tensor("inputs_5")]; + tensor input_25 = transpose(perm = input_25_perm_0, x = x_5)[name = tensor("transpose_54")]; + tensor inputs_3 = conv(bias = inner_encoder_conv_module_0_sequential_2_conv_bias, dilations = inputs_3_dilations_0, groups = inputs_3_groups_0, pad = inputs_3_pad_0, pad_type = inputs_3_pad_type_0, strides = inputs_3_strides_0, weight = inner_encoder_conv_module_0_sequential_2_conv_weight, x = input_25)[name = tensor("inputs_3")]; + tensor var_275_split_sizes_0 = const()[name = tensor("op_275_split_sizes_0"), val = tensor([256, 256])]; + tensor var_275_axis_0 = const()[name = tensor("op_275_axis_0"), val = tensor(1)]; + tensor var_275_0, tensor var_275_1 = split(axis = var_275_axis_0, split_sizes = var_275_split_sizes_0, x = inputs_3)[name = tensor("op_275")]; + tensor var_277 = sigmoid(x = var_275_1)[name = tensor("op_277")]; + tensor inputs_5 = mul(x = var_275_0, y = var_277)[name = tensor("inputs_5")]; tensor outputs_aug_1_pad_type_0 = const()[name = tensor("outputs_aug_1_pad_type_0"), val = tensor("custom")]; tensor outputs_aug_1_pad_0 = const()[name = tensor("outputs_aug_1_pad_0"), val = tensor([15, 15])]; tensor outputs_aug_1_groups_0 = const()[name = tensor("outputs_aug_1_groups_0"), val = tensor(256)]; tensor outputs_aug_1_strides_0 = const()[name = tensor("outputs_aug_1_strides_0"), val = tensor([1])]; tensor outputs_aug_1_dilations_0 = const()[name = tensor("outputs_aug_1_dilations_0"), val = tensor([1])]; - tensor outputs_aug_1 = conv(dilations = outputs_aug_1_dilations_0, groups = outputs_aug_1_groups_0, pad = outputs_aug_1_pad_0, pad_type = outputs_aug_1_pad_type_0, strides = outputs_aug_1_strides_0, weight = encoder_conv_module_0_sequential_4_conv_weight, x = inputs_5)[name = tensor("outputs_aug_1")]; - tensor input_25_begin_0 = const()[name = tensor("input_25_begin_0"), val = tensor([0, 0, 0])]; - tensor input_25_end_0 = const()[name = tensor("input_25_end_0"), val = tensor([1, 256, 16])]; - tensor input_25_end_mask_0 = const()[name = tensor("input_25_end_mask_0"), val = tensor([true, true, false])]; - tensor input_25 = slice_by_index(begin = input_25_begin_0, end = input_25_end_0, end_mask = input_25_end_mask_0, x = outputs_aug_1)[name = tensor("input_25")]; - tensor inputs_7 = batch_norm(beta = encoder_conv_module_0_sequential_5_bias, epsilon = var_28, gamma = encoder_conv_module_0_sequential_5_weight, mean = encoder_conv_module_0_sequential_5_running_mean, variance = encoder_conv_module_0_sequential_5_running_var, x = input_25)[name = tensor("inputs_7")]; - tensor input_27 = silu(x = inputs_7)[name = tensor("input_27")]; - tensor input_29_pad_type_0 = const()[name = tensor("input_29_pad_type_0"), val = tensor("valid")]; - tensor input_29_strides_0 = const()[name = tensor("input_29_strides_0"), val = tensor([1])]; - tensor input_29_pad_0 = const()[name = tensor("input_29_pad_0"), val = tensor([0, 0])]; - tensor input_29_dilations_0 = const()[name = tensor("input_29_dilations_0"), val = tensor([1])]; - tensor input_29_groups_0 = const()[name = tensor("input_29_groups_0"), val = tensor(1)]; - tensor input_29 = conv(bias = encoder_conv_module_0_sequential_7_conv_bias, dilations = input_29_dilations_0, groups = input_29_groups_0, pad = input_29_pad_0, pad_type = input_29_pad_type_0, strides = input_29_strides_0, weight = encoder_conv_module_0_sequential_7_conv_weight, x = input_27)[name = tensor("input_29")]; + tensor outputs_aug_1 = conv(dilations = outputs_aug_1_dilations_0, groups = outputs_aug_1_groups_0, pad = outputs_aug_1_pad_0, pad_type = outputs_aug_1_pad_type_0, strides = outputs_aug_1_strides_0, weight = inner_encoder_conv_module_0_sequential_4_conv_weight, x = inputs_5)[name = tensor("outputs_aug_1")]; + tensor input_27_begin_0 = const()[name = tensor("input_27_begin_0"), val = tensor([0, 0, 0])]; + tensor input_27_end_0 = const()[name = tensor("input_27_end_0"), val = tensor([1, 256, 16])]; + tensor input_27_end_mask_0 = const()[name = tensor("input_27_end_mask_0"), val = tensor([true, true, false])]; + tensor input_27 = slice_by_index(begin = input_27_begin_0, end = input_27_end_0, end_mask = input_27_end_mask_0, x = outputs_aug_1)[name = tensor("input_27")]; + tensor inputs_7 = batch_norm(beta = inner_encoder_conv_module_0_sequential_5_bias, epsilon = var_36, gamma = inner_encoder_conv_module_0_sequential_5_weight, mean = inner_encoder_conv_module_0_sequential_5_running_mean, variance = inner_encoder_conv_module_0_sequential_5_running_var, x = input_27)[name = tensor("inputs_7")]; + tensor input_29 = silu(x = inputs_7)[name = tensor("input_29")]; + tensor input_31_pad_type_0 = const()[name = tensor("input_31_pad_type_0"), val = tensor("valid")]; + tensor input_31_strides_0 = const()[name = tensor("input_31_strides_0"), val = tensor([1])]; + tensor input_31_pad_0 = const()[name = tensor("input_31_pad_0"), val = tensor([0, 0])]; + tensor input_31_dilations_0 = const()[name = tensor("input_31_dilations_0"), val = tensor([1])]; + tensor input_31_groups_0 = const()[name = tensor("input_31_groups_0"), val = tensor(1)]; + tensor input_31 = conv(bias = inner_encoder_conv_module_0_sequential_7_conv_bias, dilations = input_31_dilations_0, groups = input_31_groups_0, pad = input_31_pad_0, pad_type = input_31_pad_type_0, strides = input_31_strides_0, weight = inner_encoder_conv_module_0_sequential_7_conv_weight, x = input_29)[name = tensor("input_31")]; tensor conv_out_1_perm_0 = const()[name = tensor("conv_out_1_perm_0"), val = tensor([0, 2, 1])]; - tensor var_282_begin_0 = const()[name = tensor("op_282_begin_0"), val = tensor([0, -1, 0])]; - tensor var_282_end_0 = const()[name = tensor("op_282_end_0"), val = tensor([1, 16, 256])]; - tensor var_282_end_mask_0 = const()[name = tensor("op_282_end_mask_0"), val = tensor([true, true, true])]; - tensor conv_out_1 = transpose(perm = conv_out_1_perm_0, x = input_29)[name = tensor("transpose_53")]; - tensor var_282 = slice_by_index(begin = var_282_begin_0, end = var_282_end_0, end_mask = var_282_end_mask_0, x = conv_out_1)[name = tensor("op_282")]; - tensor var_284_perm_0 = const()[name = tensor("op_284_perm_0"), val = tensor([1, 0, 2])]; - tensor var_284 = transpose(perm = var_284_perm_0, x = var_282)[name = tensor("transpose_52")]; - tensor input_31 = add(x = x_3, y = var_284)[name = tensor("input_31")]; - tensor input_33_axes_0 = const()[name = tensor("input_33_axes_0"), val = tensor([-1])]; - tensor input_33 = layer_norm(axes = input_33_axes_0, beta = encoder_ffn2_0_module_sequential_0_bias, epsilon = var_28, gamma = encoder_ffn2_0_module_sequential_0_weight, x = input_31)[name = tensor("input_33")]; - tensor inputs_9 = linear(bias = encoder_ffn2_0_module_sequential_1_linear_bias, weight = encoder_ffn2_0_module_sequential_1_linear_weight, x = input_33)[name = tensor("linear_8")]; - tensor input_35 = silu(x = inputs_9)[name = tensor("input_35")]; - tensor input_39 = linear(bias = encoder_ffn2_0_module_sequential_4_linear_bias, weight = encoder_ffn2_0_module_sequential_4_linear_weight, x = input_35)[name = tensor("linear_9")]; - tensor var_307 = const()[name = tensor("op_307"), val = tensor(0x1p-1)]; - tensor var_308 = mul(x = input_39, y = var_307)[name = tensor("op_308")]; - tensor input_41 = add(x = var_308, y = input_31)[name = tensor("input_41")]; - tensor input_43_axes_0 = const()[name = tensor("input_43_axes_0"), val = tensor([-1])]; - tensor input_43 = layer_norm(axes = input_43_axes_0, beta = encoder_layer_norm_0_bias, epsilon = var_28, gamma = encoder_layer_norm_0_weight, x = input_41)[name = tensor("input_43")]; + tensor var_308_begin_0 = const()[name = tensor("op_308_begin_0"), val = tensor([0, -1, 0])]; + tensor var_308_end_0 = const()[name = tensor("op_308_end_0"), val = tensor([1, 16, 256])]; + tensor var_308_end_mask_0 = const()[name = tensor("op_308_end_mask_0"), val = tensor([true, true, true])]; + tensor conv_out_1 = transpose(perm = conv_out_1_perm_0, x = input_31)[name = tensor("transpose_53")]; + tensor var_308 = slice_by_index(begin = var_308_begin_0, end = var_308_end_0, end_mask = var_308_end_mask_0, x = conv_out_1)[name = tensor("op_308")]; + tensor var_310_perm_0 = const()[name = tensor("op_310_perm_0"), val = tensor([1, 0, 2])]; + tensor var_310 = transpose(perm = var_310_perm_0, x = var_308)[name = tensor("transpose_52")]; + tensor input_33 = add(x = x_3, y = var_310)[name = tensor("input_33")]; + tensor input_35_axes_0 = const()[name = tensor("input_35_axes_0"), val = tensor([-1])]; + tensor input_35 = layer_norm(axes = input_35_axes_0, beta = inner_encoder_ffn2_0_module_sequential_0_bias, epsilon = var_36, gamma = inner_encoder_ffn2_0_module_sequential_0_weight, x = input_33)[name = tensor("input_35")]; + tensor inputs_9 = linear(bias = inner_encoder_ffn2_0_module_sequential_1_linear_bias, weight = inner_encoder_ffn2_0_module_sequential_1_linear_weight, x = input_35)[name = tensor("linear_8")]; + tensor input_37 = silu(x = inputs_9)[name = tensor("input_37")]; + tensor input_41 = linear(bias = inner_encoder_ffn2_0_module_sequential_4_linear_bias, weight = inner_encoder_ffn2_0_module_sequential_4_linear_weight, x = input_37)[name = tensor("linear_9")]; + tensor var_333 = const()[name = tensor("op_333"), val = tensor(0x1p-1)]; + tensor var_334 = mul(x = input_41, y = var_333)[name = tensor("op_334")]; + tensor input_43 = add(x = var_334, y = input_33)[name = tensor("input_43")]; tensor input_45_axes_0 = const()[name = tensor("input_45_axes_0"), val = tensor([-1])]; - tensor input_45 = layer_norm(axes = input_45_axes_0, beta = encoder_ffn1_1_module_sequential_0_bias, epsilon = var_28, gamma = encoder_ffn1_1_module_sequential_0_weight, x = input_43)[name = tensor("input_45")]; - tensor inputs_11 = linear(bias = encoder_ffn1_1_module_sequential_1_linear_bias, weight = encoder_ffn1_1_module_sequential_1_linear_weight, x = input_45)[name = tensor("linear_10")]; - tensor input_47 = silu(x = inputs_11)[name = tensor("input_47")]; - tensor input_51 = linear(bias = encoder_ffn1_1_module_sequential_4_linear_bias, weight = encoder_ffn1_1_module_sequential_4_linear_weight, x = input_47)[name = tensor("linear_11")]; - tensor var_337 = const()[name = tensor("op_337"), val = tensor(0x1p-1)]; - tensor var_338 = mul(x = input_51, y = var_337)[name = tensor("op_338")]; - tensor input_53 = add(x = var_338, y = input_43)[name = tensor("input_53")]; + tensor input_45 = layer_norm(axes = input_45_axes_0, beta = inner_encoder_layer_norm_0_bias, epsilon = var_36, gamma = inner_encoder_layer_norm_0_weight, x = input_43)[name = tensor("input_45")]; + tensor input_47_axes_0 = const()[name = tensor("input_47_axes_0"), val = tensor([-1])]; + tensor input_47 = layer_norm(axes = input_47_axes_0, beta = inner_encoder_ffn1_1_module_sequential_0_bias, epsilon = var_36, gamma = inner_encoder_ffn1_1_module_sequential_0_weight, x = input_45)[name = tensor("input_47")]; + tensor inputs_11 = linear(bias = inner_encoder_ffn1_1_module_sequential_1_linear_bias, weight = inner_encoder_ffn1_1_module_sequential_1_linear_weight, x = input_47)[name = tensor("linear_10")]; + tensor input_49 = silu(x = inputs_11)[name = tensor("input_49")]; + tensor input_53 = linear(bias = inner_encoder_ffn1_1_module_sequential_4_linear_bias, weight = inner_encoder_ffn1_1_module_sequential_4_linear_weight, x = input_49)[name = tensor("linear_11")]; + tensor var_363 = const()[name = tensor("op_363"), val = tensor(0x1p-1)]; + tensor var_364 = mul(x = input_53, y = var_363)[name = tensor("op_364")]; + tensor input_55 = add(x = var_364, y = input_45)[name = tensor("input_55")]; tensor x_7_axes_0 = const()[name = tensor("x_7_axes_0"), val = tensor([-1])]; - tensor x_7 = layer_norm(axes = x_7_axes_0, beta = encoder_ret_lns_1_bias, epsilon = var_28, gamma = encoder_ret_lns_1_weight, x = input_53)[name = tensor("x_7")]; + tensor x_7 = layer_norm(axes = x_7_axes_0, beta = inner_encoder_ret_lns_1_bias, epsilon = var_36, gamma = inner_encoder_ret_lns_1_weight, x = input_55)[name = tensor("x_7")]; tensor prev_kv_3_begin_0 = const()[name = tensor("prev_kv_3_begin_0"), val = tensor([1, 0, 0, 0, 0])]; tensor prev_kv_3_end_0 = const()[name = tensor("prev_kv_3_end_0"), val = tensor([2, 1, 4, 64, 64])]; tensor prev_kv_3_end_mask_0 = const()[name = tensor("prev_kv_3_end_mask_0"), val = tensor([false, true, true, true, true])]; @@ -381,139 +387,139 @@ program(1.0) tensor prev_scale_3_end_mask_0 = const()[name = tensor("prev_scale_3_end_mask_0"), val = tensor([false, true])]; tensor prev_scale_3_squeeze_mask_0 = const()[name = tensor("prev_scale_3_squeeze_mask_0"), val = tensor([true, false])]; tensor prev_scale_3 = slice_by_index(begin = prev_scale_3_begin_0, end = prev_scale_3_end_0, end_mask = prev_scale_3_end_mask_0, squeeze_mask = prev_scale_3_squeeze_mask_0, x = enc_scale)[name = tensor("prev_scale_3")]; - tensor var_352 = linear(bias = encoder_q_proj_1_bias, weight = encoder_q_proj_1_weight, x = x_7)[name = tensor("linear_12")]; - tensor var_353 = const()[name = tensor("op_353"), val = tensor([1, 1, 4, 64])]; - tensor var_354 = reshape(shape = var_353, x = var_352)[name = tensor("op_354")]; + tensor var_378 = linear(bias = inner_encoder_q_proj_1_bias, weight = inner_encoder_q_proj_1_weight, x = x_7)[name = tensor("linear_12")]; + tensor var_379 = const()[name = tensor("op_379"), val = tensor([1, 1, 4, 64])]; + tensor var_380 = reshape(shape = var_379, x = var_378)[name = tensor("op_380")]; tensor q_3_perm_0 = const()[name = tensor("q_3_perm_0"), val = tensor([0, 2, 1, 3])]; - tensor var_358 = linear(bias = encoder_k_proj_1_bias, weight = encoder_k_proj_1_weight, x = x_7)[name = tensor("linear_13")]; - tensor var_359 = const()[name = tensor("op_359"), val = tensor(0x1p-3)]; - tensor var_360 = mul(x = var_358, y = var_359)[name = tensor("op_360")]; - tensor var_361 = const()[name = tensor("op_361"), val = tensor([1, 1, 4, 64])]; - tensor var_362 = reshape(shape = var_361, x = var_360)[name = tensor("op_362")]; + tensor var_384 = linear(bias = inner_encoder_k_proj_1_bias, weight = inner_encoder_k_proj_1_weight, x = x_7)[name = tensor("linear_13")]; + tensor var_385 = const()[name = tensor("op_385"), val = tensor(0x1p-3)]; + tensor var_386 = mul(x = var_384, y = var_385)[name = tensor("op_386")]; + tensor var_387 = const()[name = tensor("op_387"), val = tensor([1, 1, 4, 64])]; + tensor var_388 = reshape(shape = var_387, x = var_386)[name = tensor("op_388")]; tensor k_3_perm_0 = const()[name = tensor("k_3_perm_0"), val = tensor([0, 2, 1, 3])]; - tensor var_366 = linear(bias = encoder_v_proj_1_bias, weight = encoder_v_proj_1_weight, x = x_7)[name = tensor("linear_14")]; - tensor var_367 = const()[name = tensor("op_367"), val = tensor([1, 1, 4, 64])]; - tensor var_368 = reshape(shape = var_367, x = var_366)[name = tensor("op_368")]; + tensor var_392 = linear(bias = inner_encoder_v_proj_1_bias, weight = inner_encoder_v_proj_1_weight, x = x_7)[name = tensor("linear_14")]; + tensor var_393 = const()[name = tensor("op_393"), val = tensor([1, 1, 4, 64])]; + tensor var_394 = reshape(shape = var_393, x = var_392)[name = tensor("op_394")]; tensor v_3_perm_0 = const()[name = tensor("v_3_perm_0"), val = tensor([0, 2, 1, 3])]; - tensor input_57 = linear(bias = encoder_g_proj_1_bias, weight = encoder_g_proj_1_weight, x = x_7)[name = tensor("linear_15")]; + tensor input_59 = linear(bias = inner_encoder_g_proj_1_bias, weight = inner_encoder_g_proj_1_weight, x = x_7)[name = tensor("linear_15")]; tensor sqrt_s0_3 = sqrt(x = prev_scale_3)[name = tensor("sqrt_s0_3")]; - tensor s_t_3 = add(x = prev_scale_3, y = encoder__t_index)[name = tensor("s_t_3")]; + tensor s_t_3 = add(x = prev_scale_3, y = inner_encoder__t_index)[name = tensor("s_t_3")]; tensor sqrt_s_t_3 = sqrt(x = s_t_3)[name = tensor("sqrt_s_t_3")]; tensor qk_3_transpose_x_1 = const()[name = tensor("qk_3_transpose_x_1"), val = tensor(false)]; tensor qk_3_transpose_y_1 = const()[name = tensor("qk_3_transpose_y_1"), val = tensor(true)]; - tensor k_3 = transpose(perm = k_3_perm_0, x = var_362)[name = tensor("transpose_50")]; - tensor q_3 = transpose(perm = q_3_perm_0, x = var_354)[name = tensor("transpose_51")]; + tensor k_3 = transpose(perm = k_3_perm_0, x = var_388)[name = tensor("transpose_50")]; + tensor q_3 = transpose(perm = q_3_perm_0, x = var_380)[name = tensor("transpose_51")]; tensor qk_3 = matmul(transpose_x = qk_3_transpose_x_1, transpose_y = qk_3_transpose_y_1, x = q_3, y = k_3)[name = tensor("qk_3")]; - tensor var_378 = const()[name = tensor("op_378"), val = tensor([1, 1])]; - tensor var_379 = reshape(shape = var_378, x = sqrt_s_t_3)[name = tensor("op_379")]; - tensor M_3 = real_div(x = encoder__causal_mask, y = var_379)[name = tensor("M_3")]; - tensor var_381 = mul(x = qk_3, y = M_3)[name = tensor("op_381")]; + tensor var_404 = const()[name = tensor("op_404"), val = tensor([1, 1])]; + tensor var_405 = reshape(shape = var_404, x = sqrt_s_t_3)[name = tensor("op_405")]; + tensor M_3 = real_div(x = inner_encoder__causal_mask, y = var_405)[name = tensor("M_3")]; + tensor var_407 = mul(x = qk_3, y = M_3)[name = tensor("op_407")]; tensor inner_3_transpose_x_0 = const()[name = tensor("inner_3_transpose_x_0"), val = tensor(false)]; tensor inner_3_transpose_y_0 = const()[name = tensor("inner_3_transpose_y_0"), val = tensor(false)]; - tensor v_3 = transpose(perm = v_3_perm_0, x = var_368)[name = tensor("transpose_49")]; - tensor inner_3 = matmul(transpose_x = inner_3_transpose_x_0, transpose_y = inner_3_transpose_y_0, x = var_381, y = v_3)[name = tensor("inner_3")]; - tensor var_383_transpose_x_0 = const()[name = tensor("op_383_transpose_x_0"), val = tensor(false)]; - tensor var_383_transpose_y_0 = const()[name = tensor("op_383_transpose_y_0"), val = tensor(false)]; - tensor var_383 = matmul(transpose_x = var_383_transpose_x_0, transpose_y = var_383_transpose_y_0, x = q_3, y = prev_kv_3)[name = tensor("op_383")]; - tensor var_384 = real_div(x = sqrt_s0_3, y = sqrt_s_t_3)[name = tensor("op_384")]; - tensor var_385 = const()[name = tensor("op_385"), val = tensor([1, 1, 1, 1])]; - tensor var_386 = reshape(shape = var_385, x = var_384)[name = tensor("op_386")]; - tensor cross_3 = mul(x = var_383, y = var_386)[name = tensor("cross_3")]; + tensor v_3 = transpose(perm = v_3_perm_0, x = var_394)[name = tensor("transpose_49")]; + tensor inner_3 = matmul(transpose_x = inner_3_transpose_x_0, transpose_y = inner_3_transpose_y_0, x = var_407, y = v_3)[name = tensor("inner_3")]; + tensor var_409_transpose_x_0 = const()[name = tensor("op_409_transpose_x_0"), val = tensor(false)]; + tensor var_409_transpose_y_0 = const()[name = tensor("op_409_transpose_y_0"), val = tensor(false)]; + tensor var_409 = matmul(transpose_x = var_409_transpose_x_0, transpose_y = var_409_transpose_y_0, x = q_3, y = prev_kv_3)[name = tensor("op_409")]; + tensor var_410 = real_div(x = sqrt_s0_3, y = sqrt_s_t_3)[name = tensor("op_410")]; + tensor var_411 = const()[name = tensor("op_411"), val = tensor([1, 1, 1, 1])]; + tensor var_412 = reshape(shape = var_411, x = var_410)[name = tensor("op_412")]; + tensor cross_3 = mul(x = var_409, y = var_412)[name = tensor("cross_3")]; tensor out_7 = add(x = inner_3, y = cross_3)[name = tensor("out_7")]; - tensor var_389 = mul(x = prev_kv_3, y = sqrt_s0_3)[name = tensor("op_389")]; - tensor var_391_transpose_x_1 = const()[name = tensor("op_391_transpose_x_1"), val = tensor(true)]; - tensor var_391_transpose_y_1 = const()[name = tensor("op_391_transpose_y_1"), val = tensor(false)]; - tensor var_391 = matmul(transpose_x = var_391_transpose_x_1, transpose_y = var_391_transpose_y_1, x = k_3, y = v_3)[name = tensor("op_391")]; - tensor new_kv_unnorm_3 = add(x = var_389, y = var_391)[name = tensor("new_kv_unnorm_3")]; - tensor var_393 = const()[name = tensor("op_393"), val = tensor(0x1p+0)]; - tensor new_scale_3 = add(x = prev_scale_3, y = var_393)[name = tensor("new_scale_3")]; - tensor var_395 = sqrt(x = new_scale_3)[name = tensor("op_395")]; - tensor var_396 = real_div(x = new_kv_unnorm_3, y = var_395)[name = tensor("op_396")]; - tensor var_397_perm_0 = const()[name = tensor("op_397_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_415 = mul(x = prev_kv_3, y = sqrt_s0_3)[name = tensor("op_415")]; + tensor var_417_transpose_x_1 = const()[name = tensor("op_417_transpose_x_1"), val = tensor(true)]; + tensor var_417_transpose_y_1 = const()[name = tensor("op_417_transpose_y_1"), val = tensor(false)]; + tensor var_417 = matmul(transpose_x = var_417_transpose_x_1, transpose_y = var_417_transpose_y_1, x = k_3, y = v_3)[name = tensor("op_417")]; + tensor new_kv_unnorm_3 = add(x = var_415, y = var_417)[name = tensor("new_kv_unnorm_3")]; + tensor var_419 = const()[name = tensor("op_419"), val = tensor(0x1p+0)]; + tensor new_scale_3 = add(x = prev_scale_3, y = var_419)[name = tensor("new_scale_3")]; + tensor var_421 = sqrt(x = new_scale_3)[name = tensor("op_421")]; + tensor var_422 = real_div(x = new_kv_unnorm_3, y = var_421)[name = tensor("op_422")]; + tensor var_423_perm_0 = const()[name = tensor("op_423_perm_0"), val = tensor([0, 2, 1, 3])]; tensor out_9_axes_0 = const()[name = tensor("out_9_axes_0"), val = tensor([-1])]; - tensor var_397 = transpose(perm = var_397_perm_0, x = out_7)[name = tensor("transpose_48")]; - tensor out_9 = layer_norm(axes = out_9_axes_0, epsilon = var_17, x = var_397)[name = tensor("out_9")]; - tensor var_401 = const()[name = tensor("op_401"), val = tensor([1, 1, 256])]; - tensor out_11 = reshape(shape = var_401, x = out_9)[name = tensor("out_11")]; - tensor var_403 = silu(x = input_57)[name = tensor("op_403")]; - tensor input_59 = mul(x = var_403, y = out_11)[name = tensor("input_59")]; - tensor ret_out_3 = linear(bias = encoder_out_proj_1_bias, weight = encoder_out_proj_1_weight, x = input_59)[name = tensor("linear_16")]; - tensor x_9 = add(x = input_53, y = ret_out_3)[name = tensor("x_9")]; + tensor var_423 = transpose(perm = var_423_perm_0, x = out_7)[name = tensor("transpose_48")]; + tensor out_9 = layer_norm(axes = out_9_axes_0, epsilon = var_44, x = var_423)[name = tensor("out_9")]; + tensor var_427 = const()[name = tensor("op_427"), val = tensor([1, 1, 256])]; + tensor out_11 = reshape(shape = var_427, x = out_9)[name = tensor("out_11")]; + tensor var_429 = silu(x = input_59)[name = tensor("op_429")]; + tensor input_61 = mul(x = var_429, y = out_11)[name = tensor("input_61")]; + tensor ret_out_3 = linear(bias = inner_encoder_out_proj_1_bias, weight = inner_encoder_out_proj_1_weight, x = input_61)[name = tensor("linear_16")]; + tensor x_9 = add(x = input_55, y = ret_out_3)[name = tensor("x_9")]; tensor window_5_begin_0 = const()[name = tensor("window_5_begin_0"), val = tensor([1, 0, 0, 0])]; tensor window_5_end_0 = const()[name = tensor("window_5_end_0"), val = tensor([2, 1, 16, 256])]; tensor window_5_end_mask_0 = const()[name = tensor("window_5_end_mask_0"), val = tensor([false, true, true, true])]; tensor window_5_squeeze_mask_0 = const()[name = tensor("window_5_squeeze_mask_0"), val = tensor([true, false, false, false])]; tensor window_5 = slice_by_index(begin = window_5_begin_0, end = window_5_end_0, end_mask = window_5_end_mask_0, squeeze_mask = window_5_squeeze_mask_0, x = enc_conv_cache)[name = tensor("window_5")]; - tensor var_414_begin_0 = const()[name = tensor("op_414_begin_0"), val = tensor([0, 1, 0])]; - tensor var_414_end_0 = const()[name = tensor("op_414_end_0"), val = tensor([1, 16, 256])]; - tensor var_414_end_mask_0 = const()[name = tensor("op_414_end_mask_0"), val = tensor([true, true, true])]; - tensor var_414 = slice_by_index(begin = var_414_begin_0, end = var_414_end_0, end_mask = var_414_end_mask_0, x = window_5)[name = tensor("op_414")]; + tensor var_440_begin_0 = const()[name = tensor("op_440_begin_0"), val = tensor([0, 1, 0])]; + tensor var_440_end_0 = const()[name = tensor("op_440_end_0"), val = tensor([1, 16, 256])]; + tensor var_440_end_mask_0 = const()[name = tensor("op_440_end_mask_0"), val = tensor([true, true, true])]; + tensor var_440 = slice_by_index(begin = var_440_begin_0, end = var_440_end_0, end_mask = var_440_end_mask_0, x = window_5)[name = tensor("op_440")]; tensor window_7_interleave_0 = const()[name = tensor("window_7_interleave_0"), val = tensor(false)]; - tensor window_7 = concat(axis = var_26, interleave = window_7_interleave_0, values = (var_414, x_9))[name = tensor("window_7")]; - tensor input_61_interleave_0 = const()[name = tensor("input_61_interleave_0"), val = tensor(false)]; - tensor input_61 = concat(axis = var_23, interleave = input_61_interleave_0, values = window_7)[name = tensor("input_61")]; + tensor window_7 = concat(axis = var_52, interleave = window_7_interleave_0, values = (var_440, x_9))[name = tensor("window_7")]; + tensor input_63_interleave_0 = const()[name = tensor("input_63_interleave_0"), val = tensor(false)]; + tensor input_63 = concat(axis = var_39, interleave = input_63_interleave_0, values = window_7)[name = tensor("input_63")]; tensor x_11_axes_0 = const()[name = tensor("x_11_axes_0"), val = tensor([-1])]; - tensor x_11 = layer_norm(axes = x_11_axes_0, beta = encoder_conv_module_1_sequential_0_bias, epsilon = var_28, gamma = encoder_conv_module_1_sequential_0_weight, x = input_61)[name = tensor("x_11")]; - tensor input_63_perm_0 = const()[name = tensor("input_63_perm_0"), val = tensor([0, 2, 1])]; + tensor x_11 = layer_norm(axes = x_11_axes_0, beta = inner_encoder_conv_module_1_sequential_0_bias, epsilon = var_36, gamma = inner_encoder_conv_module_1_sequential_0_weight, x = input_63)[name = tensor("x_11")]; + tensor input_65_perm_0 = const()[name = tensor("input_65_perm_0"), val = tensor([0, 2, 1])]; tensor inputs_13_pad_type_0 = const()[name = tensor("inputs_13_pad_type_0"), val = tensor("valid")]; tensor inputs_13_strides_0 = const()[name = tensor("inputs_13_strides_0"), val = tensor([1])]; tensor inputs_13_pad_0 = const()[name = tensor("inputs_13_pad_0"), val = tensor([0, 0])]; tensor inputs_13_dilations_0 = const()[name = tensor("inputs_13_dilations_0"), val = tensor([1])]; tensor inputs_13_groups_0 = const()[name = tensor("inputs_13_groups_0"), val = tensor(1)]; - tensor input_63 = transpose(perm = input_63_perm_0, x = x_11)[name = tensor("transpose_47")]; - tensor inputs_13 = conv(bias = encoder_conv_module_1_sequential_2_conv_bias, dilations = inputs_13_dilations_0, groups = inputs_13_groups_0, pad = inputs_13_pad_0, pad_type = inputs_13_pad_type_0, strides = inputs_13_strides_0, weight = encoder_conv_module_1_sequential_2_conv_weight, x = input_63)[name = tensor("inputs_13")]; - tensor var_439_split_sizes_0 = const()[name = tensor("op_439_split_sizes_0"), val = tensor([256, 256])]; - tensor var_439_axis_0 = const()[name = tensor("op_439_axis_0"), val = tensor(1)]; - tensor var_439_0, tensor var_439_1 = split(axis = var_439_axis_0, split_sizes = var_439_split_sizes_0, x = inputs_13)[name = tensor("op_439")]; - tensor var_441 = sigmoid(x = var_439_1)[name = tensor("op_441")]; - tensor inputs_15 = mul(x = var_439_0, y = var_441)[name = tensor("inputs_15")]; + tensor input_65 = transpose(perm = input_65_perm_0, x = x_11)[name = tensor("transpose_47")]; + tensor inputs_13 = conv(bias = inner_encoder_conv_module_1_sequential_2_conv_bias, dilations = inputs_13_dilations_0, groups = inputs_13_groups_0, pad = inputs_13_pad_0, pad_type = inputs_13_pad_type_0, strides = inputs_13_strides_0, weight = inner_encoder_conv_module_1_sequential_2_conv_weight, x = input_65)[name = tensor("inputs_13")]; + tensor var_465_split_sizes_0 = const()[name = tensor("op_465_split_sizes_0"), val = tensor([256, 256])]; + tensor var_465_axis_0 = const()[name = tensor("op_465_axis_0"), val = tensor(1)]; + tensor var_465_0, tensor var_465_1 = split(axis = var_465_axis_0, split_sizes = var_465_split_sizes_0, x = inputs_13)[name = tensor("op_465")]; + tensor var_467 = sigmoid(x = var_465_1)[name = tensor("op_467")]; + tensor inputs_15 = mul(x = var_465_0, y = var_467)[name = tensor("inputs_15")]; tensor outputs_aug_3_pad_type_0 = const()[name = tensor("outputs_aug_3_pad_type_0"), val = tensor("custom")]; tensor outputs_aug_3_pad_0 = const()[name = tensor("outputs_aug_3_pad_0"), val = tensor([15, 15])]; tensor outputs_aug_3_groups_0 = const()[name = tensor("outputs_aug_3_groups_0"), val = tensor(256)]; tensor outputs_aug_3_strides_0 = const()[name = tensor("outputs_aug_3_strides_0"), val = tensor([1])]; tensor outputs_aug_3_dilations_0 = const()[name = tensor("outputs_aug_3_dilations_0"), val = tensor([1])]; - tensor outputs_aug_3 = conv(dilations = outputs_aug_3_dilations_0, groups = outputs_aug_3_groups_0, pad = outputs_aug_3_pad_0, pad_type = outputs_aug_3_pad_type_0, strides = outputs_aug_3_strides_0, weight = encoder_conv_module_1_sequential_4_conv_weight, x = inputs_15)[name = tensor("outputs_aug_3")]; - tensor input_65_begin_0 = const()[name = tensor("input_65_begin_0"), val = tensor([0, 0, 0])]; - tensor input_65_end_0 = const()[name = tensor("input_65_end_0"), val = tensor([1, 256, 16])]; - tensor input_65_end_mask_0 = const()[name = tensor("input_65_end_mask_0"), val = tensor([true, true, false])]; - tensor input_65 = slice_by_index(begin = input_65_begin_0, end = input_65_end_0, end_mask = input_65_end_mask_0, x = outputs_aug_3)[name = tensor("input_65")]; - tensor inputs_17 = batch_norm(beta = encoder_conv_module_1_sequential_5_bias, epsilon = var_28, gamma = encoder_conv_module_1_sequential_5_weight, mean = encoder_conv_module_1_sequential_5_running_mean, variance = encoder_conv_module_1_sequential_5_running_var, x = input_65)[name = tensor("inputs_17")]; - tensor input_67 = silu(x = inputs_17)[name = tensor("input_67")]; - tensor input_69_pad_type_0 = const()[name = tensor("input_69_pad_type_0"), val = tensor("valid")]; - tensor input_69_strides_0 = const()[name = tensor("input_69_strides_0"), val = tensor([1])]; - tensor input_69_pad_0 = const()[name = tensor("input_69_pad_0"), val = tensor([0, 0])]; - tensor input_69_dilations_0 = const()[name = tensor("input_69_dilations_0"), val = tensor([1])]; - tensor input_69_groups_0 = const()[name = tensor("input_69_groups_0"), val = tensor(1)]; - tensor input_69 = conv(bias = encoder_conv_module_1_sequential_7_conv_bias, dilations = input_69_dilations_0, groups = input_69_groups_0, pad = input_69_pad_0, pad_type = input_69_pad_type_0, strides = input_69_strides_0, weight = encoder_conv_module_1_sequential_7_conv_weight, x = input_67)[name = tensor("input_69")]; + tensor outputs_aug_3 = conv(dilations = outputs_aug_3_dilations_0, groups = outputs_aug_3_groups_0, pad = outputs_aug_3_pad_0, pad_type = outputs_aug_3_pad_type_0, strides = outputs_aug_3_strides_0, weight = inner_encoder_conv_module_1_sequential_4_conv_weight, x = inputs_15)[name = tensor("outputs_aug_3")]; + tensor input_67_begin_0 = const()[name = tensor("input_67_begin_0"), val = tensor([0, 0, 0])]; + tensor input_67_end_0 = const()[name = tensor("input_67_end_0"), val = tensor([1, 256, 16])]; + tensor input_67_end_mask_0 = const()[name = tensor("input_67_end_mask_0"), val = tensor([true, true, false])]; + tensor input_67 = slice_by_index(begin = input_67_begin_0, end = input_67_end_0, end_mask = input_67_end_mask_0, x = outputs_aug_3)[name = tensor("input_67")]; + tensor inputs_17 = batch_norm(beta = inner_encoder_conv_module_1_sequential_5_bias, epsilon = var_36, gamma = inner_encoder_conv_module_1_sequential_5_weight, mean = inner_encoder_conv_module_1_sequential_5_running_mean, variance = inner_encoder_conv_module_1_sequential_5_running_var, x = input_67)[name = tensor("inputs_17")]; + tensor input_69 = silu(x = inputs_17)[name = tensor("input_69")]; + tensor input_71_pad_type_0 = const()[name = tensor("input_71_pad_type_0"), val = tensor("valid")]; + tensor input_71_strides_0 = const()[name = tensor("input_71_strides_0"), val = tensor([1])]; + tensor input_71_pad_0 = const()[name = tensor("input_71_pad_0"), val = tensor([0, 0])]; + tensor input_71_dilations_0 = const()[name = tensor("input_71_dilations_0"), val = tensor([1])]; + tensor input_71_groups_0 = const()[name = tensor("input_71_groups_0"), val = tensor(1)]; + tensor input_71 = conv(bias = inner_encoder_conv_module_1_sequential_7_conv_bias, dilations = input_71_dilations_0, groups = input_71_groups_0, pad = input_71_pad_0, pad_type = input_71_pad_type_0, strides = input_71_strides_0, weight = inner_encoder_conv_module_1_sequential_7_conv_weight, x = input_69)[name = tensor("input_71")]; tensor conv_out_3_perm_0 = const()[name = tensor("conv_out_3_perm_0"), val = tensor([0, 2, 1])]; - tensor var_472_begin_0 = const()[name = tensor("op_472_begin_0"), val = tensor([0, -1, 0])]; - tensor var_472_end_0 = const()[name = tensor("op_472_end_0"), val = tensor([1, 16, 256])]; - tensor var_472_end_mask_0 = const()[name = tensor("op_472_end_mask_0"), val = tensor([true, true, true])]; - tensor conv_out_3 = transpose(perm = conv_out_3_perm_0, x = input_69)[name = tensor("transpose_46")]; - tensor var_472 = slice_by_index(begin = var_472_begin_0, end = var_472_end_0, end_mask = var_472_end_mask_0, x = conv_out_3)[name = tensor("op_472")]; - tensor var_474_perm_0 = const()[name = tensor("op_474_perm_0"), val = tensor([1, 0, 2])]; - tensor var_474 = transpose(perm = var_474_perm_0, x = var_472)[name = tensor("transpose_45")]; - tensor input_71 = add(x = x_9, y = var_474)[name = tensor("input_71")]; - tensor input_73_axes_0 = const()[name = tensor("input_73_axes_0"), val = tensor([-1])]; - tensor input_73 = layer_norm(axes = input_73_axes_0, beta = encoder_ffn2_1_module_sequential_0_bias, epsilon = var_28, gamma = encoder_ffn2_1_module_sequential_0_weight, x = input_71)[name = tensor("input_73")]; - tensor inputs_19 = linear(bias = encoder_ffn2_1_module_sequential_1_linear_bias, weight = encoder_ffn2_1_module_sequential_1_linear_weight, x = input_73)[name = tensor("linear_17")]; - tensor input_75 = silu(x = inputs_19)[name = tensor("input_75")]; - tensor input_79 = linear(bias = encoder_ffn2_1_module_sequential_4_linear_bias, weight = encoder_ffn2_1_module_sequential_4_linear_weight, x = input_75)[name = tensor("linear_18")]; - tensor var_497 = const()[name = tensor("op_497"), val = tensor(0x1p-1)]; - tensor var_498 = mul(x = input_79, y = var_497)[name = tensor("op_498")]; - tensor input_81 = add(x = var_498, y = input_71)[name = tensor("input_81")]; - tensor input_83_axes_0 = const()[name = tensor("input_83_axes_0"), val = tensor([-1])]; - tensor input_83 = layer_norm(axes = input_83_axes_0, beta = encoder_layer_norm_1_bias, epsilon = var_28, gamma = encoder_layer_norm_1_weight, x = input_81)[name = tensor("input_83")]; + tensor var_498_begin_0 = const()[name = tensor("op_498_begin_0"), val = tensor([0, -1, 0])]; + tensor var_498_end_0 = const()[name = tensor("op_498_end_0"), val = tensor([1, 16, 256])]; + tensor var_498_end_mask_0 = const()[name = tensor("op_498_end_mask_0"), val = tensor([true, true, true])]; + tensor conv_out_3 = transpose(perm = conv_out_3_perm_0, x = input_71)[name = tensor("transpose_46")]; + tensor var_498 = slice_by_index(begin = var_498_begin_0, end = var_498_end_0, end_mask = var_498_end_mask_0, x = conv_out_3)[name = tensor("op_498")]; + tensor var_500_perm_0 = const()[name = tensor("op_500_perm_0"), val = tensor([1, 0, 2])]; + tensor var_500 = transpose(perm = var_500_perm_0, x = var_498)[name = tensor("transpose_45")]; + tensor input_73 = add(x = x_9, y = var_500)[name = tensor("input_73")]; + tensor input_75_axes_0 = const()[name = tensor("input_75_axes_0"), val = tensor([-1])]; + tensor input_75 = layer_norm(axes = input_75_axes_0, beta = inner_encoder_ffn2_1_module_sequential_0_bias, epsilon = var_36, gamma = inner_encoder_ffn2_1_module_sequential_0_weight, x = input_73)[name = tensor("input_75")]; + tensor inputs_19 = linear(bias = inner_encoder_ffn2_1_module_sequential_1_linear_bias, weight = inner_encoder_ffn2_1_module_sequential_1_linear_weight, x = input_75)[name = tensor("linear_17")]; + tensor input_77 = silu(x = inputs_19)[name = tensor("input_77")]; + tensor input_81 = linear(bias = inner_encoder_ffn2_1_module_sequential_4_linear_bias, weight = inner_encoder_ffn2_1_module_sequential_4_linear_weight, x = input_77)[name = tensor("linear_18")]; + tensor var_523 = const()[name = tensor("op_523"), val = tensor(0x1p-1)]; + tensor var_524 = mul(x = input_81, y = var_523)[name = tensor("op_524")]; + tensor input_83 = add(x = var_524, y = input_73)[name = tensor("input_83")]; tensor input_85_axes_0 = const()[name = tensor("input_85_axes_0"), val = tensor([-1])]; - tensor input_85 = layer_norm(axes = input_85_axes_0, beta = encoder_ffn1_2_module_sequential_0_bias, epsilon = var_28, gamma = encoder_ffn1_2_module_sequential_0_weight, x = input_83)[name = tensor("input_85")]; - tensor inputs_21 = linear(bias = encoder_ffn1_2_module_sequential_1_linear_bias, weight = encoder_ffn1_2_module_sequential_1_linear_weight, x = input_85)[name = tensor("linear_19")]; - tensor input_87 = silu(x = inputs_21)[name = tensor("input_87")]; - tensor input_91 = linear(bias = encoder_ffn1_2_module_sequential_4_linear_bias, weight = encoder_ffn1_2_module_sequential_4_linear_weight, x = input_87)[name = tensor("linear_20")]; - tensor var_527 = const()[name = tensor("op_527"), val = tensor(0x1p-1)]; - tensor var_528 = mul(x = input_91, y = var_527)[name = tensor("op_528")]; - tensor input_93 = add(x = var_528, y = input_83)[name = tensor("input_93")]; + tensor input_85 = layer_norm(axes = input_85_axes_0, beta = inner_encoder_layer_norm_1_bias, epsilon = var_36, gamma = inner_encoder_layer_norm_1_weight, x = input_83)[name = tensor("input_85")]; + tensor input_87_axes_0 = const()[name = tensor("input_87_axes_0"), val = tensor([-1])]; + tensor input_87 = layer_norm(axes = input_87_axes_0, beta = inner_encoder_ffn1_2_module_sequential_0_bias, epsilon = var_36, gamma = inner_encoder_ffn1_2_module_sequential_0_weight, x = input_85)[name = tensor("input_87")]; + tensor inputs_21 = linear(bias = inner_encoder_ffn1_2_module_sequential_1_linear_bias, weight = inner_encoder_ffn1_2_module_sequential_1_linear_weight, x = input_87)[name = tensor("linear_19")]; + tensor input_89 = silu(x = inputs_21)[name = tensor("input_89")]; + tensor input_93 = linear(bias = inner_encoder_ffn1_2_module_sequential_4_linear_bias, weight = inner_encoder_ffn1_2_module_sequential_4_linear_weight, x = input_89)[name = tensor("linear_20")]; + tensor var_553 = const()[name = tensor("op_553"), val = tensor(0x1p-1)]; + tensor var_554 = mul(x = input_93, y = var_553)[name = tensor("op_554")]; + tensor input_95 = add(x = var_554, y = input_85)[name = tensor("input_95")]; tensor x_13_axes_0 = const()[name = tensor("x_13_axes_0"), val = tensor([-1])]; - tensor x_13 = layer_norm(axes = x_13_axes_0, beta = encoder_ret_lns_2_bias, epsilon = var_28, gamma = encoder_ret_lns_2_weight, x = input_93)[name = tensor("x_13")]; + tensor x_13 = layer_norm(axes = x_13_axes_0, beta = inner_encoder_ret_lns_2_bias, epsilon = var_36, gamma = inner_encoder_ret_lns_2_weight, x = input_95)[name = tensor("x_13")]; tensor prev_kv_5_begin_0 = const()[name = tensor("prev_kv_5_begin_0"), val = tensor([2, 0, 0, 0, 0])]; tensor prev_kv_5_end_0 = const()[name = tensor("prev_kv_5_end_0"), val = tensor([3, 1, 4, 64, 64])]; tensor prev_kv_5_end_mask_0 = const()[name = tensor("prev_kv_5_end_mask_0"), val = tensor([false, true, true, true, true])]; @@ -524,139 +530,139 @@ program(1.0) tensor prev_scale_5_end_mask_0 = const()[name = tensor("prev_scale_5_end_mask_0"), val = tensor([false, true])]; tensor prev_scale_5_squeeze_mask_0 = const()[name = tensor("prev_scale_5_squeeze_mask_0"), val = tensor([true, false])]; tensor prev_scale_5 = slice_by_index(begin = prev_scale_5_begin_0, end = prev_scale_5_end_0, end_mask = prev_scale_5_end_mask_0, squeeze_mask = prev_scale_5_squeeze_mask_0, x = enc_scale)[name = tensor("prev_scale_5")]; - tensor var_542 = linear(bias = encoder_q_proj_2_bias, weight = encoder_q_proj_2_weight, x = x_13)[name = tensor("linear_21")]; - tensor var_543 = const()[name = tensor("op_543"), val = tensor([1, 1, 4, 64])]; - tensor var_544 = reshape(shape = var_543, x = var_542)[name = tensor("op_544")]; + tensor var_568 = linear(bias = inner_encoder_q_proj_2_bias, weight = inner_encoder_q_proj_2_weight, x = x_13)[name = tensor("linear_21")]; + tensor var_569 = const()[name = tensor("op_569"), val = tensor([1, 1, 4, 64])]; + tensor var_570 = reshape(shape = var_569, x = var_568)[name = tensor("op_570")]; tensor q_5_perm_0 = const()[name = tensor("q_5_perm_0"), val = tensor([0, 2, 1, 3])]; - tensor var_548 = linear(bias = encoder_k_proj_2_bias, weight = encoder_k_proj_2_weight, x = x_13)[name = tensor("linear_22")]; - tensor var_549 = const()[name = tensor("op_549"), val = tensor(0x1p-3)]; - tensor var_550 = mul(x = var_548, y = var_549)[name = tensor("op_550")]; - tensor var_551 = const()[name = tensor("op_551"), val = tensor([1, 1, 4, 64])]; - tensor var_552 = reshape(shape = var_551, x = var_550)[name = tensor("op_552")]; + tensor var_574 = linear(bias = inner_encoder_k_proj_2_bias, weight = inner_encoder_k_proj_2_weight, x = x_13)[name = tensor("linear_22")]; + tensor var_575 = const()[name = tensor("op_575"), val = tensor(0x1p-3)]; + tensor var_576 = mul(x = var_574, y = var_575)[name = tensor("op_576")]; + tensor var_577 = const()[name = tensor("op_577"), val = tensor([1, 1, 4, 64])]; + tensor var_578 = reshape(shape = var_577, x = var_576)[name = tensor("op_578")]; tensor k_5_perm_0 = const()[name = tensor("k_5_perm_0"), val = tensor([0, 2, 1, 3])]; - tensor var_556 = linear(bias = encoder_v_proj_2_bias, weight = encoder_v_proj_2_weight, x = x_13)[name = tensor("linear_23")]; - tensor var_557 = const()[name = tensor("op_557"), val = tensor([1, 1, 4, 64])]; - tensor var_558 = reshape(shape = var_557, x = var_556)[name = tensor("op_558")]; + tensor var_582 = linear(bias = inner_encoder_v_proj_2_bias, weight = inner_encoder_v_proj_2_weight, x = x_13)[name = tensor("linear_23")]; + tensor var_583 = const()[name = tensor("op_583"), val = tensor([1, 1, 4, 64])]; + tensor var_584 = reshape(shape = var_583, x = var_582)[name = tensor("op_584")]; tensor v_5_perm_0 = const()[name = tensor("v_5_perm_0"), val = tensor([0, 2, 1, 3])]; - tensor input_97 = linear(bias = encoder_g_proj_2_bias, weight = encoder_g_proj_2_weight, x = x_13)[name = tensor("linear_24")]; + tensor input_99 = linear(bias = inner_encoder_g_proj_2_bias, weight = inner_encoder_g_proj_2_weight, x = x_13)[name = tensor("linear_24")]; tensor sqrt_s0_5 = sqrt(x = prev_scale_5)[name = tensor("sqrt_s0_5")]; - tensor s_t_5 = add(x = prev_scale_5, y = encoder__t_index)[name = tensor("s_t_5")]; + tensor s_t_5 = add(x = prev_scale_5, y = inner_encoder__t_index)[name = tensor("s_t_5")]; tensor sqrt_s_t_5 = sqrt(x = s_t_5)[name = tensor("sqrt_s_t_5")]; tensor qk_5_transpose_x_1 = const()[name = tensor("qk_5_transpose_x_1"), val = tensor(false)]; tensor qk_5_transpose_y_1 = const()[name = tensor("qk_5_transpose_y_1"), val = tensor(true)]; - tensor k_5 = transpose(perm = k_5_perm_0, x = var_552)[name = tensor("transpose_43")]; - tensor q_5 = transpose(perm = q_5_perm_0, x = var_544)[name = tensor("transpose_44")]; + tensor k_5 = transpose(perm = k_5_perm_0, x = var_578)[name = tensor("transpose_43")]; + tensor q_5 = transpose(perm = q_5_perm_0, x = var_570)[name = tensor("transpose_44")]; tensor qk_5 = matmul(transpose_x = qk_5_transpose_x_1, transpose_y = qk_5_transpose_y_1, x = q_5, y = k_5)[name = tensor("qk_5")]; - tensor var_568 = const()[name = tensor("op_568"), val = tensor([1, 1])]; - tensor var_569 = reshape(shape = var_568, x = sqrt_s_t_5)[name = tensor("op_569")]; - tensor M_5 = real_div(x = encoder__causal_mask, y = var_569)[name = tensor("M_5")]; - tensor var_571 = mul(x = qk_5, y = M_5)[name = tensor("op_571")]; + tensor var_594 = const()[name = tensor("op_594"), val = tensor([1, 1])]; + tensor var_595 = reshape(shape = var_594, x = sqrt_s_t_5)[name = tensor("op_595")]; + tensor M_5 = real_div(x = inner_encoder__causal_mask, y = var_595)[name = tensor("M_5")]; + tensor var_597 = mul(x = qk_5, y = M_5)[name = tensor("op_597")]; tensor inner_5_transpose_x_0 = const()[name = tensor("inner_5_transpose_x_0"), val = tensor(false)]; tensor inner_5_transpose_y_0 = const()[name = tensor("inner_5_transpose_y_0"), val = tensor(false)]; - tensor v_5 = transpose(perm = v_5_perm_0, x = var_558)[name = tensor("transpose_42")]; - tensor inner_5 = matmul(transpose_x = inner_5_transpose_x_0, transpose_y = inner_5_transpose_y_0, x = var_571, y = v_5)[name = tensor("inner_5")]; - tensor var_573_transpose_x_0 = const()[name = tensor("op_573_transpose_x_0"), val = tensor(false)]; - tensor var_573_transpose_y_0 = const()[name = tensor("op_573_transpose_y_0"), val = tensor(false)]; - tensor var_573 = matmul(transpose_x = var_573_transpose_x_0, transpose_y = var_573_transpose_y_0, x = q_5, y = prev_kv_5)[name = tensor("op_573")]; - tensor var_574 = real_div(x = sqrt_s0_5, y = sqrt_s_t_5)[name = tensor("op_574")]; - tensor var_575 = const()[name = tensor("op_575"), val = tensor([1, 1, 1, 1])]; - tensor var_576 = reshape(shape = var_575, x = var_574)[name = tensor("op_576")]; - tensor cross_5 = mul(x = var_573, y = var_576)[name = tensor("cross_5")]; + tensor v_5 = transpose(perm = v_5_perm_0, x = var_584)[name = tensor("transpose_42")]; + tensor inner_5 = matmul(transpose_x = inner_5_transpose_x_0, transpose_y = inner_5_transpose_y_0, x = var_597, y = v_5)[name = tensor("inner_5")]; + tensor var_599_transpose_x_0 = const()[name = tensor("op_599_transpose_x_0"), val = tensor(false)]; + tensor var_599_transpose_y_0 = const()[name = tensor("op_599_transpose_y_0"), val = tensor(false)]; + tensor var_599 = matmul(transpose_x = var_599_transpose_x_0, transpose_y = var_599_transpose_y_0, x = q_5, y = prev_kv_5)[name = tensor("op_599")]; + tensor var_600 = real_div(x = sqrt_s0_5, y = sqrt_s_t_5)[name = tensor("op_600")]; + tensor var_601 = const()[name = tensor("op_601"), val = tensor([1, 1, 1, 1])]; + tensor var_602 = reshape(shape = var_601, x = var_600)[name = tensor("op_602")]; + tensor cross_5 = mul(x = var_599, y = var_602)[name = tensor("cross_5")]; tensor out_13 = add(x = inner_5, y = cross_5)[name = tensor("out_13")]; - tensor var_579 = mul(x = prev_kv_5, y = sqrt_s0_5)[name = tensor("op_579")]; - tensor var_581_transpose_x_1 = const()[name = tensor("op_581_transpose_x_1"), val = tensor(true)]; - tensor var_581_transpose_y_1 = const()[name = tensor("op_581_transpose_y_1"), val = tensor(false)]; - tensor var_581 = matmul(transpose_x = var_581_transpose_x_1, transpose_y = var_581_transpose_y_1, x = k_5, y = v_5)[name = tensor("op_581")]; - tensor new_kv_unnorm_5 = add(x = var_579, y = var_581)[name = tensor("new_kv_unnorm_5")]; - tensor var_583 = const()[name = tensor("op_583"), val = tensor(0x1p+0)]; - tensor new_scale_5 = add(x = prev_scale_5, y = var_583)[name = tensor("new_scale_5")]; - tensor var_585 = sqrt(x = new_scale_5)[name = tensor("op_585")]; - tensor var_586 = real_div(x = new_kv_unnorm_5, y = var_585)[name = tensor("op_586")]; - tensor var_587_perm_0 = const()[name = tensor("op_587_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_605 = mul(x = prev_kv_5, y = sqrt_s0_5)[name = tensor("op_605")]; + tensor var_607_transpose_x_1 = const()[name = tensor("op_607_transpose_x_1"), val = tensor(true)]; + tensor var_607_transpose_y_1 = const()[name = tensor("op_607_transpose_y_1"), val = tensor(false)]; + tensor var_607 = matmul(transpose_x = var_607_transpose_x_1, transpose_y = var_607_transpose_y_1, x = k_5, y = v_5)[name = tensor("op_607")]; + tensor new_kv_unnorm_5 = add(x = var_605, y = var_607)[name = tensor("new_kv_unnorm_5")]; + tensor var_609 = const()[name = tensor("op_609"), val = tensor(0x1p+0)]; + tensor new_scale_5 = add(x = prev_scale_5, y = var_609)[name = tensor("new_scale_5")]; + tensor var_611 = sqrt(x = new_scale_5)[name = tensor("op_611")]; + tensor var_612 = real_div(x = new_kv_unnorm_5, y = var_611)[name = tensor("op_612")]; + tensor var_613_perm_0 = const()[name = tensor("op_613_perm_0"), val = tensor([0, 2, 1, 3])]; tensor out_15_axes_0 = const()[name = tensor("out_15_axes_0"), val = tensor([-1])]; - tensor var_587 = transpose(perm = var_587_perm_0, x = out_13)[name = tensor("transpose_41")]; - tensor out_15 = layer_norm(axes = out_15_axes_0, epsilon = var_17, x = var_587)[name = tensor("out_15")]; - tensor var_591 = const()[name = tensor("op_591"), val = tensor([1, 1, 256])]; - tensor out_17 = reshape(shape = var_591, x = out_15)[name = tensor("out_17")]; - tensor var_593 = silu(x = input_97)[name = tensor("op_593")]; - tensor input_99 = mul(x = var_593, y = out_17)[name = tensor("input_99")]; - tensor ret_out_5 = linear(bias = encoder_out_proj_2_bias, weight = encoder_out_proj_2_weight, x = input_99)[name = tensor("linear_25")]; - tensor x_15 = add(x = input_93, y = ret_out_5)[name = tensor("x_15")]; + tensor var_613 = transpose(perm = var_613_perm_0, x = out_13)[name = tensor("transpose_41")]; + tensor out_15 = layer_norm(axes = out_15_axes_0, epsilon = var_44, x = var_613)[name = tensor("out_15")]; + tensor var_617 = const()[name = tensor("op_617"), val = tensor([1, 1, 256])]; + tensor out_17 = reshape(shape = var_617, x = out_15)[name = tensor("out_17")]; + tensor var_619 = silu(x = input_99)[name = tensor("op_619")]; + tensor input_101 = mul(x = var_619, y = out_17)[name = tensor("input_101")]; + tensor ret_out_5 = linear(bias = inner_encoder_out_proj_2_bias, weight = inner_encoder_out_proj_2_weight, x = input_101)[name = tensor("linear_25")]; + tensor x_15 = add(x = input_95, y = ret_out_5)[name = tensor("x_15")]; tensor window_9_begin_0 = const()[name = tensor("window_9_begin_0"), val = tensor([2, 0, 0, 0])]; tensor window_9_end_0 = const()[name = tensor("window_9_end_0"), val = tensor([3, 1, 16, 256])]; tensor window_9_end_mask_0 = const()[name = tensor("window_9_end_mask_0"), val = tensor([false, true, true, true])]; tensor window_9_squeeze_mask_0 = const()[name = tensor("window_9_squeeze_mask_0"), val = tensor([true, false, false, false])]; tensor window_9 = slice_by_index(begin = window_9_begin_0, end = window_9_end_0, end_mask = window_9_end_mask_0, squeeze_mask = window_9_squeeze_mask_0, x = enc_conv_cache)[name = tensor("window_9")]; - tensor var_604_begin_0 = const()[name = tensor("op_604_begin_0"), val = tensor([0, 1, 0])]; - tensor var_604_end_0 = const()[name = tensor("op_604_end_0"), val = tensor([1, 16, 256])]; - tensor var_604_end_mask_0 = const()[name = tensor("op_604_end_mask_0"), val = tensor([true, true, true])]; - tensor var_604 = slice_by_index(begin = var_604_begin_0, end = var_604_end_0, end_mask = var_604_end_mask_0, x = window_9)[name = tensor("op_604")]; + tensor var_630_begin_0 = const()[name = tensor("op_630_begin_0"), val = tensor([0, 1, 0])]; + tensor var_630_end_0 = const()[name = tensor("op_630_end_0"), val = tensor([1, 16, 256])]; + tensor var_630_end_mask_0 = const()[name = tensor("op_630_end_mask_0"), val = tensor([true, true, true])]; + tensor var_630 = slice_by_index(begin = var_630_begin_0, end = var_630_end_0, end_mask = var_630_end_mask_0, x = window_9)[name = tensor("op_630")]; tensor window_11_interleave_0 = const()[name = tensor("window_11_interleave_0"), val = tensor(false)]; - tensor window_11 = concat(axis = var_26, interleave = window_11_interleave_0, values = (var_604, x_15))[name = tensor("window_11")]; - tensor input_101_interleave_0 = const()[name = tensor("input_101_interleave_0"), val = tensor(false)]; - tensor input_101 = concat(axis = var_23, interleave = input_101_interleave_0, values = window_11)[name = tensor("input_101")]; + tensor window_11 = concat(axis = var_52, interleave = window_11_interleave_0, values = (var_630, x_15))[name = tensor("window_11")]; + tensor input_103_interleave_0 = const()[name = tensor("input_103_interleave_0"), val = tensor(false)]; + tensor input_103 = concat(axis = var_39, interleave = input_103_interleave_0, values = window_11)[name = tensor("input_103")]; tensor x_17_axes_0 = const()[name = tensor("x_17_axes_0"), val = tensor([-1])]; - tensor x_17 = layer_norm(axes = x_17_axes_0, beta = encoder_conv_module_2_sequential_0_bias, epsilon = var_28, gamma = encoder_conv_module_2_sequential_0_weight, x = input_101)[name = tensor("x_17")]; - tensor input_103_perm_0 = const()[name = tensor("input_103_perm_0"), val = tensor([0, 2, 1])]; + tensor x_17 = layer_norm(axes = x_17_axes_0, beta = inner_encoder_conv_module_2_sequential_0_bias, epsilon = var_36, gamma = inner_encoder_conv_module_2_sequential_0_weight, x = input_103)[name = tensor("x_17")]; + tensor input_105_perm_0 = const()[name = tensor("input_105_perm_0"), val = tensor([0, 2, 1])]; tensor inputs_23_pad_type_0 = const()[name = tensor("inputs_23_pad_type_0"), val = tensor("valid")]; tensor inputs_23_strides_0 = const()[name = tensor("inputs_23_strides_0"), val = tensor([1])]; tensor inputs_23_pad_0 = const()[name = tensor("inputs_23_pad_0"), val = tensor([0, 0])]; tensor inputs_23_dilations_0 = const()[name = tensor("inputs_23_dilations_0"), val = tensor([1])]; tensor inputs_23_groups_0 = const()[name = tensor("inputs_23_groups_0"), val = tensor(1)]; - tensor input_103 = transpose(perm = input_103_perm_0, x = x_17)[name = tensor("transpose_40")]; - tensor inputs_23 = conv(bias = encoder_conv_module_2_sequential_2_conv_bias, dilations = inputs_23_dilations_0, groups = inputs_23_groups_0, pad = inputs_23_pad_0, pad_type = inputs_23_pad_type_0, strides = inputs_23_strides_0, weight = encoder_conv_module_2_sequential_2_conv_weight, x = input_103)[name = tensor("inputs_23")]; - tensor var_629_split_sizes_0 = const()[name = tensor("op_629_split_sizes_0"), val = tensor([256, 256])]; - tensor var_629_axis_0 = const()[name = tensor("op_629_axis_0"), val = tensor(1)]; - tensor var_629_0, tensor var_629_1 = split(axis = var_629_axis_0, split_sizes = var_629_split_sizes_0, x = inputs_23)[name = tensor("op_629")]; - tensor var_631 = sigmoid(x = var_629_1)[name = tensor("op_631")]; - tensor inputs_25 = mul(x = var_629_0, y = var_631)[name = tensor("inputs_25")]; + tensor input_105 = transpose(perm = input_105_perm_0, x = x_17)[name = tensor("transpose_40")]; + tensor inputs_23 = conv(bias = inner_encoder_conv_module_2_sequential_2_conv_bias, dilations = inputs_23_dilations_0, groups = inputs_23_groups_0, pad = inputs_23_pad_0, pad_type = inputs_23_pad_type_0, strides = inputs_23_strides_0, weight = inner_encoder_conv_module_2_sequential_2_conv_weight, x = input_105)[name = tensor("inputs_23")]; + tensor var_655_split_sizes_0 = const()[name = tensor("op_655_split_sizes_0"), val = tensor([256, 256])]; + tensor var_655_axis_0 = const()[name = tensor("op_655_axis_0"), val = tensor(1)]; + tensor var_655_0, tensor var_655_1 = split(axis = var_655_axis_0, split_sizes = var_655_split_sizes_0, x = inputs_23)[name = tensor("op_655")]; + tensor var_657 = sigmoid(x = var_655_1)[name = tensor("op_657")]; + tensor inputs_25 = mul(x = var_655_0, y = var_657)[name = tensor("inputs_25")]; tensor outputs_aug_5_pad_type_0 = const()[name = tensor("outputs_aug_5_pad_type_0"), val = tensor("custom")]; tensor outputs_aug_5_pad_0 = const()[name = tensor("outputs_aug_5_pad_0"), val = tensor([15, 15])]; tensor outputs_aug_5_groups_0 = const()[name = tensor("outputs_aug_5_groups_0"), val = tensor(256)]; tensor outputs_aug_5_strides_0 = const()[name = tensor("outputs_aug_5_strides_0"), val = tensor([1])]; tensor outputs_aug_5_dilations_0 = const()[name = tensor("outputs_aug_5_dilations_0"), val = tensor([1])]; - tensor outputs_aug_5 = conv(dilations = outputs_aug_5_dilations_0, groups = outputs_aug_5_groups_0, pad = outputs_aug_5_pad_0, pad_type = outputs_aug_5_pad_type_0, strides = outputs_aug_5_strides_0, weight = encoder_conv_module_2_sequential_4_conv_weight, x = inputs_25)[name = tensor("outputs_aug_5")]; - tensor input_105_begin_0 = const()[name = tensor("input_105_begin_0"), val = tensor([0, 0, 0])]; - tensor input_105_end_0 = const()[name = tensor("input_105_end_0"), val = tensor([1, 256, 16])]; - tensor input_105_end_mask_0 = const()[name = tensor("input_105_end_mask_0"), val = tensor([true, true, false])]; - tensor input_105 = slice_by_index(begin = input_105_begin_0, end = input_105_end_0, end_mask = input_105_end_mask_0, x = outputs_aug_5)[name = tensor("input_105")]; - tensor inputs_27 = batch_norm(beta = encoder_conv_module_2_sequential_5_bias, epsilon = var_28, gamma = encoder_conv_module_2_sequential_5_weight, mean = encoder_conv_module_2_sequential_5_running_mean, variance = encoder_conv_module_2_sequential_5_running_var, x = input_105)[name = tensor("inputs_27")]; - tensor input_107 = silu(x = inputs_27)[name = tensor("input_107")]; - tensor input_109_pad_type_0 = const()[name = tensor("input_109_pad_type_0"), val = tensor("valid")]; - tensor input_109_strides_0 = const()[name = tensor("input_109_strides_0"), val = tensor([1])]; - tensor input_109_pad_0 = const()[name = tensor("input_109_pad_0"), val = tensor([0, 0])]; - tensor input_109_dilations_0 = const()[name = tensor("input_109_dilations_0"), val = tensor([1])]; - tensor input_109_groups_0 = const()[name = tensor("input_109_groups_0"), val = tensor(1)]; - tensor input_109 = conv(bias = encoder_conv_module_2_sequential_7_conv_bias, dilations = input_109_dilations_0, groups = input_109_groups_0, pad = input_109_pad_0, pad_type = input_109_pad_type_0, strides = input_109_strides_0, weight = encoder_conv_module_2_sequential_7_conv_weight, x = input_107)[name = tensor("input_109")]; + tensor outputs_aug_5 = conv(dilations = outputs_aug_5_dilations_0, groups = outputs_aug_5_groups_0, pad = outputs_aug_5_pad_0, pad_type = outputs_aug_5_pad_type_0, strides = outputs_aug_5_strides_0, weight = inner_encoder_conv_module_2_sequential_4_conv_weight, x = inputs_25)[name = tensor("outputs_aug_5")]; + tensor input_107_begin_0 = const()[name = tensor("input_107_begin_0"), val = tensor([0, 0, 0])]; + tensor input_107_end_0 = const()[name = tensor("input_107_end_0"), val = tensor([1, 256, 16])]; + tensor input_107_end_mask_0 = const()[name = tensor("input_107_end_mask_0"), val = tensor([true, true, false])]; + tensor input_107 = slice_by_index(begin = input_107_begin_0, end = input_107_end_0, end_mask = input_107_end_mask_0, x = outputs_aug_5)[name = tensor("input_107")]; + tensor inputs_27 = batch_norm(beta = inner_encoder_conv_module_2_sequential_5_bias, epsilon = var_36, gamma = inner_encoder_conv_module_2_sequential_5_weight, mean = inner_encoder_conv_module_2_sequential_5_running_mean, variance = inner_encoder_conv_module_2_sequential_5_running_var, x = input_107)[name = tensor("inputs_27")]; + tensor input_109 = silu(x = inputs_27)[name = tensor("input_109")]; + tensor input_111_pad_type_0 = const()[name = tensor("input_111_pad_type_0"), val = tensor("valid")]; + tensor input_111_strides_0 = const()[name = tensor("input_111_strides_0"), val = tensor([1])]; + tensor input_111_pad_0 = const()[name = tensor("input_111_pad_0"), val = tensor([0, 0])]; + tensor input_111_dilations_0 = const()[name = tensor("input_111_dilations_0"), val = tensor([1])]; + tensor input_111_groups_0 = const()[name = tensor("input_111_groups_0"), val = tensor(1)]; + tensor input_111 = conv(bias = inner_encoder_conv_module_2_sequential_7_conv_bias, dilations = input_111_dilations_0, groups = input_111_groups_0, pad = input_111_pad_0, pad_type = input_111_pad_type_0, strides = input_111_strides_0, weight = inner_encoder_conv_module_2_sequential_7_conv_weight, x = input_109)[name = tensor("input_111")]; tensor conv_out_5_perm_0 = const()[name = tensor("conv_out_5_perm_0"), val = tensor([0, 2, 1])]; - tensor var_662_begin_0 = const()[name = tensor("op_662_begin_0"), val = tensor([0, -1, 0])]; - tensor var_662_end_0 = const()[name = tensor("op_662_end_0"), val = tensor([1, 16, 256])]; - tensor var_662_end_mask_0 = const()[name = tensor("op_662_end_mask_0"), val = tensor([true, true, true])]; - tensor conv_out_5 = transpose(perm = conv_out_5_perm_0, x = input_109)[name = tensor("transpose_39")]; - tensor var_662 = slice_by_index(begin = var_662_begin_0, end = var_662_end_0, end_mask = var_662_end_mask_0, x = conv_out_5)[name = tensor("op_662")]; - tensor var_664_perm_0 = const()[name = tensor("op_664_perm_0"), val = tensor([1, 0, 2])]; - tensor var_664 = transpose(perm = var_664_perm_0, x = var_662)[name = tensor("transpose_38")]; - tensor input_111 = add(x = x_15, y = var_664)[name = tensor("input_111")]; - tensor input_113_axes_0 = const()[name = tensor("input_113_axes_0"), val = tensor([-1])]; - tensor input_113 = layer_norm(axes = input_113_axes_0, beta = encoder_ffn2_2_module_sequential_0_bias, epsilon = var_28, gamma = encoder_ffn2_2_module_sequential_0_weight, x = input_111)[name = tensor("input_113")]; - tensor inputs_29 = linear(bias = encoder_ffn2_2_module_sequential_1_linear_bias, weight = encoder_ffn2_2_module_sequential_1_linear_weight, x = input_113)[name = tensor("linear_26")]; - tensor input_115 = silu(x = inputs_29)[name = tensor("input_115")]; - tensor input_119 = linear(bias = encoder_ffn2_2_module_sequential_4_linear_bias, weight = encoder_ffn2_2_module_sequential_4_linear_weight, x = input_115)[name = tensor("linear_27")]; - tensor var_687 = const()[name = tensor("op_687"), val = tensor(0x1p-1)]; - tensor var_688 = mul(x = input_119, y = var_687)[name = tensor("op_688")]; - tensor input_121 = add(x = var_688, y = input_111)[name = tensor("input_121")]; - tensor input_123_axes_0 = const()[name = tensor("input_123_axes_0"), val = tensor([-1])]; - tensor input_123 = layer_norm(axes = input_123_axes_0, beta = encoder_layer_norm_2_bias, epsilon = var_28, gamma = encoder_layer_norm_2_weight, x = input_121)[name = tensor("input_123")]; + tensor var_688_begin_0 = const()[name = tensor("op_688_begin_0"), val = tensor([0, -1, 0])]; + tensor var_688_end_0 = const()[name = tensor("op_688_end_0"), val = tensor([1, 16, 256])]; + tensor var_688_end_mask_0 = const()[name = tensor("op_688_end_mask_0"), val = tensor([true, true, true])]; + tensor conv_out_5 = transpose(perm = conv_out_5_perm_0, x = input_111)[name = tensor("transpose_39")]; + tensor var_688 = slice_by_index(begin = var_688_begin_0, end = var_688_end_0, end_mask = var_688_end_mask_0, x = conv_out_5)[name = tensor("op_688")]; + tensor var_690_perm_0 = const()[name = tensor("op_690_perm_0"), val = tensor([1, 0, 2])]; + tensor var_690 = transpose(perm = var_690_perm_0, x = var_688)[name = tensor("transpose_38")]; + tensor input_113 = add(x = x_15, y = var_690)[name = tensor("input_113")]; + tensor input_115_axes_0 = const()[name = tensor("input_115_axes_0"), val = tensor([-1])]; + tensor input_115 = layer_norm(axes = input_115_axes_0, beta = inner_encoder_ffn2_2_module_sequential_0_bias, epsilon = var_36, gamma = inner_encoder_ffn2_2_module_sequential_0_weight, x = input_113)[name = tensor("input_115")]; + tensor inputs_29 = linear(bias = inner_encoder_ffn2_2_module_sequential_1_linear_bias, weight = inner_encoder_ffn2_2_module_sequential_1_linear_weight, x = input_115)[name = tensor("linear_26")]; + tensor input_117 = silu(x = inputs_29)[name = tensor("input_117")]; + tensor input_121 = linear(bias = inner_encoder_ffn2_2_module_sequential_4_linear_bias, weight = inner_encoder_ffn2_2_module_sequential_4_linear_weight, x = input_117)[name = tensor("linear_27")]; + tensor var_713 = const()[name = tensor("op_713"), val = tensor(0x1p-1)]; + tensor var_714 = mul(x = input_121, y = var_713)[name = tensor("op_714")]; + tensor input_123 = add(x = var_714, y = input_113)[name = tensor("input_123")]; tensor input_125_axes_0 = const()[name = tensor("input_125_axes_0"), val = tensor([-1])]; - tensor input_125 = layer_norm(axes = input_125_axes_0, beta = encoder_ffn1_3_module_sequential_0_bias, epsilon = var_28, gamma = encoder_ffn1_3_module_sequential_0_weight, x = input_123)[name = tensor("input_125")]; - tensor inputs_31 = linear(bias = encoder_ffn1_3_module_sequential_1_linear_bias, weight = encoder_ffn1_3_module_sequential_1_linear_weight, x = input_125)[name = tensor("linear_28")]; - tensor input_127 = silu(x = inputs_31)[name = tensor("input_127")]; - tensor input_131 = linear(bias = encoder_ffn1_3_module_sequential_4_linear_bias, weight = encoder_ffn1_3_module_sequential_4_linear_weight, x = input_127)[name = tensor("linear_29")]; - tensor var_717 = const()[name = tensor("op_717"), val = tensor(0x1p-1)]; - tensor var_718 = mul(x = input_131, y = var_717)[name = tensor("op_718")]; - tensor input_133 = add(x = var_718, y = input_123)[name = tensor("input_133")]; + tensor input_125 = layer_norm(axes = input_125_axes_0, beta = inner_encoder_layer_norm_2_bias, epsilon = var_36, gamma = inner_encoder_layer_norm_2_weight, x = input_123)[name = tensor("input_125")]; + tensor input_127_axes_0 = const()[name = tensor("input_127_axes_0"), val = tensor([-1])]; + tensor input_127 = layer_norm(axes = input_127_axes_0, beta = inner_encoder_ffn1_3_module_sequential_0_bias, epsilon = var_36, gamma = inner_encoder_ffn1_3_module_sequential_0_weight, x = input_125)[name = tensor("input_127")]; + tensor inputs_31 = linear(bias = inner_encoder_ffn1_3_module_sequential_1_linear_bias, weight = inner_encoder_ffn1_3_module_sequential_1_linear_weight, x = input_127)[name = tensor("linear_28")]; + tensor input_129 = silu(x = inputs_31)[name = tensor("input_129")]; + tensor input_133 = linear(bias = inner_encoder_ffn1_3_module_sequential_4_linear_bias, weight = inner_encoder_ffn1_3_module_sequential_4_linear_weight, x = input_129)[name = tensor("linear_29")]; + tensor var_743 = const()[name = tensor("op_743"), val = tensor(0x1p-1)]; + tensor var_744 = mul(x = input_133, y = var_743)[name = tensor("op_744")]; + tensor input_135 = add(x = var_744, y = input_125)[name = tensor("input_135")]; tensor x_19_axes_0 = const()[name = tensor("x_19_axes_0"), val = tensor([-1])]; - tensor x_19 = layer_norm(axes = x_19_axes_0, beta = encoder_ret_lns_3_bias, epsilon = var_28, gamma = encoder_ret_lns_3_weight, x = input_133)[name = tensor("x_19")]; + tensor x_19 = layer_norm(axes = x_19_axes_0, beta = inner_encoder_ret_lns_3_bias, epsilon = var_36, gamma = inner_encoder_ret_lns_3_weight, x = input_135)[name = tensor("x_19")]; tensor prev_kv_7_begin_0 = const()[name = tensor("prev_kv_7_begin_0"), val = tensor([3, 0, 0, 0, 0])]; tensor prev_kv_7_end_0 = const()[name = tensor("prev_kv_7_end_0"), val = tensor([4, 1, 4, 64, 64])]; tensor prev_kv_7_end_mask_0 = const()[name = tensor("prev_kv_7_end_mask_0"), val = tensor([false, true, true, true, true])]; @@ -667,175 +673,168 @@ program(1.0) tensor prev_scale_7_end_mask_0 = const()[name = tensor("prev_scale_7_end_mask_0"), val = tensor([false, true])]; tensor prev_scale_7_squeeze_mask_0 = const()[name = tensor("prev_scale_7_squeeze_mask_0"), val = tensor([true, false])]; tensor prev_scale_7 = slice_by_index(begin = prev_scale_7_begin_0, end = prev_scale_7_end_0, end_mask = prev_scale_7_end_mask_0, squeeze_mask = prev_scale_7_squeeze_mask_0, x = enc_scale)[name = tensor("prev_scale_7")]; - tensor var_732 = linear(bias = encoder_q_proj_3_bias, weight = encoder_q_proj_3_weight, x = x_19)[name = tensor("linear_30")]; - tensor var_733 = const()[name = tensor("op_733"), val = tensor([1, 1, 4, 64])]; - tensor var_734 = reshape(shape = var_733, x = var_732)[name = tensor("op_734")]; + tensor var_758 = linear(bias = inner_encoder_q_proj_3_bias, weight = inner_encoder_q_proj_3_weight, x = x_19)[name = tensor("linear_30")]; + tensor var_759 = const()[name = tensor("op_759"), val = tensor([1, 1, 4, 64])]; + tensor var_760 = reshape(shape = var_759, x = var_758)[name = tensor("op_760")]; tensor q_7_perm_0 = const()[name = tensor("q_7_perm_0"), val = tensor([0, 2, 1, 3])]; - tensor var_738 = linear(bias = encoder_k_proj_3_bias, weight = encoder_k_proj_3_weight, x = x_19)[name = tensor("linear_31")]; - tensor var_739 = const()[name = tensor("op_739"), val = tensor(0x1p-3)]; - tensor var_740 = mul(x = var_738, y = var_739)[name = tensor("op_740")]; - tensor var_741 = const()[name = tensor("op_741"), val = tensor([1, 1, 4, 64])]; - tensor var_742 = reshape(shape = var_741, x = var_740)[name = tensor("op_742")]; + tensor var_764 = linear(bias = inner_encoder_k_proj_3_bias, weight = inner_encoder_k_proj_3_weight, x = x_19)[name = tensor("linear_31")]; + tensor var_765 = const()[name = tensor("op_765"), val = tensor(0x1p-3)]; + tensor var_766 = mul(x = var_764, y = var_765)[name = tensor("op_766")]; + tensor var_767 = const()[name = tensor("op_767"), val = tensor([1, 1, 4, 64])]; + tensor var_768 = reshape(shape = var_767, x = var_766)[name = tensor("op_768")]; tensor k_7_perm_0 = const()[name = tensor("k_7_perm_0"), val = tensor([0, 2, 1, 3])]; - tensor var_746 = linear(bias = encoder_v_proj_3_bias, weight = encoder_v_proj_3_weight, x = x_19)[name = tensor("linear_32")]; - tensor var_747 = const()[name = tensor("op_747"), val = tensor([1, 1, 4, 64])]; - tensor var_748 = reshape(shape = var_747, x = var_746)[name = tensor("op_748")]; + tensor var_772 = linear(bias = inner_encoder_v_proj_3_bias, weight = inner_encoder_v_proj_3_weight, x = x_19)[name = tensor("linear_32")]; + tensor var_773 = const()[name = tensor("op_773"), val = tensor([1, 1, 4, 64])]; + tensor var_774 = reshape(shape = var_773, x = var_772)[name = tensor("op_774")]; tensor v_7_perm_0 = const()[name = tensor("v_7_perm_0"), val = tensor([0, 2, 1, 3])]; - tensor input_137 = linear(bias = encoder_g_proj_3_bias, weight = encoder_g_proj_3_weight, x = x_19)[name = tensor("linear_33")]; + tensor input_139 = linear(bias = inner_encoder_g_proj_3_bias, weight = inner_encoder_g_proj_3_weight, x = x_19)[name = tensor("linear_33")]; tensor sqrt_s0_7 = sqrt(x = prev_scale_7)[name = tensor("sqrt_s0_7")]; - tensor s_t_7 = add(x = prev_scale_7, y = encoder__t_index)[name = tensor("s_t_7")]; + tensor s_t_7 = add(x = prev_scale_7, y = inner_encoder__t_index)[name = tensor("s_t_7")]; tensor sqrt_s_t_7 = sqrt(x = s_t_7)[name = tensor("sqrt_s_t_7")]; tensor qk_7_transpose_x_1 = const()[name = tensor("qk_7_transpose_x_1"), val = tensor(false)]; tensor qk_7_transpose_y_1 = const()[name = tensor("qk_7_transpose_y_1"), val = tensor(true)]; - tensor k_7 = transpose(perm = k_7_perm_0, x = var_742)[name = tensor("transpose_36")]; - tensor q_7 = transpose(perm = q_7_perm_0, x = var_734)[name = tensor("transpose_37")]; + tensor k_7 = transpose(perm = k_7_perm_0, x = var_768)[name = tensor("transpose_36")]; + tensor q_7 = transpose(perm = q_7_perm_0, x = var_760)[name = tensor("transpose_37")]; tensor qk_7 = matmul(transpose_x = qk_7_transpose_x_1, transpose_y = qk_7_transpose_y_1, x = q_7, y = k_7)[name = tensor("qk_7")]; - tensor var_758 = const()[name = tensor("op_758"), val = tensor([1, 1])]; - tensor var_759 = reshape(shape = var_758, x = sqrt_s_t_7)[name = tensor("op_759")]; - tensor M_7 = real_div(x = encoder__causal_mask, y = var_759)[name = tensor("M_7")]; - tensor var_761 = mul(x = qk_7, y = M_7)[name = tensor("op_761")]; + tensor var_784 = const()[name = tensor("op_784"), val = tensor([1, 1])]; + tensor var_785 = reshape(shape = var_784, x = sqrt_s_t_7)[name = tensor("op_785")]; + tensor M_7 = real_div(x = inner_encoder__causal_mask, y = var_785)[name = tensor("M_7")]; + tensor var_787 = mul(x = qk_7, y = M_7)[name = tensor("op_787")]; tensor inner_7_transpose_x_0 = const()[name = tensor("inner_7_transpose_x_0"), val = tensor(false)]; tensor inner_7_transpose_y_0 = const()[name = tensor("inner_7_transpose_y_0"), val = tensor(false)]; - tensor v_7 = transpose(perm = v_7_perm_0, x = var_748)[name = tensor("transpose_35")]; - tensor inner_7 = matmul(transpose_x = inner_7_transpose_x_0, transpose_y = inner_7_transpose_y_0, x = var_761, y = v_7)[name = tensor("inner_7")]; - tensor var_763_transpose_x_0 = const()[name = tensor("op_763_transpose_x_0"), val = tensor(false)]; - tensor var_763_transpose_y_0 = const()[name = tensor("op_763_transpose_y_0"), val = tensor(false)]; - tensor var_763 = matmul(transpose_x = var_763_transpose_x_0, transpose_y = var_763_transpose_y_0, x = q_7, y = prev_kv_7)[name = tensor("op_763")]; - tensor var_764 = real_div(x = sqrt_s0_7, y = sqrt_s_t_7)[name = tensor("op_764")]; - tensor var_765 = const()[name = tensor("op_765"), val = tensor([1, 1, 1, 1])]; - tensor var_766 = reshape(shape = var_765, x = var_764)[name = tensor("op_766")]; - tensor cross_7 = mul(x = var_763, y = var_766)[name = tensor("cross_7")]; + tensor v_7 = transpose(perm = v_7_perm_0, x = var_774)[name = tensor("transpose_35")]; + tensor inner_7 = matmul(transpose_x = inner_7_transpose_x_0, transpose_y = inner_7_transpose_y_0, x = var_787, y = v_7)[name = tensor("inner_7")]; + tensor var_789_transpose_x_0 = const()[name = tensor("op_789_transpose_x_0"), val = tensor(false)]; + tensor var_789_transpose_y_0 = const()[name = tensor("op_789_transpose_y_0"), val = tensor(false)]; + tensor var_789 = matmul(transpose_x = var_789_transpose_x_0, transpose_y = var_789_transpose_y_0, x = q_7, y = prev_kv_7)[name = tensor("op_789")]; + tensor var_790 = real_div(x = sqrt_s0_7, y = sqrt_s_t_7)[name = tensor("op_790")]; + tensor var_791 = const()[name = tensor("op_791"), val = tensor([1, 1, 1, 1])]; + tensor var_792 = reshape(shape = var_791, x = var_790)[name = tensor("op_792")]; + tensor cross_7 = mul(x = var_789, y = var_792)[name = tensor("cross_7")]; tensor out_19 = add(x = inner_7, y = cross_7)[name = tensor("out_19")]; - tensor var_769 = mul(x = prev_kv_7, y = sqrt_s0_7)[name = tensor("op_769")]; - tensor var_771_transpose_x_1 = const()[name = tensor("op_771_transpose_x_1"), val = tensor(true)]; - tensor var_771_transpose_y_1 = const()[name = tensor("op_771_transpose_y_1"), val = tensor(false)]; - tensor var_771 = matmul(transpose_x = var_771_transpose_x_1, transpose_y = var_771_transpose_y_1, x = k_7, y = v_7)[name = tensor("op_771")]; - tensor new_kv_unnorm_7 = add(x = var_769, y = var_771)[name = tensor("new_kv_unnorm_7")]; - tensor var_773 = const()[name = tensor("op_773"), val = tensor(0x1p+0)]; - tensor new_scale_7 = add(x = prev_scale_7, y = var_773)[name = tensor("new_scale_7")]; - tensor var_775 = sqrt(x = new_scale_7)[name = tensor("op_775")]; - tensor nkv_1 = real_div(x = new_kv_unnorm_7, y = var_775)[name = tensor("nkv_1")]; - tensor var_777_perm_0 = const()[name = tensor("op_777_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_795 = mul(x = prev_kv_7, y = sqrt_s0_7)[name = tensor("op_795")]; + tensor var_797_transpose_x_1 = const()[name = tensor("op_797_transpose_x_1"), val = tensor(true)]; + tensor var_797_transpose_y_1 = const()[name = tensor("op_797_transpose_y_1"), val = tensor(false)]; + tensor var_797 = matmul(transpose_x = var_797_transpose_x_1, transpose_y = var_797_transpose_y_1, x = k_7, y = v_7)[name = tensor("op_797")]; + tensor new_kv_unnorm_7 = add(x = var_795, y = var_797)[name = tensor("new_kv_unnorm_7")]; + tensor var_799 = const()[name = tensor("op_799"), val = tensor(0x1p+0)]; + tensor new_scale_7 = add(x = prev_scale_7, y = var_799)[name = tensor("new_scale_7")]; + tensor var_801 = sqrt(x = new_scale_7)[name = tensor("op_801")]; + tensor nkv_1 = real_div(x = new_kv_unnorm_7, y = var_801)[name = tensor("nkv_1")]; + tensor var_803_perm_0 = const()[name = tensor("op_803_perm_0"), val = tensor([0, 2, 1, 3])]; tensor out_21_axes_0 = const()[name = tensor("out_21_axes_0"), val = tensor([-1])]; - tensor var_777 = transpose(perm = var_777_perm_0, x = out_19)[name = tensor("transpose_34")]; - tensor out_21 = layer_norm(axes = out_21_axes_0, epsilon = var_17, x = var_777)[name = tensor("out_21")]; - tensor var_781 = const()[name = tensor("op_781"), val = tensor([1, 1, 256])]; - tensor out_23 = reshape(shape = var_781, x = out_21)[name = tensor("out_23")]; - tensor var_783 = silu(x = input_137)[name = tensor("op_783")]; - tensor input_139 = mul(x = var_783, y = out_23)[name = tensor("input_139")]; - tensor ret_out_7 = linear(bias = encoder_out_proj_3_bias, weight = encoder_out_proj_3_weight, x = input_139)[name = tensor("linear_34")]; - tensor x_21 = add(x = input_133, y = ret_out_7)[name = tensor("x_21")]; + tensor var_803 = transpose(perm = var_803_perm_0, x = out_19)[name = tensor("transpose_34")]; + tensor out_21 = layer_norm(axes = out_21_axes_0, epsilon = var_44, x = var_803)[name = tensor("out_21")]; + tensor var_807 = const()[name = tensor("op_807"), val = tensor([1, 1, 256])]; + tensor out_23 = reshape(shape = var_807, x = out_21)[name = tensor("out_23")]; + tensor var_809 = silu(x = input_139)[name = tensor("op_809")]; + tensor input_141 = mul(x = var_809, y = out_23)[name = tensor("input_141")]; + tensor ret_out_7 = linear(bias = inner_encoder_out_proj_3_bias, weight = inner_encoder_out_proj_3_weight, x = input_141)[name = tensor("linear_34")]; + tensor x_21 = add(x = input_135, y = ret_out_7)[name = tensor("x_21")]; tensor window_13_begin_0 = const()[name = tensor("window_13_begin_0"), val = tensor([3, 0, 0, 0])]; tensor window_13_end_0 = const()[name = tensor("window_13_end_0"), val = tensor([4, 1, 16, 256])]; tensor window_13_end_mask_0 = const()[name = tensor("window_13_end_mask_0"), val = tensor([false, true, true, true])]; tensor window_13_squeeze_mask_0 = const()[name = tensor("window_13_squeeze_mask_0"), val = tensor([true, false, false, false])]; tensor window_13 = slice_by_index(begin = window_13_begin_0, end = window_13_end_0, end_mask = window_13_end_mask_0, squeeze_mask = window_13_squeeze_mask_0, x = enc_conv_cache)[name = tensor("window_13")]; - tensor var_794_begin_0 = const()[name = tensor("op_794_begin_0"), val = tensor([0, 1, 0])]; - tensor var_794_end_0 = const()[name = tensor("op_794_end_0"), val = tensor([1, 16, 256])]; - tensor var_794_end_mask_0 = const()[name = tensor("op_794_end_mask_0"), val = tensor([true, true, true])]; - tensor var_794 = slice_by_index(begin = var_794_begin_0, end = var_794_end_0, end_mask = var_794_end_mask_0, x = window_13)[name = tensor("op_794")]; + tensor var_820_begin_0 = const()[name = tensor("op_820_begin_0"), val = tensor([0, 1, 0])]; + tensor var_820_end_0 = const()[name = tensor("op_820_end_0"), val = tensor([1, 16, 256])]; + tensor var_820_end_mask_0 = const()[name = tensor("op_820_end_mask_0"), val = tensor([true, true, true])]; + tensor var_820 = slice_by_index(begin = var_820_begin_0, end = var_820_end_0, end_mask = var_820_end_mask_0, x = window_13)[name = tensor("op_820")]; tensor window_interleave_0 = const()[name = tensor("window_interleave_0"), val = tensor(false)]; - tensor window = concat(axis = var_26, interleave = window_interleave_0, values = (var_794, x_21))[name = tensor("window")]; - tensor input_141_interleave_0 = const()[name = tensor("input_141_interleave_0"), val = tensor(false)]; - tensor input_141 = concat(axis = var_23, interleave = input_141_interleave_0, values = window)[name = tensor("input_141")]; + tensor window = concat(axis = var_52, interleave = window_interleave_0, values = (var_820, x_21))[name = tensor("window")]; + tensor input_143_interleave_0 = const()[name = tensor("input_143_interleave_0"), val = tensor(false)]; + tensor input_143 = concat(axis = var_39, interleave = input_143_interleave_0, values = window)[name = tensor("input_143")]; tensor x_23_axes_0 = const()[name = tensor("x_23_axes_0"), val = tensor([-1])]; - tensor x_23 = layer_norm(axes = x_23_axes_0, beta = encoder_conv_module_3_sequential_0_bias, epsilon = var_28, gamma = encoder_conv_module_3_sequential_0_weight, x = input_141)[name = tensor("x_23")]; - tensor input_143_perm_0 = const()[name = tensor("input_143_perm_0"), val = tensor([0, 2, 1])]; + tensor x_23 = layer_norm(axes = x_23_axes_0, beta = inner_encoder_conv_module_3_sequential_0_bias, epsilon = var_36, gamma = inner_encoder_conv_module_3_sequential_0_weight, x = input_143)[name = tensor("x_23")]; + tensor input_145_perm_0 = const()[name = tensor("input_145_perm_0"), val = tensor([0, 2, 1])]; tensor inputs_33_pad_type_0 = const()[name = tensor("inputs_33_pad_type_0"), val = tensor("valid")]; tensor inputs_33_strides_0 = const()[name = tensor("inputs_33_strides_0"), val = tensor([1])]; tensor inputs_33_pad_0 = const()[name = tensor("inputs_33_pad_0"), val = tensor([0, 0])]; tensor inputs_33_dilations_0 = const()[name = tensor("inputs_33_dilations_0"), val = tensor([1])]; tensor inputs_33_groups_0 = const()[name = tensor("inputs_33_groups_0"), val = tensor(1)]; - tensor input_143 = transpose(perm = input_143_perm_0, x = x_23)[name = tensor("transpose_33")]; - tensor inputs_33 = conv(bias = encoder_conv_module_3_sequential_2_conv_bias, dilations = inputs_33_dilations_0, groups = inputs_33_groups_0, pad = inputs_33_pad_0, pad_type = inputs_33_pad_type_0, strides = inputs_33_strides_0, weight = encoder_conv_module_3_sequential_2_conv_weight, x = input_143)[name = tensor("inputs_33")]; - tensor var_819_split_sizes_0 = const()[name = tensor("op_819_split_sizes_0"), val = tensor([256, 256])]; - tensor var_819_axis_0 = const()[name = tensor("op_819_axis_0"), val = tensor(1)]; - tensor var_819_0, tensor var_819_1 = split(axis = var_819_axis_0, split_sizes = var_819_split_sizes_0, x = inputs_33)[name = tensor("op_819")]; - tensor var_821 = sigmoid(x = var_819_1)[name = tensor("op_821")]; - tensor inputs_35 = mul(x = var_819_0, y = var_821)[name = tensor("inputs_35")]; + tensor input_145 = transpose(perm = input_145_perm_0, x = x_23)[name = tensor("transpose_33")]; + tensor inputs_33 = conv(bias = inner_encoder_conv_module_3_sequential_2_conv_bias, dilations = inputs_33_dilations_0, groups = inputs_33_groups_0, pad = inputs_33_pad_0, pad_type = inputs_33_pad_type_0, strides = inputs_33_strides_0, weight = inner_encoder_conv_module_3_sequential_2_conv_weight, x = input_145)[name = tensor("inputs_33")]; + tensor var_845_split_sizes_0 = const()[name = tensor("op_845_split_sizes_0"), val = tensor([256, 256])]; + tensor var_845_axis_0 = const()[name = tensor("op_845_axis_0"), val = tensor(1)]; + tensor var_845_0, tensor var_845_1 = split(axis = var_845_axis_0, split_sizes = var_845_split_sizes_0, x = inputs_33)[name = tensor("op_845")]; + tensor var_847 = sigmoid(x = var_845_1)[name = tensor("op_847")]; + tensor inputs_35 = mul(x = var_845_0, y = var_847)[name = tensor("inputs_35")]; tensor outputs_aug_pad_type_0 = const()[name = tensor("outputs_aug_pad_type_0"), val = tensor("custom")]; tensor outputs_aug_pad_0 = const()[name = tensor("outputs_aug_pad_0"), val = tensor([15, 15])]; tensor outputs_aug_groups_0 = const()[name = tensor("outputs_aug_groups_0"), val = tensor(256)]; tensor outputs_aug_strides_0 = const()[name = tensor("outputs_aug_strides_0"), val = tensor([1])]; tensor outputs_aug_dilations_0 = const()[name = tensor("outputs_aug_dilations_0"), val = tensor([1])]; - tensor outputs_aug = conv(dilations = outputs_aug_dilations_0, groups = outputs_aug_groups_0, pad = outputs_aug_pad_0, pad_type = outputs_aug_pad_type_0, strides = outputs_aug_strides_0, weight = encoder_conv_module_3_sequential_4_conv_weight, x = inputs_35)[name = tensor("outputs_aug")]; - tensor input_145_begin_0 = const()[name = tensor("input_145_begin_0"), val = tensor([0, 0, 0])]; - tensor input_145_end_0 = const()[name = tensor("input_145_end_0"), val = tensor([1, 256, 16])]; - tensor input_145_end_mask_0 = const()[name = tensor("input_145_end_mask_0"), val = tensor([true, true, false])]; - tensor input_145 = slice_by_index(begin = input_145_begin_0, end = input_145_end_0, end_mask = input_145_end_mask_0, x = outputs_aug)[name = tensor("input_145")]; - tensor inputs_37 = batch_norm(beta = encoder_conv_module_3_sequential_5_bias, epsilon = var_28, gamma = encoder_conv_module_3_sequential_5_weight, mean = encoder_conv_module_3_sequential_5_running_mean, variance = encoder_conv_module_3_sequential_5_running_var, x = input_145)[name = tensor("inputs_37")]; - tensor input_147 = silu(x = inputs_37)[name = tensor("input_147")]; - tensor input_149_pad_type_0 = const()[name = tensor("input_149_pad_type_0"), val = tensor("valid")]; - tensor input_149_strides_0 = const()[name = tensor("input_149_strides_0"), val = tensor([1])]; - tensor input_149_pad_0 = const()[name = tensor("input_149_pad_0"), val = tensor([0, 0])]; - tensor input_149_dilations_0 = const()[name = tensor("input_149_dilations_0"), val = tensor([1])]; - tensor input_149_groups_0 = const()[name = tensor("input_149_groups_0"), val = tensor(1)]; - tensor input_149 = conv(bias = encoder_conv_module_3_sequential_7_conv_bias, dilations = input_149_dilations_0, groups = input_149_groups_0, pad = input_149_pad_0, pad_type = input_149_pad_type_0, strides = input_149_strides_0, weight = encoder_conv_module_3_sequential_7_conv_weight, x = input_147)[name = tensor("input_149")]; + tensor outputs_aug = conv(dilations = outputs_aug_dilations_0, groups = outputs_aug_groups_0, pad = outputs_aug_pad_0, pad_type = outputs_aug_pad_type_0, strides = outputs_aug_strides_0, weight = inner_encoder_conv_module_3_sequential_4_conv_weight, x = inputs_35)[name = tensor("outputs_aug")]; + tensor input_147_begin_0 = const()[name = tensor("input_147_begin_0"), val = tensor([0, 0, 0])]; + tensor input_147_end_0 = const()[name = tensor("input_147_end_0"), val = tensor([1, 256, 16])]; + tensor input_147_end_mask_0 = const()[name = tensor("input_147_end_mask_0"), val = tensor([true, true, false])]; + tensor input_147 = slice_by_index(begin = input_147_begin_0, end = input_147_end_0, end_mask = input_147_end_mask_0, x = outputs_aug)[name = tensor("input_147")]; + tensor inputs_37 = batch_norm(beta = inner_encoder_conv_module_3_sequential_5_bias, epsilon = var_36, gamma = inner_encoder_conv_module_3_sequential_5_weight, mean = inner_encoder_conv_module_3_sequential_5_running_mean, variance = inner_encoder_conv_module_3_sequential_5_running_var, x = input_147)[name = tensor("inputs_37")]; + tensor input_149 = silu(x = inputs_37)[name = tensor("input_149")]; + tensor input_151_pad_type_0 = const()[name = tensor("input_151_pad_type_0"), val = tensor("valid")]; + tensor input_151_strides_0 = const()[name = tensor("input_151_strides_0"), val = tensor([1])]; + tensor input_151_pad_0 = const()[name = tensor("input_151_pad_0"), val = tensor([0, 0])]; + tensor input_151_dilations_0 = const()[name = tensor("input_151_dilations_0"), val = tensor([1])]; + tensor input_151_groups_0 = const()[name = tensor("input_151_groups_0"), val = tensor(1)]; + tensor input_151 = conv(bias = inner_encoder_conv_module_3_sequential_7_conv_bias, dilations = input_151_dilations_0, groups = input_151_groups_0, pad = input_151_pad_0, pad_type = input_151_pad_type_0, strides = input_151_strides_0, weight = inner_encoder_conv_module_3_sequential_7_conv_weight, x = input_149)[name = tensor("input_151")]; tensor conv_out_7_perm_0 = const()[name = tensor("conv_out_7_perm_0"), val = tensor([0, 2, 1])]; - tensor var_852_begin_0 = const()[name = tensor("op_852_begin_0"), val = tensor([0, -1, 0])]; - tensor var_852_end_0 = const()[name = tensor("op_852_end_0"), val = tensor([1, 16, 256])]; - tensor var_852_end_mask_0 = const()[name = tensor("op_852_end_mask_0"), val = tensor([true, true, true])]; - tensor conv_out_7 = transpose(perm = conv_out_7_perm_0, x = input_149)[name = tensor("transpose_32")]; - tensor var_852 = slice_by_index(begin = var_852_begin_0, end = var_852_end_0, end_mask = var_852_end_mask_0, x = conv_out_7)[name = tensor("op_852")]; - tensor var_854_perm_0 = const()[name = tensor("op_854_perm_0"), val = tensor([1, 0, 2])]; - tensor var_854 = transpose(perm = var_854_perm_0, x = var_852)[name = tensor("transpose_31")]; - tensor input_151 = add(x = x_21, y = var_854)[name = tensor("input_151")]; - tensor input_153_axes_0 = const()[name = tensor("input_153_axes_0"), val = tensor([-1])]; - tensor input_153 = layer_norm(axes = input_153_axes_0, beta = encoder_ffn2_3_module_sequential_0_bias, epsilon = var_28, gamma = encoder_ffn2_3_module_sequential_0_weight, x = input_151)[name = tensor("input_153")]; - tensor inputs = linear(bias = encoder_ffn2_3_module_sequential_1_linear_bias, weight = encoder_ffn2_3_module_sequential_1_linear_weight, x = input_153)[name = tensor("linear_35")]; - tensor input_155 = silu(x = inputs)[name = tensor("input_155")]; - tensor input_159 = linear(bias = encoder_ffn2_3_module_sequential_4_linear_bias, weight = encoder_ffn2_3_module_sequential_4_linear_weight, x = input_155)[name = tensor("linear_36")]; - tensor var_877 = const()[name = tensor("op_877"), val = tensor(0x1p-1)]; - tensor var_878 = mul(x = input_159, y = var_877)[name = tensor("op_878")]; - tensor input_161 = add(x = var_878, y = input_151)[name = tensor("input_161")]; + tensor var_878_begin_0 = const()[name = tensor("op_878_begin_0"), val = tensor([0, -1, 0])]; + tensor var_878_end_0 = const()[name = tensor("op_878_end_0"), val = tensor([1, 16, 256])]; + tensor var_878_end_mask_0 = const()[name = tensor("op_878_end_mask_0"), val = tensor([true, true, true])]; + tensor conv_out_7 = transpose(perm = conv_out_7_perm_0, x = input_151)[name = tensor("transpose_32")]; + tensor var_878 = slice_by_index(begin = var_878_begin_0, end = var_878_end_0, end_mask = var_878_end_mask_0, x = conv_out_7)[name = tensor("op_878")]; + tensor var_880_perm_0 = const()[name = tensor("op_880_perm_0"), val = tensor([1, 0, 2])]; + tensor var_880 = transpose(perm = var_880_perm_0, x = var_878)[name = tensor("transpose_31")]; + tensor input_153 = add(x = x_21, y = var_880)[name = tensor("input_153")]; + tensor input_155_axes_0 = const()[name = tensor("input_155_axes_0"), val = tensor([-1])]; + tensor input_155 = layer_norm(axes = input_155_axes_0, beta = inner_encoder_ffn2_3_module_sequential_0_bias, epsilon = var_36, gamma = inner_encoder_ffn2_3_module_sequential_0_weight, x = input_153)[name = tensor("input_155")]; + tensor inputs = linear(bias = inner_encoder_ffn2_3_module_sequential_1_linear_bias, weight = inner_encoder_ffn2_3_module_sequential_1_linear_weight, x = input_155)[name = tensor("linear_35")]; + tensor input_157 = silu(x = inputs)[name = tensor("input_157")]; + tensor input_161 = linear(bias = inner_encoder_ffn2_3_module_sequential_4_linear_bias, weight = inner_encoder_ffn2_3_module_sequential_4_linear_weight, x = input_157)[name = tensor("linear_36")]; + tensor var_903 = const()[name = tensor("op_903"), val = tensor(0x1p-1)]; + tensor var_904 = mul(x = input_161, y = var_903)[name = tensor("op_904")]; + tensor input_163 = add(x = var_904, y = input_153)[name = tensor("input_163")]; tensor x_25_axes_0 = const()[name = tensor("x_25_axes_0"), val = tensor([-1])]; - tensor x_25 = layer_norm(axes = x_25_axes_0, beta = encoder_layer_norm_3_bias, epsilon = var_28, gamma = encoder_layer_norm_3_weight, x = input_161)[name = tensor("x_25")]; + tensor x_25 = layer_norm(axes = x_25_axes_0, beta = inner_encoder_layer_norm_3_bias, epsilon = var_36, gamma = inner_encoder_layer_norm_3_weight, x = input_163)[name = tensor("x_25")]; tensor x_ct_perm_0 = const()[name = tensor("x_ct_perm_0"), val = tensor([0, 2, 1])]; tensor cat_interleave_0 = const()[name = tensor("cat_interleave_0"), val = tensor(false)]; tensor x_ct = transpose(perm = x_ct_perm_0, x = x_25)[name = tensor("transpose_30")]; - tensor cat = concat(axis = var_20, interleave = cat_interleave_0, values = (cnn_window, x_ct))[name = tensor("cat")]; + tensor cat = concat(axis = var_41, interleave = cat_interleave_0, values = (cnn_window, x_ct))[name = tensor("cat")]; tensor conv_out_pad_type_0 = const()[name = tensor("conv_out_pad_type_0"), val = tensor("valid")]; tensor conv_out_strides_0 = const()[name = tensor("conv_out_strides_0"), val = tensor([1])]; tensor conv_out_pad_0 = const()[name = tensor("conv_out_pad_0"), val = tensor([0, 0])]; tensor conv_out_dilations_0 = const()[name = tensor("conv_out_dilations_0"), val = tensor([1])]; tensor conv_out_groups_0 = const()[name = tensor("conv_out_groups_0"), val = tensor(1)]; - tensor conv_out = conv(bias = encoder_cnn_bias, dilations = conv_out_dilations_0, groups = conv_out_groups_0, pad = conv_out_pad_0, pad_type = conv_out_pad_type_0, strides = conv_out_strides_0, weight = encoder_cnn_weight, x = cat)[name = tensor("conv_out")]; - tensor var_896_begin_0 = const()[name = tensor("op_896_begin_0"), val = tensor([0, 0, 1])]; - tensor var_896_end_0 = const()[name = tensor("op_896_end_0"), val = tensor([1, 256, 19])]; - tensor var_896_end_mask_0 = const()[name = tensor("op_896_end_mask_0"), val = tensor([true, true, true])]; - tensor cnn_window_new = slice_by_index(begin = var_896_begin_0, end = var_896_end_0, end_mask = var_896_end_mask_0, x = cat)[name = tensor("op_896")]; - tensor input_163_perm_0 = const()[name = tensor("input_163_perm_0"), val = tensor([0, 2, 1])]; - tensor var_898 = const()[name = tensor("op_898"), val = tensor([-1])]; - tensor input_163 = transpose(perm = input_163_perm_0, x = conv_out)[name = tensor("transpose_29")]; - tensor var_899 = reduce_l2_norm(axes = var_898, keep_dims = var_29, x = input_163)[name = tensor("op_899")]; + tensor conv_out = conv(bias = inner_encoder_cnn_bias, dilations = conv_out_dilations_0, groups = conv_out_groups_0, pad = conv_out_pad_0, pad_type = conv_out_pad_type_0, strides = conv_out_strides_0, weight = inner_encoder_cnn_weight, x = cat)[name = tensor("conv_out")]; + tensor var_922_begin_0 = const()[name = tensor("op_922_begin_0"), val = tensor([0, 0, 1])]; + tensor var_922_end_0 = const()[name = tensor("op_922_end_0"), val = tensor([1, 256, 19])]; + tensor var_922_end_mask_0 = const()[name = tensor("op_922_end_mask_0"), val = tensor([true, true, true])]; + tensor cnn_window_new = slice_by_index(begin = var_922_begin_0, end = var_922_end_0, end_mask = var_922_end_mask_0, x = cat)[name = tensor("op_922")]; + tensor input_165_perm_0 = const()[name = tensor("input_165_perm_0"), val = tensor([0, 2, 1])]; + tensor var_924 = const()[name = tensor("op_924"), val = tensor([-1])]; + tensor input_165 = transpose(perm = input_165_perm_0, x = conv_out)[name = tensor("transpose_29")]; + tensor var_925 = reduce_l2_norm(axes = var_924, keep_dims = var_35, x = input_165)[name = tensor("op_925")]; tensor const_12 = const()[name = tensor("const_12"), val = tensor(0x1.fffffep+127)]; - tensor clip_0 = clip(alpha = var_11, beta = const_12, x = var_899)[name = tensor("clip_0")]; - tensor emb = real_div(x = input_163, y = clip_0)[name = tensor("emb")]; - tensor var_903_axis_0 = const()[name = tensor("op_903_axis_0"), val = tensor(0)]; - tensor enc_kv_new = stack(axis = var_903_axis_0, values = (var_206, var_396, var_586, nkv_1))[name = tensor("op_903")]; - tensor var_905_axis_0 = const()[name = tensor("op_905_axis_0"), val = tensor(0)]; - tensor enc_scale_new = stack(axis = var_905_axis_0, values = (new_scale_1, new_scale_3, new_scale_5, new_scale_7))[name = tensor("op_905")]; - tensor var_907_axis_0 = const()[name = tensor("op_907_axis_0"), val = tensor(0)]; - tensor enc_conv_cache_new = stack(axis = var_907_axis_0, values = (window_3, window_7, window_11, window))[name = tensor("op_907")]; - tensor var_916 = const()[name = tensor("op_916"), val = tensor(0x1.5798eep-27)]; - tensor var_921 = const()[name = tensor("op_921"), val = tensor(0x1.4f8b58p-17)]; - tensor var_923 = const()[name = tensor("op_923"), val = tensor(0x1.0c6f7ap-20)]; - tensor var_924 = const()[name = tensor("op_924"), val = tensor(true)]; - tensor var_926 = const()[name = tensor("op_926"), val = tensor(0x1p+0)]; - tensor var_930 = const()[name = tensor("op_930"), val = tensor(-1)]; - tensor var_936 = const()[name = tensor("op_936"), val = tensor(0)]; - tensor var_993 = const()[name = tensor("op_993"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(44395328)))]; - tensor var_998_axes_0 = const()[name = tensor("op_998_axes_0"), val = tensor([2])]; - tensor var_998 = expand_dims(axes = var_998_axes_0, x = emb)[name = tensor("op_998")]; + tensor clip_0 = clip(alpha = var_49, beta = const_12, x = var_925)[name = tensor("clip_0")]; + tensor emb = real_div(x = input_165, y = clip_0)[name = tensor("emb")]; + tensor var_929_axis_0 = const()[name = tensor("op_929_axis_0"), val = tensor(0)]; + tensor enc_kv_new = stack(axis = var_929_axis_0, values = (var_232, var_422, var_612, nkv_1))[name = tensor("op_929")]; + tensor var_931_axis_0 = const()[name = tensor("op_931_axis_0"), val = tensor(0)]; + tensor enc_scale_new = stack(axis = var_931_axis_0, values = (new_scale_1, new_scale_3, new_scale_5, new_scale_7))[name = tensor("op_931")]; + tensor var_933_axis_0 = const()[name = tensor("op_933_axis_0"), val = tensor(0)]; + tensor enc_conv_cache_new = stack(axis = var_933_axis_0, values = (window_3, window_7, window_11, window))[name = tensor("op_933")]; + tensor var_996 = const()[name = tensor("op_996"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(44395328)))]; + tensor var_1001_axes_0 = const()[name = tensor("op_1001_axes_0"), val = tensor([2])]; + tensor var_1001 = expand_dims(axes = var_1001_axes_0, x = emb)[name = tensor("op_1001")]; tensor emb_exp_reps_0 = const()[name = tensor("emb_exp_reps_0"), val = tensor([1, 1, 12, 1])]; - tensor emb_exp = tile(reps = emb_exp_reps_0, x = var_998)[name = tensor("emb_exp")]; - tensor input_165_interleave_0 = const()[name = tensor("input_165_interleave_0"), val = tensor(false)]; - tensor input_165 = concat(axis = var_930, interleave = input_165_interleave_0, values = (emb_exp, var_993))[name = tensor("input_165")]; - tensor x_27 = linear(bias = decoder_convert_bias, weight = decoder_convert_weight, x = input_165)[name = tensor("linear_37")]; - tensor var_1006_perm_0 = const()[name = tensor("op_1006_perm_0"), val = tensor([0, 2, 1, 3])]; - tensor var_1010 = const()[name = tensor("op_1010"), val = tensor([12, 1, 256])]; - tensor var_1006 = transpose(perm = var_1006_perm_0, x = x_27)[name = tensor("transpose_28")]; - tensor x_29 = reshape(shape = var_1010, x = var_1006)[name = tensor("x_29")]; + tensor emb_exp = tile(reps = emb_exp_reps_0, x = var_1001)[name = tensor("emb_exp")]; + tensor input_167_interleave_0 = const()[name = tensor("input_167_interleave_0"), val = tensor(false)]; + tensor input_167 = concat(axis = var_42, interleave = input_167_interleave_0, values = (emb_exp, var_996))[name = tensor("input_167")]; + tensor x_27 = linear(bias = inner_decoder_convert_bias, weight = inner_decoder_convert_weight, x = input_167)[name = tensor("linear_37")]; + tensor var_1009_perm_0 = const()[name = tensor("op_1009_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_1013 = const()[name = tensor("op_1013"), val = tensor([12, 1, 256])]; + tensor var_1009 = transpose(perm = var_1009_perm_0, x = x_27)[name = tensor("transpose_28")]; + tensor x_29 = reshape(shape = var_1013, x = var_1009)[name = tensor("x_29")]; tensor prev_kv_9_begin_0 = const()[name = tensor("prev_kv_9_begin_0"), val = tensor([0, 0, 0, 0, 0])]; tensor prev_kv_9_end_0 = const()[name = tensor("prev_kv_9_end_0"), val = tensor([1, 12, 4, 64, 64])]; tensor prev_kv_9_end_mask_0 = const()[name = tensor("prev_kv_9_end_mask_0"), val = tensor([false, true, true, true, true])]; @@ -846,131 +845,131 @@ program(1.0) tensor prev_scale_9_end_mask_0 = const()[name = tensor("prev_scale_9_end_mask_0"), val = tensor([false, true])]; tensor prev_scale_9_squeeze_mask_0 = const()[name = tensor("prev_scale_9_squeeze_mask_0"), val = tensor([true, false])]; tensor prev_scale_9 = slice_by_index(begin = prev_scale_9_begin_0, end = prev_scale_9_end_0, end_mask = prev_scale_9_end_mask_0, squeeze_mask = prev_scale_9_squeeze_mask_0, x = dec_scale)[name = tensor("prev_scale_9")]; - tensor var_1018 = linear(bias = decoder_q_proj_0_bias, weight = decoder_q_proj_0_weight, x = x_29)[name = tensor("linear_38")]; - tensor var_1019 = const()[name = tensor("op_1019"), val = tensor([12, 1, 4, 64])]; - tensor var_1020 = reshape(shape = var_1019, x = var_1018)[name = tensor("op_1020")]; + tensor var_1021 = linear(bias = inner_decoder_q_proj_0_bias, weight = inner_decoder_q_proj_0_weight, x = x_29)[name = tensor("linear_38")]; + tensor var_1022 = const()[name = tensor("op_1022"), val = tensor([12, 1, 4, 64])]; + tensor var_1023 = reshape(shape = var_1022, x = var_1021)[name = tensor("op_1023")]; tensor q_9_perm_0 = const()[name = tensor("q_9_perm_0"), val = tensor([0, 2, 1, 3])]; - tensor var_1024 = linear(bias = decoder_k_proj_0_bias, weight = decoder_k_proj_0_weight, x = x_29)[name = tensor("linear_39")]; - tensor var_1025 = const()[name = tensor("op_1025"), val = tensor(0x1p-3)]; - tensor var_1026 = mul(x = var_1024, y = var_1025)[name = tensor("op_1026")]; - tensor var_1027 = const()[name = tensor("op_1027"), val = tensor([12, 1, 4, 64])]; - tensor var_1028 = reshape(shape = var_1027, x = var_1026)[name = tensor("op_1028")]; + tensor var_1027 = linear(bias = inner_decoder_k_proj_0_bias, weight = inner_decoder_k_proj_0_weight, x = x_29)[name = tensor("linear_39")]; + tensor var_1028 = const()[name = tensor("op_1028"), val = tensor(0x1p-3)]; + tensor var_1029 = mul(x = var_1027, y = var_1028)[name = tensor("op_1029")]; + tensor var_1030 = const()[name = tensor("op_1030"), val = tensor([12, 1, 4, 64])]; + tensor var_1031 = reshape(shape = var_1030, x = var_1029)[name = tensor("op_1031")]; tensor k_9_perm_0 = const()[name = tensor("k_9_perm_0"), val = tensor([0, 2, 1, 3])]; - tensor var_1032 = linear(bias = decoder_v_proj_0_bias, weight = decoder_v_proj_0_weight, x = x_29)[name = tensor("linear_40")]; - tensor var_1033 = const()[name = tensor("op_1033"), val = tensor([12, 1, 4, 64])]; - tensor var_1034 = reshape(shape = var_1033, x = var_1032)[name = tensor("op_1034")]; + tensor var_1035 = linear(bias = inner_decoder_v_proj_0_bias, weight = inner_decoder_v_proj_0_weight, x = x_29)[name = tensor("linear_40")]; + tensor var_1036 = const()[name = tensor("op_1036"), val = tensor([12, 1, 4, 64])]; + tensor var_1037 = reshape(shape = var_1036, x = var_1035)[name = tensor("op_1037")]; tensor v_9_perm_0 = const()[name = tensor("v_9_perm_0"), val = tensor([0, 2, 1, 3])]; - tensor input_169 = linear(bias = decoder_g_proj_0_bias, weight = decoder_g_proj_0_weight, x = x_29)[name = tensor("linear_41")]; + tensor input_171 = linear(bias = inner_decoder_g_proj_0_bias, weight = inner_decoder_g_proj_0_weight, x = x_29)[name = tensor("linear_41")]; tensor cumsum_mask_1_exclusive_0 = const()[name = tensor("cumsum_mask_1_exclusive_0"), val = tensor(false)]; tensor cumsum_mask_1_reverse_0 = const()[name = tensor("cumsum_mask_1_reverse_0"), val = tensor(false)]; - tensor cumsum_mask_1 = cumsum(axis = var_936, exclusive = cumsum_mask_1_exclusive_0, reverse = cumsum_mask_1_reverse_0, x = valid_mask)[name = tensor("cumsum_mask_1")]; + tensor cumsum_mask_1 = cumsum(axis = var_39, exclusive = cumsum_mask_1_exclusive_0, reverse = cumsum_mask_1_reverse_0, x = valid_mask)[name = tensor("cumsum_mask_1")]; tensor sqrt_s0_9 = sqrt(x = prev_scale_9)[name = tensor("sqrt_s0_9")]; tensor s_t_9 = add(x = prev_scale_9, y = cumsum_mask_1)[name = tensor("s_t_9")]; tensor const_20 = const()[name = tensor("const_20"), val = tensor(0x1.fffffep+127)]; - tensor clip_1 = clip(alpha = var_926, beta = const_20, x = s_t_9)[name = tensor("clip_1")]; + tensor clip_1 = clip(alpha = var_29, beta = const_20, x = s_t_9)[name = tensor("clip_1")]; tensor sqrt_s_t_9 = sqrt(x = clip_1)[name = tensor("sqrt_s_t_9")]; tensor qk_9_transpose_x_1 = const()[name = tensor("qk_9_transpose_x_1"), val = tensor(false)]; tensor qk_9_transpose_y_1 = const()[name = tensor("qk_9_transpose_y_1"), val = tensor(true)]; - tensor k_9 = transpose(perm = k_9_perm_0, x = var_1028)[name = tensor("transpose_26")]; - tensor q_9 = transpose(perm = q_9_perm_0, x = var_1020)[name = tensor("transpose_27")]; + tensor k_9 = transpose(perm = k_9_perm_0, x = var_1031)[name = tensor("transpose_26")]; + tensor q_9 = transpose(perm = q_9_perm_0, x = var_1023)[name = tensor("transpose_27")]; tensor qk_9 = matmul(transpose_x = qk_9_transpose_x_1, transpose_y = qk_9_transpose_y_1, x = q_9, y = k_9)[name = tensor("qk_9")]; - tensor var_1046 = const()[name = tensor("op_1046"), val = tensor([1, 1])]; - tensor var_1047 = reshape(shape = var_1046, x = valid_mask)[name = tensor("op_1047")]; tensor var_1049 = const()[name = tensor("op_1049"), val = tensor([1, 1])]; - tensor var_1050 = reshape(shape = var_1049, x = sqrt_s_t_9)[name = tensor("op_1050")]; - tensor M_9 = real_div(x = var_1047, y = var_1050)[name = tensor("M_9")]; - tensor var_1052 = mul(x = qk_9, y = M_9)[name = tensor("op_1052")]; + tensor var_1050 = reshape(shape = var_1049, x = valid_mask)[name = tensor("op_1050")]; + tensor var_1052 = const()[name = tensor("op_1052"), val = tensor([1, 1])]; + tensor var_1053 = reshape(shape = var_1052, x = sqrt_s_t_9)[name = tensor("op_1053")]; + tensor M_9 = real_div(x = var_1050, y = var_1053)[name = tensor("M_9")]; + tensor var_1055 = mul(x = qk_9, y = M_9)[name = tensor("op_1055")]; tensor inner_9_transpose_x_0 = const()[name = tensor("inner_9_transpose_x_0"), val = tensor(false)]; tensor inner_9_transpose_y_0 = const()[name = tensor("inner_9_transpose_y_0"), val = tensor(false)]; - tensor v_9 = transpose(perm = v_9_perm_0, x = var_1034)[name = tensor("transpose_25")]; - tensor inner_9 = matmul(transpose_x = inner_9_transpose_x_0, transpose_y = inner_9_transpose_y_0, x = var_1052, y = v_9)[name = tensor("inner_9")]; - tensor var_1054_transpose_x_0 = const()[name = tensor("op_1054_transpose_x_0"), val = tensor(false)]; - tensor var_1054_transpose_y_0 = const()[name = tensor("op_1054_transpose_y_0"), val = tensor(false)]; - tensor var_1054 = matmul(transpose_x = var_1054_transpose_x_0, transpose_y = var_1054_transpose_y_0, x = q_9, y = prev_kv_9)[name = tensor("op_1054")]; - tensor var_1055 = real_div(x = sqrt_s0_9, y = sqrt_s_t_9)[name = tensor("op_1055")]; - tensor var_1056 = const()[name = tensor("op_1056"), val = tensor([1, 1, 1, 1])]; - tensor var_1057 = reshape(shape = var_1056, x = var_1055)[name = tensor("op_1057")]; - tensor cross_9 = mul(x = var_1054, y = var_1057)[name = tensor("cross_9")]; + tensor v_9 = transpose(perm = v_9_perm_0, x = var_1037)[name = tensor("transpose_25")]; + tensor inner_9 = matmul(transpose_x = inner_9_transpose_x_0, transpose_y = inner_9_transpose_y_0, x = var_1055, y = v_9)[name = tensor("inner_9")]; + tensor var_1057_transpose_x_0 = const()[name = tensor("op_1057_transpose_x_0"), val = tensor(false)]; + tensor var_1057_transpose_y_0 = const()[name = tensor("op_1057_transpose_y_0"), val = tensor(false)]; + tensor var_1057 = matmul(transpose_x = var_1057_transpose_x_0, transpose_y = var_1057_transpose_y_0, x = q_9, y = prev_kv_9)[name = tensor("op_1057")]; + tensor var_1058 = real_div(x = sqrt_s0_9, y = sqrt_s_t_9)[name = tensor("op_1058")]; + tensor var_1059 = const()[name = tensor("op_1059"), val = tensor([1, 1, 1, 1])]; + tensor var_1060 = reshape(shape = var_1059, x = var_1058)[name = tensor("op_1060")]; + tensor cross_9 = mul(x = var_1057, y = var_1060)[name = tensor("cross_9")]; tensor out_25 = add(x = inner_9, y = cross_9)[name = tensor("out_25")]; - tensor var_1060 = const()[name = tensor("op_1060"), val = tensor([1, 1, 1, 1])]; - tensor var_1061 = reshape(shape = var_1060, x = valid_mask)[name = tensor("op_1061")]; - tensor v_masked_1 = mul(x = v_9, y = var_1061)[name = tensor("v_masked_1")]; - tensor var_1063 = mul(x = prev_kv_9, y = sqrt_s0_9)[name = tensor("op_1063")]; - tensor var_1065_transpose_x_1 = const()[name = tensor("op_1065_transpose_x_1"), val = tensor(true)]; - tensor var_1065_transpose_y_1 = const()[name = tensor("op_1065_transpose_y_1"), val = tensor(false)]; - tensor var_1065 = matmul(transpose_x = var_1065_transpose_x_1, transpose_y = var_1065_transpose_y_1, x = k_9, y = v_masked_1)[name = tensor("op_1065")]; - tensor new_kv_unnorm_9 = add(x = var_1063, y = var_1065)[name = tensor("new_kv_unnorm_9")]; - tensor var_1067_keep_dims_0 = const()[name = tensor("op_1067_keep_dims_0"), val = tensor(false)]; - tensor var_1067 = reduce_sum(keep_dims = var_1067_keep_dims_0, x = valid_mask)[name = tensor("op_1067")]; - tensor var_1068 = const()[name = tensor("op_1068"), val = tensor([1])]; - tensor var_1069 = reshape(shape = var_1068, x = var_1067)[name = tensor("op_1069")]; - tensor new_scale_9 = add(x = prev_scale_9, y = var_1069)[name = tensor("new_scale_9")]; + tensor var_1063 = const()[name = tensor("op_1063"), val = tensor([1, 1, 1, 1])]; + tensor var_1064 = reshape(shape = var_1063, x = valid_mask)[name = tensor("op_1064")]; + tensor v_masked_1 = mul(x = v_9, y = var_1064)[name = tensor("v_masked_1")]; + tensor var_1066 = mul(x = prev_kv_9, y = sqrt_s0_9)[name = tensor("op_1066")]; + tensor var_1068_transpose_x_1 = const()[name = tensor("op_1068_transpose_x_1"), val = tensor(true)]; + tensor var_1068_transpose_y_1 = const()[name = tensor("op_1068_transpose_y_1"), val = tensor(false)]; + tensor var_1068 = matmul(transpose_x = var_1068_transpose_x_1, transpose_y = var_1068_transpose_y_1, x = k_9, y = v_masked_1)[name = tensor("op_1068")]; + tensor new_kv_unnorm_9 = add(x = var_1066, y = var_1068)[name = tensor("new_kv_unnorm_9")]; + tensor var_1070_keep_dims_0 = const()[name = tensor("op_1070_keep_dims_0"), val = tensor(false)]; + tensor var_1070 = reduce_sum(keep_dims = var_1070_keep_dims_0, x = valid_mask)[name = tensor("op_1070")]; + tensor var_1071 = const()[name = tensor("op_1071"), val = tensor([1])]; + tensor var_1072 = reshape(shape = var_1071, x = var_1070)[name = tensor("op_1072")]; + tensor new_scale_9 = add(x = prev_scale_9, y = var_1072)[name = tensor("new_scale_9")]; tensor const_21 = const()[name = tensor("const_21"), val = tensor(0x1.fffffep+127)]; - tensor clip_2 = clip(alpha = var_926, beta = const_21, x = new_scale_9)[name = tensor("clip_2")]; + tensor clip_2 = clip(alpha = var_29, beta = const_21, x = new_scale_9)[name = tensor("clip_2")]; tensor sqrt_new_scale_1 = sqrt(x = clip_2)[name = tensor("sqrt_new_scale_1")]; - tensor var_1073 = real_div(x = new_kv_unnorm_9, y = sqrt_new_scale_1)[name = tensor("op_1073")]; - tensor var_1074_perm_0 = const()[name = tensor("op_1074_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_1076 = real_div(x = new_kv_unnorm_9, y = sqrt_new_scale_1)[name = tensor("op_1076")]; + tensor var_1077_perm_0 = const()[name = tensor("op_1077_perm_0"), val = tensor([0, 2, 1, 3])]; tensor out_27_axes_0 = const()[name = tensor("out_27_axes_0"), val = tensor([-1])]; - tensor var_1074 = transpose(perm = var_1074_perm_0, x = out_25)[name = tensor("transpose_24")]; - tensor out_27 = layer_norm(axes = out_27_axes_0, epsilon = var_923, x = var_1074)[name = tensor("out_27")]; - tensor var_1078 = const()[name = tensor("op_1078"), val = tensor([12, 1, 256])]; - tensor out_29 = reshape(shape = var_1078, x = out_27)[name = tensor("out_29")]; - tensor var_1080 = silu(x = input_169)[name = tensor("op_1080")]; - tensor input_171 = mul(x = var_1080, y = out_29)[name = tensor("input_171")]; - tensor ret_out_9 = linear(bias = decoder_out_proj_0_bias, weight = decoder_out_proj_0_weight, x = input_171)[name = tensor("linear_42")]; - tensor input_173 = add(x = x_29, y = ret_out_9)[name = tensor("input_173")]; + tensor var_1077 = transpose(perm = var_1077_perm_0, x = out_25)[name = tensor("transpose_24")]; + tensor out_27 = layer_norm(axes = out_27_axes_0, epsilon = var_44, x = var_1077)[name = tensor("out_27")]; + tensor var_1081 = const()[name = tensor("op_1081"), val = tensor([12, 1, 256])]; + tensor out_29 = reshape(shape = var_1081, x = out_27)[name = tensor("out_29")]; + tensor var_1083 = silu(x = input_171)[name = tensor("op_1083")]; + tensor input_173 = mul(x = var_1083, y = out_29)[name = tensor("input_173")]; + tensor ret_out_9 = linear(bias = inner_decoder_out_proj_0_bias, weight = inner_decoder_out_proj_0_weight, x = input_173)[name = tensor("linear_42")]; + tensor input_175 = add(x = x_29, y = ret_out_9)[name = tensor("input_175")]; tensor xt_1_axes_0 = const()[name = tensor("xt_1_axes_0"), val = tensor([-1])]; - tensor xt_1 = layer_norm(axes = xt_1_axes_0, beta = decoder_norm11_0_bias, epsilon = var_921, gamma = decoder_norm11_0_weight, x = input_173)[name = tensor("xt_1")]; - tensor var_1090 = const()[name = tensor("op_1090"), val = tensor([1, 12, 1, 256])]; - tensor var_1091 = reshape(shape = var_1090, x = xt_1)[name = tensor("op_1091")]; - tensor var_1092_perm_0 = const()[name = tensor("op_1092_perm_0"), val = tensor([0, 2, 1, 3])]; - tensor var_1095 = const()[name = tensor("op_1095"), val = tensor([1, 12, 256])]; - tensor var_1092 = transpose(perm = var_1092_perm_0, x = var_1091)[name = tensor("transpose_23")]; - tensor query_1 = reshape(shape = var_1095, x = var_1092)[name = tensor("query_1")]; + tensor xt_1 = layer_norm(axes = xt_1_axes_0, beta = inner_decoder_norm11_0_bias, epsilon = var_36, gamma = inner_decoder_norm11_0_weight, x = input_175)[name = tensor("xt_1")]; + tensor var_1093 = const()[name = tensor("op_1093"), val = tensor([1, 12, 1, 256])]; + tensor var_1094 = reshape(shape = var_1093, x = xt_1)[name = tensor("op_1094")]; + tensor var_1095_perm_0 = const()[name = tensor("op_1095_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_1098 = const()[name = tensor("op_1098"), val = tensor([1, 12, 256])]; + tensor var_1095 = transpose(perm = var_1095_perm_0, x = var_1094)[name = tensor("transpose_23")]; + tensor query_1 = reshape(shape = var_1098, x = var_1095)[name = tensor("query_1")]; tensor query_3_perm_0 = const()[name = tensor("query_3_perm_0"), val = tensor([1, 0, 2])]; tensor query_3 = transpose(perm = query_3_perm_0, x = query_1)[name = tensor("transpose_22")]; - tensor var_1118 = linear(bias = decoder_self_attn2_0_in_proj_bias, weight = decoder_self_attn2_0_in_proj_weight, x = query_3)[name = tensor("linear_43")]; + tensor var_1121 = linear(bias = inner_decoder_self_attn2_0_in_proj_bias, weight = inner_decoder_self_attn2_0_in_proj_weight, x = query_3)[name = tensor("linear_43")]; tensor concat_1 = const()[name = tensor("concat_1"), val = tensor([12, 1, 3, 256])]; - tensor var_1120 = reshape(shape = concat_1, x = var_1118)[name = tensor("op_1120")]; - tensor var_1121_axes_0 = const()[name = tensor("op_1121_axes_0"), val = tensor([0])]; - tensor var_1121 = expand_dims(axes = var_1121_axes_0, x = var_1120)[name = tensor("op_1121")]; - tensor var_1122_perm_0 = const()[name = tensor("op_1122_perm_0"), val = tensor([-2, 1, 2, 0, 4])]; - tensor var_1123_axes_0 = const()[name = tensor("op_1123_axes_0"), val = tensor([-2])]; - tensor var_1122 = transpose(perm = var_1122_perm_0, x = var_1121)[name = tensor("transpose_21")]; - tensor var_1123 = squeeze(axes = var_1123_axes_0, x = var_1122)[name = tensor("op_1123")]; + tensor var_1123 = reshape(shape = concat_1, x = var_1121)[name = tensor("op_1123")]; + tensor var_1124_axes_0 = const()[name = tensor("op_1124_axes_0"), val = tensor([0])]; + tensor var_1124 = expand_dims(axes = var_1124_axes_0, x = var_1123)[name = tensor("op_1124")]; + tensor var_1125_perm_0 = const()[name = tensor("op_1125_perm_0"), val = tensor([-2, 1, 2, 0, 4])]; + tensor var_1126_axes_0 = const()[name = tensor("op_1126_axes_0"), val = tensor([-2])]; + tensor var_1125 = transpose(perm = var_1125_perm_0, x = var_1124)[name = tensor("transpose_21")]; + tensor var_1126 = squeeze(axes = var_1126_axes_0, x = var_1125)[name = tensor("op_1126")]; tensor q_11_begin_0 = const()[name = tensor("q_11_begin_0"), val = tensor([0, 0, 0, 0])]; tensor q_11_end_0 = const()[name = tensor("q_11_end_0"), val = tensor([1, 12, 1, 256])]; tensor q_11_end_mask_0 = const()[name = tensor("q_11_end_mask_0"), val = tensor([false, true, true, true])]; tensor q_11_squeeze_mask_0 = const()[name = tensor("q_11_squeeze_mask_0"), val = tensor([true, false, false, false])]; - tensor q_11 = slice_by_index(begin = q_11_begin_0, end = q_11_end_0, end_mask = q_11_end_mask_0, squeeze_mask = q_11_squeeze_mask_0, x = var_1123)[name = tensor("q_11")]; + tensor q_11 = slice_by_index(begin = q_11_begin_0, end = q_11_end_0, end_mask = q_11_end_mask_0, squeeze_mask = q_11_squeeze_mask_0, x = var_1126)[name = tensor("q_11")]; tensor k_11_begin_0 = const()[name = tensor("k_11_begin_0"), val = tensor([1, 0, 0, 0])]; tensor k_11_end_0 = const()[name = tensor("k_11_end_0"), val = tensor([2, 12, 1, 256])]; tensor k_11_end_mask_0 = const()[name = tensor("k_11_end_mask_0"), val = tensor([false, true, true, true])]; tensor k_11_squeeze_mask_0 = const()[name = tensor("k_11_squeeze_mask_0"), val = tensor([true, false, false, false])]; - tensor k_11 = slice_by_index(begin = k_11_begin_0, end = k_11_end_0, end_mask = k_11_end_mask_0, squeeze_mask = k_11_squeeze_mask_0, x = var_1123)[name = tensor("k_11")]; + tensor k_11 = slice_by_index(begin = k_11_begin_0, end = k_11_end_0, end_mask = k_11_end_mask_0, squeeze_mask = k_11_squeeze_mask_0, x = var_1126)[name = tensor("k_11")]; tensor v_11_begin_0 = const()[name = tensor("v_11_begin_0"), val = tensor([2, 0, 0, 0])]; tensor v_11_end_0 = const()[name = tensor("v_11_end_0"), val = tensor([3, 12, 1, 256])]; tensor v_11_end_mask_0 = const()[name = tensor("v_11_end_mask_0"), val = tensor([false, true, true, true])]; tensor v_11_squeeze_mask_0 = const()[name = tensor("v_11_squeeze_mask_0"), val = tensor([true, false, false, false])]; - tensor v_11 = slice_by_index(begin = v_11_begin_0, end = v_11_end_0, end_mask = v_11_end_mask_0, squeeze_mask = v_11_squeeze_mask_0, x = var_1123)[name = tensor("v_11")]; - tensor var_1131 = const()[name = tensor("op_1131"), val = tensor([12, 4, 64])]; - tensor var_1132 = reshape(shape = var_1131, x = q_11)[name = tensor("op_1132")]; + tensor v_11 = slice_by_index(begin = v_11_begin_0, end = v_11_end_0, end_mask = v_11_end_mask_0, squeeze_mask = v_11_squeeze_mask_0, x = var_1126)[name = tensor("v_11")]; + tensor var_1134 = const()[name = tensor("op_1134"), val = tensor([12, 4, 64])]; + tensor var_1135 = reshape(shape = var_1134, x = q_11)[name = tensor("op_1135")]; tensor q_13_perm_0 = const()[name = tensor("q_13_perm_0"), val = tensor([1, 0, 2])]; - tensor var_1138 = const()[name = tensor("op_1138"), val = tensor([12, 4, 64])]; - tensor var_1139 = reshape(shape = var_1138, x = k_11)[name = tensor("op_1139")]; + tensor var_1141 = const()[name = tensor("op_1141"), val = tensor([12, 4, 64])]; + tensor var_1142 = reshape(shape = var_1141, x = k_11)[name = tensor("op_1142")]; tensor k_13_perm_0 = const()[name = tensor("k_13_perm_0"), val = tensor([1, 0, 2])]; - tensor var_1145 = const()[name = tensor("op_1145"), val = tensor([12, 4, 64])]; - tensor var_1146 = reshape(shape = var_1145, x = v_11)[name = tensor("op_1146")]; + tensor var_1148 = const()[name = tensor("op_1148"), val = tensor([12, 4, 64])]; + tensor var_1149 = reshape(shape = var_1148, x = v_11)[name = tensor("op_1149")]; tensor v_13_perm_0 = const()[name = tensor("v_13_perm_0"), val = tensor([1, 0, 2])]; - tensor var_1149 = const()[name = tensor("op_1149"), val = tensor([1, 4, 12, 64])]; - tensor q_13 = transpose(perm = q_13_perm_0, x = var_1132)[name = tensor("transpose_20")]; - tensor q_15 = reshape(shape = var_1149, x = q_13)[name = tensor("q_15")]; - tensor var_1151 = const()[name = tensor("op_1151"), val = tensor([1, 4, 12, 64])]; - tensor k_13 = transpose(perm = k_13_perm_0, x = var_1139)[name = tensor("transpose_19")]; - tensor k_15 = reshape(shape = var_1151, x = k_13)[name = tensor("k_15")]; - tensor var_1153 = const()[name = tensor("op_1153"), val = tensor([1, 4, 12, 64])]; - tensor v_13 = transpose(perm = v_13_perm_0, x = var_1146)[name = tensor("transpose_18")]; - tensor v_15 = reshape(shape = var_1153, x = v_13)[name = tensor("v_15")]; + tensor var_1152 = const()[name = tensor("op_1152"), val = tensor([1, 4, 12, 64])]; + tensor q_13 = transpose(perm = q_13_perm_0, x = var_1135)[name = tensor("transpose_20")]; + tensor q_15 = reshape(shape = var_1152, x = q_13)[name = tensor("q_15")]; + tensor var_1154 = const()[name = tensor("op_1154"), val = tensor([1, 4, 12, 64])]; + tensor k_13 = transpose(perm = k_13_perm_0, x = var_1142)[name = tensor("transpose_19")]; + tensor k_15 = reshape(shape = var_1154, x = k_13)[name = tensor("k_15")]; + tensor var_1156 = const()[name = tensor("op_1156"), val = tensor([1, 4, 12, 64])]; + tensor v_13 = transpose(perm = v_13_perm_0, x = var_1149)[name = tensor("transpose_18")]; + tensor v_15 = reshape(shape = var_1156, x = v_13)[name = tensor("v_15")]; tensor mul_1_y_0 = const()[name = tensor("mul_1_y_0"), val = tensor(0x1p-3)]; tensor mul_1 = mul(x = q_15, y = mul_1_y_0)[name = tensor("mul_1")]; tensor matmul_0_transpose_y_0 = const()[name = tensor("matmul_0_transpose_y_0"), val = tensor(true)]; @@ -981,30 +980,30 @@ program(1.0) tensor attn_output_1_transpose_x_0 = const()[name = tensor("attn_output_1_transpose_x_0"), val = tensor(false)]; tensor attn_output_1_transpose_y_0 = const()[name = tensor("attn_output_1_transpose_y_0"), val = tensor(false)]; tensor attn_output_1 = matmul(transpose_x = attn_output_1_transpose_x_0, transpose_y = attn_output_1_transpose_y_0, x = softmax_0, y = v_15)[name = tensor("attn_output_1")]; - tensor var_1156 = const()[name = tensor("op_1156"), val = tensor([2, 0, 1, 3])]; - tensor var_1161 = const()[name = tensor("op_1161"), val = tensor([12, 256])]; - tensor var_1157 = transpose(perm = var_1156, x = attn_output_1)[name = tensor("transpose_17")]; - tensor attn_output_3 = reshape(shape = var_1161, x = var_1157)[name = tensor("attn_output_3")]; - tensor attn_output_5 = linear(bias = decoder_self_attn2_0_out_proj_bias, weight = decoder_self_attn2_0_out_proj_weight, x = attn_output_3)[name = tensor("linear_44")]; - tensor var_1165 = const()[name = tensor("op_1165"), val = tensor([12, 1, 256])]; - tensor attn_output_7 = reshape(shape = var_1165, x = attn_output_5)[name = tensor("attn_output_7")]; + tensor var_1159 = const()[name = tensor("op_1159"), val = tensor([2, 0, 1, 3])]; + tensor var_1164 = const()[name = tensor("op_1164"), val = tensor([12, 256])]; + tensor var_1160 = transpose(perm = var_1159, x = attn_output_1)[name = tensor("transpose_17")]; + tensor attn_output_3 = reshape(shape = var_1164, x = var_1160)[name = tensor("attn_output_3")]; + tensor attn_output_5 = linear(bias = inner_decoder_self_attn2_0_out_proj_bias, weight = inner_decoder_self_attn2_0_out_proj_weight, x = attn_output_3)[name = tensor("linear_44")]; + tensor var_1168 = const()[name = tensor("op_1168"), val = tensor([12, 1, 256])]; + tensor attn_output_7 = reshape(shape = var_1168, x = attn_output_5)[name = tensor("attn_output_7")]; tensor sa2_out_1_perm_0 = const()[name = tensor("sa2_out_1_perm_0"), val = tensor([1, 0, 2])]; tensor sa2_out_1 = transpose(perm = sa2_out_1_perm_0, x = attn_output_7)[name = tensor("transpose_16")]; - tensor input_175 = add(x = query_1, y = sa2_out_1)[name = tensor("input_175")]; - tensor input_177_axes_0 = const()[name = tensor("input_177_axes_0"), val = tensor([-1])]; - tensor input_177 = layer_norm(axes = input_177_axes_0, beta = decoder_norm21_0_bias, epsilon = var_921, gamma = decoder_norm21_0_weight, x = input_175)[name = tensor("input_177")]; - tensor input_179 = linear(bias = decoder_linear1_0_bias, weight = decoder_linear1_0_weight, x = input_177)[name = tensor("linear_45")]; - tensor input_181 = relu(x = input_179)[name = tensor("input_181")]; - tensor ffn_out_1 = linear(bias = decoder_linear2_0_bias, weight = decoder_linear2_0_weight, x = input_181)[name = tensor("linear_46")]; - tensor input_183 = add(x = input_177, y = ffn_out_1)[name = tensor("input_183")]; + tensor input_177 = add(x = query_1, y = sa2_out_1)[name = tensor("input_177")]; + tensor input_179_axes_0 = const()[name = tensor("input_179_axes_0"), val = tensor([-1])]; + tensor input_179 = layer_norm(axes = input_179_axes_0, beta = inner_decoder_norm21_0_bias, epsilon = var_36, gamma = inner_decoder_norm21_0_weight, x = input_177)[name = tensor("input_179")]; + tensor input_181 = linear(bias = inner_decoder_linear1_0_bias, weight = inner_decoder_linear1_0_weight, x = input_179)[name = tensor("linear_45")]; + tensor input_183 = relu(x = input_181)[name = tensor("input_183")]; + tensor ffn_out_1 = linear(bias = inner_decoder_linear2_0_bias, weight = inner_decoder_linear2_0_weight, x = input_183)[name = tensor("linear_46")]; + tensor input_185 = add(x = input_179, y = ffn_out_1)[name = tensor("input_185")]; tensor xt_3_axes_0 = const()[name = tensor("xt_3_axes_0"), val = tensor([-1])]; - tensor xt_3 = layer_norm(axes = xt_3_axes_0, beta = decoder_norm22_0_bias, epsilon = var_921, gamma = decoder_norm22_0_weight, x = input_183)[name = tensor("xt_3")]; - tensor var_1185 = const()[name = tensor("op_1185"), val = tensor([1, 1, 12, 256])]; - tensor x_31 = reshape(shape = var_1185, x = xt_3)[name = tensor("x_31")]; - tensor var_1187_perm_0 = const()[name = tensor("op_1187_perm_0"), val = tensor([0, 2, 1, 3])]; - tensor var_1191 = const()[name = tensor("op_1191"), val = tensor([12, 1, 256])]; - tensor var_1187 = transpose(perm = var_1187_perm_0, x = x_31)[name = tensor("transpose_15")]; - tensor x = reshape(shape = var_1191, x = var_1187)[name = tensor("x")]; + tensor xt_3 = layer_norm(axes = xt_3_axes_0, beta = inner_decoder_norm22_0_bias, epsilon = var_36, gamma = inner_decoder_norm22_0_weight, x = input_185)[name = tensor("xt_3")]; + tensor var_1188 = const()[name = tensor("op_1188"), val = tensor([1, 1, 12, 256])]; + tensor x_31 = reshape(shape = var_1188, x = xt_3)[name = tensor("x_31")]; + tensor var_1190_perm_0 = const()[name = tensor("op_1190_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_1194 = const()[name = tensor("op_1194"), val = tensor([12, 1, 256])]; + tensor var_1190 = transpose(perm = var_1190_perm_0, x = x_31)[name = tensor("transpose_15")]; + tensor x = reshape(shape = var_1194, x = var_1190)[name = tensor("x")]; tensor prev_kv_begin_0 = const()[name = tensor("prev_kv_begin_0"), val = tensor([1, 0, 0, 0, 0])]; tensor prev_kv_end_0 = const()[name = tensor("prev_kv_end_0"), val = tensor([2, 12, 4, 64, 64])]; tensor prev_kv_end_mask_0 = const()[name = tensor("prev_kv_end_mask_0"), val = tensor([false, true, true, true, true])]; @@ -1015,120 +1014,120 @@ program(1.0) tensor prev_scale_end_mask_0 = const()[name = tensor("prev_scale_end_mask_0"), val = tensor([false, true])]; tensor prev_scale_squeeze_mask_0 = const()[name = tensor("prev_scale_squeeze_mask_0"), val = tensor([true, false])]; tensor prev_scale = slice_by_index(begin = prev_scale_begin_0, end = prev_scale_end_0, end_mask = prev_scale_end_mask_0, squeeze_mask = prev_scale_squeeze_mask_0, x = dec_scale)[name = tensor("prev_scale")]; - tensor var_1199 = linear(bias = decoder_q_proj_1_bias, weight = decoder_q_proj_1_weight, x = x)[name = tensor("linear_47")]; - tensor var_1200 = const()[name = tensor("op_1200"), val = tensor([12, 1, 4, 64])]; - tensor var_1201 = reshape(shape = var_1200, x = var_1199)[name = tensor("op_1201")]; + tensor var_1202 = linear(bias = inner_decoder_q_proj_1_bias, weight = inner_decoder_q_proj_1_weight, x = x)[name = tensor("linear_47")]; + tensor var_1203 = const()[name = tensor("op_1203"), val = tensor([12, 1, 4, 64])]; + tensor var_1204 = reshape(shape = var_1203, x = var_1202)[name = tensor("op_1204")]; tensor q_17_perm_0 = const()[name = tensor("q_17_perm_0"), val = tensor([0, 2, 1, 3])]; - tensor var_1205 = linear(bias = decoder_k_proj_1_bias, weight = decoder_k_proj_1_weight, x = x)[name = tensor("linear_48")]; - tensor var_1206 = const()[name = tensor("op_1206"), val = tensor(0x1p-3)]; - tensor var_1207 = mul(x = var_1205, y = var_1206)[name = tensor("op_1207")]; - tensor var_1208 = const()[name = tensor("op_1208"), val = tensor([12, 1, 4, 64])]; - tensor var_1209 = reshape(shape = var_1208, x = var_1207)[name = tensor("op_1209")]; + tensor var_1208 = linear(bias = inner_decoder_k_proj_1_bias, weight = inner_decoder_k_proj_1_weight, x = x)[name = tensor("linear_48")]; + tensor var_1209 = const()[name = tensor("op_1209"), val = tensor(0x1p-3)]; + tensor var_1210 = mul(x = var_1208, y = var_1209)[name = tensor("op_1210")]; + tensor var_1211 = const()[name = tensor("op_1211"), val = tensor([12, 1, 4, 64])]; + tensor var_1212 = reshape(shape = var_1211, x = var_1210)[name = tensor("op_1212")]; tensor k_17_perm_0 = const()[name = tensor("k_17_perm_0"), val = tensor([0, 2, 1, 3])]; - tensor var_1213 = linear(bias = decoder_v_proj_1_bias, weight = decoder_v_proj_1_weight, x = x)[name = tensor("linear_49")]; - tensor var_1214 = const()[name = tensor("op_1214"), val = tensor([12, 1, 4, 64])]; - tensor var_1215 = reshape(shape = var_1214, x = var_1213)[name = tensor("op_1215")]; + tensor var_1216 = linear(bias = inner_decoder_v_proj_1_bias, weight = inner_decoder_v_proj_1_weight, x = x)[name = tensor("linear_49")]; + tensor var_1217 = const()[name = tensor("op_1217"), val = tensor([12, 1, 4, 64])]; + tensor var_1218 = reshape(shape = var_1217, x = var_1216)[name = tensor("op_1218")]; tensor v_17_perm_0 = const()[name = tensor("v_17_perm_0"), val = tensor([0, 2, 1, 3])]; - tensor input_187 = linear(bias = decoder_g_proj_1_bias, weight = decoder_g_proj_1_weight, x = x)[name = tensor("linear_50")]; + tensor input_189 = linear(bias = inner_decoder_g_proj_1_bias, weight = inner_decoder_g_proj_1_weight, x = x)[name = tensor("linear_50")]; tensor sqrt_s0 = sqrt(x = prev_scale)[name = tensor("sqrt_s0")]; tensor s_t = add(x = prev_scale, y = cumsum_mask_1)[name = tensor("s_t")]; tensor const_32 = const()[name = tensor("const_32"), val = tensor(0x1.fffffep+127)]; - tensor clip_3 = clip(alpha = var_926, beta = const_32, x = s_t)[name = tensor("clip_3")]; + tensor clip_3 = clip(alpha = var_29, beta = const_32, x = s_t)[name = tensor("clip_3")]; tensor sqrt_s_t = sqrt(x = clip_3)[name = tensor("sqrt_s_t")]; tensor qk_transpose_x_1 = const()[name = tensor("qk_transpose_x_1"), val = tensor(false)]; tensor qk_transpose_y_1 = const()[name = tensor("qk_transpose_y_1"), val = tensor(true)]; - tensor k_17 = transpose(perm = k_17_perm_0, x = var_1209)[name = tensor("transpose_13")]; - tensor q_17 = transpose(perm = q_17_perm_0, x = var_1201)[name = tensor("transpose_14")]; + tensor k_17 = transpose(perm = k_17_perm_0, x = var_1212)[name = tensor("transpose_13")]; + tensor q_17 = transpose(perm = q_17_perm_0, x = var_1204)[name = tensor("transpose_14")]; tensor qk = matmul(transpose_x = qk_transpose_x_1, transpose_y = qk_transpose_y_1, x = q_17, y = k_17)[name = tensor("qk")]; - tensor var_1230 = const()[name = tensor("op_1230"), val = tensor([1, 1])]; - tensor var_1231 = reshape(shape = var_1230, x = sqrt_s_t)[name = tensor("op_1231")]; - tensor M = real_div(x = var_1047, y = var_1231)[name = tensor("M")]; - tensor var_1233 = mul(x = qk, y = M)[name = tensor("op_1233")]; - tensor inner_transpose_x_0 = const()[name = tensor("inner_transpose_x_0"), val = tensor(false)]; - tensor inner_transpose_y_0 = const()[name = tensor("inner_transpose_y_0"), val = tensor(false)]; - tensor v_17 = transpose(perm = v_17_perm_0, x = var_1215)[name = tensor("transpose_12")]; - tensor inner = matmul(transpose_x = inner_transpose_x_0, transpose_y = inner_transpose_y_0, x = var_1233, y = v_17)[name = tensor("inner")]; - tensor var_1235_transpose_x_0 = const()[name = tensor("op_1235_transpose_x_0"), val = tensor(false)]; - tensor var_1235_transpose_y_0 = const()[name = tensor("op_1235_transpose_y_0"), val = tensor(false)]; - tensor var_1235 = matmul(transpose_x = var_1235_transpose_x_0, transpose_y = var_1235_transpose_y_0, x = q_17, y = prev_kv)[name = tensor("op_1235")]; - tensor var_1236 = real_div(x = sqrt_s0, y = sqrt_s_t)[name = tensor("op_1236")]; - tensor var_1237 = const()[name = tensor("op_1237"), val = tensor([1, 1, 1, 1])]; - tensor var_1238 = reshape(shape = var_1237, x = var_1236)[name = tensor("op_1238")]; - tensor cross = mul(x = var_1235, y = var_1238)[name = tensor("cross")]; - tensor out_31 = add(x = inner, y = cross)[name = tensor("out_31")]; - tensor v_masked = mul(x = v_17, y = var_1061)[name = tensor("v_masked")]; - tensor var_1244 = mul(x = prev_kv, y = sqrt_s0)[name = tensor("op_1244")]; - tensor var_1246_transpose_x_1 = const()[name = tensor("op_1246_transpose_x_1"), val = tensor(true)]; - tensor var_1246_transpose_y_1 = const()[name = tensor("op_1246_transpose_y_1"), val = tensor(false)]; - tensor var_1246 = matmul(transpose_x = var_1246_transpose_x_1, transpose_y = var_1246_transpose_y_1, x = k_17, y = v_masked)[name = tensor("op_1246")]; - tensor new_kv_unnorm = add(x = var_1244, y = var_1246)[name = tensor("new_kv_unnorm")]; - tensor new_scale = add(x = prev_scale, y = var_1069)[name = tensor("new_scale")]; + tensor var_1233 = const()[name = tensor("op_1233"), val = tensor([1, 1])]; + tensor var_1234 = reshape(shape = var_1233, x = sqrt_s_t)[name = tensor("op_1234")]; + tensor M = real_div(x = var_1050, y = var_1234)[name = tensor("M")]; + tensor var_1236 = mul(x = qk, y = M)[name = tensor("op_1236")]; + tensor inner_11_transpose_x_0 = const()[name = tensor("inner_11_transpose_x_0"), val = tensor(false)]; + tensor inner_11_transpose_y_0 = const()[name = tensor("inner_11_transpose_y_0"), val = tensor(false)]; + tensor v_17 = transpose(perm = v_17_perm_0, x = var_1218)[name = tensor("transpose_12")]; + tensor inner_11 = matmul(transpose_x = inner_11_transpose_x_0, transpose_y = inner_11_transpose_y_0, x = var_1236, y = v_17)[name = tensor("inner_11")]; + tensor var_1238_transpose_x_0 = const()[name = tensor("op_1238_transpose_x_0"), val = tensor(false)]; + tensor var_1238_transpose_y_0 = const()[name = tensor("op_1238_transpose_y_0"), val = tensor(false)]; + tensor var_1238 = matmul(transpose_x = var_1238_transpose_x_0, transpose_y = var_1238_transpose_y_0, x = q_17, y = prev_kv)[name = tensor("op_1238")]; + tensor var_1239 = real_div(x = sqrt_s0, y = sqrt_s_t)[name = tensor("op_1239")]; + tensor var_1240 = const()[name = tensor("op_1240"), val = tensor([1, 1, 1, 1])]; + tensor var_1241 = reshape(shape = var_1240, x = var_1239)[name = tensor("op_1241")]; + tensor cross = mul(x = var_1238, y = var_1241)[name = tensor("cross")]; + tensor out_31 = add(x = inner_11, y = cross)[name = tensor("out_31")]; + tensor v_masked = mul(x = v_17, y = var_1064)[name = tensor("v_masked")]; + tensor var_1247 = mul(x = prev_kv, y = sqrt_s0)[name = tensor("op_1247")]; + tensor var_1249_transpose_x_1 = const()[name = tensor("op_1249_transpose_x_1"), val = tensor(true)]; + tensor var_1249_transpose_y_1 = const()[name = tensor("op_1249_transpose_y_1"), val = tensor(false)]; + tensor var_1249 = matmul(transpose_x = var_1249_transpose_x_1, transpose_y = var_1249_transpose_y_1, x = k_17, y = v_masked)[name = tensor("op_1249")]; + tensor new_kv_unnorm = add(x = var_1247, y = var_1249)[name = tensor("new_kv_unnorm")]; + tensor new_scale = add(x = prev_scale, y = var_1072)[name = tensor("new_scale")]; tensor const_33 = const()[name = tensor("const_33"), val = tensor(0x1.fffffep+127)]; - tensor clip_4 = clip(alpha = var_926, beta = const_33, x = new_scale)[name = tensor("clip_4")]; + tensor clip_4 = clip(alpha = var_29, beta = const_33, x = new_scale)[name = tensor("clip_4")]; tensor sqrt_new_scale = sqrt(x = clip_4)[name = tensor("sqrt_new_scale")]; tensor nkv = real_div(x = new_kv_unnorm, y = sqrt_new_scale)[name = tensor("nkv")]; - tensor var_1255_perm_0 = const()[name = tensor("op_1255_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_1258_perm_0 = const()[name = tensor("op_1258_perm_0"), val = tensor([0, 2, 1, 3])]; tensor out_33_axes_0 = const()[name = tensor("out_33_axes_0"), val = tensor([-1])]; - tensor var_1255 = transpose(perm = var_1255_perm_0, x = out_31)[name = tensor("transpose_11")]; - tensor out_33 = layer_norm(axes = out_33_axes_0, epsilon = var_923, x = var_1255)[name = tensor("out_33")]; - tensor var_1259 = const()[name = tensor("op_1259"), val = tensor([12, 1, 256])]; - tensor out = reshape(shape = var_1259, x = out_33)[name = tensor("out")]; - tensor var_1261 = silu(x = input_187)[name = tensor("op_1261")]; - tensor input_189 = mul(x = var_1261, y = out)[name = tensor("input_189")]; - tensor ret_out = linear(bias = decoder_out_proj_1_bias, weight = decoder_out_proj_1_weight, x = input_189)[name = tensor("linear_51")]; - tensor input_191 = add(x = x, y = ret_out)[name = tensor("input_191")]; + tensor var_1258 = transpose(perm = var_1258_perm_0, x = out_31)[name = tensor("transpose_11")]; + tensor out_33 = layer_norm(axes = out_33_axes_0, epsilon = var_44, x = var_1258)[name = tensor("out_33")]; + tensor var_1262 = const()[name = tensor("op_1262"), val = tensor([12, 1, 256])]; + tensor out = reshape(shape = var_1262, x = out_33)[name = tensor("out")]; + tensor var_1264 = silu(x = input_189)[name = tensor("op_1264")]; + tensor input_191 = mul(x = var_1264, y = out)[name = tensor("input_191")]; + tensor ret_out = linear(bias = inner_decoder_out_proj_1_bias, weight = inner_decoder_out_proj_1_weight, x = input_191)[name = tensor("linear_51")]; + tensor input_193 = add(x = x, y = ret_out)[name = tensor("input_193")]; tensor xt_5_axes_0 = const()[name = tensor("xt_5_axes_0"), val = tensor([-1])]; - tensor xt_5 = layer_norm(axes = xt_5_axes_0, beta = decoder_norm11_1_bias, epsilon = var_921, gamma = decoder_norm11_1_weight, x = input_191)[name = tensor("xt_5")]; - tensor var_1271 = const()[name = tensor("op_1271"), val = tensor([1, 12, 1, 256])]; - tensor var_1272 = reshape(shape = var_1271, x = xt_5)[name = tensor("op_1272")]; - tensor var_1273_perm_0 = const()[name = tensor("op_1273_perm_0"), val = tensor([0, 2, 1, 3])]; - tensor var_1276 = const()[name = tensor("op_1276"), val = tensor([1, 12, 256])]; - tensor var_1273 = transpose(perm = var_1273_perm_0, x = var_1272)[name = tensor("transpose_10")]; - tensor query_5 = reshape(shape = var_1276, x = var_1273)[name = tensor("query_5")]; + tensor xt_5 = layer_norm(axes = xt_5_axes_0, beta = inner_decoder_norm11_1_bias, epsilon = var_36, gamma = inner_decoder_norm11_1_weight, x = input_193)[name = tensor("xt_5")]; + tensor var_1274 = const()[name = tensor("op_1274"), val = tensor([1, 12, 1, 256])]; + tensor var_1275 = reshape(shape = var_1274, x = xt_5)[name = tensor("op_1275")]; + tensor var_1276_perm_0 = const()[name = tensor("op_1276_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_1279 = const()[name = tensor("op_1279"), val = tensor([1, 12, 256])]; + tensor var_1276 = transpose(perm = var_1276_perm_0, x = var_1275)[name = tensor("transpose_10")]; + tensor query_5 = reshape(shape = var_1279, x = var_1276)[name = tensor("query_5")]; tensor query_perm_0 = const()[name = tensor("query_perm_0"), val = tensor([1, 0, 2])]; tensor query = transpose(perm = query_perm_0, x = query_5)[name = tensor("transpose_9")]; - tensor var_1299 = linear(bias = decoder_self_attn2_1_in_proj_bias, weight = decoder_self_attn2_1_in_proj_weight, x = query)[name = tensor("linear_52")]; + tensor var_1302 = linear(bias = inner_decoder_self_attn2_1_in_proj_bias, weight = inner_decoder_self_attn2_1_in_proj_weight, x = query)[name = tensor("linear_52")]; tensor concat_2 = const()[name = tensor("concat_2"), val = tensor([12, 1, 3, 256])]; - tensor var_1301 = reshape(shape = concat_2, x = var_1299)[name = tensor("op_1301")]; - tensor var_1302_axes_0 = const()[name = tensor("op_1302_axes_0"), val = tensor([0])]; - tensor var_1302 = expand_dims(axes = var_1302_axes_0, x = var_1301)[name = tensor("op_1302")]; - tensor var_1303_perm_0 = const()[name = tensor("op_1303_perm_0"), val = tensor([-2, 1, 2, 0, 4])]; - tensor var_1304_axes_0 = const()[name = tensor("op_1304_axes_0"), val = tensor([-2])]; - tensor var_1303 = transpose(perm = var_1303_perm_0, x = var_1302)[name = tensor("transpose_8")]; - tensor var_1304 = squeeze(axes = var_1304_axes_0, x = var_1303)[name = tensor("op_1304")]; + tensor var_1304 = reshape(shape = concat_2, x = var_1302)[name = tensor("op_1304")]; + tensor var_1305_axes_0 = const()[name = tensor("op_1305_axes_0"), val = tensor([0])]; + tensor var_1305 = expand_dims(axes = var_1305_axes_0, x = var_1304)[name = tensor("op_1305")]; + tensor var_1306_perm_0 = const()[name = tensor("op_1306_perm_0"), val = tensor([-2, 1, 2, 0, 4])]; + tensor var_1307_axes_0 = const()[name = tensor("op_1307_axes_0"), val = tensor([-2])]; + tensor var_1306 = transpose(perm = var_1306_perm_0, x = var_1305)[name = tensor("transpose_8")]; + tensor var_1307 = squeeze(axes = var_1307_axes_0, x = var_1306)[name = tensor("op_1307")]; tensor q_19_begin_0 = const()[name = tensor("q_19_begin_0"), val = tensor([0, 0, 0, 0])]; tensor q_19_end_0 = const()[name = tensor("q_19_end_0"), val = tensor([1, 12, 1, 256])]; tensor q_19_end_mask_0 = const()[name = tensor("q_19_end_mask_0"), val = tensor([false, true, true, true])]; tensor q_19_squeeze_mask_0 = const()[name = tensor("q_19_squeeze_mask_0"), val = tensor([true, false, false, false])]; - tensor q_19 = slice_by_index(begin = q_19_begin_0, end = q_19_end_0, end_mask = q_19_end_mask_0, squeeze_mask = q_19_squeeze_mask_0, x = var_1304)[name = tensor("q_19")]; + tensor q_19 = slice_by_index(begin = q_19_begin_0, end = q_19_end_0, end_mask = q_19_end_mask_0, squeeze_mask = q_19_squeeze_mask_0, x = var_1307)[name = tensor("q_19")]; tensor k_19_begin_0 = const()[name = tensor("k_19_begin_0"), val = tensor([1, 0, 0, 0])]; tensor k_19_end_0 = const()[name = tensor("k_19_end_0"), val = tensor([2, 12, 1, 256])]; tensor k_19_end_mask_0 = const()[name = tensor("k_19_end_mask_0"), val = tensor([false, true, true, true])]; tensor k_19_squeeze_mask_0 = const()[name = tensor("k_19_squeeze_mask_0"), val = tensor([true, false, false, false])]; - tensor k_19 = slice_by_index(begin = k_19_begin_0, end = k_19_end_0, end_mask = k_19_end_mask_0, squeeze_mask = k_19_squeeze_mask_0, x = var_1304)[name = tensor("k_19")]; + tensor k_19 = slice_by_index(begin = k_19_begin_0, end = k_19_end_0, end_mask = k_19_end_mask_0, squeeze_mask = k_19_squeeze_mask_0, x = var_1307)[name = tensor("k_19")]; tensor v_19_begin_0 = const()[name = tensor("v_19_begin_0"), val = tensor([2, 0, 0, 0])]; tensor v_19_end_0 = const()[name = tensor("v_19_end_0"), val = tensor([3, 12, 1, 256])]; tensor v_19_end_mask_0 = const()[name = tensor("v_19_end_mask_0"), val = tensor([false, true, true, true])]; tensor v_19_squeeze_mask_0 = const()[name = tensor("v_19_squeeze_mask_0"), val = tensor([true, false, false, false])]; - tensor v_19 = slice_by_index(begin = v_19_begin_0, end = v_19_end_0, end_mask = v_19_end_mask_0, squeeze_mask = v_19_squeeze_mask_0, x = var_1304)[name = tensor("v_19")]; - tensor var_1312 = const()[name = tensor("op_1312"), val = tensor([12, 4, 64])]; - tensor var_1313 = reshape(shape = var_1312, x = q_19)[name = tensor("op_1313")]; + tensor v_19 = slice_by_index(begin = v_19_begin_0, end = v_19_end_0, end_mask = v_19_end_mask_0, squeeze_mask = v_19_squeeze_mask_0, x = var_1307)[name = tensor("v_19")]; + tensor var_1315 = const()[name = tensor("op_1315"), val = tensor([12, 4, 64])]; + tensor var_1316 = reshape(shape = var_1315, x = q_19)[name = tensor("op_1316")]; tensor q_21_perm_0 = const()[name = tensor("q_21_perm_0"), val = tensor([1, 0, 2])]; - tensor var_1319 = const()[name = tensor("op_1319"), val = tensor([12, 4, 64])]; - tensor var_1320 = reshape(shape = var_1319, x = k_19)[name = tensor("op_1320")]; + tensor var_1322 = const()[name = tensor("op_1322"), val = tensor([12, 4, 64])]; + tensor var_1323 = reshape(shape = var_1322, x = k_19)[name = tensor("op_1323")]; tensor k_21_perm_0 = const()[name = tensor("k_21_perm_0"), val = tensor([1, 0, 2])]; - tensor var_1326 = const()[name = tensor("op_1326"), val = tensor([12, 4, 64])]; - tensor var_1327 = reshape(shape = var_1326, x = v_19)[name = tensor("op_1327")]; + tensor var_1329 = const()[name = tensor("op_1329"), val = tensor([12, 4, 64])]; + tensor var_1330 = reshape(shape = var_1329, x = v_19)[name = tensor("op_1330")]; tensor v_21_perm_0 = const()[name = tensor("v_21_perm_0"), val = tensor([1, 0, 2])]; - tensor var_1330 = const()[name = tensor("op_1330"), val = tensor([1, 4, 12, 64])]; - tensor q_21 = transpose(perm = q_21_perm_0, x = var_1313)[name = tensor("transpose_7")]; - tensor q = reshape(shape = var_1330, x = q_21)[name = tensor("q")]; - tensor var_1332 = const()[name = tensor("op_1332"), val = tensor([1, 4, 12, 64])]; - tensor k_21 = transpose(perm = k_21_perm_0, x = var_1320)[name = tensor("transpose_6")]; - tensor k = reshape(shape = var_1332, x = k_21)[name = tensor("k")]; - tensor var_1334 = const()[name = tensor("op_1334"), val = tensor([1, 4, 12, 64])]; - tensor v_21 = transpose(perm = v_21_perm_0, x = var_1327)[name = tensor("transpose_5")]; - tensor v = reshape(shape = var_1334, x = v_21)[name = tensor("v")]; + tensor var_1333 = const()[name = tensor("op_1333"), val = tensor([1, 4, 12, 64])]; + tensor q_21 = transpose(perm = q_21_perm_0, x = var_1316)[name = tensor("transpose_7")]; + tensor q = reshape(shape = var_1333, x = q_21)[name = tensor("q")]; + tensor var_1335 = const()[name = tensor("op_1335"), val = tensor([1, 4, 12, 64])]; + tensor k_21 = transpose(perm = k_21_perm_0, x = var_1323)[name = tensor("transpose_6")]; + tensor k = reshape(shape = var_1335, x = k_21)[name = tensor("k")]; + tensor var_1337 = const()[name = tensor("op_1337"), val = tensor([1, 4, 12, 64])]; + tensor v_21 = transpose(perm = v_21_perm_0, x = var_1330)[name = tensor("transpose_5")]; + tensor v = reshape(shape = var_1337, x = v_21)[name = tensor("v")]; tensor mul_3_y_0 = const()[name = tensor("mul_3_y_0"), val = tensor(0x1p-3)]; tensor mul_3 = mul(x = q, y = mul_3_y_0)[name = tensor("mul_3")]; tensor matmul_1_transpose_y_0 = const()[name = tensor("matmul_1_transpose_y_0"), val = tensor(true)]; @@ -1139,34 +1138,34 @@ program(1.0) tensor attn_output_9_transpose_x_0 = const()[name = tensor("attn_output_9_transpose_x_0"), val = tensor(false)]; tensor attn_output_9_transpose_y_0 = const()[name = tensor("attn_output_9_transpose_y_0"), val = tensor(false)]; tensor attn_output_9 = matmul(transpose_x = attn_output_9_transpose_x_0, transpose_y = attn_output_9_transpose_y_0, x = softmax_1, y = v)[name = tensor("attn_output_9")]; - tensor var_1337 = const()[name = tensor("op_1337"), val = tensor([2, 0, 1, 3])]; - tensor var_1342 = const()[name = tensor("op_1342"), val = tensor([12, 256])]; - tensor var_1338 = transpose(perm = var_1337, x = attn_output_9)[name = tensor("transpose_4")]; - tensor attn_output_11 = reshape(shape = var_1342, x = var_1338)[name = tensor("attn_output_11")]; - tensor attn_output_13 = linear(bias = decoder_self_attn2_1_out_proj_bias, weight = decoder_self_attn2_1_out_proj_weight, x = attn_output_11)[name = tensor("linear_53")]; - tensor var_1346 = const()[name = tensor("op_1346"), val = tensor([12, 1, 256])]; - tensor attn_output = reshape(shape = var_1346, x = attn_output_13)[name = tensor("attn_output")]; + tensor var_1340 = const()[name = tensor("op_1340"), val = tensor([2, 0, 1, 3])]; + tensor var_1345 = const()[name = tensor("op_1345"), val = tensor([12, 256])]; + tensor var_1341 = transpose(perm = var_1340, x = attn_output_9)[name = tensor("transpose_4")]; + tensor attn_output_11 = reshape(shape = var_1345, x = var_1341)[name = tensor("attn_output_11")]; + tensor attn_output_13 = linear(bias = inner_decoder_self_attn2_1_out_proj_bias, weight = inner_decoder_self_attn2_1_out_proj_weight, x = attn_output_11)[name = tensor("linear_53")]; + tensor var_1349 = const()[name = tensor("op_1349"), val = tensor([12, 1, 256])]; + tensor attn_output = reshape(shape = var_1349, x = attn_output_13)[name = tensor("attn_output")]; tensor sa2_out_perm_0 = const()[name = tensor("sa2_out_perm_0"), val = tensor([1, 0, 2])]; tensor sa2_out = transpose(perm = sa2_out_perm_0, x = attn_output)[name = tensor("transpose_3")]; - tensor input_193 = add(x = query_5, y = sa2_out)[name = tensor("input_193")]; - tensor input_195_axes_0 = const()[name = tensor("input_195_axes_0"), val = tensor([-1])]; - tensor input_195 = layer_norm(axes = input_195_axes_0, beta = decoder_norm21_1_bias, epsilon = var_921, gamma = decoder_norm21_1_weight, x = input_193)[name = tensor("input_195")]; - tensor input_197 = linear(bias = decoder_linear1_1_bias, weight = decoder_linear1_1_weight, x = input_195)[name = tensor("linear_54")]; - tensor input_199 = relu(x = input_197)[name = tensor("input_199")]; - tensor ffn_out = linear(bias = decoder_linear2_1_bias, weight = decoder_linear2_1_weight, x = input_199)[name = tensor("linear_55")]; - tensor input_201 = add(x = input_195, y = ffn_out)[name = tensor("input_201")]; + tensor input_195 = add(x = query_5, y = sa2_out)[name = tensor("input_195")]; + tensor input_197_axes_0 = const()[name = tensor("input_197_axes_0"), val = tensor([-1])]; + tensor input_197 = layer_norm(axes = input_197_axes_0, beta = inner_decoder_norm21_1_bias, epsilon = var_36, gamma = inner_decoder_norm21_1_weight, x = input_195)[name = tensor("input_197")]; + tensor input_199 = linear(bias = inner_decoder_linear1_1_bias, weight = inner_decoder_linear1_1_weight, x = input_197)[name = tensor("linear_54")]; + tensor input_201 = relu(x = input_199)[name = tensor("input_201")]; + tensor ffn_out = linear(bias = inner_decoder_linear2_1_bias, weight = inner_decoder_linear2_1_weight, x = input_201)[name = tensor("linear_55")]; + tensor input_203 = add(x = input_197, y = ffn_out)[name = tensor("input_203")]; tensor xt_axes_0 = const()[name = tensor("xt_axes_0"), val = tensor([-1])]; - tensor xt = layer_norm(axes = xt_axes_0, beta = decoder_norm22_1_bias, epsilon = var_921, gamma = decoder_norm22_1_weight, x = input_201)[name = tensor("xt")]; - tensor var_1366 = const()[name = tensor("op_1366"), val = tensor([1, 1, 12, 256])]; - tensor input = reshape(shape = var_1366, x = xt)[name = tensor("input")]; - tensor var_1368 = const()[name = tensor("op_1368"), val = tensor([-1])]; - tensor var_1369 = reduce_l2_norm(axes = var_1368, keep_dims = var_924, x = input)[name = tensor("op_1369")]; + tensor xt = layer_norm(axes = xt_axes_0, beta = inner_decoder_norm22_1_bias, epsilon = var_36, gamma = inner_decoder_norm22_1_weight, x = input_203)[name = tensor("xt")]; + tensor var_1369 = const()[name = tensor("op_1369"), val = tensor([1, 1, 12, 256])]; + tensor input = reshape(shape = var_1369, x = xt)[name = tensor("input")]; + tensor var_1371 = const()[name = tensor("op_1371"), val = tensor([-1])]; + tensor var_1372 = reduce_l2_norm(axes = var_1371, keep_dims = var_35, x = input)[name = tensor("op_1372")]; tensor const_42 = const()[name = tensor("const_42"), val = tensor(0x1.fffffep+127)]; - tensor clip_5 = clip(alpha = var_916, beta = const_42, x = var_1369)[name = tensor("clip_5")]; - tensor var_1371 = real_div(x = input, y = clip_5)[name = tensor("op_1371")]; + tensor clip_5 = clip(alpha = var_49, beta = const_42, x = var_1372)[name = tensor("clip_5")]; + tensor var_1374 = real_div(x = input, y = clip_5)[name = tensor("op_1374")]; tensor transpose_1_perm_0 = const()[name = tensor("transpose_1_perm_0"), val = tensor([0, 1, 3, 2])]; tensor concat_7 = const()[name = tensor("concat_7"), val = tensor([1, 256, 12])]; - tensor transpose_1 = transpose(perm = transpose_1_perm_0, x = var_1371)[name = tensor("transpose_2")]; + tensor transpose_1 = transpose(perm = transpose_1_perm_0, x = var_1374)[name = tensor("transpose_2")]; tensor reshape_1 = reshape(shape = concat_7, x = transpose_1)[name = tensor("reshape_1")]; tensor matmul_2_transpose_x_0 = const()[name = tensor("matmul_2_transpose_x_0"), val = tensor(false)]; tensor matmul_2_transpose_y_0 = const()[name = tensor("matmul_2_transpose_y_0"), val = tensor(false)]; @@ -1175,10 +1174,10 @@ program(1.0) tensor output_end_0 = const()[name = tensor("output_end_0"), val = tensor([1, 1, 11])]; tensor output_end_mask_0 = const()[name = tensor("output_end_mask_0"), val = tensor([true, true, false])]; tensor output = slice_by_index(begin = output_begin_0, end = output_end_0, end_mask = output_end_mask_0, x = matmul_2)[name = tensor("output")]; - tensor probs = sigmoid(x = output)[name = tensor("op_1375")]; - tensor var_1377_axis_0 = const()[name = tensor("op_1377_axis_0"), val = tensor(0)]; - tensor dec_kv_new = stack(axis = var_1377_axis_0, values = (var_1073, nkv))[name = tensor("op_1377")]; - tensor var_1379_axis_0 = const()[name = tensor("op_1379_axis_0"), val = tensor(0)]; - tensor dec_scale_new = stack(axis = var_1379_axis_0, values = (new_scale_9, new_scale))[name = tensor("op_1379")]; + tensor probs = sigmoid(x = output)[name = tensor("op_1378")]; + tensor var_1380_axis_0 = const()[name = tensor("op_1380_axis_0"), val = tensor(0)]; + tensor dec_kv_new = stack(axis = var_1380_axis_0, values = (var_1076, nkv))[name = tensor("op_1380")]; + tensor var_1382_axis_0 = const()[name = tensor("op_1382_axis_0"), val = tensor(0)]; + tensor dec_scale_new = stack(axis = var_1382_axis_0, values = (new_scale_9, new_scale))[name = tensor("op_1382")]; } -> (probs, enc_kv_new, enc_scale_new, enc_conv_cache_new, cnn_window_new, dec_kv_new, dec_scale_new); } \ No newline at end of file diff --git a/optimized/dih2/100ms/ls_eend_dih2_100ms.mlpackage/Data/com.apple.CoreML/model.mlmodel b/optimized/dih2/100ms/ls_eend_dih2_100ms.mlpackage/Data/com.apple.CoreML/model.mlmodel index 2b188342f30379c189c2ef1ec6d955b9033ba14d..b188b3ecab52bd0e3ef6ff0a638d3cdfe256dc40 100644 --- a/optimized/dih2/100ms/ls_eend_dih2_100ms.mlpackage/Data/com.apple.CoreML/model.mlmodel +++ b/optimized/dih2/100ms/ls_eend_dih2_100ms.mlpackage/Data/com.apple.CoreML/model.mlmodel @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:f0e1bf629771d4cad61a66e7932eb77dae2f0231edb3f0070409798b2d0c643c -size 171364 +oid sha256:9e9f5785b2d4260e790dcbe4aa4afb5c33fb5377866338c24ddf9e39bf75a1c0 +size 175282 diff --git a/optimized/dih2/100ms/ls_eend_dih2_100ms.mlpackage/Manifest.json b/optimized/dih2/100ms/ls_eend_dih2_100ms.mlpackage/Manifest.json index 84f6f3de60aa59bdc9ffe7ddbf987d76c5f22beb..16f3cfcaecb66465ec8f40bc6e1e4c580806a317 100644 --- a/optimized/dih2/100ms/ls_eend_dih2_100ms.mlpackage/Manifest.json +++ b/optimized/dih2/100ms/ls_eend_dih2_100ms.mlpackage/Manifest.json @@ -1,18 +1,18 @@ { "fileFormatVersion": "1.0.0", "itemInfoEntries": { - "377FDFC9-ABF8-4C96-8917-8CBDA39D46CD": { + "34D13A76-D1CC-4F3E-BACD-855110612CCB": { "author": "com.apple.CoreML", "description": "CoreML Model Specification", "name": "model.mlmodel", "path": "com.apple.CoreML/model.mlmodel" }, - "EEC7467E-63A3-4C31-9B54-74606F035FAA": { + "59FE4953-2547-41D5-BACE-CC39E25C6CE9": { "author": "com.apple.CoreML", "description": "CoreML Model Weights", "name": "weights", "path": "com.apple.CoreML/weights" } }, - "rootModelIdentifier": "377FDFC9-ABF8-4C96-8917-8CBDA39D46CD" + "rootModelIdentifier": "34D13A76-D1CC-4F3E-BACD-855110612CCB" } diff --git a/optimized/dih2/200ms/ls_eend_dih2_200ms.mlmodelc/analytics/coremldata.bin b/optimized/dih2/200ms/ls_eend_dih2_200ms.mlmodelc/analytics/coremldata.bin index 397635a9a0e1cd59d8cf6642abba267486b68f69..da9d72c722d78161a3ac653368f6509b3416670b 100644 --- a/optimized/dih2/200ms/ls_eend_dih2_200ms.mlmodelc/analytics/coremldata.bin +++ b/optimized/dih2/200ms/ls_eend_dih2_200ms.mlmodelc/analytics/coremldata.bin @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:f1b0bab18a5f1a234fd3b8f86af03830354b0ff7d715753b069b582fd0dac6de +oid sha256:4a16507ded5530171695a9e2509707ad428a6321635d34dd07ec69fd35b1d8d5 size 243 diff --git a/optimized/dih2/200ms/ls_eend_dih2_200ms.mlmodelc/coremldata.bin b/optimized/dih2/200ms/ls_eend_dih2_200ms.mlmodelc/coremldata.bin index e804d1c5d44837686767f82aef6a0244dc12e26c..781371005b7381affae84182486c13a36628874b 100644 --- a/optimized/dih2/200ms/ls_eend_dih2_200ms.mlmodelc/coremldata.bin +++ b/optimized/dih2/200ms/ls_eend_dih2_200ms.mlmodelc/coremldata.bin @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:76832edb1250c28bb6886d9fd633957adb70b61d6525ec6e5ee1131c50305350 -size 1308 +oid sha256:a160a2ed2d67341409c7837e2856513edf10369b35df3aaea2ee36ecb094050a +size 1411 diff --git a/optimized/dih2/200ms/ls_eend_dih2_200ms.mlmodelc/metadata.json b/optimized/dih2/200ms/ls_eend_dih2_200ms.mlmodelc/metadata.json index fcec287e71dc3d3cfdc7a74070099ad04b2b2ae4..e6cf77d52791285f08f6a72989700edc635598d0 100644 --- a/optimized/dih2/200ms/ls_eend_dih2_200ms.mlmodelc/metadata.json +++ b/optimized/dih2/200ms/ls_eend_dih2_200ms.mlmodelc/metadata.json @@ -1,7 +1,7 @@ [ { "metadataOutputVersion" : "3.0", - "shortDescription" : "LS-EEND DIHARD II streaming diarizer (pipeline, T=2, max_speakers=10)", + "shortDescription" : "LS-EEND DIHARD II streaming diarizer (pipeline, T=2, max_speakers=10, layout=raw_mel, cu=all)", "outputSchema" : [ { "hasShapeFlexibility" : "0", @@ -81,7 +81,7 @@ "specificationVersion" : 8, "mlProgramOperationTypeHistogram" : { "Ios16.reduceL2Norm" : 2, - "Ios17.reshape" : 68, + "Ios17.reshape" : 69, "Ios16.softmax" : 2, "Ios17.matmul" : 29, "Ios17.transpose" : 57, @@ -89,7 +89,7 @@ "Ios17.expandDims" : 3, "Ios17.add" : 46, "Ios16.sigmoid" : 5, - "Ios17.sliceByIndex" : 48, + "Ios17.sliceByIndex" : 50, "Tile" : 1, "Ios16.reduceSum" : 1, "Ios17.squeeze" : 2, @@ -101,7 +101,7 @@ "Ios16.silu" : 18, "Ios17.realDiv" : 20, "Ios17.linear" : 56, - "Stack" : 5, + "Stack" : 6, "Ios17.concat" : 14, "Ios16.relu" : 2, "Ios16.cumsum" : 1, @@ -128,9 +128,9 @@ "hasShapeFlexibility" : "0", "isOptional" : "0", "dataType" : "Float32", - "formattedType" : "MultiArray (Float32 1 × 2 × 345)", + "formattedType" : "MultiArray (Float32 1 × 25 × 23)", "shortDescription" : "", - "shape" : "[1, 2, 345]", + "shape" : "[1, 25, 23]", "name" : "features", "type" : "MultiArray" }, @@ -206,8 +206,8 @@ } ], "userDefinedMetadata" : { - "com.github.apple.coremltools.conversion_date" : "2026-04-16", - "config" : "{\"model_name\": \"dih2\", \"model_label\": \"DIHARD II\", \"variant\": \"pipeline\", \"chunk_size\": 2, \"step_duration_ms\": 200, \"frame_duration_ms\": 100, \"frame_duration_seconds\": 0.1, \"max_speakers\": 10, \"max_nspks\": 12, \"n_units\": 256, \"n_heads\": 4, \"enc_n_layers\": 4, \"dec_n_layers\": 2, \"conv_kernel_size\": 16, \"conv_delay\": 9, \"sample_rate\": 8000, \"win_length\": 200, \"hop_length\": 80, \"n_mels\": 23, \"context_size\": 7, \"subsampling\": 10, \"feat_type\": \"logmel23_cummn\", \"pure_roll\": true}", + "com.github.apple.coremltools.conversion_date" : "2026-04-18", + "config" : "{\"model_name\": \"dih2\", \"model_label\": \"DIHARD II\", \"variant\": \"pipeline\", \"chunk_size\": 2, \"step_duration_ms\": 200, \"frame_duration_ms\": 100, \"frame_duration_seconds\": 0.1, \"max_speakers\": 10, \"max_nspks\": 12, \"n_units\": 256, \"n_heads\": 4, \"enc_n_layers\": 4, \"dec_n_layers\": 2, \"conv_kernel_size\": 16, \"conv_delay\": 9, \"sample_rate\": 8000, \"win_length\": 200, \"hop_length\": 80, \"n_mels\": 23, \"context_size\": 7, \"subsampling\": 10, \"feat_type\": \"logmel23_cummn\", \"pure_roll\": true, \"input_layout\": \"raw_mel\", \"compute_units_export\": \"all\", \"raw_mel_length\": 25}", "com.github.apple.coremltools.source" : "torch==2.6.0", "com.github.apple.coremltools.version" : "9.0", "com.github.apple.coremltools.source_dialect" : "TorchScript" diff --git a/optimized/dih2/200ms/ls_eend_dih2_200ms.mlmodelc/model.mil b/optimized/dih2/200ms/ls_eend_dih2_200ms.mlmodelc/model.mil index 5651eaf71d9adaeffd8f5bfe34de4f9950314977..2ccb23e6a99a5f049025ab2b67fb75bd7b7827a2 100644 --- a/optimized/dih2/200ms/ls_eend_dih2_200ms.mlmodelc/model.mil +++ b/optimized/dih2/200ms/ls_eend_dih2_200ms.mlmodelc/model.mil @@ -1,234 +1,248 @@ program(1.0) [buildInfo = dict, tensor>({{"coremlc-component-MIL", "3510.2.1"}, {"coremlc-version", "3500.32.1"}, {"coremltools-component-torch", "2.6.0"}, {"coremltools-source-dialect", "TorchScript"}, {"coremltools-version", "9.0"}})] { - func main(tensor cnn_window, tensor dec_kv, tensor dec_scale, tensor enc_conv_cache, tensor enc_kv, tensor enc_scale, tensor features, tensor valid_mask) { - tensor encoder_cnn_bias = const()[name = tensor("encoder_cnn_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(64)))]; - tensor encoder_cnn_weight = const()[name = tensor("encoder_cnn_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1152)))]; - tensor encoder__causal_mask = const()[name = tensor("encoder__causal_mask"), val = tensor([[0x1p+0, 0x0p+0], [0x1p+0, 0x1p+0]])]; - tensor encoder__t_index = const()[name = tensor("encoder__t_index"), val = tensor([0x1p+0, 0x1p+1])]; - tensor encoder_input_projection_linear_bias = const()[name = tensor("encoder_input_projection_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4981952)))]; - tensor encoder_input_projection_linear_weight = const()[name = tensor("encoder_input_projection_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4983040)))]; - tensor encoder_pre_layer_norm_bias = const()[name = tensor("encoder_pre_layer_norm_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(5336384)))]; - tensor encoder_pre_layer_norm_weight = const()[name = tensor("encoder_pre_layer_norm_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(5337472)))]; - tensor encoder_ffn1_0_module_sequential_0_bias = const()[name = tensor("encoder_ffn1_0_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(5338560)))]; - tensor encoder_ffn1_0_module_sequential_0_weight = const()[name = tensor("encoder_ffn1_0_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(5339648)))]; - tensor encoder_ffn1_0_module_sequential_1_linear_bias = const()[name = tensor("encoder_ffn1_0_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(5340736)))]; - tensor encoder_ffn1_0_module_sequential_1_linear_weight = const()[name = tensor("encoder_ffn1_0_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(5344896)))]; - tensor encoder_ffn1_0_module_sequential_4_linear_bias = const()[name = tensor("encoder_ffn1_0_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(6393536)))]; - tensor encoder_ffn1_0_module_sequential_4_linear_weight = const()[name = tensor("encoder_ffn1_0_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(6394624)))]; - tensor encoder_ret_lns_0_bias = const()[name = tensor("encoder_ret_lns_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7443264)))]; - tensor encoder_ret_lns_0_weight = const()[name = tensor("encoder_ret_lns_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7444352)))]; - tensor encoder_q_proj_0_bias = const()[name = tensor("encoder_q_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7445440)))]; - tensor encoder_q_proj_0_weight = const()[name = tensor("encoder_q_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7446528)))]; - tensor encoder_k_proj_0_bias = const()[name = tensor("encoder_k_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7708736)))]; - tensor encoder_k_proj_0_weight = const()[name = tensor("encoder_k_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7709824)))]; - tensor encoder_v_proj_0_bias = const()[name = tensor("encoder_v_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7972032)))]; - tensor encoder_v_proj_0_weight = const()[name = tensor("encoder_v_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7973120)))]; - tensor encoder_g_proj_0_bias = const()[name = tensor("encoder_g_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8235328)))]; - tensor encoder_g_proj_0_weight = const()[name = tensor("encoder_g_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8236416)))]; - tensor encoder_out_proj_0_bias = const()[name = tensor("encoder_out_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8498624)))]; - tensor encoder_out_proj_0_weight = const()[name = tensor("encoder_out_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8499712)))]; - tensor encoder_conv_module_0_sequential_0_bias = const()[name = tensor("encoder_conv_module_0_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8761920)))]; - tensor encoder_conv_module_0_sequential_0_weight = const()[name = tensor("encoder_conv_module_0_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8763008)))]; - tensor encoder_conv_module_0_sequential_2_conv_bias = const()[name = tensor("encoder_conv_module_0_sequential_2_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8764096)))]; - tensor encoder_conv_module_0_sequential_2_conv_weight = const()[name = tensor("encoder_conv_module_0_sequential_2_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8766208)))]; - tensor encoder_conv_module_0_sequential_4_conv_weight = const()[name = tensor("encoder_conv_module_0_sequential_4_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9290560)))]; - tensor encoder_conv_module_0_sequential_5_running_var = const()[name = tensor("encoder_conv_module_0_sequential_5_running_var"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9307008)))]; - tensor encoder_conv_module_0_sequential_5_running_mean = const()[name = tensor("encoder_conv_module_0_sequential_5_running_mean"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9308096)))]; - tensor encoder_conv_module_0_sequential_5_bias = const()[name = tensor("encoder_conv_module_0_sequential_5_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9309184)))]; - tensor encoder_conv_module_0_sequential_5_weight = const()[name = tensor("encoder_conv_module_0_sequential_5_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9310272)))]; - tensor encoder_conv_module_0_sequential_7_conv_bias = const()[name = tensor("encoder_conv_module_0_sequential_7_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9311360)))]; - tensor encoder_conv_module_0_sequential_7_conv_weight = const()[name = tensor("encoder_conv_module_0_sequential_7_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9312448)))]; - tensor encoder_ffn2_0_module_sequential_0_bias = const()[name = tensor("encoder_ffn2_0_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9574656)))]; - tensor encoder_ffn2_0_module_sequential_0_weight = const()[name = tensor("encoder_ffn2_0_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9575744)))]; - tensor encoder_ffn2_0_module_sequential_1_linear_bias = const()[name = tensor("encoder_ffn2_0_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9576832)))]; - tensor encoder_ffn2_0_module_sequential_1_linear_weight = const()[name = tensor("encoder_ffn2_0_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9580992)))]; - tensor encoder_ffn2_0_module_sequential_4_linear_bias = const()[name = tensor("encoder_ffn2_0_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(10629632)))]; - tensor encoder_ffn2_0_module_sequential_4_linear_weight = const()[name = tensor("encoder_ffn2_0_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(10630720)))]; - tensor encoder_layer_norm_0_bias = const()[name = tensor("encoder_layer_norm_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11679360)))]; - tensor encoder_layer_norm_0_weight = const()[name = tensor("encoder_layer_norm_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11680448)))]; - tensor encoder_ffn1_1_module_sequential_0_bias = const()[name = tensor("encoder_ffn1_1_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11681536)))]; - tensor encoder_ffn1_1_module_sequential_0_weight = const()[name = tensor("encoder_ffn1_1_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11682624)))]; - tensor encoder_ffn1_1_module_sequential_1_linear_bias = const()[name = tensor("encoder_ffn1_1_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11683712)))]; - tensor encoder_ffn1_1_module_sequential_1_linear_weight = const()[name = tensor("encoder_ffn1_1_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11687872)))]; - tensor encoder_ffn1_1_module_sequential_4_linear_bias = const()[name = tensor("encoder_ffn1_1_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(12736512)))]; - tensor encoder_ffn1_1_module_sequential_4_linear_weight = const()[name = tensor("encoder_ffn1_1_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(12737600)))]; - tensor encoder_ret_lns_1_bias = const()[name = tensor("encoder_ret_lns_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(13786240)))]; - tensor encoder_ret_lns_1_weight = const()[name = tensor("encoder_ret_lns_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(13787328)))]; - tensor encoder_q_proj_1_bias = const()[name = tensor("encoder_q_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(13788416)))]; - tensor encoder_q_proj_1_weight = const()[name = tensor("encoder_q_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(13789504)))]; - tensor encoder_k_proj_1_bias = const()[name = tensor("encoder_k_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14051712)))]; - tensor encoder_k_proj_1_weight = const()[name = tensor("encoder_k_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14052800)))]; - tensor encoder_v_proj_1_bias = const()[name = tensor("encoder_v_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14315008)))]; - tensor encoder_v_proj_1_weight = const()[name = tensor("encoder_v_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14316096)))]; - tensor encoder_g_proj_1_bias = const()[name = tensor("encoder_g_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14578304)))]; - tensor encoder_g_proj_1_weight = const()[name = tensor("encoder_g_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14579392)))]; - tensor encoder_out_proj_1_bias = const()[name = tensor("encoder_out_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14841600)))]; - tensor encoder_out_proj_1_weight = const()[name = tensor("encoder_out_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14842688)))]; - tensor encoder_conv_module_1_sequential_0_bias = const()[name = tensor("encoder_conv_module_1_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15104896)))]; - tensor encoder_conv_module_1_sequential_0_weight = const()[name = tensor("encoder_conv_module_1_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15105984)))]; - tensor encoder_conv_module_1_sequential_2_conv_bias = const()[name = tensor("encoder_conv_module_1_sequential_2_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15107072)))]; - tensor encoder_conv_module_1_sequential_2_conv_weight = const()[name = tensor("encoder_conv_module_1_sequential_2_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15109184)))]; - tensor encoder_conv_module_1_sequential_4_conv_weight = const()[name = tensor("encoder_conv_module_1_sequential_4_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15633536)))]; - tensor encoder_conv_module_1_sequential_5_running_var = const()[name = tensor("encoder_conv_module_1_sequential_5_running_var"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15649984)))]; - tensor encoder_conv_module_1_sequential_5_running_mean = const()[name = tensor("encoder_conv_module_1_sequential_5_running_mean"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15651072)))]; - tensor encoder_conv_module_1_sequential_5_bias = const()[name = tensor("encoder_conv_module_1_sequential_5_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15652160)))]; - tensor encoder_conv_module_1_sequential_5_weight = const()[name = tensor("encoder_conv_module_1_sequential_5_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15653248)))]; - tensor encoder_conv_module_1_sequential_7_conv_bias = const()[name = tensor("encoder_conv_module_1_sequential_7_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15654336)))]; - tensor encoder_conv_module_1_sequential_7_conv_weight = const()[name = tensor("encoder_conv_module_1_sequential_7_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15655424)))]; - tensor encoder_ffn2_1_module_sequential_0_bias = const()[name = tensor("encoder_ffn2_1_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15917632)))]; - tensor encoder_ffn2_1_module_sequential_0_weight = const()[name = tensor("encoder_ffn2_1_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15918720)))]; - tensor encoder_ffn2_1_module_sequential_1_linear_bias = const()[name = tensor("encoder_ffn2_1_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15919808)))]; - tensor encoder_ffn2_1_module_sequential_1_linear_weight = const()[name = tensor("encoder_ffn2_1_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15923968)))]; - tensor encoder_ffn2_1_module_sequential_4_linear_bias = const()[name = tensor("encoder_ffn2_1_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(16972608)))]; - tensor encoder_ffn2_1_module_sequential_4_linear_weight = const()[name = tensor("encoder_ffn2_1_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(16973696)))]; - tensor encoder_layer_norm_1_bias = const()[name = tensor("encoder_layer_norm_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18022336)))]; - tensor encoder_layer_norm_1_weight = const()[name = tensor("encoder_layer_norm_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18023424)))]; - tensor encoder_ffn1_2_module_sequential_0_bias = const()[name = tensor("encoder_ffn1_2_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18024512)))]; - tensor encoder_ffn1_2_module_sequential_0_weight = const()[name = tensor("encoder_ffn1_2_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18025600)))]; - tensor encoder_ffn1_2_module_sequential_1_linear_bias = const()[name = tensor("encoder_ffn1_2_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18026688)))]; - tensor encoder_ffn1_2_module_sequential_1_linear_weight = const()[name = tensor("encoder_ffn1_2_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18030848)))]; - tensor encoder_ffn1_2_module_sequential_4_linear_bias = const()[name = tensor("encoder_ffn1_2_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(19079488)))]; - tensor encoder_ffn1_2_module_sequential_4_linear_weight = const()[name = tensor("encoder_ffn1_2_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(19080576)))]; - tensor encoder_ret_lns_2_bias = const()[name = tensor("encoder_ret_lns_2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20129216)))]; - tensor encoder_ret_lns_2_weight = const()[name = tensor("encoder_ret_lns_2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20130304)))]; - tensor encoder_q_proj_2_bias = const()[name = tensor("encoder_q_proj_2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20131392)))]; - tensor encoder_q_proj_2_weight = const()[name = tensor("encoder_q_proj_2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20132480)))]; - tensor encoder_k_proj_2_bias = const()[name = tensor("encoder_k_proj_2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20394688)))]; - tensor encoder_k_proj_2_weight = const()[name = tensor("encoder_k_proj_2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20395776)))]; - tensor encoder_v_proj_2_bias = const()[name = tensor("encoder_v_proj_2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20657984)))]; - tensor encoder_v_proj_2_weight = const()[name = tensor("encoder_v_proj_2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20659072)))]; - tensor encoder_g_proj_2_bias = const()[name = tensor("encoder_g_proj_2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20921280)))]; - tensor encoder_g_proj_2_weight = const()[name = tensor("encoder_g_proj_2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20922368)))]; - tensor encoder_out_proj_2_bias = const()[name = tensor("encoder_out_proj_2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21184576)))]; - tensor encoder_out_proj_2_weight = const()[name = tensor("encoder_out_proj_2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21185664)))]; - tensor encoder_conv_module_2_sequential_0_bias = const()[name = tensor("encoder_conv_module_2_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21447872)))]; - tensor encoder_conv_module_2_sequential_0_weight = const()[name = tensor("encoder_conv_module_2_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21448960)))]; - tensor encoder_conv_module_2_sequential_2_conv_bias = const()[name = tensor("encoder_conv_module_2_sequential_2_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21450048)))]; - tensor encoder_conv_module_2_sequential_2_conv_weight = const()[name = tensor("encoder_conv_module_2_sequential_2_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21452160)))]; - tensor encoder_conv_module_2_sequential_4_conv_weight = const()[name = tensor("encoder_conv_module_2_sequential_4_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21976512)))]; - tensor encoder_conv_module_2_sequential_5_running_var = const()[name = tensor("encoder_conv_module_2_sequential_5_running_var"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21992960)))]; - tensor encoder_conv_module_2_sequential_5_running_mean = const()[name = tensor("encoder_conv_module_2_sequential_5_running_mean"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21994048)))]; - tensor encoder_conv_module_2_sequential_5_bias = const()[name = tensor("encoder_conv_module_2_sequential_5_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21995136)))]; - tensor encoder_conv_module_2_sequential_5_weight = const()[name = tensor("encoder_conv_module_2_sequential_5_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21996224)))]; - tensor encoder_conv_module_2_sequential_7_conv_bias = const()[name = tensor("encoder_conv_module_2_sequential_7_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21997312)))]; - tensor encoder_conv_module_2_sequential_7_conv_weight = const()[name = tensor("encoder_conv_module_2_sequential_7_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21998400)))]; - tensor encoder_ffn2_2_module_sequential_0_bias = const()[name = tensor("encoder_ffn2_2_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(22260608)))]; - tensor encoder_ffn2_2_module_sequential_0_weight = const()[name = tensor("encoder_ffn2_2_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(22261696)))]; - tensor encoder_ffn2_2_module_sequential_1_linear_bias = const()[name = tensor("encoder_ffn2_2_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(22262784)))]; - tensor encoder_ffn2_2_module_sequential_1_linear_weight = const()[name = tensor("encoder_ffn2_2_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(22266944)))]; - tensor encoder_ffn2_2_module_sequential_4_linear_bias = const()[name = tensor("encoder_ffn2_2_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(23315584)))]; - tensor encoder_ffn2_2_module_sequential_4_linear_weight = const()[name = tensor("encoder_ffn2_2_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(23316672)))]; - tensor encoder_layer_norm_2_bias = const()[name = tensor("encoder_layer_norm_2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(24365312)))]; - tensor encoder_layer_norm_2_weight = const()[name = tensor("encoder_layer_norm_2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(24366400)))]; - tensor encoder_ffn1_3_module_sequential_0_bias = const()[name = tensor("encoder_ffn1_3_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(24367488)))]; - tensor encoder_ffn1_3_module_sequential_0_weight = const()[name = tensor("encoder_ffn1_3_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(24368576)))]; - tensor encoder_ffn1_3_module_sequential_1_linear_bias = const()[name = tensor("encoder_ffn1_3_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(24369664)))]; - tensor encoder_ffn1_3_module_sequential_1_linear_weight = const()[name = tensor("encoder_ffn1_3_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(24373824)))]; - tensor encoder_ffn1_3_module_sequential_4_linear_bias = const()[name = tensor("encoder_ffn1_3_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(25422464)))]; - tensor encoder_ffn1_3_module_sequential_4_linear_weight = const()[name = tensor("encoder_ffn1_3_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(25423552)))]; - tensor encoder_ret_lns_3_bias = const()[name = tensor("encoder_ret_lns_3_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(26472192)))]; - tensor encoder_ret_lns_3_weight = const()[name = tensor("encoder_ret_lns_3_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(26473280)))]; - tensor encoder_q_proj_3_bias = const()[name = tensor("encoder_q_proj_3_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(26474368)))]; - tensor encoder_q_proj_3_weight = const()[name = tensor("encoder_q_proj_3_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(26475456)))]; - tensor encoder_k_proj_3_bias = const()[name = tensor("encoder_k_proj_3_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(26737664)))]; - tensor encoder_k_proj_3_weight = const()[name = tensor("encoder_k_proj_3_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(26738752)))]; - tensor encoder_v_proj_3_bias = const()[name = tensor("encoder_v_proj_3_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27000960)))]; - tensor encoder_v_proj_3_weight = const()[name = tensor("encoder_v_proj_3_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27002048)))]; - tensor encoder_g_proj_3_bias = const()[name = tensor("encoder_g_proj_3_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27264256)))]; - tensor encoder_g_proj_3_weight = const()[name = tensor("encoder_g_proj_3_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27265344)))]; - tensor encoder_out_proj_3_bias = const()[name = tensor("encoder_out_proj_3_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27527552)))]; - tensor encoder_out_proj_3_weight = const()[name = tensor("encoder_out_proj_3_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27528640)))]; - tensor encoder_conv_module_3_sequential_0_bias = const()[name = tensor("encoder_conv_module_3_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27790848)))]; - tensor encoder_conv_module_3_sequential_0_weight = const()[name = tensor("encoder_conv_module_3_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27791936)))]; - tensor encoder_conv_module_3_sequential_2_conv_bias = const()[name = tensor("encoder_conv_module_3_sequential_2_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27793024)))]; - tensor encoder_conv_module_3_sequential_2_conv_weight = const()[name = tensor("encoder_conv_module_3_sequential_2_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27795136)))]; - tensor encoder_conv_module_3_sequential_4_conv_weight = const()[name = tensor("encoder_conv_module_3_sequential_4_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28319488)))]; - tensor encoder_conv_module_3_sequential_5_running_var = const()[name = tensor("encoder_conv_module_3_sequential_5_running_var"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28335936)))]; - tensor encoder_conv_module_3_sequential_5_running_mean = const()[name = tensor("encoder_conv_module_3_sequential_5_running_mean"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28337024)))]; - tensor encoder_conv_module_3_sequential_5_bias = const()[name = tensor("encoder_conv_module_3_sequential_5_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28338112)))]; - tensor encoder_conv_module_3_sequential_5_weight = const()[name = tensor("encoder_conv_module_3_sequential_5_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28339200)))]; - tensor encoder_conv_module_3_sequential_7_conv_bias = const()[name = tensor("encoder_conv_module_3_sequential_7_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28340288)))]; - tensor encoder_conv_module_3_sequential_7_conv_weight = const()[name = tensor("encoder_conv_module_3_sequential_7_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28341376)))]; - tensor encoder_ffn2_3_module_sequential_0_bias = const()[name = tensor("encoder_ffn2_3_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28603584)))]; - tensor encoder_ffn2_3_module_sequential_0_weight = const()[name = tensor("encoder_ffn2_3_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28604672)))]; - tensor encoder_ffn2_3_module_sequential_1_linear_bias = const()[name = tensor("encoder_ffn2_3_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28605760)))]; - tensor encoder_ffn2_3_module_sequential_1_linear_weight = const()[name = tensor("encoder_ffn2_3_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28609920)))]; - tensor encoder_ffn2_3_module_sequential_4_linear_bias = const()[name = tensor("encoder_ffn2_3_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(29658560)))]; - tensor encoder_ffn2_3_module_sequential_4_linear_weight = const()[name = tensor("encoder_ffn2_3_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(29659648)))]; - tensor encoder_layer_norm_3_bias = const()[name = tensor("encoder_layer_norm_3_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(30708288)))]; - tensor encoder_layer_norm_3_weight = const()[name = tensor("encoder_layer_norm_3_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(30709376)))]; - tensor decoder__causal_mask = const()[name = tensor("decoder__causal_mask"), val = tensor([[0x1p+0, 0x0p+0], [0x1p+0, 0x1p+0]])]; - tensor decoder_convert_bias = const()[name = tensor("decoder_convert_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(30710464)))]; - tensor decoder_convert_weight = const()[name = tensor("decoder_convert_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(30711552)))]; - tensor decoder_q_proj_0_bias = const()[name = tensor("decoder_q_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(31235904)))]; - tensor decoder_q_proj_0_weight = const()[name = tensor("decoder_q_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(31236992)))]; - tensor decoder_k_proj_0_bias = const()[name = tensor("decoder_k_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(31499200)))]; - tensor decoder_k_proj_0_weight = const()[name = tensor("decoder_k_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(31500288)))]; - tensor decoder_v_proj_0_bias = const()[name = tensor("decoder_v_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(31762496)))]; - tensor decoder_v_proj_0_weight = const()[name = tensor("decoder_v_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(31763584)))]; - tensor decoder_g_proj_0_bias = const()[name = tensor("decoder_g_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32025792)))]; - tensor decoder_g_proj_0_weight = const()[name = tensor("decoder_g_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32026880)))]; - tensor decoder_out_proj_0_bias = const()[name = tensor("decoder_out_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32289088)))]; - tensor decoder_out_proj_0_weight = const()[name = tensor("decoder_out_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32290176)))]; - tensor decoder_norm11_0_bias = const()[name = tensor("decoder_norm11_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32552384)))]; - tensor decoder_norm11_0_weight = const()[name = tensor("decoder_norm11_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32553472)))]; - tensor decoder_self_attn2_0_out_proj_bias = const()[name = tensor("decoder_self_attn2_0_out_proj_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32554560)))]; - tensor decoder_self_attn2_0_out_proj_weight = const()[name = tensor("decoder_self_attn2_0_out_proj_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32555648)))]; - tensor decoder_self_attn2_0_in_proj_bias = const()[name = tensor("decoder_self_attn2_0_in_proj_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32817856)))]; - tensor decoder_self_attn2_0_in_proj_weight = const()[name = tensor("decoder_self_attn2_0_in_proj_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32820992)))]; - tensor decoder_norm21_0_bias = const()[name = tensor("decoder_norm21_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(33607488)))]; - tensor decoder_norm21_0_weight = const()[name = tensor("decoder_norm21_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(33608576)))]; - tensor decoder_linear1_0_bias = const()[name = tensor("decoder_linear1_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(33609664)))]; - tensor decoder_linear1_0_weight = const()[name = tensor("decoder_linear1_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(33617920)))]; - tensor decoder_linear2_0_bias = const()[name = tensor("decoder_linear2_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(35715136)))]; - tensor decoder_linear2_0_weight = const()[name = tensor("decoder_linear2_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(35716224)))]; - tensor decoder_norm22_0_bias = const()[name = tensor("decoder_norm22_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(37813440)))]; - tensor decoder_norm22_0_weight = const()[name = tensor("decoder_norm22_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(37814528)))]; - tensor decoder_q_proj_1_bias = const()[name = tensor("decoder_q_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(37815616)))]; - tensor decoder_q_proj_1_weight = const()[name = tensor("decoder_q_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(37816704)))]; - tensor decoder_k_proj_1_bias = const()[name = tensor("decoder_k_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38078912)))]; - tensor decoder_k_proj_1_weight = const()[name = tensor("decoder_k_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38080000)))]; - tensor decoder_v_proj_1_bias = const()[name = tensor("decoder_v_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38342208)))]; - tensor decoder_v_proj_1_weight = const()[name = tensor("decoder_v_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38343296)))]; - tensor decoder_g_proj_1_bias = const()[name = tensor("decoder_g_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38605504)))]; - tensor decoder_g_proj_1_weight = const()[name = tensor("decoder_g_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38606592)))]; - tensor decoder_out_proj_1_bias = const()[name = tensor("decoder_out_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38868800)))]; - tensor decoder_out_proj_1_weight = const()[name = tensor("decoder_out_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38869888)))]; - tensor decoder_norm11_1_bias = const()[name = tensor("decoder_norm11_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(39132096)))]; - tensor decoder_norm11_1_weight = const()[name = tensor("decoder_norm11_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(39133184)))]; - tensor decoder_self_attn2_1_out_proj_bias = const()[name = tensor("decoder_self_attn2_1_out_proj_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(39134272)))]; - tensor decoder_self_attn2_1_out_proj_weight = const()[name = tensor("decoder_self_attn2_1_out_proj_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(39135360)))]; - tensor decoder_self_attn2_1_in_proj_bias = const()[name = tensor("decoder_self_attn2_1_in_proj_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(39397568)))]; - tensor decoder_self_attn2_1_in_proj_weight = const()[name = tensor("decoder_self_attn2_1_in_proj_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(39400704)))]; - tensor decoder_norm21_1_bias = const()[name = tensor("decoder_norm21_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(40187200)))]; - tensor decoder_norm21_1_weight = const()[name = tensor("decoder_norm21_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(40188288)))]; - tensor decoder_linear1_1_bias = const()[name = tensor("decoder_linear1_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(40189376)))]; - tensor decoder_linear1_1_weight = const()[name = tensor("decoder_linear1_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(40197632)))]; - tensor decoder_linear2_1_bias = const()[name = tensor("decoder_linear2_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(42294848)))]; - tensor decoder_linear2_1_weight = const()[name = tensor("decoder_linear2_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(42295936)))]; - tensor decoder_norm22_1_bias = const()[name = tensor("decoder_norm22_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(44393152)))]; - tensor decoder_norm22_1_weight = const()[name = tensor("decoder_norm22_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(44394240)))]; - tensor var_11 = const()[name = tensor("op_11"), val = tensor(0x1.5798eep-27)]; - tensor var_17 = const()[name = tensor("op_17"), val = tensor(0x1.0c6f7ap-20)]; - tensor var_20 = const()[name = tensor("op_20"), val = tensor(2)]; - tensor var_23 = const()[name = tensor("op_23"), val = tensor(0)]; - tensor var_26 = const()[name = tensor("op_26"), val = tensor(1)]; - tensor var_28 = const()[name = tensor("op_28"), val = tensor(0x1.4f8b58p-17)]; - tensor var_29 = const()[name = tensor("op_29"), val = tensor(true)]; - tensor input_1 = linear(bias = encoder_input_projection_linear_bias, weight = encoder_input_projection_linear_weight, x = features)[name = tensor("linear_0")]; - tensor input_3_axes_0 = const()[name = tensor("input_3_axes_0"), val = tensor([-1])]; - tensor input_3 = layer_norm(axes = input_3_axes_0, beta = encoder_pre_layer_norm_bias, epsilon = var_28, gamma = encoder_pre_layer_norm_weight, x = input_1)[name = tensor("input_3")]; + func main(tensor cnn_window, tensor dec_kv, tensor dec_scale, tensor enc_conv_cache, tensor enc_kv, tensor enc_scale, tensor features, tensor valid_mask) { + tensor inner_encoder_cnn_bias = const()[name = tensor("inner_encoder_cnn_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(64)))]; + tensor inner_encoder_cnn_weight = const()[name = tensor("inner_encoder_cnn_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1152)))]; + tensor inner_encoder__causal_mask = const()[name = tensor("inner_encoder__causal_mask"), val = tensor([[0x1p+0, 0x0p+0], [0x1p+0, 0x1p+0]])]; + tensor inner_encoder__t_index = const()[name = tensor("inner_encoder__t_index"), val = tensor([0x1p+0, 0x1p+1])]; + tensor inner_encoder_input_projection_linear_bias = const()[name = tensor("inner_encoder_input_projection_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4981952)))]; + tensor inner_encoder_input_projection_linear_weight = const()[name = tensor("inner_encoder_input_projection_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4983040)))]; + tensor inner_encoder_pre_layer_norm_bias = const()[name = tensor("inner_encoder_pre_layer_norm_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(5336384)))]; + tensor inner_encoder_pre_layer_norm_weight = const()[name = tensor("inner_encoder_pre_layer_norm_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(5337472)))]; + tensor inner_encoder_ffn1_0_module_sequential_0_bias = const()[name = tensor("inner_encoder_ffn1_0_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(5338560)))]; + tensor inner_encoder_ffn1_0_module_sequential_0_weight = const()[name = tensor("inner_encoder_ffn1_0_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(5339648)))]; + tensor inner_encoder_ffn1_0_module_sequential_1_linear_bias = const()[name = tensor("inner_encoder_ffn1_0_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(5340736)))]; + tensor inner_encoder_ffn1_0_module_sequential_1_linear_weight = const()[name = tensor("inner_encoder_ffn1_0_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(5344896)))]; + tensor inner_encoder_ffn1_0_module_sequential_4_linear_bias = const()[name = tensor("inner_encoder_ffn1_0_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(6393536)))]; + tensor inner_encoder_ffn1_0_module_sequential_4_linear_weight = const()[name = tensor("inner_encoder_ffn1_0_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(6394624)))]; + tensor inner_encoder_ret_lns_0_bias = const()[name = tensor("inner_encoder_ret_lns_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7443264)))]; + tensor inner_encoder_ret_lns_0_weight = const()[name = tensor("inner_encoder_ret_lns_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7444352)))]; + tensor inner_encoder_q_proj_0_bias = const()[name = tensor("inner_encoder_q_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7445440)))]; + tensor inner_encoder_q_proj_0_weight = const()[name = tensor("inner_encoder_q_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7446528)))]; + tensor inner_encoder_k_proj_0_bias = const()[name = tensor("inner_encoder_k_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7708736)))]; + tensor inner_encoder_k_proj_0_weight = const()[name = tensor("inner_encoder_k_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7709824)))]; + tensor inner_encoder_v_proj_0_bias = const()[name = tensor("inner_encoder_v_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7972032)))]; + tensor inner_encoder_v_proj_0_weight = const()[name = tensor("inner_encoder_v_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7973120)))]; + tensor inner_encoder_g_proj_0_bias = const()[name = tensor("inner_encoder_g_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8235328)))]; + tensor inner_encoder_g_proj_0_weight = const()[name = tensor("inner_encoder_g_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8236416)))]; + tensor inner_encoder_out_proj_0_bias = const()[name = tensor("inner_encoder_out_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8498624)))]; + tensor inner_encoder_out_proj_0_weight = const()[name = tensor("inner_encoder_out_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8499712)))]; + tensor inner_encoder_conv_module_0_sequential_0_bias = const()[name = tensor("inner_encoder_conv_module_0_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8761920)))]; + tensor inner_encoder_conv_module_0_sequential_0_weight = const()[name = tensor("inner_encoder_conv_module_0_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8763008)))]; + tensor inner_encoder_conv_module_0_sequential_2_conv_bias = const()[name = tensor("inner_encoder_conv_module_0_sequential_2_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8764096)))]; + tensor inner_encoder_conv_module_0_sequential_2_conv_weight = const()[name = tensor("inner_encoder_conv_module_0_sequential_2_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8766208)))]; + tensor inner_encoder_conv_module_0_sequential_4_conv_weight = const()[name = tensor("inner_encoder_conv_module_0_sequential_4_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9290560)))]; + tensor inner_encoder_conv_module_0_sequential_5_running_var = const()[name = tensor("inner_encoder_conv_module_0_sequential_5_running_var"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9307008)))]; + tensor inner_encoder_conv_module_0_sequential_5_running_mean = const()[name = tensor("inner_encoder_conv_module_0_sequential_5_running_mean"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9308096)))]; + tensor inner_encoder_conv_module_0_sequential_5_bias = const()[name = tensor("inner_encoder_conv_module_0_sequential_5_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9309184)))]; + tensor inner_encoder_conv_module_0_sequential_5_weight = const()[name = tensor("inner_encoder_conv_module_0_sequential_5_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9310272)))]; + tensor inner_encoder_conv_module_0_sequential_7_conv_bias = const()[name = tensor("inner_encoder_conv_module_0_sequential_7_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9311360)))]; + tensor inner_encoder_conv_module_0_sequential_7_conv_weight = const()[name = tensor("inner_encoder_conv_module_0_sequential_7_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9312448)))]; + tensor inner_encoder_ffn2_0_module_sequential_0_bias = const()[name = tensor("inner_encoder_ffn2_0_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9574656)))]; + tensor inner_encoder_ffn2_0_module_sequential_0_weight = const()[name = tensor("inner_encoder_ffn2_0_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9575744)))]; + tensor inner_encoder_ffn2_0_module_sequential_1_linear_bias = const()[name = tensor("inner_encoder_ffn2_0_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9576832)))]; + tensor inner_encoder_ffn2_0_module_sequential_1_linear_weight = const()[name = tensor("inner_encoder_ffn2_0_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9580992)))]; + tensor inner_encoder_ffn2_0_module_sequential_4_linear_bias = const()[name = tensor("inner_encoder_ffn2_0_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(10629632)))]; + tensor inner_encoder_ffn2_0_module_sequential_4_linear_weight = const()[name = tensor("inner_encoder_ffn2_0_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(10630720)))]; + tensor inner_encoder_layer_norm_0_bias = const()[name = tensor("inner_encoder_layer_norm_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11679360)))]; + tensor inner_encoder_layer_norm_0_weight = const()[name = tensor("inner_encoder_layer_norm_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11680448)))]; + tensor inner_encoder_ffn1_1_module_sequential_0_bias = const()[name = tensor("inner_encoder_ffn1_1_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11681536)))]; + tensor inner_encoder_ffn1_1_module_sequential_0_weight = const()[name = tensor("inner_encoder_ffn1_1_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11682624)))]; + tensor inner_encoder_ffn1_1_module_sequential_1_linear_bias = const()[name = tensor("inner_encoder_ffn1_1_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11683712)))]; + tensor inner_encoder_ffn1_1_module_sequential_1_linear_weight = const()[name = tensor("inner_encoder_ffn1_1_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11687872)))]; + tensor inner_encoder_ffn1_1_module_sequential_4_linear_bias = const()[name = tensor("inner_encoder_ffn1_1_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(12736512)))]; + tensor inner_encoder_ffn1_1_module_sequential_4_linear_weight = const()[name = tensor("inner_encoder_ffn1_1_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(12737600)))]; + tensor inner_encoder_ret_lns_1_bias = const()[name = tensor("inner_encoder_ret_lns_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(13786240)))]; + tensor inner_encoder_ret_lns_1_weight = const()[name = tensor("inner_encoder_ret_lns_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(13787328)))]; + tensor inner_encoder_q_proj_1_bias = const()[name = tensor("inner_encoder_q_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(13788416)))]; + tensor inner_encoder_q_proj_1_weight = const()[name = tensor("inner_encoder_q_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(13789504)))]; + tensor inner_encoder_k_proj_1_bias = const()[name = tensor("inner_encoder_k_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14051712)))]; + tensor inner_encoder_k_proj_1_weight = const()[name = tensor("inner_encoder_k_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14052800)))]; + tensor inner_encoder_v_proj_1_bias = const()[name = tensor("inner_encoder_v_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14315008)))]; + tensor inner_encoder_v_proj_1_weight = const()[name = tensor("inner_encoder_v_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14316096)))]; + tensor inner_encoder_g_proj_1_bias = const()[name = tensor("inner_encoder_g_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14578304)))]; + tensor inner_encoder_g_proj_1_weight = const()[name = tensor("inner_encoder_g_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14579392)))]; + tensor inner_encoder_out_proj_1_bias = const()[name = tensor("inner_encoder_out_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14841600)))]; + tensor inner_encoder_out_proj_1_weight = const()[name = tensor("inner_encoder_out_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14842688)))]; + tensor inner_encoder_conv_module_1_sequential_0_bias = const()[name = tensor("inner_encoder_conv_module_1_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15104896)))]; + tensor inner_encoder_conv_module_1_sequential_0_weight = const()[name = tensor("inner_encoder_conv_module_1_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15105984)))]; + tensor inner_encoder_conv_module_1_sequential_2_conv_bias = const()[name = tensor("inner_encoder_conv_module_1_sequential_2_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15107072)))]; + tensor inner_encoder_conv_module_1_sequential_2_conv_weight = const()[name = tensor("inner_encoder_conv_module_1_sequential_2_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15109184)))]; + tensor inner_encoder_conv_module_1_sequential_4_conv_weight = const()[name = tensor("inner_encoder_conv_module_1_sequential_4_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15633536)))]; + tensor inner_encoder_conv_module_1_sequential_5_running_var = const()[name = tensor("inner_encoder_conv_module_1_sequential_5_running_var"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15649984)))]; + tensor inner_encoder_conv_module_1_sequential_5_running_mean = const()[name = tensor("inner_encoder_conv_module_1_sequential_5_running_mean"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15651072)))]; + tensor inner_encoder_conv_module_1_sequential_5_bias = const()[name = tensor("inner_encoder_conv_module_1_sequential_5_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15652160)))]; + tensor inner_encoder_conv_module_1_sequential_5_weight = const()[name = tensor("inner_encoder_conv_module_1_sequential_5_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15653248)))]; + tensor inner_encoder_conv_module_1_sequential_7_conv_bias = const()[name = tensor("inner_encoder_conv_module_1_sequential_7_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15654336)))]; + tensor inner_encoder_conv_module_1_sequential_7_conv_weight = const()[name = tensor("inner_encoder_conv_module_1_sequential_7_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15655424)))]; + tensor inner_encoder_ffn2_1_module_sequential_0_bias = const()[name = tensor("inner_encoder_ffn2_1_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15917632)))]; + tensor inner_encoder_ffn2_1_module_sequential_0_weight = const()[name = tensor("inner_encoder_ffn2_1_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15918720)))]; + tensor inner_encoder_ffn2_1_module_sequential_1_linear_bias = const()[name = tensor("inner_encoder_ffn2_1_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15919808)))]; + tensor inner_encoder_ffn2_1_module_sequential_1_linear_weight = const()[name = tensor("inner_encoder_ffn2_1_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15923968)))]; + tensor inner_encoder_ffn2_1_module_sequential_4_linear_bias = const()[name = tensor("inner_encoder_ffn2_1_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(16972608)))]; + tensor inner_encoder_ffn2_1_module_sequential_4_linear_weight = const()[name = tensor("inner_encoder_ffn2_1_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(16973696)))]; + tensor inner_encoder_layer_norm_1_bias = const()[name = tensor("inner_encoder_layer_norm_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18022336)))]; + tensor inner_encoder_layer_norm_1_weight = const()[name = tensor("inner_encoder_layer_norm_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18023424)))]; + tensor inner_encoder_ffn1_2_module_sequential_0_bias = const()[name = tensor("inner_encoder_ffn1_2_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18024512)))]; + tensor inner_encoder_ffn1_2_module_sequential_0_weight = const()[name = tensor("inner_encoder_ffn1_2_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18025600)))]; + tensor inner_encoder_ffn1_2_module_sequential_1_linear_bias = const()[name = tensor("inner_encoder_ffn1_2_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18026688)))]; + tensor inner_encoder_ffn1_2_module_sequential_1_linear_weight = const()[name = tensor("inner_encoder_ffn1_2_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18030848)))]; + tensor inner_encoder_ffn1_2_module_sequential_4_linear_bias = const()[name = tensor("inner_encoder_ffn1_2_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(19079488)))]; + tensor inner_encoder_ffn1_2_module_sequential_4_linear_weight = const()[name = tensor("inner_encoder_ffn1_2_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(19080576)))]; + tensor inner_encoder_ret_lns_2_bias = const()[name = tensor("inner_encoder_ret_lns_2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20129216)))]; + tensor inner_encoder_ret_lns_2_weight = const()[name = tensor("inner_encoder_ret_lns_2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20130304)))]; + tensor inner_encoder_q_proj_2_bias = const()[name = tensor("inner_encoder_q_proj_2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20131392)))]; + tensor inner_encoder_q_proj_2_weight = const()[name = tensor("inner_encoder_q_proj_2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20132480)))]; + tensor inner_encoder_k_proj_2_bias = const()[name = tensor("inner_encoder_k_proj_2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20394688)))]; + tensor inner_encoder_k_proj_2_weight = const()[name = tensor("inner_encoder_k_proj_2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20395776)))]; + tensor inner_encoder_v_proj_2_bias = const()[name = tensor("inner_encoder_v_proj_2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20657984)))]; + tensor inner_encoder_v_proj_2_weight = const()[name = tensor("inner_encoder_v_proj_2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20659072)))]; + tensor inner_encoder_g_proj_2_bias = const()[name = tensor("inner_encoder_g_proj_2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20921280)))]; + tensor inner_encoder_g_proj_2_weight = const()[name = tensor("inner_encoder_g_proj_2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20922368)))]; + tensor inner_encoder_out_proj_2_bias = const()[name = tensor("inner_encoder_out_proj_2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21184576)))]; + tensor inner_encoder_out_proj_2_weight = const()[name = tensor("inner_encoder_out_proj_2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21185664)))]; + tensor inner_encoder_conv_module_2_sequential_0_bias = const()[name = tensor("inner_encoder_conv_module_2_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21447872)))]; + tensor inner_encoder_conv_module_2_sequential_0_weight = const()[name = tensor("inner_encoder_conv_module_2_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21448960)))]; + tensor inner_encoder_conv_module_2_sequential_2_conv_bias = const()[name = tensor("inner_encoder_conv_module_2_sequential_2_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21450048)))]; + tensor inner_encoder_conv_module_2_sequential_2_conv_weight = const()[name = tensor("inner_encoder_conv_module_2_sequential_2_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21452160)))]; + tensor inner_encoder_conv_module_2_sequential_4_conv_weight = const()[name = tensor("inner_encoder_conv_module_2_sequential_4_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21976512)))]; + tensor inner_encoder_conv_module_2_sequential_5_running_var = const()[name = tensor("inner_encoder_conv_module_2_sequential_5_running_var"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21992960)))]; + tensor inner_encoder_conv_module_2_sequential_5_running_mean = const()[name = tensor("inner_encoder_conv_module_2_sequential_5_running_mean"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21994048)))]; + tensor inner_encoder_conv_module_2_sequential_5_bias = const()[name = tensor("inner_encoder_conv_module_2_sequential_5_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21995136)))]; + tensor inner_encoder_conv_module_2_sequential_5_weight = const()[name = tensor("inner_encoder_conv_module_2_sequential_5_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21996224)))]; + tensor inner_encoder_conv_module_2_sequential_7_conv_bias = const()[name = tensor("inner_encoder_conv_module_2_sequential_7_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21997312)))]; + tensor inner_encoder_conv_module_2_sequential_7_conv_weight = const()[name = tensor("inner_encoder_conv_module_2_sequential_7_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21998400)))]; + tensor inner_encoder_ffn2_2_module_sequential_0_bias = const()[name = tensor("inner_encoder_ffn2_2_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(22260608)))]; + tensor inner_encoder_ffn2_2_module_sequential_0_weight = const()[name = tensor("inner_encoder_ffn2_2_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(22261696)))]; + tensor inner_encoder_ffn2_2_module_sequential_1_linear_bias = const()[name = tensor("inner_encoder_ffn2_2_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(22262784)))]; + tensor inner_encoder_ffn2_2_module_sequential_1_linear_weight = const()[name = tensor("inner_encoder_ffn2_2_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(22266944)))]; + tensor inner_encoder_ffn2_2_module_sequential_4_linear_bias = const()[name = tensor("inner_encoder_ffn2_2_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(23315584)))]; + tensor inner_encoder_ffn2_2_module_sequential_4_linear_weight = const()[name = tensor("inner_encoder_ffn2_2_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(23316672)))]; + tensor inner_encoder_layer_norm_2_bias = const()[name = tensor("inner_encoder_layer_norm_2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(24365312)))]; + tensor inner_encoder_layer_norm_2_weight = const()[name = tensor("inner_encoder_layer_norm_2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(24366400)))]; + tensor inner_encoder_ffn1_3_module_sequential_0_bias = const()[name = tensor("inner_encoder_ffn1_3_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(24367488)))]; + tensor inner_encoder_ffn1_3_module_sequential_0_weight = const()[name = tensor("inner_encoder_ffn1_3_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(24368576)))]; + tensor inner_encoder_ffn1_3_module_sequential_1_linear_bias = const()[name = tensor("inner_encoder_ffn1_3_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(24369664)))]; + tensor inner_encoder_ffn1_3_module_sequential_1_linear_weight = const()[name = tensor("inner_encoder_ffn1_3_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(24373824)))]; + tensor inner_encoder_ffn1_3_module_sequential_4_linear_bias = const()[name = tensor("inner_encoder_ffn1_3_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(25422464)))]; + tensor inner_encoder_ffn1_3_module_sequential_4_linear_weight = const()[name = tensor("inner_encoder_ffn1_3_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(25423552)))]; + tensor inner_encoder_ret_lns_3_bias = const()[name = tensor("inner_encoder_ret_lns_3_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(26472192)))]; + tensor inner_encoder_ret_lns_3_weight = const()[name = tensor("inner_encoder_ret_lns_3_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(26473280)))]; + tensor inner_encoder_q_proj_3_bias = const()[name = tensor("inner_encoder_q_proj_3_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(26474368)))]; + tensor inner_encoder_q_proj_3_weight = const()[name = tensor("inner_encoder_q_proj_3_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(26475456)))]; + tensor inner_encoder_k_proj_3_bias = const()[name = tensor("inner_encoder_k_proj_3_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(26737664)))]; + tensor inner_encoder_k_proj_3_weight = const()[name = tensor("inner_encoder_k_proj_3_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(26738752)))]; + tensor inner_encoder_v_proj_3_bias = const()[name = tensor("inner_encoder_v_proj_3_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27000960)))]; + tensor inner_encoder_v_proj_3_weight = const()[name = tensor("inner_encoder_v_proj_3_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27002048)))]; + tensor inner_encoder_g_proj_3_bias = const()[name = tensor("inner_encoder_g_proj_3_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27264256)))]; + tensor inner_encoder_g_proj_3_weight = const()[name = tensor("inner_encoder_g_proj_3_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27265344)))]; + tensor inner_encoder_out_proj_3_bias = const()[name = tensor("inner_encoder_out_proj_3_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27527552)))]; + tensor inner_encoder_out_proj_3_weight = const()[name = tensor("inner_encoder_out_proj_3_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27528640)))]; + tensor inner_encoder_conv_module_3_sequential_0_bias = const()[name = tensor("inner_encoder_conv_module_3_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27790848)))]; + tensor inner_encoder_conv_module_3_sequential_0_weight = const()[name = tensor("inner_encoder_conv_module_3_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27791936)))]; + tensor inner_encoder_conv_module_3_sequential_2_conv_bias = const()[name = tensor("inner_encoder_conv_module_3_sequential_2_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27793024)))]; + tensor inner_encoder_conv_module_3_sequential_2_conv_weight = const()[name = tensor("inner_encoder_conv_module_3_sequential_2_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27795136)))]; + tensor inner_encoder_conv_module_3_sequential_4_conv_weight = const()[name = tensor("inner_encoder_conv_module_3_sequential_4_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28319488)))]; + tensor inner_encoder_conv_module_3_sequential_5_running_var = const()[name = tensor("inner_encoder_conv_module_3_sequential_5_running_var"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28335936)))]; + tensor inner_encoder_conv_module_3_sequential_5_running_mean = const()[name = tensor("inner_encoder_conv_module_3_sequential_5_running_mean"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28337024)))]; + tensor inner_encoder_conv_module_3_sequential_5_bias = const()[name = tensor("inner_encoder_conv_module_3_sequential_5_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28338112)))]; + tensor inner_encoder_conv_module_3_sequential_5_weight = const()[name = tensor("inner_encoder_conv_module_3_sequential_5_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28339200)))]; + tensor inner_encoder_conv_module_3_sequential_7_conv_bias = const()[name = tensor("inner_encoder_conv_module_3_sequential_7_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28340288)))]; + tensor inner_encoder_conv_module_3_sequential_7_conv_weight = const()[name = tensor("inner_encoder_conv_module_3_sequential_7_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28341376)))]; + tensor inner_encoder_ffn2_3_module_sequential_0_bias = const()[name = tensor("inner_encoder_ffn2_3_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28603584)))]; + tensor inner_encoder_ffn2_3_module_sequential_0_weight = const()[name = tensor("inner_encoder_ffn2_3_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28604672)))]; + tensor inner_encoder_ffn2_3_module_sequential_1_linear_bias = const()[name = tensor("inner_encoder_ffn2_3_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28605760)))]; + tensor inner_encoder_ffn2_3_module_sequential_1_linear_weight = const()[name = tensor("inner_encoder_ffn2_3_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28609920)))]; + tensor inner_encoder_ffn2_3_module_sequential_4_linear_bias = const()[name = tensor("inner_encoder_ffn2_3_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(29658560)))]; + tensor inner_encoder_ffn2_3_module_sequential_4_linear_weight = const()[name = tensor("inner_encoder_ffn2_3_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(29659648)))]; + tensor inner_encoder_layer_norm_3_bias = const()[name = tensor("inner_encoder_layer_norm_3_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(30708288)))]; + tensor inner_encoder_layer_norm_3_weight = const()[name = tensor("inner_encoder_layer_norm_3_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(30709376)))]; + tensor inner_decoder__causal_mask = const()[name = tensor("inner_decoder__causal_mask"), val = tensor([[0x1p+0, 0x0p+0], [0x1p+0, 0x1p+0]])]; + tensor inner_decoder_convert_bias = const()[name = tensor("inner_decoder_convert_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(30710464)))]; + tensor inner_decoder_convert_weight = const()[name = tensor("inner_decoder_convert_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(30711552)))]; + tensor inner_decoder_q_proj_0_bias = const()[name = tensor("inner_decoder_q_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(31235904)))]; + tensor inner_decoder_q_proj_0_weight = const()[name = tensor("inner_decoder_q_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(31236992)))]; + tensor inner_decoder_k_proj_0_bias = const()[name = tensor("inner_decoder_k_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(31499200)))]; + tensor inner_decoder_k_proj_0_weight = const()[name = tensor("inner_decoder_k_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(31500288)))]; + tensor inner_decoder_v_proj_0_bias = const()[name = tensor("inner_decoder_v_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(31762496)))]; + tensor inner_decoder_v_proj_0_weight = const()[name = tensor("inner_decoder_v_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(31763584)))]; + tensor inner_decoder_g_proj_0_bias = const()[name = tensor("inner_decoder_g_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32025792)))]; + tensor inner_decoder_g_proj_0_weight = const()[name = tensor("inner_decoder_g_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32026880)))]; + tensor inner_decoder_out_proj_0_bias = const()[name = tensor("inner_decoder_out_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32289088)))]; + tensor inner_decoder_out_proj_0_weight = const()[name = tensor("inner_decoder_out_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32290176)))]; + tensor inner_decoder_norm11_0_bias = const()[name = tensor("inner_decoder_norm11_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32552384)))]; + tensor inner_decoder_norm11_0_weight = const()[name = tensor("inner_decoder_norm11_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32553472)))]; + tensor inner_decoder_self_attn2_0_out_proj_bias = const()[name = tensor("inner_decoder_self_attn2_0_out_proj_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32554560)))]; + tensor inner_decoder_self_attn2_0_out_proj_weight = const()[name = tensor("inner_decoder_self_attn2_0_out_proj_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32555648)))]; + tensor inner_decoder_self_attn2_0_in_proj_bias = const()[name = tensor("inner_decoder_self_attn2_0_in_proj_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32817856)))]; + tensor inner_decoder_self_attn2_0_in_proj_weight = const()[name = tensor("inner_decoder_self_attn2_0_in_proj_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32820992)))]; + tensor inner_decoder_norm21_0_bias = const()[name = tensor("inner_decoder_norm21_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(33607488)))]; + tensor inner_decoder_norm21_0_weight = const()[name = tensor("inner_decoder_norm21_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(33608576)))]; + tensor inner_decoder_linear1_0_bias = const()[name = tensor("inner_decoder_linear1_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(33609664)))]; + tensor inner_decoder_linear1_0_weight = const()[name = tensor("inner_decoder_linear1_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(33617920)))]; + tensor inner_decoder_linear2_0_bias = const()[name = tensor("inner_decoder_linear2_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(35715136)))]; + tensor inner_decoder_linear2_0_weight = const()[name = tensor("inner_decoder_linear2_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(35716224)))]; + tensor inner_decoder_norm22_0_bias = const()[name = tensor("inner_decoder_norm22_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(37813440)))]; + tensor inner_decoder_norm22_0_weight = const()[name = tensor("inner_decoder_norm22_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(37814528)))]; + tensor inner_decoder_q_proj_1_bias = const()[name = tensor("inner_decoder_q_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(37815616)))]; + tensor inner_decoder_q_proj_1_weight = const()[name = tensor("inner_decoder_q_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(37816704)))]; + tensor inner_decoder_k_proj_1_bias = const()[name = tensor("inner_decoder_k_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38078912)))]; + tensor inner_decoder_k_proj_1_weight = const()[name = tensor("inner_decoder_k_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38080000)))]; + tensor inner_decoder_v_proj_1_bias = const()[name = tensor("inner_decoder_v_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38342208)))]; + tensor inner_decoder_v_proj_1_weight = const()[name = tensor("inner_decoder_v_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38343296)))]; + tensor inner_decoder_g_proj_1_bias = const()[name = tensor("inner_decoder_g_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38605504)))]; + tensor inner_decoder_g_proj_1_weight = const()[name = tensor("inner_decoder_g_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38606592)))]; + tensor inner_decoder_out_proj_1_bias = const()[name = tensor("inner_decoder_out_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38868800)))]; + tensor inner_decoder_out_proj_1_weight = const()[name = tensor("inner_decoder_out_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38869888)))]; + tensor inner_decoder_norm11_1_bias = const()[name = tensor("inner_decoder_norm11_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(39132096)))]; + tensor inner_decoder_norm11_1_weight = const()[name = tensor("inner_decoder_norm11_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(39133184)))]; + tensor inner_decoder_self_attn2_1_out_proj_bias = const()[name = tensor("inner_decoder_self_attn2_1_out_proj_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(39134272)))]; + tensor inner_decoder_self_attn2_1_out_proj_weight = const()[name = tensor("inner_decoder_self_attn2_1_out_proj_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(39135360)))]; + tensor inner_decoder_self_attn2_1_in_proj_bias = const()[name = tensor("inner_decoder_self_attn2_1_in_proj_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(39397568)))]; + tensor inner_decoder_self_attn2_1_in_proj_weight = const()[name = tensor("inner_decoder_self_attn2_1_in_proj_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(39400704)))]; + tensor inner_decoder_norm21_1_bias = const()[name = tensor("inner_decoder_norm21_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(40187200)))]; + tensor inner_decoder_norm21_1_weight = const()[name = tensor("inner_decoder_norm21_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(40188288)))]; + tensor inner_decoder_linear1_1_bias = const()[name = tensor("inner_decoder_linear1_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(40189376)))]; + tensor inner_decoder_linear1_1_weight = const()[name = tensor("inner_decoder_linear1_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(40197632)))]; + tensor inner_decoder_linear2_1_bias = const()[name = tensor("inner_decoder_linear2_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(42294848)))]; + tensor inner_decoder_linear2_1_weight = const()[name = tensor("inner_decoder_linear2_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(42295936)))]; + tensor inner_decoder_norm22_1_bias = const()[name = tensor("inner_decoder_norm22_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(44393152)))]; + tensor inner_decoder_norm22_1_weight = const()[name = tensor("inner_decoder_norm22_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(44394240)))]; + tensor var_19_begin_0 = const()[name = tensor("op_19_begin_0"), val = tensor([0, 0, 0])]; + tensor var_19_end_0 = const()[name = tensor("op_19_end_0"), val = tensor([1, 15, 23])]; + tensor var_19_end_mask_0 = const()[name = tensor("op_19_end_mask_0"), val = tensor([true, false, true])]; + tensor var_19 = slice_by_index(begin = var_19_begin_0, end = var_19_end_0, end_mask = var_19_end_mask_0, x = features)[name = tensor("op_19")]; + tensor var_29_begin_0 = const()[name = tensor("op_29_begin_0"), val = tensor([0, 10, 0])]; + tensor var_29_end_0 = const()[name = tensor("op_29_end_0"), val = tensor([1, 1, 23])]; + tensor var_29_end_mask_0 = const()[name = tensor("op_29_end_mask_0"), val = tensor([true, true, true])]; + tensor var_29 = slice_by_index(begin = var_29_begin_0, end = var_29_end_0, end_mask = var_29_end_mask_0, x = features)[name = tensor("op_29")]; + tensor stacked_axis_0 = const()[name = tensor("stacked_axis_0"), val = tensor(1)]; + tensor stacked = stack(axis = stacked_axis_0, values = (var_19, var_29))[name = tensor("stacked")]; + tensor var_36 = const()[name = tensor("op_36"), val = tensor([1, 2, 345])]; + tensor input_1 = reshape(shape = var_36, x = stacked)[name = tensor("input_1")]; + tensor var_39 = const()[name = tensor("op_39"), val = tensor(0x1p+0)]; + tensor var_45 = const()[name = tensor("op_45"), val = tensor(true)]; + tensor var_46 = const()[name = tensor("op_46"), val = tensor(0x1.4f8b58p-17)]; + tensor var_49 = const()[name = tensor("op_49"), val = tensor(0)]; + tensor var_51 = const()[name = tensor("op_51"), val = tensor(2)]; + tensor var_52 = const()[name = tensor("op_52"), val = tensor(-1)]; + tensor var_54 = const()[name = tensor("op_54"), val = tensor(0x1.0c6f7ap-20)]; + tensor var_59 = const()[name = tensor("op_59"), val = tensor(0x1.5798eep-27)]; + tensor var_62 = const()[name = tensor("op_62"), val = tensor(1)]; + tensor input_3 = linear(bias = inner_encoder_input_projection_linear_bias, weight = inner_encoder_input_projection_linear_weight, x = input_1)[name = tensor("linear_0")]; tensor input_5_axes_0 = const()[name = tensor("input_5_axes_0"), val = tensor([-1])]; - tensor input_5 = layer_norm(axes = input_5_axes_0, beta = encoder_ffn1_0_module_sequential_0_bias, epsilon = var_28, gamma = encoder_ffn1_0_module_sequential_0_weight, x = input_3)[name = tensor("input_5")]; - tensor inputs_1 = linear(bias = encoder_ffn1_0_module_sequential_1_linear_bias, weight = encoder_ffn1_0_module_sequential_1_linear_weight, x = input_5)[name = tensor("linear_1")]; - tensor input_7 = silu(x = inputs_1)[name = tensor("input_7")]; - tensor input_11 = linear(bias = encoder_ffn1_0_module_sequential_4_linear_bias, weight = encoder_ffn1_0_module_sequential_4_linear_weight, x = input_7)[name = tensor("linear_2")]; - tensor var_147 = const()[name = tensor("op_147"), val = tensor(0x1p-1)]; - tensor var_148 = mul(x = input_11, y = var_147)[name = tensor("op_148")]; - tensor input_13 = add(x = var_148, y = input_3)[name = tensor("input_13")]; + tensor input_5 = layer_norm(axes = input_5_axes_0, beta = inner_encoder_pre_layer_norm_bias, epsilon = var_46, gamma = inner_encoder_pre_layer_norm_weight, x = input_3)[name = tensor("input_5")]; + tensor input_7_axes_0 = const()[name = tensor("input_7_axes_0"), val = tensor([-1])]; + tensor input_7 = layer_norm(axes = input_7_axes_0, beta = inner_encoder_ffn1_0_module_sequential_0_bias, epsilon = var_46, gamma = inner_encoder_ffn1_0_module_sequential_0_weight, x = input_5)[name = tensor("input_7")]; + tensor inputs_1 = linear(bias = inner_encoder_ffn1_0_module_sequential_1_linear_bias, weight = inner_encoder_ffn1_0_module_sequential_1_linear_weight, x = input_7)[name = tensor("linear_1")]; + tensor input_9 = silu(x = inputs_1)[name = tensor("input_9")]; + tensor input_13 = linear(bias = inner_encoder_ffn1_0_module_sequential_4_linear_bias, weight = inner_encoder_ffn1_0_module_sequential_4_linear_weight, x = input_9)[name = tensor("linear_2")]; + tensor var_183 = const()[name = tensor("op_183"), val = tensor(0x1p-1)]; + tensor var_184 = mul(x = input_13, y = var_183)[name = tensor("op_184")]; + tensor input_15 = add(x = var_184, y = input_5)[name = tensor("input_15")]; tensor x_1_axes_0 = const()[name = tensor("x_1_axes_0"), val = tensor([-1])]; - tensor x_1 = layer_norm(axes = x_1_axes_0, beta = encoder_ret_lns_0_bias, epsilon = var_28, gamma = encoder_ret_lns_0_weight, x = input_13)[name = tensor("x_1")]; + tensor x_1 = layer_norm(axes = x_1_axes_0, beta = inner_encoder_ret_lns_0_bias, epsilon = var_46, gamma = inner_encoder_ret_lns_0_weight, x = input_15)[name = tensor("x_1")]; tensor prev_kv_1_begin_0 = const()[name = tensor("prev_kv_1_begin_0"), val = tensor([0, 0, 0, 0, 0])]; tensor prev_kv_1_end_0 = const()[name = tensor("prev_kv_1_end_0"), val = tensor([1, 1, 4, 64, 64])]; tensor prev_kv_1_end_mask_0 = const()[name = tensor("prev_kv_1_end_mask_0"), val = tensor([false, true, true, true, true])]; @@ -239,153 +253,153 @@ program(1.0) tensor prev_scale_1_end_mask_0 = const()[name = tensor("prev_scale_1_end_mask_0"), val = tensor([false, true])]; tensor prev_scale_1_squeeze_mask_0 = const()[name = tensor("prev_scale_1_squeeze_mask_0"), val = tensor([true, false])]; tensor prev_scale_1 = slice_by_index(begin = prev_scale_1_begin_0, end = prev_scale_1_end_0, end_mask = prev_scale_1_end_mask_0, squeeze_mask = prev_scale_1_squeeze_mask_0, x = enc_scale)[name = tensor("prev_scale_1")]; - tensor var_162 = linear(bias = encoder_q_proj_0_bias, weight = encoder_q_proj_0_weight, x = x_1)[name = tensor("linear_3")]; - tensor var_163 = const()[name = tensor("op_163"), val = tensor([1, 2, 4, 64])]; - tensor var_164 = reshape(shape = var_163, x = var_162)[name = tensor("op_164")]; + tensor var_198 = linear(bias = inner_encoder_q_proj_0_bias, weight = inner_encoder_q_proj_0_weight, x = x_1)[name = tensor("linear_3")]; + tensor var_199 = const()[name = tensor("op_199"), val = tensor([1, 2, 4, 64])]; + tensor var_200 = reshape(shape = var_199, x = var_198)[name = tensor("op_200")]; tensor q_1_perm_0 = const()[name = tensor("q_1_perm_0"), val = tensor([0, 2, 1, 3])]; - tensor var_168 = linear(bias = encoder_k_proj_0_bias, weight = encoder_k_proj_0_weight, x = x_1)[name = tensor("linear_4")]; - tensor var_169 = const()[name = tensor("op_169"), val = tensor(0x1p-3)]; - tensor var_170 = mul(x = var_168, y = var_169)[name = tensor("op_170")]; - tensor var_171 = const()[name = tensor("op_171"), val = tensor([1, 2, 4, 64])]; - tensor var_172 = reshape(shape = var_171, x = var_170)[name = tensor("op_172")]; + tensor var_204 = linear(bias = inner_encoder_k_proj_0_bias, weight = inner_encoder_k_proj_0_weight, x = x_1)[name = tensor("linear_4")]; + tensor var_205 = const()[name = tensor("op_205"), val = tensor(0x1p-3)]; + tensor var_206 = mul(x = var_204, y = var_205)[name = tensor("op_206")]; + tensor var_207 = const()[name = tensor("op_207"), val = tensor([1, 2, 4, 64])]; + tensor var_208 = reshape(shape = var_207, x = var_206)[name = tensor("op_208")]; tensor k_1_perm_0 = const()[name = tensor("k_1_perm_0"), val = tensor([0, 2, 1, 3])]; - tensor var_176 = linear(bias = encoder_v_proj_0_bias, weight = encoder_v_proj_0_weight, x = x_1)[name = tensor("linear_5")]; - tensor var_177 = const()[name = tensor("op_177"), val = tensor([1, 2, 4, 64])]; - tensor var_178 = reshape(shape = var_177, x = var_176)[name = tensor("op_178")]; + tensor var_212 = linear(bias = inner_encoder_v_proj_0_bias, weight = inner_encoder_v_proj_0_weight, x = x_1)[name = tensor("linear_5")]; + tensor var_213 = const()[name = tensor("op_213"), val = tensor([1, 2, 4, 64])]; + tensor var_214 = reshape(shape = var_213, x = var_212)[name = tensor("op_214")]; tensor v_1_perm_0 = const()[name = tensor("v_1_perm_0"), val = tensor([0, 2, 1, 3])]; - tensor input_17 = linear(bias = encoder_g_proj_0_bias, weight = encoder_g_proj_0_weight, x = x_1)[name = tensor("linear_6")]; + tensor input_19 = linear(bias = inner_encoder_g_proj_0_bias, weight = inner_encoder_g_proj_0_weight, x = x_1)[name = tensor("linear_6")]; tensor sqrt_s0_1 = sqrt(x = prev_scale_1)[name = tensor("sqrt_s0_1")]; - tensor s_t_1 = add(x = prev_scale_1, y = encoder__t_index)[name = tensor("s_t_1")]; + tensor s_t_1 = add(x = prev_scale_1, y = inner_encoder__t_index)[name = tensor("s_t_1")]; tensor sqrt_s_t_1 = sqrt(x = s_t_1)[name = tensor("sqrt_s_t_1")]; tensor qk_1_transpose_x_1 = const()[name = tensor("qk_1_transpose_x_1"), val = tensor(false)]; tensor qk_1_transpose_y_1 = const()[name = tensor("qk_1_transpose_y_1"), val = tensor(true)]; - tensor k_1 = transpose(perm = k_1_perm_0, x = var_172)[name = tensor("transpose_57")]; - tensor q_1 = transpose(perm = q_1_perm_0, x = var_164)[name = tensor("transpose_58")]; + tensor k_1 = transpose(perm = k_1_perm_0, x = var_208)[name = tensor("transpose_57")]; + tensor q_1 = transpose(perm = q_1_perm_0, x = var_200)[name = tensor("transpose_58")]; tensor qk_1 = matmul(transpose_x = qk_1_transpose_x_1, transpose_y = qk_1_transpose_y_1, x = q_1, y = k_1)[name = tensor("qk_1")]; - tensor var_188 = const()[name = tensor("op_188"), val = tensor([2, 1])]; - tensor var_189 = reshape(shape = var_188, x = sqrt_s_t_1)[name = tensor("op_189")]; - tensor M_1 = real_div(x = encoder__causal_mask, y = var_189)[name = tensor("M_1")]; - tensor var_191 = mul(x = qk_1, y = M_1)[name = tensor("op_191")]; + tensor var_224 = const()[name = tensor("op_224"), val = tensor([2, 1])]; + tensor var_225 = reshape(shape = var_224, x = sqrt_s_t_1)[name = tensor("op_225")]; + tensor M_1 = real_div(x = inner_encoder__causal_mask, y = var_225)[name = tensor("M_1")]; + tensor var_227 = mul(x = qk_1, y = M_1)[name = tensor("op_227")]; tensor inner_1_transpose_x_0 = const()[name = tensor("inner_1_transpose_x_0"), val = tensor(false)]; tensor inner_1_transpose_y_0 = const()[name = tensor("inner_1_transpose_y_0"), val = tensor(false)]; - tensor v_1 = transpose(perm = v_1_perm_0, x = var_178)[name = tensor("transpose_56")]; - tensor inner_1 = matmul(transpose_x = inner_1_transpose_x_0, transpose_y = inner_1_transpose_y_0, x = var_191, y = v_1)[name = tensor("inner_1")]; - tensor var_193_transpose_x_0 = const()[name = tensor("op_193_transpose_x_0"), val = tensor(false)]; - tensor var_193_transpose_y_0 = const()[name = tensor("op_193_transpose_y_0"), val = tensor(false)]; - tensor var_193 = matmul(transpose_x = var_193_transpose_x_0, transpose_y = var_193_transpose_y_0, x = q_1, y = prev_kv_1)[name = tensor("op_193")]; - tensor var_194 = real_div(x = sqrt_s0_1, y = sqrt_s_t_1)[name = tensor("op_194")]; - tensor var_195 = const()[name = tensor("op_195"), val = tensor([1, 1, 2, 1])]; - tensor var_196 = reshape(shape = var_195, x = var_194)[name = tensor("op_196")]; - tensor cross_1 = mul(x = var_193, y = var_196)[name = tensor("cross_1")]; + tensor v_1 = transpose(perm = v_1_perm_0, x = var_214)[name = tensor("transpose_56")]; + tensor inner_1 = matmul(transpose_x = inner_1_transpose_x_0, transpose_y = inner_1_transpose_y_0, x = var_227, y = v_1)[name = tensor("inner_1")]; + tensor var_229_transpose_x_0 = const()[name = tensor("op_229_transpose_x_0"), val = tensor(false)]; + tensor var_229_transpose_y_0 = const()[name = tensor("op_229_transpose_y_0"), val = tensor(false)]; + tensor var_229 = matmul(transpose_x = var_229_transpose_x_0, transpose_y = var_229_transpose_y_0, x = q_1, y = prev_kv_1)[name = tensor("op_229")]; + tensor var_230 = real_div(x = sqrt_s0_1, y = sqrt_s_t_1)[name = tensor("op_230")]; + tensor var_231 = const()[name = tensor("op_231"), val = tensor([1, 1, 2, 1])]; + tensor var_232 = reshape(shape = var_231, x = var_230)[name = tensor("op_232")]; + tensor cross_1 = mul(x = var_229, y = var_232)[name = tensor("cross_1")]; tensor out_1 = add(x = inner_1, y = cross_1)[name = tensor("out_1")]; - tensor var_199 = mul(x = prev_kv_1, y = sqrt_s0_1)[name = tensor("op_199")]; - tensor var_201_transpose_x_1 = const()[name = tensor("op_201_transpose_x_1"), val = tensor(true)]; - tensor var_201_transpose_y_1 = const()[name = tensor("op_201_transpose_y_1"), val = tensor(false)]; - tensor var_201 = matmul(transpose_x = var_201_transpose_x_1, transpose_y = var_201_transpose_y_1, x = k_1, y = v_1)[name = tensor("op_201")]; - tensor new_kv_unnorm_1 = add(x = var_199, y = var_201)[name = tensor("new_kv_unnorm_1")]; - tensor var_203 = const()[name = tensor("op_203"), val = tensor(0x1p+1)]; - tensor new_scale_1 = add(x = prev_scale_1, y = var_203)[name = tensor("new_scale_1")]; - tensor var_205 = sqrt(x = new_scale_1)[name = tensor("op_205")]; - tensor var_206 = real_div(x = new_kv_unnorm_1, y = var_205)[name = tensor("op_206")]; - tensor var_207_perm_0 = const()[name = tensor("op_207_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_235 = mul(x = prev_kv_1, y = sqrt_s0_1)[name = tensor("op_235")]; + tensor var_237_transpose_x_1 = const()[name = tensor("op_237_transpose_x_1"), val = tensor(true)]; + tensor var_237_transpose_y_1 = const()[name = tensor("op_237_transpose_y_1"), val = tensor(false)]; + tensor var_237 = matmul(transpose_x = var_237_transpose_x_1, transpose_y = var_237_transpose_y_1, x = k_1, y = v_1)[name = tensor("op_237")]; + tensor new_kv_unnorm_1 = add(x = var_235, y = var_237)[name = tensor("new_kv_unnorm_1")]; + tensor var_239 = const()[name = tensor("op_239"), val = tensor(0x1p+1)]; + tensor new_scale_1 = add(x = prev_scale_1, y = var_239)[name = tensor("new_scale_1")]; + tensor var_241 = sqrt(x = new_scale_1)[name = tensor("op_241")]; + tensor var_242 = real_div(x = new_kv_unnorm_1, y = var_241)[name = tensor("op_242")]; + tensor var_243_perm_0 = const()[name = tensor("op_243_perm_0"), val = tensor([0, 2, 1, 3])]; tensor out_3_axes_0 = const()[name = tensor("out_3_axes_0"), val = tensor([-1])]; - tensor var_207 = transpose(perm = var_207_perm_0, x = out_1)[name = tensor("transpose_55")]; - tensor out_3 = layer_norm(axes = out_3_axes_0, epsilon = var_17, x = var_207)[name = tensor("out_3")]; - tensor var_211 = const()[name = tensor("op_211"), val = tensor([1, 2, 256])]; - tensor out_5 = reshape(shape = var_211, x = out_3)[name = tensor("out_5")]; - tensor var_213 = silu(x = input_17)[name = tensor("op_213")]; - tensor input_19 = mul(x = var_213, y = out_5)[name = tensor("input_19")]; - tensor ret_out_1 = linear(bias = encoder_out_proj_0_bias, weight = encoder_out_proj_0_weight, x = input_19)[name = tensor("linear_7")]; - tensor x_3 = add(x = input_13, y = ret_out_1)[name = tensor("x_3")]; + tensor var_243 = transpose(perm = var_243_perm_0, x = out_1)[name = tensor("transpose_55")]; + tensor out_3 = layer_norm(axes = out_3_axes_0, epsilon = var_54, x = var_243)[name = tensor("out_3")]; + tensor var_247 = const()[name = tensor("op_247"), val = tensor([1, 2, 256])]; + tensor out_5 = reshape(shape = var_247, x = out_3)[name = tensor("out_5")]; + tensor var_249 = silu(x = input_19)[name = tensor("op_249")]; + tensor input_21 = mul(x = var_249, y = out_5)[name = tensor("input_21")]; + tensor ret_out_1 = linear(bias = inner_encoder_out_proj_0_bias, weight = inner_encoder_out_proj_0_weight, x = input_21)[name = tensor("linear_7")]; + tensor x_3 = add(x = input_15, y = ret_out_1)[name = tensor("x_3")]; tensor window_1_begin_0 = const()[name = tensor("window_1_begin_0"), val = tensor([0, 0, 0, 0])]; tensor window_1_end_0 = const()[name = tensor("window_1_end_0"), val = tensor([1, 1, 16, 256])]; tensor window_1_end_mask_0 = const()[name = tensor("window_1_end_mask_0"), val = tensor([false, true, true, true])]; tensor window_1_squeeze_mask_0 = const()[name = tensor("window_1_squeeze_mask_0"), val = tensor([true, false, false, false])]; tensor window_1 = slice_by_index(begin = window_1_begin_0, end = window_1_end_0, end_mask = window_1_end_mask_0, squeeze_mask = window_1_squeeze_mask_0, x = enc_conv_cache)[name = tensor("window_1")]; - tensor var_221_begin_0 = const()[name = tensor("op_221_begin_0"), val = tensor([0, 0, 0])]; - tensor var_221_end_0 = const()[name = tensor("op_221_end_0"), val = tensor([1, 1, 256])]; - tensor var_221_end_mask_0 = const()[name = tensor("op_221_end_mask_0"), val = tensor([true, false, true])]; - tensor var_221 = slice_by_index(begin = var_221_begin_0, end = var_221_end_0, end_mask = var_221_end_mask_0, x = x_3)[name = tensor("op_221")]; - tensor var_224_begin_0 = const()[name = tensor("op_224_begin_0"), val = tensor([0, 1, 0])]; - tensor var_224_end_0 = const()[name = tensor("op_224_end_0"), val = tensor([1, 16, 256])]; - tensor var_224_end_mask_0 = const()[name = tensor("op_224_end_mask_0"), val = tensor([true, true, true])]; - tensor var_224 = slice_by_index(begin = var_224_begin_0, end = var_224_end_0, end_mask = var_224_end_mask_0, x = window_1)[name = tensor("op_224")]; + tensor var_257_begin_0 = const()[name = tensor("op_257_begin_0"), val = tensor([0, 0, 0])]; + tensor var_257_end_0 = const()[name = tensor("op_257_end_0"), val = tensor([1, 1, 256])]; + tensor var_257_end_mask_0 = const()[name = tensor("op_257_end_mask_0"), val = tensor([true, false, true])]; + tensor var_257 = slice_by_index(begin = var_257_begin_0, end = var_257_end_0, end_mask = var_257_end_mask_0, x = x_3)[name = tensor("op_257")]; + tensor var_260_begin_0 = const()[name = tensor("op_260_begin_0"), val = tensor([0, 1, 0])]; + tensor var_260_end_0 = const()[name = tensor("op_260_end_0"), val = tensor([1, 16, 256])]; + tensor var_260_end_mask_0 = const()[name = tensor("op_260_end_mask_0"), val = tensor([true, true, true])]; + tensor var_260 = slice_by_index(begin = var_260_begin_0, end = var_260_end_0, end_mask = var_260_end_mask_0, x = window_1)[name = tensor("op_260")]; tensor window_3_interleave_0 = const()[name = tensor("window_3_interleave_0"), val = tensor(false)]; - tensor window_3 = concat(axis = var_26, interleave = window_3_interleave_0, values = (var_224, var_221))[name = tensor("window_3")]; - tensor var_229_begin_0 = const()[name = tensor("op_229_begin_0"), val = tensor([0, 1, 0])]; - tensor var_229_end_0 = const()[name = tensor("op_229_end_0"), val = tensor([1, 1, 256])]; - tensor var_229_end_mask_0 = const()[name = tensor("op_229_end_mask_0"), val = tensor([true, true, true])]; - tensor var_229 = slice_by_index(begin = var_229_begin_0, end = var_229_end_0, end_mask = var_229_end_mask_0, x = x_3)[name = tensor("op_229")]; - tensor var_232_begin_0 = const()[name = tensor("op_232_begin_0"), val = tensor([0, 1, 0])]; - tensor var_232_end_0 = const()[name = tensor("op_232_end_0"), val = tensor([1, 16, 256])]; - tensor var_232_end_mask_0 = const()[name = tensor("op_232_end_mask_0"), val = tensor([true, true, true])]; - tensor var_232 = slice_by_index(begin = var_232_begin_0, end = var_232_end_0, end_mask = var_232_end_mask_0, x = window_3)[name = tensor("op_232")]; + tensor window_3 = concat(axis = var_62, interleave = window_3_interleave_0, values = (var_260, var_257))[name = tensor("window_3")]; + tensor var_265_begin_0 = const()[name = tensor("op_265_begin_0"), val = tensor([0, 1, 0])]; + tensor var_265_end_0 = const()[name = tensor("op_265_end_0"), val = tensor([1, 1, 256])]; + tensor var_265_end_mask_0 = const()[name = tensor("op_265_end_mask_0"), val = tensor([true, true, true])]; + tensor var_265 = slice_by_index(begin = var_265_begin_0, end = var_265_end_0, end_mask = var_265_end_mask_0, x = x_3)[name = tensor("op_265")]; + tensor var_268_begin_0 = const()[name = tensor("op_268_begin_0"), val = tensor([0, 1, 0])]; + tensor var_268_end_0 = const()[name = tensor("op_268_end_0"), val = tensor([1, 16, 256])]; + tensor var_268_end_mask_0 = const()[name = tensor("op_268_end_mask_0"), val = tensor([true, true, true])]; + tensor var_268 = slice_by_index(begin = var_268_begin_0, end = var_268_end_0, end_mask = var_268_end_mask_0, x = window_3)[name = tensor("op_268")]; tensor window_5_interleave_0 = const()[name = tensor("window_5_interleave_0"), val = tensor(false)]; - tensor window_5 = concat(axis = var_26, interleave = window_5_interleave_0, values = (var_232, var_229))[name = tensor("window_5")]; - tensor input_21_interleave_0 = const()[name = tensor("input_21_interleave_0"), val = tensor(false)]; - tensor input_21 = concat(axis = var_23, interleave = input_21_interleave_0, values = (window_3, window_5))[name = tensor("input_21")]; + tensor window_5 = concat(axis = var_62, interleave = window_5_interleave_0, values = (var_268, var_265))[name = tensor("window_5")]; + tensor input_23_interleave_0 = const()[name = tensor("input_23_interleave_0"), val = tensor(false)]; + tensor input_23 = concat(axis = var_49, interleave = input_23_interleave_0, values = (window_3, window_5))[name = tensor("input_23")]; tensor x_5_axes_0 = const()[name = tensor("x_5_axes_0"), val = tensor([-1])]; - tensor x_5 = layer_norm(axes = x_5_axes_0, beta = encoder_conv_module_0_sequential_0_bias, epsilon = var_28, gamma = encoder_conv_module_0_sequential_0_weight, x = input_21)[name = tensor("x_5")]; - tensor input_23_perm_0 = const()[name = tensor("input_23_perm_0"), val = tensor([0, 2, 1])]; + tensor x_5 = layer_norm(axes = x_5_axes_0, beta = inner_encoder_conv_module_0_sequential_0_bias, epsilon = var_46, gamma = inner_encoder_conv_module_0_sequential_0_weight, x = input_23)[name = tensor("x_5")]; + tensor input_25_perm_0 = const()[name = tensor("input_25_perm_0"), val = tensor([0, 2, 1])]; tensor inputs_3_pad_type_0 = const()[name = tensor("inputs_3_pad_type_0"), val = tensor("valid")]; tensor inputs_3_strides_0 = const()[name = tensor("inputs_3_strides_0"), val = tensor([1])]; tensor inputs_3_pad_0 = const()[name = tensor("inputs_3_pad_0"), val = tensor([0, 0])]; tensor inputs_3_dilations_0 = const()[name = tensor("inputs_3_dilations_0"), val = tensor([1])]; tensor inputs_3_groups_0 = const()[name = tensor("inputs_3_groups_0"), val = tensor(1)]; - tensor input_23 = transpose(perm = input_23_perm_0, x = x_5)[name = tensor("transpose_54")]; - tensor inputs_3 = conv(bias = encoder_conv_module_0_sequential_2_conv_bias, dilations = inputs_3_dilations_0, groups = inputs_3_groups_0, pad = inputs_3_pad_0, pad_type = inputs_3_pad_type_0, strides = inputs_3_strides_0, weight = encoder_conv_module_0_sequential_2_conv_weight, x = input_23)[name = tensor("inputs_3")]; - tensor var_257_split_sizes_0 = const()[name = tensor("op_257_split_sizes_0"), val = tensor([256, 256])]; - tensor var_257_axis_0 = const()[name = tensor("op_257_axis_0"), val = tensor(1)]; - tensor var_257_0, tensor var_257_1 = split(axis = var_257_axis_0, split_sizes = var_257_split_sizes_0, x = inputs_3)[name = tensor("op_257")]; - tensor var_259 = sigmoid(x = var_257_1)[name = tensor("op_259")]; - tensor inputs_5 = mul(x = var_257_0, y = var_259)[name = tensor("inputs_5")]; + tensor input_25 = transpose(perm = input_25_perm_0, x = x_5)[name = tensor("transpose_54")]; + tensor inputs_3 = conv(bias = inner_encoder_conv_module_0_sequential_2_conv_bias, dilations = inputs_3_dilations_0, groups = inputs_3_groups_0, pad = inputs_3_pad_0, pad_type = inputs_3_pad_type_0, strides = inputs_3_strides_0, weight = inner_encoder_conv_module_0_sequential_2_conv_weight, x = input_25)[name = tensor("inputs_3")]; + tensor var_293_split_sizes_0 = const()[name = tensor("op_293_split_sizes_0"), val = tensor([256, 256])]; + tensor var_293_axis_0 = const()[name = tensor("op_293_axis_0"), val = tensor(1)]; + tensor var_293_0, tensor var_293_1 = split(axis = var_293_axis_0, split_sizes = var_293_split_sizes_0, x = inputs_3)[name = tensor("op_293")]; + tensor var_295 = sigmoid(x = var_293_1)[name = tensor("op_295")]; + tensor inputs_5 = mul(x = var_293_0, y = var_295)[name = tensor("inputs_5")]; tensor outputs_aug_1_pad_type_0 = const()[name = tensor("outputs_aug_1_pad_type_0"), val = tensor("custom")]; tensor outputs_aug_1_pad_0 = const()[name = tensor("outputs_aug_1_pad_0"), val = tensor([15, 15])]; tensor outputs_aug_1_groups_0 = const()[name = tensor("outputs_aug_1_groups_0"), val = tensor(256)]; tensor outputs_aug_1_strides_0 = const()[name = tensor("outputs_aug_1_strides_0"), val = tensor([1])]; tensor outputs_aug_1_dilations_0 = const()[name = tensor("outputs_aug_1_dilations_0"), val = tensor([1])]; - tensor outputs_aug_1 = conv(dilations = outputs_aug_1_dilations_0, groups = outputs_aug_1_groups_0, pad = outputs_aug_1_pad_0, pad_type = outputs_aug_1_pad_type_0, strides = outputs_aug_1_strides_0, weight = encoder_conv_module_0_sequential_4_conv_weight, x = inputs_5)[name = tensor("outputs_aug_1")]; - tensor input_25_begin_0 = const()[name = tensor("input_25_begin_0"), val = tensor([0, 0, 0])]; - tensor input_25_end_0 = const()[name = tensor("input_25_end_0"), val = tensor([2, 256, 16])]; - tensor input_25_end_mask_0 = const()[name = tensor("input_25_end_mask_0"), val = tensor([true, true, false])]; - tensor input_25 = slice_by_index(begin = input_25_begin_0, end = input_25_end_0, end_mask = input_25_end_mask_0, x = outputs_aug_1)[name = tensor("input_25")]; - tensor inputs_7 = batch_norm(beta = encoder_conv_module_0_sequential_5_bias, epsilon = var_28, gamma = encoder_conv_module_0_sequential_5_weight, mean = encoder_conv_module_0_sequential_5_running_mean, variance = encoder_conv_module_0_sequential_5_running_var, x = input_25)[name = tensor("inputs_7")]; - tensor input_27 = silu(x = inputs_7)[name = tensor("input_27")]; - tensor input_29_pad_type_0 = const()[name = tensor("input_29_pad_type_0"), val = tensor("valid")]; - tensor input_29_strides_0 = const()[name = tensor("input_29_strides_0"), val = tensor([1])]; - tensor input_29_pad_0 = const()[name = tensor("input_29_pad_0"), val = tensor([0, 0])]; - tensor input_29_dilations_0 = const()[name = tensor("input_29_dilations_0"), val = tensor([1])]; - tensor input_29_groups_0 = const()[name = tensor("input_29_groups_0"), val = tensor(1)]; - tensor input_29 = conv(bias = encoder_conv_module_0_sequential_7_conv_bias, dilations = input_29_dilations_0, groups = input_29_groups_0, pad = input_29_pad_0, pad_type = input_29_pad_type_0, strides = input_29_strides_0, weight = encoder_conv_module_0_sequential_7_conv_weight, x = input_27)[name = tensor("input_29")]; + tensor outputs_aug_1 = conv(dilations = outputs_aug_1_dilations_0, groups = outputs_aug_1_groups_0, pad = outputs_aug_1_pad_0, pad_type = outputs_aug_1_pad_type_0, strides = outputs_aug_1_strides_0, weight = inner_encoder_conv_module_0_sequential_4_conv_weight, x = inputs_5)[name = tensor("outputs_aug_1")]; + tensor input_27_begin_0 = const()[name = tensor("input_27_begin_0"), val = tensor([0, 0, 0])]; + tensor input_27_end_0 = const()[name = tensor("input_27_end_0"), val = tensor([2, 256, 16])]; + tensor input_27_end_mask_0 = const()[name = tensor("input_27_end_mask_0"), val = tensor([true, true, false])]; + tensor input_27 = slice_by_index(begin = input_27_begin_0, end = input_27_end_0, end_mask = input_27_end_mask_0, x = outputs_aug_1)[name = tensor("input_27")]; + tensor inputs_7 = batch_norm(beta = inner_encoder_conv_module_0_sequential_5_bias, epsilon = var_46, gamma = inner_encoder_conv_module_0_sequential_5_weight, mean = inner_encoder_conv_module_0_sequential_5_running_mean, variance = inner_encoder_conv_module_0_sequential_5_running_var, x = input_27)[name = tensor("inputs_7")]; + tensor input_29 = silu(x = inputs_7)[name = tensor("input_29")]; + tensor input_31_pad_type_0 = const()[name = tensor("input_31_pad_type_0"), val = tensor("valid")]; + tensor input_31_strides_0 = const()[name = tensor("input_31_strides_0"), val = tensor([1])]; + tensor input_31_pad_0 = const()[name = tensor("input_31_pad_0"), val = tensor([0, 0])]; + tensor input_31_dilations_0 = const()[name = tensor("input_31_dilations_0"), val = tensor([1])]; + tensor input_31_groups_0 = const()[name = tensor("input_31_groups_0"), val = tensor(1)]; + tensor input_31 = conv(bias = inner_encoder_conv_module_0_sequential_7_conv_bias, dilations = input_31_dilations_0, groups = input_31_groups_0, pad = input_31_pad_0, pad_type = input_31_pad_type_0, strides = input_31_strides_0, weight = inner_encoder_conv_module_0_sequential_7_conv_weight, x = input_29)[name = tensor("input_31")]; tensor conv_out_1_perm_0 = const()[name = tensor("conv_out_1_perm_0"), val = tensor([0, 2, 1])]; - tensor var_290_begin_0 = const()[name = tensor("op_290_begin_0"), val = tensor([0, -1, 0])]; - tensor var_290_end_0 = const()[name = tensor("op_290_end_0"), val = tensor([2, 16, 256])]; - tensor var_290_end_mask_0 = const()[name = tensor("op_290_end_mask_0"), val = tensor([true, true, true])]; - tensor conv_out_1 = transpose(perm = conv_out_1_perm_0, x = input_29)[name = tensor("transpose_53")]; - tensor var_290 = slice_by_index(begin = var_290_begin_0, end = var_290_end_0, end_mask = var_290_end_mask_0, x = conv_out_1)[name = tensor("op_290")]; - tensor var_292_perm_0 = const()[name = tensor("op_292_perm_0"), val = tensor([1, 0, 2])]; - tensor var_292 = transpose(perm = var_292_perm_0, x = var_290)[name = tensor("transpose_52")]; - tensor input_31 = add(x = x_3, y = var_292)[name = tensor("input_31")]; - tensor input_33_axes_0 = const()[name = tensor("input_33_axes_0"), val = tensor([-1])]; - tensor input_33 = layer_norm(axes = input_33_axes_0, beta = encoder_ffn2_0_module_sequential_0_bias, epsilon = var_28, gamma = encoder_ffn2_0_module_sequential_0_weight, x = input_31)[name = tensor("input_33")]; - tensor inputs_9 = linear(bias = encoder_ffn2_0_module_sequential_1_linear_bias, weight = encoder_ffn2_0_module_sequential_1_linear_weight, x = input_33)[name = tensor("linear_8")]; - tensor input_35 = silu(x = inputs_9)[name = tensor("input_35")]; - tensor input_39 = linear(bias = encoder_ffn2_0_module_sequential_4_linear_bias, weight = encoder_ffn2_0_module_sequential_4_linear_weight, x = input_35)[name = tensor("linear_9")]; - tensor var_315 = const()[name = tensor("op_315"), val = tensor(0x1p-1)]; - tensor var_316 = mul(x = input_39, y = var_315)[name = tensor("op_316")]; - tensor input_41 = add(x = var_316, y = input_31)[name = tensor("input_41")]; - tensor input_43_axes_0 = const()[name = tensor("input_43_axes_0"), val = tensor([-1])]; - tensor input_43 = layer_norm(axes = input_43_axes_0, beta = encoder_layer_norm_0_bias, epsilon = var_28, gamma = encoder_layer_norm_0_weight, x = input_41)[name = tensor("input_43")]; + tensor var_326_begin_0 = const()[name = tensor("op_326_begin_0"), val = tensor([0, -1, 0])]; + tensor var_326_end_0 = const()[name = tensor("op_326_end_0"), val = tensor([2, 16, 256])]; + tensor var_326_end_mask_0 = const()[name = tensor("op_326_end_mask_0"), val = tensor([true, true, true])]; + tensor conv_out_1 = transpose(perm = conv_out_1_perm_0, x = input_31)[name = tensor("transpose_53")]; + tensor var_326 = slice_by_index(begin = var_326_begin_0, end = var_326_end_0, end_mask = var_326_end_mask_0, x = conv_out_1)[name = tensor("op_326")]; + tensor var_328_perm_0 = const()[name = tensor("op_328_perm_0"), val = tensor([1, 0, 2])]; + tensor var_328 = transpose(perm = var_328_perm_0, x = var_326)[name = tensor("transpose_52")]; + tensor input_33 = add(x = x_3, y = var_328)[name = tensor("input_33")]; + tensor input_35_axes_0 = const()[name = tensor("input_35_axes_0"), val = tensor([-1])]; + tensor input_35 = layer_norm(axes = input_35_axes_0, beta = inner_encoder_ffn2_0_module_sequential_0_bias, epsilon = var_46, gamma = inner_encoder_ffn2_0_module_sequential_0_weight, x = input_33)[name = tensor("input_35")]; + tensor inputs_9 = linear(bias = inner_encoder_ffn2_0_module_sequential_1_linear_bias, weight = inner_encoder_ffn2_0_module_sequential_1_linear_weight, x = input_35)[name = tensor("linear_8")]; + tensor input_37 = silu(x = inputs_9)[name = tensor("input_37")]; + tensor input_41 = linear(bias = inner_encoder_ffn2_0_module_sequential_4_linear_bias, weight = inner_encoder_ffn2_0_module_sequential_4_linear_weight, x = input_37)[name = tensor("linear_9")]; + tensor var_351 = const()[name = tensor("op_351"), val = tensor(0x1p-1)]; + tensor var_352 = mul(x = input_41, y = var_351)[name = tensor("op_352")]; + tensor input_43 = add(x = var_352, y = input_33)[name = tensor("input_43")]; tensor input_45_axes_0 = const()[name = tensor("input_45_axes_0"), val = tensor([-1])]; - tensor input_45 = layer_norm(axes = input_45_axes_0, beta = encoder_ffn1_1_module_sequential_0_bias, epsilon = var_28, gamma = encoder_ffn1_1_module_sequential_0_weight, x = input_43)[name = tensor("input_45")]; - tensor inputs_11 = linear(bias = encoder_ffn1_1_module_sequential_1_linear_bias, weight = encoder_ffn1_1_module_sequential_1_linear_weight, x = input_45)[name = tensor("linear_10")]; - tensor input_47 = silu(x = inputs_11)[name = tensor("input_47")]; - tensor input_51 = linear(bias = encoder_ffn1_1_module_sequential_4_linear_bias, weight = encoder_ffn1_1_module_sequential_4_linear_weight, x = input_47)[name = tensor("linear_11")]; - tensor var_345 = const()[name = tensor("op_345"), val = tensor(0x1p-1)]; - tensor var_346 = mul(x = input_51, y = var_345)[name = tensor("op_346")]; - tensor input_53 = add(x = var_346, y = input_43)[name = tensor("input_53")]; + tensor input_45 = layer_norm(axes = input_45_axes_0, beta = inner_encoder_layer_norm_0_bias, epsilon = var_46, gamma = inner_encoder_layer_norm_0_weight, x = input_43)[name = tensor("input_45")]; + tensor input_47_axes_0 = const()[name = tensor("input_47_axes_0"), val = tensor([-1])]; + tensor input_47 = layer_norm(axes = input_47_axes_0, beta = inner_encoder_ffn1_1_module_sequential_0_bias, epsilon = var_46, gamma = inner_encoder_ffn1_1_module_sequential_0_weight, x = input_45)[name = tensor("input_47")]; + tensor inputs_11 = linear(bias = inner_encoder_ffn1_1_module_sequential_1_linear_bias, weight = inner_encoder_ffn1_1_module_sequential_1_linear_weight, x = input_47)[name = tensor("linear_10")]; + tensor input_49 = silu(x = inputs_11)[name = tensor("input_49")]; + tensor input_53 = linear(bias = inner_encoder_ffn1_1_module_sequential_4_linear_bias, weight = inner_encoder_ffn1_1_module_sequential_4_linear_weight, x = input_49)[name = tensor("linear_11")]; + tensor var_381 = const()[name = tensor("op_381"), val = tensor(0x1p-1)]; + tensor var_382 = mul(x = input_53, y = var_381)[name = tensor("op_382")]; + tensor input_55 = add(x = var_382, y = input_45)[name = tensor("input_55")]; tensor x_7_axes_0 = const()[name = tensor("x_7_axes_0"), val = tensor([-1])]; - tensor x_7 = layer_norm(axes = x_7_axes_0, beta = encoder_ret_lns_1_bias, epsilon = var_28, gamma = encoder_ret_lns_1_weight, x = input_53)[name = tensor("x_7")]; + tensor x_7 = layer_norm(axes = x_7_axes_0, beta = inner_encoder_ret_lns_1_bias, epsilon = var_46, gamma = inner_encoder_ret_lns_1_weight, x = input_55)[name = tensor("x_7")]; tensor prev_kv_3_begin_0 = const()[name = tensor("prev_kv_3_begin_0"), val = tensor([1, 0, 0, 0, 0])]; tensor prev_kv_3_end_0 = const()[name = tensor("prev_kv_3_end_0"), val = tensor([2, 1, 4, 64, 64])]; tensor prev_kv_3_end_mask_0 = const()[name = tensor("prev_kv_3_end_mask_0"), val = tensor([false, true, true, true, true])]; @@ -396,153 +410,153 @@ program(1.0) tensor prev_scale_3_end_mask_0 = const()[name = tensor("prev_scale_3_end_mask_0"), val = tensor([false, true])]; tensor prev_scale_3_squeeze_mask_0 = const()[name = tensor("prev_scale_3_squeeze_mask_0"), val = tensor([true, false])]; tensor prev_scale_3 = slice_by_index(begin = prev_scale_3_begin_0, end = prev_scale_3_end_0, end_mask = prev_scale_3_end_mask_0, squeeze_mask = prev_scale_3_squeeze_mask_0, x = enc_scale)[name = tensor("prev_scale_3")]; - tensor var_360 = linear(bias = encoder_q_proj_1_bias, weight = encoder_q_proj_1_weight, x = x_7)[name = tensor("linear_12")]; - tensor var_361 = const()[name = tensor("op_361"), val = tensor([1, 2, 4, 64])]; - tensor var_362 = reshape(shape = var_361, x = var_360)[name = tensor("op_362")]; + tensor var_396 = linear(bias = inner_encoder_q_proj_1_bias, weight = inner_encoder_q_proj_1_weight, x = x_7)[name = tensor("linear_12")]; + tensor var_397 = const()[name = tensor("op_397"), val = tensor([1, 2, 4, 64])]; + tensor var_398 = reshape(shape = var_397, x = var_396)[name = tensor("op_398")]; tensor q_3_perm_0 = const()[name = tensor("q_3_perm_0"), val = tensor([0, 2, 1, 3])]; - tensor var_366 = linear(bias = encoder_k_proj_1_bias, weight = encoder_k_proj_1_weight, x = x_7)[name = tensor("linear_13")]; - tensor var_367 = const()[name = tensor("op_367"), val = tensor(0x1p-3)]; - tensor var_368 = mul(x = var_366, y = var_367)[name = tensor("op_368")]; - tensor var_369 = const()[name = tensor("op_369"), val = tensor([1, 2, 4, 64])]; - tensor var_370 = reshape(shape = var_369, x = var_368)[name = tensor("op_370")]; + tensor var_402 = linear(bias = inner_encoder_k_proj_1_bias, weight = inner_encoder_k_proj_1_weight, x = x_7)[name = tensor("linear_13")]; + tensor var_403 = const()[name = tensor("op_403"), val = tensor(0x1p-3)]; + tensor var_404 = mul(x = var_402, y = var_403)[name = tensor("op_404")]; + tensor var_405 = const()[name = tensor("op_405"), val = tensor([1, 2, 4, 64])]; + tensor var_406 = reshape(shape = var_405, x = var_404)[name = tensor("op_406")]; tensor k_3_perm_0 = const()[name = tensor("k_3_perm_0"), val = tensor([0, 2, 1, 3])]; - tensor var_374 = linear(bias = encoder_v_proj_1_bias, weight = encoder_v_proj_1_weight, x = x_7)[name = tensor("linear_14")]; - tensor var_375 = const()[name = tensor("op_375"), val = tensor([1, 2, 4, 64])]; - tensor var_376 = reshape(shape = var_375, x = var_374)[name = tensor("op_376")]; + tensor var_410 = linear(bias = inner_encoder_v_proj_1_bias, weight = inner_encoder_v_proj_1_weight, x = x_7)[name = tensor("linear_14")]; + tensor var_411 = const()[name = tensor("op_411"), val = tensor([1, 2, 4, 64])]; + tensor var_412 = reshape(shape = var_411, x = var_410)[name = tensor("op_412")]; tensor v_3_perm_0 = const()[name = tensor("v_3_perm_0"), val = tensor([0, 2, 1, 3])]; - tensor input_57 = linear(bias = encoder_g_proj_1_bias, weight = encoder_g_proj_1_weight, x = x_7)[name = tensor("linear_15")]; + tensor input_59 = linear(bias = inner_encoder_g_proj_1_bias, weight = inner_encoder_g_proj_1_weight, x = x_7)[name = tensor("linear_15")]; tensor sqrt_s0_3 = sqrt(x = prev_scale_3)[name = tensor("sqrt_s0_3")]; - tensor s_t_3 = add(x = prev_scale_3, y = encoder__t_index)[name = tensor("s_t_3")]; + tensor s_t_3 = add(x = prev_scale_3, y = inner_encoder__t_index)[name = tensor("s_t_3")]; tensor sqrt_s_t_3 = sqrt(x = s_t_3)[name = tensor("sqrt_s_t_3")]; tensor qk_3_transpose_x_1 = const()[name = tensor("qk_3_transpose_x_1"), val = tensor(false)]; tensor qk_3_transpose_y_1 = const()[name = tensor("qk_3_transpose_y_1"), val = tensor(true)]; - tensor k_3 = transpose(perm = k_3_perm_0, x = var_370)[name = tensor("transpose_50")]; - tensor q_3 = transpose(perm = q_3_perm_0, x = var_362)[name = tensor("transpose_51")]; + tensor k_3 = transpose(perm = k_3_perm_0, x = var_406)[name = tensor("transpose_50")]; + tensor q_3 = transpose(perm = q_3_perm_0, x = var_398)[name = tensor("transpose_51")]; tensor qk_3 = matmul(transpose_x = qk_3_transpose_x_1, transpose_y = qk_3_transpose_y_1, x = q_3, y = k_3)[name = tensor("qk_3")]; - tensor var_386 = const()[name = tensor("op_386"), val = tensor([2, 1])]; - tensor var_387 = reshape(shape = var_386, x = sqrt_s_t_3)[name = tensor("op_387")]; - tensor M_3 = real_div(x = encoder__causal_mask, y = var_387)[name = tensor("M_3")]; - tensor var_389 = mul(x = qk_3, y = M_3)[name = tensor("op_389")]; + tensor var_422 = const()[name = tensor("op_422"), val = tensor([2, 1])]; + tensor var_423 = reshape(shape = var_422, x = sqrt_s_t_3)[name = tensor("op_423")]; + tensor M_3 = real_div(x = inner_encoder__causal_mask, y = var_423)[name = tensor("M_3")]; + tensor var_425 = mul(x = qk_3, y = M_3)[name = tensor("op_425")]; tensor inner_3_transpose_x_0 = const()[name = tensor("inner_3_transpose_x_0"), val = tensor(false)]; tensor inner_3_transpose_y_0 = const()[name = tensor("inner_3_transpose_y_0"), val = tensor(false)]; - tensor v_3 = transpose(perm = v_3_perm_0, x = var_376)[name = tensor("transpose_49")]; - tensor inner_3 = matmul(transpose_x = inner_3_transpose_x_0, transpose_y = inner_3_transpose_y_0, x = var_389, y = v_3)[name = tensor("inner_3")]; - tensor var_391_transpose_x_0 = const()[name = tensor("op_391_transpose_x_0"), val = tensor(false)]; - tensor var_391_transpose_y_0 = const()[name = tensor("op_391_transpose_y_0"), val = tensor(false)]; - tensor var_391 = matmul(transpose_x = var_391_transpose_x_0, transpose_y = var_391_transpose_y_0, x = q_3, y = prev_kv_3)[name = tensor("op_391")]; - tensor var_392 = real_div(x = sqrt_s0_3, y = sqrt_s_t_3)[name = tensor("op_392")]; - tensor var_393 = const()[name = tensor("op_393"), val = tensor([1, 1, 2, 1])]; - tensor var_394 = reshape(shape = var_393, x = var_392)[name = tensor("op_394")]; - tensor cross_3 = mul(x = var_391, y = var_394)[name = tensor("cross_3")]; + tensor v_3 = transpose(perm = v_3_perm_0, x = var_412)[name = tensor("transpose_49")]; + tensor inner_3 = matmul(transpose_x = inner_3_transpose_x_0, transpose_y = inner_3_transpose_y_0, x = var_425, y = v_3)[name = tensor("inner_3")]; + tensor var_427_transpose_x_0 = const()[name = tensor("op_427_transpose_x_0"), val = tensor(false)]; + tensor var_427_transpose_y_0 = const()[name = tensor("op_427_transpose_y_0"), val = tensor(false)]; + tensor var_427 = matmul(transpose_x = var_427_transpose_x_0, transpose_y = var_427_transpose_y_0, x = q_3, y = prev_kv_3)[name = tensor("op_427")]; + tensor var_428 = real_div(x = sqrt_s0_3, y = sqrt_s_t_3)[name = tensor("op_428")]; + tensor var_429 = const()[name = tensor("op_429"), val = tensor([1, 1, 2, 1])]; + tensor var_430 = reshape(shape = var_429, x = var_428)[name = tensor("op_430")]; + tensor cross_3 = mul(x = var_427, y = var_430)[name = tensor("cross_3")]; tensor out_7 = add(x = inner_3, y = cross_3)[name = tensor("out_7")]; - tensor var_397 = mul(x = prev_kv_3, y = sqrt_s0_3)[name = tensor("op_397")]; - tensor var_399_transpose_x_1 = const()[name = tensor("op_399_transpose_x_1"), val = tensor(true)]; - tensor var_399_transpose_y_1 = const()[name = tensor("op_399_transpose_y_1"), val = tensor(false)]; - tensor var_399 = matmul(transpose_x = var_399_transpose_x_1, transpose_y = var_399_transpose_y_1, x = k_3, y = v_3)[name = tensor("op_399")]; - tensor new_kv_unnorm_3 = add(x = var_397, y = var_399)[name = tensor("new_kv_unnorm_3")]; - tensor var_401 = const()[name = tensor("op_401"), val = tensor(0x1p+1)]; - tensor new_scale_3 = add(x = prev_scale_3, y = var_401)[name = tensor("new_scale_3")]; - tensor var_403 = sqrt(x = new_scale_3)[name = tensor("op_403")]; - tensor var_404 = real_div(x = new_kv_unnorm_3, y = var_403)[name = tensor("op_404")]; - tensor var_405_perm_0 = const()[name = tensor("op_405_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_433 = mul(x = prev_kv_3, y = sqrt_s0_3)[name = tensor("op_433")]; + tensor var_435_transpose_x_1 = const()[name = tensor("op_435_transpose_x_1"), val = tensor(true)]; + tensor var_435_transpose_y_1 = const()[name = tensor("op_435_transpose_y_1"), val = tensor(false)]; + tensor var_435 = matmul(transpose_x = var_435_transpose_x_1, transpose_y = var_435_transpose_y_1, x = k_3, y = v_3)[name = tensor("op_435")]; + tensor new_kv_unnorm_3 = add(x = var_433, y = var_435)[name = tensor("new_kv_unnorm_3")]; + tensor var_437 = const()[name = tensor("op_437"), val = tensor(0x1p+1)]; + tensor new_scale_3 = add(x = prev_scale_3, y = var_437)[name = tensor("new_scale_3")]; + tensor var_439 = sqrt(x = new_scale_3)[name = tensor("op_439")]; + tensor var_440 = real_div(x = new_kv_unnorm_3, y = var_439)[name = tensor("op_440")]; + tensor var_441_perm_0 = const()[name = tensor("op_441_perm_0"), val = tensor([0, 2, 1, 3])]; tensor out_9_axes_0 = const()[name = tensor("out_9_axes_0"), val = tensor([-1])]; - tensor var_405 = transpose(perm = var_405_perm_0, x = out_7)[name = tensor("transpose_48")]; - tensor out_9 = layer_norm(axes = out_9_axes_0, epsilon = var_17, x = var_405)[name = tensor("out_9")]; - tensor var_409 = const()[name = tensor("op_409"), val = tensor([1, 2, 256])]; - tensor out_11 = reshape(shape = var_409, x = out_9)[name = tensor("out_11")]; - tensor var_411 = silu(x = input_57)[name = tensor("op_411")]; - tensor input_59 = mul(x = var_411, y = out_11)[name = tensor("input_59")]; - tensor ret_out_3 = linear(bias = encoder_out_proj_1_bias, weight = encoder_out_proj_1_weight, x = input_59)[name = tensor("linear_16")]; - tensor x_9 = add(x = input_53, y = ret_out_3)[name = tensor("x_9")]; + tensor var_441 = transpose(perm = var_441_perm_0, x = out_7)[name = tensor("transpose_48")]; + tensor out_9 = layer_norm(axes = out_9_axes_0, epsilon = var_54, x = var_441)[name = tensor("out_9")]; + tensor var_445 = const()[name = tensor("op_445"), val = tensor([1, 2, 256])]; + tensor out_11 = reshape(shape = var_445, x = out_9)[name = tensor("out_11")]; + tensor var_447 = silu(x = input_59)[name = tensor("op_447")]; + tensor input_61 = mul(x = var_447, y = out_11)[name = tensor("input_61")]; + tensor ret_out_3 = linear(bias = inner_encoder_out_proj_1_bias, weight = inner_encoder_out_proj_1_weight, x = input_61)[name = tensor("linear_16")]; + tensor x_9 = add(x = input_55, y = ret_out_3)[name = tensor("x_9")]; tensor window_7_begin_0 = const()[name = tensor("window_7_begin_0"), val = tensor([1, 0, 0, 0])]; tensor window_7_end_0 = const()[name = tensor("window_7_end_0"), val = tensor([2, 1, 16, 256])]; tensor window_7_end_mask_0 = const()[name = tensor("window_7_end_mask_0"), val = tensor([false, true, true, true])]; tensor window_7_squeeze_mask_0 = const()[name = tensor("window_7_squeeze_mask_0"), val = tensor([true, false, false, false])]; tensor window_7 = slice_by_index(begin = window_7_begin_0, end = window_7_end_0, end_mask = window_7_end_mask_0, squeeze_mask = window_7_squeeze_mask_0, x = enc_conv_cache)[name = tensor("window_7")]; - tensor var_419_begin_0 = const()[name = tensor("op_419_begin_0"), val = tensor([0, 0, 0])]; - tensor var_419_end_0 = const()[name = tensor("op_419_end_0"), val = tensor([1, 1, 256])]; - tensor var_419_end_mask_0 = const()[name = tensor("op_419_end_mask_0"), val = tensor([true, false, true])]; - tensor var_419 = slice_by_index(begin = var_419_begin_0, end = var_419_end_0, end_mask = var_419_end_mask_0, x = x_9)[name = tensor("op_419")]; - tensor var_422_begin_0 = const()[name = tensor("op_422_begin_0"), val = tensor([0, 1, 0])]; - tensor var_422_end_0 = const()[name = tensor("op_422_end_0"), val = tensor([1, 16, 256])]; - tensor var_422_end_mask_0 = const()[name = tensor("op_422_end_mask_0"), val = tensor([true, true, true])]; - tensor var_422 = slice_by_index(begin = var_422_begin_0, end = var_422_end_0, end_mask = var_422_end_mask_0, x = window_7)[name = tensor("op_422")]; + tensor var_455_begin_0 = const()[name = tensor("op_455_begin_0"), val = tensor([0, 0, 0])]; + tensor var_455_end_0 = const()[name = tensor("op_455_end_0"), val = tensor([1, 1, 256])]; + tensor var_455_end_mask_0 = const()[name = tensor("op_455_end_mask_0"), val = tensor([true, false, true])]; + tensor var_455 = slice_by_index(begin = var_455_begin_0, end = var_455_end_0, end_mask = var_455_end_mask_0, x = x_9)[name = tensor("op_455")]; + tensor var_458_begin_0 = const()[name = tensor("op_458_begin_0"), val = tensor([0, 1, 0])]; + tensor var_458_end_0 = const()[name = tensor("op_458_end_0"), val = tensor([1, 16, 256])]; + tensor var_458_end_mask_0 = const()[name = tensor("op_458_end_mask_0"), val = tensor([true, true, true])]; + tensor var_458 = slice_by_index(begin = var_458_begin_0, end = var_458_end_0, end_mask = var_458_end_mask_0, x = window_7)[name = tensor("op_458")]; tensor window_9_interleave_0 = const()[name = tensor("window_9_interleave_0"), val = tensor(false)]; - tensor window_9 = concat(axis = var_26, interleave = window_9_interleave_0, values = (var_422, var_419))[name = tensor("window_9")]; - tensor var_427_begin_0 = const()[name = tensor("op_427_begin_0"), val = tensor([0, 1, 0])]; - tensor var_427_end_0 = const()[name = tensor("op_427_end_0"), val = tensor([1, 1, 256])]; - tensor var_427_end_mask_0 = const()[name = tensor("op_427_end_mask_0"), val = tensor([true, true, true])]; - tensor var_427 = slice_by_index(begin = var_427_begin_0, end = var_427_end_0, end_mask = var_427_end_mask_0, x = x_9)[name = tensor("op_427")]; - tensor var_430_begin_0 = const()[name = tensor("op_430_begin_0"), val = tensor([0, 1, 0])]; - tensor var_430_end_0 = const()[name = tensor("op_430_end_0"), val = tensor([1, 16, 256])]; - tensor var_430_end_mask_0 = const()[name = tensor("op_430_end_mask_0"), val = tensor([true, true, true])]; - tensor var_430 = slice_by_index(begin = var_430_begin_0, end = var_430_end_0, end_mask = var_430_end_mask_0, x = window_9)[name = tensor("op_430")]; + tensor window_9 = concat(axis = var_62, interleave = window_9_interleave_0, values = (var_458, var_455))[name = tensor("window_9")]; + tensor var_463_begin_0 = const()[name = tensor("op_463_begin_0"), val = tensor([0, 1, 0])]; + tensor var_463_end_0 = const()[name = tensor("op_463_end_0"), val = tensor([1, 1, 256])]; + tensor var_463_end_mask_0 = const()[name = tensor("op_463_end_mask_0"), val = tensor([true, true, true])]; + tensor var_463 = slice_by_index(begin = var_463_begin_0, end = var_463_end_0, end_mask = var_463_end_mask_0, x = x_9)[name = tensor("op_463")]; + tensor var_466_begin_0 = const()[name = tensor("op_466_begin_0"), val = tensor([0, 1, 0])]; + tensor var_466_end_0 = const()[name = tensor("op_466_end_0"), val = tensor([1, 16, 256])]; + tensor var_466_end_mask_0 = const()[name = tensor("op_466_end_mask_0"), val = tensor([true, true, true])]; + tensor var_466 = slice_by_index(begin = var_466_begin_0, end = var_466_end_0, end_mask = var_466_end_mask_0, x = window_9)[name = tensor("op_466")]; tensor window_11_interleave_0 = const()[name = tensor("window_11_interleave_0"), val = tensor(false)]; - tensor window_11 = concat(axis = var_26, interleave = window_11_interleave_0, values = (var_430, var_427))[name = tensor("window_11")]; - tensor input_61_interleave_0 = const()[name = tensor("input_61_interleave_0"), val = tensor(false)]; - tensor input_61 = concat(axis = var_23, interleave = input_61_interleave_0, values = (window_9, window_11))[name = tensor("input_61")]; + tensor window_11 = concat(axis = var_62, interleave = window_11_interleave_0, values = (var_466, var_463))[name = tensor("window_11")]; + tensor input_63_interleave_0 = const()[name = tensor("input_63_interleave_0"), val = tensor(false)]; + tensor input_63 = concat(axis = var_49, interleave = input_63_interleave_0, values = (window_9, window_11))[name = tensor("input_63")]; tensor x_11_axes_0 = const()[name = tensor("x_11_axes_0"), val = tensor([-1])]; - tensor x_11 = layer_norm(axes = x_11_axes_0, beta = encoder_conv_module_1_sequential_0_bias, epsilon = var_28, gamma = encoder_conv_module_1_sequential_0_weight, x = input_61)[name = tensor("x_11")]; - tensor input_63_perm_0 = const()[name = tensor("input_63_perm_0"), val = tensor([0, 2, 1])]; + tensor x_11 = layer_norm(axes = x_11_axes_0, beta = inner_encoder_conv_module_1_sequential_0_bias, epsilon = var_46, gamma = inner_encoder_conv_module_1_sequential_0_weight, x = input_63)[name = tensor("x_11")]; + tensor input_65_perm_0 = const()[name = tensor("input_65_perm_0"), val = tensor([0, 2, 1])]; tensor inputs_13_pad_type_0 = const()[name = tensor("inputs_13_pad_type_0"), val = tensor("valid")]; tensor inputs_13_strides_0 = const()[name = tensor("inputs_13_strides_0"), val = tensor([1])]; tensor inputs_13_pad_0 = const()[name = tensor("inputs_13_pad_0"), val = tensor([0, 0])]; tensor inputs_13_dilations_0 = const()[name = tensor("inputs_13_dilations_0"), val = tensor([1])]; tensor inputs_13_groups_0 = const()[name = tensor("inputs_13_groups_0"), val = tensor(1)]; - tensor input_63 = transpose(perm = input_63_perm_0, x = x_11)[name = tensor("transpose_47")]; - tensor inputs_13 = conv(bias = encoder_conv_module_1_sequential_2_conv_bias, dilations = inputs_13_dilations_0, groups = inputs_13_groups_0, pad = inputs_13_pad_0, pad_type = inputs_13_pad_type_0, strides = inputs_13_strides_0, weight = encoder_conv_module_1_sequential_2_conv_weight, x = input_63)[name = tensor("inputs_13")]; - tensor var_455_split_sizes_0 = const()[name = tensor("op_455_split_sizes_0"), val = tensor([256, 256])]; - tensor var_455_axis_0 = const()[name = tensor("op_455_axis_0"), val = tensor(1)]; - tensor var_455_0, tensor var_455_1 = split(axis = var_455_axis_0, split_sizes = var_455_split_sizes_0, x = inputs_13)[name = tensor("op_455")]; - tensor var_457 = sigmoid(x = var_455_1)[name = tensor("op_457")]; - tensor inputs_15 = mul(x = var_455_0, y = var_457)[name = tensor("inputs_15")]; + tensor input_65 = transpose(perm = input_65_perm_0, x = x_11)[name = tensor("transpose_47")]; + tensor inputs_13 = conv(bias = inner_encoder_conv_module_1_sequential_2_conv_bias, dilations = inputs_13_dilations_0, groups = inputs_13_groups_0, pad = inputs_13_pad_0, pad_type = inputs_13_pad_type_0, strides = inputs_13_strides_0, weight = inner_encoder_conv_module_1_sequential_2_conv_weight, x = input_65)[name = tensor("inputs_13")]; + tensor var_491_split_sizes_0 = const()[name = tensor("op_491_split_sizes_0"), val = tensor([256, 256])]; + tensor var_491_axis_0 = const()[name = tensor("op_491_axis_0"), val = tensor(1)]; + tensor var_491_0, tensor var_491_1 = split(axis = var_491_axis_0, split_sizes = var_491_split_sizes_0, x = inputs_13)[name = tensor("op_491")]; + tensor var_493 = sigmoid(x = var_491_1)[name = tensor("op_493")]; + tensor inputs_15 = mul(x = var_491_0, y = var_493)[name = tensor("inputs_15")]; tensor outputs_aug_3_pad_type_0 = const()[name = tensor("outputs_aug_3_pad_type_0"), val = tensor("custom")]; tensor outputs_aug_3_pad_0 = const()[name = tensor("outputs_aug_3_pad_0"), val = tensor([15, 15])]; tensor outputs_aug_3_groups_0 = const()[name = tensor("outputs_aug_3_groups_0"), val = tensor(256)]; tensor outputs_aug_3_strides_0 = const()[name = tensor("outputs_aug_3_strides_0"), val = tensor([1])]; tensor outputs_aug_3_dilations_0 = const()[name = tensor("outputs_aug_3_dilations_0"), val = tensor([1])]; - tensor outputs_aug_3 = conv(dilations = outputs_aug_3_dilations_0, groups = outputs_aug_3_groups_0, pad = outputs_aug_3_pad_0, pad_type = outputs_aug_3_pad_type_0, strides = outputs_aug_3_strides_0, weight = encoder_conv_module_1_sequential_4_conv_weight, x = inputs_15)[name = tensor("outputs_aug_3")]; - tensor input_65_begin_0 = const()[name = tensor("input_65_begin_0"), val = tensor([0, 0, 0])]; - tensor input_65_end_0 = const()[name = tensor("input_65_end_0"), val = tensor([2, 256, 16])]; - tensor input_65_end_mask_0 = const()[name = tensor("input_65_end_mask_0"), val = tensor([true, true, false])]; - tensor input_65 = slice_by_index(begin = input_65_begin_0, end = input_65_end_0, end_mask = input_65_end_mask_0, x = outputs_aug_3)[name = tensor("input_65")]; - tensor inputs_17 = batch_norm(beta = encoder_conv_module_1_sequential_5_bias, epsilon = var_28, gamma = encoder_conv_module_1_sequential_5_weight, mean = encoder_conv_module_1_sequential_5_running_mean, variance = encoder_conv_module_1_sequential_5_running_var, x = input_65)[name = tensor("inputs_17")]; - tensor input_67 = silu(x = inputs_17)[name = tensor("input_67")]; - tensor input_69_pad_type_0 = const()[name = tensor("input_69_pad_type_0"), val = tensor("valid")]; - tensor input_69_strides_0 = const()[name = tensor("input_69_strides_0"), val = tensor([1])]; - tensor input_69_pad_0 = const()[name = tensor("input_69_pad_0"), val = tensor([0, 0])]; - tensor input_69_dilations_0 = const()[name = tensor("input_69_dilations_0"), val = tensor([1])]; - tensor input_69_groups_0 = const()[name = tensor("input_69_groups_0"), val = tensor(1)]; - tensor input_69 = conv(bias = encoder_conv_module_1_sequential_7_conv_bias, dilations = input_69_dilations_0, groups = input_69_groups_0, pad = input_69_pad_0, pad_type = input_69_pad_type_0, strides = input_69_strides_0, weight = encoder_conv_module_1_sequential_7_conv_weight, x = input_67)[name = tensor("input_69")]; + tensor outputs_aug_3 = conv(dilations = outputs_aug_3_dilations_0, groups = outputs_aug_3_groups_0, pad = outputs_aug_3_pad_0, pad_type = outputs_aug_3_pad_type_0, strides = outputs_aug_3_strides_0, weight = inner_encoder_conv_module_1_sequential_4_conv_weight, x = inputs_15)[name = tensor("outputs_aug_3")]; + tensor input_67_begin_0 = const()[name = tensor("input_67_begin_0"), val = tensor([0, 0, 0])]; + tensor input_67_end_0 = const()[name = tensor("input_67_end_0"), val = tensor([2, 256, 16])]; + tensor input_67_end_mask_0 = const()[name = tensor("input_67_end_mask_0"), val = tensor([true, true, false])]; + tensor input_67 = slice_by_index(begin = input_67_begin_0, end = input_67_end_0, end_mask = input_67_end_mask_0, x = outputs_aug_3)[name = tensor("input_67")]; + tensor inputs_17 = batch_norm(beta = inner_encoder_conv_module_1_sequential_5_bias, epsilon = var_46, gamma = inner_encoder_conv_module_1_sequential_5_weight, mean = inner_encoder_conv_module_1_sequential_5_running_mean, variance = inner_encoder_conv_module_1_sequential_5_running_var, x = input_67)[name = tensor("inputs_17")]; + tensor input_69 = silu(x = inputs_17)[name = tensor("input_69")]; + tensor input_71_pad_type_0 = const()[name = tensor("input_71_pad_type_0"), val = tensor("valid")]; + tensor input_71_strides_0 = const()[name = tensor("input_71_strides_0"), val = tensor([1])]; + tensor input_71_pad_0 = const()[name = tensor("input_71_pad_0"), val = tensor([0, 0])]; + tensor input_71_dilations_0 = const()[name = tensor("input_71_dilations_0"), val = tensor([1])]; + tensor input_71_groups_0 = const()[name = tensor("input_71_groups_0"), val = tensor(1)]; + tensor input_71 = conv(bias = inner_encoder_conv_module_1_sequential_7_conv_bias, dilations = input_71_dilations_0, groups = input_71_groups_0, pad = input_71_pad_0, pad_type = input_71_pad_type_0, strides = input_71_strides_0, weight = inner_encoder_conv_module_1_sequential_7_conv_weight, x = input_69)[name = tensor("input_71")]; tensor conv_out_3_perm_0 = const()[name = tensor("conv_out_3_perm_0"), val = tensor([0, 2, 1])]; - tensor var_488_begin_0 = const()[name = tensor("op_488_begin_0"), val = tensor([0, -1, 0])]; - tensor var_488_end_0 = const()[name = tensor("op_488_end_0"), val = tensor([2, 16, 256])]; - tensor var_488_end_mask_0 = const()[name = tensor("op_488_end_mask_0"), val = tensor([true, true, true])]; - tensor conv_out_3 = transpose(perm = conv_out_3_perm_0, x = input_69)[name = tensor("transpose_46")]; - tensor var_488 = slice_by_index(begin = var_488_begin_0, end = var_488_end_0, end_mask = var_488_end_mask_0, x = conv_out_3)[name = tensor("op_488")]; - tensor var_490_perm_0 = const()[name = tensor("op_490_perm_0"), val = tensor([1, 0, 2])]; - tensor var_490 = transpose(perm = var_490_perm_0, x = var_488)[name = tensor("transpose_45")]; - tensor input_71 = add(x = x_9, y = var_490)[name = tensor("input_71")]; - tensor input_73_axes_0 = const()[name = tensor("input_73_axes_0"), val = tensor([-1])]; - tensor input_73 = layer_norm(axes = input_73_axes_0, beta = encoder_ffn2_1_module_sequential_0_bias, epsilon = var_28, gamma = encoder_ffn2_1_module_sequential_0_weight, x = input_71)[name = tensor("input_73")]; - tensor inputs_19 = linear(bias = encoder_ffn2_1_module_sequential_1_linear_bias, weight = encoder_ffn2_1_module_sequential_1_linear_weight, x = input_73)[name = tensor("linear_17")]; - tensor input_75 = silu(x = inputs_19)[name = tensor("input_75")]; - tensor input_79 = linear(bias = encoder_ffn2_1_module_sequential_4_linear_bias, weight = encoder_ffn2_1_module_sequential_4_linear_weight, x = input_75)[name = tensor("linear_18")]; - tensor var_513 = const()[name = tensor("op_513"), val = tensor(0x1p-1)]; - tensor var_514 = mul(x = input_79, y = var_513)[name = tensor("op_514")]; - tensor input_81 = add(x = var_514, y = input_71)[name = tensor("input_81")]; - tensor input_83_axes_0 = const()[name = tensor("input_83_axes_0"), val = tensor([-1])]; - tensor input_83 = layer_norm(axes = input_83_axes_0, beta = encoder_layer_norm_1_bias, epsilon = var_28, gamma = encoder_layer_norm_1_weight, x = input_81)[name = tensor("input_83")]; + tensor var_524_begin_0 = const()[name = tensor("op_524_begin_0"), val = tensor([0, -1, 0])]; + tensor var_524_end_0 = const()[name = tensor("op_524_end_0"), val = tensor([2, 16, 256])]; + tensor var_524_end_mask_0 = const()[name = tensor("op_524_end_mask_0"), val = tensor([true, true, true])]; + tensor conv_out_3 = transpose(perm = conv_out_3_perm_0, x = input_71)[name = tensor("transpose_46")]; + tensor var_524 = slice_by_index(begin = var_524_begin_0, end = var_524_end_0, end_mask = var_524_end_mask_0, x = conv_out_3)[name = tensor("op_524")]; + tensor var_526_perm_0 = const()[name = tensor("op_526_perm_0"), val = tensor([1, 0, 2])]; + tensor var_526 = transpose(perm = var_526_perm_0, x = var_524)[name = tensor("transpose_45")]; + tensor input_73 = add(x = x_9, y = var_526)[name = tensor("input_73")]; + tensor input_75_axes_0 = const()[name = tensor("input_75_axes_0"), val = tensor([-1])]; + tensor input_75 = layer_norm(axes = input_75_axes_0, beta = inner_encoder_ffn2_1_module_sequential_0_bias, epsilon = var_46, gamma = inner_encoder_ffn2_1_module_sequential_0_weight, x = input_73)[name = tensor("input_75")]; + tensor inputs_19 = linear(bias = inner_encoder_ffn2_1_module_sequential_1_linear_bias, weight = inner_encoder_ffn2_1_module_sequential_1_linear_weight, x = input_75)[name = tensor("linear_17")]; + tensor input_77 = silu(x = inputs_19)[name = tensor("input_77")]; + tensor input_81 = linear(bias = inner_encoder_ffn2_1_module_sequential_4_linear_bias, weight = inner_encoder_ffn2_1_module_sequential_4_linear_weight, x = input_77)[name = tensor("linear_18")]; + tensor var_549 = const()[name = tensor("op_549"), val = tensor(0x1p-1)]; + tensor var_550 = mul(x = input_81, y = var_549)[name = tensor("op_550")]; + tensor input_83 = add(x = var_550, y = input_73)[name = tensor("input_83")]; tensor input_85_axes_0 = const()[name = tensor("input_85_axes_0"), val = tensor([-1])]; - tensor input_85 = layer_norm(axes = input_85_axes_0, beta = encoder_ffn1_2_module_sequential_0_bias, epsilon = var_28, gamma = encoder_ffn1_2_module_sequential_0_weight, x = input_83)[name = tensor("input_85")]; - tensor inputs_21 = linear(bias = encoder_ffn1_2_module_sequential_1_linear_bias, weight = encoder_ffn1_2_module_sequential_1_linear_weight, x = input_85)[name = tensor("linear_19")]; - tensor input_87 = silu(x = inputs_21)[name = tensor("input_87")]; - tensor input_91 = linear(bias = encoder_ffn1_2_module_sequential_4_linear_bias, weight = encoder_ffn1_2_module_sequential_4_linear_weight, x = input_87)[name = tensor("linear_20")]; - tensor var_543 = const()[name = tensor("op_543"), val = tensor(0x1p-1)]; - tensor var_544 = mul(x = input_91, y = var_543)[name = tensor("op_544")]; - tensor input_93 = add(x = var_544, y = input_83)[name = tensor("input_93")]; + tensor input_85 = layer_norm(axes = input_85_axes_0, beta = inner_encoder_layer_norm_1_bias, epsilon = var_46, gamma = inner_encoder_layer_norm_1_weight, x = input_83)[name = tensor("input_85")]; + tensor input_87_axes_0 = const()[name = tensor("input_87_axes_0"), val = tensor([-1])]; + tensor input_87 = layer_norm(axes = input_87_axes_0, beta = inner_encoder_ffn1_2_module_sequential_0_bias, epsilon = var_46, gamma = inner_encoder_ffn1_2_module_sequential_0_weight, x = input_85)[name = tensor("input_87")]; + tensor inputs_21 = linear(bias = inner_encoder_ffn1_2_module_sequential_1_linear_bias, weight = inner_encoder_ffn1_2_module_sequential_1_linear_weight, x = input_87)[name = tensor("linear_19")]; + tensor input_89 = silu(x = inputs_21)[name = tensor("input_89")]; + tensor input_93 = linear(bias = inner_encoder_ffn1_2_module_sequential_4_linear_bias, weight = inner_encoder_ffn1_2_module_sequential_4_linear_weight, x = input_89)[name = tensor("linear_20")]; + tensor var_579 = const()[name = tensor("op_579"), val = tensor(0x1p-1)]; + tensor var_580 = mul(x = input_93, y = var_579)[name = tensor("op_580")]; + tensor input_95 = add(x = var_580, y = input_85)[name = tensor("input_95")]; tensor x_13_axes_0 = const()[name = tensor("x_13_axes_0"), val = tensor([-1])]; - tensor x_13 = layer_norm(axes = x_13_axes_0, beta = encoder_ret_lns_2_bias, epsilon = var_28, gamma = encoder_ret_lns_2_weight, x = input_93)[name = tensor("x_13")]; + tensor x_13 = layer_norm(axes = x_13_axes_0, beta = inner_encoder_ret_lns_2_bias, epsilon = var_46, gamma = inner_encoder_ret_lns_2_weight, x = input_95)[name = tensor("x_13")]; tensor prev_kv_5_begin_0 = const()[name = tensor("prev_kv_5_begin_0"), val = tensor([2, 0, 0, 0, 0])]; tensor prev_kv_5_end_0 = const()[name = tensor("prev_kv_5_end_0"), val = tensor([3, 1, 4, 64, 64])]; tensor prev_kv_5_end_mask_0 = const()[name = tensor("prev_kv_5_end_mask_0"), val = tensor([false, true, true, true, true])]; @@ -553,153 +567,153 @@ program(1.0) tensor prev_scale_5_end_mask_0 = const()[name = tensor("prev_scale_5_end_mask_0"), val = tensor([false, true])]; tensor prev_scale_5_squeeze_mask_0 = const()[name = tensor("prev_scale_5_squeeze_mask_0"), val = tensor([true, false])]; tensor prev_scale_5 = slice_by_index(begin = prev_scale_5_begin_0, end = prev_scale_5_end_0, end_mask = prev_scale_5_end_mask_0, squeeze_mask = prev_scale_5_squeeze_mask_0, x = enc_scale)[name = tensor("prev_scale_5")]; - tensor var_558 = linear(bias = encoder_q_proj_2_bias, weight = encoder_q_proj_2_weight, x = x_13)[name = tensor("linear_21")]; - tensor var_559 = const()[name = tensor("op_559"), val = tensor([1, 2, 4, 64])]; - tensor var_560 = reshape(shape = var_559, x = var_558)[name = tensor("op_560")]; + tensor var_594 = linear(bias = inner_encoder_q_proj_2_bias, weight = inner_encoder_q_proj_2_weight, x = x_13)[name = tensor("linear_21")]; + tensor var_595 = const()[name = tensor("op_595"), val = tensor([1, 2, 4, 64])]; + tensor var_596 = reshape(shape = var_595, x = var_594)[name = tensor("op_596")]; tensor q_5_perm_0 = const()[name = tensor("q_5_perm_0"), val = tensor([0, 2, 1, 3])]; - tensor var_564 = linear(bias = encoder_k_proj_2_bias, weight = encoder_k_proj_2_weight, x = x_13)[name = tensor("linear_22")]; - tensor var_565 = const()[name = tensor("op_565"), val = tensor(0x1p-3)]; - tensor var_566 = mul(x = var_564, y = var_565)[name = tensor("op_566")]; - tensor var_567 = const()[name = tensor("op_567"), val = tensor([1, 2, 4, 64])]; - tensor var_568 = reshape(shape = var_567, x = var_566)[name = tensor("op_568")]; + tensor var_600 = linear(bias = inner_encoder_k_proj_2_bias, weight = inner_encoder_k_proj_2_weight, x = x_13)[name = tensor("linear_22")]; + tensor var_601 = const()[name = tensor("op_601"), val = tensor(0x1p-3)]; + tensor var_602 = mul(x = var_600, y = var_601)[name = tensor("op_602")]; + tensor var_603 = const()[name = tensor("op_603"), val = tensor([1, 2, 4, 64])]; + tensor var_604 = reshape(shape = var_603, x = var_602)[name = tensor("op_604")]; tensor k_5_perm_0 = const()[name = tensor("k_5_perm_0"), val = tensor([0, 2, 1, 3])]; - tensor var_572 = linear(bias = encoder_v_proj_2_bias, weight = encoder_v_proj_2_weight, x = x_13)[name = tensor("linear_23")]; - tensor var_573 = const()[name = tensor("op_573"), val = tensor([1, 2, 4, 64])]; - tensor var_574 = reshape(shape = var_573, x = var_572)[name = tensor("op_574")]; + tensor var_608 = linear(bias = inner_encoder_v_proj_2_bias, weight = inner_encoder_v_proj_2_weight, x = x_13)[name = tensor("linear_23")]; + tensor var_609 = const()[name = tensor("op_609"), val = tensor([1, 2, 4, 64])]; + tensor var_610 = reshape(shape = var_609, x = var_608)[name = tensor("op_610")]; tensor v_5_perm_0 = const()[name = tensor("v_5_perm_0"), val = tensor([0, 2, 1, 3])]; - tensor input_97 = linear(bias = encoder_g_proj_2_bias, weight = encoder_g_proj_2_weight, x = x_13)[name = tensor("linear_24")]; + tensor input_99 = linear(bias = inner_encoder_g_proj_2_bias, weight = inner_encoder_g_proj_2_weight, x = x_13)[name = tensor("linear_24")]; tensor sqrt_s0_5 = sqrt(x = prev_scale_5)[name = tensor("sqrt_s0_5")]; - tensor s_t_5 = add(x = prev_scale_5, y = encoder__t_index)[name = tensor("s_t_5")]; + tensor s_t_5 = add(x = prev_scale_5, y = inner_encoder__t_index)[name = tensor("s_t_5")]; tensor sqrt_s_t_5 = sqrt(x = s_t_5)[name = tensor("sqrt_s_t_5")]; tensor qk_5_transpose_x_1 = const()[name = tensor("qk_5_transpose_x_1"), val = tensor(false)]; tensor qk_5_transpose_y_1 = const()[name = tensor("qk_5_transpose_y_1"), val = tensor(true)]; - tensor k_5 = transpose(perm = k_5_perm_0, x = var_568)[name = tensor("transpose_43")]; - tensor q_5 = transpose(perm = q_5_perm_0, x = var_560)[name = tensor("transpose_44")]; + tensor k_5 = transpose(perm = k_5_perm_0, x = var_604)[name = tensor("transpose_43")]; + tensor q_5 = transpose(perm = q_5_perm_0, x = var_596)[name = tensor("transpose_44")]; tensor qk_5 = matmul(transpose_x = qk_5_transpose_x_1, transpose_y = qk_5_transpose_y_1, x = q_5, y = k_5)[name = tensor("qk_5")]; - tensor var_584 = const()[name = tensor("op_584"), val = tensor([2, 1])]; - tensor var_585 = reshape(shape = var_584, x = sqrt_s_t_5)[name = tensor("op_585")]; - tensor M_5 = real_div(x = encoder__causal_mask, y = var_585)[name = tensor("M_5")]; - tensor var_587 = mul(x = qk_5, y = M_5)[name = tensor("op_587")]; + tensor var_620 = const()[name = tensor("op_620"), val = tensor([2, 1])]; + tensor var_621 = reshape(shape = var_620, x = sqrt_s_t_5)[name = tensor("op_621")]; + tensor M_5 = real_div(x = inner_encoder__causal_mask, y = var_621)[name = tensor("M_5")]; + tensor var_623 = mul(x = qk_5, y = M_5)[name = tensor("op_623")]; tensor inner_5_transpose_x_0 = const()[name = tensor("inner_5_transpose_x_0"), val = tensor(false)]; tensor inner_5_transpose_y_0 = const()[name = tensor("inner_5_transpose_y_0"), val = tensor(false)]; - tensor v_5 = transpose(perm = v_5_perm_0, x = var_574)[name = tensor("transpose_42")]; - tensor inner_5 = matmul(transpose_x = inner_5_transpose_x_0, transpose_y = inner_5_transpose_y_0, x = var_587, y = v_5)[name = tensor("inner_5")]; - tensor var_589_transpose_x_0 = const()[name = tensor("op_589_transpose_x_0"), val = tensor(false)]; - tensor var_589_transpose_y_0 = const()[name = tensor("op_589_transpose_y_0"), val = tensor(false)]; - tensor var_589 = matmul(transpose_x = var_589_transpose_x_0, transpose_y = var_589_transpose_y_0, x = q_5, y = prev_kv_5)[name = tensor("op_589")]; - tensor var_590 = real_div(x = sqrt_s0_5, y = sqrt_s_t_5)[name = tensor("op_590")]; - tensor var_591 = const()[name = tensor("op_591"), val = tensor([1, 1, 2, 1])]; - tensor var_592 = reshape(shape = var_591, x = var_590)[name = tensor("op_592")]; - tensor cross_5 = mul(x = var_589, y = var_592)[name = tensor("cross_5")]; + tensor v_5 = transpose(perm = v_5_perm_0, x = var_610)[name = tensor("transpose_42")]; + tensor inner_5 = matmul(transpose_x = inner_5_transpose_x_0, transpose_y = inner_5_transpose_y_0, x = var_623, y = v_5)[name = tensor("inner_5")]; + tensor var_625_transpose_x_0 = const()[name = tensor("op_625_transpose_x_0"), val = tensor(false)]; + tensor var_625_transpose_y_0 = const()[name = tensor("op_625_transpose_y_0"), val = tensor(false)]; + tensor var_625 = matmul(transpose_x = var_625_transpose_x_0, transpose_y = var_625_transpose_y_0, x = q_5, y = prev_kv_5)[name = tensor("op_625")]; + tensor var_626 = real_div(x = sqrt_s0_5, y = sqrt_s_t_5)[name = tensor("op_626")]; + tensor var_627 = const()[name = tensor("op_627"), val = tensor([1, 1, 2, 1])]; + tensor var_628 = reshape(shape = var_627, x = var_626)[name = tensor("op_628")]; + tensor cross_5 = mul(x = var_625, y = var_628)[name = tensor("cross_5")]; tensor out_13 = add(x = inner_5, y = cross_5)[name = tensor("out_13")]; - tensor var_595 = mul(x = prev_kv_5, y = sqrt_s0_5)[name = tensor("op_595")]; - tensor var_597_transpose_x_1 = const()[name = tensor("op_597_transpose_x_1"), val = tensor(true)]; - tensor var_597_transpose_y_1 = const()[name = tensor("op_597_transpose_y_1"), val = tensor(false)]; - tensor var_597 = matmul(transpose_x = var_597_transpose_x_1, transpose_y = var_597_transpose_y_1, x = k_5, y = v_5)[name = tensor("op_597")]; - tensor new_kv_unnorm_5 = add(x = var_595, y = var_597)[name = tensor("new_kv_unnorm_5")]; - tensor var_599 = const()[name = tensor("op_599"), val = tensor(0x1p+1)]; - tensor new_scale_5 = add(x = prev_scale_5, y = var_599)[name = tensor("new_scale_5")]; - tensor var_601 = sqrt(x = new_scale_5)[name = tensor("op_601")]; - tensor var_602 = real_div(x = new_kv_unnorm_5, y = var_601)[name = tensor("op_602")]; - tensor var_603_perm_0 = const()[name = tensor("op_603_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_631 = mul(x = prev_kv_5, y = sqrt_s0_5)[name = tensor("op_631")]; + tensor var_633_transpose_x_1 = const()[name = tensor("op_633_transpose_x_1"), val = tensor(true)]; + tensor var_633_transpose_y_1 = const()[name = tensor("op_633_transpose_y_1"), val = tensor(false)]; + tensor var_633 = matmul(transpose_x = var_633_transpose_x_1, transpose_y = var_633_transpose_y_1, x = k_5, y = v_5)[name = tensor("op_633")]; + tensor new_kv_unnorm_5 = add(x = var_631, y = var_633)[name = tensor("new_kv_unnorm_5")]; + tensor var_635 = const()[name = tensor("op_635"), val = tensor(0x1p+1)]; + tensor new_scale_5 = add(x = prev_scale_5, y = var_635)[name = tensor("new_scale_5")]; + tensor var_637 = sqrt(x = new_scale_5)[name = tensor("op_637")]; + tensor var_638 = real_div(x = new_kv_unnorm_5, y = var_637)[name = tensor("op_638")]; + tensor var_639_perm_0 = const()[name = tensor("op_639_perm_0"), val = tensor([0, 2, 1, 3])]; tensor out_15_axes_0 = const()[name = tensor("out_15_axes_0"), val = tensor([-1])]; - tensor var_603 = transpose(perm = var_603_perm_0, x = out_13)[name = tensor("transpose_41")]; - tensor out_15 = layer_norm(axes = out_15_axes_0, epsilon = var_17, x = var_603)[name = tensor("out_15")]; - tensor var_607 = const()[name = tensor("op_607"), val = tensor([1, 2, 256])]; - tensor out_17 = reshape(shape = var_607, x = out_15)[name = tensor("out_17")]; - tensor var_609 = silu(x = input_97)[name = tensor("op_609")]; - tensor input_99 = mul(x = var_609, y = out_17)[name = tensor("input_99")]; - tensor ret_out_5 = linear(bias = encoder_out_proj_2_bias, weight = encoder_out_proj_2_weight, x = input_99)[name = tensor("linear_25")]; - tensor x_15 = add(x = input_93, y = ret_out_5)[name = tensor("x_15")]; + tensor var_639 = transpose(perm = var_639_perm_0, x = out_13)[name = tensor("transpose_41")]; + tensor out_15 = layer_norm(axes = out_15_axes_0, epsilon = var_54, x = var_639)[name = tensor("out_15")]; + tensor var_643 = const()[name = tensor("op_643"), val = tensor([1, 2, 256])]; + tensor out_17 = reshape(shape = var_643, x = out_15)[name = tensor("out_17")]; + tensor var_645 = silu(x = input_99)[name = tensor("op_645")]; + tensor input_101 = mul(x = var_645, y = out_17)[name = tensor("input_101")]; + tensor ret_out_5 = linear(bias = inner_encoder_out_proj_2_bias, weight = inner_encoder_out_proj_2_weight, x = input_101)[name = tensor("linear_25")]; + tensor x_15 = add(x = input_95, y = ret_out_5)[name = tensor("x_15")]; tensor window_13_begin_0 = const()[name = tensor("window_13_begin_0"), val = tensor([2, 0, 0, 0])]; tensor window_13_end_0 = const()[name = tensor("window_13_end_0"), val = tensor([3, 1, 16, 256])]; tensor window_13_end_mask_0 = const()[name = tensor("window_13_end_mask_0"), val = tensor([false, true, true, true])]; tensor window_13_squeeze_mask_0 = const()[name = tensor("window_13_squeeze_mask_0"), val = tensor([true, false, false, false])]; tensor window_13 = slice_by_index(begin = window_13_begin_0, end = window_13_end_0, end_mask = window_13_end_mask_0, squeeze_mask = window_13_squeeze_mask_0, x = enc_conv_cache)[name = tensor("window_13")]; - tensor var_617_begin_0 = const()[name = tensor("op_617_begin_0"), val = tensor([0, 0, 0])]; - tensor var_617_end_0 = const()[name = tensor("op_617_end_0"), val = tensor([1, 1, 256])]; - tensor var_617_end_mask_0 = const()[name = tensor("op_617_end_mask_0"), val = tensor([true, false, true])]; - tensor var_617 = slice_by_index(begin = var_617_begin_0, end = var_617_end_0, end_mask = var_617_end_mask_0, x = x_15)[name = tensor("op_617")]; - tensor var_620_begin_0 = const()[name = tensor("op_620_begin_0"), val = tensor([0, 1, 0])]; - tensor var_620_end_0 = const()[name = tensor("op_620_end_0"), val = tensor([1, 16, 256])]; - tensor var_620_end_mask_0 = const()[name = tensor("op_620_end_mask_0"), val = tensor([true, true, true])]; - tensor var_620 = slice_by_index(begin = var_620_begin_0, end = var_620_end_0, end_mask = var_620_end_mask_0, x = window_13)[name = tensor("op_620")]; + tensor var_653_begin_0 = const()[name = tensor("op_653_begin_0"), val = tensor([0, 0, 0])]; + tensor var_653_end_0 = const()[name = tensor("op_653_end_0"), val = tensor([1, 1, 256])]; + tensor var_653_end_mask_0 = const()[name = tensor("op_653_end_mask_0"), val = tensor([true, false, true])]; + tensor var_653 = slice_by_index(begin = var_653_begin_0, end = var_653_end_0, end_mask = var_653_end_mask_0, x = x_15)[name = tensor("op_653")]; + tensor var_656_begin_0 = const()[name = tensor("op_656_begin_0"), val = tensor([0, 1, 0])]; + tensor var_656_end_0 = const()[name = tensor("op_656_end_0"), val = tensor([1, 16, 256])]; + tensor var_656_end_mask_0 = const()[name = tensor("op_656_end_mask_0"), val = tensor([true, true, true])]; + tensor var_656 = slice_by_index(begin = var_656_begin_0, end = var_656_end_0, end_mask = var_656_end_mask_0, x = window_13)[name = tensor("op_656")]; tensor window_15_interleave_0 = const()[name = tensor("window_15_interleave_0"), val = tensor(false)]; - tensor window_15 = concat(axis = var_26, interleave = window_15_interleave_0, values = (var_620, var_617))[name = tensor("window_15")]; - tensor var_625_begin_0 = const()[name = tensor("op_625_begin_0"), val = tensor([0, 1, 0])]; - tensor var_625_end_0 = const()[name = tensor("op_625_end_0"), val = tensor([1, 1, 256])]; - tensor var_625_end_mask_0 = const()[name = tensor("op_625_end_mask_0"), val = tensor([true, true, true])]; - tensor var_625 = slice_by_index(begin = var_625_begin_0, end = var_625_end_0, end_mask = var_625_end_mask_0, x = x_15)[name = tensor("op_625")]; - tensor var_628_begin_0 = const()[name = tensor("op_628_begin_0"), val = tensor([0, 1, 0])]; - tensor var_628_end_0 = const()[name = tensor("op_628_end_0"), val = tensor([1, 16, 256])]; - tensor var_628_end_mask_0 = const()[name = tensor("op_628_end_mask_0"), val = tensor([true, true, true])]; - tensor var_628 = slice_by_index(begin = var_628_begin_0, end = var_628_end_0, end_mask = var_628_end_mask_0, x = window_15)[name = tensor("op_628")]; + tensor window_15 = concat(axis = var_62, interleave = window_15_interleave_0, values = (var_656, var_653))[name = tensor("window_15")]; + tensor var_661_begin_0 = const()[name = tensor("op_661_begin_0"), val = tensor([0, 1, 0])]; + tensor var_661_end_0 = const()[name = tensor("op_661_end_0"), val = tensor([1, 1, 256])]; + tensor var_661_end_mask_0 = const()[name = tensor("op_661_end_mask_0"), val = tensor([true, true, true])]; + tensor var_661 = slice_by_index(begin = var_661_begin_0, end = var_661_end_0, end_mask = var_661_end_mask_0, x = x_15)[name = tensor("op_661")]; + tensor var_664_begin_0 = const()[name = tensor("op_664_begin_0"), val = tensor([0, 1, 0])]; + tensor var_664_end_0 = const()[name = tensor("op_664_end_0"), val = tensor([1, 16, 256])]; + tensor var_664_end_mask_0 = const()[name = tensor("op_664_end_mask_0"), val = tensor([true, true, true])]; + tensor var_664 = slice_by_index(begin = var_664_begin_0, end = var_664_end_0, end_mask = var_664_end_mask_0, x = window_15)[name = tensor("op_664")]; tensor window_17_interleave_0 = const()[name = tensor("window_17_interleave_0"), val = tensor(false)]; - tensor window_17 = concat(axis = var_26, interleave = window_17_interleave_0, values = (var_628, var_625))[name = tensor("window_17")]; - tensor input_101_interleave_0 = const()[name = tensor("input_101_interleave_0"), val = tensor(false)]; - tensor input_101 = concat(axis = var_23, interleave = input_101_interleave_0, values = (window_15, window_17))[name = tensor("input_101")]; + tensor window_17 = concat(axis = var_62, interleave = window_17_interleave_0, values = (var_664, var_661))[name = tensor("window_17")]; + tensor input_103_interleave_0 = const()[name = tensor("input_103_interleave_0"), val = tensor(false)]; + tensor input_103 = concat(axis = var_49, interleave = input_103_interleave_0, values = (window_15, window_17))[name = tensor("input_103")]; tensor x_17_axes_0 = const()[name = tensor("x_17_axes_0"), val = tensor([-1])]; - tensor x_17 = layer_norm(axes = x_17_axes_0, beta = encoder_conv_module_2_sequential_0_bias, epsilon = var_28, gamma = encoder_conv_module_2_sequential_0_weight, x = input_101)[name = tensor("x_17")]; - tensor input_103_perm_0 = const()[name = tensor("input_103_perm_0"), val = tensor([0, 2, 1])]; + tensor x_17 = layer_norm(axes = x_17_axes_0, beta = inner_encoder_conv_module_2_sequential_0_bias, epsilon = var_46, gamma = inner_encoder_conv_module_2_sequential_0_weight, x = input_103)[name = tensor("x_17")]; + tensor input_105_perm_0 = const()[name = tensor("input_105_perm_0"), val = tensor([0, 2, 1])]; tensor inputs_23_pad_type_0 = const()[name = tensor("inputs_23_pad_type_0"), val = tensor("valid")]; tensor inputs_23_strides_0 = const()[name = tensor("inputs_23_strides_0"), val = tensor([1])]; tensor inputs_23_pad_0 = const()[name = tensor("inputs_23_pad_0"), val = tensor([0, 0])]; tensor inputs_23_dilations_0 = const()[name = tensor("inputs_23_dilations_0"), val = tensor([1])]; tensor inputs_23_groups_0 = const()[name = tensor("inputs_23_groups_0"), val = tensor(1)]; - tensor input_103 = transpose(perm = input_103_perm_0, x = x_17)[name = tensor("transpose_40")]; - tensor inputs_23 = conv(bias = encoder_conv_module_2_sequential_2_conv_bias, dilations = inputs_23_dilations_0, groups = inputs_23_groups_0, pad = inputs_23_pad_0, pad_type = inputs_23_pad_type_0, strides = inputs_23_strides_0, weight = encoder_conv_module_2_sequential_2_conv_weight, x = input_103)[name = tensor("inputs_23")]; - tensor var_653_split_sizes_0 = const()[name = tensor("op_653_split_sizes_0"), val = tensor([256, 256])]; - tensor var_653_axis_0 = const()[name = tensor("op_653_axis_0"), val = tensor(1)]; - tensor var_653_0, tensor var_653_1 = split(axis = var_653_axis_0, split_sizes = var_653_split_sizes_0, x = inputs_23)[name = tensor("op_653")]; - tensor var_655 = sigmoid(x = var_653_1)[name = tensor("op_655")]; - tensor inputs_25 = mul(x = var_653_0, y = var_655)[name = tensor("inputs_25")]; + tensor input_105 = transpose(perm = input_105_perm_0, x = x_17)[name = tensor("transpose_40")]; + tensor inputs_23 = conv(bias = inner_encoder_conv_module_2_sequential_2_conv_bias, dilations = inputs_23_dilations_0, groups = inputs_23_groups_0, pad = inputs_23_pad_0, pad_type = inputs_23_pad_type_0, strides = inputs_23_strides_0, weight = inner_encoder_conv_module_2_sequential_2_conv_weight, x = input_105)[name = tensor("inputs_23")]; + tensor var_689_split_sizes_0 = const()[name = tensor("op_689_split_sizes_0"), val = tensor([256, 256])]; + tensor var_689_axis_0 = const()[name = tensor("op_689_axis_0"), val = tensor(1)]; + tensor var_689_0, tensor var_689_1 = split(axis = var_689_axis_0, split_sizes = var_689_split_sizes_0, x = inputs_23)[name = tensor("op_689")]; + tensor var_691 = sigmoid(x = var_689_1)[name = tensor("op_691")]; + tensor inputs_25 = mul(x = var_689_0, y = var_691)[name = tensor("inputs_25")]; tensor outputs_aug_5_pad_type_0 = const()[name = tensor("outputs_aug_5_pad_type_0"), val = tensor("custom")]; tensor outputs_aug_5_pad_0 = const()[name = tensor("outputs_aug_5_pad_0"), val = tensor([15, 15])]; tensor outputs_aug_5_groups_0 = const()[name = tensor("outputs_aug_5_groups_0"), val = tensor(256)]; tensor outputs_aug_5_strides_0 = const()[name = tensor("outputs_aug_5_strides_0"), val = tensor([1])]; tensor outputs_aug_5_dilations_0 = const()[name = tensor("outputs_aug_5_dilations_0"), val = tensor([1])]; - tensor outputs_aug_5 = conv(dilations = outputs_aug_5_dilations_0, groups = outputs_aug_5_groups_0, pad = outputs_aug_5_pad_0, pad_type = outputs_aug_5_pad_type_0, strides = outputs_aug_5_strides_0, weight = encoder_conv_module_2_sequential_4_conv_weight, x = inputs_25)[name = tensor("outputs_aug_5")]; - tensor input_105_begin_0 = const()[name = tensor("input_105_begin_0"), val = tensor([0, 0, 0])]; - tensor input_105_end_0 = const()[name = tensor("input_105_end_0"), val = tensor([2, 256, 16])]; - tensor input_105_end_mask_0 = const()[name = tensor("input_105_end_mask_0"), val = tensor([true, true, false])]; - tensor input_105 = slice_by_index(begin = input_105_begin_0, end = input_105_end_0, end_mask = input_105_end_mask_0, x = outputs_aug_5)[name = tensor("input_105")]; - tensor inputs_27 = batch_norm(beta = encoder_conv_module_2_sequential_5_bias, epsilon = var_28, gamma = encoder_conv_module_2_sequential_5_weight, mean = encoder_conv_module_2_sequential_5_running_mean, variance = encoder_conv_module_2_sequential_5_running_var, x = input_105)[name = tensor("inputs_27")]; - tensor input_107 = silu(x = inputs_27)[name = tensor("input_107")]; - tensor input_109_pad_type_0 = const()[name = tensor("input_109_pad_type_0"), val = tensor("valid")]; - tensor input_109_strides_0 = const()[name = tensor("input_109_strides_0"), val = tensor([1])]; - tensor input_109_pad_0 = const()[name = tensor("input_109_pad_0"), val = tensor([0, 0])]; - tensor input_109_dilations_0 = const()[name = tensor("input_109_dilations_0"), val = tensor([1])]; - tensor input_109_groups_0 = const()[name = tensor("input_109_groups_0"), val = tensor(1)]; - tensor input_109 = conv(bias = encoder_conv_module_2_sequential_7_conv_bias, dilations = input_109_dilations_0, groups = input_109_groups_0, pad = input_109_pad_0, pad_type = input_109_pad_type_0, strides = input_109_strides_0, weight = encoder_conv_module_2_sequential_7_conv_weight, x = input_107)[name = tensor("input_109")]; + tensor outputs_aug_5 = conv(dilations = outputs_aug_5_dilations_0, groups = outputs_aug_5_groups_0, pad = outputs_aug_5_pad_0, pad_type = outputs_aug_5_pad_type_0, strides = outputs_aug_5_strides_0, weight = inner_encoder_conv_module_2_sequential_4_conv_weight, x = inputs_25)[name = tensor("outputs_aug_5")]; + tensor input_107_begin_0 = const()[name = tensor("input_107_begin_0"), val = tensor([0, 0, 0])]; + tensor input_107_end_0 = const()[name = tensor("input_107_end_0"), val = tensor([2, 256, 16])]; + tensor input_107_end_mask_0 = const()[name = tensor("input_107_end_mask_0"), val = tensor([true, true, false])]; + tensor input_107 = slice_by_index(begin = input_107_begin_0, end = input_107_end_0, end_mask = input_107_end_mask_0, x = outputs_aug_5)[name = tensor("input_107")]; + tensor inputs_27 = batch_norm(beta = inner_encoder_conv_module_2_sequential_5_bias, epsilon = var_46, gamma = inner_encoder_conv_module_2_sequential_5_weight, mean = inner_encoder_conv_module_2_sequential_5_running_mean, variance = inner_encoder_conv_module_2_sequential_5_running_var, x = input_107)[name = tensor("inputs_27")]; + tensor input_109 = silu(x = inputs_27)[name = tensor("input_109")]; + tensor input_111_pad_type_0 = const()[name = tensor("input_111_pad_type_0"), val = tensor("valid")]; + tensor input_111_strides_0 = const()[name = tensor("input_111_strides_0"), val = tensor([1])]; + tensor input_111_pad_0 = const()[name = tensor("input_111_pad_0"), val = tensor([0, 0])]; + tensor input_111_dilations_0 = const()[name = tensor("input_111_dilations_0"), val = tensor([1])]; + tensor input_111_groups_0 = const()[name = tensor("input_111_groups_0"), val = tensor(1)]; + tensor input_111 = conv(bias = inner_encoder_conv_module_2_sequential_7_conv_bias, dilations = input_111_dilations_0, groups = input_111_groups_0, pad = input_111_pad_0, pad_type = input_111_pad_type_0, strides = input_111_strides_0, weight = inner_encoder_conv_module_2_sequential_7_conv_weight, x = input_109)[name = tensor("input_111")]; tensor conv_out_5_perm_0 = const()[name = tensor("conv_out_5_perm_0"), val = tensor([0, 2, 1])]; - tensor var_686_begin_0 = const()[name = tensor("op_686_begin_0"), val = tensor([0, -1, 0])]; - tensor var_686_end_0 = const()[name = tensor("op_686_end_0"), val = tensor([2, 16, 256])]; - tensor var_686_end_mask_0 = const()[name = tensor("op_686_end_mask_0"), val = tensor([true, true, true])]; - tensor conv_out_5 = transpose(perm = conv_out_5_perm_0, x = input_109)[name = tensor("transpose_39")]; - tensor var_686 = slice_by_index(begin = var_686_begin_0, end = var_686_end_0, end_mask = var_686_end_mask_0, x = conv_out_5)[name = tensor("op_686")]; - tensor var_688_perm_0 = const()[name = tensor("op_688_perm_0"), val = tensor([1, 0, 2])]; - tensor var_688 = transpose(perm = var_688_perm_0, x = var_686)[name = tensor("transpose_38")]; - tensor input_111 = add(x = x_15, y = var_688)[name = tensor("input_111")]; - tensor input_113_axes_0 = const()[name = tensor("input_113_axes_0"), val = tensor([-1])]; - tensor input_113 = layer_norm(axes = input_113_axes_0, beta = encoder_ffn2_2_module_sequential_0_bias, epsilon = var_28, gamma = encoder_ffn2_2_module_sequential_0_weight, x = input_111)[name = tensor("input_113")]; - tensor inputs_29 = linear(bias = encoder_ffn2_2_module_sequential_1_linear_bias, weight = encoder_ffn2_2_module_sequential_1_linear_weight, x = input_113)[name = tensor("linear_26")]; - tensor input_115 = silu(x = inputs_29)[name = tensor("input_115")]; - tensor input_119 = linear(bias = encoder_ffn2_2_module_sequential_4_linear_bias, weight = encoder_ffn2_2_module_sequential_4_linear_weight, x = input_115)[name = tensor("linear_27")]; - tensor var_711 = const()[name = tensor("op_711"), val = tensor(0x1p-1)]; - tensor var_712 = mul(x = input_119, y = var_711)[name = tensor("op_712")]; - tensor input_121 = add(x = var_712, y = input_111)[name = tensor("input_121")]; - tensor input_123_axes_0 = const()[name = tensor("input_123_axes_0"), val = tensor([-1])]; - tensor input_123 = layer_norm(axes = input_123_axes_0, beta = encoder_layer_norm_2_bias, epsilon = var_28, gamma = encoder_layer_norm_2_weight, x = input_121)[name = tensor("input_123")]; + tensor var_722_begin_0 = const()[name = tensor("op_722_begin_0"), val = tensor([0, -1, 0])]; + tensor var_722_end_0 = const()[name = tensor("op_722_end_0"), val = tensor([2, 16, 256])]; + tensor var_722_end_mask_0 = const()[name = tensor("op_722_end_mask_0"), val = tensor([true, true, true])]; + tensor conv_out_5 = transpose(perm = conv_out_5_perm_0, x = input_111)[name = tensor("transpose_39")]; + tensor var_722 = slice_by_index(begin = var_722_begin_0, end = var_722_end_0, end_mask = var_722_end_mask_0, x = conv_out_5)[name = tensor("op_722")]; + tensor var_724_perm_0 = const()[name = tensor("op_724_perm_0"), val = tensor([1, 0, 2])]; + tensor var_724 = transpose(perm = var_724_perm_0, x = var_722)[name = tensor("transpose_38")]; + tensor input_113 = add(x = x_15, y = var_724)[name = tensor("input_113")]; + tensor input_115_axes_0 = const()[name = tensor("input_115_axes_0"), val = tensor([-1])]; + tensor input_115 = layer_norm(axes = input_115_axes_0, beta = inner_encoder_ffn2_2_module_sequential_0_bias, epsilon = var_46, gamma = inner_encoder_ffn2_2_module_sequential_0_weight, x = input_113)[name = tensor("input_115")]; + tensor inputs_29 = linear(bias = inner_encoder_ffn2_2_module_sequential_1_linear_bias, weight = inner_encoder_ffn2_2_module_sequential_1_linear_weight, x = input_115)[name = tensor("linear_26")]; + tensor input_117 = silu(x = inputs_29)[name = tensor("input_117")]; + tensor input_121 = linear(bias = inner_encoder_ffn2_2_module_sequential_4_linear_bias, weight = inner_encoder_ffn2_2_module_sequential_4_linear_weight, x = input_117)[name = tensor("linear_27")]; + tensor var_747 = const()[name = tensor("op_747"), val = tensor(0x1p-1)]; + tensor var_748 = mul(x = input_121, y = var_747)[name = tensor("op_748")]; + tensor input_123 = add(x = var_748, y = input_113)[name = tensor("input_123")]; tensor input_125_axes_0 = const()[name = tensor("input_125_axes_0"), val = tensor([-1])]; - tensor input_125 = layer_norm(axes = input_125_axes_0, beta = encoder_ffn1_3_module_sequential_0_bias, epsilon = var_28, gamma = encoder_ffn1_3_module_sequential_0_weight, x = input_123)[name = tensor("input_125")]; - tensor inputs_31 = linear(bias = encoder_ffn1_3_module_sequential_1_linear_bias, weight = encoder_ffn1_3_module_sequential_1_linear_weight, x = input_125)[name = tensor("linear_28")]; - tensor input_127 = silu(x = inputs_31)[name = tensor("input_127")]; - tensor input_131 = linear(bias = encoder_ffn1_3_module_sequential_4_linear_bias, weight = encoder_ffn1_3_module_sequential_4_linear_weight, x = input_127)[name = tensor("linear_29")]; - tensor var_741 = const()[name = tensor("op_741"), val = tensor(0x1p-1)]; - tensor var_742 = mul(x = input_131, y = var_741)[name = tensor("op_742")]; - tensor input_133 = add(x = var_742, y = input_123)[name = tensor("input_133")]; + tensor input_125 = layer_norm(axes = input_125_axes_0, beta = inner_encoder_layer_norm_2_bias, epsilon = var_46, gamma = inner_encoder_layer_norm_2_weight, x = input_123)[name = tensor("input_125")]; + tensor input_127_axes_0 = const()[name = tensor("input_127_axes_0"), val = tensor([-1])]; + tensor input_127 = layer_norm(axes = input_127_axes_0, beta = inner_encoder_ffn1_3_module_sequential_0_bias, epsilon = var_46, gamma = inner_encoder_ffn1_3_module_sequential_0_weight, x = input_125)[name = tensor("input_127")]; + tensor inputs_31 = linear(bias = inner_encoder_ffn1_3_module_sequential_1_linear_bias, weight = inner_encoder_ffn1_3_module_sequential_1_linear_weight, x = input_127)[name = tensor("linear_28")]; + tensor input_129 = silu(x = inputs_31)[name = tensor("input_129")]; + tensor input_133 = linear(bias = inner_encoder_ffn1_3_module_sequential_4_linear_bias, weight = inner_encoder_ffn1_3_module_sequential_4_linear_weight, x = input_129)[name = tensor("linear_29")]; + tensor var_777 = const()[name = tensor("op_777"), val = tensor(0x1p-1)]; + tensor var_778 = mul(x = input_133, y = var_777)[name = tensor("op_778")]; + tensor input_135 = add(x = var_778, y = input_125)[name = tensor("input_135")]; tensor x_19_axes_0 = const()[name = tensor("x_19_axes_0"), val = tensor([-1])]; - tensor x_19 = layer_norm(axes = x_19_axes_0, beta = encoder_ret_lns_3_bias, epsilon = var_28, gamma = encoder_ret_lns_3_weight, x = input_133)[name = tensor("x_19")]; + tensor x_19 = layer_norm(axes = x_19_axes_0, beta = inner_encoder_ret_lns_3_bias, epsilon = var_46, gamma = inner_encoder_ret_lns_3_weight, x = input_135)[name = tensor("x_19")]; tensor prev_kv_7_begin_0 = const()[name = tensor("prev_kv_7_begin_0"), val = tensor([3, 0, 0, 0, 0])]; tensor prev_kv_7_end_0 = const()[name = tensor("prev_kv_7_end_0"), val = tensor([4, 1, 4, 64, 64])]; tensor prev_kv_7_end_mask_0 = const()[name = tensor("prev_kv_7_end_mask_0"), val = tensor([false, true, true, true, true])]; @@ -710,189 +724,182 @@ program(1.0) tensor prev_scale_7_end_mask_0 = const()[name = tensor("prev_scale_7_end_mask_0"), val = tensor([false, true])]; tensor prev_scale_7_squeeze_mask_0 = const()[name = tensor("prev_scale_7_squeeze_mask_0"), val = tensor([true, false])]; tensor prev_scale_7 = slice_by_index(begin = prev_scale_7_begin_0, end = prev_scale_7_end_0, end_mask = prev_scale_7_end_mask_0, squeeze_mask = prev_scale_7_squeeze_mask_0, x = enc_scale)[name = tensor("prev_scale_7")]; - tensor var_756 = linear(bias = encoder_q_proj_3_bias, weight = encoder_q_proj_3_weight, x = x_19)[name = tensor("linear_30")]; - tensor var_757 = const()[name = tensor("op_757"), val = tensor([1, 2, 4, 64])]; - tensor var_758 = reshape(shape = var_757, x = var_756)[name = tensor("op_758")]; + tensor var_792 = linear(bias = inner_encoder_q_proj_3_bias, weight = inner_encoder_q_proj_3_weight, x = x_19)[name = tensor("linear_30")]; + tensor var_793 = const()[name = tensor("op_793"), val = tensor([1, 2, 4, 64])]; + tensor var_794 = reshape(shape = var_793, x = var_792)[name = tensor("op_794")]; tensor q_7_perm_0 = const()[name = tensor("q_7_perm_0"), val = tensor([0, 2, 1, 3])]; - tensor var_762 = linear(bias = encoder_k_proj_3_bias, weight = encoder_k_proj_3_weight, x = x_19)[name = tensor("linear_31")]; - tensor var_763 = const()[name = tensor("op_763"), val = tensor(0x1p-3)]; - tensor var_764 = mul(x = var_762, y = var_763)[name = tensor("op_764")]; - tensor var_765 = const()[name = tensor("op_765"), val = tensor([1, 2, 4, 64])]; - tensor var_766 = reshape(shape = var_765, x = var_764)[name = tensor("op_766")]; + tensor var_798 = linear(bias = inner_encoder_k_proj_3_bias, weight = inner_encoder_k_proj_3_weight, x = x_19)[name = tensor("linear_31")]; + tensor var_799 = const()[name = tensor("op_799"), val = tensor(0x1p-3)]; + tensor var_800 = mul(x = var_798, y = var_799)[name = tensor("op_800")]; + tensor var_801 = const()[name = tensor("op_801"), val = tensor([1, 2, 4, 64])]; + tensor var_802 = reshape(shape = var_801, x = var_800)[name = tensor("op_802")]; tensor k_7_perm_0 = const()[name = tensor("k_7_perm_0"), val = tensor([0, 2, 1, 3])]; - tensor var_770 = linear(bias = encoder_v_proj_3_bias, weight = encoder_v_proj_3_weight, x = x_19)[name = tensor("linear_32")]; - tensor var_771 = const()[name = tensor("op_771"), val = tensor([1, 2, 4, 64])]; - tensor var_772 = reshape(shape = var_771, x = var_770)[name = tensor("op_772")]; + tensor var_806 = linear(bias = inner_encoder_v_proj_3_bias, weight = inner_encoder_v_proj_3_weight, x = x_19)[name = tensor("linear_32")]; + tensor var_807 = const()[name = tensor("op_807"), val = tensor([1, 2, 4, 64])]; + tensor var_808 = reshape(shape = var_807, x = var_806)[name = tensor("op_808")]; tensor v_7_perm_0 = const()[name = tensor("v_7_perm_0"), val = tensor([0, 2, 1, 3])]; - tensor input_137 = linear(bias = encoder_g_proj_3_bias, weight = encoder_g_proj_3_weight, x = x_19)[name = tensor("linear_33")]; + tensor input_139 = linear(bias = inner_encoder_g_proj_3_bias, weight = inner_encoder_g_proj_3_weight, x = x_19)[name = tensor("linear_33")]; tensor sqrt_s0_7 = sqrt(x = prev_scale_7)[name = tensor("sqrt_s0_7")]; - tensor s_t_7 = add(x = prev_scale_7, y = encoder__t_index)[name = tensor("s_t_7")]; + tensor s_t_7 = add(x = prev_scale_7, y = inner_encoder__t_index)[name = tensor("s_t_7")]; tensor sqrt_s_t_7 = sqrt(x = s_t_7)[name = tensor("sqrt_s_t_7")]; tensor qk_7_transpose_x_1 = const()[name = tensor("qk_7_transpose_x_1"), val = tensor(false)]; tensor qk_7_transpose_y_1 = const()[name = tensor("qk_7_transpose_y_1"), val = tensor(true)]; - tensor k_7 = transpose(perm = k_7_perm_0, x = var_766)[name = tensor("transpose_36")]; - tensor q_7 = transpose(perm = q_7_perm_0, x = var_758)[name = tensor("transpose_37")]; + tensor k_7 = transpose(perm = k_7_perm_0, x = var_802)[name = tensor("transpose_36")]; + tensor q_7 = transpose(perm = q_7_perm_0, x = var_794)[name = tensor("transpose_37")]; tensor qk_7 = matmul(transpose_x = qk_7_transpose_x_1, transpose_y = qk_7_transpose_y_1, x = q_7, y = k_7)[name = tensor("qk_7")]; - tensor var_782 = const()[name = tensor("op_782"), val = tensor([2, 1])]; - tensor var_783 = reshape(shape = var_782, x = sqrt_s_t_7)[name = tensor("op_783")]; - tensor M_7 = real_div(x = encoder__causal_mask, y = var_783)[name = tensor("M_7")]; - tensor var_785 = mul(x = qk_7, y = M_7)[name = tensor("op_785")]; + tensor var_818 = const()[name = tensor("op_818"), val = tensor([2, 1])]; + tensor var_819 = reshape(shape = var_818, x = sqrt_s_t_7)[name = tensor("op_819")]; + tensor M_7 = real_div(x = inner_encoder__causal_mask, y = var_819)[name = tensor("M_7")]; + tensor var_821 = mul(x = qk_7, y = M_7)[name = tensor("op_821")]; tensor inner_7_transpose_x_0 = const()[name = tensor("inner_7_transpose_x_0"), val = tensor(false)]; tensor inner_7_transpose_y_0 = const()[name = tensor("inner_7_transpose_y_0"), val = tensor(false)]; - tensor v_7 = transpose(perm = v_7_perm_0, x = var_772)[name = tensor("transpose_35")]; - tensor inner_7 = matmul(transpose_x = inner_7_transpose_x_0, transpose_y = inner_7_transpose_y_0, x = var_785, y = v_7)[name = tensor("inner_7")]; - tensor var_787_transpose_x_0 = const()[name = tensor("op_787_transpose_x_0"), val = tensor(false)]; - tensor var_787_transpose_y_0 = const()[name = tensor("op_787_transpose_y_0"), val = tensor(false)]; - tensor var_787 = matmul(transpose_x = var_787_transpose_x_0, transpose_y = var_787_transpose_y_0, x = q_7, y = prev_kv_7)[name = tensor("op_787")]; - tensor var_788 = real_div(x = sqrt_s0_7, y = sqrt_s_t_7)[name = tensor("op_788")]; - tensor var_789 = const()[name = tensor("op_789"), val = tensor([1, 1, 2, 1])]; - tensor var_790 = reshape(shape = var_789, x = var_788)[name = tensor("op_790")]; - tensor cross_7 = mul(x = var_787, y = var_790)[name = tensor("cross_7")]; + tensor v_7 = transpose(perm = v_7_perm_0, x = var_808)[name = tensor("transpose_35")]; + tensor inner_7 = matmul(transpose_x = inner_7_transpose_x_0, transpose_y = inner_7_transpose_y_0, x = var_821, y = v_7)[name = tensor("inner_7")]; + tensor var_823_transpose_x_0 = const()[name = tensor("op_823_transpose_x_0"), val = tensor(false)]; + tensor var_823_transpose_y_0 = const()[name = tensor("op_823_transpose_y_0"), val = tensor(false)]; + tensor var_823 = matmul(transpose_x = var_823_transpose_x_0, transpose_y = var_823_transpose_y_0, x = q_7, y = prev_kv_7)[name = tensor("op_823")]; + tensor var_824 = real_div(x = sqrt_s0_7, y = sqrt_s_t_7)[name = tensor("op_824")]; + tensor var_825 = const()[name = tensor("op_825"), val = tensor([1, 1, 2, 1])]; + tensor var_826 = reshape(shape = var_825, x = var_824)[name = tensor("op_826")]; + tensor cross_7 = mul(x = var_823, y = var_826)[name = tensor("cross_7")]; tensor out_19 = add(x = inner_7, y = cross_7)[name = tensor("out_19")]; - tensor var_793 = mul(x = prev_kv_7, y = sqrt_s0_7)[name = tensor("op_793")]; - tensor var_795_transpose_x_1 = const()[name = tensor("op_795_transpose_x_1"), val = tensor(true)]; - tensor var_795_transpose_y_1 = const()[name = tensor("op_795_transpose_y_1"), val = tensor(false)]; - tensor var_795 = matmul(transpose_x = var_795_transpose_x_1, transpose_y = var_795_transpose_y_1, x = k_7, y = v_7)[name = tensor("op_795")]; - tensor new_kv_unnorm_7 = add(x = var_793, y = var_795)[name = tensor("new_kv_unnorm_7")]; - tensor var_797 = const()[name = tensor("op_797"), val = tensor(0x1p+1)]; - tensor new_scale_7 = add(x = prev_scale_7, y = var_797)[name = tensor("new_scale_7")]; - tensor var_799 = sqrt(x = new_scale_7)[name = tensor("op_799")]; - tensor nkv_1 = real_div(x = new_kv_unnorm_7, y = var_799)[name = tensor("nkv_1")]; - tensor var_801_perm_0 = const()[name = tensor("op_801_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_829 = mul(x = prev_kv_7, y = sqrt_s0_7)[name = tensor("op_829")]; + tensor var_831_transpose_x_1 = const()[name = tensor("op_831_transpose_x_1"), val = tensor(true)]; + tensor var_831_transpose_y_1 = const()[name = tensor("op_831_transpose_y_1"), val = tensor(false)]; + tensor var_831 = matmul(transpose_x = var_831_transpose_x_1, transpose_y = var_831_transpose_y_1, x = k_7, y = v_7)[name = tensor("op_831")]; + tensor new_kv_unnorm_7 = add(x = var_829, y = var_831)[name = tensor("new_kv_unnorm_7")]; + tensor var_833 = const()[name = tensor("op_833"), val = tensor(0x1p+1)]; + tensor new_scale_7 = add(x = prev_scale_7, y = var_833)[name = tensor("new_scale_7")]; + tensor var_835 = sqrt(x = new_scale_7)[name = tensor("op_835")]; + tensor nkv_1 = real_div(x = new_kv_unnorm_7, y = var_835)[name = tensor("nkv_1")]; + tensor var_837_perm_0 = const()[name = tensor("op_837_perm_0"), val = tensor([0, 2, 1, 3])]; tensor out_21_axes_0 = const()[name = tensor("out_21_axes_0"), val = tensor([-1])]; - tensor var_801 = transpose(perm = var_801_perm_0, x = out_19)[name = tensor("transpose_34")]; - tensor out_21 = layer_norm(axes = out_21_axes_0, epsilon = var_17, x = var_801)[name = tensor("out_21")]; - tensor var_805 = const()[name = tensor("op_805"), val = tensor([1, 2, 256])]; - tensor out_23 = reshape(shape = var_805, x = out_21)[name = tensor("out_23")]; - tensor var_807 = silu(x = input_137)[name = tensor("op_807")]; - tensor input_139 = mul(x = var_807, y = out_23)[name = tensor("input_139")]; - tensor ret_out_7 = linear(bias = encoder_out_proj_3_bias, weight = encoder_out_proj_3_weight, x = input_139)[name = tensor("linear_34")]; - tensor x_21 = add(x = input_133, y = ret_out_7)[name = tensor("x_21")]; + tensor var_837 = transpose(perm = var_837_perm_0, x = out_19)[name = tensor("transpose_34")]; + tensor out_21 = layer_norm(axes = out_21_axes_0, epsilon = var_54, x = var_837)[name = tensor("out_21")]; + tensor var_841 = const()[name = tensor("op_841"), val = tensor([1, 2, 256])]; + tensor out_23 = reshape(shape = var_841, x = out_21)[name = tensor("out_23")]; + tensor var_843 = silu(x = input_139)[name = tensor("op_843")]; + tensor input_141 = mul(x = var_843, y = out_23)[name = tensor("input_141")]; + tensor ret_out_7 = linear(bias = inner_encoder_out_proj_3_bias, weight = inner_encoder_out_proj_3_weight, x = input_141)[name = tensor("linear_34")]; + tensor x_21 = add(x = input_135, y = ret_out_7)[name = tensor("x_21")]; tensor window_19_begin_0 = const()[name = tensor("window_19_begin_0"), val = tensor([3, 0, 0, 0])]; tensor window_19_end_0 = const()[name = tensor("window_19_end_0"), val = tensor([4, 1, 16, 256])]; tensor window_19_end_mask_0 = const()[name = tensor("window_19_end_mask_0"), val = tensor([false, true, true, true])]; tensor window_19_squeeze_mask_0 = const()[name = tensor("window_19_squeeze_mask_0"), val = tensor([true, false, false, false])]; tensor window_19 = slice_by_index(begin = window_19_begin_0, end = window_19_end_0, end_mask = window_19_end_mask_0, squeeze_mask = window_19_squeeze_mask_0, x = enc_conv_cache)[name = tensor("window_19")]; - tensor var_815_begin_0 = const()[name = tensor("op_815_begin_0"), val = tensor([0, 0, 0])]; - tensor var_815_end_0 = const()[name = tensor("op_815_end_0"), val = tensor([1, 1, 256])]; - tensor var_815_end_mask_0 = const()[name = tensor("op_815_end_mask_0"), val = tensor([true, false, true])]; - tensor var_815 = slice_by_index(begin = var_815_begin_0, end = var_815_end_0, end_mask = var_815_end_mask_0, x = x_21)[name = tensor("op_815")]; - tensor var_818_begin_0 = const()[name = tensor("op_818_begin_0"), val = tensor([0, 1, 0])]; - tensor var_818_end_0 = const()[name = tensor("op_818_end_0"), val = tensor([1, 16, 256])]; - tensor var_818_end_mask_0 = const()[name = tensor("op_818_end_mask_0"), val = tensor([true, true, true])]; - tensor var_818 = slice_by_index(begin = var_818_begin_0, end = var_818_end_0, end_mask = var_818_end_mask_0, x = window_19)[name = tensor("op_818")]; + tensor var_851_begin_0 = const()[name = tensor("op_851_begin_0"), val = tensor([0, 0, 0])]; + tensor var_851_end_0 = const()[name = tensor("op_851_end_0"), val = tensor([1, 1, 256])]; + tensor var_851_end_mask_0 = const()[name = tensor("op_851_end_mask_0"), val = tensor([true, false, true])]; + tensor var_851 = slice_by_index(begin = var_851_begin_0, end = var_851_end_0, end_mask = var_851_end_mask_0, x = x_21)[name = tensor("op_851")]; + tensor var_854_begin_0 = const()[name = tensor("op_854_begin_0"), val = tensor([0, 1, 0])]; + tensor var_854_end_0 = const()[name = tensor("op_854_end_0"), val = tensor([1, 16, 256])]; + tensor var_854_end_mask_0 = const()[name = tensor("op_854_end_mask_0"), val = tensor([true, true, true])]; + tensor var_854 = slice_by_index(begin = var_854_begin_0, end = var_854_end_0, end_mask = var_854_end_mask_0, x = window_19)[name = tensor("op_854")]; tensor window_21_interleave_0 = const()[name = tensor("window_21_interleave_0"), val = tensor(false)]; - tensor window_21 = concat(axis = var_26, interleave = window_21_interleave_0, values = (var_818, var_815))[name = tensor("window_21")]; - tensor var_823_begin_0 = const()[name = tensor("op_823_begin_0"), val = tensor([0, 1, 0])]; - tensor var_823_end_0 = const()[name = tensor("op_823_end_0"), val = tensor([1, 1, 256])]; - tensor var_823_end_mask_0 = const()[name = tensor("op_823_end_mask_0"), val = tensor([true, true, true])]; - tensor var_823 = slice_by_index(begin = var_823_begin_0, end = var_823_end_0, end_mask = var_823_end_mask_0, x = x_21)[name = tensor("op_823")]; - tensor var_826_begin_0 = const()[name = tensor("op_826_begin_0"), val = tensor([0, 1, 0])]; - tensor var_826_end_0 = const()[name = tensor("op_826_end_0"), val = tensor([1, 16, 256])]; - tensor var_826_end_mask_0 = const()[name = tensor("op_826_end_mask_0"), val = tensor([true, true, true])]; - tensor var_826 = slice_by_index(begin = var_826_begin_0, end = var_826_end_0, end_mask = var_826_end_mask_0, x = window_21)[name = tensor("op_826")]; + tensor window_21 = concat(axis = var_62, interleave = window_21_interleave_0, values = (var_854, var_851))[name = tensor("window_21")]; + tensor var_859_begin_0 = const()[name = tensor("op_859_begin_0"), val = tensor([0, 1, 0])]; + tensor var_859_end_0 = const()[name = tensor("op_859_end_0"), val = tensor([1, 1, 256])]; + tensor var_859_end_mask_0 = const()[name = tensor("op_859_end_mask_0"), val = tensor([true, true, true])]; + tensor var_859 = slice_by_index(begin = var_859_begin_0, end = var_859_end_0, end_mask = var_859_end_mask_0, x = x_21)[name = tensor("op_859")]; + tensor var_862_begin_0 = const()[name = tensor("op_862_begin_0"), val = tensor([0, 1, 0])]; + tensor var_862_end_0 = const()[name = tensor("op_862_end_0"), val = tensor([1, 16, 256])]; + tensor var_862_end_mask_0 = const()[name = tensor("op_862_end_mask_0"), val = tensor([true, true, true])]; + tensor var_862 = slice_by_index(begin = var_862_begin_0, end = var_862_end_0, end_mask = var_862_end_mask_0, x = window_21)[name = tensor("op_862")]; tensor window_interleave_0 = const()[name = tensor("window_interleave_0"), val = tensor(false)]; - tensor window = concat(axis = var_26, interleave = window_interleave_0, values = (var_826, var_823))[name = tensor("window")]; - tensor input_141_interleave_0 = const()[name = tensor("input_141_interleave_0"), val = tensor(false)]; - tensor input_141 = concat(axis = var_23, interleave = input_141_interleave_0, values = (window_21, window))[name = tensor("input_141")]; + tensor window = concat(axis = var_62, interleave = window_interleave_0, values = (var_862, var_859))[name = tensor("window")]; + tensor input_143_interleave_0 = const()[name = tensor("input_143_interleave_0"), val = tensor(false)]; + tensor input_143 = concat(axis = var_49, interleave = input_143_interleave_0, values = (window_21, window))[name = tensor("input_143")]; tensor x_23_axes_0 = const()[name = tensor("x_23_axes_0"), val = tensor([-1])]; - tensor x_23 = layer_norm(axes = x_23_axes_0, beta = encoder_conv_module_3_sequential_0_bias, epsilon = var_28, gamma = encoder_conv_module_3_sequential_0_weight, x = input_141)[name = tensor("x_23")]; - tensor input_143_perm_0 = const()[name = tensor("input_143_perm_0"), val = tensor([0, 2, 1])]; + tensor x_23 = layer_norm(axes = x_23_axes_0, beta = inner_encoder_conv_module_3_sequential_0_bias, epsilon = var_46, gamma = inner_encoder_conv_module_3_sequential_0_weight, x = input_143)[name = tensor("x_23")]; + tensor input_145_perm_0 = const()[name = tensor("input_145_perm_0"), val = tensor([0, 2, 1])]; tensor inputs_33_pad_type_0 = const()[name = tensor("inputs_33_pad_type_0"), val = tensor("valid")]; tensor inputs_33_strides_0 = const()[name = tensor("inputs_33_strides_0"), val = tensor([1])]; tensor inputs_33_pad_0 = const()[name = tensor("inputs_33_pad_0"), val = tensor([0, 0])]; tensor inputs_33_dilations_0 = const()[name = tensor("inputs_33_dilations_0"), val = tensor([1])]; tensor inputs_33_groups_0 = const()[name = tensor("inputs_33_groups_0"), val = tensor(1)]; - tensor input_143 = transpose(perm = input_143_perm_0, x = x_23)[name = tensor("transpose_33")]; - tensor inputs_33 = conv(bias = encoder_conv_module_3_sequential_2_conv_bias, dilations = inputs_33_dilations_0, groups = inputs_33_groups_0, pad = inputs_33_pad_0, pad_type = inputs_33_pad_type_0, strides = inputs_33_strides_0, weight = encoder_conv_module_3_sequential_2_conv_weight, x = input_143)[name = tensor("inputs_33")]; - tensor var_851_split_sizes_0 = const()[name = tensor("op_851_split_sizes_0"), val = tensor([256, 256])]; - tensor var_851_axis_0 = const()[name = tensor("op_851_axis_0"), val = tensor(1)]; - tensor var_851_0, tensor var_851_1 = split(axis = var_851_axis_0, split_sizes = var_851_split_sizes_0, x = inputs_33)[name = tensor("op_851")]; - tensor var_853 = sigmoid(x = var_851_1)[name = tensor("op_853")]; - tensor inputs_35 = mul(x = var_851_0, y = var_853)[name = tensor("inputs_35")]; + tensor input_145 = transpose(perm = input_145_perm_0, x = x_23)[name = tensor("transpose_33")]; + tensor inputs_33 = conv(bias = inner_encoder_conv_module_3_sequential_2_conv_bias, dilations = inputs_33_dilations_0, groups = inputs_33_groups_0, pad = inputs_33_pad_0, pad_type = inputs_33_pad_type_0, strides = inputs_33_strides_0, weight = inner_encoder_conv_module_3_sequential_2_conv_weight, x = input_145)[name = tensor("inputs_33")]; + tensor var_887_split_sizes_0 = const()[name = tensor("op_887_split_sizes_0"), val = tensor([256, 256])]; + tensor var_887_axis_0 = const()[name = tensor("op_887_axis_0"), val = tensor(1)]; + tensor var_887_0, tensor var_887_1 = split(axis = var_887_axis_0, split_sizes = var_887_split_sizes_0, x = inputs_33)[name = tensor("op_887")]; + tensor var_889 = sigmoid(x = var_887_1)[name = tensor("op_889")]; + tensor inputs_35 = mul(x = var_887_0, y = var_889)[name = tensor("inputs_35")]; tensor outputs_aug_pad_type_0 = const()[name = tensor("outputs_aug_pad_type_0"), val = tensor("custom")]; tensor outputs_aug_pad_0 = const()[name = tensor("outputs_aug_pad_0"), val = tensor([15, 15])]; tensor outputs_aug_groups_0 = const()[name = tensor("outputs_aug_groups_0"), val = tensor(256)]; tensor outputs_aug_strides_0 = const()[name = tensor("outputs_aug_strides_0"), val = tensor([1])]; tensor outputs_aug_dilations_0 = const()[name = tensor("outputs_aug_dilations_0"), val = tensor([1])]; - tensor outputs_aug = conv(dilations = outputs_aug_dilations_0, groups = outputs_aug_groups_0, pad = outputs_aug_pad_0, pad_type = outputs_aug_pad_type_0, strides = outputs_aug_strides_0, weight = encoder_conv_module_3_sequential_4_conv_weight, x = inputs_35)[name = tensor("outputs_aug")]; - tensor input_145_begin_0 = const()[name = tensor("input_145_begin_0"), val = tensor([0, 0, 0])]; - tensor input_145_end_0 = const()[name = tensor("input_145_end_0"), val = tensor([2, 256, 16])]; - tensor input_145_end_mask_0 = const()[name = tensor("input_145_end_mask_0"), val = tensor([true, true, false])]; - tensor input_145 = slice_by_index(begin = input_145_begin_0, end = input_145_end_0, end_mask = input_145_end_mask_0, x = outputs_aug)[name = tensor("input_145")]; - tensor inputs_37 = batch_norm(beta = encoder_conv_module_3_sequential_5_bias, epsilon = var_28, gamma = encoder_conv_module_3_sequential_5_weight, mean = encoder_conv_module_3_sequential_5_running_mean, variance = encoder_conv_module_3_sequential_5_running_var, x = input_145)[name = tensor("inputs_37")]; - tensor input_147 = silu(x = inputs_37)[name = tensor("input_147")]; - tensor input_149_pad_type_0 = const()[name = tensor("input_149_pad_type_0"), val = tensor("valid")]; - tensor input_149_strides_0 = const()[name = tensor("input_149_strides_0"), val = tensor([1])]; - tensor input_149_pad_0 = const()[name = tensor("input_149_pad_0"), val = tensor([0, 0])]; - tensor input_149_dilations_0 = const()[name = tensor("input_149_dilations_0"), val = tensor([1])]; - tensor input_149_groups_0 = const()[name = tensor("input_149_groups_0"), val = tensor(1)]; - tensor input_149 = conv(bias = encoder_conv_module_3_sequential_7_conv_bias, dilations = input_149_dilations_0, groups = input_149_groups_0, pad = input_149_pad_0, pad_type = input_149_pad_type_0, strides = input_149_strides_0, weight = encoder_conv_module_3_sequential_7_conv_weight, x = input_147)[name = tensor("input_149")]; + tensor outputs_aug = conv(dilations = outputs_aug_dilations_0, groups = outputs_aug_groups_0, pad = outputs_aug_pad_0, pad_type = outputs_aug_pad_type_0, strides = outputs_aug_strides_0, weight = inner_encoder_conv_module_3_sequential_4_conv_weight, x = inputs_35)[name = tensor("outputs_aug")]; + tensor input_147_begin_0 = const()[name = tensor("input_147_begin_0"), val = tensor([0, 0, 0])]; + tensor input_147_end_0 = const()[name = tensor("input_147_end_0"), val = tensor([2, 256, 16])]; + tensor input_147_end_mask_0 = const()[name = tensor("input_147_end_mask_0"), val = tensor([true, true, false])]; + tensor input_147 = slice_by_index(begin = input_147_begin_0, end = input_147_end_0, end_mask = input_147_end_mask_0, x = outputs_aug)[name = tensor("input_147")]; + tensor inputs_37 = batch_norm(beta = inner_encoder_conv_module_3_sequential_5_bias, epsilon = var_46, gamma = inner_encoder_conv_module_3_sequential_5_weight, mean = inner_encoder_conv_module_3_sequential_5_running_mean, variance = inner_encoder_conv_module_3_sequential_5_running_var, x = input_147)[name = tensor("inputs_37")]; + tensor input_149 = silu(x = inputs_37)[name = tensor("input_149")]; + tensor input_151_pad_type_0 = const()[name = tensor("input_151_pad_type_0"), val = tensor("valid")]; + tensor input_151_strides_0 = const()[name = tensor("input_151_strides_0"), val = tensor([1])]; + tensor input_151_pad_0 = const()[name = tensor("input_151_pad_0"), val = tensor([0, 0])]; + tensor input_151_dilations_0 = const()[name = tensor("input_151_dilations_0"), val = tensor([1])]; + tensor input_151_groups_0 = const()[name = tensor("input_151_groups_0"), val = tensor(1)]; + tensor input_151 = conv(bias = inner_encoder_conv_module_3_sequential_7_conv_bias, dilations = input_151_dilations_0, groups = input_151_groups_0, pad = input_151_pad_0, pad_type = input_151_pad_type_0, strides = input_151_strides_0, weight = inner_encoder_conv_module_3_sequential_7_conv_weight, x = input_149)[name = tensor("input_151")]; tensor conv_out_7_perm_0 = const()[name = tensor("conv_out_7_perm_0"), val = tensor([0, 2, 1])]; - tensor var_884_begin_0 = const()[name = tensor("op_884_begin_0"), val = tensor([0, -1, 0])]; - tensor var_884_end_0 = const()[name = tensor("op_884_end_0"), val = tensor([2, 16, 256])]; - tensor var_884_end_mask_0 = const()[name = tensor("op_884_end_mask_0"), val = tensor([true, true, true])]; - tensor conv_out_7 = transpose(perm = conv_out_7_perm_0, x = input_149)[name = tensor("transpose_32")]; - tensor var_884 = slice_by_index(begin = var_884_begin_0, end = var_884_end_0, end_mask = var_884_end_mask_0, x = conv_out_7)[name = tensor("op_884")]; - tensor var_886_perm_0 = const()[name = tensor("op_886_perm_0"), val = tensor([1, 0, 2])]; - tensor var_886 = transpose(perm = var_886_perm_0, x = var_884)[name = tensor("transpose_31")]; - tensor input_151 = add(x = x_21, y = var_886)[name = tensor("input_151")]; - tensor input_153_axes_0 = const()[name = tensor("input_153_axes_0"), val = tensor([-1])]; - tensor input_153 = layer_norm(axes = input_153_axes_0, beta = encoder_ffn2_3_module_sequential_0_bias, epsilon = var_28, gamma = encoder_ffn2_3_module_sequential_0_weight, x = input_151)[name = tensor("input_153")]; - tensor inputs = linear(bias = encoder_ffn2_3_module_sequential_1_linear_bias, weight = encoder_ffn2_3_module_sequential_1_linear_weight, x = input_153)[name = tensor("linear_35")]; - tensor input_155 = silu(x = inputs)[name = tensor("input_155")]; - tensor input_159 = linear(bias = encoder_ffn2_3_module_sequential_4_linear_bias, weight = encoder_ffn2_3_module_sequential_4_linear_weight, x = input_155)[name = tensor("linear_36")]; - tensor var_909 = const()[name = tensor("op_909"), val = tensor(0x1p-1)]; - tensor var_910 = mul(x = input_159, y = var_909)[name = tensor("op_910")]; - tensor input_161 = add(x = var_910, y = input_151)[name = tensor("input_161")]; + tensor var_920_begin_0 = const()[name = tensor("op_920_begin_0"), val = tensor([0, -1, 0])]; + tensor var_920_end_0 = const()[name = tensor("op_920_end_0"), val = tensor([2, 16, 256])]; + tensor var_920_end_mask_0 = const()[name = tensor("op_920_end_mask_0"), val = tensor([true, true, true])]; + tensor conv_out_7 = transpose(perm = conv_out_7_perm_0, x = input_151)[name = tensor("transpose_32")]; + tensor var_920 = slice_by_index(begin = var_920_begin_0, end = var_920_end_0, end_mask = var_920_end_mask_0, x = conv_out_7)[name = tensor("op_920")]; + tensor var_922_perm_0 = const()[name = tensor("op_922_perm_0"), val = tensor([1, 0, 2])]; + tensor var_922 = transpose(perm = var_922_perm_0, x = var_920)[name = tensor("transpose_31")]; + tensor input_153 = add(x = x_21, y = var_922)[name = tensor("input_153")]; + tensor input_155_axes_0 = const()[name = tensor("input_155_axes_0"), val = tensor([-1])]; + tensor input_155 = layer_norm(axes = input_155_axes_0, beta = inner_encoder_ffn2_3_module_sequential_0_bias, epsilon = var_46, gamma = inner_encoder_ffn2_3_module_sequential_0_weight, x = input_153)[name = tensor("input_155")]; + tensor inputs = linear(bias = inner_encoder_ffn2_3_module_sequential_1_linear_bias, weight = inner_encoder_ffn2_3_module_sequential_1_linear_weight, x = input_155)[name = tensor("linear_35")]; + tensor input_157 = silu(x = inputs)[name = tensor("input_157")]; + tensor input_161 = linear(bias = inner_encoder_ffn2_3_module_sequential_4_linear_bias, weight = inner_encoder_ffn2_3_module_sequential_4_linear_weight, x = input_157)[name = tensor("linear_36")]; + tensor var_945 = const()[name = tensor("op_945"), val = tensor(0x1p-1)]; + tensor var_946 = mul(x = input_161, y = var_945)[name = tensor("op_946")]; + tensor input_163 = add(x = var_946, y = input_153)[name = tensor("input_163")]; tensor x_25_axes_0 = const()[name = tensor("x_25_axes_0"), val = tensor([-1])]; - tensor x_25 = layer_norm(axes = x_25_axes_0, beta = encoder_layer_norm_3_bias, epsilon = var_28, gamma = encoder_layer_norm_3_weight, x = input_161)[name = tensor("x_25")]; + tensor x_25 = layer_norm(axes = x_25_axes_0, beta = inner_encoder_layer_norm_3_bias, epsilon = var_46, gamma = inner_encoder_layer_norm_3_weight, x = input_163)[name = tensor("x_25")]; tensor x_ct_perm_0 = const()[name = tensor("x_ct_perm_0"), val = tensor([0, 2, 1])]; tensor cat_interleave_0 = const()[name = tensor("cat_interleave_0"), val = tensor(false)]; tensor x_ct = transpose(perm = x_ct_perm_0, x = x_25)[name = tensor("transpose_30")]; - tensor cat = concat(axis = var_20, interleave = cat_interleave_0, values = (cnn_window, x_ct))[name = tensor("cat")]; + tensor cat = concat(axis = var_51, interleave = cat_interleave_0, values = (cnn_window, x_ct))[name = tensor("cat")]; tensor conv_out_pad_type_0 = const()[name = tensor("conv_out_pad_type_0"), val = tensor("valid")]; tensor conv_out_strides_0 = const()[name = tensor("conv_out_strides_0"), val = tensor([1])]; tensor conv_out_pad_0 = const()[name = tensor("conv_out_pad_0"), val = tensor([0, 0])]; tensor conv_out_dilations_0 = const()[name = tensor("conv_out_dilations_0"), val = tensor([1])]; tensor conv_out_groups_0 = const()[name = tensor("conv_out_groups_0"), val = tensor(1)]; - tensor conv_out = conv(bias = encoder_cnn_bias, dilations = conv_out_dilations_0, groups = conv_out_groups_0, pad = conv_out_pad_0, pad_type = conv_out_pad_type_0, strides = conv_out_strides_0, weight = encoder_cnn_weight, x = cat)[name = tensor("conv_out")]; - tensor var_928_begin_0 = const()[name = tensor("op_928_begin_0"), val = tensor([0, 0, 2])]; - tensor var_928_end_0 = const()[name = tensor("op_928_end_0"), val = tensor([1, 256, 20])]; - tensor var_928_end_mask_0 = const()[name = tensor("op_928_end_mask_0"), val = tensor([true, true, true])]; - tensor cnn_window_new = slice_by_index(begin = var_928_begin_0, end = var_928_end_0, end_mask = var_928_end_mask_0, x = cat)[name = tensor("op_928")]; - tensor input_163_perm_0 = const()[name = tensor("input_163_perm_0"), val = tensor([0, 2, 1])]; - tensor var_930 = const()[name = tensor("op_930"), val = tensor([-1])]; - tensor input_163 = transpose(perm = input_163_perm_0, x = conv_out)[name = tensor("transpose_29")]; - tensor var_931 = reduce_l2_norm(axes = var_930, keep_dims = var_29, x = input_163)[name = tensor("op_931")]; + tensor conv_out = conv(bias = inner_encoder_cnn_bias, dilations = conv_out_dilations_0, groups = conv_out_groups_0, pad = conv_out_pad_0, pad_type = conv_out_pad_type_0, strides = conv_out_strides_0, weight = inner_encoder_cnn_weight, x = cat)[name = tensor("conv_out")]; + tensor var_964_begin_0 = const()[name = tensor("op_964_begin_0"), val = tensor([0, 0, 2])]; + tensor var_964_end_0 = const()[name = tensor("op_964_end_0"), val = tensor([1, 256, 20])]; + tensor var_964_end_mask_0 = const()[name = tensor("op_964_end_mask_0"), val = tensor([true, true, true])]; + tensor cnn_window_new = slice_by_index(begin = var_964_begin_0, end = var_964_end_0, end_mask = var_964_end_mask_0, x = cat)[name = tensor("op_964")]; + tensor input_165_perm_0 = const()[name = tensor("input_165_perm_0"), val = tensor([0, 2, 1])]; + tensor var_966 = const()[name = tensor("op_966"), val = tensor([-1])]; + tensor input_165 = transpose(perm = input_165_perm_0, x = conv_out)[name = tensor("transpose_29")]; + tensor var_967 = reduce_l2_norm(axes = var_966, keep_dims = var_45, x = input_165)[name = tensor("op_967")]; tensor const_12 = const()[name = tensor("const_12"), val = tensor(0x1.fffffep+127)]; - tensor clip_0 = clip(alpha = var_11, beta = const_12, x = var_931)[name = tensor("clip_0")]; - tensor emb = real_div(x = input_163, y = clip_0)[name = tensor("emb")]; - tensor var_935_axis_0 = const()[name = tensor("op_935_axis_0"), val = tensor(0)]; - tensor enc_kv_new = stack(axis = var_935_axis_0, values = (var_206, var_404, var_602, nkv_1))[name = tensor("op_935")]; - tensor var_937_axis_0 = const()[name = tensor("op_937_axis_0"), val = tensor(0)]; - tensor enc_scale_new = stack(axis = var_937_axis_0, values = (new_scale_1, new_scale_3, new_scale_5, new_scale_7))[name = tensor("op_937")]; - tensor var_939_axis_0 = const()[name = tensor("op_939_axis_0"), val = tensor(0)]; - tensor enc_conv_cache_new = stack(axis = var_939_axis_0, values = (window_5, window_11, window_17, window))[name = tensor("op_939")]; - tensor var_948 = const()[name = tensor("op_948"), val = tensor(0x1.5798eep-27)]; - tensor var_953 = const()[name = tensor("op_953"), val = tensor(0x1.4f8b58p-17)]; - tensor var_955 = const()[name = tensor("op_955"), val = tensor(0x1.0c6f7ap-20)]; - tensor var_956 = const()[name = tensor("op_956"), val = tensor(true)]; - tensor var_958 = const()[name = tensor("op_958"), val = tensor(0x1p+0)]; - tensor var_962 = const()[name = tensor("op_962"), val = tensor(-1)]; - tensor var_968 = const()[name = tensor("op_968"), val = tensor(0)]; + tensor clip_0 = clip(alpha = var_59, beta = const_12, x = var_967)[name = tensor("clip_0")]; + tensor emb = real_div(x = input_165, y = clip_0)[name = tensor("emb")]; + tensor var_971_axis_0 = const()[name = tensor("op_971_axis_0"), val = tensor(0)]; + tensor enc_kv_new = stack(axis = var_971_axis_0, values = (var_242, var_440, var_638, nkv_1))[name = tensor("op_971")]; + tensor var_973_axis_0 = const()[name = tensor("op_973_axis_0"), val = tensor(0)]; + tensor enc_scale_new = stack(axis = var_973_axis_0, values = (new_scale_1, new_scale_3, new_scale_5, new_scale_7))[name = tensor("op_973")]; + tensor var_975_axis_0 = const()[name = tensor("op_975_axis_0"), val = tensor(0)]; + tensor enc_conv_cache_new = stack(axis = var_975_axis_0, values = (window_5, window_11, window_17, window))[name = tensor("op_975")]; tensor pos = const()[name = tensor("pos"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(44395328)))]; - tensor var_1030_axes_0 = const()[name = tensor("op_1030_axes_0"), val = tensor([2])]; - tensor var_1030 = expand_dims(axes = var_1030_axes_0, x = emb)[name = tensor("op_1030")]; + tensor var_1043_axes_0 = const()[name = tensor("op_1043_axes_0"), val = tensor([2])]; + tensor var_1043 = expand_dims(axes = var_1043_axes_0, x = emb)[name = tensor("op_1043")]; tensor emb_exp_reps_0 = const()[name = tensor("emb_exp_reps_0"), val = tensor([1, 1, 12, 1])]; - tensor emb_exp = tile(reps = emb_exp_reps_0, x = var_1030)[name = tensor("emb_exp")]; - tensor input_165_interleave_0 = const()[name = tensor("input_165_interleave_0"), val = tensor(false)]; - tensor input_165 = concat(axis = var_962, interleave = input_165_interleave_0, values = (emb_exp, pos))[name = tensor("input_165")]; - tensor x_27 = linear(bias = decoder_convert_bias, weight = decoder_convert_weight, x = input_165)[name = tensor("linear_37")]; - tensor var_1038_perm_0 = const()[name = tensor("op_1038_perm_0"), val = tensor([0, 2, 1, 3])]; - tensor var_1042 = const()[name = tensor("op_1042"), val = tensor([12, 2, 256])]; - tensor var_1038 = transpose(perm = var_1038_perm_0, x = x_27)[name = tensor("transpose_28")]; - tensor x_29 = reshape(shape = var_1042, x = var_1038)[name = tensor("x_29")]; + tensor emb_exp = tile(reps = emb_exp_reps_0, x = var_1043)[name = tensor("emb_exp")]; + tensor input_167_interleave_0 = const()[name = tensor("input_167_interleave_0"), val = tensor(false)]; + tensor input_167 = concat(axis = var_52, interleave = input_167_interleave_0, values = (emb_exp, pos))[name = tensor("input_167")]; + tensor x_27 = linear(bias = inner_decoder_convert_bias, weight = inner_decoder_convert_weight, x = input_167)[name = tensor("linear_37")]; + tensor var_1051_perm_0 = const()[name = tensor("op_1051_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_1055 = const()[name = tensor("op_1055"), val = tensor([12, 2, 256])]; + tensor var_1051 = transpose(perm = var_1051_perm_0, x = x_27)[name = tensor("transpose_28")]; + tensor x_29 = reshape(shape = var_1055, x = var_1051)[name = tensor("x_29")]; tensor prev_kv_9_begin_0 = const()[name = tensor("prev_kv_9_begin_0"), val = tensor([0, 0, 0, 0, 0])]; tensor prev_kv_9_end_0 = const()[name = tensor("prev_kv_9_end_0"), val = tensor([1, 12, 4, 64, 64])]; tensor prev_kv_9_end_mask_0 = const()[name = tensor("prev_kv_9_end_mask_0"), val = tensor([false, true, true, true, true])]; @@ -903,132 +910,132 @@ program(1.0) tensor prev_scale_9_end_mask_0 = const()[name = tensor("prev_scale_9_end_mask_0"), val = tensor([false, true])]; tensor prev_scale_9_squeeze_mask_0 = const()[name = tensor("prev_scale_9_squeeze_mask_0"), val = tensor([true, false])]; tensor prev_scale_9 = slice_by_index(begin = prev_scale_9_begin_0, end = prev_scale_9_end_0, end_mask = prev_scale_9_end_mask_0, squeeze_mask = prev_scale_9_squeeze_mask_0, x = dec_scale)[name = tensor("prev_scale_9")]; - tensor var_1050 = linear(bias = decoder_q_proj_0_bias, weight = decoder_q_proj_0_weight, x = x_29)[name = tensor("linear_38")]; - tensor var_1051 = const()[name = tensor("op_1051"), val = tensor([12, 2, 4, 64])]; - tensor var_1052 = reshape(shape = var_1051, x = var_1050)[name = tensor("op_1052")]; + tensor var_1063 = linear(bias = inner_decoder_q_proj_0_bias, weight = inner_decoder_q_proj_0_weight, x = x_29)[name = tensor("linear_38")]; + tensor var_1064 = const()[name = tensor("op_1064"), val = tensor([12, 2, 4, 64])]; + tensor var_1065 = reshape(shape = var_1064, x = var_1063)[name = tensor("op_1065")]; tensor q_9_perm_0 = const()[name = tensor("q_9_perm_0"), val = tensor([0, 2, 1, 3])]; - tensor var_1056 = linear(bias = decoder_k_proj_0_bias, weight = decoder_k_proj_0_weight, x = x_29)[name = tensor("linear_39")]; - tensor var_1057 = const()[name = tensor("op_1057"), val = tensor(0x1p-3)]; - tensor var_1058 = mul(x = var_1056, y = var_1057)[name = tensor("op_1058")]; - tensor var_1059 = const()[name = tensor("op_1059"), val = tensor([12, 2, 4, 64])]; - tensor var_1060 = reshape(shape = var_1059, x = var_1058)[name = tensor("op_1060")]; + tensor var_1069 = linear(bias = inner_decoder_k_proj_0_bias, weight = inner_decoder_k_proj_0_weight, x = x_29)[name = tensor("linear_39")]; + tensor var_1070 = const()[name = tensor("op_1070"), val = tensor(0x1p-3)]; + tensor var_1071 = mul(x = var_1069, y = var_1070)[name = tensor("op_1071")]; + tensor var_1072 = const()[name = tensor("op_1072"), val = tensor([12, 2, 4, 64])]; + tensor var_1073 = reshape(shape = var_1072, x = var_1071)[name = tensor("op_1073")]; tensor k_9_perm_0 = const()[name = tensor("k_9_perm_0"), val = tensor([0, 2, 1, 3])]; - tensor var_1064 = linear(bias = decoder_v_proj_0_bias, weight = decoder_v_proj_0_weight, x = x_29)[name = tensor("linear_40")]; - tensor var_1065 = const()[name = tensor("op_1065"), val = tensor([12, 2, 4, 64])]; - tensor var_1066 = reshape(shape = var_1065, x = var_1064)[name = tensor("op_1066")]; + tensor var_1077 = linear(bias = inner_decoder_v_proj_0_bias, weight = inner_decoder_v_proj_0_weight, x = x_29)[name = tensor("linear_40")]; + tensor var_1078 = const()[name = tensor("op_1078"), val = tensor([12, 2, 4, 64])]; + tensor var_1079 = reshape(shape = var_1078, x = var_1077)[name = tensor("op_1079")]; tensor v_9_perm_0 = const()[name = tensor("v_9_perm_0"), val = tensor([0, 2, 1, 3])]; - tensor input_169 = linear(bias = decoder_g_proj_0_bias, weight = decoder_g_proj_0_weight, x = x_29)[name = tensor("linear_41")]; + tensor input_171 = linear(bias = inner_decoder_g_proj_0_bias, weight = inner_decoder_g_proj_0_weight, x = x_29)[name = tensor("linear_41")]; tensor cumsum_mask_1_exclusive_0 = const()[name = tensor("cumsum_mask_1_exclusive_0"), val = tensor(false)]; tensor cumsum_mask_1_reverse_0 = const()[name = tensor("cumsum_mask_1_reverse_0"), val = tensor(false)]; - tensor cumsum_mask_1 = cumsum(axis = var_968, exclusive = cumsum_mask_1_exclusive_0, reverse = cumsum_mask_1_reverse_0, x = valid_mask)[name = tensor("cumsum_mask_1")]; + tensor cumsum_mask_1 = cumsum(axis = var_49, exclusive = cumsum_mask_1_exclusive_0, reverse = cumsum_mask_1_reverse_0, x = valid_mask)[name = tensor("cumsum_mask_1")]; tensor sqrt_s0_9 = sqrt(x = prev_scale_9)[name = tensor("sqrt_s0_9")]; tensor s_t_9 = add(x = prev_scale_9, y = cumsum_mask_1)[name = tensor("s_t_9")]; tensor const_20 = const()[name = tensor("const_20"), val = tensor(0x1.fffffep+127)]; - tensor clip_1 = clip(alpha = var_958, beta = const_20, x = s_t_9)[name = tensor("clip_1")]; + tensor clip_1 = clip(alpha = var_39, beta = const_20, x = s_t_9)[name = tensor("clip_1")]; tensor sqrt_s_t_9 = sqrt(x = clip_1)[name = tensor("sqrt_s_t_9")]; tensor qk_9_transpose_x_1 = const()[name = tensor("qk_9_transpose_x_1"), val = tensor(false)]; tensor qk_9_transpose_y_1 = const()[name = tensor("qk_9_transpose_y_1"), val = tensor(true)]; - tensor k_9 = transpose(perm = k_9_perm_0, x = var_1060)[name = tensor("transpose_26")]; - tensor q_9 = transpose(perm = q_9_perm_0, x = var_1052)[name = tensor("transpose_27")]; + tensor k_9 = transpose(perm = k_9_perm_0, x = var_1073)[name = tensor("transpose_26")]; + tensor q_9 = transpose(perm = q_9_perm_0, x = var_1065)[name = tensor("transpose_27")]; tensor qk_9 = matmul(transpose_x = qk_9_transpose_x_1, transpose_y = qk_9_transpose_y_1, x = q_9, y = k_9)[name = tensor("qk_9")]; - tensor var_1078 = const()[name = tensor("op_1078"), val = tensor([1, 2])]; - tensor var_1079 = reshape(shape = var_1078, x = valid_mask)[name = tensor("op_1079")]; - tensor causal_with_valid_1 = mul(x = decoder__causal_mask, y = var_1079)[name = tensor("causal_with_valid_1")]; - tensor var_1081 = const()[name = tensor("op_1081"), val = tensor([2, 1])]; - tensor var_1082 = reshape(shape = var_1081, x = sqrt_s_t_9)[name = tensor("op_1082")]; - tensor M_9 = real_div(x = causal_with_valid_1, y = var_1082)[name = tensor("M_9")]; - tensor var_1084 = mul(x = qk_9, y = M_9)[name = tensor("op_1084")]; + tensor var_1091 = const()[name = tensor("op_1091"), val = tensor([1, 2])]; + tensor var_1092 = reshape(shape = var_1091, x = valid_mask)[name = tensor("op_1092")]; + tensor causal_with_valid_1 = mul(x = inner_decoder__causal_mask, y = var_1092)[name = tensor("causal_with_valid_1")]; + tensor var_1094 = const()[name = tensor("op_1094"), val = tensor([2, 1])]; + tensor var_1095 = reshape(shape = var_1094, x = sqrt_s_t_9)[name = tensor("op_1095")]; + tensor M_9 = real_div(x = causal_with_valid_1, y = var_1095)[name = tensor("M_9")]; + tensor var_1097 = mul(x = qk_9, y = M_9)[name = tensor("op_1097")]; tensor inner_9_transpose_x_0 = const()[name = tensor("inner_9_transpose_x_0"), val = tensor(false)]; tensor inner_9_transpose_y_0 = const()[name = tensor("inner_9_transpose_y_0"), val = tensor(false)]; - tensor v_9 = transpose(perm = v_9_perm_0, x = var_1066)[name = tensor("transpose_25")]; - tensor inner_9 = matmul(transpose_x = inner_9_transpose_x_0, transpose_y = inner_9_transpose_y_0, x = var_1084, y = v_9)[name = tensor("inner_9")]; - tensor var_1086_transpose_x_0 = const()[name = tensor("op_1086_transpose_x_0"), val = tensor(false)]; - tensor var_1086_transpose_y_0 = const()[name = tensor("op_1086_transpose_y_0"), val = tensor(false)]; - tensor var_1086 = matmul(transpose_x = var_1086_transpose_x_0, transpose_y = var_1086_transpose_y_0, x = q_9, y = prev_kv_9)[name = tensor("op_1086")]; - tensor var_1087 = real_div(x = sqrt_s0_9, y = sqrt_s_t_9)[name = tensor("op_1087")]; - tensor var_1088 = const()[name = tensor("op_1088"), val = tensor([1, 1, 2, 1])]; - tensor var_1089 = reshape(shape = var_1088, x = var_1087)[name = tensor("op_1089")]; - tensor cross_9 = mul(x = var_1086, y = var_1089)[name = tensor("cross_9")]; + tensor v_9 = transpose(perm = v_9_perm_0, x = var_1079)[name = tensor("transpose_25")]; + tensor inner_9 = matmul(transpose_x = inner_9_transpose_x_0, transpose_y = inner_9_transpose_y_0, x = var_1097, y = v_9)[name = tensor("inner_9")]; + tensor var_1099_transpose_x_0 = const()[name = tensor("op_1099_transpose_x_0"), val = tensor(false)]; + tensor var_1099_transpose_y_0 = const()[name = tensor("op_1099_transpose_y_0"), val = tensor(false)]; + tensor var_1099 = matmul(transpose_x = var_1099_transpose_x_0, transpose_y = var_1099_transpose_y_0, x = q_9, y = prev_kv_9)[name = tensor("op_1099")]; + tensor var_1100 = real_div(x = sqrt_s0_9, y = sqrt_s_t_9)[name = tensor("op_1100")]; + tensor var_1101 = const()[name = tensor("op_1101"), val = tensor([1, 1, 2, 1])]; + tensor var_1102 = reshape(shape = var_1101, x = var_1100)[name = tensor("op_1102")]; + tensor cross_9 = mul(x = var_1099, y = var_1102)[name = tensor("cross_9")]; tensor out_25 = add(x = inner_9, y = cross_9)[name = tensor("out_25")]; - tensor var_1092 = const()[name = tensor("op_1092"), val = tensor([1, 1, 2, 1])]; - tensor var_1093 = reshape(shape = var_1092, x = valid_mask)[name = tensor("op_1093")]; - tensor v_masked_1 = mul(x = v_9, y = var_1093)[name = tensor("v_masked_1")]; - tensor var_1095 = mul(x = prev_kv_9, y = sqrt_s0_9)[name = tensor("op_1095")]; - tensor var_1097_transpose_x_1 = const()[name = tensor("op_1097_transpose_x_1"), val = tensor(true)]; - tensor var_1097_transpose_y_1 = const()[name = tensor("op_1097_transpose_y_1"), val = tensor(false)]; - tensor var_1097 = matmul(transpose_x = var_1097_transpose_x_1, transpose_y = var_1097_transpose_y_1, x = k_9, y = v_masked_1)[name = tensor("op_1097")]; - tensor new_kv_unnorm_9 = add(x = var_1095, y = var_1097)[name = tensor("new_kv_unnorm_9")]; - tensor var_1099_keep_dims_0 = const()[name = tensor("op_1099_keep_dims_0"), val = tensor(false)]; - tensor var_1099 = reduce_sum(keep_dims = var_1099_keep_dims_0, x = valid_mask)[name = tensor("op_1099")]; - tensor var_1100 = const()[name = tensor("op_1100"), val = tensor([1])]; - tensor var_1101 = reshape(shape = var_1100, x = var_1099)[name = tensor("op_1101")]; - tensor new_scale_9 = add(x = prev_scale_9, y = var_1101)[name = tensor("new_scale_9")]; + tensor var_1105 = const()[name = tensor("op_1105"), val = tensor([1, 1, 2, 1])]; + tensor var_1106 = reshape(shape = var_1105, x = valid_mask)[name = tensor("op_1106")]; + tensor v_masked_1 = mul(x = v_9, y = var_1106)[name = tensor("v_masked_1")]; + tensor var_1108 = mul(x = prev_kv_9, y = sqrt_s0_9)[name = tensor("op_1108")]; + tensor var_1110_transpose_x_1 = const()[name = tensor("op_1110_transpose_x_1"), val = tensor(true)]; + tensor var_1110_transpose_y_1 = const()[name = tensor("op_1110_transpose_y_1"), val = tensor(false)]; + tensor var_1110 = matmul(transpose_x = var_1110_transpose_x_1, transpose_y = var_1110_transpose_y_1, x = k_9, y = v_masked_1)[name = tensor("op_1110")]; + tensor new_kv_unnorm_9 = add(x = var_1108, y = var_1110)[name = tensor("new_kv_unnorm_9")]; + tensor var_1112_keep_dims_0 = const()[name = tensor("op_1112_keep_dims_0"), val = tensor(false)]; + tensor var_1112 = reduce_sum(keep_dims = var_1112_keep_dims_0, x = valid_mask)[name = tensor("op_1112")]; + tensor var_1113 = const()[name = tensor("op_1113"), val = tensor([1])]; + tensor var_1114 = reshape(shape = var_1113, x = var_1112)[name = tensor("op_1114")]; + tensor new_scale_9 = add(x = prev_scale_9, y = var_1114)[name = tensor("new_scale_9")]; tensor const_21 = const()[name = tensor("const_21"), val = tensor(0x1.fffffep+127)]; - tensor clip_2 = clip(alpha = var_958, beta = const_21, x = new_scale_9)[name = tensor("clip_2")]; + tensor clip_2 = clip(alpha = var_39, beta = const_21, x = new_scale_9)[name = tensor("clip_2")]; tensor sqrt_new_scale_1 = sqrt(x = clip_2)[name = tensor("sqrt_new_scale_1")]; - tensor var_1105 = real_div(x = new_kv_unnorm_9, y = sqrt_new_scale_1)[name = tensor("op_1105")]; - tensor var_1106_perm_0 = const()[name = tensor("op_1106_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_1118 = real_div(x = new_kv_unnorm_9, y = sqrt_new_scale_1)[name = tensor("op_1118")]; + tensor var_1119_perm_0 = const()[name = tensor("op_1119_perm_0"), val = tensor([0, 2, 1, 3])]; tensor out_27_axes_0 = const()[name = tensor("out_27_axes_0"), val = tensor([-1])]; - tensor var_1106 = transpose(perm = var_1106_perm_0, x = out_25)[name = tensor("transpose_24")]; - tensor out_27 = layer_norm(axes = out_27_axes_0, epsilon = var_955, x = var_1106)[name = tensor("out_27")]; - tensor var_1110 = const()[name = tensor("op_1110"), val = tensor([12, 2, 256])]; - tensor out_29 = reshape(shape = var_1110, x = out_27)[name = tensor("out_29")]; - tensor var_1112 = silu(x = input_169)[name = tensor("op_1112")]; - tensor input_171 = mul(x = var_1112, y = out_29)[name = tensor("input_171")]; - tensor ret_out_9 = linear(bias = decoder_out_proj_0_bias, weight = decoder_out_proj_0_weight, x = input_171)[name = tensor("linear_42")]; - tensor input_173 = add(x = x_29, y = ret_out_9)[name = tensor("input_173")]; + tensor var_1119 = transpose(perm = var_1119_perm_0, x = out_25)[name = tensor("transpose_24")]; + tensor out_27 = layer_norm(axes = out_27_axes_0, epsilon = var_54, x = var_1119)[name = tensor("out_27")]; + tensor var_1123 = const()[name = tensor("op_1123"), val = tensor([12, 2, 256])]; + tensor out_29 = reshape(shape = var_1123, x = out_27)[name = tensor("out_29")]; + tensor var_1125 = silu(x = input_171)[name = tensor("op_1125")]; + tensor input_173 = mul(x = var_1125, y = out_29)[name = tensor("input_173")]; + tensor ret_out_9 = linear(bias = inner_decoder_out_proj_0_bias, weight = inner_decoder_out_proj_0_weight, x = input_173)[name = tensor("linear_42")]; + tensor input_175 = add(x = x_29, y = ret_out_9)[name = tensor("input_175")]; tensor xt_1_axes_0 = const()[name = tensor("xt_1_axes_0"), val = tensor([-1])]; - tensor xt_1 = layer_norm(axes = xt_1_axes_0, beta = decoder_norm11_0_bias, epsilon = var_953, gamma = decoder_norm11_0_weight, x = input_173)[name = tensor("xt_1")]; - tensor var_1122 = const()[name = tensor("op_1122"), val = tensor([1, 12, 2, 256])]; - tensor var_1123 = reshape(shape = var_1122, x = xt_1)[name = tensor("op_1123")]; - tensor var_1124_perm_0 = const()[name = tensor("op_1124_perm_0"), val = tensor([0, 2, 1, 3])]; - tensor var_1127 = const()[name = tensor("op_1127"), val = tensor([2, 12, 256])]; - tensor var_1124 = transpose(perm = var_1124_perm_0, x = var_1123)[name = tensor("transpose_23")]; - tensor query_1 = reshape(shape = var_1127, x = var_1124)[name = tensor("query_1")]; + tensor xt_1 = layer_norm(axes = xt_1_axes_0, beta = inner_decoder_norm11_0_bias, epsilon = var_46, gamma = inner_decoder_norm11_0_weight, x = input_175)[name = tensor("xt_1")]; + tensor var_1135 = const()[name = tensor("op_1135"), val = tensor([1, 12, 2, 256])]; + tensor var_1136 = reshape(shape = var_1135, x = xt_1)[name = tensor("op_1136")]; + tensor var_1137_perm_0 = const()[name = tensor("op_1137_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_1140 = const()[name = tensor("op_1140"), val = tensor([2, 12, 256])]; + tensor var_1137 = transpose(perm = var_1137_perm_0, x = var_1136)[name = tensor("transpose_23")]; + tensor query_1 = reshape(shape = var_1140, x = var_1137)[name = tensor("query_1")]; tensor query_3_perm_0 = const()[name = tensor("query_3_perm_0"), val = tensor([1, 0, 2])]; tensor query_3 = transpose(perm = query_3_perm_0, x = query_1)[name = tensor("transpose_22")]; - tensor var_1150 = linear(bias = decoder_self_attn2_0_in_proj_bias, weight = decoder_self_attn2_0_in_proj_weight, x = query_3)[name = tensor("linear_43")]; + tensor var_1163 = linear(bias = inner_decoder_self_attn2_0_in_proj_bias, weight = inner_decoder_self_attn2_0_in_proj_weight, x = query_3)[name = tensor("linear_43")]; tensor concat_1 = const()[name = tensor("concat_1"), val = tensor([12, 2, 3, 256])]; - tensor var_1152 = reshape(shape = concat_1, x = var_1150)[name = tensor("op_1152")]; - tensor var_1153_axes_0 = const()[name = tensor("op_1153_axes_0"), val = tensor([0])]; - tensor var_1153 = expand_dims(axes = var_1153_axes_0, x = var_1152)[name = tensor("op_1153")]; - tensor var_1154_perm_0 = const()[name = tensor("op_1154_perm_0"), val = tensor([-2, 1, 2, 0, 4])]; - tensor var_1155_axes_0 = const()[name = tensor("op_1155_axes_0"), val = tensor([-2])]; - tensor var_1154 = transpose(perm = var_1154_perm_0, x = var_1153)[name = tensor("transpose_21")]; - tensor var_1155 = squeeze(axes = var_1155_axes_0, x = var_1154)[name = tensor("op_1155")]; + tensor var_1165 = reshape(shape = concat_1, x = var_1163)[name = tensor("op_1165")]; + tensor var_1166_axes_0 = const()[name = tensor("op_1166_axes_0"), val = tensor([0])]; + tensor var_1166 = expand_dims(axes = var_1166_axes_0, x = var_1165)[name = tensor("op_1166")]; + tensor var_1167_perm_0 = const()[name = tensor("op_1167_perm_0"), val = tensor([-2, 1, 2, 0, 4])]; + tensor var_1168_axes_0 = const()[name = tensor("op_1168_axes_0"), val = tensor([-2])]; + tensor var_1167 = transpose(perm = var_1167_perm_0, x = var_1166)[name = tensor("transpose_21")]; + tensor var_1168 = squeeze(axes = var_1168_axes_0, x = var_1167)[name = tensor("op_1168")]; tensor q_11_begin_0 = const()[name = tensor("q_11_begin_0"), val = tensor([0, 0, 0, 0])]; tensor q_11_end_0 = const()[name = tensor("q_11_end_0"), val = tensor([1, 12, 2, 256])]; tensor q_11_end_mask_0 = const()[name = tensor("q_11_end_mask_0"), val = tensor([false, true, true, true])]; tensor q_11_squeeze_mask_0 = const()[name = tensor("q_11_squeeze_mask_0"), val = tensor([true, false, false, false])]; - tensor q_11 = slice_by_index(begin = q_11_begin_0, end = q_11_end_0, end_mask = q_11_end_mask_0, squeeze_mask = q_11_squeeze_mask_0, x = var_1155)[name = tensor("q_11")]; + tensor q_11 = slice_by_index(begin = q_11_begin_0, end = q_11_end_0, end_mask = q_11_end_mask_0, squeeze_mask = q_11_squeeze_mask_0, x = var_1168)[name = tensor("q_11")]; tensor k_11_begin_0 = const()[name = tensor("k_11_begin_0"), val = tensor([1, 0, 0, 0])]; tensor k_11_end_0 = const()[name = tensor("k_11_end_0"), val = tensor([2, 12, 2, 256])]; tensor k_11_end_mask_0 = const()[name = tensor("k_11_end_mask_0"), val = tensor([false, true, true, true])]; tensor k_11_squeeze_mask_0 = const()[name = tensor("k_11_squeeze_mask_0"), val = tensor([true, false, false, false])]; - tensor k_11 = slice_by_index(begin = k_11_begin_0, end = k_11_end_0, end_mask = k_11_end_mask_0, squeeze_mask = k_11_squeeze_mask_0, x = var_1155)[name = tensor("k_11")]; + tensor k_11 = slice_by_index(begin = k_11_begin_0, end = k_11_end_0, end_mask = k_11_end_mask_0, squeeze_mask = k_11_squeeze_mask_0, x = var_1168)[name = tensor("k_11")]; tensor v_11_begin_0 = const()[name = tensor("v_11_begin_0"), val = tensor([2, 0, 0, 0])]; tensor v_11_end_0 = const()[name = tensor("v_11_end_0"), val = tensor([3, 12, 2, 256])]; tensor v_11_end_mask_0 = const()[name = tensor("v_11_end_mask_0"), val = tensor([false, true, true, true])]; tensor v_11_squeeze_mask_0 = const()[name = tensor("v_11_squeeze_mask_0"), val = tensor([true, false, false, false])]; - tensor v_11 = slice_by_index(begin = v_11_begin_0, end = v_11_end_0, end_mask = v_11_end_mask_0, squeeze_mask = v_11_squeeze_mask_0, x = var_1155)[name = tensor("v_11")]; - tensor var_1163 = const()[name = tensor("op_1163"), val = tensor([12, 8, 64])]; - tensor var_1164 = reshape(shape = var_1163, x = q_11)[name = tensor("op_1164")]; + tensor v_11 = slice_by_index(begin = v_11_begin_0, end = v_11_end_0, end_mask = v_11_end_mask_0, squeeze_mask = v_11_squeeze_mask_0, x = var_1168)[name = tensor("v_11")]; + tensor var_1176 = const()[name = tensor("op_1176"), val = tensor([12, 8, 64])]; + tensor var_1177 = reshape(shape = var_1176, x = q_11)[name = tensor("op_1177")]; tensor q_13_perm_0 = const()[name = tensor("q_13_perm_0"), val = tensor([1, 0, 2])]; - tensor var_1170 = const()[name = tensor("op_1170"), val = tensor([12, 8, 64])]; - tensor var_1171 = reshape(shape = var_1170, x = k_11)[name = tensor("op_1171")]; + tensor var_1183 = const()[name = tensor("op_1183"), val = tensor([12, 8, 64])]; + tensor var_1184 = reshape(shape = var_1183, x = k_11)[name = tensor("op_1184")]; tensor k_13_perm_0 = const()[name = tensor("k_13_perm_0"), val = tensor([1, 0, 2])]; - tensor var_1177 = const()[name = tensor("op_1177"), val = tensor([12, 8, 64])]; - tensor var_1178 = reshape(shape = var_1177, x = v_11)[name = tensor("op_1178")]; + tensor var_1190 = const()[name = tensor("op_1190"), val = tensor([12, 8, 64])]; + tensor var_1191 = reshape(shape = var_1190, x = v_11)[name = tensor("op_1191")]; tensor v_13_perm_0 = const()[name = tensor("v_13_perm_0"), val = tensor([1, 0, 2])]; - tensor var_1181 = const()[name = tensor("op_1181"), val = tensor([2, 4, 12, 64])]; - tensor q_13 = transpose(perm = q_13_perm_0, x = var_1164)[name = tensor("transpose_20")]; - tensor q_15 = reshape(shape = var_1181, x = q_13)[name = tensor("q_15")]; - tensor var_1183 = const()[name = tensor("op_1183"), val = tensor([2, 4, 12, 64])]; - tensor k_13 = transpose(perm = k_13_perm_0, x = var_1171)[name = tensor("transpose_19")]; - tensor k_15 = reshape(shape = var_1183, x = k_13)[name = tensor("k_15")]; - tensor var_1185 = const()[name = tensor("op_1185"), val = tensor([2, 4, 12, 64])]; - tensor v_13 = transpose(perm = v_13_perm_0, x = var_1178)[name = tensor("transpose_18")]; - tensor v_15 = reshape(shape = var_1185, x = v_13)[name = tensor("v_15")]; + tensor var_1194 = const()[name = tensor("op_1194"), val = tensor([2, 4, 12, 64])]; + tensor q_13 = transpose(perm = q_13_perm_0, x = var_1177)[name = tensor("transpose_20")]; + tensor q_15 = reshape(shape = var_1194, x = q_13)[name = tensor("q_15")]; + tensor var_1196 = const()[name = tensor("op_1196"), val = tensor([2, 4, 12, 64])]; + tensor k_13 = transpose(perm = k_13_perm_0, x = var_1184)[name = tensor("transpose_19")]; + tensor k_15 = reshape(shape = var_1196, x = k_13)[name = tensor("k_15")]; + tensor var_1198 = const()[name = tensor("op_1198"), val = tensor([2, 4, 12, 64])]; + tensor v_13 = transpose(perm = v_13_perm_0, x = var_1191)[name = tensor("transpose_18")]; + tensor v_15 = reshape(shape = var_1198, x = v_13)[name = tensor("v_15")]; tensor mul_1_y_0 = const()[name = tensor("mul_1_y_0"), val = tensor(0x1p-3)]; tensor mul_1 = mul(x = q_15, y = mul_1_y_0)[name = tensor("mul_1")]; tensor matmul_0_transpose_y_0 = const()[name = tensor("matmul_0_transpose_y_0"), val = tensor(true)]; @@ -1039,30 +1046,30 @@ program(1.0) tensor attn_output_1_transpose_x_0 = const()[name = tensor("attn_output_1_transpose_x_0"), val = tensor(false)]; tensor attn_output_1_transpose_y_0 = const()[name = tensor("attn_output_1_transpose_y_0"), val = tensor(false)]; tensor attn_output_1 = matmul(transpose_x = attn_output_1_transpose_x_0, transpose_y = attn_output_1_transpose_y_0, x = softmax_0, y = v_15)[name = tensor("attn_output_1")]; - tensor var_1188 = const()[name = tensor("op_1188"), val = tensor([2, 0, 1, 3])]; - tensor var_1193 = const()[name = tensor("op_1193"), val = tensor([24, 256])]; - tensor var_1189 = transpose(perm = var_1188, x = attn_output_1)[name = tensor("transpose_17")]; - tensor attn_output_3 = reshape(shape = var_1193, x = var_1189)[name = tensor("attn_output_3")]; - tensor attn_output_5 = linear(bias = decoder_self_attn2_0_out_proj_bias, weight = decoder_self_attn2_0_out_proj_weight, x = attn_output_3)[name = tensor("linear_44")]; - tensor var_1197 = const()[name = tensor("op_1197"), val = tensor([12, 2, 256])]; - tensor attn_output_7 = reshape(shape = var_1197, x = attn_output_5)[name = tensor("attn_output_7")]; + tensor var_1201 = const()[name = tensor("op_1201"), val = tensor([2, 0, 1, 3])]; + tensor var_1206 = const()[name = tensor("op_1206"), val = tensor([24, 256])]; + tensor var_1202 = transpose(perm = var_1201, x = attn_output_1)[name = tensor("transpose_17")]; + tensor attn_output_3 = reshape(shape = var_1206, x = var_1202)[name = tensor("attn_output_3")]; + tensor attn_output_5 = linear(bias = inner_decoder_self_attn2_0_out_proj_bias, weight = inner_decoder_self_attn2_0_out_proj_weight, x = attn_output_3)[name = tensor("linear_44")]; + tensor var_1210 = const()[name = tensor("op_1210"), val = tensor([12, 2, 256])]; + tensor attn_output_7 = reshape(shape = var_1210, x = attn_output_5)[name = tensor("attn_output_7")]; tensor sa2_out_1_perm_0 = const()[name = tensor("sa2_out_1_perm_0"), val = tensor([1, 0, 2])]; tensor sa2_out_1 = transpose(perm = sa2_out_1_perm_0, x = attn_output_7)[name = tensor("transpose_16")]; - tensor input_175 = add(x = query_1, y = sa2_out_1)[name = tensor("input_175")]; - tensor input_177_axes_0 = const()[name = tensor("input_177_axes_0"), val = tensor([-1])]; - tensor input_177 = layer_norm(axes = input_177_axes_0, beta = decoder_norm21_0_bias, epsilon = var_953, gamma = decoder_norm21_0_weight, x = input_175)[name = tensor("input_177")]; - tensor input_179 = linear(bias = decoder_linear1_0_bias, weight = decoder_linear1_0_weight, x = input_177)[name = tensor("linear_45")]; - tensor input_181 = relu(x = input_179)[name = tensor("input_181")]; - tensor ffn_out_1 = linear(bias = decoder_linear2_0_bias, weight = decoder_linear2_0_weight, x = input_181)[name = tensor("linear_46")]; - tensor input_183 = add(x = input_177, y = ffn_out_1)[name = tensor("input_183")]; + tensor input_177 = add(x = query_1, y = sa2_out_1)[name = tensor("input_177")]; + tensor input_179_axes_0 = const()[name = tensor("input_179_axes_0"), val = tensor([-1])]; + tensor input_179 = layer_norm(axes = input_179_axes_0, beta = inner_decoder_norm21_0_bias, epsilon = var_46, gamma = inner_decoder_norm21_0_weight, x = input_177)[name = tensor("input_179")]; + tensor input_181 = linear(bias = inner_decoder_linear1_0_bias, weight = inner_decoder_linear1_0_weight, x = input_179)[name = tensor("linear_45")]; + tensor input_183 = relu(x = input_181)[name = tensor("input_183")]; + tensor ffn_out_1 = linear(bias = inner_decoder_linear2_0_bias, weight = inner_decoder_linear2_0_weight, x = input_183)[name = tensor("linear_46")]; + tensor input_185 = add(x = input_179, y = ffn_out_1)[name = tensor("input_185")]; tensor xt_3_axes_0 = const()[name = tensor("xt_3_axes_0"), val = tensor([-1])]; - tensor xt_3 = layer_norm(axes = xt_3_axes_0, beta = decoder_norm22_0_bias, epsilon = var_953, gamma = decoder_norm22_0_weight, x = input_183)[name = tensor("xt_3")]; - tensor var_1217 = const()[name = tensor("op_1217"), val = tensor([1, 2, 12, 256])]; - tensor x_31 = reshape(shape = var_1217, x = xt_3)[name = tensor("x_31")]; - tensor var_1219_perm_0 = const()[name = tensor("op_1219_perm_0"), val = tensor([0, 2, 1, 3])]; - tensor var_1223 = const()[name = tensor("op_1223"), val = tensor([12, 2, 256])]; - tensor var_1219 = transpose(perm = var_1219_perm_0, x = x_31)[name = tensor("transpose_15")]; - tensor x = reshape(shape = var_1223, x = var_1219)[name = tensor("x")]; + tensor xt_3 = layer_norm(axes = xt_3_axes_0, beta = inner_decoder_norm22_0_bias, epsilon = var_46, gamma = inner_decoder_norm22_0_weight, x = input_185)[name = tensor("xt_3")]; + tensor var_1230 = const()[name = tensor("op_1230"), val = tensor([1, 2, 12, 256])]; + tensor x_31 = reshape(shape = var_1230, x = xt_3)[name = tensor("x_31")]; + tensor var_1232_perm_0 = const()[name = tensor("op_1232_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_1236 = const()[name = tensor("op_1236"), val = tensor([12, 2, 256])]; + tensor var_1232 = transpose(perm = var_1232_perm_0, x = x_31)[name = tensor("transpose_15")]; + tensor x = reshape(shape = var_1236, x = var_1232)[name = tensor("x")]; tensor prev_kv_begin_0 = const()[name = tensor("prev_kv_begin_0"), val = tensor([1, 0, 0, 0, 0])]; tensor prev_kv_end_0 = const()[name = tensor("prev_kv_end_0"), val = tensor([2, 12, 4, 64, 64])]; tensor prev_kv_end_mask_0 = const()[name = tensor("prev_kv_end_mask_0"), val = tensor([false, true, true, true, true])]; @@ -1073,120 +1080,120 @@ program(1.0) tensor prev_scale_end_mask_0 = const()[name = tensor("prev_scale_end_mask_0"), val = tensor([false, true])]; tensor prev_scale_squeeze_mask_0 = const()[name = tensor("prev_scale_squeeze_mask_0"), val = tensor([true, false])]; tensor prev_scale = slice_by_index(begin = prev_scale_begin_0, end = prev_scale_end_0, end_mask = prev_scale_end_mask_0, squeeze_mask = prev_scale_squeeze_mask_0, x = dec_scale)[name = tensor("prev_scale")]; - tensor var_1231 = linear(bias = decoder_q_proj_1_bias, weight = decoder_q_proj_1_weight, x = x)[name = tensor("linear_47")]; - tensor var_1232 = const()[name = tensor("op_1232"), val = tensor([12, 2, 4, 64])]; - tensor var_1233 = reshape(shape = var_1232, x = var_1231)[name = tensor("op_1233")]; + tensor var_1244 = linear(bias = inner_decoder_q_proj_1_bias, weight = inner_decoder_q_proj_1_weight, x = x)[name = tensor("linear_47")]; + tensor var_1245 = const()[name = tensor("op_1245"), val = tensor([12, 2, 4, 64])]; + tensor var_1246 = reshape(shape = var_1245, x = var_1244)[name = tensor("op_1246")]; tensor q_17_perm_0 = const()[name = tensor("q_17_perm_0"), val = tensor([0, 2, 1, 3])]; - tensor var_1237 = linear(bias = decoder_k_proj_1_bias, weight = decoder_k_proj_1_weight, x = x)[name = tensor("linear_48")]; - tensor var_1238 = const()[name = tensor("op_1238"), val = tensor(0x1p-3)]; - tensor var_1239 = mul(x = var_1237, y = var_1238)[name = tensor("op_1239")]; - tensor var_1240 = const()[name = tensor("op_1240"), val = tensor([12, 2, 4, 64])]; - tensor var_1241 = reshape(shape = var_1240, x = var_1239)[name = tensor("op_1241")]; + tensor var_1250 = linear(bias = inner_decoder_k_proj_1_bias, weight = inner_decoder_k_proj_1_weight, x = x)[name = tensor("linear_48")]; + tensor var_1251 = const()[name = tensor("op_1251"), val = tensor(0x1p-3)]; + tensor var_1252 = mul(x = var_1250, y = var_1251)[name = tensor("op_1252")]; + tensor var_1253 = const()[name = tensor("op_1253"), val = tensor([12, 2, 4, 64])]; + tensor var_1254 = reshape(shape = var_1253, x = var_1252)[name = tensor("op_1254")]; tensor k_17_perm_0 = const()[name = tensor("k_17_perm_0"), val = tensor([0, 2, 1, 3])]; - tensor var_1245 = linear(bias = decoder_v_proj_1_bias, weight = decoder_v_proj_1_weight, x = x)[name = tensor("linear_49")]; - tensor var_1246 = const()[name = tensor("op_1246"), val = tensor([12, 2, 4, 64])]; - tensor var_1247 = reshape(shape = var_1246, x = var_1245)[name = tensor("op_1247")]; + tensor var_1258 = linear(bias = inner_decoder_v_proj_1_bias, weight = inner_decoder_v_proj_1_weight, x = x)[name = tensor("linear_49")]; + tensor var_1259 = const()[name = tensor("op_1259"), val = tensor([12, 2, 4, 64])]; + tensor var_1260 = reshape(shape = var_1259, x = var_1258)[name = tensor("op_1260")]; tensor v_17_perm_0 = const()[name = tensor("v_17_perm_0"), val = tensor([0, 2, 1, 3])]; - tensor input_187 = linear(bias = decoder_g_proj_1_bias, weight = decoder_g_proj_1_weight, x = x)[name = tensor("linear_50")]; + tensor input_189 = linear(bias = inner_decoder_g_proj_1_bias, weight = inner_decoder_g_proj_1_weight, x = x)[name = tensor("linear_50")]; tensor sqrt_s0 = sqrt(x = prev_scale)[name = tensor("sqrt_s0")]; tensor s_t = add(x = prev_scale, y = cumsum_mask_1)[name = tensor("s_t")]; tensor const_32 = const()[name = tensor("const_32"), val = tensor(0x1.fffffep+127)]; - tensor clip_3 = clip(alpha = var_958, beta = const_32, x = s_t)[name = tensor("clip_3")]; + tensor clip_3 = clip(alpha = var_39, beta = const_32, x = s_t)[name = tensor("clip_3")]; tensor sqrt_s_t = sqrt(x = clip_3)[name = tensor("sqrt_s_t")]; tensor qk_transpose_x_1 = const()[name = tensor("qk_transpose_x_1"), val = tensor(false)]; tensor qk_transpose_y_1 = const()[name = tensor("qk_transpose_y_1"), val = tensor(true)]; - tensor k_17 = transpose(perm = k_17_perm_0, x = var_1241)[name = tensor("transpose_13")]; - tensor q_17 = transpose(perm = q_17_perm_0, x = var_1233)[name = tensor("transpose_14")]; + tensor k_17 = transpose(perm = k_17_perm_0, x = var_1254)[name = tensor("transpose_13")]; + tensor q_17 = transpose(perm = q_17_perm_0, x = var_1246)[name = tensor("transpose_14")]; tensor qk = matmul(transpose_x = qk_transpose_x_1, transpose_y = qk_transpose_y_1, x = q_17, y = k_17)[name = tensor("qk")]; - tensor var_1262 = const()[name = tensor("op_1262"), val = tensor([2, 1])]; - tensor var_1263 = reshape(shape = var_1262, x = sqrt_s_t)[name = tensor("op_1263")]; - tensor M = real_div(x = causal_with_valid_1, y = var_1263)[name = tensor("M")]; - tensor var_1265 = mul(x = qk, y = M)[name = tensor("op_1265")]; - tensor inner_transpose_x_0 = const()[name = tensor("inner_transpose_x_0"), val = tensor(false)]; - tensor inner_transpose_y_0 = const()[name = tensor("inner_transpose_y_0"), val = tensor(false)]; - tensor v_17 = transpose(perm = v_17_perm_0, x = var_1247)[name = tensor("transpose_12")]; - tensor inner = matmul(transpose_x = inner_transpose_x_0, transpose_y = inner_transpose_y_0, x = var_1265, y = v_17)[name = tensor("inner")]; - tensor var_1267_transpose_x_0 = const()[name = tensor("op_1267_transpose_x_0"), val = tensor(false)]; - tensor var_1267_transpose_y_0 = const()[name = tensor("op_1267_transpose_y_0"), val = tensor(false)]; - tensor var_1267 = matmul(transpose_x = var_1267_transpose_x_0, transpose_y = var_1267_transpose_y_0, x = q_17, y = prev_kv)[name = tensor("op_1267")]; - tensor var_1268 = real_div(x = sqrt_s0, y = sqrt_s_t)[name = tensor("op_1268")]; - tensor var_1269 = const()[name = tensor("op_1269"), val = tensor([1, 1, 2, 1])]; - tensor var_1270 = reshape(shape = var_1269, x = var_1268)[name = tensor("op_1270")]; - tensor cross = mul(x = var_1267, y = var_1270)[name = tensor("cross")]; - tensor out_31 = add(x = inner, y = cross)[name = tensor("out_31")]; - tensor v_masked = mul(x = v_17, y = var_1093)[name = tensor("v_masked")]; - tensor var_1276 = mul(x = prev_kv, y = sqrt_s0)[name = tensor("op_1276")]; - tensor var_1278_transpose_x_1 = const()[name = tensor("op_1278_transpose_x_1"), val = tensor(true)]; - tensor var_1278_transpose_y_1 = const()[name = tensor("op_1278_transpose_y_1"), val = tensor(false)]; - tensor var_1278 = matmul(transpose_x = var_1278_transpose_x_1, transpose_y = var_1278_transpose_y_1, x = k_17, y = v_masked)[name = tensor("op_1278")]; - tensor new_kv_unnorm = add(x = var_1276, y = var_1278)[name = tensor("new_kv_unnorm")]; - tensor new_scale = add(x = prev_scale, y = var_1101)[name = tensor("new_scale")]; + tensor var_1275 = const()[name = tensor("op_1275"), val = tensor([2, 1])]; + tensor var_1276 = reshape(shape = var_1275, x = sqrt_s_t)[name = tensor("op_1276")]; + tensor M = real_div(x = causal_with_valid_1, y = var_1276)[name = tensor("M")]; + tensor var_1278 = mul(x = qk, y = M)[name = tensor("op_1278")]; + tensor inner_11_transpose_x_0 = const()[name = tensor("inner_11_transpose_x_0"), val = tensor(false)]; + tensor inner_11_transpose_y_0 = const()[name = tensor("inner_11_transpose_y_0"), val = tensor(false)]; + tensor v_17 = transpose(perm = v_17_perm_0, x = var_1260)[name = tensor("transpose_12")]; + tensor inner_11 = matmul(transpose_x = inner_11_transpose_x_0, transpose_y = inner_11_transpose_y_0, x = var_1278, y = v_17)[name = tensor("inner_11")]; + tensor var_1280_transpose_x_0 = const()[name = tensor("op_1280_transpose_x_0"), val = tensor(false)]; + tensor var_1280_transpose_y_0 = const()[name = tensor("op_1280_transpose_y_0"), val = tensor(false)]; + tensor var_1280 = matmul(transpose_x = var_1280_transpose_x_0, transpose_y = var_1280_transpose_y_0, x = q_17, y = prev_kv)[name = tensor("op_1280")]; + tensor var_1281 = real_div(x = sqrt_s0, y = sqrt_s_t)[name = tensor("op_1281")]; + tensor var_1282 = const()[name = tensor("op_1282"), val = tensor([1, 1, 2, 1])]; + tensor var_1283 = reshape(shape = var_1282, x = var_1281)[name = tensor("op_1283")]; + tensor cross = mul(x = var_1280, y = var_1283)[name = tensor("cross")]; + tensor out_31 = add(x = inner_11, y = cross)[name = tensor("out_31")]; + tensor v_masked = mul(x = v_17, y = var_1106)[name = tensor("v_masked")]; + tensor var_1289 = mul(x = prev_kv, y = sqrt_s0)[name = tensor("op_1289")]; + tensor var_1291_transpose_x_1 = const()[name = tensor("op_1291_transpose_x_1"), val = tensor(true)]; + tensor var_1291_transpose_y_1 = const()[name = tensor("op_1291_transpose_y_1"), val = tensor(false)]; + tensor var_1291 = matmul(transpose_x = var_1291_transpose_x_1, transpose_y = var_1291_transpose_y_1, x = k_17, y = v_masked)[name = tensor("op_1291")]; + tensor new_kv_unnorm = add(x = var_1289, y = var_1291)[name = tensor("new_kv_unnorm")]; + tensor new_scale = add(x = prev_scale, y = var_1114)[name = tensor("new_scale")]; tensor const_33 = const()[name = tensor("const_33"), val = tensor(0x1.fffffep+127)]; - tensor clip_4 = clip(alpha = var_958, beta = const_33, x = new_scale)[name = tensor("clip_4")]; + tensor clip_4 = clip(alpha = var_39, beta = const_33, x = new_scale)[name = tensor("clip_4")]; tensor sqrt_new_scale = sqrt(x = clip_4)[name = tensor("sqrt_new_scale")]; tensor nkv = real_div(x = new_kv_unnorm, y = sqrt_new_scale)[name = tensor("nkv")]; - tensor var_1287_perm_0 = const()[name = tensor("op_1287_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_1300_perm_0 = const()[name = tensor("op_1300_perm_0"), val = tensor([0, 2, 1, 3])]; tensor out_33_axes_0 = const()[name = tensor("out_33_axes_0"), val = tensor([-1])]; - tensor var_1287 = transpose(perm = var_1287_perm_0, x = out_31)[name = tensor("transpose_11")]; - tensor out_33 = layer_norm(axes = out_33_axes_0, epsilon = var_955, x = var_1287)[name = tensor("out_33")]; - tensor var_1291 = const()[name = tensor("op_1291"), val = tensor([12, 2, 256])]; - tensor out = reshape(shape = var_1291, x = out_33)[name = tensor("out")]; - tensor var_1293 = silu(x = input_187)[name = tensor("op_1293")]; - tensor input_189 = mul(x = var_1293, y = out)[name = tensor("input_189")]; - tensor ret_out = linear(bias = decoder_out_proj_1_bias, weight = decoder_out_proj_1_weight, x = input_189)[name = tensor("linear_51")]; - tensor input_191 = add(x = x, y = ret_out)[name = tensor("input_191")]; + tensor var_1300 = transpose(perm = var_1300_perm_0, x = out_31)[name = tensor("transpose_11")]; + tensor out_33 = layer_norm(axes = out_33_axes_0, epsilon = var_54, x = var_1300)[name = tensor("out_33")]; + tensor var_1304 = const()[name = tensor("op_1304"), val = tensor([12, 2, 256])]; + tensor out = reshape(shape = var_1304, x = out_33)[name = tensor("out")]; + tensor var_1306 = silu(x = input_189)[name = tensor("op_1306")]; + tensor input_191 = mul(x = var_1306, y = out)[name = tensor("input_191")]; + tensor ret_out = linear(bias = inner_decoder_out_proj_1_bias, weight = inner_decoder_out_proj_1_weight, x = input_191)[name = tensor("linear_51")]; + tensor input_193 = add(x = x, y = ret_out)[name = tensor("input_193")]; tensor xt_5_axes_0 = const()[name = tensor("xt_5_axes_0"), val = tensor([-1])]; - tensor xt_5 = layer_norm(axes = xt_5_axes_0, beta = decoder_norm11_1_bias, epsilon = var_953, gamma = decoder_norm11_1_weight, x = input_191)[name = tensor("xt_5")]; - tensor var_1303 = const()[name = tensor("op_1303"), val = tensor([1, 12, 2, 256])]; - tensor var_1304 = reshape(shape = var_1303, x = xt_5)[name = tensor("op_1304")]; - tensor var_1305_perm_0 = const()[name = tensor("op_1305_perm_0"), val = tensor([0, 2, 1, 3])]; - tensor var_1308 = const()[name = tensor("op_1308"), val = tensor([2, 12, 256])]; - tensor var_1305 = transpose(perm = var_1305_perm_0, x = var_1304)[name = tensor("transpose_10")]; - tensor query_5 = reshape(shape = var_1308, x = var_1305)[name = tensor("query_5")]; + tensor xt_5 = layer_norm(axes = xt_5_axes_0, beta = inner_decoder_norm11_1_bias, epsilon = var_46, gamma = inner_decoder_norm11_1_weight, x = input_193)[name = tensor("xt_5")]; + tensor var_1316 = const()[name = tensor("op_1316"), val = tensor([1, 12, 2, 256])]; + tensor var_1317 = reshape(shape = var_1316, x = xt_5)[name = tensor("op_1317")]; + tensor var_1318_perm_0 = const()[name = tensor("op_1318_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_1321 = const()[name = tensor("op_1321"), val = tensor([2, 12, 256])]; + tensor var_1318 = transpose(perm = var_1318_perm_0, x = var_1317)[name = tensor("transpose_10")]; + tensor query_5 = reshape(shape = var_1321, x = var_1318)[name = tensor("query_5")]; tensor query_perm_0 = const()[name = tensor("query_perm_0"), val = tensor([1, 0, 2])]; tensor query = transpose(perm = query_perm_0, x = query_5)[name = tensor("transpose_9")]; - tensor var_1331 = linear(bias = decoder_self_attn2_1_in_proj_bias, weight = decoder_self_attn2_1_in_proj_weight, x = query)[name = tensor("linear_52")]; + tensor var_1344 = linear(bias = inner_decoder_self_attn2_1_in_proj_bias, weight = inner_decoder_self_attn2_1_in_proj_weight, x = query)[name = tensor("linear_52")]; tensor concat_2 = const()[name = tensor("concat_2"), val = tensor([12, 2, 3, 256])]; - tensor var_1333 = reshape(shape = concat_2, x = var_1331)[name = tensor("op_1333")]; - tensor var_1334_axes_0 = const()[name = tensor("op_1334_axes_0"), val = tensor([0])]; - tensor var_1334 = expand_dims(axes = var_1334_axes_0, x = var_1333)[name = tensor("op_1334")]; - tensor var_1335_perm_0 = const()[name = tensor("op_1335_perm_0"), val = tensor([-2, 1, 2, 0, 4])]; - tensor var_1336_axes_0 = const()[name = tensor("op_1336_axes_0"), val = tensor([-2])]; - tensor var_1335 = transpose(perm = var_1335_perm_0, x = var_1334)[name = tensor("transpose_8")]; - tensor var_1336 = squeeze(axes = var_1336_axes_0, x = var_1335)[name = tensor("op_1336")]; + tensor var_1346 = reshape(shape = concat_2, x = var_1344)[name = tensor("op_1346")]; + tensor var_1347_axes_0 = const()[name = tensor("op_1347_axes_0"), val = tensor([0])]; + tensor var_1347 = expand_dims(axes = var_1347_axes_0, x = var_1346)[name = tensor("op_1347")]; + tensor var_1348_perm_0 = const()[name = tensor("op_1348_perm_0"), val = tensor([-2, 1, 2, 0, 4])]; + tensor var_1349_axes_0 = const()[name = tensor("op_1349_axes_0"), val = tensor([-2])]; + tensor var_1348 = transpose(perm = var_1348_perm_0, x = var_1347)[name = tensor("transpose_8")]; + tensor var_1349 = squeeze(axes = var_1349_axes_0, x = var_1348)[name = tensor("op_1349")]; tensor q_19_begin_0 = const()[name = tensor("q_19_begin_0"), val = tensor([0, 0, 0, 0])]; tensor q_19_end_0 = const()[name = tensor("q_19_end_0"), val = tensor([1, 12, 2, 256])]; tensor q_19_end_mask_0 = const()[name = tensor("q_19_end_mask_0"), val = tensor([false, true, true, true])]; tensor q_19_squeeze_mask_0 = const()[name = tensor("q_19_squeeze_mask_0"), val = tensor([true, false, false, false])]; - tensor q_19 = slice_by_index(begin = q_19_begin_0, end = q_19_end_0, end_mask = q_19_end_mask_0, squeeze_mask = q_19_squeeze_mask_0, x = var_1336)[name = tensor("q_19")]; + tensor q_19 = slice_by_index(begin = q_19_begin_0, end = q_19_end_0, end_mask = q_19_end_mask_0, squeeze_mask = q_19_squeeze_mask_0, x = var_1349)[name = tensor("q_19")]; tensor k_19_begin_0 = const()[name = tensor("k_19_begin_0"), val = tensor([1, 0, 0, 0])]; tensor k_19_end_0 = const()[name = tensor("k_19_end_0"), val = tensor([2, 12, 2, 256])]; tensor k_19_end_mask_0 = const()[name = tensor("k_19_end_mask_0"), val = tensor([false, true, true, true])]; tensor k_19_squeeze_mask_0 = const()[name = tensor("k_19_squeeze_mask_0"), val = tensor([true, false, false, false])]; - tensor k_19 = slice_by_index(begin = k_19_begin_0, end = k_19_end_0, end_mask = k_19_end_mask_0, squeeze_mask = k_19_squeeze_mask_0, x = var_1336)[name = tensor("k_19")]; + tensor k_19 = slice_by_index(begin = k_19_begin_0, end = k_19_end_0, end_mask = k_19_end_mask_0, squeeze_mask = k_19_squeeze_mask_0, x = var_1349)[name = tensor("k_19")]; tensor v_19_begin_0 = const()[name = tensor("v_19_begin_0"), val = tensor([2, 0, 0, 0])]; tensor v_19_end_0 = const()[name = tensor("v_19_end_0"), val = tensor([3, 12, 2, 256])]; tensor v_19_end_mask_0 = const()[name = tensor("v_19_end_mask_0"), val = tensor([false, true, true, true])]; tensor v_19_squeeze_mask_0 = const()[name = tensor("v_19_squeeze_mask_0"), val = tensor([true, false, false, false])]; - tensor v_19 = slice_by_index(begin = v_19_begin_0, end = v_19_end_0, end_mask = v_19_end_mask_0, squeeze_mask = v_19_squeeze_mask_0, x = var_1336)[name = tensor("v_19")]; - tensor var_1344 = const()[name = tensor("op_1344"), val = tensor([12, 8, 64])]; - tensor var_1345 = reshape(shape = var_1344, x = q_19)[name = tensor("op_1345")]; + tensor v_19 = slice_by_index(begin = v_19_begin_0, end = v_19_end_0, end_mask = v_19_end_mask_0, squeeze_mask = v_19_squeeze_mask_0, x = var_1349)[name = tensor("v_19")]; + tensor var_1357 = const()[name = tensor("op_1357"), val = tensor([12, 8, 64])]; + tensor var_1358 = reshape(shape = var_1357, x = q_19)[name = tensor("op_1358")]; tensor q_21_perm_0 = const()[name = tensor("q_21_perm_0"), val = tensor([1, 0, 2])]; - tensor var_1351 = const()[name = tensor("op_1351"), val = tensor([12, 8, 64])]; - tensor var_1352 = reshape(shape = var_1351, x = k_19)[name = tensor("op_1352")]; + tensor var_1364 = const()[name = tensor("op_1364"), val = tensor([12, 8, 64])]; + tensor var_1365 = reshape(shape = var_1364, x = k_19)[name = tensor("op_1365")]; tensor k_21_perm_0 = const()[name = tensor("k_21_perm_0"), val = tensor([1, 0, 2])]; - tensor var_1358 = const()[name = tensor("op_1358"), val = tensor([12, 8, 64])]; - tensor var_1359 = reshape(shape = var_1358, x = v_19)[name = tensor("op_1359")]; + tensor var_1371 = const()[name = tensor("op_1371"), val = tensor([12, 8, 64])]; + tensor var_1372 = reshape(shape = var_1371, x = v_19)[name = tensor("op_1372")]; tensor v_21_perm_0 = const()[name = tensor("v_21_perm_0"), val = tensor([1, 0, 2])]; - tensor var_1362 = const()[name = tensor("op_1362"), val = tensor([2, 4, 12, 64])]; - tensor q_21 = transpose(perm = q_21_perm_0, x = var_1345)[name = tensor("transpose_7")]; - tensor q = reshape(shape = var_1362, x = q_21)[name = tensor("q")]; - tensor var_1364 = const()[name = tensor("op_1364"), val = tensor([2, 4, 12, 64])]; - tensor k_21 = transpose(perm = k_21_perm_0, x = var_1352)[name = tensor("transpose_6")]; - tensor k = reshape(shape = var_1364, x = k_21)[name = tensor("k")]; - tensor var_1366 = const()[name = tensor("op_1366"), val = tensor([2, 4, 12, 64])]; - tensor v_21 = transpose(perm = v_21_perm_0, x = var_1359)[name = tensor("transpose_5")]; - tensor v = reshape(shape = var_1366, x = v_21)[name = tensor("v")]; + tensor var_1375 = const()[name = tensor("op_1375"), val = tensor([2, 4, 12, 64])]; + tensor q_21 = transpose(perm = q_21_perm_0, x = var_1358)[name = tensor("transpose_7")]; + tensor q = reshape(shape = var_1375, x = q_21)[name = tensor("q")]; + tensor var_1377 = const()[name = tensor("op_1377"), val = tensor([2, 4, 12, 64])]; + tensor k_21 = transpose(perm = k_21_perm_0, x = var_1365)[name = tensor("transpose_6")]; + tensor k = reshape(shape = var_1377, x = k_21)[name = tensor("k")]; + tensor var_1379 = const()[name = tensor("op_1379"), val = tensor([2, 4, 12, 64])]; + tensor v_21 = transpose(perm = v_21_perm_0, x = var_1372)[name = tensor("transpose_5")]; + tensor v = reshape(shape = var_1379, x = v_21)[name = tensor("v")]; tensor mul_3_y_0 = const()[name = tensor("mul_3_y_0"), val = tensor(0x1p-3)]; tensor mul_3 = mul(x = q, y = mul_3_y_0)[name = tensor("mul_3")]; tensor matmul_1_transpose_y_0 = const()[name = tensor("matmul_1_transpose_y_0"), val = tensor(true)]; @@ -1197,36 +1204,36 @@ program(1.0) tensor attn_output_9_transpose_x_0 = const()[name = tensor("attn_output_9_transpose_x_0"), val = tensor(false)]; tensor attn_output_9_transpose_y_0 = const()[name = tensor("attn_output_9_transpose_y_0"), val = tensor(false)]; tensor attn_output_9 = matmul(transpose_x = attn_output_9_transpose_x_0, transpose_y = attn_output_9_transpose_y_0, x = softmax_1, y = v)[name = tensor("attn_output_9")]; - tensor var_1369 = const()[name = tensor("op_1369"), val = tensor([2, 0, 1, 3])]; - tensor var_1374 = const()[name = tensor("op_1374"), val = tensor([24, 256])]; - tensor var_1370 = transpose(perm = var_1369, x = attn_output_9)[name = tensor("transpose_4")]; - tensor attn_output_11 = reshape(shape = var_1374, x = var_1370)[name = tensor("attn_output_11")]; - tensor attn_output_13 = linear(bias = decoder_self_attn2_1_out_proj_bias, weight = decoder_self_attn2_1_out_proj_weight, x = attn_output_11)[name = tensor("linear_53")]; - tensor var_1378 = const()[name = tensor("op_1378"), val = tensor([12, 2, 256])]; - tensor attn_output = reshape(shape = var_1378, x = attn_output_13)[name = tensor("attn_output")]; + tensor var_1382 = const()[name = tensor("op_1382"), val = tensor([2, 0, 1, 3])]; + tensor var_1387 = const()[name = tensor("op_1387"), val = tensor([24, 256])]; + tensor var_1383 = transpose(perm = var_1382, x = attn_output_9)[name = tensor("transpose_4")]; + tensor attn_output_11 = reshape(shape = var_1387, x = var_1383)[name = tensor("attn_output_11")]; + tensor attn_output_13 = linear(bias = inner_decoder_self_attn2_1_out_proj_bias, weight = inner_decoder_self_attn2_1_out_proj_weight, x = attn_output_11)[name = tensor("linear_53")]; + tensor var_1391 = const()[name = tensor("op_1391"), val = tensor([12, 2, 256])]; + tensor attn_output = reshape(shape = var_1391, x = attn_output_13)[name = tensor("attn_output")]; tensor sa2_out_perm_0 = const()[name = tensor("sa2_out_perm_0"), val = tensor([1, 0, 2])]; tensor sa2_out = transpose(perm = sa2_out_perm_0, x = attn_output)[name = tensor("transpose_3")]; - tensor input_193 = add(x = query_5, y = sa2_out)[name = tensor("input_193")]; - tensor input_195_axes_0 = const()[name = tensor("input_195_axes_0"), val = tensor([-1])]; - tensor input_195 = layer_norm(axes = input_195_axes_0, beta = decoder_norm21_1_bias, epsilon = var_953, gamma = decoder_norm21_1_weight, x = input_193)[name = tensor("input_195")]; - tensor input_197 = linear(bias = decoder_linear1_1_bias, weight = decoder_linear1_1_weight, x = input_195)[name = tensor("linear_54")]; - tensor input_199 = relu(x = input_197)[name = tensor("input_199")]; - tensor ffn_out = linear(bias = decoder_linear2_1_bias, weight = decoder_linear2_1_weight, x = input_199)[name = tensor("linear_55")]; - tensor input_201 = add(x = input_195, y = ffn_out)[name = tensor("input_201")]; + tensor input_195 = add(x = query_5, y = sa2_out)[name = tensor("input_195")]; + tensor input_197_axes_0 = const()[name = tensor("input_197_axes_0"), val = tensor([-1])]; + tensor input_197 = layer_norm(axes = input_197_axes_0, beta = inner_decoder_norm21_1_bias, epsilon = var_46, gamma = inner_decoder_norm21_1_weight, x = input_195)[name = tensor("input_197")]; + tensor input_199 = linear(bias = inner_decoder_linear1_1_bias, weight = inner_decoder_linear1_1_weight, x = input_197)[name = tensor("linear_54")]; + tensor input_201 = relu(x = input_199)[name = tensor("input_201")]; + tensor ffn_out = linear(bias = inner_decoder_linear2_1_bias, weight = inner_decoder_linear2_1_weight, x = input_201)[name = tensor("linear_55")]; + tensor input_203 = add(x = input_197, y = ffn_out)[name = tensor("input_203")]; tensor xt_axes_0 = const()[name = tensor("xt_axes_0"), val = tensor([-1])]; - tensor xt = layer_norm(axes = xt_axes_0, beta = decoder_norm22_1_bias, epsilon = var_953, gamma = decoder_norm22_1_weight, x = input_201)[name = tensor("xt")]; - tensor var_1398 = const()[name = tensor("op_1398"), val = tensor([1, 2, 12, 256])]; - tensor input = reshape(shape = var_1398, x = xt)[name = tensor("input")]; - tensor var_1400 = const()[name = tensor("op_1400"), val = tensor([-1])]; - tensor var_1401 = reduce_l2_norm(axes = var_1400, keep_dims = var_956, x = input)[name = tensor("op_1401")]; + tensor xt = layer_norm(axes = xt_axes_0, beta = inner_decoder_norm22_1_bias, epsilon = var_46, gamma = inner_decoder_norm22_1_weight, x = input_203)[name = tensor("xt")]; + tensor var_1411 = const()[name = tensor("op_1411"), val = tensor([1, 2, 12, 256])]; + tensor input = reshape(shape = var_1411, x = xt)[name = tensor("input")]; + tensor var_1413 = const()[name = tensor("op_1413"), val = tensor([-1])]; + tensor var_1414 = reduce_l2_norm(axes = var_1413, keep_dims = var_45, x = input)[name = tensor("op_1414")]; tensor const_42 = const()[name = tensor("const_42"), val = tensor(0x1.fffffep+127)]; - tensor clip_5 = clip(alpha = var_948, beta = const_42, x = var_1401)[name = tensor("clip_5")]; - tensor var_1403 = real_div(x = input, y = clip_5)[name = tensor("op_1403")]; + tensor clip_5 = clip(alpha = var_59, beta = const_42, x = var_1414)[name = tensor("clip_5")]; + tensor var_1416 = real_div(x = input, y = clip_5)[name = tensor("op_1416")]; tensor concat_6 = const()[name = tensor("concat_6"), val = tensor([2, 1, 256])]; tensor reshape_0 = reshape(shape = concat_6, x = emb)[name = tensor("reshape_0")]; tensor transpose_1_perm_0 = const()[name = tensor("transpose_1_perm_0"), val = tensor([0, 1, 3, 2])]; tensor concat_7 = const()[name = tensor("concat_7"), val = tensor([2, 256, 12])]; - tensor transpose_1 = transpose(perm = transpose_1_perm_0, x = var_1403)[name = tensor("transpose_2")]; + tensor transpose_1 = transpose(perm = transpose_1_perm_0, x = var_1416)[name = tensor("transpose_2")]; tensor reshape_1 = reshape(shape = concat_7, x = transpose_1)[name = tensor("reshape_1")]; tensor matmul_2_transpose_x_0 = const()[name = tensor("matmul_2_transpose_x_0"), val = tensor(false)]; tensor matmul_2_transpose_y_0 = const()[name = tensor("matmul_2_transpose_y_0"), val = tensor(false)]; @@ -1237,10 +1244,10 @@ program(1.0) tensor output_end_0 = const()[name = tensor("output_end_0"), val = tensor([1, 2, 11])]; tensor output_end_mask_0 = const()[name = tensor("output_end_mask_0"), val = tensor([true, true, false])]; tensor output = slice_by_index(begin = output_begin_0, end = output_end_0, end_mask = output_end_mask_0, x = reshape_2)[name = tensor("output")]; - tensor probs = sigmoid(x = output)[name = tensor("op_1407")]; - tensor var_1409_axis_0 = const()[name = tensor("op_1409_axis_0"), val = tensor(0)]; - tensor dec_kv_new = stack(axis = var_1409_axis_0, values = (var_1105, nkv))[name = tensor("op_1409")]; - tensor var_1411_axis_0 = const()[name = tensor("op_1411_axis_0"), val = tensor(0)]; - tensor dec_scale_new = stack(axis = var_1411_axis_0, values = (new_scale_9, new_scale))[name = tensor("op_1411")]; + tensor probs = sigmoid(x = output)[name = tensor("op_1420")]; + tensor var_1422_axis_0 = const()[name = tensor("op_1422_axis_0"), val = tensor(0)]; + tensor dec_kv_new = stack(axis = var_1422_axis_0, values = (var_1118, nkv))[name = tensor("op_1422")]; + tensor var_1424_axis_0 = const()[name = tensor("op_1424_axis_0"), val = tensor(0)]; + tensor dec_scale_new = stack(axis = var_1424_axis_0, values = (new_scale_9, new_scale))[name = tensor("op_1424")]; } -> (probs, enc_kv_new, enc_scale_new, enc_conv_cache_new, cnn_window_new, dec_kv_new, dec_scale_new); } \ No newline at end of file diff --git a/optimized/dih2/200ms/ls_eend_dih2_200ms.mlpackage/Data/com.apple.CoreML/model.mlmodel b/optimized/dih2/200ms/ls_eend_dih2_200ms.mlpackage/Data/com.apple.CoreML/model.mlmodel index 51b7851ed05bebc0051a092a84d8dcde58dd2ed9..ffbe0985dc3832581dca4b17786a3207330562f7 100644 --- a/optimized/dih2/200ms/ls_eend_dih2_200ms.mlpackage/Data/com.apple.CoreML/model.mlmodel +++ b/optimized/dih2/200ms/ls_eend_dih2_200ms.mlpackage/Data/com.apple.CoreML/model.mlmodel @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:2de52b68412da2b73e916665ebf41cce5eacd2010c99e37de2722c6ce40094a2 -size 179874 +oid sha256:9fd639729c2e10e76067e27f6f5a186e23712630cbb548e9b7620438e1ac94c6 +size 184854 diff --git a/optimized/dih2/200ms/ls_eend_dih2_200ms.mlpackage/Manifest.json b/optimized/dih2/200ms/ls_eend_dih2_200ms.mlpackage/Manifest.json index 439619156a11723a4eb1eb25e302786f8f2998b0..ef6ea4a36e08efa69e926663bf9bddbfb17bdfd9 100644 --- a/optimized/dih2/200ms/ls_eend_dih2_200ms.mlpackage/Manifest.json +++ b/optimized/dih2/200ms/ls_eend_dih2_200ms.mlpackage/Manifest.json @@ -1,18 +1,18 @@ { "fileFormatVersion": "1.0.0", "itemInfoEntries": { - "07858CB5-65AA-49FE-837C-C3CED8A0B958": { - "author": "com.apple.CoreML", - "description": "CoreML Model Weights", - "name": "weights", - "path": "com.apple.CoreML/weights" - }, - "E331E090-E6BB-40EA-B66D-83ED2654BAA6": { + "5F736C82-ABFF-4D26-B541-9C4CCF47CDCA": { "author": "com.apple.CoreML", "description": "CoreML Model Specification", "name": "model.mlmodel", "path": "com.apple.CoreML/model.mlmodel" + }, + "B05FC407-28F8-4F04-B6D6-27CDC5F0D3B1": { + "author": "com.apple.CoreML", + "description": "CoreML Model Weights", + "name": "weights", + "path": "com.apple.CoreML/weights" } }, - "rootModelIdentifier": "E331E090-E6BB-40EA-B66D-83ED2654BAA6" + "rootModelIdentifier": "5F736C82-ABFF-4D26-B541-9C4CCF47CDCA" } diff --git a/optimized/dih2/300ms/ls_eend_dih2_300ms.mlmodelc/analytics/coremldata.bin b/optimized/dih2/300ms/ls_eend_dih2_300ms.mlmodelc/analytics/coremldata.bin index aa2356ddff010ed9351b885712064ec8c4872b59..65a87f95a21450aecd95f34377a2914d0593c27e 100644 --- a/optimized/dih2/300ms/ls_eend_dih2_300ms.mlmodelc/analytics/coremldata.bin +++ b/optimized/dih2/300ms/ls_eend_dih2_300ms.mlmodelc/analytics/coremldata.bin @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:a0e2c05f0fa28ecf43578092711f1d39b3176b08681e17319b4ff36eac9caff0 +oid sha256:16ca3f2bb974e8c6e8d39726f5a028c8f9b01e4b894b4fae3109c92c5f67b136 size 243 diff --git a/optimized/dih2/300ms/ls_eend_dih2_300ms.mlmodelc/coremldata.bin b/optimized/dih2/300ms/ls_eend_dih2_300ms.mlmodelc/coremldata.bin index 4bac83cc94d2161567ef0eaff6090eac0493e52a..36fbd09f1caa54a582a266c828e716be7bb6c777 100644 --- a/optimized/dih2/300ms/ls_eend_dih2_300ms.mlmodelc/coremldata.bin +++ b/optimized/dih2/300ms/ls_eend_dih2_300ms.mlmodelc/coremldata.bin @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:f7894d3847f177d05bc7f718737302118c6e2735e776cb8d02cf573f134d75b4 -size 1308 +oid sha256:f0494f0c13f34614bb55bd5d8c5498411214bdef6540124ea7f63c9a8515f6c7 +size 1411 diff --git a/optimized/dih2/300ms/ls_eend_dih2_300ms.mlmodelc/metadata.json b/optimized/dih2/300ms/ls_eend_dih2_300ms.mlmodelc/metadata.json index 61a23767da180facd27d20100346d45d4522e1f9..fbdf79e0cb1eaf56046b0f5273fb89d1118c174b 100644 --- a/optimized/dih2/300ms/ls_eend_dih2_300ms.mlmodelc/metadata.json +++ b/optimized/dih2/300ms/ls_eend_dih2_300ms.mlmodelc/metadata.json @@ -1,7 +1,7 @@ [ { "metadataOutputVersion" : "3.0", - "shortDescription" : "LS-EEND DIHARD II streaming diarizer (pipeline, T=3, max_speakers=10)", + "shortDescription" : "LS-EEND DIHARD II streaming diarizer (pipeline, T=3, max_speakers=10, layout=raw_mel, cu=all)", "outputSchema" : [ { "hasShapeFlexibility" : "0", @@ -81,7 +81,7 @@ "specificationVersion" : 8, "mlProgramOperationTypeHistogram" : { "Ios16.reduceL2Norm" : 2, - "Ios17.reshape" : 68, + "Ios17.reshape" : 69, "Ios16.softmax" : 2, "Ios17.matmul" : 29, "Ios17.transpose" : 57, @@ -89,7 +89,7 @@ "Ios17.expandDims" : 3, "Ios17.add" : 46, "Ios16.sigmoid" : 5, - "Ios17.sliceByIndex" : 56, + "Ios17.sliceByIndex" : 59, "Tile" : 1, "Ios16.reduceSum" : 1, "Ios17.squeeze" : 2, @@ -101,7 +101,7 @@ "Ios16.silu" : 18, "Ios17.realDiv" : 20, "Ios17.linear" : 56, - "Stack" : 5, + "Stack" : 6, "Ios17.concat" : 18, "Ios16.relu" : 2, "Ios16.cumsum" : 1, @@ -128,9 +128,9 @@ "hasShapeFlexibility" : "0", "isOptional" : "0", "dataType" : "Float32", - "formattedType" : "MultiArray (Float32 1 × 3 × 345)", + "formattedType" : "MultiArray (Float32 1 × 35 × 23)", "shortDescription" : "", - "shape" : "[1, 3, 345]", + "shape" : "[1, 35, 23]", "name" : "features", "type" : "MultiArray" }, @@ -206,8 +206,8 @@ } ], "userDefinedMetadata" : { - "com.github.apple.coremltools.conversion_date" : "2026-04-16", - "config" : "{\"model_name\": \"dih2\", \"model_label\": \"DIHARD II\", \"variant\": \"pipeline\", \"chunk_size\": 3, \"step_duration_ms\": 300, \"frame_duration_ms\": 100, \"frame_duration_seconds\": 0.1, \"max_speakers\": 10, \"max_nspks\": 12, \"n_units\": 256, \"n_heads\": 4, \"enc_n_layers\": 4, \"dec_n_layers\": 2, \"conv_kernel_size\": 16, \"conv_delay\": 9, \"sample_rate\": 8000, \"win_length\": 200, \"hop_length\": 80, \"n_mels\": 23, \"context_size\": 7, \"subsampling\": 10, \"feat_type\": \"logmel23_cummn\", \"pure_roll\": true}", + "com.github.apple.coremltools.conversion_date" : "2026-04-18", + "config" : "{\"model_name\": \"dih2\", \"model_label\": \"DIHARD II\", \"variant\": \"pipeline\", \"chunk_size\": 3, \"step_duration_ms\": 300, \"frame_duration_ms\": 100, \"frame_duration_seconds\": 0.1, \"max_speakers\": 10, \"max_nspks\": 12, \"n_units\": 256, \"n_heads\": 4, \"enc_n_layers\": 4, \"dec_n_layers\": 2, \"conv_kernel_size\": 16, \"conv_delay\": 9, \"sample_rate\": 8000, \"win_length\": 200, \"hop_length\": 80, \"n_mels\": 23, \"context_size\": 7, \"subsampling\": 10, \"feat_type\": \"logmel23_cummn\", \"pure_roll\": true, \"input_layout\": \"raw_mel\", \"compute_units_export\": \"all\", \"raw_mel_length\": 35}", "com.github.apple.coremltools.source" : "torch==2.6.0", "com.github.apple.coremltools.version" : "9.0", "com.github.apple.coremltools.source_dialect" : "TorchScript" diff --git a/optimized/dih2/300ms/ls_eend_dih2_300ms.mlmodelc/model.mil b/optimized/dih2/300ms/ls_eend_dih2_300ms.mlmodelc/model.mil index 48f1c415ceeabfb14a7edfb748becb4366d6de9f..d54a8f41d62f449a9ccc5f0085a42da328ef56f7 100644 --- a/optimized/dih2/300ms/ls_eend_dih2_300ms.mlmodelc/model.mil +++ b/optimized/dih2/300ms/ls_eend_dih2_300ms.mlmodelc/model.mil @@ -1,234 +1,252 @@ program(1.0) [buildInfo = dict, tensor>({{"coremlc-component-MIL", "3510.2.1"}, {"coremlc-version", "3500.32.1"}, {"coremltools-component-torch", "2.6.0"}, {"coremltools-source-dialect", "TorchScript"}, {"coremltools-version", "9.0"}})] { - func main(tensor cnn_window, tensor dec_kv, tensor dec_scale, tensor enc_conv_cache, tensor enc_kv, tensor enc_scale, tensor features, tensor valid_mask) { - tensor encoder_cnn_bias = const()[name = tensor("encoder_cnn_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(64)))]; - tensor encoder_cnn_weight = const()[name = tensor("encoder_cnn_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1152)))]; - tensor encoder__causal_mask = const()[name = tensor("encoder__causal_mask"), val = tensor([[0x1p+0, 0x0p+0, 0x0p+0], [0x1p+0, 0x1p+0, 0x0p+0], [0x1p+0, 0x1p+0, 0x1p+0]])]; - tensor encoder__t_index = const()[name = tensor("encoder__t_index"), val = tensor([0x1p+0, 0x1p+1, 0x1.8p+1])]; - tensor encoder_input_projection_linear_bias = const()[name = tensor("encoder_input_projection_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4981952)))]; - tensor encoder_input_projection_linear_weight = const()[name = tensor("encoder_input_projection_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4983040)))]; - tensor encoder_pre_layer_norm_bias = const()[name = tensor("encoder_pre_layer_norm_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(5336384)))]; - tensor encoder_pre_layer_norm_weight = const()[name = tensor("encoder_pre_layer_norm_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(5337472)))]; - tensor encoder_ffn1_0_module_sequential_0_bias = const()[name = tensor("encoder_ffn1_0_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(5338560)))]; - tensor encoder_ffn1_0_module_sequential_0_weight = const()[name = tensor("encoder_ffn1_0_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(5339648)))]; - tensor encoder_ffn1_0_module_sequential_1_linear_bias = const()[name = tensor("encoder_ffn1_0_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(5340736)))]; - tensor encoder_ffn1_0_module_sequential_1_linear_weight = const()[name = tensor("encoder_ffn1_0_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(5344896)))]; - tensor encoder_ffn1_0_module_sequential_4_linear_bias = const()[name = tensor("encoder_ffn1_0_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(6393536)))]; - tensor encoder_ffn1_0_module_sequential_4_linear_weight = const()[name = tensor("encoder_ffn1_0_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(6394624)))]; - tensor encoder_ret_lns_0_bias = const()[name = tensor("encoder_ret_lns_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7443264)))]; - tensor encoder_ret_lns_0_weight = const()[name = tensor("encoder_ret_lns_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7444352)))]; - tensor encoder_q_proj_0_bias = const()[name = tensor("encoder_q_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7445440)))]; - tensor encoder_q_proj_0_weight = const()[name = tensor("encoder_q_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7446528)))]; - tensor encoder_k_proj_0_bias = const()[name = tensor("encoder_k_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7708736)))]; - tensor encoder_k_proj_0_weight = const()[name = tensor("encoder_k_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7709824)))]; - tensor encoder_v_proj_0_bias = const()[name = tensor("encoder_v_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7972032)))]; - tensor encoder_v_proj_0_weight = const()[name = tensor("encoder_v_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7973120)))]; - tensor encoder_g_proj_0_bias = const()[name = tensor("encoder_g_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8235328)))]; - tensor encoder_g_proj_0_weight = const()[name = tensor("encoder_g_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8236416)))]; - tensor encoder_out_proj_0_bias = const()[name = tensor("encoder_out_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8498624)))]; - tensor encoder_out_proj_0_weight = const()[name = tensor("encoder_out_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8499712)))]; - tensor encoder_conv_module_0_sequential_0_bias = const()[name = tensor("encoder_conv_module_0_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8761920)))]; - tensor encoder_conv_module_0_sequential_0_weight = const()[name = tensor("encoder_conv_module_0_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8763008)))]; - tensor encoder_conv_module_0_sequential_2_conv_bias = const()[name = tensor("encoder_conv_module_0_sequential_2_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8764096)))]; - tensor encoder_conv_module_0_sequential_2_conv_weight = const()[name = tensor("encoder_conv_module_0_sequential_2_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8766208)))]; - tensor encoder_conv_module_0_sequential_4_conv_weight = const()[name = tensor("encoder_conv_module_0_sequential_4_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9290560)))]; - tensor encoder_conv_module_0_sequential_5_running_var = const()[name = tensor("encoder_conv_module_0_sequential_5_running_var"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9307008)))]; - tensor encoder_conv_module_0_sequential_5_running_mean = const()[name = tensor("encoder_conv_module_0_sequential_5_running_mean"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9308096)))]; - tensor encoder_conv_module_0_sequential_5_bias = const()[name = tensor("encoder_conv_module_0_sequential_5_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9309184)))]; - tensor encoder_conv_module_0_sequential_5_weight = const()[name = tensor("encoder_conv_module_0_sequential_5_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9310272)))]; - tensor encoder_conv_module_0_sequential_7_conv_bias = const()[name = tensor("encoder_conv_module_0_sequential_7_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9311360)))]; - tensor encoder_conv_module_0_sequential_7_conv_weight = const()[name = tensor("encoder_conv_module_0_sequential_7_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9312448)))]; - tensor encoder_ffn2_0_module_sequential_0_bias = const()[name = tensor("encoder_ffn2_0_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9574656)))]; - tensor encoder_ffn2_0_module_sequential_0_weight = const()[name = tensor("encoder_ffn2_0_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9575744)))]; - tensor encoder_ffn2_0_module_sequential_1_linear_bias = const()[name = tensor("encoder_ffn2_0_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9576832)))]; - tensor encoder_ffn2_0_module_sequential_1_linear_weight = const()[name = tensor("encoder_ffn2_0_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9580992)))]; - tensor encoder_ffn2_0_module_sequential_4_linear_bias = const()[name = tensor("encoder_ffn2_0_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(10629632)))]; - tensor encoder_ffn2_0_module_sequential_4_linear_weight = const()[name = tensor("encoder_ffn2_0_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(10630720)))]; - tensor encoder_layer_norm_0_bias = const()[name = tensor("encoder_layer_norm_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11679360)))]; - tensor encoder_layer_norm_0_weight = const()[name = tensor("encoder_layer_norm_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11680448)))]; - tensor encoder_ffn1_1_module_sequential_0_bias = const()[name = tensor("encoder_ffn1_1_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11681536)))]; - tensor encoder_ffn1_1_module_sequential_0_weight = const()[name = tensor("encoder_ffn1_1_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11682624)))]; - tensor encoder_ffn1_1_module_sequential_1_linear_bias = const()[name = tensor("encoder_ffn1_1_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11683712)))]; - tensor encoder_ffn1_1_module_sequential_1_linear_weight = const()[name = tensor("encoder_ffn1_1_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11687872)))]; - tensor encoder_ffn1_1_module_sequential_4_linear_bias = const()[name = tensor("encoder_ffn1_1_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(12736512)))]; - tensor encoder_ffn1_1_module_sequential_4_linear_weight = const()[name = tensor("encoder_ffn1_1_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(12737600)))]; - tensor encoder_ret_lns_1_bias = const()[name = tensor("encoder_ret_lns_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(13786240)))]; - tensor encoder_ret_lns_1_weight = const()[name = tensor("encoder_ret_lns_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(13787328)))]; - tensor encoder_q_proj_1_bias = const()[name = tensor("encoder_q_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(13788416)))]; - tensor encoder_q_proj_1_weight = const()[name = tensor("encoder_q_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(13789504)))]; - tensor encoder_k_proj_1_bias = const()[name = tensor("encoder_k_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14051712)))]; - tensor encoder_k_proj_1_weight = const()[name = tensor("encoder_k_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14052800)))]; - tensor encoder_v_proj_1_bias = const()[name = tensor("encoder_v_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14315008)))]; - tensor encoder_v_proj_1_weight = const()[name = tensor("encoder_v_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14316096)))]; - tensor encoder_g_proj_1_bias = const()[name = tensor("encoder_g_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14578304)))]; - tensor encoder_g_proj_1_weight = const()[name = tensor("encoder_g_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14579392)))]; - tensor encoder_out_proj_1_bias = const()[name = tensor("encoder_out_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14841600)))]; - tensor encoder_out_proj_1_weight = const()[name = tensor("encoder_out_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14842688)))]; - tensor encoder_conv_module_1_sequential_0_bias = const()[name = tensor("encoder_conv_module_1_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15104896)))]; - tensor encoder_conv_module_1_sequential_0_weight = const()[name = tensor("encoder_conv_module_1_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15105984)))]; - tensor encoder_conv_module_1_sequential_2_conv_bias = const()[name = tensor("encoder_conv_module_1_sequential_2_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15107072)))]; - tensor encoder_conv_module_1_sequential_2_conv_weight = const()[name = tensor("encoder_conv_module_1_sequential_2_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15109184)))]; - tensor encoder_conv_module_1_sequential_4_conv_weight = const()[name = tensor("encoder_conv_module_1_sequential_4_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15633536)))]; - tensor encoder_conv_module_1_sequential_5_running_var = const()[name = tensor("encoder_conv_module_1_sequential_5_running_var"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15649984)))]; - tensor encoder_conv_module_1_sequential_5_running_mean = const()[name = tensor("encoder_conv_module_1_sequential_5_running_mean"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15651072)))]; - tensor encoder_conv_module_1_sequential_5_bias = const()[name = tensor("encoder_conv_module_1_sequential_5_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15652160)))]; - tensor encoder_conv_module_1_sequential_5_weight = const()[name = tensor("encoder_conv_module_1_sequential_5_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15653248)))]; - tensor encoder_conv_module_1_sequential_7_conv_bias = const()[name = tensor("encoder_conv_module_1_sequential_7_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15654336)))]; - tensor encoder_conv_module_1_sequential_7_conv_weight = const()[name = tensor("encoder_conv_module_1_sequential_7_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15655424)))]; - tensor encoder_ffn2_1_module_sequential_0_bias = const()[name = tensor("encoder_ffn2_1_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15917632)))]; - tensor encoder_ffn2_1_module_sequential_0_weight = const()[name = tensor("encoder_ffn2_1_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15918720)))]; - tensor encoder_ffn2_1_module_sequential_1_linear_bias = const()[name = tensor("encoder_ffn2_1_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15919808)))]; - tensor encoder_ffn2_1_module_sequential_1_linear_weight = const()[name = tensor("encoder_ffn2_1_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15923968)))]; - tensor encoder_ffn2_1_module_sequential_4_linear_bias = const()[name = tensor("encoder_ffn2_1_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(16972608)))]; - tensor encoder_ffn2_1_module_sequential_4_linear_weight = const()[name = tensor("encoder_ffn2_1_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(16973696)))]; - tensor encoder_layer_norm_1_bias = const()[name = tensor("encoder_layer_norm_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18022336)))]; - tensor encoder_layer_norm_1_weight = const()[name = tensor("encoder_layer_norm_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18023424)))]; - tensor encoder_ffn1_2_module_sequential_0_bias = const()[name = tensor("encoder_ffn1_2_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18024512)))]; - tensor encoder_ffn1_2_module_sequential_0_weight = const()[name = tensor("encoder_ffn1_2_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18025600)))]; - tensor encoder_ffn1_2_module_sequential_1_linear_bias = const()[name = tensor("encoder_ffn1_2_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18026688)))]; - tensor encoder_ffn1_2_module_sequential_1_linear_weight = const()[name = tensor("encoder_ffn1_2_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18030848)))]; - tensor encoder_ffn1_2_module_sequential_4_linear_bias = const()[name = tensor("encoder_ffn1_2_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(19079488)))]; - tensor encoder_ffn1_2_module_sequential_4_linear_weight = const()[name = tensor("encoder_ffn1_2_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(19080576)))]; - tensor encoder_ret_lns_2_bias = const()[name = tensor("encoder_ret_lns_2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20129216)))]; - tensor encoder_ret_lns_2_weight = const()[name = tensor("encoder_ret_lns_2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20130304)))]; - tensor encoder_q_proj_2_bias = const()[name = tensor("encoder_q_proj_2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20131392)))]; - tensor encoder_q_proj_2_weight = const()[name = tensor("encoder_q_proj_2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20132480)))]; - tensor encoder_k_proj_2_bias = const()[name = tensor("encoder_k_proj_2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20394688)))]; - tensor encoder_k_proj_2_weight = const()[name = tensor("encoder_k_proj_2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20395776)))]; - tensor encoder_v_proj_2_bias = const()[name = tensor("encoder_v_proj_2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20657984)))]; - tensor encoder_v_proj_2_weight = const()[name = tensor("encoder_v_proj_2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20659072)))]; - tensor encoder_g_proj_2_bias = const()[name = tensor("encoder_g_proj_2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20921280)))]; - tensor encoder_g_proj_2_weight = const()[name = tensor("encoder_g_proj_2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20922368)))]; - tensor encoder_out_proj_2_bias = const()[name = tensor("encoder_out_proj_2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21184576)))]; - tensor encoder_out_proj_2_weight = const()[name = tensor("encoder_out_proj_2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21185664)))]; - tensor encoder_conv_module_2_sequential_0_bias = const()[name = tensor("encoder_conv_module_2_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21447872)))]; - tensor encoder_conv_module_2_sequential_0_weight = const()[name = tensor("encoder_conv_module_2_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21448960)))]; - tensor encoder_conv_module_2_sequential_2_conv_bias = const()[name = tensor("encoder_conv_module_2_sequential_2_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21450048)))]; - tensor encoder_conv_module_2_sequential_2_conv_weight = const()[name = tensor("encoder_conv_module_2_sequential_2_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21452160)))]; - tensor encoder_conv_module_2_sequential_4_conv_weight = const()[name = tensor("encoder_conv_module_2_sequential_4_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21976512)))]; - tensor encoder_conv_module_2_sequential_5_running_var = const()[name = tensor("encoder_conv_module_2_sequential_5_running_var"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21992960)))]; - tensor encoder_conv_module_2_sequential_5_running_mean = const()[name = tensor("encoder_conv_module_2_sequential_5_running_mean"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21994048)))]; - tensor encoder_conv_module_2_sequential_5_bias = const()[name = tensor("encoder_conv_module_2_sequential_5_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21995136)))]; - tensor encoder_conv_module_2_sequential_5_weight = const()[name = tensor("encoder_conv_module_2_sequential_5_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21996224)))]; - tensor encoder_conv_module_2_sequential_7_conv_bias = const()[name = tensor("encoder_conv_module_2_sequential_7_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21997312)))]; - tensor encoder_conv_module_2_sequential_7_conv_weight = const()[name = tensor("encoder_conv_module_2_sequential_7_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21998400)))]; - tensor encoder_ffn2_2_module_sequential_0_bias = const()[name = tensor("encoder_ffn2_2_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(22260608)))]; - tensor encoder_ffn2_2_module_sequential_0_weight = const()[name = tensor("encoder_ffn2_2_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(22261696)))]; - tensor encoder_ffn2_2_module_sequential_1_linear_bias = const()[name = tensor("encoder_ffn2_2_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(22262784)))]; - tensor encoder_ffn2_2_module_sequential_1_linear_weight = const()[name = tensor("encoder_ffn2_2_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(22266944)))]; - tensor encoder_ffn2_2_module_sequential_4_linear_bias = const()[name = tensor("encoder_ffn2_2_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(23315584)))]; - tensor encoder_ffn2_2_module_sequential_4_linear_weight = const()[name = tensor("encoder_ffn2_2_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(23316672)))]; - tensor encoder_layer_norm_2_bias = const()[name = tensor("encoder_layer_norm_2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(24365312)))]; - tensor encoder_layer_norm_2_weight = const()[name = tensor("encoder_layer_norm_2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(24366400)))]; - tensor encoder_ffn1_3_module_sequential_0_bias = const()[name = tensor("encoder_ffn1_3_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(24367488)))]; - tensor encoder_ffn1_3_module_sequential_0_weight = const()[name = tensor("encoder_ffn1_3_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(24368576)))]; - tensor encoder_ffn1_3_module_sequential_1_linear_bias = const()[name = tensor("encoder_ffn1_3_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(24369664)))]; - tensor encoder_ffn1_3_module_sequential_1_linear_weight = const()[name = tensor("encoder_ffn1_3_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(24373824)))]; - tensor encoder_ffn1_3_module_sequential_4_linear_bias = const()[name = tensor("encoder_ffn1_3_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(25422464)))]; - tensor encoder_ffn1_3_module_sequential_4_linear_weight = const()[name = tensor("encoder_ffn1_3_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(25423552)))]; - tensor encoder_ret_lns_3_bias = const()[name = tensor("encoder_ret_lns_3_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(26472192)))]; - tensor encoder_ret_lns_3_weight = const()[name = tensor("encoder_ret_lns_3_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(26473280)))]; - tensor encoder_q_proj_3_bias = const()[name = tensor("encoder_q_proj_3_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(26474368)))]; - tensor encoder_q_proj_3_weight = const()[name = tensor("encoder_q_proj_3_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(26475456)))]; - tensor encoder_k_proj_3_bias = const()[name = tensor("encoder_k_proj_3_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(26737664)))]; - tensor encoder_k_proj_3_weight = const()[name = tensor("encoder_k_proj_3_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(26738752)))]; - tensor encoder_v_proj_3_bias = const()[name = tensor("encoder_v_proj_3_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27000960)))]; - tensor encoder_v_proj_3_weight = const()[name = tensor("encoder_v_proj_3_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27002048)))]; - tensor encoder_g_proj_3_bias = const()[name = tensor("encoder_g_proj_3_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27264256)))]; - tensor encoder_g_proj_3_weight = const()[name = tensor("encoder_g_proj_3_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27265344)))]; - tensor encoder_out_proj_3_bias = const()[name = tensor("encoder_out_proj_3_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27527552)))]; - tensor encoder_out_proj_3_weight = const()[name = tensor("encoder_out_proj_3_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27528640)))]; - tensor encoder_conv_module_3_sequential_0_bias = const()[name = tensor("encoder_conv_module_3_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27790848)))]; - tensor encoder_conv_module_3_sequential_0_weight = const()[name = tensor("encoder_conv_module_3_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27791936)))]; - tensor encoder_conv_module_3_sequential_2_conv_bias = const()[name = tensor("encoder_conv_module_3_sequential_2_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27793024)))]; - tensor encoder_conv_module_3_sequential_2_conv_weight = const()[name = tensor("encoder_conv_module_3_sequential_2_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27795136)))]; - tensor encoder_conv_module_3_sequential_4_conv_weight = const()[name = tensor("encoder_conv_module_3_sequential_4_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28319488)))]; - tensor encoder_conv_module_3_sequential_5_running_var = const()[name = tensor("encoder_conv_module_3_sequential_5_running_var"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28335936)))]; - tensor encoder_conv_module_3_sequential_5_running_mean = const()[name = tensor("encoder_conv_module_3_sequential_5_running_mean"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28337024)))]; - tensor encoder_conv_module_3_sequential_5_bias = const()[name = tensor("encoder_conv_module_3_sequential_5_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28338112)))]; - tensor encoder_conv_module_3_sequential_5_weight = const()[name = tensor("encoder_conv_module_3_sequential_5_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28339200)))]; - tensor encoder_conv_module_3_sequential_7_conv_bias = const()[name = tensor("encoder_conv_module_3_sequential_7_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28340288)))]; - tensor encoder_conv_module_3_sequential_7_conv_weight = const()[name = tensor("encoder_conv_module_3_sequential_7_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28341376)))]; - tensor encoder_ffn2_3_module_sequential_0_bias = const()[name = tensor("encoder_ffn2_3_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28603584)))]; - tensor encoder_ffn2_3_module_sequential_0_weight = const()[name = tensor("encoder_ffn2_3_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28604672)))]; - tensor encoder_ffn2_3_module_sequential_1_linear_bias = const()[name = tensor("encoder_ffn2_3_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28605760)))]; - tensor encoder_ffn2_3_module_sequential_1_linear_weight = const()[name = tensor("encoder_ffn2_3_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28609920)))]; - tensor encoder_ffn2_3_module_sequential_4_linear_bias = const()[name = tensor("encoder_ffn2_3_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(29658560)))]; - tensor encoder_ffn2_3_module_sequential_4_linear_weight = const()[name = tensor("encoder_ffn2_3_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(29659648)))]; - tensor encoder_layer_norm_3_bias = const()[name = tensor("encoder_layer_norm_3_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(30708288)))]; - tensor encoder_layer_norm_3_weight = const()[name = tensor("encoder_layer_norm_3_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(30709376)))]; - tensor decoder__causal_mask = const()[name = tensor("decoder__causal_mask"), val = tensor([[0x1p+0, 0x0p+0, 0x0p+0], [0x1p+0, 0x1p+0, 0x0p+0], [0x1p+0, 0x1p+0, 0x1p+0]])]; - tensor decoder_convert_bias = const()[name = tensor("decoder_convert_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(30710464)))]; - tensor decoder_convert_weight = const()[name = tensor("decoder_convert_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(30711552)))]; - tensor decoder_q_proj_0_bias = const()[name = tensor("decoder_q_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(31235904)))]; - tensor decoder_q_proj_0_weight = const()[name = tensor("decoder_q_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(31236992)))]; - tensor decoder_k_proj_0_bias = const()[name = tensor("decoder_k_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(31499200)))]; - tensor decoder_k_proj_0_weight = const()[name = tensor("decoder_k_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(31500288)))]; - tensor decoder_v_proj_0_bias = const()[name = tensor("decoder_v_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(31762496)))]; - tensor decoder_v_proj_0_weight = const()[name = tensor("decoder_v_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(31763584)))]; - tensor decoder_g_proj_0_bias = const()[name = tensor("decoder_g_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32025792)))]; - tensor decoder_g_proj_0_weight = const()[name = tensor("decoder_g_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32026880)))]; - tensor decoder_out_proj_0_bias = const()[name = tensor("decoder_out_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32289088)))]; - tensor decoder_out_proj_0_weight = const()[name = tensor("decoder_out_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32290176)))]; - tensor decoder_norm11_0_bias = const()[name = tensor("decoder_norm11_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32552384)))]; - tensor decoder_norm11_0_weight = const()[name = tensor("decoder_norm11_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32553472)))]; - tensor decoder_self_attn2_0_out_proj_bias = const()[name = tensor("decoder_self_attn2_0_out_proj_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32554560)))]; - tensor decoder_self_attn2_0_out_proj_weight = const()[name = tensor("decoder_self_attn2_0_out_proj_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32555648)))]; - tensor decoder_self_attn2_0_in_proj_bias = const()[name = tensor("decoder_self_attn2_0_in_proj_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32817856)))]; - tensor decoder_self_attn2_0_in_proj_weight = const()[name = tensor("decoder_self_attn2_0_in_proj_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32820992)))]; - tensor decoder_norm21_0_bias = const()[name = tensor("decoder_norm21_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(33607488)))]; - tensor decoder_norm21_0_weight = const()[name = tensor("decoder_norm21_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(33608576)))]; - tensor decoder_linear1_0_bias = const()[name = tensor("decoder_linear1_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(33609664)))]; - tensor decoder_linear1_0_weight = const()[name = tensor("decoder_linear1_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(33617920)))]; - tensor decoder_linear2_0_bias = const()[name = tensor("decoder_linear2_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(35715136)))]; - tensor decoder_linear2_0_weight = const()[name = tensor("decoder_linear2_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(35716224)))]; - tensor decoder_norm22_0_bias = const()[name = tensor("decoder_norm22_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(37813440)))]; - tensor decoder_norm22_0_weight = const()[name = tensor("decoder_norm22_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(37814528)))]; - tensor decoder_q_proj_1_bias = const()[name = tensor("decoder_q_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(37815616)))]; - tensor decoder_q_proj_1_weight = const()[name = tensor("decoder_q_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(37816704)))]; - tensor decoder_k_proj_1_bias = const()[name = tensor("decoder_k_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38078912)))]; - tensor decoder_k_proj_1_weight = const()[name = tensor("decoder_k_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38080000)))]; - tensor decoder_v_proj_1_bias = const()[name = tensor("decoder_v_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38342208)))]; - tensor decoder_v_proj_1_weight = const()[name = tensor("decoder_v_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38343296)))]; - tensor decoder_g_proj_1_bias = const()[name = tensor("decoder_g_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38605504)))]; - tensor decoder_g_proj_1_weight = const()[name = tensor("decoder_g_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38606592)))]; - tensor decoder_out_proj_1_bias = const()[name = tensor("decoder_out_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38868800)))]; - tensor decoder_out_proj_1_weight = const()[name = tensor("decoder_out_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38869888)))]; - tensor decoder_norm11_1_bias = const()[name = tensor("decoder_norm11_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(39132096)))]; - tensor decoder_norm11_1_weight = const()[name = tensor("decoder_norm11_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(39133184)))]; - tensor decoder_self_attn2_1_out_proj_bias = const()[name = tensor("decoder_self_attn2_1_out_proj_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(39134272)))]; - tensor decoder_self_attn2_1_out_proj_weight = const()[name = tensor("decoder_self_attn2_1_out_proj_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(39135360)))]; - tensor decoder_self_attn2_1_in_proj_bias = const()[name = tensor("decoder_self_attn2_1_in_proj_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(39397568)))]; - tensor decoder_self_attn2_1_in_proj_weight = const()[name = tensor("decoder_self_attn2_1_in_proj_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(39400704)))]; - tensor decoder_norm21_1_bias = const()[name = tensor("decoder_norm21_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(40187200)))]; - tensor decoder_norm21_1_weight = const()[name = tensor("decoder_norm21_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(40188288)))]; - tensor decoder_linear1_1_bias = const()[name = tensor("decoder_linear1_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(40189376)))]; - tensor decoder_linear1_1_weight = const()[name = tensor("decoder_linear1_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(40197632)))]; - tensor decoder_linear2_1_bias = const()[name = tensor("decoder_linear2_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(42294848)))]; - tensor decoder_linear2_1_weight = const()[name = tensor("decoder_linear2_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(42295936)))]; - tensor decoder_norm22_1_bias = const()[name = tensor("decoder_norm22_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(44393152)))]; - tensor decoder_norm22_1_weight = const()[name = tensor("decoder_norm22_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(44394240)))]; - tensor var_11 = const()[name = tensor("op_11"), val = tensor(0x1.5798eep-27)]; - tensor var_17 = const()[name = tensor("op_17"), val = tensor(0x1.0c6f7ap-20)]; - tensor var_20 = const()[name = tensor("op_20"), val = tensor(2)]; - tensor var_23 = const()[name = tensor("op_23"), val = tensor(0)]; - tensor var_26 = const()[name = tensor("op_26"), val = tensor(1)]; - tensor var_28 = const()[name = tensor("op_28"), val = tensor(0x1.4f8b58p-17)]; - tensor var_29 = const()[name = tensor("op_29"), val = tensor(true)]; - tensor input_1 = linear(bias = encoder_input_projection_linear_bias, weight = encoder_input_projection_linear_weight, x = features)[name = tensor("linear_0")]; - tensor input_3_axes_0 = const()[name = tensor("input_3_axes_0"), val = tensor([-1])]; - tensor input_3 = layer_norm(axes = input_3_axes_0, beta = encoder_pre_layer_norm_bias, epsilon = var_28, gamma = encoder_pre_layer_norm_weight, x = input_1)[name = tensor("input_3")]; + func main(tensor cnn_window, tensor dec_kv, tensor dec_scale, tensor enc_conv_cache, tensor enc_kv, tensor enc_scale, tensor features, tensor valid_mask) { + tensor inner_encoder_cnn_bias = const()[name = tensor("inner_encoder_cnn_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(64)))]; + tensor inner_encoder_cnn_weight = const()[name = tensor("inner_encoder_cnn_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1152)))]; + tensor inner_encoder__causal_mask = const()[name = tensor("inner_encoder__causal_mask"), val = tensor([[0x1p+0, 0x0p+0, 0x0p+0], [0x1p+0, 0x1p+0, 0x0p+0], [0x1p+0, 0x1p+0, 0x1p+0]])]; + tensor inner_encoder__t_index = const()[name = tensor("inner_encoder__t_index"), val = tensor([0x1p+0, 0x1p+1, 0x1.8p+1])]; + tensor inner_encoder_input_projection_linear_bias = const()[name = tensor("inner_encoder_input_projection_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4981952)))]; + tensor inner_encoder_input_projection_linear_weight = const()[name = tensor("inner_encoder_input_projection_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4983040)))]; + tensor inner_encoder_pre_layer_norm_bias = const()[name = tensor("inner_encoder_pre_layer_norm_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(5336384)))]; + tensor inner_encoder_pre_layer_norm_weight = const()[name = tensor("inner_encoder_pre_layer_norm_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(5337472)))]; + tensor inner_encoder_ffn1_0_module_sequential_0_bias = const()[name = tensor("inner_encoder_ffn1_0_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(5338560)))]; + tensor inner_encoder_ffn1_0_module_sequential_0_weight = const()[name = tensor("inner_encoder_ffn1_0_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(5339648)))]; + tensor inner_encoder_ffn1_0_module_sequential_1_linear_bias = const()[name = tensor("inner_encoder_ffn1_0_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(5340736)))]; + tensor inner_encoder_ffn1_0_module_sequential_1_linear_weight = const()[name = tensor("inner_encoder_ffn1_0_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(5344896)))]; + tensor inner_encoder_ffn1_0_module_sequential_4_linear_bias = const()[name = tensor("inner_encoder_ffn1_0_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(6393536)))]; + tensor inner_encoder_ffn1_0_module_sequential_4_linear_weight = const()[name = tensor("inner_encoder_ffn1_0_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(6394624)))]; + tensor inner_encoder_ret_lns_0_bias = const()[name = tensor("inner_encoder_ret_lns_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7443264)))]; + tensor inner_encoder_ret_lns_0_weight = const()[name = tensor("inner_encoder_ret_lns_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7444352)))]; + tensor inner_encoder_q_proj_0_bias = const()[name = tensor("inner_encoder_q_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7445440)))]; + tensor inner_encoder_q_proj_0_weight = const()[name = tensor("inner_encoder_q_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7446528)))]; + tensor inner_encoder_k_proj_0_bias = const()[name = tensor("inner_encoder_k_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7708736)))]; + tensor inner_encoder_k_proj_0_weight = const()[name = tensor("inner_encoder_k_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7709824)))]; + tensor inner_encoder_v_proj_0_bias = const()[name = tensor("inner_encoder_v_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7972032)))]; + tensor inner_encoder_v_proj_0_weight = const()[name = tensor("inner_encoder_v_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7973120)))]; + tensor inner_encoder_g_proj_0_bias = const()[name = tensor("inner_encoder_g_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8235328)))]; + tensor inner_encoder_g_proj_0_weight = const()[name = tensor("inner_encoder_g_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8236416)))]; + tensor inner_encoder_out_proj_0_bias = const()[name = tensor("inner_encoder_out_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8498624)))]; + tensor inner_encoder_out_proj_0_weight = const()[name = tensor("inner_encoder_out_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8499712)))]; + tensor inner_encoder_conv_module_0_sequential_0_bias = const()[name = tensor("inner_encoder_conv_module_0_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8761920)))]; + tensor inner_encoder_conv_module_0_sequential_0_weight = const()[name = tensor("inner_encoder_conv_module_0_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8763008)))]; + tensor inner_encoder_conv_module_0_sequential_2_conv_bias = const()[name = tensor("inner_encoder_conv_module_0_sequential_2_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8764096)))]; + tensor inner_encoder_conv_module_0_sequential_2_conv_weight = const()[name = tensor("inner_encoder_conv_module_0_sequential_2_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8766208)))]; + tensor inner_encoder_conv_module_0_sequential_4_conv_weight = const()[name = tensor("inner_encoder_conv_module_0_sequential_4_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9290560)))]; + tensor inner_encoder_conv_module_0_sequential_5_running_var = const()[name = tensor("inner_encoder_conv_module_0_sequential_5_running_var"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9307008)))]; + tensor inner_encoder_conv_module_0_sequential_5_running_mean = const()[name = tensor("inner_encoder_conv_module_0_sequential_5_running_mean"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9308096)))]; + tensor inner_encoder_conv_module_0_sequential_5_bias = const()[name = tensor("inner_encoder_conv_module_0_sequential_5_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9309184)))]; + tensor inner_encoder_conv_module_0_sequential_5_weight = const()[name = tensor("inner_encoder_conv_module_0_sequential_5_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9310272)))]; + tensor inner_encoder_conv_module_0_sequential_7_conv_bias = const()[name = tensor("inner_encoder_conv_module_0_sequential_7_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9311360)))]; + tensor inner_encoder_conv_module_0_sequential_7_conv_weight = const()[name = tensor("inner_encoder_conv_module_0_sequential_7_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9312448)))]; + tensor inner_encoder_ffn2_0_module_sequential_0_bias = const()[name = tensor("inner_encoder_ffn2_0_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9574656)))]; + tensor inner_encoder_ffn2_0_module_sequential_0_weight = const()[name = tensor("inner_encoder_ffn2_0_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9575744)))]; + tensor inner_encoder_ffn2_0_module_sequential_1_linear_bias = const()[name = tensor("inner_encoder_ffn2_0_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9576832)))]; + tensor inner_encoder_ffn2_0_module_sequential_1_linear_weight = const()[name = tensor("inner_encoder_ffn2_0_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9580992)))]; + tensor inner_encoder_ffn2_0_module_sequential_4_linear_bias = const()[name = tensor("inner_encoder_ffn2_0_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(10629632)))]; + tensor inner_encoder_ffn2_0_module_sequential_4_linear_weight = const()[name = tensor("inner_encoder_ffn2_0_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(10630720)))]; + tensor inner_encoder_layer_norm_0_bias = const()[name = tensor("inner_encoder_layer_norm_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11679360)))]; + tensor inner_encoder_layer_norm_0_weight = const()[name = tensor("inner_encoder_layer_norm_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11680448)))]; + tensor inner_encoder_ffn1_1_module_sequential_0_bias = const()[name = tensor("inner_encoder_ffn1_1_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11681536)))]; + tensor inner_encoder_ffn1_1_module_sequential_0_weight = const()[name = tensor("inner_encoder_ffn1_1_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11682624)))]; + tensor inner_encoder_ffn1_1_module_sequential_1_linear_bias = const()[name = tensor("inner_encoder_ffn1_1_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11683712)))]; + tensor inner_encoder_ffn1_1_module_sequential_1_linear_weight = const()[name = tensor("inner_encoder_ffn1_1_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11687872)))]; + tensor inner_encoder_ffn1_1_module_sequential_4_linear_bias = const()[name = tensor("inner_encoder_ffn1_1_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(12736512)))]; + tensor inner_encoder_ffn1_1_module_sequential_4_linear_weight = const()[name = tensor("inner_encoder_ffn1_1_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(12737600)))]; + tensor inner_encoder_ret_lns_1_bias = const()[name = tensor("inner_encoder_ret_lns_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(13786240)))]; + tensor inner_encoder_ret_lns_1_weight = const()[name = tensor("inner_encoder_ret_lns_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(13787328)))]; + tensor inner_encoder_q_proj_1_bias = const()[name = tensor("inner_encoder_q_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(13788416)))]; + tensor inner_encoder_q_proj_1_weight = const()[name = tensor("inner_encoder_q_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(13789504)))]; + tensor inner_encoder_k_proj_1_bias = const()[name = tensor("inner_encoder_k_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14051712)))]; + tensor inner_encoder_k_proj_1_weight = const()[name = tensor("inner_encoder_k_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14052800)))]; + tensor inner_encoder_v_proj_1_bias = const()[name = tensor("inner_encoder_v_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14315008)))]; + tensor inner_encoder_v_proj_1_weight = const()[name = tensor("inner_encoder_v_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14316096)))]; + tensor inner_encoder_g_proj_1_bias = const()[name = tensor("inner_encoder_g_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14578304)))]; + tensor inner_encoder_g_proj_1_weight = const()[name = tensor("inner_encoder_g_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14579392)))]; + tensor inner_encoder_out_proj_1_bias = const()[name = tensor("inner_encoder_out_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14841600)))]; + tensor inner_encoder_out_proj_1_weight = const()[name = tensor("inner_encoder_out_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14842688)))]; + tensor inner_encoder_conv_module_1_sequential_0_bias = const()[name = tensor("inner_encoder_conv_module_1_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15104896)))]; + tensor inner_encoder_conv_module_1_sequential_0_weight = const()[name = tensor("inner_encoder_conv_module_1_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15105984)))]; + tensor inner_encoder_conv_module_1_sequential_2_conv_bias = const()[name = tensor("inner_encoder_conv_module_1_sequential_2_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15107072)))]; + tensor inner_encoder_conv_module_1_sequential_2_conv_weight = const()[name = tensor("inner_encoder_conv_module_1_sequential_2_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15109184)))]; + tensor inner_encoder_conv_module_1_sequential_4_conv_weight = const()[name = tensor("inner_encoder_conv_module_1_sequential_4_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15633536)))]; + tensor inner_encoder_conv_module_1_sequential_5_running_var = const()[name = tensor("inner_encoder_conv_module_1_sequential_5_running_var"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15649984)))]; + tensor inner_encoder_conv_module_1_sequential_5_running_mean = const()[name = tensor("inner_encoder_conv_module_1_sequential_5_running_mean"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15651072)))]; + tensor inner_encoder_conv_module_1_sequential_5_bias = const()[name = tensor("inner_encoder_conv_module_1_sequential_5_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15652160)))]; + tensor inner_encoder_conv_module_1_sequential_5_weight = const()[name = tensor("inner_encoder_conv_module_1_sequential_5_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15653248)))]; + tensor inner_encoder_conv_module_1_sequential_7_conv_bias = const()[name = tensor("inner_encoder_conv_module_1_sequential_7_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15654336)))]; + tensor inner_encoder_conv_module_1_sequential_7_conv_weight = const()[name = tensor("inner_encoder_conv_module_1_sequential_7_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15655424)))]; + tensor inner_encoder_ffn2_1_module_sequential_0_bias = const()[name = tensor("inner_encoder_ffn2_1_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15917632)))]; + tensor inner_encoder_ffn2_1_module_sequential_0_weight = const()[name = tensor("inner_encoder_ffn2_1_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15918720)))]; + tensor inner_encoder_ffn2_1_module_sequential_1_linear_bias = const()[name = tensor("inner_encoder_ffn2_1_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15919808)))]; + tensor inner_encoder_ffn2_1_module_sequential_1_linear_weight = const()[name = tensor("inner_encoder_ffn2_1_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15923968)))]; + tensor inner_encoder_ffn2_1_module_sequential_4_linear_bias = const()[name = tensor("inner_encoder_ffn2_1_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(16972608)))]; + tensor inner_encoder_ffn2_1_module_sequential_4_linear_weight = const()[name = tensor("inner_encoder_ffn2_1_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(16973696)))]; + tensor inner_encoder_layer_norm_1_bias = const()[name = tensor("inner_encoder_layer_norm_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18022336)))]; + tensor inner_encoder_layer_norm_1_weight = const()[name = tensor("inner_encoder_layer_norm_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18023424)))]; + tensor inner_encoder_ffn1_2_module_sequential_0_bias = const()[name = tensor("inner_encoder_ffn1_2_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18024512)))]; + tensor inner_encoder_ffn1_2_module_sequential_0_weight = const()[name = tensor("inner_encoder_ffn1_2_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18025600)))]; + tensor inner_encoder_ffn1_2_module_sequential_1_linear_bias = const()[name = tensor("inner_encoder_ffn1_2_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18026688)))]; + tensor inner_encoder_ffn1_2_module_sequential_1_linear_weight = const()[name = tensor("inner_encoder_ffn1_2_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18030848)))]; + tensor inner_encoder_ffn1_2_module_sequential_4_linear_bias = const()[name = tensor("inner_encoder_ffn1_2_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(19079488)))]; + tensor inner_encoder_ffn1_2_module_sequential_4_linear_weight = const()[name = tensor("inner_encoder_ffn1_2_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(19080576)))]; + tensor inner_encoder_ret_lns_2_bias = const()[name = tensor("inner_encoder_ret_lns_2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20129216)))]; + tensor inner_encoder_ret_lns_2_weight = const()[name = tensor("inner_encoder_ret_lns_2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20130304)))]; + tensor inner_encoder_q_proj_2_bias = const()[name = tensor("inner_encoder_q_proj_2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20131392)))]; + tensor inner_encoder_q_proj_2_weight = const()[name = tensor("inner_encoder_q_proj_2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20132480)))]; + tensor inner_encoder_k_proj_2_bias = const()[name = tensor("inner_encoder_k_proj_2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20394688)))]; + tensor inner_encoder_k_proj_2_weight = const()[name = tensor("inner_encoder_k_proj_2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20395776)))]; + tensor inner_encoder_v_proj_2_bias = const()[name = tensor("inner_encoder_v_proj_2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20657984)))]; + tensor inner_encoder_v_proj_2_weight = const()[name = tensor("inner_encoder_v_proj_2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20659072)))]; + tensor inner_encoder_g_proj_2_bias = const()[name = tensor("inner_encoder_g_proj_2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20921280)))]; + tensor inner_encoder_g_proj_2_weight = const()[name = tensor("inner_encoder_g_proj_2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20922368)))]; + tensor inner_encoder_out_proj_2_bias = const()[name = tensor("inner_encoder_out_proj_2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21184576)))]; + tensor inner_encoder_out_proj_2_weight = const()[name = tensor("inner_encoder_out_proj_2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21185664)))]; + tensor inner_encoder_conv_module_2_sequential_0_bias = const()[name = tensor("inner_encoder_conv_module_2_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21447872)))]; + tensor inner_encoder_conv_module_2_sequential_0_weight = const()[name = tensor("inner_encoder_conv_module_2_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21448960)))]; + tensor inner_encoder_conv_module_2_sequential_2_conv_bias = const()[name = tensor("inner_encoder_conv_module_2_sequential_2_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21450048)))]; + tensor inner_encoder_conv_module_2_sequential_2_conv_weight = const()[name = tensor("inner_encoder_conv_module_2_sequential_2_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21452160)))]; + tensor inner_encoder_conv_module_2_sequential_4_conv_weight = const()[name = tensor("inner_encoder_conv_module_2_sequential_4_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21976512)))]; + tensor inner_encoder_conv_module_2_sequential_5_running_var = const()[name = tensor("inner_encoder_conv_module_2_sequential_5_running_var"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21992960)))]; + tensor inner_encoder_conv_module_2_sequential_5_running_mean = const()[name = tensor("inner_encoder_conv_module_2_sequential_5_running_mean"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21994048)))]; + tensor inner_encoder_conv_module_2_sequential_5_bias = const()[name = tensor("inner_encoder_conv_module_2_sequential_5_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21995136)))]; + tensor inner_encoder_conv_module_2_sequential_5_weight = const()[name = tensor("inner_encoder_conv_module_2_sequential_5_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21996224)))]; + tensor inner_encoder_conv_module_2_sequential_7_conv_bias = const()[name = tensor("inner_encoder_conv_module_2_sequential_7_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21997312)))]; + tensor inner_encoder_conv_module_2_sequential_7_conv_weight = const()[name = tensor("inner_encoder_conv_module_2_sequential_7_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21998400)))]; + tensor inner_encoder_ffn2_2_module_sequential_0_bias = const()[name = tensor("inner_encoder_ffn2_2_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(22260608)))]; + tensor inner_encoder_ffn2_2_module_sequential_0_weight = const()[name = tensor("inner_encoder_ffn2_2_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(22261696)))]; + tensor inner_encoder_ffn2_2_module_sequential_1_linear_bias = const()[name = tensor("inner_encoder_ffn2_2_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(22262784)))]; + tensor inner_encoder_ffn2_2_module_sequential_1_linear_weight = const()[name = tensor("inner_encoder_ffn2_2_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(22266944)))]; + tensor inner_encoder_ffn2_2_module_sequential_4_linear_bias = const()[name = tensor("inner_encoder_ffn2_2_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(23315584)))]; + tensor inner_encoder_ffn2_2_module_sequential_4_linear_weight = const()[name = tensor("inner_encoder_ffn2_2_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(23316672)))]; + tensor inner_encoder_layer_norm_2_bias = const()[name = tensor("inner_encoder_layer_norm_2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(24365312)))]; + tensor inner_encoder_layer_norm_2_weight = const()[name = tensor("inner_encoder_layer_norm_2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(24366400)))]; + tensor inner_encoder_ffn1_3_module_sequential_0_bias = const()[name = tensor("inner_encoder_ffn1_3_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(24367488)))]; + tensor inner_encoder_ffn1_3_module_sequential_0_weight = const()[name = tensor("inner_encoder_ffn1_3_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(24368576)))]; + tensor inner_encoder_ffn1_3_module_sequential_1_linear_bias = const()[name = tensor("inner_encoder_ffn1_3_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(24369664)))]; + tensor inner_encoder_ffn1_3_module_sequential_1_linear_weight = const()[name = tensor("inner_encoder_ffn1_3_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(24373824)))]; + tensor inner_encoder_ffn1_3_module_sequential_4_linear_bias = const()[name = tensor("inner_encoder_ffn1_3_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(25422464)))]; + tensor inner_encoder_ffn1_3_module_sequential_4_linear_weight = const()[name = tensor("inner_encoder_ffn1_3_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(25423552)))]; + tensor inner_encoder_ret_lns_3_bias = const()[name = tensor("inner_encoder_ret_lns_3_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(26472192)))]; + tensor inner_encoder_ret_lns_3_weight = const()[name = tensor("inner_encoder_ret_lns_3_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(26473280)))]; + tensor inner_encoder_q_proj_3_bias = const()[name = tensor("inner_encoder_q_proj_3_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(26474368)))]; + tensor inner_encoder_q_proj_3_weight = const()[name = tensor("inner_encoder_q_proj_3_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(26475456)))]; + tensor inner_encoder_k_proj_3_bias = const()[name = tensor("inner_encoder_k_proj_3_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(26737664)))]; + tensor inner_encoder_k_proj_3_weight = const()[name = tensor("inner_encoder_k_proj_3_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(26738752)))]; + tensor inner_encoder_v_proj_3_bias = const()[name = tensor("inner_encoder_v_proj_3_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27000960)))]; + tensor inner_encoder_v_proj_3_weight = const()[name = tensor("inner_encoder_v_proj_3_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27002048)))]; + tensor inner_encoder_g_proj_3_bias = const()[name = tensor("inner_encoder_g_proj_3_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27264256)))]; + tensor inner_encoder_g_proj_3_weight = const()[name = tensor("inner_encoder_g_proj_3_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27265344)))]; + tensor inner_encoder_out_proj_3_bias = const()[name = tensor("inner_encoder_out_proj_3_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27527552)))]; + tensor inner_encoder_out_proj_3_weight = const()[name = tensor("inner_encoder_out_proj_3_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27528640)))]; + tensor inner_encoder_conv_module_3_sequential_0_bias = const()[name = tensor("inner_encoder_conv_module_3_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27790848)))]; + tensor inner_encoder_conv_module_3_sequential_0_weight = const()[name = tensor("inner_encoder_conv_module_3_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27791936)))]; + tensor inner_encoder_conv_module_3_sequential_2_conv_bias = const()[name = tensor("inner_encoder_conv_module_3_sequential_2_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27793024)))]; + tensor inner_encoder_conv_module_3_sequential_2_conv_weight = const()[name = tensor("inner_encoder_conv_module_3_sequential_2_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27795136)))]; + tensor inner_encoder_conv_module_3_sequential_4_conv_weight = const()[name = tensor("inner_encoder_conv_module_3_sequential_4_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28319488)))]; + tensor inner_encoder_conv_module_3_sequential_5_running_var = const()[name = tensor("inner_encoder_conv_module_3_sequential_5_running_var"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28335936)))]; + tensor inner_encoder_conv_module_3_sequential_5_running_mean = const()[name = tensor("inner_encoder_conv_module_3_sequential_5_running_mean"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28337024)))]; + tensor inner_encoder_conv_module_3_sequential_5_bias = const()[name = tensor("inner_encoder_conv_module_3_sequential_5_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28338112)))]; + tensor inner_encoder_conv_module_3_sequential_5_weight = const()[name = tensor("inner_encoder_conv_module_3_sequential_5_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28339200)))]; + tensor inner_encoder_conv_module_3_sequential_7_conv_bias = const()[name = tensor("inner_encoder_conv_module_3_sequential_7_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28340288)))]; + tensor inner_encoder_conv_module_3_sequential_7_conv_weight = const()[name = tensor("inner_encoder_conv_module_3_sequential_7_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28341376)))]; + tensor inner_encoder_ffn2_3_module_sequential_0_bias = const()[name = tensor("inner_encoder_ffn2_3_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28603584)))]; + tensor inner_encoder_ffn2_3_module_sequential_0_weight = const()[name = tensor("inner_encoder_ffn2_3_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28604672)))]; + tensor inner_encoder_ffn2_3_module_sequential_1_linear_bias = const()[name = tensor("inner_encoder_ffn2_3_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28605760)))]; + tensor inner_encoder_ffn2_3_module_sequential_1_linear_weight = const()[name = tensor("inner_encoder_ffn2_3_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28609920)))]; + tensor inner_encoder_ffn2_3_module_sequential_4_linear_bias = const()[name = tensor("inner_encoder_ffn2_3_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(29658560)))]; + tensor inner_encoder_ffn2_3_module_sequential_4_linear_weight = const()[name = tensor("inner_encoder_ffn2_3_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(29659648)))]; + tensor inner_encoder_layer_norm_3_bias = const()[name = tensor("inner_encoder_layer_norm_3_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(30708288)))]; + tensor inner_encoder_layer_norm_3_weight = const()[name = tensor("inner_encoder_layer_norm_3_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(30709376)))]; + tensor inner_decoder__causal_mask = const()[name = tensor("inner_decoder__causal_mask"), val = tensor([[0x1p+0, 0x0p+0, 0x0p+0], [0x1p+0, 0x1p+0, 0x0p+0], [0x1p+0, 0x1p+0, 0x1p+0]])]; + tensor inner_decoder_convert_bias = const()[name = tensor("inner_decoder_convert_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(30710464)))]; + tensor inner_decoder_convert_weight = const()[name = tensor("inner_decoder_convert_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(30711552)))]; + tensor inner_decoder_q_proj_0_bias = const()[name = tensor("inner_decoder_q_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(31235904)))]; + tensor inner_decoder_q_proj_0_weight = const()[name = tensor("inner_decoder_q_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(31236992)))]; + tensor inner_decoder_k_proj_0_bias = const()[name = tensor("inner_decoder_k_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(31499200)))]; + tensor inner_decoder_k_proj_0_weight = const()[name = tensor("inner_decoder_k_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(31500288)))]; + tensor inner_decoder_v_proj_0_bias = const()[name = tensor("inner_decoder_v_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(31762496)))]; + tensor inner_decoder_v_proj_0_weight = const()[name = tensor("inner_decoder_v_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(31763584)))]; + tensor inner_decoder_g_proj_0_bias = const()[name = tensor("inner_decoder_g_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32025792)))]; + tensor inner_decoder_g_proj_0_weight = const()[name = tensor("inner_decoder_g_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32026880)))]; + tensor inner_decoder_out_proj_0_bias = const()[name = tensor("inner_decoder_out_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32289088)))]; + tensor inner_decoder_out_proj_0_weight = const()[name = tensor("inner_decoder_out_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32290176)))]; + tensor inner_decoder_norm11_0_bias = const()[name = tensor("inner_decoder_norm11_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32552384)))]; + tensor inner_decoder_norm11_0_weight = const()[name = tensor("inner_decoder_norm11_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32553472)))]; + tensor inner_decoder_self_attn2_0_out_proj_bias = const()[name = tensor("inner_decoder_self_attn2_0_out_proj_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32554560)))]; + tensor inner_decoder_self_attn2_0_out_proj_weight = const()[name = tensor("inner_decoder_self_attn2_0_out_proj_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32555648)))]; + tensor inner_decoder_self_attn2_0_in_proj_bias = const()[name = tensor("inner_decoder_self_attn2_0_in_proj_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32817856)))]; + tensor inner_decoder_self_attn2_0_in_proj_weight = const()[name = tensor("inner_decoder_self_attn2_0_in_proj_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32820992)))]; + tensor inner_decoder_norm21_0_bias = const()[name = tensor("inner_decoder_norm21_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(33607488)))]; + tensor inner_decoder_norm21_0_weight = const()[name = tensor("inner_decoder_norm21_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(33608576)))]; + tensor inner_decoder_linear1_0_bias = const()[name = tensor("inner_decoder_linear1_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(33609664)))]; + tensor inner_decoder_linear1_0_weight = const()[name = tensor("inner_decoder_linear1_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(33617920)))]; + tensor inner_decoder_linear2_0_bias = const()[name = tensor("inner_decoder_linear2_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(35715136)))]; + tensor inner_decoder_linear2_0_weight = const()[name = tensor("inner_decoder_linear2_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(35716224)))]; + tensor inner_decoder_norm22_0_bias = const()[name = tensor("inner_decoder_norm22_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(37813440)))]; + tensor inner_decoder_norm22_0_weight = const()[name = tensor("inner_decoder_norm22_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(37814528)))]; + tensor inner_decoder_q_proj_1_bias = const()[name = tensor("inner_decoder_q_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(37815616)))]; + tensor inner_decoder_q_proj_1_weight = const()[name = tensor("inner_decoder_q_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(37816704)))]; + tensor inner_decoder_k_proj_1_bias = const()[name = tensor("inner_decoder_k_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38078912)))]; + tensor inner_decoder_k_proj_1_weight = const()[name = tensor("inner_decoder_k_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38080000)))]; + tensor inner_decoder_v_proj_1_bias = const()[name = tensor("inner_decoder_v_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38342208)))]; + tensor inner_decoder_v_proj_1_weight = const()[name = tensor("inner_decoder_v_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38343296)))]; + tensor inner_decoder_g_proj_1_bias = const()[name = tensor("inner_decoder_g_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38605504)))]; + tensor inner_decoder_g_proj_1_weight = const()[name = tensor("inner_decoder_g_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38606592)))]; + tensor inner_decoder_out_proj_1_bias = const()[name = tensor("inner_decoder_out_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38868800)))]; + tensor inner_decoder_out_proj_1_weight = const()[name = tensor("inner_decoder_out_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38869888)))]; + tensor inner_decoder_norm11_1_bias = const()[name = tensor("inner_decoder_norm11_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(39132096)))]; + tensor inner_decoder_norm11_1_weight = const()[name = tensor("inner_decoder_norm11_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(39133184)))]; + tensor inner_decoder_self_attn2_1_out_proj_bias = const()[name = tensor("inner_decoder_self_attn2_1_out_proj_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(39134272)))]; + tensor inner_decoder_self_attn2_1_out_proj_weight = const()[name = tensor("inner_decoder_self_attn2_1_out_proj_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(39135360)))]; + tensor inner_decoder_self_attn2_1_in_proj_bias = const()[name = tensor("inner_decoder_self_attn2_1_in_proj_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(39397568)))]; + tensor inner_decoder_self_attn2_1_in_proj_weight = const()[name = tensor("inner_decoder_self_attn2_1_in_proj_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(39400704)))]; + tensor inner_decoder_norm21_1_bias = const()[name = tensor("inner_decoder_norm21_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(40187200)))]; + tensor inner_decoder_norm21_1_weight = const()[name = tensor("inner_decoder_norm21_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(40188288)))]; + tensor inner_decoder_linear1_1_bias = const()[name = tensor("inner_decoder_linear1_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(40189376)))]; + tensor inner_decoder_linear1_1_weight = const()[name = tensor("inner_decoder_linear1_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(40197632)))]; + tensor inner_decoder_linear2_1_bias = const()[name = tensor("inner_decoder_linear2_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(42294848)))]; + tensor inner_decoder_linear2_1_weight = const()[name = tensor("inner_decoder_linear2_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(42295936)))]; + tensor inner_decoder_norm22_1_bias = const()[name = tensor("inner_decoder_norm22_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(44393152)))]; + tensor inner_decoder_norm22_1_weight = const()[name = tensor("inner_decoder_norm22_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(44394240)))]; + tensor var_19_begin_0 = const()[name = tensor("op_19_begin_0"), val = tensor([0, 0, 0])]; + tensor var_19_end_0 = const()[name = tensor("op_19_end_0"), val = tensor([1, 15, 23])]; + tensor var_19_end_mask_0 = const()[name = tensor("op_19_end_mask_0"), val = tensor([true, false, true])]; + tensor var_19 = slice_by_index(begin = var_19_begin_0, end = var_19_end_0, end_mask = var_19_end_mask_0, x = features)[name = tensor("op_19")]; + tensor var_29_begin_0 = const()[name = tensor("op_29_begin_0"), val = tensor([0, 10, 0])]; + tensor var_29_end_0 = const()[name = tensor("op_29_end_0"), val = tensor([1, 25, 23])]; + tensor var_29_end_mask_0 = const()[name = tensor("op_29_end_mask_0"), val = tensor([true, false, true])]; + tensor var_29 = slice_by_index(begin = var_29_begin_0, end = var_29_end_0, end_mask = var_29_end_mask_0, x = features)[name = tensor("op_29")]; + tensor var_39_begin_0 = const()[name = tensor("op_39_begin_0"), val = tensor([0, 20, 0])]; + tensor var_39_end_0 = const()[name = tensor("op_39_end_0"), val = tensor([1, 1, 23])]; + tensor var_39_end_mask_0 = const()[name = tensor("op_39_end_mask_0"), val = tensor([true, true, true])]; + tensor var_39 = slice_by_index(begin = var_39_begin_0, end = var_39_end_0, end_mask = var_39_end_mask_0, x = features)[name = tensor("op_39")]; + tensor stacked_axis_0 = const()[name = tensor("stacked_axis_0"), val = tensor(1)]; + tensor stacked = stack(axis = stacked_axis_0, values = (var_19, var_29, var_39))[name = tensor("stacked")]; + tensor var_46 = const()[name = tensor("op_46"), val = tensor([1, 3, 345])]; + tensor input_1 = reshape(shape = var_46, x = stacked)[name = tensor("input_1")]; + tensor var_49 = const()[name = tensor("op_49"), val = tensor(0x1p+0)]; + tensor var_55 = const()[name = tensor("op_55"), val = tensor(true)]; + tensor var_56 = const()[name = tensor("op_56"), val = tensor(0x1.4f8b58p-17)]; + tensor var_59 = const()[name = tensor("op_59"), val = tensor(0)]; + tensor var_61 = const()[name = tensor("op_61"), val = tensor(2)]; + tensor var_62 = const()[name = tensor("op_62"), val = tensor(-1)]; + tensor var_64 = const()[name = tensor("op_64"), val = tensor(0x1.0c6f7ap-20)]; + tensor var_69 = const()[name = tensor("op_69"), val = tensor(0x1.5798eep-27)]; + tensor var_72 = const()[name = tensor("op_72"), val = tensor(1)]; + tensor input_3 = linear(bias = inner_encoder_input_projection_linear_bias, weight = inner_encoder_input_projection_linear_weight, x = input_1)[name = tensor("linear_0")]; tensor input_5_axes_0 = const()[name = tensor("input_5_axes_0"), val = tensor([-1])]; - tensor input_5 = layer_norm(axes = input_5_axes_0, beta = encoder_ffn1_0_module_sequential_0_bias, epsilon = var_28, gamma = encoder_ffn1_0_module_sequential_0_weight, x = input_3)[name = tensor("input_5")]; - tensor inputs_1 = linear(bias = encoder_ffn1_0_module_sequential_1_linear_bias, weight = encoder_ffn1_0_module_sequential_1_linear_weight, x = input_5)[name = tensor("linear_1")]; - tensor input_7 = silu(x = inputs_1)[name = tensor("input_7")]; - tensor input_11 = linear(bias = encoder_ffn1_0_module_sequential_4_linear_bias, weight = encoder_ffn1_0_module_sequential_4_linear_weight, x = input_7)[name = tensor("linear_2")]; - tensor var_147 = const()[name = tensor("op_147"), val = tensor(0x1p-1)]; - tensor var_148 = mul(x = input_11, y = var_147)[name = tensor("op_148")]; - tensor input_13 = add(x = var_148, y = input_3)[name = tensor("input_13")]; + tensor input_5 = layer_norm(axes = input_5_axes_0, beta = inner_encoder_pre_layer_norm_bias, epsilon = var_56, gamma = inner_encoder_pre_layer_norm_weight, x = input_3)[name = tensor("input_5")]; + tensor input_7_axes_0 = const()[name = tensor("input_7_axes_0"), val = tensor([-1])]; + tensor input_7 = layer_norm(axes = input_7_axes_0, beta = inner_encoder_ffn1_0_module_sequential_0_bias, epsilon = var_56, gamma = inner_encoder_ffn1_0_module_sequential_0_weight, x = input_5)[name = tensor("input_7")]; + tensor inputs_1 = linear(bias = inner_encoder_ffn1_0_module_sequential_1_linear_bias, weight = inner_encoder_ffn1_0_module_sequential_1_linear_weight, x = input_7)[name = tensor("linear_1")]; + tensor input_9 = silu(x = inputs_1)[name = tensor("input_9")]; + tensor input_13 = linear(bias = inner_encoder_ffn1_0_module_sequential_4_linear_bias, weight = inner_encoder_ffn1_0_module_sequential_4_linear_weight, x = input_9)[name = tensor("linear_2")]; + tensor var_193 = const()[name = tensor("op_193"), val = tensor(0x1p-1)]; + tensor var_194 = mul(x = input_13, y = var_193)[name = tensor("op_194")]; + tensor input_15 = add(x = var_194, y = input_5)[name = tensor("input_15")]; tensor x_1_axes_0 = const()[name = tensor("x_1_axes_0"), val = tensor([-1])]; - tensor x_1 = layer_norm(axes = x_1_axes_0, beta = encoder_ret_lns_0_bias, epsilon = var_28, gamma = encoder_ret_lns_0_weight, x = input_13)[name = tensor("x_1")]; + tensor x_1 = layer_norm(axes = x_1_axes_0, beta = inner_encoder_ret_lns_0_bias, epsilon = var_56, gamma = inner_encoder_ret_lns_0_weight, x = input_15)[name = tensor("x_1")]; tensor prev_kv_1_begin_0 = const()[name = tensor("prev_kv_1_begin_0"), val = tensor([0, 0, 0, 0, 0])]; tensor prev_kv_1_end_0 = const()[name = tensor("prev_kv_1_end_0"), val = tensor([1, 1, 4, 64, 64])]; tensor prev_kv_1_end_mask_0 = const()[name = tensor("prev_kv_1_end_mask_0"), val = tensor([false, true, true, true, true])]; @@ -239,163 +257,163 @@ program(1.0) tensor prev_scale_1_end_mask_0 = const()[name = tensor("prev_scale_1_end_mask_0"), val = tensor([false, true])]; tensor prev_scale_1_squeeze_mask_0 = const()[name = tensor("prev_scale_1_squeeze_mask_0"), val = tensor([true, false])]; tensor prev_scale_1 = slice_by_index(begin = prev_scale_1_begin_0, end = prev_scale_1_end_0, end_mask = prev_scale_1_end_mask_0, squeeze_mask = prev_scale_1_squeeze_mask_0, x = enc_scale)[name = tensor("prev_scale_1")]; - tensor var_162 = linear(bias = encoder_q_proj_0_bias, weight = encoder_q_proj_0_weight, x = x_1)[name = tensor("linear_3")]; - tensor var_163 = const()[name = tensor("op_163"), val = tensor([1, 3, 4, 64])]; - tensor var_164 = reshape(shape = var_163, x = var_162)[name = tensor("op_164")]; + tensor var_208 = linear(bias = inner_encoder_q_proj_0_bias, weight = inner_encoder_q_proj_0_weight, x = x_1)[name = tensor("linear_3")]; + tensor var_209 = const()[name = tensor("op_209"), val = tensor([1, 3, 4, 64])]; + tensor var_210 = reshape(shape = var_209, x = var_208)[name = tensor("op_210")]; tensor q_1_perm_0 = const()[name = tensor("q_1_perm_0"), val = tensor([0, 2, 1, 3])]; - tensor var_168 = linear(bias = encoder_k_proj_0_bias, weight = encoder_k_proj_0_weight, x = x_1)[name = tensor("linear_4")]; - tensor var_169 = const()[name = tensor("op_169"), val = tensor(0x1p-3)]; - tensor var_170 = mul(x = var_168, y = var_169)[name = tensor("op_170")]; - tensor var_171 = const()[name = tensor("op_171"), val = tensor([1, 3, 4, 64])]; - tensor var_172 = reshape(shape = var_171, x = var_170)[name = tensor("op_172")]; + tensor var_214 = linear(bias = inner_encoder_k_proj_0_bias, weight = inner_encoder_k_proj_0_weight, x = x_1)[name = tensor("linear_4")]; + tensor var_215 = const()[name = tensor("op_215"), val = tensor(0x1p-3)]; + tensor var_216 = mul(x = var_214, y = var_215)[name = tensor("op_216")]; + tensor var_217 = const()[name = tensor("op_217"), val = tensor([1, 3, 4, 64])]; + tensor var_218 = reshape(shape = var_217, x = var_216)[name = tensor("op_218")]; tensor k_1_perm_0 = const()[name = tensor("k_1_perm_0"), val = tensor([0, 2, 1, 3])]; - tensor var_176 = linear(bias = encoder_v_proj_0_bias, weight = encoder_v_proj_0_weight, x = x_1)[name = tensor("linear_5")]; - tensor var_177 = const()[name = tensor("op_177"), val = tensor([1, 3, 4, 64])]; - tensor var_178 = reshape(shape = var_177, x = var_176)[name = tensor("op_178")]; + tensor var_222 = linear(bias = inner_encoder_v_proj_0_bias, weight = inner_encoder_v_proj_0_weight, x = x_1)[name = tensor("linear_5")]; + tensor var_223 = const()[name = tensor("op_223"), val = tensor([1, 3, 4, 64])]; + tensor var_224 = reshape(shape = var_223, x = var_222)[name = tensor("op_224")]; tensor v_1_perm_0 = const()[name = tensor("v_1_perm_0"), val = tensor([0, 2, 1, 3])]; - tensor input_17 = linear(bias = encoder_g_proj_0_bias, weight = encoder_g_proj_0_weight, x = x_1)[name = tensor("linear_6")]; + tensor input_19 = linear(bias = inner_encoder_g_proj_0_bias, weight = inner_encoder_g_proj_0_weight, x = x_1)[name = tensor("linear_6")]; tensor sqrt_s0_1 = sqrt(x = prev_scale_1)[name = tensor("sqrt_s0_1")]; - tensor s_t_1 = add(x = prev_scale_1, y = encoder__t_index)[name = tensor("s_t_1")]; + tensor s_t_1 = add(x = prev_scale_1, y = inner_encoder__t_index)[name = tensor("s_t_1")]; tensor sqrt_s_t_1 = sqrt(x = s_t_1)[name = tensor("sqrt_s_t_1")]; tensor qk_1_transpose_x_1 = const()[name = tensor("qk_1_transpose_x_1"), val = tensor(false)]; tensor qk_1_transpose_y_1 = const()[name = tensor("qk_1_transpose_y_1"), val = tensor(true)]; - tensor k_1 = transpose(perm = k_1_perm_0, x = var_172)[name = tensor("transpose_57")]; - tensor q_1 = transpose(perm = q_1_perm_0, x = var_164)[name = tensor("transpose_58")]; + tensor k_1 = transpose(perm = k_1_perm_0, x = var_218)[name = tensor("transpose_57")]; + tensor q_1 = transpose(perm = q_1_perm_0, x = var_210)[name = tensor("transpose_58")]; tensor qk_1 = matmul(transpose_x = qk_1_transpose_x_1, transpose_y = qk_1_transpose_y_1, x = q_1, y = k_1)[name = tensor("qk_1")]; - tensor var_188 = const()[name = tensor("op_188"), val = tensor([3, 1])]; - tensor var_189 = reshape(shape = var_188, x = sqrt_s_t_1)[name = tensor("op_189")]; - tensor M_1 = real_div(x = encoder__causal_mask, y = var_189)[name = tensor("M_1")]; - tensor var_191 = mul(x = qk_1, y = M_1)[name = tensor("op_191")]; + tensor var_234 = const()[name = tensor("op_234"), val = tensor([3, 1])]; + tensor var_235 = reshape(shape = var_234, x = sqrt_s_t_1)[name = tensor("op_235")]; + tensor M_1 = real_div(x = inner_encoder__causal_mask, y = var_235)[name = tensor("M_1")]; + tensor var_237 = mul(x = qk_1, y = M_1)[name = tensor("op_237")]; tensor inner_1_transpose_x_0 = const()[name = tensor("inner_1_transpose_x_0"), val = tensor(false)]; tensor inner_1_transpose_y_0 = const()[name = tensor("inner_1_transpose_y_0"), val = tensor(false)]; - tensor v_1 = transpose(perm = v_1_perm_0, x = var_178)[name = tensor("transpose_56")]; - tensor inner_1 = matmul(transpose_x = inner_1_transpose_x_0, transpose_y = inner_1_transpose_y_0, x = var_191, y = v_1)[name = tensor("inner_1")]; - tensor var_193_transpose_x_0 = const()[name = tensor("op_193_transpose_x_0"), val = tensor(false)]; - tensor var_193_transpose_y_0 = const()[name = tensor("op_193_transpose_y_0"), val = tensor(false)]; - tensor var_193 = matmul(transpose_x = var_193_transpose_x_0, transpose_y = var_193_transpose_y_0, x = q_1, y = prev_kv_1)[name = tensor("op_193")]; - tensor var_194 = real_div(x = sqrt_s0_1, y = sqrt_s_t_1)[name = tensor("op_194")]; - tensor var_195 = const()[name = tensor("op_195"), val = tensor([1, 1, 3, 1])]; - tensor var_196 = reshape(shape = var_195, x = var_194)[name = tensor("op_196")]; - tensor cross_1 = mul(x = var_193, y = var_196)[name = tensor("cross_1")]; + tensor v_1 = transpose(perm = v_1_perm_0, x = var_224)[name = tensor("transpose_56")]; + tensor inner_1 = matmul(transpose_x = inner_1_transpose_x_0, transpose_y = inner_1_transpose_y_0, x = var_237, y = v_1)[name = tensor("inner_1")]; + tensor var_239_transpose_x_0 = const()[name = tensor("op_239_transpose_x_0"), val = tensor(false)]; + tensor var_239_transpose_y_0 = const()[name = tensor("op_239_transpose_y_0"), val = tensor(false)]; + tensor var_239 = matmul(transpose_x = var_239_transpose_x_0, transpose_y = var_239_transpose_y_0, x = q_1, y = prev_kv_1)[name = tensor("op_239")]; + tensor var_240 = real_div(x = sqrt_s0_1, y = sqrt_s_t_1)[name = tensor("op_240")]; + tensor var_241 = const()[name = tensor("op_241"), val = tensor([1, 1, 3, 1])]; + tensor var_242 = reshape(shape = var_241, x = var_240)[name = tensor("op_242")]; + tensor cross_1 = mul(x = var_239, y = var_242)[name = tensor("cross_1")]; tensor out_1 = add(x = inner_1, y = cross_1)[name = tensor("out_1")]; - tensor var_199 = mul(x = prev_kv_1, y = sqrt_s0_1)[name = tensor("op_199")]; - tensor var_201_transpose_x_1 = const()[name = tensor("op_201_transpose_x_1"), val = tensor(true)]; - tensor var_201_transpose_y_1 = const()[name = tensor("op_201_transpose_y_1"), val = tensor(false)]; - tensor var_201 = matmul(transpose_x = var_201_transpose_x_1, transpose_y = var_201_transpose_y_1, x = k_1, y = v_1)[name = tensor("op_201")]; - tensor new_kv_unnorm_1 = add(x = var_199, y = var_201)[name = tensor("new_kv_unnorm_1")]; - tensor var_203 = const()[name = tensor("op_203"), val = tensor(0x1.8p+1)]; - tensor new_scale_1 = add(x = prev_scale_1, y = var_203)[name = tensor("new_scale_1")]; - tensor var_205 = sqrt(x = new_scale_1)[name = tensor("op_205")]; - tensor var_206 = real_div(x = new_kv_unnorm_1, y = var_205)[name = tensor("op_206")]; - tensor var_207_perm_0 = const()[name = tensor("op_207_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_245 = mul(x = prev_kv_1, y = sqrt_s0_1)[name = tensor("op_245")]; + tensor var_247_transpose_x_1 = const()[name = tensor("op_247_transpose_x_1"), val = tensor(true)]; + tensor var_247_transpose_y_1 = const()[name = tensor("op_247_transpose_y_1"), val = tensor(false)]; + tensor var_247 = matmul(transpose_x = var_247_transpose_x_1, transpose_y = var_247_transpose_y_1, x = k_1, y = v_1)[name = tensor("op_247")]; + tensor new_kv_unnorm_1 = add(x = var_245, y = var_247)[name = tensor("new_kv_unnorm_1")]; + tensor var_249 = const()[name = tensor("op_249"), val = tensor(0x1.8p+1)]; + tensor new_scale_1 = add(x = prev_scale_1, y = var_249)[name = tensor("new_scale_1")]; + tensor var_251 = sqrt(x = new_scale_1)[name = tensor("op_251")]; + tensor var_252 = real_div(x = new_kv_unnorm_1, y = var_251)[name = tensor("op_252")]; + tensor var_253_perm_0 = const()[name = tensor("op_253_perm_0"), val = tensor([0, 2, 1, 3])]; tensor out_3_axes_0 = const()[name = tensor("out_3_axes_0"), val = tensor([-1])]; - tensor var_207 = transpose(perm = var_207_perm_0, x = out_1)[name = tensor("transpose_55")]; - tensor out_3 = layer_norm(axes = out_3_axes_0, epsilon = var_17, x = var_207)[name = tensor("out_3")]; - tensor var_211 = const()[name = tensor("op_211"), val = tensor([1, 3, 256])]; - tensor out_5 = reshape(shape = var_211, x = out_3)[name = tensor("out_5")]; - tensor var_213 = silu(x = input_17)[name = tensor("op_213")]; - tensor input_19 = mul(x = var_213, y = out_5)[name = tensor("input_19")]; - tensor ret_out_1 = linear(bias = encoder_out_proj_0_bias, weight = encoder_out_proj_0_weight, x = input_19)[name = tensor("linear_7")]; - tensor x_3 = add(x = input_13, y = ret_out_1)[name = tensor("x_3")]; + tensor var_253 = transpose(perm = var_253_perm_0, x = out_1)[name = tensor("transpose_55")]; + tensor out_3 = layer_norm(axes = out_3_axes_0, epsilon = var_64, x = var_253)[name = tensor("out_3")]; + tensor var_257 = const()[name = tensor("op_257"), val = tensor([1, 3, 256])]; + tensor out_5 = reshape(shape = var_257, x = out_3)[name = tensor("out_5")]; + tensor var_259 = silu(x = input_19)[name = tensor("op_259")]; + tensor input_21 = mul(x = var_259, y = out_5)[name = tensor("input_21")]; + tensor ret_out_1 = linear(bias = inner_encoder_out_proj_0_bias, weight = inner_encoder_out_proj_0_weight, x = input_21)[name = tensor("linear_7")]; + tensor x_3 = add(x = input_15, y = ret_out_1)[name = tensor("x_3")]; tensor window_1_begin_0 = const()[name = tensor("window_1_begin_0"), val = tensor([0, 0, 0, 0])]; tensor window_1_end_0 = const()[name = tensor("window_1_end_0"), val = tensor([1, 1, 16, 256])]; tensor window_1_end_mask_0 = const()[name = tensor("window_1_end_mask_0"), val = tensor([false, true, true, true])]; tensor window_1_squeeze_mask_0 = const()[name = tensor("window_1_squeeze_mask_0"), val = tensor([true, false, false, false])]; tensor window_1 = slice_by_index(begin = window_1_begin_0, end = window_1_end_0, end_mask = window_1_end_mask_0, squeeze_mask = window_1_squeeze_mask_0, x = enc_conv_cache)[name = tensor("window_1")]; - tensor var_221_begin_0 = const()[name = tensor("op_221_begin_0"), val = tensor([0, 0, 0])]; - tensor var_221_end_0 = const()[name = tensor("op_221_end_0"), val = tensor([1, 1, 256])]; - tensor var_221_end_mask_0 = const()[name = tensor("op_221_end_mask_0"), val = tensor([true, false, true])]; - tensor var_221 = slice_by_index(begin = var_221_begin_0, end = var_221_end_0, end_mask = var_221_end_mask_0, x = x_3)[name = tensor("op_221")]; - tensor var_224_begin_0 = const()[name = tensor("op_224_begin_0"), val = tensor([0, 1, 0])]; - tensor var_224_end_0 = const()[name = tensor("op_224_end_0"), val = tensor([1, 16, 256])]; - tensor var_224_end_mask_0 = const()[name = tensor("op_224_end_mask_0"), val = tensor([true, true, true])]; - tensor var_224 = slice_by_index(begin = var_224_begin_0, end = var_224_end_0, end_mask = var_224_end_mask_0, x = window_1)[name = tensor("op_224")]; + tensor var_267_begin_0 = const()[name = tensor("op_267_begin_0"), val = tensor([0, 0, 0])]; + tensor var_267_end_0 = const()[name = tensor("op_267_end_0"), val = tensor([1, 1, 256])]; + tensor var_267_end_mask_0 = const()[name = tensor("op_267_end_mask_0"), val = tensor([true, false, true])]; + tensor var_267 = slice_by_index(begin = var_267_begin_0, end = var_267_end_0, end_mask = var_267_end_mask_0, x = x_3)[name = tensor("op_267")]; + tensor var_270_begin_0 = const()[name = tensor("op_270_begin_0"), val = tensor([0, 1, 0])]; + tensor var_270_end_0 = const()[name = tensor("op_270_end_0"), val = tensor([1, 16, 256])]; + tensor var_270_end_mask_0 = const()[name = tensor("op_270_end_mask_0"), val = tensor([true, true, true])]; + tensor var_270 = slice_by_index(begin = var_270_begin_0, end = var_270_end_0, end_mask = var_270_end_mask_0, x = window_1)[name = tensor("op_270")]; tensor window_3_interleave_0 = const()[name = tensor("window_3_interleave_0"), val = tensor(false)]; - tensor window_3 = concat(axis = var_26, interleave = window_3_interleave_0, values = (var_224, var_221))[name = tensor("window_3")]; - tensor var_229_begin_0 = const()[name = tensor("op_229_begin_0"), val = tensor([0, 1, 0])]; - tensor var_229_end_0 = const()[name = tensor("op_229_end_0"), val = tensor([1, 2, 256])]; - tensor var_229_end_mask_0 = const()[name = tensor("op_229_end_mask_0"), val = tensor([true, false, true])]; - tensor var_229 = slice_by_index(begin = var_229_begin_0, end = var_229_end_0, end_mask = var_229_end_mask_0, x = x_3)[name = tensor("op_229")]; - tensor var_232_begin_0 = const()[name = tensor("op_232_begin_0"), val = tensor([0, 1, 0])]; - tensor var_232_end_0 = const()[name = tensor("op_232_end_0"), val = tensor([1, 16, 256])]; - tensor var_232_end_mask_0 = const()[name = tensor("op_232_end_mask_0"), val = tensor([true, true, true])]; - tensor var_232 = slice_by_index(begin = var_232_begin_0, end = var_232_end_0, end_mask = var_232_end_mask_0, x = window_3)[name = tensor("op_232")]; + tensor window_3 = concat(axis = var_72, interleave = window_3_interleave_0, values = (var_270, var_267))[name = tensor("window_3")]; + tensor var_275_begin_0 = const()[name = tensor("op_275_begin_0"), val = tensor([0, 1, 0])]; + tensor var_275_end_0 = const()[name = tensor("op_275_end_0"), val = tensor([1, 2, 256])]; + tensor var_275_end_mask_0 = const()[name = tensor("op_275_end_mask_0"), val = tensor([true, false, true])]; + tensor var_275 = slice_by_index(begin = var_275_begin_0, end = var_275_end_0, end_mask = var_275_end_mask_0, x = x_3)[name = tensor("op_275")]; + tensor var_278_begin_0 = const()[name = tensor("op_278_begin_0"), val = tensor([0, 1, 0])]; + tensor var_278_end_0 = const()[name = tensor("op_278_end_0"), val = tensor([1, 16, 256])]; + tensor var_278_end_mask_0 = const()[name = tensor("op_278_end_mask_0"), val = tensor([true, true, true])]; + tensor var_278 = slice_by_index(begin = var_278_begin_0, end = var_278_end_0, end_mask = var_278_end_mask_0, x = window_3)[name = tensor("op_278")]; tensor window_5_interleave_0 = const()[name = tensor("window_5_interleave_0"), val = tensor(false)]; - tensor window_5 = concat(axis = var_26, interleave = window_5_interleave_0, values = (var_232, var_229))[name = tensor("window_5")]; - tensor var_237_begin_0 = const()[name = tensor("op_237_begin_0"), val = tensor([0, 2, 0])]; - tensor var_237_end_0 = const()[name = tensor("op_237_end_0"), val = tensor([1, 1, 256])]; - tensor var_237_end_mask_0 = const()[name = tensor("op_237_end_mask_0"), val = tensor([true, true, true])]; - tensor var_237 = slice_by_index(begin = var_237_begin_0, end = var_237_end_0, end_mask = var_237_end_mask_0, x = x_3)[name = tensor("op_237")]; - tensor var_240_begin_0 = const()[name = tensor("op_240_begin_0"), val = tensor([0, 1, 0])]; - tensor var_240_end_0 = const()[name = tensor("op_240_end_0"), val = tensor([1, 16, 256])]; - tensor var_240_end_mask_0 = const()[name = tensor("op_240_end_mask_0"), val = tensor([true, true, true])]; - tensor var_240 = slice_by_index(begin = var_240_begin_0, end = var_240_end_0, end_mask = var_240_end_mask_0, x = window_5)[name = tensor("op_240")]; + tensor window_5 = concat(axis = var_72, interleave = window_5_interleave_0, values = (var_278, var_275))[name = tensor("window_5")]; + tensor var_283_begin_0 = const()[name = tensor("op_283_begin_0"), val = tensor([0, 2, 0])]; + tensor var_283_end_0 = const()[name = tensor("op_283_end_0"), val = tensor([1, 1, 256])]; + tensor var_283_end_mask_0 = const()[name = tensor("op_283_end_mask_0"), val = tensor([true, true, true])]; + tensor var_283 = slice_by_index(begin = var_283_begin_0, end = var_283_end_0, end_mask = var_283_end_mask_0, x = x_3)[name = tensor("op_283")]; + tensor var_286_begin_0 = const()[name = tensor("op_286_begin_0"), val = tensor([0, 1, 0])]; + tensor var_286_end_0 = const()[name = tensor("op_286_end_0"), val = tensor([1, 16, 256])]; + tensor var_286_end_mask_0 = const()[name = tensor("op_286_end_mask_0"), val = tensor([true, true, true])]; + tensor var_286 = slice_by_index(begin = var_286_begin_0, end = var_286_end_0, end_mask = var_286_end_mask_0, x = window_5)[name = tensor("op_286")]; tensor window_7_interleave_0 = const()[name = tensor("window_7_interleave_0"), val = tensor(false)]; - tensor window_7 = concat(axis = var_26, interleave = window_7_interleave_0, values = (var_240, var_237))[name = tensor("window_7")]; - tensor input_21_interleave_0 = const()[name = tensor("input_21_interleave_0"), val = tensor(false)]; - tensor input_21 = concat(axis = var_23, interleave = input_21_interleave_0, values = (window_3, window_5, window_7))[name = tensor("input_21")]; + tensor window_7 = concat(axis = var_72, interleave = window_7_interleave_0, values = (var_286, var_283))[name = tensor("window_7")]; + tensor input_23_interleave_0 = const()[name = tensor("input_23_interleave_0"), val = tensor(false)]; + tensor input_23 = concat(axis = var_59, interleave = input_23_interleave_0, values = (window_3, window_5, window_7))[name = tensor("input_23")]; tensor x_5_axes_0 = const()[name = tensor("x_5_axes_0"), val = tensor([-1])]; - tensor x_5 = layer_norm(axes = x_5_axes_0, beta = encoder_conv_module_0_sequential_0_bias, epsilon = var_28, gamma = encoder_conv_module_0_sequential_0_weight, x = input_21)[name = tensor("x_5")]; - tensor input_23_perm_0 = const()[name = tensor("input_23_perm_0"), val = tensor([0, 2, 1])]; + tensor x_5 = layer_norm(axes = x_5_axes_0, beta = inner_encoder_conv_module_0_sequential_0_bias, epsilon = var_56, gamma = inner_encoder_conv_module_0_sequential_0_weight, x = input_23)[name = tensor("x_5")]; + tensor input_25_perm_0 = const()[name = tensor("input_25_perm_0"), val = tensor([0, 2, 1])]; tensor inputs_3_pad_type_0 = const()[name = tensor("inputs_3_pad_type_0"), val = tensor("valid")]; tensor inputs_3_strides_0 = const()[name = tensor("inputs_3_strides_0"), val = tensor([1])]; tensor inputs_3_pad_0 = const()[name = tensor("inputs_3_pad_0"), val = tensor([0, 0])]; tensor inputs_3_dilations_0 = const()[name = tensor("inputs_3_dilations_0"), val = tensor([1])]; tensor inputs_3_groups_0 = const()[name = tensor("inputs_3_groups_0"), val = tensor(1)]; - tensor input_23 = transpose(perm = input_23_perm_0, x = x_5)[name = tensor("transpose_54")]; - tensor inputs_3 = conv(bias = encoder_conv_module_0_sequential_2_conv_bias, dilations = inputs_3_dilations_0, groups = inputs_3_groups_0, pad = inputs_3_pad_0, pad_type = inputs_3_pad_type_0, strides = inputs_3_strides_0, weight = encoder_conv_module_0_sequential_2_conv_weight, x = input_23)[name = tensor("inputs_3")]; - tensor var_265_split_sizes_0 = const()[name = tensor("op_265_split_sizes_0"), val = tensor([256, 256])]; - tensor var_265_axis_0 = const()[name = tensor("op_265_axis_0"), val = tensor(1)]; - tensor var_265_0, tensor var_265_1 = split(axis = var_265_axis_0, split_sizes = var_265_split_sizes_0, x = inputs_3)[name = tensor("op_265")]; - tensor var_267 = sigmoid(x = var_265_1)[name = tensor("op_267")]; - tensor inputs_5 = mul(x = var_265_0, y = var_267)[name = tensor("inputs_5")]; + tensor input_25 = transpose(perm = input_25_perm_0, x = x_5)[name = tensor("transpose_54")]; + tensor inputs_3 = conv(bias = inner_encoder_conv_module_0_sequential_2_conv_bias, dilations = inputs_3_dilations_0, groups = inputs_3_groups_0, pad = inputs_3_pad_0, pad_type = inputs_3_pad_type_0, strides = inputs_3_strides_0, weight = inner_encoder_conv_module_0_sequential_2_conv_weight, x = input_25)[name = tensor("inputs_3")]; + tensor var_311_split_sizes_0 = const()[name = tensor("op_311_split_sizes_0"), val = tensor([256, 256])]; + tensor var_311_axis_0 = const()[name = tensor("op_311_axis_0"), val = tensor(1)]; + tensor var_311_0, tensor var_311_1 = split(axis = var_311_axis_0, split_sizes = var_311_split_sizes_0, x = inputs_3)[name = tensor("op_311")]; + tensor var_313 = sigmoid(x = var_311_1)[name = tensor("op_313")]; + tensor inputs_5 = mul(x = var_311_0, y = var_313)[name = tensor("inputs_5")]; tensor outputs_aug_1_pad_type_0 = const()[name = tensor("outputs_aug_1_pad_type_0"), val = tensor("custom")]; tensor outputs_aug_1_pad_0 = const()[name = tensor("outputs_aug_1_pad_0"), val = tensor([15, 15])]; tensor outputs_aug_1_groups_0 = const()[name = tensor("outputs_aug_1_groups_0"), val = tensor(256)]; tensor outputs_aug_1_strides_0 = const()[name = tensor("outputs_aug_1_strides_0"), val = tensor([1])]; tensor outputs_aug_1_dilations_0 = const()[name = tensor("outputs_aug_1_dilations_0"), val = tensor([1])]; - tensor outputs_aug_1 = conv(dilations = outputs_aug_1_dilations_0, groups = outputs_aug_1_groups_0, pad = outputs_aug_1_pad_0, pad_type = outputs_aug_1_pad_type_0, strides = outputs_aug_1_strides_0, weight = encoder_conv_module_0_sequential_4_conv_weight, x = inputs_5)[name = tensor("outputs_aug_1")]; - tensor input_25_begin_0 = const()[name = tensor("input_25_begin_0"), val = tensor([0, 0, 0])]; - tensor input_25_end_0 = const()[name = tensor("input_25_end_0"), val = tensor([3, 256, 16])]; - tensor input_25_end_mask_0 = const()[name = tensor("input_25_end_mask_0"), val = tensor([true, true, false])]; - tensor input_25 = slice_by_index(begin = input_25_begin_0, end = input_25_end_0, end_mask = input_25_end_mask_0, x = outputs_aug_1)[name = tensor("input_25")]; - tensor inputs_7 = batch_norm(beta = encoder_conv_module_0_sequential_5_bias, epsilon = var_28, gamma = encoder_conv_module_0_sequential_5_weight, mean = encoder_conv_module_0_sequential_5_running_mean, variance = encoder_conv_module_0_sequential_5_running_var, x = input_25)[name = tensor("inputs_7")]; - tensor input_27 = silu(x = inputs_7)[name = tensor("input_27")]; - tensor input_29_pad_type_0 = const()[name = tensor("input_29_pad_type_0"), val = tensor("valid")]; - tensor input_29_strides_0 = const()[name = tensor("input_29_strides_0"), val = tensor([1])]; - tensor input_29_pad_0 = const()[name = tensor("input_29_pad_0"), val = tensor([0, 0])]; - tensor input_29_dilations_0 = const()[name = tensor("input_29_dilations_0"), val = tensor([1])]; - tensor input_29_groups_0 = const()[name = tensor("input_29_groups_0"), val = tensor(1)]; - tensor input_29 = conv(bias = encoder_conv_module_0_sequential_7_conv_bias, dilations = input_29_dilations_0, groups = input_29_groups_0, pad = input_29_pad_0, pad_type = input_29_pad_type_0, strides = input_29_strides_0, weight = encoder_conv_module_0_sequential_7_conv_weight, x = input_27)[name = tensor("input_29")]; + tensor outputs_aug_1 = conv(dilations = outputs_aug_1_dilations_0, groups = outputs_aug_1_groups_0, pad = outputs_aug_1_pad_0, pad_type = outputs_aug_1_pad_type_0, strides = outputs_aug_1_strides_0, weight = inner_encoder_conv_module_0_sequential_4_conv_weight, x = inputs_5)[name = tensor("outputs_aug_1")]; + tensor input_27_begin_0 = const()[name = tensor("input_27_begin_0"), val = tensor([0, 0, 0])]; + tensor input_27_end_0 = const()[name = tensor("input_27_end_0"), val = tensor([3, 256, 16])]; + tensor input_27_end_mask_0 = const()[name = tensor("input_27_end_mask_0"), val = tensor([true, true, false])]; + tensor input_27 = slice_by_index(begin = input_27_begin_0, end = input_27_end_0, end_mask = input_27_end_mask_0, x = outputs_aug_1)[name = tensor("input_27")]; + tensor inputs_7 = batch_norm(beta = inner_encoder_conv_module_0_sequential_5_bias, epsilon = var_56, gamma = inner_encoder_conv_module_0_sequential_5_weight, mean = inner_encoder_conv_module_0_sequential_5_running_mean, variance = inner_encoder_conv_module_0_sequential_5_running_var, x = input_27)[name = tensor("inputs_7")]; + tensor input_29 = silu(x = inputs_7)[name = tensor("input_29")]; + tensor input_31_pad_type_0 = const()[name = tensor("input_31_pad_type_0"), val = tensor("valid")]; + tensor input_31_strides_0 = const()[name = tensor("input_31_strides_0"), val = tensor([1])]; + tensor input_31_pad_0 = const()[name = tensor("input_31_pad_0"), val = tensor([0, 0])]; + tensor input_31_dilations_0 = const()[name = tensor("input_31_dilations_0"), val = tensor([1])]; + tensor input_31_groups_0 = const()[name = tensor("input_31_groups_0"), val = tensor(1)]; + tensor input_31 = conv(bias = inner_encoder_conv_module_0_sequential_7_conv_bias, dilations = input_31_dilations_0, groups = input_31_groups_0, pad = input_31_pad_0, pad_type = input_31_pad_type_0, strides = input_31_strides_0, weight = inner_encoder_conv_module_0_sequential_7_conv_weight, x = input_29)[name = tensor("input_31")]; tensor conv_out_1_perm_0 = const()[name = tensor("conv_out_1_perm_0"), val = tensor([0, 2, 1])]; - tensor var_298_begin_0 = const()[name = tensor("op_298_begin_0"), val = tensor([0, -1, 0])]; - tensor var_298_end_0 = const()[name = tensor("op_298_end_0"), val = tensor([3, 16, 256])]; - tensor var_298_end_mask_0 = const()[name = tensor("op_298_end_mask_0"), val = tensor([true, true, true])]; - tensor conv_out_1 = transpose(perm = conv_out_1_perm_0, x = input_29)[name = tensor("transpose_53")]; - tensor var_298 = slice_by_index(begin = var_298_begin_0, end = var_298_end_0, end_mask = var_298_end_mask_0, x = conv_out_1)[name = tensor("op_298")]; - tensor var_300_perm_0 = const()[name = tensor("op_300_perm_0"), val = tensor([1, 0, 2])]; - tensor var_300 = transpose(perm = var_300_perm_0, x = var_298)[name = tensor("transpose_52")]; - tensor input_31 = add(x = x_3, y = var_300)[name = tensor("input_31")]; - tensor input_33_axes_0 = const()[name = tensor("input_33_axes_0"), val = tensor([-1])]; - tensor input_33 = layer_norm(axes = input_33_axes_0, beta = encoder_ffn2_0_module_sequential_0_bias, epsilon = var_28, gamma = encoder_ffn2_0_module_sequential_0_weight, x = input_31)[name = tensor("input_33")]; - tensor inputs_9 = linear(bias = encoder_ffn2_0_module_sequential_1_linear_bias, weight = encoder_ffn2_0_module_sequential_1_linear_weight, x = input_33)[name = tensor("linear_8")]; - tensor input_35 = silu(x = inputs_9)[name = tensor("input_35")]; - tensor input_39 = linear(bias = encoder_ffn2_0_module_sequential_4_linear_bias, weight = encoder_ffn2_0_module_sequential_4_linear_weight, x = input_35)[name = tensor("linear_9")]; - tensor var_323 = const()[name = tensor("op_323"), val = tensor(0x1p-1)]; - tensor var_324 = mul(x = input_39, y = var_323)[name = tensor("op_324")]; - tensor input_41 = add(x = var_324, y = input_31)[name = tensor("input_41")]; - tensor input_43_axes_0 = const()[name = tensor("input_43_axes_0"), val = tensor([-1])]; - tensor input_43 = layer_norm(axes = input_43_axes_0, beta = encoder_layer_norm_0_bias, epsilon = var_28, gamma = encoder_layer_norm_0_weight, x = input_41)[name = tensor("input_43")]; + tensor var_344_begin_0 = const()[name = tensor("op_344_begin_0"), val = tensor([0, -1, 0])]; + tensor var_344_end_0 = const()[name = tensor("op_344_end_0"), val = tensor([3, 16, 256])]; + tensor var_344_end_mask_0 = const()[name = tensor("op_344_end_mask_0"), val = tensor([true, true, true])]; + tensor conv_out_1 = transpose(perm = conv_out_1_perm_0, x = input_31)[name = tensor("transpose_53")]; + tensor var_344 = slice_by_index(begin = var_344_begin_0, end = var_344_end_0, end_mask = var_344_end_mask_0, x = conv_out_1)[name = tensor("op_344")]; + tensor var_346_perm_0 = const()[name = tensor("op_346_perm_0"), val = tensor([1, 0, 2])]; + tensor var_346 = transpose(perm = var_346_perm_0, x = var_344)[name = tensor("transpose_52")]; + tensor input_33 = add(x = x_3, y = var_346)[name = tensor("input_33")]; + tensor input_35_axes_0 = const()[name = tensor("input_35_axes_0"), val = tensor([-1])]; + tensor input_35 = layer_norm(axes = input_35_axes_0, beta = inner_encoder_ffn2_0_module_sequential_0_bias, epsilon = var_56, gamma = inner_encoder_ffn2_0_module_sequential_0_weight, x = input_33)[name = tensor("input_35")]; + tensor inputs_9 = linear(bias = inner_encoder_ffn2_0_module_sequential_1_linear_bias, weight = inner_encoder_ffn2_0_module_sequential_1_linear_weight, x = input_35)[name = tensor("linear_8")]; + tensor input_37 = silu(x = inputs_9)[name = tensor("input_37")]; + tensor input_41 = linear(bias = inner_encoder_ffn2_0_module_sequential_4_linear_bias, weight = inner_encoder_ffn2_0_module_sequential_4_linear_weight, x = input_37)[name = tensor("linear_9")]; + tensor var_369 = const()[name = tensor("op_369"), val = tensor(0x1p-1)]; + tensor var_370 = mul(x = input_41, y = var_369)[name = tensor("op_370")]; + tensor input_43 = add(x = var_370, y = input_33)[name = tensor("input_43")]; tensor input_45_axes_0 = const()[name = tensor("input_45_axes_0"), val = tensor([-1])]; - tensor input_45 = layer_norm(axes = input_45_axes_0, beta = encoder_ffn1_1_module_sequential_0_bias, epsilon = var_28, gamma = encoder_ffn1_1_module_sequential_0_weight, x = input_43)[name = tensor("input_45")]; - tensor inputs_11 = linear(bias = encoder_ffn1_1_module_sequential_1_linear_bias, weight = encoder_ffn1_1_module_sequential_1_linear_weight, x = input_45)[name = tensor("linear_10")]; - tensor input_47 = silu(x = inputs_11)[name = tensor("input_47")]; - tensor input_51 = linear(bias = encoder_ffn1_1_module_sequential_4_linear_bias, weight = encoder_ffn1_1_module_sequential_4_linear_weight, x = input_47)[name = tensor("linear_11")]; - tensor var_353 = const()[name = tensor("op_353"), val = tensor(0x1p-1)]; - tensor var_354 = mul(x = input_51, y = var_353)[name = tensor("op_354")]; - tensor input_53 = add(x = var_354, y = input_43)[name = tensor("input_53")]; + tensor input_45 = layer_norm(axes = input_45_axes_0, beta = inner_encoder_layer_norm_0_bias, epsilon = var_56, gamma = inner_encoder_layer_norm_0_weight, x = input_43)[name = tensor("input_45")]; + tensor input_47_axes_0 = const()[name = tensor("input_47_axes_0"), val = tensor([-1])]; + tensor input_47 = layer_norm(axes = input_47_axes_0, beta = inner_encoder_ffn1_1_module_sequential_0_bias, epsilon = var_56, gamma = inner_encoder_ffn1_1_module_sequential_0_weight, x = input_45)[name = tensor("input_47")]; + tensor inputs_11 = linear(bias = inner_encoder_ffn1_1_module_sequential_1_linear_bias, weight = inner_encoder_ffn1_1_module_sequential_1_linear_weight, x = input_47)[name = tensor("linear_10")]; + tensor input_49 = silu(x = inputs_11)[name = tensor("input_49")]; + tensor input_53 = linear(bias = inner_encoder_ffn1_1_module_sequential_4_linear_bias, weight = inner_encoder_ffn1_1_module_sequential_4_linear_weight, x = input_49)[name = tensor("linear_11")]; + tensor var_399 = const()[name = tensor("op_399"), val = tensor(0x1p-1)]; + tensor var_400 = mul(x = input_53, y = var_399)[name = tensor("op_400")]; + tensor input_55 = add(x = var_400, y = input_45)[name = tensor("input_55")]; tensor x_7_axes_0 = const()[name = tensor("x_7_axes_0"), val = tensor([-1])]; - tensor x_7 = layer_norm(axes = x_7_axes_0, beta = encoder_ret_lns_1_bias, epsilon = var_28, gamma = encoder_ret_lns_1_weight, x = input_53)[name = tensor("x_7")]; + tensor x_7 = layer_norm(axes = x_7_axes_0, beta = inner_encoder_ret_lns_1_bias, epsilon = var_56, gamma = inner_encoder_ret_lns_1_weight, x = input_55)[name = tensor("x_7")]; tensor prev_kv_3_begin_0 = const()[name = tensor("prev_kv_3_begin_0"), val = tensor([1, 0, 0, 0, 0])]; tensor prev_kv_3_end_0 = const()[name = tensor("prev_kv_3_end_0"), val = tensor([2, 1, 4, 64, 64])]; tensor prev_kv_3_end_mask_0 = const()[name = tensor("prev_kv_3_end_mask_0"), val = tensor([false, true, true, true, true])]; @@ -406,163 +424,163 @@ program(1.0) tensor prev_scale_3_end_mask_0 = const()[name = tensor("prev_scale_3_end_mask_0"), val = tensor([false, true])]; tensor prev_scale_3_squeeze_mask_0 = const()[name = tensor("prev_scale_3_squeeze_mask_0"), val = tensor([true, false])]; tensor prev_scale_3 = slice_by_index(begin = prev_scale_3_begin_0, end = prev_scale_3_end_0, end_mask = prev_scale_3_end_mask_0, squeeze_mask = prev_scale_3_squeeze_mask_0, x = enc_scale)[name = tensor("prev_scale_3")]; - tensor var_368 = linear(bias = encoder_q_proj_1_bias, weight = encoder_q_proj_1_weight, x = x_7)[name = tensor("linear_12")]; - tensor var_369 = const()[name = tensor("op_369"), val = tensor([1, 3, 4, 64])]; - tensor var_370 = reshape(shape = var_369, x = var_368)[name = tensor("op_370")]; + tensor var_414 = linear(bias = inner_encoder_q_proj_1_bias, weight = inner_encoder_q_proj_1_weight, x = x_7)[name = tensor("linear_12")]; + tensor var_415 = const()[name = tensor("op_415"), val = tensor([1, 3, 4, 64])]; + tensor var_416 = reshape(shape = var_415, x = var_414)[name = tensor("op_416")]; tensor q_3_perm_0 = const()[name = tensor("q_3_perm_0"), val = tensor([0, 2, 1, 3])]; - tensor var_374 = linear(bias = encoder_k_proj_1_bias, weight = encoder_k_proj_1_weight, x = x_7)[name = tensor("linear_13")]; - tensor var_375 = const()[name = tensor("op_375"), val = tensor(0x1p-3)]; - tensor var_376 = mul(x = var_374, y = var_375)[name = tensor("op_376")]; - tensor var_377 = const()[name = tensor("op_377"), val = tensor([1, 3, 4, 64])]; - tensor var_378 = reshape(shape = var_377, x = var_376)[name = tensor("op_378")]; + tensor var_420 = linear(bias = inner_encoder_k_proj_1_bias, weight = inner_encoder_k_proj_1_weight, x = x_7)[name = tensor("linear_13")]; + tensor var_421 = const()[name = tensor("op_421"), val = tensor(0x1p-3)]; + tensor var_422 = mul(x = var_420, y = var_421)[name = tensor("op_422")]; + tensor var_423 = const()[name = tensor("op_423"), val = tensor([1, 3, 4, 64])]; + tensor var_424 = reshape(shape = var_423, x = var_422)[name = tensor("op_424")]; tensor k_3_perm_0 = const()[name = tensor("k_3_perm_0"), val = tensor([0, 2, 1, 3])]; - tensor var_382 = linear(bias = encoder_v_proj_1_bias, weight = encoder_v_proj_1_weight, x = x_7)[name = tensor("linear_14")]; - tensor var_383 = const()[name = tensor("op_383"), val = tensor([1, 3, 4, 64])]; - tensor var_384 = reshape(shape = var_383, x = var_382)[name = tensor("op_384")]; + tensor var_428 = linear(bias = inner_encoder_v_proj_1_bias, weight = inner_encoder_v_proj_1_weight, x = x_7)[name = tensor("linear_14")]; + tensor var_429 = const()[name = tensor("op_429"), val = tensor([1, 3, 4, 64])]; + tensor var_430 = reshape(shape = var_429, x = var_428)[name = tensor("op_430")]; tensor v_3_perm_0 = const()[name = tensor("v_3_perm_0"), val = tensor([0, 2, 1, 3])]; - tensor input_57 = linear(bias = encoder_g_proj_1_bias, weight = encoder_g_proj_1_weight, x = x_7)[name = tensor("linear_15")]; + tensor input_59 = linear(bias = inner_encoder_g_proj_1_bias, weight = inner_encoder_g_proj_1_weight, x = x_7)[name = tensor("linear_15")]; tensor sqrt_s0_3 = sqrt(x = prev_scale_3)[name = tensor("sqrt_s0_3")]; - tensor s_t_3 = add(x = prev_scale_3, y = encoder__t_index)[name = tensor("s_t_3")]; + tensor s_t_3 = add(x = prev_scale_3, y = inner_encoder__t_index)[name = tensor("s_t_3")]; tensor sqrt_s_t_3 = sqrt(x = s_t_3)[name = tensor("sqrt_s_t_3")]; tensor qk_3_transpose_x_1 = const()[name = tensor("qk_3_transpose_x_1"), val = tensor(false)]; tensor qk_3_transpose_y_1 = const()[name = tensor("qk_3_transpose_y_1"), val = tensor(true)]; - tensor k_3 = transpose(perm = k_3_perm_0, x = var_378)[name = tensor("transpose_50")]; - tensor q_3 = transpose(perm = q_3_perm_0, x = var_370)[name = tensor("transpose_51")]; + tensor k_3 = transpose(perm = k_3_perm_0, x = var_424)[name = tensor("transpose_50")]; + tensor q_3 = transpose(perm = q_3_perm_0, x = var_416)[name = tensor("transpose_51")]; tensor qk_3 = matmul(transpose_x = qk_3_transpose_x_1, transpose_y = qk_3_transpose_y_1, x = q_3, y = k_3)[name = tensor("qk_3")]; - tensor var_394 = const()[name = tensor("op_394"), val = tensor([3, 1])]; - tensor var_395 = reshape(shape = var_394, x = sqrt_s_t_3)[name = tensor("op_395")]; - tensor M_3 = real_div(x = encoder__causal_mask, y = var_395)[name = tensor("M_3")]; - tensor var_397 = mul(x = qk_3, y = M_3)[name = tensor("op_397")]; + tensor var_440 = const()[name = tensor("op_440"), val = tensor([3, 1])]; + tensor var_441 = reshape(shape = var_440, x = sqrt_s_t_3)[name = tensor("op_441")]; + tensor M_3 = real_div(x = inner_encoder__causal_mask, y = var_441)[name = tensor("M_3")]; + tensor var_443 = mul(x = qk_3, y = M_3)[name = tensor("op_443")]; tensor inner_3_transpose_x_0 = const()[name = tensor("inner_3_transpose_x_0"), val = tensor(false)]; tensor inner_3_transpose_y_0 = const()[name = tensor("inner_3_transpose_y_0"), val = tensor(false)]; - tensor v_3 = transpose(perm = v_3_perm_0, x = var_384)[name = tensor("transpose_49")]; - tensor inner_3 = matmul(transpose_x = inner_3_transpose_x_0, transpose_y = inner_3_transpose_y_0, x = var_397, y = v_3)[name = tensor("inner_3")]; - tensor var_399_transpose_x_0 = const()[name = tensor("op_399_transpose_x_0"), val = tensor(false)]; - tensor var_399_transpose_y_0 = const()[name = tensor("op_399_transpose_y_0"), val = tensor(false)]; - tensor var_399 = matmul(transpose_x = var_399_transpose_x_0, transpose_y = var_399_transpose_y_0, x = q_3, y = prev_kv_3)[name = tensor("op_399")]; - tensor var_400 = real_div(x = sqrt_s0_3, y = sqrt_s_t_3)[name = tensor("op_400")]; - tensor var_401 = const()[name = tensor("op_401"), val = tensor([1, 1, 3, 1])]; - tensor var_402 = reshape(shape = var_401, x = var_400)[name = tensor("op_402")]; - tensor cross_3 = mul(x = var_399, y = var_402)[name = tensor("cross_3")]; + tensor v_3 = transpose(perm = v_3_perm_0, x = var_430)[name = tensor("transpose_49")]; + tensor inner_3 = matmul(transpose_x = inner_3_transpose_x_0, transpose_y = inner_3_transpose_y_0, x = var_443, y = v_3)[name = tensor("inner_3")]; + tensor var_445_transpose_x_0 = const()[name = tensor("op_445_transpose_x_0"), val = tensor(false)]; + tensor var_445_transpose_y_0 = const()[name = tensor("op_445_transpose_y_0"), val = tensor(false)]; + tensor var_445 = matmul(transpose_x = var_445_transpose_x_0, transpose_y = var_445_transpose_y_0, x = q_3, y = prev_kv_3)[name = tensor("op_445")]; + tensor var_446 = real_div(x = sqrt_s0_3, y = sqrt_s_t_3)[name = tensor("op_446")]; + tensor var_447 = const()[name = tensor("op_447"), val = tensor([1, 1, 3, 1])]; + tensor var_448 = reshape(shape = var_447, x = var_446)[name = tensor("op_448")]; + tensor cross_3 = mul(x = var_445, y = var_448)[name = tensor("cross_3")]; tensor out_7 = add(x = inner_3, y = cross_3)[name = tensor("out_7")]; - tensor var_405 = mul(x = prev_kv_3, y = sqrt_s0_3)[name = tensor("op_405")]; - tensor var_407_transpose_x_1 = const()[name = tensor("op_407_transpose_x_1"), val = tensor(true)]; - tensor var_407_transpose_y_1 = const()[name = tensor("op_407_transpose_y_1"), val = tensor(false)]; - tensor var_407 = matmul(transpose_x = var_407_transpose_x_1, transpose_y = var_407_transpose_y_1, x = k_3, y = v_3)[name = tensor("op_407")]; - tensor new_kv_unnorm_3 = add(x = var_405, y = var_407)[name = tensor("new_kv_unnorm_3")]; - tensor var_409 = const()[name = tensor("op_409"), val = tensor(0x1.8p+1)]; - tensor new_scale_3 = add(x = prev_scale_3, y = var_409)[name = tensor("new_scale_3")]; - tensor var_411 = sqrt(x = new_scale_3)[name = tensor("op_411")]; - tensor var_412 = real_div(x = new_kv_unnorm_3, y = var_411)[name = tensor("op_412")]; - tensor var_413_perm_0 = const()[name = tensor("op_413_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_451 = mul(x = prev_kv_3, y = sqrt_s0_3)[name = tensor("op_451")]; + tensor var_453_transpose_x_1 = const()[name = tensor("op_453_transpose_x_1"), val = tensor(true)]; + tensor var_453_transpose_y_1 = const()[name = tensor("op_453_transpose_y_1"), val = tensor(false)]; + tensor var_453 = matmul(transpose_x = var_453_transpose_x_1, transpose_y = var_453_transpose_y_1, x = k_3, y = v_3)[name = tensor("op_453")]; + tensor new_kv_unnorm_3 = add(x = var_451, y = var_453)[name = tensor("new_kv_unnorm_3")]; + tensor var_455 = const()[name = tensor("op_455"), val = tensor(0x1.8p+1)]; + tensor new_scale_3 = add(x = prev_scale_3, y = var_455)[name = tensor("new_scale_3")]; + tensor var_457 = sqrt(x = new_scale_3)[name = tensor("op_457")]; + tensor var_458 = real_div(x = new_kv_unnorm_3, y = var_457)[name = tensor("op_458")]; + tensor var_459_perm_0 = const()[name = tensor("op_459_perm_0"), val = tensor([0, 2, 1, 3])]; tensor out_9_axes_0 = const()[name = tensor("out_9_axes_0"), val = tensor([-1])]; - tensor var_413 = transpose(perm = var_413_perm_0, x = out_7)[name = tensor("transpose_48")]; - tensor out_9 = layer_norm(axes = out_9_axes_0, epsilon = var_17, x = var_413)[name = tensor("out_9")]; - tensor var_417 = const()[name = tensor("op_417"), val = tensor([1, 3, 256])]; - tensor out_11 = reshape(shape = var_417, x = out_9)[name = tensor("out_11")]; - tensor var_419 = silu(x = input_57)[name = tensor("op_419")]; - tensor input_59 = mul(x = var_419, y = out_11)[name = tensor("input_59")]; - tensor ret_out_3 = linear(bias = encoder_out_proj_1_bias, weight = encoder_out_proj_1_weight, x = input_59)[name = tensor("linear_16")]; - tensor x_9 = add(x = input_53, y = ret_out_3)[name = tensor("x_9")]; + tensor var_459 = transpose(perm = var_459_perm_0, x = out_7)[name = tensor("transpose_48")]; + tensor out_9 = layer_norm(axes = out_9_axes_0, epsilon = var_64, x = var_459)[name = tensor("out_9")]; + tensor var_463 = const()[name = tensor("op_463"), val = tensor([1, 3, 256])]; + tensor out_11 = reshape(shape = var_463, x = out_9)[name = tensor("out_11")]; + tensor var_465 = silu(x = input_59)[name = tensor("op_465")]; + tensor input_61 = mul(x = var_465, y = out_11)[name = tensor("input_61")]; + tensor ret_out_3 = linear(bias = inner_encoder_out_proj_1_bias, weight = inner_encoder_out_proj_1_weight, x = input_61)[name = tensor("linear_16")]; + tensor x_9 = add(x = input_55, y = ret_out_3)[name = tensor("x_9")]; tensor window_9_begin_0 = const()[name = tensor("window_9_begin_0"), val = tensor([1, 0, 0, 0])]; tensor window_9_end_0 = const()[name = tensor("window_9_end_0"), val = tensor([2, 1, 16, 256])]; tensor window_9_end_mask_0 = const()[name = tensor("window_9_end_mask_0"), val = tensor([false, true, true, true])]; tensor window_9_squeeze_mask_0 = const()[name = tensor("window_9_squeeze_mask_0"), val = tensor([true, false, false, false])]; tensor window_9 = slice_by_index(begin = window_9_begin_0, end = window_9_end_0, end_mask = window_9_end_mask_0, squeeze_mask = window_9_squeeze_mask_0, x = enc_conv_cache)[name = tensor("window_9")]; - tensor var_427_begin_0 = const()[name = tensor("op_427_begin_0"), val = tensor([0, 0, 0])]; - tensor var_427_end_0 = const()[name = tensor("op_427_end_0"), val = tensor([1, 1, 256])]; - tensor var_427_end_mask_0 = const()[name = tensor("op_427_end_mask_0"), val = tensor([true, false, true])]; - tensor var_427 = slice_by_index(begin = var_427_begin_0, end = var_427_end_0, end_mask = var_427_end_mask_0, x = x_9)[name = tensor("op_427")]; - tensor var_430_begin_0 = const()[name = tensor("op_430_begin_0"), val = tensor([0, 1, 0])]; - tensor var_430_end_0 = const()[name = tensor("op_430_end_0"), val = tensor([1, 16, 256])]; - tensor var_430_end_mask_0 = const()[name = tensor("op_430_end_mask_0"), val = tensor([true, true, true])]; - tensor var_430 = slice_by_index(begin = var_430_begin_0, end = var_430_end_0, end_mask = var_430_end_mask_0, x = window_9)[name = tensor("op_430")]; + tensor var_473_begin_0 = const()[name = tensor("op_473_begin_0"), val = tensor([0, 0, 0])]; + tensor var_473_end_0 = const()[name = tensor("op_473_end_0"), val = tensor([1, 1, 256])]; + tensor var_473_end_mask_0 = const()[name = tensor("op_473_end_mask_0"), val = tensor([true, false, true])]; + tensor var_473 = slice_by_index(begin = var_473_begin_0, end = var_473_end_0, end_mask = var_473_end_mask_0, x = x_9)[name = tensor("op_473")]; + tensor var_476_begin_0 = const()[name = tensor("op_476_begin_0"), val = tensor([0, 1, 0])]; + tensor var_476_end_0 = const()[name = tensor("op_476_end_0"), val = tensor([1, 16, 256])]; + tensor var_476_end_mask_0 = const()[name = tensor("op_476_end_mask_0"), val = tensor([true, true, true])]; + tensor var_476 = slice_by_index(begin = var_476_begin_0, end = var_476_end_0, end_mask = var_476_end_mask_0, x = window_9)[name = tensor("op_476")]; tensor window_11_interleave_0 = const()[name = tensor("window_11_interleave_0"), val = tensor(false)]; - tensor window_11 = concat(axis = var_26, interleave = window_11_interleave_0, values = (var_430, var_427))[name = tensor("window_11")]; - tensor var_435_begin_0 = const()[name = tensor("op_435_begin_0"), val = tensor([0, 1, 0])]; - tensor var_435_end_0 = const()[name = tensor("op_435_end_0"), val = tensor([1, 2, 256])]; - tensor var_435_end_mask_0 = const()[name = tensor("op_435_end_mask_0"), val = tensor([true, false, true])]; - tensor var_435 = slice_by_index(begin = var_435_begin_0, end = var_435_end_0, end_mask = var_435_end_mask_0, x = x_9)[name = tensor("op_435")]; - tensor var_438_begin_0 = const()[name = tensor("op_438_begin_0"), val = tensor([0, 1, 0])]; - tensor var_438_end_0 = const()[name = tensor("op_438_end_0"), val = tensor([1, 16, 256])]; - tensor var_438_end_mask_0 = const()[name = tensor("op_438_end_mask_0"), val = tensor([true, true, true])]; - tensor var_438 = slice_by_index(begin = var_438_begin_0, end = var_438_end_0, end_mask = var_438_end_mask_0, x = window_11)[name = tensor("op_438")]; + tensor window_11 = concat(axis = var_72, interleave = window_11_interleave_0, values = (var_476, var_473))[name = tensor("window_11")]; + tensor var_481_begin_0 = const()[name = tensor("op_481_begin_0"), val = tensor([0, 1, 0])]; + tensor var_481_end_0 = const()[name = tensor("op_481_end_0"), val = tensor([1, 2, 256])]; + tensor var_481_end_mask_0 = const()[name = tensor("op_481_end_mask_0"), val = tensor([true, false, true])]; + tensor var_481 = slice_by_index(begin = var_481_begin_0, end = var_481_end_0, end_mask = var_481_end_mask_0, x = x_9)[name = tensor("op_481")]; + tensor var_484_begin_0 = const()[name = tensor("op_484_begin_0"), val = tensor([0, 1, 0])]; + tensor var_484_end_0 = const()[name = tensor("op_484_end_0"), val = tensor([1, 16, 256])]; + tensor var_484_end_mask_0 = const()[name = tensor("op_484_end_mask_0"), val = tensor([true, true, true])]; + tensor var_484 = slice_by_index(begin = var_484_begin_0, end = var_484_end_0, end_mask = var_484_end_mask_0, x = window_11)[name = tensor("op_484")]; tensor window_13_interleave_0 = const()[name = tensor("window_13_interleave_0"), val = tensor(false)]; - tensor window_13 = concat(axis = var_26, interleave = window_13_interleave_0, values = (var_438, var_435))[name = tensor("window_13")]; - tensor var_443_begin_0 = const()[name = tensor("op_443_begin_0"), val = tensor([0, 2, 0])]; - tensor var_443_end_0 = const()[name = tensor("op_443_end_0"), val = tensor([1, 1, 256])]; - tensor var_443_end_mask_0 = const()[name = tensor("op_443_end_mask_0"), val = tensor([true, true, true])]; - tensor var_443 = slice_by_index(begin = var_443_begin_0, end = var_443_end_0, end_mask = var_443_end_mask_0, x = x_9)[name = tensor("op_443")]; - tensor var_446_begin_0 = const()[name = tensor("op_446_begin_0"), val = tensor([0, 1, 0])]; - tensor var_446_end_0 = const()[name = tensor("op_446_end_0"), val = tensor([1, 16, 256])]; - tensor var_446_end_mask_0 = const()[name = tensor("op_446_end_mask_0"), val = tensor([true, true, true])]; - tensor var_446 = slice_by_index(begin = var_446_begin_0, end = var_446_end_0, end_mask = var_446_end_mask_0, x = window_13)[name = tensor("op_446")]; + tensor window_13 = concat(axis = var_72, interleave = window_13_interleave_0, values = (var_484, var_481))[name = tensor("window_13")]; + tensor var_489_begin_0 = const()[name = tensor("op_489_begin_0"), val = tensor([0, 2, 0])]; + tensor var_489_end_0 = const()[name = tensor("op_489_end_0"), val = tensor([1, 1, 256])]; + tensor var_489_end_mask_0 = const()[name = tensor("op_489_end_mask_0"), val = tensor([true, true, true])]; + tensor var_489 = slice_by_index(begin = var_489_begin_0, end = var_489_end_0, end_mask = var_489_end_mask_0, x = x_9)[name = tensor("op_489")]; + tensor var_492_begin_0 = const()[name = tensor("op_492_begin_0"), val = tensor([0, 1, 0])]; + tensor var_492_end_0 = const()[name = tensor("op_492_end_0"), val = tensor([1, 16, 256])]; + tensor var_492_end_mask_0 = const()[name = tensor("op_492_end_mask_0"), val = tensor([true, true, true])]; + tensor var_492 = slice_by_index(begin = var_492_begin_0, end = var_492_end_0, end_mask = var_492_end_mask_0, x = window_13)[name = tensor("op_492")]; tensor window_15_interleave_0 = const()[name = tensor("window_15_interleave_0"), val = tensor(false)]; - tensor window_15 = concat(axis = var_26, interleave = window_15_interleave_0, values = (var_446, var_443))[name = tensor("window_15")]; - tensor input_61_interleave_0 = const()[name = tensor("input_61_interleave_0"), val = tensor(false)]; - tensor input_61 = concat(axis = var_23, interleave = input_61_interleave_0, values = (window_11, window_13, window_15))[name = tensor("input_61")]; + tensor window_15 = concat(axis = var_72, interleave = window_15_interleave_0, values = (var_492, var_489))[name = tensor("window_15")]; + tensor input_63_interleave_0 = const()[name = tensor("input_63_interleave_0"), val = tensor(false)]; + tensor input_63 = concat(axis = var_59, interleave = input_63_interleave_0, values = (window_11, window_13, window_15))[name = tensor("input_63")]; tensor x_11_axes_0 = const()[name = tensor("x_11_axes_0"), val = tensor([-1])]; - tensor x_11 = layer_norm(axes = x_11_axes_0, beta = encoder_conv_module_1_sequential_0_bias, epsilon = var_28, gamma = encoder_conv_module_1_sequential_0_weight, x = input_61)[name = tensor("x_11")]; - tensor input_63_perm_0 = const()[name = tensor("input_63_perm_0"), val = tensor([0, 2, 1])]; + tensor x_11 = layer_norm(axes = x_11_axes_0, beta = inner_encoder_conv_module_1_sequential_0_bias, epsilon = var_56, gamma = inner_encoder_conv_module_1_sequential_0_weight, x = input_63)[name = tensor("x_11")]; + tensor input_65_perm_0 = const()[name = tensor("input_65_perm_0"), val = tensor([0, 2, 1])]; tensor inputs_13_pad_type_0 = const()[name = tensor("inputs_13_pad_type_0"), val = tensor("valid")]; tensor inputs_13_strides_0 = const()[name = tensor("inputs_13_strides_0"), val = tensor([1])]; tensor inputs_13_pad_0 = const()[name = tensor("inputs_13_pad_0"), val = tensor([0, 0])]; tensor inputs_13_dilations_0 = const()[name = tensor("inputs_13_dilations_0"), val = tensor([1])]; tensor inputs_13_groups_0 = const()[name = tensor("inputs_13_groups_0"), val = tensor(1)]; - tensor input_63 = transpose(perm = input_63_perm_0, x = x_11)[name = tensor("transpose_47")]; - tensor inputs_13 = conv(bias = encoder_conv_module_1_sequential_2_conv_bias, dilations = inputs_13_dilations_0, groups = inputs_13_groups_0, pad = inputs_13_pad_0, pad_type = inputs_13_pad_type_0, strides = inputs_13_strides_0, weight = encoder_conv_module_1_sequential_2_conv_weight, x = input_63)[name = tensor("inputs_13")]; - tensor var_471_split_sizes_0 = const()[name = tensor("op_471_split_sizes_0"), val = tensor([256, 256])]; - tensor var_471_axis_0 = const()[name = tensor("op_471_axis_0"), val = tensor(1)]; - tensor var_471_0, tensor var_471_1 = split(axis = var_471_axis_0, split_sizes = var_471_split_sizes_0, x = inputs_13)[name = tensor("op_471")]; - tensor var_473 = sigmoid(x = var_471_1)[name = tensor("op_473")]; - tensor inputs_15 = mul(x = var_471_0, y = var_473)[name = tensor("inputs_15")]; + tensor input_65 = transpose(perm = input_65_perm_0, x = x_11)[name = tensor("transpose_47")]; + tensor inputs_13 = conv(bias = inner_encoder_conv_module_1_sequential_2_conv_bias, dilations = inputs_13_dilations_0, groups = inputs_13_groups_0, pad = inputs_13_pad_0, pad_type = inputs_13_pad_type_0, strides = inputs_13_strides_0, weight = inner_encoder_conv_module_1_sequential_2_conv_weight, x = input_65)[name = tensor("inputs_13")]; + tensor var_517_split_sizes_0 = const()[name = tensor("op_517_split_sizes_0"), val = tensor([256, 256])]; + tensor var_517_axis_0 = const()[name = tensor("op_517_axis_0"), val = tensor(1)]; + tensor var_517_0, tensor var_517_1 = split(axis = var_517_axis_0, split_sizes = var_517_split_sizes_0, x = inputs_13)[name = tensor("op_517")]; + tensor var_519 = sigmoid(x = var_517_1)[name = tensor("op_519")]; + tensor inputs_15 = mul(x = var_517_0, y = var_519)[name = tensor("inputs_15")]; tensor outputs_aug_3_pad_type_0 = const()[name = tensor("outputs_aug_3_pad_type_0"), val = tensor("custom")]; tensor outputs_aug_3_pad_0 = const()[name = tensor("outputs_aug_3_pad_0"), val = tensor([15, 15])]; tensor outputs_aug_3_groups_0 = const()[name = tensor("outputs_aug_3_groups_0"), val = tensor(256)]; tensor outputs_aug_3_strides_0 = const()[name = tensor("outputs_aug_3_strides_0"), val = tensor([1])]; tensor outputs_aug_3_dilations_0 = const()[name = tensor("outputs_aug_3_dilations_0"), val = tensor([1])]; - tensor outputs_aug_3 = conv(dilations = outputs_aug_3_dilations_0, groups = outputs_aug_3_groups_0, pad = outputs_aug_3_pad_0, pad_type = outputs_aug_3_pad_type_0, strides = outputs_aug_3_strides_0, weight = encoder_conv_module_1_sequential_4_conv_weight, x = inputs_15)[name = tensor("outputs_aug_3")]; - tensor input_65_begin_0 = const()[name = tensor("input_65_begin_0"), val = tensor([0, 0, 0])]; - tensor input_65_end_0 = const()[name = tensor("input_65_end_0"), val = tensor([3, 256, 16])]; - tensor input_65_end_mask_0 = const()[name = tensor("input_65_end_mask_0"), val = tensor([true, true, false])]; - tensor input_65 = slice_by_index(begin = input_65_begin_0, end = input_65_end_0, end_mask = input_65_end_mask_0, x = outputs_aug_3)[name = tensor("input_65")]; - tensor inputs_17 = batch_norm(beta = encoder_conv_module_1_sequential_5_bias, epsilon = var_28, gamma = encoder_conv_module_1_sequential_5_weight, mean = encoder_conv_module_1_sequential_5_running_mean, variance = encoder_conv_module_1_sequential_5_running_var, x = input_65)[name = tensor("inputs_17")]; - tensor input_67 = silu(x = inputs_17)[name = tensor("input_67")]; - tensor input_69_pad_type_0 = const()[name = tensor("input_69_pad_type_0"), val = tensor("valid")]; - tensor input_69_strides_0 = const()[name = tensor("input_69_strides_0"), val = tensor([1])]; - tensor input_69_pad_0 = const()[name = tensor("input_69_pad_0"), val = tensor([0, 0])]; - tensor input_69_dilations_0 = const()[name = tensor("input_69_dilations_0"), val = tensor([1])]; - tensor input_69_groups_0 = const()[name = tensor("input_69_groups_0"), val = tensor(1)]; - tensor input_69 = conv(bias = encoder_conv_module_1_sequential_7_conv_bias, dilations = input_69_dilations_0, groups = input_69_groups_0, pad = input_69_pad_0, pad_type = input_69_pad_type_0, strides = input_69_strides_0, weight = encoder_conv_module_1_sequential_7_conv_weight, x = input_67)[name = tensor("input_69")]; + tensor outputs_aug_3 = conv(dilations = outputs_aug_3_dilations_0, groups = outputs_aug_3_groups_0, pad = outputs_aug_3_pad_0, pad_type = outputs_aug_3_pad_type_0, strides = outputs_aug_3_strides_0, weight = inner_encoder_conv_module_1_sequential_4_conv_weight, x = inputs_15)[name = tensor("outputs_aug_3")]; + tensor input_67_begin_0 = const()[name = tensor("input_67_begin_0"), val = tensor([0, 0, 0])]; + tensor input_67_end_0 = const()[name = tensor("input_67_end_0"), val = tensor([3, 256, 16])]; + tensor input_67_end_mask_0 = const()[name = tensor("input_67_end_mask_0"), val = tensor([true, true, false])]; + tensor input_67 = slice_by_index(begin = input_67_begin_0, end = input_67_end_0, end_mask = input_67_end_mask_0, x = outputs_aug_3)[name = tensor("input_67")]; + tensor inputs_17 = batch_norm(beta = inner_encoder_conv_module_1_sequential_5_bias, epsilon = var_56, gamma = inner_encoder_conv_module_1_sequential_5_weight, mean = inner_encoder_conv_module_1_sequential_5_running_mean, variance = inner_encoder_conv_module_1_sequential_5_running_var, x = input_67)[name = tensor("inputs_17")]; + tensor input_69 = silu(x = inputs_17)[name = tensor("input_69")]; + tensor input_71_pad_type_0 = const()[name = tensor("input_71_pad_type_0"), val = tensor("valid")]; + tensor input_71_strides_0 = const()[name = tensor("input_71_strides_0"), val = tensor([1])]; + tensor input_71_pad_0 = const()[name = tensor("input_71_pad_0"), val = tensor([0, 0])]; + tensor input_71_dilations_0 = const()[name = tensor("input_71_dilations_0"), val = tensor([1])]; + tensor input_71_groups_0 = const()[name = tensor("input_71_groups_0"), val = tensor(1)]; + tensor input_71 = conv(bias = inner_encoder_conv_module_1_sequential_7_conv_bias, dilations = input_71_dilations_0, groups = input_71_groups_0, pad = input_71_pad_0, pad_type = input_71_pad_type_0, strides = input_71_strides_0, weight = inner_encoder_conv_module_1_sequential_7_conv_weight, x = input_69)[name = tensor("input_71")]; tensor conv_out_3_perm_0 = const()[name = tensor("conv_out_3_perm_0"), val = tensor([0, 2, 1])]; - tensor var_504_begin_0 = const()[name = tensor("op_504_begin_0"), val = tensor([0, -1, 0])]; - tensor var_504_end_0 = const()[name = tensor("op_504_end_0"), val = tensor([3, 16, 256])]; - tensor var_504_end_mask_0 = const()[name = tensor("op_504_end_mask_0"), val = tensor([true, true, true])]; - tensor conv_out_3 = transpose(perm = conv_out_3_perm_0, x = input_69)[name = tensor("transpose_46")]; - tensor var_504 = slice_by_index(begin = var_504_begin_0, end = var_504_end_0, end_mask = var_504_end_mask_0, x = conv_out_3)[name = tensor("op_504")]; - tensor var_506_perm_0 = const()[name = tensor("op_506_perm_0"), val = tensor([1, 0, 2])]; - tensor var_506 = transpose(perm = var_506_perm_0, x = var_504)[name = tensor("transpose_45")]; - tensor input_71 = add(x = x_9, y = var_506)[name = tensor("input_71")]; - tensor input_73_axes_0 = const()[name = tensor("input_73_axes_0"), val = tensor([-1])]; - tensor input_73 = layer_norm(axes = input_73_axes_0, beta = encoder_ffn2_1_module_sequential_0_bias, epsilon = var_28, gamma = encoder_ffn2_1_module_sequential_0_weight, x = input_71)[name = tensor("input_73")]; - tensor inputs_19 = linear(bias = encoder_ffn2_1_module_sequential_1_linear_bias, weight = encoder_ffn2_1_module_sequential_1_linear_weight, x = input_73)[name = tensor("linear_17")]; - tensor input_75 = silu(x = inputs_19)[name = tensor("input_75")]; - tensor input_79 = linear(bias = encoder_ffn2_1_module_sequential_4_linear_bias, weight = encoder_ffn2_1_module_sequential_4_linear_weight, x = input_75)[name = tensor("linear_18")]; - tensor var_529 = const()[name = tensor("op_529"), val = tensor(0x1p-1)]; - tensor var_530 = mul(x = input_79, y = var_529)[name = tensor("op_530")]; - tensor input_81 = add(x = var_530, y = input_71)[name = tensor("input_81")]; - tensor input_83_axes_0 = const()[name = tensor("input_83_axes_0"), val = tensor([-1])]; - tensor input_83 = layer_norm(axes = input_83_axes_0, beta = encoder_layer_norm_1_bias, epsilon = var_28, gamma = encoder_layer_norm_1_weight, x = input_81)[name = tensor("input_83")]; + tensor var_550_begin_0 = const()[name = tensor("op_550_begin_0"), val = tensor([0, -1, 0])]; + tensor var_550_end_0 = const()[name = tensor("op_550_end_0"), val = tensor([3, 16, 256])]; + tensor var_550_end_mask_0 = const()[name = tensor("op_550_end_mask_0"), val = tensor([true, true, true])]; + tensor conv_out_3 = transpose(perm = conv_out_3_perm_0, x = input_71)[name = tensor("transpose_46")]; + tensor var_550 = slice_by_index(begin = var_550_begin_0, end = var_550_end_0, end_mask = var_550_end_mask_0, x = conv_out_3)[name = tensor("op_550")]; + tensor var_552_perm_0 = const()[name = tensor("op_552_perm_0"), val = tensor([1, 0, 2])]; + tensor var_552 = transpose(perm = var_552_perm_0, x = var_550)[name = tensor("transpose_45")]; + tensor input_73 = add(x = x_9, y = var_552)[name = tensor("input_73")]; + tensor input_75_axes_0 = const()[name = tensor("input_75_axes_0"), val = tensor([-1])]; + tensor input_75 = layer_norm(axes = input_75_axes_0, beta = inner_encoder_ffn2_1_module_sequential_0_bias, epsilon = var_56, gamma = inner_encoder_ffn2_1_module_sequential_0_weight, x = input_73)[name = tensor("input_75")]; + tensor inputs_19 = linear(bias = inner_encoder_ffn2_1_module_sequential_1_linear_bias, weight = inner_encoder_ffn2_1_module_sequential_1_linear_weight, x = input_75)[name = tensor("linear_17")]; + tensor input_77 = silu(x = inputs_19)[name = tensor("input_77")]; + tensor input_81 = linear(bias = inner_encoder_ffn2_1_module_sequential_4_linear_bias, weight = inner_encoder_ffn2_1_module_sequential_4_linear_weight, x = input_77)[name = tensor("linear_18")]; + tensor var_575 = const()[name = tensor("op_575"), val = tensor(0x1p-1)]; + tensor var_576 = mul(x = input_81, y = var_575)[name = tensor("op_576")]; + tensor input_83 = add(x = var_576, y = input_73)[name = tensor("input_83")]; tensor input_85_axes_0 = const()[name = tensor("input_85_axes_0"), val = tensor([-1])]; - tensor input_85 = layer_norm(axes = input_85_axes_0, beta = encoder_ffn1_2_module_sequential_0_bias, epsilon = var_28, gamma = encoder_ffn1_2_module_sequential_0_weight, x = input_83)[name = tensor("input_85")]; - tensor inputs_21 = linear(bias = encoder_ffn1_2_module_sequential_1_linear_bias, weight = encoder_ffn1_2_module_sequential_1_linear_weight, x = input_85)[name = tensor("linear_19")]; - tensor input_87 = silu(x = inputs_21)[name = tensor("input_87")]; - tensor input_91 = linear(bias = encoder_ffn1_2_module_sequential_4_linear_bias, weight = encoder_ffn1_2_module_sequential_4_linear_weight, x = input_87)[name = tensor("linear_20")]; - tensor var_559 = const()[name = tensor("op_559"), val = tensor(0x1p-1)]; - tensor var_560 = mul(x = input_91, y = var_559)[name = tensor("op_560")]; - tensor input_93 = add(x = var_560, y = input_83)[name = tensor("input_93")]; + tensor input_85 = layer_norm(axes = input_85_axes_0, beta = inner_encoder_layer_norm_1_bias, epsilon = var_56, gamma = inner_encoder_layer_norm_1_weight, x = input_83)[name = tensor("input_85")]; + tensor input_87_axes_0 = const()[name = tensor("input_87_axes_0"), val = tensor([-1])]; + tensor input_87 = layer_norm(axes = input_87_axes_0, beta = inner_encoder_ffn1_2_module_sequential_0_bias, epsilon = var_56, gamma = inner_encoder_ffn1_2_module_sequential_0_weight, x = input_85)[name = tensor("input_87")]; + tensor inputs_21 = linear(bias = inner_encoder_ffn1_2_module_sequential_1_linear_bias, weight = inner_encoder_ffn1_2_module_sequential_1_linear_weight, x = input_87)[name = tensor("linear_19")]; + tensor input_89 = silu(x = inputs_21)[name = tensor("input_89")]; + tensor input_93 = linear(bias = inner_encoder_ffn1_2_module_sequential_4_linear_bias, weight = inner_encoder_ffn1_2_module_sequential_4_linear_weight, x = input_89)[name = tensor("linear_20")]; + tensor var_605 = const()[name = tensor("op_605"), val = tensor(0x1p-1)]; + tensor var_606 = mul(x = input_93, y = var_605)[name = tensor("op_606")]; + tensor input_95 = add(x = var_606, y = input_85)[name = tensor("input_95")]; tensor x_13_axes_0 = const()[name = tensor("x_13_axes_0"), val = tensor([-1])]; - tensor x_13 = layer_norm(axes = x_13_axes_0, beta = encoder_ret_lns_2_bias, epsilon = var_28, gamma = encoder_ret_lns_2_weight, x = input_93)[name = tensor("x_13")]; + tensor x_13 = layer_norm(axes = x_13_axes_0, beta = inner_encoder_ret_lns_2_bias, epsilon = var_56, gamma = inner_encoder_ret_lns_2_weight, x = input_95)[name = tensor("x_13")]; tensor prev_kv_5_begin_0 = const()[name = tensor("prev_kv_5_begin_0"), val = tensor([2, 0, 0, 0, 0])]; tensor prev_kv_5_end_0 = const()[name = tensor("prev_kv_5_end_0"), val = tensor([3, 1, 4, 64, 64])]; tensor prev_kv_5_end_mask_0 = const()[name = tensor("prev_kv_5_end_mask_0"), val = tensor([false, true, true, true, true])]; @@ -573,163 +591,163 @@ program(1.0) tensor prev_scale_5_end_mask_0 = const()[name = tensor("prev_scale_5_end_mask_0"), val = tensor([false, true])]; tensor prev_scale_5_squeeze_mask_0 = const()[name = tensor("prev_scale_5_squeeze_mask_0"), val = tensor([true, false])]; tensor prev_scale_5 = slice_by_index(begin = prev_scale_5_begin_0, end = prev_scale_5_end_0, end_mask = prev_scale_5_end_mask_0, squeeze_mask = prev_scale_5_squeeze_mask_0, x = enc_scale)[name = tensor("prev_scale_5")]; - tensor var_574 = linear(bias = encoder_q_proj_2_bias, weight = encoder_q_proj_2_weight, x = x_13)[name = tensor("linear_21")]; - tensor var_575 = const()[name = tensor("op_575"), val = tensor([1, 3, 4, 64])]; - tensor var_576 = reshape(shape = var_575, x = var_574)[name = tensor("op_576")]; + tensor var_620 = linear(bias = inner_encoder_q_proj_2_bias, weight = inner_encoder_q_proj_2_weight, x = x_13)[name = tensor("linear_21")]; + tensor var_621 = const()[name = tensor("op_621"), val = tensor([1, 3, 4, 64])]; + tensor var_622 = reshape(shape = var_621, x = var_620)[name = tensor("op_622")]; tensor q_5_perm_0 = const()[name = tensor("q_5_perm_0"), val = tensor([0, 2, 1, 3])]; - tensor var_580 = linear(bias = encoder_k_proj_2_bias, weight = encoder_k_proj_2_weight, x = x_13)[name = tensor("linear_22")]; - tensor var_581 = const()[name = tensor("op_581"), val = tensor(0x1p-3)]; - tensor var_582 = mul(x = var_580, y = var_581)[name = tensor("op_582")]; - tensor var_583 = const()[name = tensor("op_583"), val = tensor([1, 3, 4, 64])]; - tensor var_584 = reshape(shape = var_583, x = var_582)[name = tensor("op_584")]; + tensor var_626 = linear(bias = inner_encoder_k_proj_2_bias, weight = inner_encoder_k_proj_2_weight, x = x_13)[name = tensor("linear_22")]; + tensor var_627 = const()[name = tensor("op_627"), val = tensor(0x1p-3)]; + tensor var_628 = mul(x = var_626, y = var_627)[name = tensor("op_628")]; + tensor var_629 = const()[name = tensor("op_629"), val = tensor([1, 3, 4, 64])]; + tensor var_630 = reshape(shape = var_629, x = var_628)[name = tensor("op_630")]; tensor k_5_perm_0 = const()[name = tensor("k_5_perm_0"), val = tensor([0, 2, 1, 3])]; - tensor var_588 = linear(bias = encoder_v_proj_2_bias, weight = encoder_v_proj_2_weight, x = x_13)[name = tensor("linear_23")]; - tensor var_589 = const()[name = tensor("op_589"), val = tensor([1, 3, 4, 64])]; - tensor var_590 = reshape(shape = var_589, x = var_588)[name = tensor("op_590")]; + tensor var_634 = linear(bias = inner_encoder_v_proj_2_bias, weight = inner_encoder_v_proj_2_weight, x = x_13)[name = tensor("linear_23")]; + tensor var_635 = const()[name = tensor("op_635"), val = tensor([1, 3, 4, 64])]; + tensor var_636 = reshape(shape = var_635, x = var_634)[name = tensor("op_636")]; tensor v_5_perm_0 = const()[name = tensor("v_5_perm_0"), val = tensor([0, 2, 1, 3])]; - tensor input_97 = linear(bias = encoder_g_proj_2_bias, weight = encoder_g_proj_2_weight, x = x_13)[name = tensor("linear_24")]; + tensor input_99 = linear(bias = inner_encoder_g_proj_2_bias, weight = inner_encoder_g_proj_2_weight, x = x_13)[name = tensor("linear_24")]; tensor sqrt_s0_5 = sqrt(x = prev_scale_5)[name = tensor("sqrt_s0_5")]; - tensor s_t_5 = add(x = prev_scale_5, y = encoder__t_index)[name = tensor("s_t_5")]; + tensor s_t_5 = add(x = prev_scale_5, y = inner_encoder__t_index)[name = tensor("s_t_5")]; tensor sqrt_s_t_5 = sqrt(x = s_t_5)[name = tensor("sqrt_s_t_5")]; tensor qk_5_transpose_x_1 = const()[name = tensor("qk_5_transpose_x_1"), val = tensor(false)]; tensor qk_5_transpose_y_1 = const()[name = tensor("qk_5_transpose_y_1"), val = tensor(true)]; - tensor k_5 = transpose(perm = k_5_perm_0, x = var_584)[name = tensor("transpose_43")]; - tensor q_5 = transpose(perm = q_5_perm_0, x = var_576)[name = tensor("transpose_44")]; + tensor k_5 = transpose(perm = k_5_perm_0, x = var_630)[name = tensor("transpose_43")]; + tensor q_5 = transpose(perm = q_5_perm_0, x = var_622)[name = tensor("transpose_44")]; tensor qk_5 = matmul(transpose_x = qk_5_transpose_x_1, transpose_y = qk_5_transpose_y_1, x = q_5, y = k_5)[name = tensor("qk_5")]; - tensor var_600 = const()[name = tensor("op_600"), val = tensor([3, 1])]; - tensor var_601 = reshape(shape = var_600, x = sqrt_s_t_5)[name = tensor("op_601")]; - tensor M_5 = real_div(x = encoder__causal_mask, y = var_601)[name = tensor("M_5")]; - tensor var_603 = mul(x = qk_5, y = M_5)[name = tensor("op_603")]; + tensor var_646 = const()[name = tensor("op_646"), val = tensor([3, 1])]; + tensor var_647 = reshape(shape = var_646, x = sqrt_s_t_5)[name = tensor("op_647")]; + tensor M_5 = real_div(x = inner_encoder__causal_mask, y = var_647)[name = tensor("M_5")]; + tensor var_649 = mul(x = qk_5, y = M_5)[name = tensor("op_649")]; tensor inner_5_transpose_x_0 = const()[name = tensor("inner_5_transpose_x_0"), val = tensor(false)]; tensor inner_5_transpose_y_0 = const()[name = tensor("inner_5_transpose_y_0"), val = tensor(false)]; - tensor v_5 = transpose(perm = v_5_perm_0, x = var_590)[name = tensor("transpose_42")]; - tensor inner_5 = matmul(transpose_x = inner_5_transpose_x_0, transpose_y = inner_5_transpose_y_0, x = var_603, y = v_5)[name = tensor("inner_5")]; - tensor var_605_transpose_x_0 = const()[name = tensor("op_605_transpose_x_0"), val = tensor(false)]; - tensor var_605_transpose_y_0 = const()[name = tensor("op_605_transpose_y_0"), val = tensor(false)]; - tensor var_605 = matmul(transpose_x = var_605_transpose_x_0, transpose_y = var_605_transpose_y_0, x = q_5, y = prev_kv_5)[name = tensor("op_605")]; - tensor var_606 = real_div(x = sqrt_s0_5, y = sqrt_s_t_5)[name = tensor("op_606")]; - tensor var_607 = const()[name = tensor("op_607"), val = tensor([1, 1, 3, 1])]; - tensor var_608 = reshape(shape = var_607, x = var_606)[name = tensor("op_608")]; - tensor cross_5 = mul(x = var_605, y = var_608)[name = tensor("cross_5")]; + tensor v_5 = transpose(perm = v_5_perm_0, x = var_636)[name = tensor("transpose_42")]; + tensor inner_5 = matmul(transpose_x = inner_5_transpose_x_0, transpose_y = inner_5_transpose_y_0, x = var_649, y = v_5)[name = tensor("inner_5")]; + tensor var_651_transpose_x_0 = const()[name = tensor("op_651_transpose_x_0"), val = tensor(false)]; + tensor var_651_transpose_y_0 = const()[name = tensor("op_651_transpose_y_0"), val = tensor(false)]; + tensor var_651 = matmul(transpose_x = var_651_transpose_x_0, transpose_y = var_651_transpose_y_0, x = q_5, y = prev_kv_5)[name = tensor("op_651")]; + tensor var_652 = real_div(x = sqrt_s0_5, y = sqrt_s_t_5)[name = tensor("op_652")]; + tensor var_653 = const()[name = tensor("op_653"), val = tensor([1, 1, 3, 1])]; + tensor var_654 = reshape(shape = var_653, x = var_652)[name = tensor("op_654")]; + tensor cross_5 = mul(x = var_651, y = var_654)[name = tensor("cross_5")]; tensor out_13 = add(x = inner_5, y = cross_5)[name = tensor("out_13")]; - tensor var_611 = mul(x = prev_kv_5, y = sqrt_s0_5)[name = tensor("op_611")]; - tensor var_613_transpose_x_1 = const()[name = tensor("op_613_transpose_x_1"), val = tensor(true)]; - tensor var_613_transpose_y_1 = const()[name = tensor("op_613_transpose_y_1"), val = tensor(false)]; - tensor var_613 = matmul(transpose_x = var_613_transpose_x_1, transpose_y = var_613_transpose_y_1, x = k_5, y = v_5)[name = tensor("op_613")]; - tensor new_kv_unnorm_5 = add(x = var_611, y = var_613)[name = tensor("new_kv_unnorm_5")]; - tensor var_615 = const()[name = tensor("op_615"), val = tensor(0x1.8p+1)]; - tensor new_scale_5 = add(x = prev_scale_5, y = var_615)[name = tensor("new_scale_5")]; - tensor var_617 = sqrt(x = new_scale_5)[name = tensor("op_617")]; - tensor var_618 = real_div(x = new_kv_unnorm_5, y = var_617)[name = tensor("op_618")]; - tensor var_619_perm_0 = const()[name = tensor("op_619_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_657 = mul(x = prev_kv_5, y = sqrt_s0_5)[name = tensor("op_657")]; + tensor var_659_transpose_x_1 = const()[name = tensor("op_659_transpose_x_1"), val = tensor(true)]; + tensor var_659_transpose_y_1 = const()[name = tensor("op_659_transpose_y_1"), val = tensor(false)]; + tensor var_659 = matmul(transpose_x = var_659_transpose_x_1, transpose_y = var_659_transpose_y_1, x = k_5, y = v_5)[name = tensor("op_659")]; + tensor new_kv_unnorm_5 = add(x = var_657, y = var_659)[name = tensor("new_kv_unnorm_5")]; + tensor var_661 = const()[name = tensor("op_661"), val = tensor(0x1.8p+1)]; + tensor new_scale_5 = add(x = prev_scale_5, y = var_661)[name = tensor("new_scale_5")]; + tensor var_663 = sqrt(x = new_scale_5)[name = tensor("op_663")]; + tensor var_664 = real_div(x = new_kv_unnorm_5, y = var_663)[name = tensor("op_664")]; + tensor var_665_perm_0 = const()[name = tensor("op_665_perm_0"), val = tensor([0, 2, 1, 3])]; tensor out_15_axes_0 = const()[name = tensor("out_15_axes_0"), val = tensor([-1])]; - tensor var_619 = transpose(perm = var_619_perm_0, x = out_13)[name = tensor("transpose_41")]; - tensor out_15 = layer_norm(axes = out_15_axes_0, epsilon = var_17, x = var_619)[name = tensor("out_15")]; - tensor var_623 = const()[name = tensor("op_623"), val = tensor([1, 3, 256])]; - tensor out_17 = reshape(shape = var_623, x = out_15)[name = tensor("out_17")]; - tensor var_625 = silu(x = input_97)[name = tensor("op_625")]; - tensor input_99 = mul(x = var_625, y = out_17)[name = tensor("input_99")]; - tensor ret_out_5 = linear(bias = encoder_out_proj_2_bias, weight = encoder_out_proj_2_weight, x = input_99)[name = tensor("linear_25")]; - tensor x_15 = add(x = input_93, y = ret_out_5)[name = tensor("x_15")]; + tensor var_665 = transpose(perm = var_665_perm_0, x = out_13)[name = tensor("transpose_41")]; + tensor out_15 = layer_norm(axes = out_15_axes_0, epsilon = var_64, x = var_665)[name = tensor("out_15")]; + tensor var_669 = const()[name = tensor("op_669"), val = tensor([1, 3, 256])]; + tensor out_17 = reshape(shape = var_669, x = out_15)[name = tensor("out_17")]; + tensor var_671 = silu(x = input_99)[name = tensor("op_671")]; + tensor input_101 = mul(x = var_671, y = out_17)[name = tensor("input_101")]; + tensor ret_out_5 = linear(bias = inner_encoder_out_proj_2_bias, weight = inner_encoder_out_proj_2_weight, x = input_101)[name = tensor("linear_25")]; + tensor x_15 = add(x = input_95, y = ret_out_5)[name = tensor("x_15")]; tensor window_17_begin_0 = const()[name = tensor("window_17_begin_0"), val = tensor([2, 0, 0, 0])]; tensor window_17_end_0 = const()[name = tensor("window_17_end_0"), val = tensor([3, 1, 16, 256])]; tensor window_17_end_mask_0 = const()[name = tensor("window_17_end_mask_0"), val = tensor([false, true, true, true])]; tensor window_17_squeeze_mask_0 = const()[name = tensor("window_17_squeeze_mask_0"), val = tensor([true, false, false, false])]; tensor window_17 = slice_by_index(begin = window_17_begin_0, end = window_17_end_0, end_mask = window_17_end_mask_0, squeeze_mask = window_17_squeeze_mask_0, x = enc_conv_cache)[name = tensor("window_17")]; - tensor var_633_begin_0 = const()[name = tensor("op_633_begin_0"), val = tensor([0, 0, 0])]; - tensor var_633_end_0 = const()[name = tensor("op_633_end_0"), val = tensor([1, 1, 256])]; - tensor var_633_end_mask_0 = const()[name = tensor("op_633_end_mask_0"), val = tensor([true, false, true])]; - tensor var_633 = slice_by_index(begin = var_633_begin_0, end = var_633_end_0, end_mask = var_633_end_mask_0, x = x_15)[name = tensor("op_633")]; - tensor var_636_begin_0 = const()[name = tensor("op_636_begin_0"), val = tensor([0, 1, 0])]; - tensor var_636_end_0 = const()[name = tensor("op_636_end_0"), val = tensor([1, 16, 256])]; - tensor var_636_end_mask_0 = const()[name = tensor("op_636_end_mask_0"), val = tensor([true, true, true])]; - tensor var_636 = slice_by_index(begin = var_636_begin_0, end = var_636_end_0, end_mask = var_636_end_mask_0, x = window_17)[name = tensor("op_636")]; + tensor var_679_begin_0 = const()[name = tensor("op_679_begin_0"), val = tensor([0, 0, 0])]; + tensor var_679_end_0 = const()[name = tensor("op_679_end_0"), val = tensor([1, 1, 256])]; + tensor var_679_end_mask_0 = const()[name = tensor("op_679_end_mask_0"), val = tensor([true, false, true])]; + tensor var_679 = slice_by_index(begin = var_679_begin_0, end = var_679_end_0, end_mask = var_679_end_mask_0, x = x_15)[name = tensor("op_679")]; + tensor var_682_begin_0 = const()[name = tensor("op_682_begin_0"), val = tensor([0, 1, 0])]; + tensor var_682_end_0 = const()[name = tensor("op_682_end_0"), val = tensor([1, 16, 256])]; + tensor var_682_end_mask_0 = const()[name = tensor("op_682_end_mask_0"), val = tensor([true, true, true])]; + tensor var_682 = slice_by_index(begin = var_682_begin_0, end = var_682_end_0, end_mask = var_682_end_mask_0, x = window_17)[name = tensor("op_682")]; tensor window_19_interleave_0 = const()[name = tensor("window_19_interleave_0"), val = tensor(false)]; - tensor window_19 = concat(axis = var_26, interleave = window_19_interleave_0, values = (var_636, var_633))[name = tensor("window_19")]; - tensor var_641_begin_0 = const()[name = tensor("op_641_begin_0"), val = tensor([0, 1, 0])]; - tensor var_641_end_0 = const()[name = tensor("op_641_end_0"), val = tensor([1, 2, 256])]; - tensor var_641_end_mask_0 = const()[name = tensor("op_641_end_mask_0"), val = tensor([true, false, true])]; - tensor var_641 = slice_by_index(begin = var_641_begin_0, end = var_641_end_0, end_mask = var_641_end_mask_0, x = x_15)[name = tensor("op_641")]; - tensor var_644_begin_0 = const()[name = tensor("op_644_begin_0"), val = tensor([0, 1, 0])]; - tensor var_644_end_0 = const()[name = tensor("op_644_end_0"), val = tensor([1, 16, 256])]; - tensor var_644_end_mask_0 = const()[name = tensor("op_644_end_mask_0"), val = tensor([true, true, true])]; - tensor var_644 = slice_by_index(begin = var_644_begin_0, end = var_644_end_0, end_mask = var_644_end_mask_0, x = window_19)[name = tensor("op_644")]; + tensor window_19 = concat(axis = var_72, interleave = window_19_interleave_0, values = (var_682, var_679))[name = tensor("window_19")]; + tensor var_687_begin_0 = const()[name = tensor("op_687_begin_0"), val = tensor([0, 1, 0])]; + tensor var_687_end_0 = const()[name = tensor("op_687_end_0"), val = tensor([1, 2, 256])]; + tensor var_687_end_mask_0 = const()[name = tensor("op_687_end_mask_0"), val = tensor([true, false, true])]; + tensor var_687 = slice_by_index(begin = var_687_begin_0, end = var_687_end_0, end_mask = var_687_end_mask_0, x = x_15)[name = tensor("op_687")]; + tensor var_690_begin_0 = const()[name = tensor("op_690_begin_0"), val = tensor([0, 1, 0])]; + tensor var_690_end_0 = const()[name = tensor("op_690_end_0"), val = tensor([1, 16, 256])]; + tensor var_690_end_mask_0 = const()[name = tensor("op_690_end_mask_0"), val = tensor([true, true, true])]; + tensor var_690 = slice_by_index(begin = var_690_begin_0, end = var_690_end_0, end_mask = var_690_end_mask_0, x = window_19)[name = tensor("op_690")]; tensor window_21_interleave_0 = const()[name = tensor("window_21_interleave_0"), val = tensor(false)]; - tensor window_21 = concat(axis = var_26, interleave = window_21_interleave_0, values = (var_644, var_641))[name = tensor("window_21")]; - tensor var_649_begin_0 = const()[name = tensor("op_649_begin_0"), val = tensor([0, 2, 0])]; - tensor var_649_end_0 = const()[name = tensor("op_649_end_0"), val = tensor([1, 1, 256])]; - tensor var_649_end_mask_0 = const()[name = tensor("op_649_end_mask_0"), val = tensor([true, true, true])]; - tensor var_649 = slice_by_index(begin = var_649_begin_0, end = var_649_end_0, end_mask = var_649_end_mask_0, x = x_15)[name = tensor("op_649")]; - tensor var_652_begin_0 = const()[name = tensor("op_652_begin_0"), val = tensor([0, 1, 0])]; - tensor var_652_end_0 = const()[name = tensor("op_652_end_0"), val = tensor([1, 16, 256])]; - tensor var_652_end_mask_0 = const()[name = tensor("op_652_end_mask_0"), val = tensor([true, true, true])]; - tensor var_652 = slice_by_index(begin = var_652_begin_0, end = var_652_end_0, end_mask = var_652_end_mask_0, x = window_21)[name = tensor("op_652")]; + tensor window_21 = concat(axis = var_72, interleave = window_21_interleave_0, values = (var_690, var_687))[name = tensor("window_21")]; + tensor var_695_begin_0 = const()[name = tensor("op_695_begin_0"), val = tensor([0, 2, 0])]; + tensor var_695_end_0 = const()[name = tensor("op_695_end_0"), val = tensor([1, 1, 256])]; + tensor var_695_end_mask_0 = const()[name = tensor("op_695_end_mask_0"), val = tensor([true, true, true])]; + tensor var_695 = slice_by_index(begin = var_695_begin_0, end = var_695_end_0, end_mask = var_695_end_mask_0, x = x_15)[name = tensor("op_695")]; + tensor var_698_begin_0 = const()[name = tensor("op_698_begin_0"), val = tensor([0, 1, 0])]; + tensor var_698_end_0 = const()[name = tensor("op_698_end_0"), val = tensor([1, 16, 256])]; + tensor var_698_end_mask_0 = const()[name = tensor("op_698_end_mask_0"), val = tensor([true, true, true])]; + tensor var_698 = slice_by_index(begin = var_698_begin_0, end = var_698_end_0, end_mask = var_698_end_mask_0, x = window_21)[name = tensor("op_698")]; tensor window_23_interleave_0 = const()[name = tensor("window_23_interleave_0"), val = tensor(false)]; - tensor window_23 = concat(axis = var_26, interleave = window_23_interleave_0, values = (var_652, var_649))[name = tensor("window_23")]; - tensor input_101_interleave_0 = const()[name = tensor("input_101_interleave_0"), val = tensor(false)]; - tensor input_101 = concat(axis = var_23, interleave = input_101_interleave_0, values = (window_19, window_21, window_23))[name = tensor("input_101")]; + tensor window_23 = concat(axis = var_72, interleave = window_23_interleave_0, values = (var_698, var_695))[name = tensor("window_23")]; + tensor input_103_interleave_0 = const()[name = tensor("input_103_interleave_0"), val = tensor(false)]; + tensor input_103 = concat(axis = var_59, interleave = input_103_interleave_0, values = (window_19, window_21, window_23))[name = tensor("input_103")]; tensor x_17_axes_0 = const()[name = tensor("x_17_axes_0"), val = tensor([-1])]; - tensor x_17 = layer_norm(axes = x_17_axes_0, beta = encoder_conv_module_2_sequential_0_bias, epsilon = var_28, gamma = encoder_conv_module_2_sequential_0_weight, x = input_101)[name = tensor("x_17")]; - tensor input_103_perm_0 = const()[name = tensor("input_103_perm_0"), val = tensor([0, 2, 1])]; + tensor x_17 = layer_norm(axes = x_17_axes_0, beta = inner_encoder_conv_module_2_sequential_0_bias, epsilon = var_56, gamma = inner_encoder_conv_module_2_sequential_0_weight, x = input_103)[name = tensor("x_17")]; + tensor input_105_perm_0 = const()[name = tensor("input_105_perm_0"), val = tensor([0, 2, 1])]; tensor inputs_23_pad_type_0 = const()[name = tensor("inputs_23_pad_type_0"), val = tensor("valid")]; tensor inputs_23_strides_0 = const()[name = tensor("inputs_23_strides_0"), val = tensor([1])]; tensor inputs_23_pad_0 = const()[name = tensor("inputs_23_pad_0"), val = tensor([0, 0])]; tensor inputs_23_dilations_0 = const()[name = tensor("inputs_23_dilations_0"), val = tensor([1])]; tensor inputs_23_groups_0 = const()[name = tensor("inputs_23_groups_0"), val = tensor(1)]; - tensor input_103 = transpose(perm = input_103_perm_0, x = x_17)[name = tensor("transpose_40")]; - tensor inputs_23 = conv(bias = encoder_conv_module_2_sequential_2_conv_bias, dilations = inputs_23_dilations_0, groups = inputs_23_groups_0, pad = inputs_23_pad_0, pad_type = inputs_23_pad_type_0, strides = inputs_23_strides_0, weight = encoder_conv_module_2_sequential_2_conv_weight, x = input_103)[name = tensor("inputs_23")]; - tensor var_677_split_sizes_0 = const()[name = tensor("op_677_split_sizes_0"), val = tensor([256, 256])]; - tensor var_677_axis_0 = const()[name = tensor("op_677_axis_0"), val = tensor(1)]; - tensor var_677_0, tensor var_677_1 = split(axis = var_677_axis_0, split_sizes = var_677_split_sizes_0, x = inputs_23)[name = tensor("op_677")]; - tensor var_679 = sigmoid(x = var_677_1)[name = tensor("op_679")]; - tensor inputs_25 = mul(x = var_677_0, y = var_679)[name = tensor("inputs_25")]; + tensor input_105 = transpose(perm = input_105_perm_0, x = x_17)[name = tensor("transpose_40")]; + tensor inputs_23 = conv(bias = inner_encoder_conv_module_2_sequential_2_conv_bias, dilations = inputs_23_dilations_0, groups = inputs_23_groups_0, pad = inputs_23_pad_0, pad_type = inputs_23_pad_type_0, strides = inputs_23_strides_0, weight = inner_encoder_conv_module_2_sequential_2_conv_weight, x = input_105)[name = tensor("inputs_23")]; + tensor var_723_split_sizes_0 = const()[name = tensor("op_723_split_sizes_0"), val = tensor([256, 256])]; + tensor var_723_axis_0 = const()[name = tensor("op_723_axis_0"), val = tensor(1)]; + tensor var_723_0, tensor var_723_1 = split(axis = var_723_axis_0, split_sizes = var_723_split_sizes_0, x = inputs_23)[name = tensor("op_723")]; + tensor var_725 = sigmoid(x = var_723_1)[name = tensor("op_725")]; + tensor inputs_25 = mul(x = var_723_0, y = var_725)[name = tensor("inputs_25")]; tensor outputs_aug_5_pad_type_0 = const()[name = tensor("outputs_aug_5_pad_type_0"), val = tensor("custom")]; tensor outputs_aug_5_pad_0 = const()[name = tensor("outputs_aug_5_pad_0"), val = tensor([15, 15])]; tensor outputs_aug_5_groups_0 = const()[name = tensor("outputs_aug_5_groups_0"), val = tensor(256)]; tensor outputs_aug_5_strides_0 = const()[name = tensor("outputs_aug_5_strides_0"), val = tensor([1])]; tensor outputs_aug_5_dilations_0 = const()[name = tensor("outputs_aug_5_dilations_0"), val = tensor([1])]; - tensor outputs_aug_5 = conv(dilations = outputs_aug_5_dilations_0, groups = outputs_aug_5_groups_0, pad = outputs_aug_5_pad_0, pad_type = outputs_aug_5_pad_type_0, strides = outputs_aug_5_strides_0, weight = encoder_conv_module_2_sequential_4_conv_weight, x = inputs_25)[name = tensor("outputs_aug_5")]; - tensor input_105_begin_0 = const()[name = tensor("input_105_begin_0"), val = tensor([0, 0, 0])]; - tensor input_105_end_0 = const()[name = tensor("input_105_end_0"), val = tensor([3, 256, 16])]; - tensor input_105_end_mask_0 = const()[name = tensor("input_105_end_mask_0"), val = tensor([true, true, false])]; - tensor input_105 = slice_by_index(begin = input_105_begin_0, end = input_105_end_0, end_mask = input_105_end_mask_0, x = outputs_aug_5)[name = tensor("input_105")]; - tensor inputs_27 = batch_norm(beta = encoder_conv_module_2_sequential_5_bias, epsilon = var_28, gamma = encoder_conv_module_2_sequential_5_weight, mean = encoder_conv_module_2_sequential_5_running_mean, variance = encoder_conv_module_2_sequential_5_running_var, x = input_105)[name = tensor("inputs_27")]; - tensor input_107 = silu(x = inputs_27)[name = tensor("input_107")]; - tensor input_109_pad_type_0 = const()[name = tensor("input_109_pad_type_0"), val = tensor("valid")]; - tensor input_109_strides_0 = const()[name = tensor("input_109_strides_0"), val = tensor([1])]; - tensor input_109_pad_0 = const()[name = tensor("input_109_pad_0"), val = tensor([0, 0])]; - tensor input_109_dilations_0 = const()[name = tensor("input_109_dilations_0"), val = tensor([1])]; - tensor input_109_groups_0 = const()[name = tensor("input_109_groups_0"), val = tensor(1)]; - tensor input_109 = conv(bias = encoder_conv_module_2_sequential_7_conv_bias, dilations = input_109_dilations_0, groups = input_109_groups_0, pad = input_109_pad_0, pad_type = input_109_pad_type_0, strides = input_109_strides_0, weight = encoder_conv_module_2_sequential_7_conv_weight, x = input_107)[name = tensor("input_109")]; + tensor outputs_aug_5 = conv(dilations = outputs_aug_5_dilations_0, groups = outputs_aug_5_groups_0, pad = outputs_aug_5_pad_0, pad_type = outputs_aug_5_pad_type_0, strides = outputs_aug_5_strides_0, weight = inner_encoder_conv_module_2_sequential_4_conv_weight, x = inputs_25)[name = tensor("outputs_aug_5")]; + tensor input_107_begin_0 = const()[name = tensor("input_107_begin_0"), val = tensor([0, 0, 0])]; + tensor input_107_end_0 = const()[name = tensor("input_107_end_0"), val = tensor([3, 256, 16])]; + tensor input_107_end_mask_0 = const()[name = tensor("input_107_end_mask_0"), val = tensor([true, true, false])]; + tensor input_107 = slice_by_index(begin = input_107_begin_0, end = input_107_end_0, end_mask = input_107_end_mask_0, x = outputs_aug_5)[name = tensor("input_107")]; + tensor inputs_27 = batch_norm(beta = inner_encoder_conv_module_2_sequential_5_bias, epsilon = var_56, gamma = inner_encoder_conv_module_2_sequential_5_weight, mean = inner_encoder_conv_module_2_sequential_5_running_mean, variance = inner_encoder_conv_module_2_sequential_5_running_var, x = input_107)[name = tensor("inputs_27")]; + tensor input_109 = silu(x = inputs_27)[name = tensor("input_109")]; + tensor input_111_pad_type_0 = const()[name = tensor("input_111_pad_type_0"), val = tensor("valid")]; + tensor input_111_strides_0 = const()[name = tensor("input_111_strides_0"), val = tensor([1])]; + tensor input_111_pad_0 = const()[name = tensor("input_111_pad_0"), val = tensor([0, 0])]; + tensor input_111_dilations_0 = const()[name = tensor("input_111_dilations_0"), val = tensor([1])]; + tensor input_111_groups_0 = const()[name = tensor("input_111_groups_0"), val = tensor(1)]; + tensor input_111 = conv(bias = inner_encoder_conv_module_2_sequential_7_conv_bias, dilations = input_111_dilations_0, groups = input_111_groups_0, pad = input_111_pad_0, pad_type = input_111_pad_type_0, strides = input_111_strides_0, weight = inner_encoder_conv_module_2_sequential_7_conv_weight, x = input_109)[name = tensor("input_111")]; tensor conv_out_5_perm_0 = const()[name = tensor("conv_out_5_perm_0"), val = tensor([0, 2, 1])]; - tensor var_710_begin_0 = const()[name = tensor("op_710_begin_0"), val = tensor([0, -1, 0])]; - tensor var_710_end_0 = const()[name = tensor("op_710_end_0"), val = tensor([3, 16, 256])]; - tensor var_710_end_mask_0 = const()[name = tensor("op_710_end_mask_0"), val = tensor([true, true, true])]; - tensor conv_out_5 = transpose(perm = conv_out_5_perm_0, x = input_109)[name = tensor("transpose_39")]; - tensor var_710 = slice_by_index(begin = var_710_begin_0, end = var_710_end_0, end_mask = var_710_end_mask_0, x = conv_out_5)[name = tensor("op_710")]; - tensor var_712_perm_0 = const()[name = tensor("op_712_perm_0"), val = tensor([1, 0, 2])]; - tensor var_712 = transpose(perm = var_712_perm_0, x = var_710)[name = tensor("transpose_38")]; - tensor input_111 = add(x = x_15, y = var_712)[name = tensor("input_111")]; - tensor input_113_axes_0 = const()[name = tensor("input_113_axes_0"), val = tensor([-1])]; - tensor input_113 = layer_norm(axes = input_113_axes_0, beta = encoder_ffn2_2_module_sequential_0_bias, epsilon = var_28, gamma = encoder_ffn2_2_module_sequential_0_weight, x = input_111)[name = tensor("input_113")]; - tensor inputs_29 = linear(bias = encoder_ffn2_2_module_sequential_1_linear_bias, weight = encoder_ffn2_2_module_sequential_1_linear_weight, x = input_113)[name = tensor("linear_26")]; - tensor input_115 = silu(x = inputs_29)[name = tensor("input_115")]; - tensor input_119 = linear(bias = encoder_ffn2_2_module_sequential_4_linear_bias, weight = encoder_ffn2_2_module_sequential_4_linear_weight, x = input_115)[name = tensor("linear_27")]; - tensor var_735 = const()[name = tensor("op_735"), val = tensor(0x1p-1)]; - tensor var_736 = mul(x = input_119, y = var_735)[name = tensor("op_736")]; - tensor input_121 = add(x = var_736, y = input_111)[name = tensor("input_121")]; - tensor input_123_axes_0 = const()[name = tensor("input_123_axes_0"), val = tensor([-1])]; - tensor input_123 = layer_norm(axes = input_123_axes_0, beta = encoder_layer_norm_2_bias, epsilon = var_28, gamma = encoder_layer_norm_2_weight, x = input_121)[name = tensor("input_123")]; + tensor var_756_begin_0 = const()[name = tensor("op_756_begin_0"), val = tensor([0, -1, 0])]; + tensor var_756_end_0 = const()[name = tensor("op_756_end_0"), val = tensor([3, 16, 256])]; + tensor var_756_end_mask_0 = const()[name = tensor("op_756_end_mask_0"), val = tensor([true, true, true])]; + tensor conv_out_5 = transpose(perm = conv_out_5_perm_0, x = input_111)[name = tensor("transpose_39")]; + tensor var_756 = slice_by_index(begin = var_756_begin_0, end = var_756_end_0, end_mask = var_756_end_mask_0, x = conv_out_5)[name = tensor("op_756")]; + tensor var_758_perm_0 = const()[name = tensor("op_758_perm_0"), val = tensor([1, 0, 2])]; + tensor var_758 = transpose(perm = var_758_perm_0, x = var_756)[name = tensor("transpose_38")]; + tensor input_113 = add(x = x_15, y = var_758)[name = tensor("input_113")]; + tensor input_115_axes_0 = const()[name = tensor("input_115_axes_0"), val = tensor([-1])]; + tensor input_115 = layer_norm(axes = input_115_axes_0, beta = inner_encoder_ffn2_2_module_sequential_0_bias, epsilon = var_56, gamma = inner_encoder_ffn2_2_module_sequential_0_weight, x = input_113)[name = tensor("input_115")]; + tensor inputs_29 = linear(bias = inner_encoder_ffn2_2_module_sequential_1_linear_bias, weight = inner_encoder_ffn2_2_module_sequential_1_linear_weight, x = input_115)[name = tensor("linear_26")]; + tensor input_117 = silu(x = inputs_29)[name = tensor("input_117")]; + tensor input_121 = linear(bias = inner_encoder_ffn2_2_module_sequential_4_linear_bias, weight = inner_encoder_ffn2_2_module_sequential_4_linear_weight, x = input_117)[name = tensor("linear_27")]; + tensor var_781 = const()[name = tensor("op_781"), val = tensor(0x1p-1)]; + tensor var_782 = mul(x = input_121, y = var_781)[name = tensor("op_782")]; + tensor input_123 = add(x = var_782, y = input_113)[name = tensor("input_123")]; tensor input_125_axes_0 = const()[name = tensor("input_125_axes_0"), val = tensor([-1])]; - tensor input_125 = layer_norm(axes = input_125_axes_0, beta = encoder_ffn1_3_module_sequential_0_bias, epsilon = var_28, gamma = encoder_ffn1_3_module_sequential_0_weight, x = input_123)[name = tensor("input_125")]; - tensor inputs_31 = linear(bias = encoder_ffn1_3_module_sequential_1_linear_bias, weight = encoder_ffn1_3_module_sequential_1_linear_weight, x = input_125)[name = tensor("linear_28")]; - tensor input_127 = silu(x = inputs_31)[name = tensor("input_127")]; - tensor input_131 = linear(bias = encoder_ffn1_3_module_sequential_4_linear_bias, weight = encoder_ffn1_3_module_sequential_4_linear_weight, x = input_127)[name = tensor("linear_29")]; - tensor var_765 = const()[name = tensor("op_765"), val = tensor(0x1p-1)]; - tensor var_766 = mul(x = input_131, y = var_765)[name = tensor("op_766")]; - tensor input_133 = add(x = var_766, y = input_123)[name = tensor("input_133")]; + tensor input_125 = layer_norm(axes = input_125_axes_0, beta = inner_encoder_layer_norm_2_bias, epsilon = var_56, gamma = inner_encoder_layer_norm_2_weight, x = input_123)[name = tensor("input_125")]; + tensor input_127_axes_0 = const()[name = tensor("input_127_axes_0"), val = tensor([-1])]; + tensor input_127 = layer_norm(axes = input_127_axes_0, beta = inner_encoder_ffn1_3_module_sequential_0_bias, epsilon = var_56, gamma = inner_encoder_ffn1_3_module_sequential_0_weight, x = input_125)[name = tensor("input_127")]; + tensor inputs_31 = linear(bias = inner_encoder_ffn1_3_module_sequential_1_linear_bias, weight = inner_encoder_ffn1_3_module_sequential_1_linear_weight, x = input_127)[name = tensor("linear_28")]; + tensor input_129 = silu(x = inputs_31)[name = tensor("input_129")]; + tensor input_133 = linear(bias = inner_encoder_ffn1_3_module_sequential_4_linear_bias, weight = inner_encoder_ffn1_3_module_sequential_4_linear_weight, x = input_129)[name = tensor("linear_29")]; + tensor var_811 = const()[name = tensor("op_811"), val = tensor(0x1p-1)]; + tensor var_812 = mul(x = input_133, y = var_811)[name = tensor("op_812")]; + tensor input_135 = add(x = var_812, y = input_125)[name = tensor("input_135")]; tensor x_19_axes_0 = const()[name = tensor("x_19_axes_0"), val = tensor([-1])]; - tensor x_19 = layer_norm(axes = x_19_axes_0, beta = encoder_ret_lns_3_bias, epsilon = var_28, gamma = encoder_ret_lns_3_weight, x = input_133)[name = tensor("x_19")]; + tensor x_19 = layer_norm(axes = x_19_axes_0, beta = inner_encoder_ret_lns_3_bias, epsilon = var_56, gamma = inner_encoder_ret_lns_3_weight, x = input_135)[name = tensor("x_19")]; tensor prev_kv_7_begin_0 = const()[name = tensor("prev_kv_7_begin_0"), val = tensor([3, 0, 0, 0, 0])]; tensor prev_kv_7_end_0 = const()[name = tensor("prev_kv_7_end_0"), val = tensor([4, 1, 4, 64, 64])]; tensor prev_kv_7_end_mask_0 = const()[name = tensor("prev_kv_7_end_mask_0"), val = tensor([false, true, true, true, true])]; @@ -740,199 +758,192 @@ program(1.0) tensor prev_scale_7_end_mask_0 = const()[name = tensor("prev_scale_7_end_mask_0"), val = tensor([false, true])]; tensor prev_scale_7_squeeze_mask_0 = const()[name = tensor("prev_scale_7_squeeze_mask_0"), val = tensor([true, false])]; tensor prev_scale_7 = slice_by_index(begin = prev_scale_7_begin_0, end = prev_scale_7_end_0, end_mask = prev_scale_7_end_mask_0, squeeze_mask = prev_scale_7_squeeze_mask_0, x = enc_scale)[name = tensor("prev_scale_7")]; - tensor var_780 = linear(bias = encoder_q_proj_3_bias, weight = encoder_q_proj_3_weight, x = x_19)[name = tensor("linear_30")]; - tensor var_781 = const()[name = tensor("op_781"), val = tensor([1, 3, 4, 64])]; - tensor var_782 = reshape(shape = var_781, x = var_780)[name = tensor("op_782")]; + tensor var_826 = linear(bias = inner_encoder_q_proj_3_bias, weight = inner_encoder_q_proj_3_weight, x = x_19)[name = tensor("linear_30")]; + tensor var_827 = const()[name = tensor("op_827"), val = tensor([1, 3, 4, 64])]; + tensor var_828 = reshape(shape = var_827, x = var_826)[name = tensor("op_828")]; tensor q_7_perm_0 = const()[name = tensor("q_7_perm_0"), val = tensor([0, 2, 1, 3])]; - tensor var_786 = linear(bias = encoder_k_proj_3_bias, weight = encoder_k_proj_3_weight, x = x_19)[name = tensor("linear_31")]; - tensor var_787 = const()[name = tensor("op_787"), val = tensor(0x1p-3)]; - tensor var_788 = mul(x = var_786, y = var_787)[name = tensor("op_788")]; - tensor var_789 = const()[name = tensor("op_789"), val = tensor([1, 3, 4, 64])]; - tensor var_790 = reshape(shape = var_789, x = var_788)[name = tensor("op_790")]; + tensor var_832 = linear(bias = inner_encoder_k_proj_3_bias, weight = inner_encoder_k_proj_3_weight, x = x_19)[name = tensor("linear_31")]; + tensor var_833 = const()[name = tensor("op_833"), val = tensor(0x1p-3)]; + tensor var_834 = mul(x = var_832, y = var_833)[name = tensor("op_834")]; + tensor var_835 = const()[name = tensor("op_835"), val = tensor([1, 3, 4, 64])]; + tensor var_836 = reshape(shape = var_835, x = var_834)[name = tensor("op_836")]; tensor k_7_perm_0 = const()[name = tensor("k_7_perm_0"), val = tensor([0, 2, 1, 3])]; - tensor var_794 = linear(bias = encoder_v_proj_3_bias, weight = encoder_v_proj_3_weight, x = x_19)[name = tensor("linear_32")]; - tensor var_795 = const()[name = tensor("op_795"), val = tensor([1, 3, 4, 64])]; - tensor var_796 = reshape(shape = var_795, x = var_794)[name = tensor("op_796")]; + tensor var_840 = linear(bias = inner_encoder_v_proj_3_bias, weight = inner_encoder_v_proj_3_weight, x = x_19)[name = tensor("linear_32")]; + tensor var_841 = const()[name = tensor("op_841"), val = tensor([1, 3, 4, 64])]; + tensor var_842 = reshape(shape = var_841, x = var_840)[name = tensor("op_842")]; tensor v_7_perm_0 = const()[name = tensor("v_7_perm_0"), val = tensor([0, 2, 1, 3])]; - tensor input_137 = linear(bias = encoder_g_proj_3_bias, weight = encoder_g_proj_3_weight, x = x_19)[name = tensor("linear_33")]; + tensor input_139 = linear(bias = inner_encoder_g_proj_3_bias, weight = inner_encoder_g_proj_3_weight, x = x_19)[name = tensor("linear_33")]; tensor sqrt_s0_7 = sqrt(x = prev_scale_7)[name = tensor("sqrt_s0_7")]; - tensor s_t_7 = add(x = prev_scale_7, y = encoder__t_index)[name = tensor("s_t_7")]; + tensor s_t_7 = add(x = prev_scale_7, y = inner_encoder__t_index)[name = tensor("s_t_7")]; tensor sqrt_s_t_7 = sqrt(x = s_t_7)[name = tensor("sqrt_s_t_7")]; tensor qk_7_transpose_x_1 = const()[name = tensor("qk_7_transpose_x_1"), val = tensor(false)]; tensor qk_7_transpose_y_1 = const()[name = tensor("qk_7_transpose_y_1"), val = tensor(true)]; - tensor k_7 = transpose(perm = k_7_perm_0, x = var_790)[name = tensor("transpose_36")]; - tensor q_7 = transpose(perm = q_7_perm_0, x = var_782)[name = tensor("transpose_37")]; + tensor k_7 = transpose(perm = k_7_perm_0, x = var_836)[name = tensor("transpose_36")]; + tensor q_7 = transpose(perm = q_7_perm_0, x = var_828)[name = tensor("transpose_37")]; tensor qk_7 = matmul(transpose_x = qk_7_transpose_x_1, transpose_y = qk_7_transpose_y_1, x = q_7, y = k_7)[name = tensor("qk_7")]; - tensor var_806 = const()[name = tensor("op_806"), val = tensor([3, 1])]; - tensor var_807 = reshape(shape = var_806, x = sqrt_s_t_7)[name = tensor("op_807")]; - tensor M_7 = real_div(x = encoder__causal_mask, y = var_807)[name = tensor("M_7")]; - tensor var_809 = mul(x = qk_7, y = M_7)[name = tensor("op_809")]; + tensor var_852 = const()[name = tensor("op_852"), val = tensor([3, 1])]; + tensor var_853 = reshape(shape = var_852, x = sqrt_s_t_7)[name = tensor("op_853")]; + tensor M_7 = real_div(x = inner_encoder__causal_mask, y = var_853)[name = tensor("M_7")]; + tensor var_855 = mul(x = qk_7, y = M_7)[name = tensor("op_855")]; tensor inner_7_transpose_x_0 = const()[name = tensor("inner_7_transpose_x_0"), val = tensor(false)]; tensor inner_7_transpose_y_0 = const()[name = tensor("inner_7_transpose_y_0"), val = tensor(false)]; - tensor v_7 = transpose(perm = v_7_perm_0, x = var_796)[name = tensor("transpose_35")]; - tensor inner_7 = matmul(transpose_x = inner_7_transpose_x_0, transpose_y = inner_7_transpose_y_0, x = var_809, y = v_7)[name = tensor("inner_7")]; - tensor var_811_transpose_x_0 = const()[name = tensor("op_811_transpose_x_0"), val = tensor(false)]; - tensor var_811_transpose_y_0 = const()[name = tensor("op_811_transpose_y_0"), val = tensor(false)]; - tensor var_811 = matmul(transpose_x = var_811_transpose_x_0, transpose_y = var_811_transpose_y_0, x = q_7, y = prev_kv_7)[name = tensor("op_811")]; - tensor var_812 = real_div(x = sqrt_s0_7, y = sqrt_s_t_7)[name = tensor("op_812")]; - tensor var_813 = const()[name = tensor("op_813"), val = tensor([1, 1, 3, 1])]; - tensor var_814 = reshape(shape = var_813, x = var_812)[name = tensor("op_814")]; - tensor cross_7 = mul(x = var_811, y = var_814)[name = tensor("cross_7")]; + tensor v_7 = transpose(perm = v_7_perm_0, x = var_842)[name = tensor("transpose_35")]; + tensor inner_7 = matmul(transpose_x = inner_7_transpose_x_0, transpose_y = inner_7_transpose_y_0, x = var_855, y = v_7)[name = tensor("inner_7")]; + tensor var_857_transpose_x_0 = const()[name = tensor("op_857_transpose_x_0"), val = tensor(false)]; + tensor var_857_transpose_y_0 = const()[name = tensor("op_857_transpose_y_0"), val = tensor(false)]; + tensor var_857 = matmul(transpose_x = var_857_transpose_x_0, transpose_y = var_857_transpose_y_0, x = q_7, y = prev_kv_7)[name = tensor("op_857")]; + tensor var_858 = real_div(x = sqrt_s0_7, y = sqrt_s_t_7)[name = tensor("op_858")]; + tensor var_859 = const()[name = tensor("op_859"), val = tensor([1, 1, 3, 1])]; + tensor var_860 = reshape(shape = var_859, x = var_858)[name = tensor("op_860")]; + tensor cross_7 = mul(x = var_857, y = var_860)[name = tensor("cross_7")]; tensor out_19 = add(x = inner_7, y = cross_7)[name = tensor("out_19")]; - tensor var_817 = mul(x = prev_kv_7, y = sqrt_s0_7)[name = tensor("op_817")]; - tensor var_819_transpose_x_1 = const()[name = tensor("op_819_transpose_x_1"), val = tensor(true)]; - tensor var_819_transpose_y_1 = const()[name = tensor("op_819_transpose_y_1"), val = tensor(false)]; - tensor var_819 = matmul(transpose_x = var_819_transpose_x_1, transpose_y = var_819_transpose_y_1, x = k_7, y = v_7)[name = tensor("op_819")]; - tensor new_kv_unnorm_7 = add(x = var_817, y = var_819)[name = tensor("new_kv_unnorm_7")]; - tensor var_821 = const()[name = tensor("op_821"), val = tensor(0x1.8p+1)]; - tensor new_scale_7 = add(x = prev_scale_7, y = var_821)[name = tensor("new_scale_7")]; - tensor var_823 = sqrt(x = new_scale_7)[name = tensor("op_823")]; - tensor nkv_1 = real_div(x = new_kv_unnorm_7, y = var_823)[name = tensor("nkv_1")]; - tensor var_825_perm_0 = const()[name = tensor("op_825_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_863 = mul(x = prev_kv_7, y = sqrt_s0_7)[name = tensor("op_863")]; + tensor var_865_transpose_x_1 = const()[name = tensor("op_865_transpose_x_1"), val = tensor(true)]; + tensor var_865_transpose_y_1 = const()[name = tensor("op_865_transpose_y_1"), val = tensor(false)]; + tensor var_865 = matmul(transpose_x = var_865_transpose_x_1, transpose_y = var_865_transpose_y_1, x = k_7, y = v_7)[name = tensor("op_865")]; + tensor new_kv_unnorm_7 = add(x = var_863, y = var_865)[name = tensor("new_kv_unnorm_7")]; + tensor var_867 = const()[name = tensor("op_867"), val = tensor(0x1.8p+1)]; + tensor new_scale_7 = add(x = prev_scale_7, y = var_867)[name = tensor("new_scale_7")]; + tensor var_869 = sqrt(x = new_scale_7)[name = tensor("op_869")]; + tensor nkv_1 = real_div(x = new_kv_unnorm_7, y = var_869)[name = tensor("nkv_1")]; + tensor var_871_perm_0 = const()[name = tensor("op_871_perm_0"), val = tensor([0, 2, 1, 3])]; tensor out_21_axes_0 = const()[name = tensor("out_21_axes_0"), val = tensor([-1])]; - tensor var_825 = transpose(perm = var_825_perm_0, x = out_19)[name = tensor("transpose_34")]; - tensor out_21 = layer_norm(axes = out_21_axes_0, epsilon = var_17, x = var_825)[name = tensor("out_21")]; - tensor var_829 = const()[name = tensor("op_829"), val = tensor([1, 3, 256])]; - tensor out_23 = reshape(shape = var_829, x = out_21)[name = tensor("out_23")]; - tensor var_831 = silu(x = input_137)[name = tensor("op_831")]; - tensor input_139 = mul(x = var_831, y = out_23)[name = tensor("input_139")]; - tensor ret_out_7 = linear(bias = encoder_out_proj_3_bias, weight = encoder_out_proj_3_weight, x = input_139)[name = tensor("linear_34")]; - tensor x_21 = add(x = input_133, y = ret_out_7)[name = tensor("x_21")]; + tensor var_871 = transpose(perm = var_871_perm_0, x = out_19)[name = tensor("transpose_34")]; + tensor out_21 = layer_norm(axes = out_21_axes_0, epsilon = var_64, x = var_871)[name = tensor("out_21")]; + tensor var_875 = const()[name = tensor("op_875"), val = tensor([1, 3, 256])]; + tensor out_23 = reshape(shape = var_875, x = out_21)[name = tensor("out_23")]; + tensor var_877 = silu(x = input_139)[name = tensor("op_877")]; + tensor input_141 = mul(x = var_877, y = out_23)[name = tensor("input_141")]; + tensor ret_out_7 = linear(bias = inner_encoder_out_proj_3_bias, weight = inner_encoder_out_proj_3_weight, x = input_141)[name = tensor("linear_34")]; + tensor x_21 = add(x = input_135, y = ret_out_7)[name = tensor("x_21")]; tensor window_25_begin_0 = const()[name = tensor("window_25_begin_0"), val = tensor([3, 0, 0, 0])]; tensor window_25_end_0 = const()[name = tensor("window_25_end_0"), val = tensor([4, 1, 16, 256])]; tensor window_25_end_mask_0 = const()[name = tensor("window_25_end_mask_0"), val = tensor([false, true, true, true])]; tensor window_25_squeeze_mask_0 = const()[name = tensor("window_25_squeeze_mask_0"), val = tensor([true, false, false, false])]; tensor window_25 = slice_by_index(begin = window_25_begin_0, end = window_25_end_0, end_mask = window_25_end_mask_0, squeeze_mask = window_25_squeeze_mask_0, x = enc_conv_cache)[name = tensor("window_25")]; - tensor var_839_begin_0 = const()[name = tensor("op_839_begin_0"), val = tensor([0, 0, 0])]; - tensor var_839_end_0 = const()[name = tensor("op_839_end_0"), val = tensor([1, 1, 256])]; - tensor var_839_end_mask_0 = const()[name = tensor("op_839_end_mask_0"), val = tensor([true, false, true])]; - tensor var_839 = slice_by_index(begin = var_839_begin_0, end = var_839_end_0, end_mask = var_839_end_mask_0, x = x_21)[name = tensor("op_839")]; - tensor var_842_begin_0 = const()[name = tensor("op_842_begin_0"), val = tensor([0, 1, 0])]; - tensor var_842_end_0 = const()[name = tensor("op_842_end_0"), val = tensor([1, 16, 256])]; - tensor var_842_end_mask_0 = const()[name = tensor("op_842_end_mask_0"), val = tensor([true, true, true])]; - tensor var_842 = slice_by_index(begin = var_842_begin_0, end = var_842_end_0, end_mask = var_842_end_mask_0, x = window_25)[name = tensor("op_842")]; + tensor var_885_begin_0 = const()[name = tensor("op_885_begin_0"), val = tensor([0, 0, 0])]; + tensor var_885_end_0 = const()[name = tensor("op_885_end_0"), val = tensor([1, 1, 256])]; + tensor var_885_end_mask_0 = const()[name = tensor("op_885_end_mask_0"), val = tensor([true, false, true])]; + tensor var_885 = slice_by_index(begin = var_885_begin_0, end = var_885_end_0, end_mask = var_885_end_mask_0, x = x_21)[name = tensor("op_885")]; + tensor var_888_begin_0 = const()[name = tensor("op_888_begin_0"), val = tensor([0, 1, 0])]; + tensor var_888_end_0 = const()[name = tensor("op_888_end_0"), val = tensor([1, 16, 256])]; + tensor var_888_end_mask_0 = const()[name = tensor("op_888_end_mask_0"), val = tensor([true, true, true])]; + tensor var_888 = slice_by_index(begin = var_888_begin_0, end = var_888_end_0, end_mask = var_888_end_mask_0, x = window_25)[name = tensor("op_888")]; tensor window_27_interleave_0 = const()[name = tensor("window_27_interleave_0"), val = tensor(false)]; - tensor window_27 = concat(axis = var_26, interleave = window_27_interleave_0, values = (var_842, var_839))[name = tensor("window_27")]; - tensor var_847_begin_0 = const()[name = tensor("op_847_begin_0"), val = tensor([0, 1, 0])]; - tensor var_847_end_0 = const()[name = tensor("op_847_end_0"), val = tensor([1, 2, 256])]; - tensor var_847_end_mask_0 = const()[name = tensor("op_847_end_mask_0"), val = tensor([true, false, true])]; - tensor var_847 = slice_by_index(begin = var_847_begin_0, end = var_847_end_0, end_mask = var_847_end_mask_0, x = x_21)[name = tensor("op_847")]; - tensor var_850_begin_0 = const()[name = tensor("op_850_begin_0"), val = tensor([0, 1, 0])]; - tensor var_850_end_0 = const()[name = tensor("op_850_end_0"), val = tensor([1, 16, 256])]; - tensor var_850_end_mask_0 = const()[name = tensor("op_850_end_mask_0"), val = tensor([true, true, true])]; - tensor var_850 = slice_by_index(begin = var_850_begin_0, end = var_850_end_0, end_mask = var_850_end_mask_0, x = window_27)[name = tensor("op_850")]; + tensor window_27 = concat(axis = var_72, interleave = window_27_interleave_0, values = (var_888, var_885))[name = tensor("window_27")]; + tensor var_893_begin_0 = const()[name = tensor("op_893_begin_0"), val = tensor([0, 1, 0])]; + tensor var_893_end_0 = const()[name = tensor("op_893_end_0"), val = tensor([1, 2, 256])]; + tensor var_893_end_mask_0 = const()[name = tensor("op_893_end_mask_0"), val = tensor([true, false, true])]; + tensor var_893 = slice_by_index(begin = var_893_begin_0, end = var_893_end_0, end_mask = var_893_end_mask_0, x = x_21)[name = tensor("op_893")]; + tensor var_896_begin_0 = const()[name = tensor("op_896_begin_0"), val = tensor([0, 1, 0])]; + tensor var_896_end_0 = const()[name = tensor("op_896_end_0"), val = tensor([1, 16, 256])]; + tensor var_896_end_mask_0 = const()[name = tensor("op_896_end_mask_0"), val = tensor([true, true, true])]; + tensor var_896 = slice_by_index(begin = var_896_begin_0, end = var_896_end_0, end_mask = var_896_end_mask_0, x = window_27)[name = tensor("op_896")]; tensor window_29_interleave_0 = const()[name = tensor("window_29_interleave_0"), val = tensor(false)]; - tensor window_29 = concat(axis = var_26, interleave = window_29_interleave_0, values = (var_850, var_847))[name = tensor("window_29")]; - tensor var_855_begin_0 = const()[name = tensor("op_855_begin_0"), val = tensor([0, 2, 0])]; - tensor var_855_end_0 = const()[name = tensor("op_855_end_0"), val = tensor([1, 1, 256])]; - tensor var_855_end_mask_0 = const()[name = tensor("op_855_end_mask_0"), val = tensor([true, true, true])]; - tensor var_855 = slice_by_index(begin = var_855_begin_0, end = var_855_end_0, end_mask = var_855_end_mask_0, x = x_21)[name = tensor("op_855")]; - tensor var_858_begin_0 = const()[name = tensor("op_858_begin_0"), val = tensor([0, 1, 0])]; - tensor var_858_end_0 = const()[name = tensor("op_858_end_0"), val = tensor([1, 16, 256])]; - tensor var_858_end_mask_0 = const()[name = tensor("op_858_end_mask_0"), val = tensor([true, true, true])]; - tensor var_858 = slice_by_index(begin = var_858_begin_0, end = var_858_end_0, end_mask = var_858_end_mask_0, x = window_29)[name = tensor("op_858")]; + tensor window_29 = concat(axis = var_72, interleave = window_29_interleave_0, values = (var_896, var_893))[name = tensor("window_29")]; + tensor var_901_begin_0 = const()[name = tensor("op_901_begin_0"), val = tensor([0, 2, 0])]; + tensor var_901_end_0 = const()[name = tensor("op_901_end_0"), val = tensor([1, 1, 256])]; + tensor var_901_end_mask_0 = const()[name = tensor("op_901_end_mask_0"), val = tensor([true, true, true])]; + tensor var_901 = slice_by_index(begin = var_901_begin_0, end = var_901_end_0, end_mask = var_901_end_mask_0, x = x_21)[name = tensor("op_901")]; + tensor var_904_begin_0 = const()[name = tensor("op_904_begin_0"), val = tensor([0, 1, 0])]; + tensor var_904_end_0 = const()[name = tensor("op_904_end_0"), val = tensor([1, 16, 256])]; + tensor var_904_end_mask_0 = const()[name = tensor("op_904_end_mask_0"), val = tensor([true, true, true])]; + tensor var_904 = slice_by_index(begin = var_904_begin_0, end = var_904_end_0, end_mask = var_904_end_mask_0, x = window_29)[name = tensor("op_904")]; tensor window_interleave_0 = const()[name = tensor("window_interleave_0"), val = tensor(false)]; - tensor window = concat(axis = var_26, interleave = window_interleave_0, values = (var_858, var_855))[name = tensor("window")]; - tensor input_141_interleave_0 = const()[name = tensor("input_141_interleave_0"), val = tensor(false)]; - tensor input_141 = concat(axis = var_23, interleave = input_141_interleave_0, values = (window_27, window_29, window))[name = tensor("input_141")]; + tensor window = concat(axis = var_72, interleave = window_interleave_0, values = (var_904, var_901))[name = tensor("window")]; + tensor input_143_interleave_0 = const()[name = tensor("input_143_interleave_0"), val = tensor(false)]; + tensor input_143 = concat(axis = var_59, interleave = input_143_interleave_0, values = (window_27, window_29, window))[name = tensor("input_143")]; tensor x_23_axes_0 = const()[name = tensor("x_23_axes_0"), val = tensor([-1])]; - tensor x_23 = layer_norm(axes = x_23_axes_0, beta = encoder_conv_module_3_sequential_0_bias, epsilon = var_28, gamma = encoder_conv_module_3_sequential_0_weight, x = input_141)[name = tensor("x_23")]; - tensor input_143_perm_0 = const()[name = tensor("input_143_perm_0"), val = tensor([0, 2, 1])]; + tensor x_23 = layer_norm(axes = x_23_axes_0, beta = inner_encoder_conv_module_3_sequential_0_bias, epsilon = var_56, gamma = inner_encoder_conv_module_3_sequential_0_weight, x = input_143)[name = tensor("x_23")]; + tensor input_145_perm_0 = const()[name = tensor("input_145_perm_0"), val = tensor([0, 2, 1])]; tensor inputs_33_pad_type_0 = const()[name = tensor("inputs_33_pad_type_0"), val = tensor("valid")]; tensor inputs_33_strides_0 = const()[name = tensor("inputs_33_strides_0"), val = tensor([1])]; tensor inputs_33_pad_0 = const()[name = tensor("inputs_33_pad_0"), val = tensor([0, 0])]; tensor inputs_33_dilations_0 = const()[name = tensor("inputs_33_dilations_0"), val = tensor([1])]; tensor inputs_33_groups_0 = const()[name = tensor("inputs_33_groups_0"), val = tensor(1)]; - tensor input_143 = transpose(perm = input_143_perm_0, x = x_23)[name = tensor("transpose_33")]; - tensor inputs_33 = conv(bias = encoder_conv_module_3_sequential_2_conv_bias, dilations = inputs_33_dilations_0, groups = inputs_33_groups_0, pad = inputs_33_pad_0, pad_type = inputs_33_pad_type_0, strides = inputs_33_strides_0, weight = encoder_conv_module_3_sequential_2_conv_weight, x = input_143)[name = tensor("inputs_33")]; - tensor var_883_split_sizes_0 = const()[name = tensor("op_883_split_sizes_0"), val = tensor([256, 256])]; - tensor var_883_axis_0 = const()[name = tensor("op_883_axis_0"), val = tensor(1)]; - tensor var_883_0, tensor var_883_1 = split(axis = var_883_axis_0, split_sizes = var_883_split_sizes_0, x = inputs_33)[name = tensor("op_883")]; - tensor var_885 = sigmoid(x = var_883_1)[name = tensor("op_885")]; - tensor inputs_35 = mul(x = var_883_0, y = var_885)[name = tensor("inputs_35")]; + tensor input_145 = transpose(perm = input_145_perm_0, x = x_23)[name = tensor("transpose_33")]; + tensor inputs_33 = conv(bias = inner_encoder_conv_module_3_sequential_2_conv_bias, dilations = inputs_33_dilations_0, groups = inputs_33_groups_0, pad = inputs_33_pad_0, pad_type = inputs_33_pad_type_0, strides = inputs_33_strides_0, weight = inner_encoder_conv_module_3_sequential_2_conv_weight, x = input_145)[name = tensor("inputs_33")]; + tensor var_929_split_sizes_0 = const()[name = tensor("op_929_split_sizes_0"), val = tensor([256, 256])]; + tensor var_929_axis_0 = const()[name = tensor("op_929_axis_0"), val = tensor(1)]; + tensor var_929_0, tensor var_929_1 = split(axis = var_929_axis_0, split_sizes = var_929_split_sizes_0, x = inputs_33)[name = tensor("op_929")]; + tensor var_931 = sigmoid(x = var_929_1)[name = tensor("op_931")]; + tensor inputs_35 = mul(x = var_929_0, y = var_931)[name = tensor("inputs_35")]; tensor outputs_aug_pad_type_0 = const()[name = tensor("outputs_aug_pad_type_0"), val = tensor("custom")]; tensor outputs_aug_pad_0 = const()[name = tensor("outputs_aug_pad_0"), val = tensor([15, 15])]; tensor outputs_aug_groups_0 = const()[name = tensor("outputs_aug_groups_0"), val = tensor(256)]; tensor outputs_aug_strides_0 = const()[name = tensor("outputs_aug_strides_0"), val = tensor([1])]; tensor outputs_aug_dilations_0 = const()[name = tensor("outputs_aug_dilations_0"), val = tensor([1])]; - tensor outputs_aug = conv(dilations = outputs_aug_dilations_0, groups = outputs_aug_groups_0, pad = outputs_aug_pad_0, pad_type = outputs_aug_pad_type_0, strides = outputs_aug_strides_0, weight = encoder_conv_module_3_sequential_4_conv_weight, x = inputs_35)[name = tensor("outputs_aug")]; - tensor input_145_begin_0 = const()[name = tensor("input_145_begin_0"), val = tensor([0, 0, 0])]; - tensor input_145_end_0 = const()[name = tensor("input_145_end_0"), val = tensor([3, 256, 16])]; - tensor input_145_end_mask_0 = const()[name = tensor("input_145_end_mask_0"), val = tensor([true, true, false])]; - tensor input_145 = slice_by_index(begin = input_145_begin_0, end = input_145_end_0, end_mask = input_145_end_mask_0, x = outputs_aug)[name = tensor("input_145")]; - tensor inputs_37 = batch_norm(beta = encoder_conv_module_3_sequential_5_bias, epsilon = var_28, gamma = encoder_conv_module_3_sequential_5_weight, mean = encoder_conv_module_3_sequential_5_running_mean, variance = encoder_conv_module_3_sequential_5_running_var, x = input_145)[name = tensor("inputs_37")]; - tensor input_147 = silu(x = inputs_37)[name = tensor("input_147")]; - tensor input_149_pad_type_0 = const()[name = tensor("input_149_pad_type_0"), val = tensor("valid")]; - tensor input_149_strides_0 = const()[name = tensor("input_149_strides_0"), val = tensor([1])]; - tensor input_149_pad_0 = const()[name = tensor("input_149_pad_0"), val = tensor([0, 0])]; - tensor input_149_dilations_0 = const()[name = tensor("input_149_dilations_0"), val = tensor([1])]; - tensor input_149_groups_0 = const()[name = tensor("input_149_groups_0"), val = tensor(1)]; - tensor input_149 = conv(bias = encoder_conv_module_3_sequential_7_conv_bias, dilations = input_149_dilations_0, groups = input_149_groups_0, pad = input_149_pad_0, pad_type = input_149_pad_type_0, strides = input_149_strides_0, weight = encoder_conv_module_3_sequential_7_conv_weight, x = input_147)[name = tensor("input_149")]; + tensor outputs_aug = conv(dilations = outputs_aug_dilations_0, groups = outputs_aug_groups_0, pad = outputs_aug_pad_0, pad_type = outputs_aug_pad_type_0, strides = outputs_aug_strides_0, weight = inner_encoder_conv_module_3_sequential_4_conv_weight, x = inputs_35)[name = tensor("outputs_aug")]; + tensor input_147_begin_0 = const()[name = tensor("input_147_begin_0"), val = tensor([0, 0, 0])]; + tensor input_147_end_0 = const()[name = tensor("input_147_end_0"), val = tensor([3, 256, 16])]; + tensor input_147_end_mask_0 = const()[name = tensor("input_147_end_mask_0"), val = tensor([true, true, false])]; + tensor input_147 = slice_by_index(begin = input_147_begin_0, end = input_147_end_0, end_mask = input_147_end_mask_0, x = outputs_aug)[name = tensor("input_147")]; + tensor inputs_37 = batch_norm(beta = inner_encoder_conv_module_3_sequential_5_bias, epsilon = var_56, gamma = inner_encoder_conv_module_3_sequential_5_weight, mean = inner_encoder_conv_module_3_sequential_5_running_mean, variance = inner_encoder_conv_module_3_sequential_5_running_var, x = input_147)[name = tensor("inputs_37")]; + tensor input_149 = silu(x = inputs_37)[name = tensor("input_149")]; + tensor input_151_pad_type_0 = const()[name = tensor("input_151_pad_type_0"), val = tensor("valid")]; + tensor input_151_strides_0 = const()[name = tensor("input_151_strides_0"), val = tensor([1])]; + tensor input_151_pad_0 = const()[name = tensor("input_151_pad_0"), val = tensor([0, 0])]; + tensor input_151_dilations_0 = const()[name = tensor("input_151_dilations_0"), val = tensor([1])]; + tensor input_151_groups_0 = const()[name = tensor("input_151_groups_0"), val = tensor(1)]; + tensor input_151 = conv(bias = inner_encoder_conv_module_3_sequential_7_conv_bias, dilations = input_151_dilations_0, groups = input_151_groups_0, pad = input_151_pad_0, pad_type = input_151_pad_type_0, strides = input_151_strides_0, weight = inner_encoder_conv_module_3_sequential_7_conv_weight, x = input_149)[name = tensor("input_151")]; tensor conv_out_7_perm_0 = const()[name = tensor("conv_out_7_perm_0"), val = tensor([0, 2, 1])]; - tensor var_916_begin_0 = const()[name = tensor("op_916_begin_0"), val = tensor([0, -1, 0])]; - tensor var_916_end_0 = const()[name = tensor("op_916_end_0"), val = tensor([3, 16, 256])]; - tensor var_916_end_mask_0 = const()[name = tensor("op_916_end_mask_0"), val = tensor([true, true, true])]; - tensor conv_out_7 = transpose(perm = conv_out_7_perm_0, x = input_149)[name = tensor("transpose_32")]; - tensor var_916 = slice_by_index(begin = var_916_begin_0, end = var_916_end_0, end_mask = var_916_end_mask_0, x = conv_out_7)[name = tensor("op_916")]; - tensor var_918_perm_0 = const()[name = tensor("op_918_perm_0"), val = tensor([1, 0, 2])]; - tensor var_918 = transpose(perm = var_918_perm_0, x = var_916)[name = tensor("transpose_31")]; - tensor input_151 = add(x = x_21, y = var_918)[name = tensor("input_151")]; - tensor input_153_axes_0 = const()[name = tensor("input_153_axes_0"), val = tensor([-1])]; - tensor input_153 = layer_norm(axes = input_153_axes_0, beta = encoder_ffn2_3_module_sequential_0_bias, epsilon = var_28, gamma = encoder_ffn2_3_module_sequential_0_weight, x = input_151)[name = tensor("input_153")]; - tensor inputs = linear(bias = encoder_ffn2_3_module_sequential_1_linear_bias, weight = encoder_ffn2_3_module_sequential_1_linear_weight, x = input_153)[name = tensor("linear_35")]; - tensor input_155 = silu(x = inputs)[name = tensor("input_155")]; - tensor input_159 = linear(bias = encoder_ffn2_3_module_sequential_4_linear_bias, weight = encoder_ffn2_3_module_sequential_4_linear_weight, x = input_155)[name = tensor("linear_36")]; - tensor var_941 = const()[name = tensor("op_941"), val = tensor(0x1p-1)]; - tensor var_942 = mul(x = input_159, y = var_941)[name = tensor("op_942")]; - tensor input_161 = add(x = var_942, y = input_151)[name = tensor("input_161")]; + tensor var_962_begin_0 = const()[name = tensor("op_962_begin_0"), val = tensor([0, -1, 0])]; + tensor var_962_end_0 = const()[name = tensor("op_962_end_0"), val = tensor([3, 16, 256])]; + tensor var_962_end_mask_0 = const()[name = tensor("op_962_end_mask_0"), val = tensor([true, true, true])]; + tensor conv_out_7 = transpose(perm = conv_out_7_perm_0, x = input_151)[name = tensor("transpose_32")]; + tensor var_962 = slice_by_index(begin = var_962_begin_0, end = var_962_end_0, end_mask = var_962_end_mask_0, x = conv_out_7)[name = tensor("op_962")]; + tensor var_964_perm_0 = const()[name = tensor("op_964_perm_0"), val = tensor([1, 0, 2])]; + tensor var_964 = transpose(perm = var_964_perm_0, x = var_962)[name = tensor("transpose_31")]; + tensor input_153 = add(x = x_21, y = var_964)[name = tensor("input_153")]; + tensor input_155_axes_0 = const()[name = tensor("input_155_axes_0"), val = tensor([-1])]; + tensor input_155 = layer_norm(axes = input_155_axes_0, beta = inner_encoder_ffn2_3_module_sequential_0_bias, epsilon = var_56, gamma = inner_encoder_ffn2_3_module_sequential_0_weight, x = input_153)[name = tensor("input_155")]; + tensor inputs = linear(bias = inner_encoder_ffn2_3_module_sequential_1_linear_bias, weight = inner_encoder_ffn2_3_module_sequential_1_linear_weight, x = input_155)[name = tensor("linear_35")]; + tensor input_157 = silu(x = inputs)[name = tensor("input_157")]; + tensor input_161 = linear(bias = inner_encoder_ffn2_3_module_sequential_4_linear_bias, weight = inner_encoder_ffn2_3_module_sequential_4_linear_weight, x = input_157)[name = tensor("linear_36")]; + tensor var_987 = const()[name = tensor("op_987"), val = tensor(0x1p-1)]; + tensor var_988 = mul(x = input_161, y = var_987)[name = tensor("op_988")]; + tensor input_163 = add(x = var_988, y = input_153)[name = tensor("input_163")]; tensor x_25_axes_0 = const()[name = tensor("x_25_axes_0"), val = tensor([-1])]; - tensor x_25 = layer_norm(axes = x_25_axes_0, beta = encoder_layer_norm_3_bias, epsilon = var_28, gamma = encoder_layer_norm_3_weight, x = input_161)[name = tensor("x_25")]; + tensor x_25 = layer_norm(axes = x_25_axes_0, beta = inner_encoder_layer_norm_3_bias, epsilon = var_56, gamma = inner_encoder_layer_norm_3_weight, x = input_163)[name = tensor("x_25")]; tensor x_ct_perm_0 = const()[name = tensor("x_ct_perm_0"), val = tensor([0, 2, 1])]; tensor cat_interleave_0 = const()[name = tensor("cat_interleave_0"), val = tensor(false)]; tensor x_ct = transpose(perm = x_ct_perm_0, x = x_25)[name = tensor("transpose_30")]; - tensor cat = concat(axis = var_20, interleave = cat_interleave_0, values = (cnn_window, x_ct))[name = tensor("cat")]; + tensor cat = concat(axis = var_61, interleave = cat_interleave_0, values = (cnn_window, x_ct))[name = tensor("cat")]; tensor conv_out_pad_type_0 = const()[name = tensor("conv_out_pad_type_0"), val = tensor("valid")]; tensor conv_out_strides_0 = const()[name = tensor("conv_out_strides_0"), val = tensor([1])]; tensor conv_out_pad_0 = const()[name = tensor("conv_out_pad_0"), val = tensor([0, 0])]; tensor conv_out_dilations_0 = const()[name = tensor("conv_out_dilations_0"), val = tensor([1])]; tensor conv_out_groups_0 = const()[name = tensor("conv_out_groups_0"), val = tensor(1)]; - tensor conv_out = conv(bias = encoder_cnn_bias, dilations = conv_out_dilations_0, groups = conv_out_groups_0, pad = conv_out_pad_0, pad_type = conv_out_pad_type_0, strides = conv_out_strides_0, weight = encoder_cnn_weight, x = cat)[name = tensor("conv_out")]; - tensor var_960_begin_0 = const()[name = tensor("op_960_begin_0"), val = tensor([0, 0, 3])]; - tensor var_960_end_0 = const()[name = tensor("op_960_end_0"), val = tensor([1, 256, 21])]; - tensor var_960_end_mask_0 = const()[name = tensor("op_960_end_mask_0"), val = tensor([true, true, true])]; - tensor cnn_window_new = slice_by_index(begin = var_960_begin_0, end = var_960_end_0, end_mask = var_960_end_mask_0, x = cat)[name = tensor("op_960")]; - tensor input_163_perm_0 = const()[name = tensor("input_163_perm_0"), val = tensor([0, 2, 1])]; - tensor var_962 = const()[name = tensor("op_962"), val = tensor([-1])]; - tensor input_163 = transpose(perm = input_163_perm_0, x = conv_out)[name = tensor("transpose_29")]; - tensor var_963 = reduce_l2_norm(axes = var_962, keep_dims = var_29, x = input_163)[name = tensor("op_963")]; + tensor conv_out = conv(bias = inner_encoder_cnn_bias, dilations = conv_out_dilations_0, groups = conv_out_groups_0, pad = conv_out_pad_0, pad_type = conv_out_pad_type_0, strides = conv_out_strides_0, weight = inner_encoder_cnn_weight, x = cat)[name = tensor("conv_out")]; + tensor var_1006_begin_0 = const()[name = tensor("op_1006_begin_0"), val = tensor([0, 0, 3])]; + tensor var_1006_end_0 = const()[name = tensor("op_1006_end_0"), val = tensor([1, 256, 21])]; + tensor var_1006_end_mask_0 = const()[name = tensor("op_1006_end_mask_0"), val = tensor([true, true, true])]; + tensor cnn_window_new = slice_by_index(begin = var_1006_begin_0, end = var_1006_end_0, end_mask = var_1006_end_mask_0, x = cat)[name = tensor("op_1006")]; + tensor input_165_perm_0 = const()[name = tensor("input_165_perm_0"), val = tensor([0, 2, 1])]; + tensor var_1008 = const()[name = tensor("op_1008"), val = tensor([-1])]; + tensor input_165 = transpose(perm = input_165_perm_0, x = conv_out)[name = tensor("transpose_29")]; + tensor var_1009 = reduce_l2_norm(axes = var_1008, keep_dims = var_55, x = input_165)[name = tensor("op_1009")]; tensor const_12 = const()[name = tensor("const_12"), val = tensor(0x1.fffffep+127)]; - tensor clip_0 = clip(alpha = var_11, beta = const_12, x = var_963)[name = tensor("clip_0")]; - tensor emb = real_div(x = input_163, y = clip_0)[name = tensor("emb")]; - tensor var_967_axis_0 = const()[name = tensor("op_967_axis_0"), val = tensor(0)]; - tensor enc_kv_new = stack(axis = var_967_axis_0, values = (var_206, var_412, var_618, nkv_1))[name = tensor("op_967")]; - tensor var_969_axis_0 = const()[name = tensor("op_969_axis_0"), val = tensor(0)]; - tensor enc_scale_new = stack(axis = var_969_axis_0, values = (new_scale_1, new_scale_3, new_scale_5, new_scale_7))[name = tensor("op_969")]; - tensor var_971_axis_0 = const()[name = tensor("op_971_axis_0"), val = tensor(0)]; - tensor enc_conv_cache_new = stack(axis = var_971_axis_0, values = (window_7, window_15, window_23, window))[name = tensor("op_971")]; - tensor var_980 = const()[name = tensor("op_980"), val = tensor(0x1.5798eep-27)]; - tensor var_985 = const()[name = tensor("op_985"), val = tensor(0x1.4f8b58p-17)]; - tensor var_987 = const()[name = tensor("op_987"), val = tensor(0x1.0c6f7ap-20)]; - tensor var_988 = const()[name = tensor("op_988"), val = tensor(true)]; - tensor var_990 = const()[name = tensor("op_990"), val = tensor(0x1p+0)]; - tensor var_994 = const()[name = tensor("op_994"), val = tensor(-1)]; - tensor var_1000 = const()[name = tensor("op_1000"), val = tensor(0)]; + tensor clip_0 = clip(alpha = var_69, beta = const_12, x = var_1009)[name = tensor("clip_0")]; + tensor emb = real_div(x = input_165, y = clip_0)[name = tensor("emb")]; + tensor var_1013_axis_0 = const()[name = tensor("op_1013_axis_0"), val = tensor(0)]; + tensor enc_kv_new = stack(axis = var_1013_axis_0, values = (var_252, var_458, var_664, nkv_1))[name = tensor("op_1013")]; + tensor var_1015_axis_0 = const()[name = tensor("op_1015_axis_0"), val = tensor(0)]; + tensor enc_scale_new = stack(axis = var_1015_axis_0, values = (new_scale_1, new_scale_3, new_scale_5, new_scale_7))[name = tensor("op_1015")]; + tensor var_1017_axis_0 = const()[name = tensor("op_1017_axis_0"), val = tensor(0)]; + tensor enc_conv_cache_new = stack(axis = var_1017_axis_0, values = (window_7, window_15, window_23, window))[name = tensor("op_1017")]; tensor pos = const()[name = tensor("pos"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(44395328)))]; - tensor var_1062_axes_0 = const()[name = tensor("op_1062_axes_0"), val = tensor([2])]; - tensor var_1062 = expand_dims(axes = var_1062_axes_0, x = emb)[name = tensor("op_1062")]; + tensor var_1085_axes_0 = const()[name = tensor("op_1085_axes_0"), val = tensor([2])]; + tensor var_1085 = expand_dims(axes = var_1085_axes_0, x = emb)[name = tensor("op_1085")]; tensor emb_exp_reps_0 = const()[name = tensor("emb_exp_reps_0"), val = tensor([1, 1, 12, 1])]; - tensor emb_exp = tile(reps = emb_exp_reps_0, x = var_1062)[name = tensor("emb_exp")]; - tensor input_165_interleave_0 = const()[name = tensor("input_165_interleave_0"), val = tensor(false)]; - tensor input_165 = concat(axis = var_994, interleave = input_165_interleave_0, values = (emb_exp, pos))[name = tensor("input_165")]; - tensor x_27 = linear(bias = decoder_convert_bias, weight = decoder_convert_weight, x = input_165)[name = tensor("linear_37")]; - tensor var_1070_perm_0 = const()[name = tensor("op_1070_perm_0"), val = tensor([0, 2, 1, 3])]; - tensor var_1074 = const()[name = tensor("op_1074"), val = tensor([12, 3, 256])]; - tensor var_1070 = transpose(perm = var_1070_perm_0, x = x_27)[name = tensor("transpose_28")]; - tensor x_29 = reshape(shape = var_1074, x = var_1070)[name = tensor("x_29")]; + tensor emb_exp = tile(reps = emb_exp_reps_0, x = var_1085)[name = tensor("emb_exp")]; + tensor input_167_interleave_0 = const()[name = tensor("input_167_interleave_0"), val = tensor(false)]; + tensor input_167 = concat(axis = var_62, interleave = input_167_interleave_0, values = (emb_exp, pos))[name = tensor("input_167")]; + tensor x_27 = linear(bias = inner_decoder_convert_bias, weight = inner_decoder_convert_weight, x = input_167)[name = tensor("linear_37")]; + tensor var_1093_perm_0 = const()[name = tensor("op_1093_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_1097 = const()[name = tensor("op_1097"), val = tensor([12, 3, 256])]; + tensor var_1093 = transpose(perm = var_1093_perm_0, x = x_27)[name = tensor("transpose_28")]; + tensor x_29 = reshape(shape = var_1097, x = var_1093)[name = tensor("x_29")]; tensor prev_kv_9_begin_0 = const()[name = tensor("prev_kv_9_begin_0"), val = tensor([0, 0, 0, 0, 0])]; tensor prev_kv_9_end_0 = const()[name = tensor("prev_kv_9_end_0"), val = tensor([1, 12, 4, 64, 64])]; tensor prev_kv_9_end_mask_0 = const()[name = tensor("prev_kv_9_end_mask_0"), val = tensor([false, true, true, true, true])]; @@ -943,132 +954,132 @@ program(1.0) tensor prev_scale_9_end_mask_0 = const()[name = tensor("prev_scale_9_end_mask_0"), val = tensor([false, true])]; tensor prev_scale_9_squeeze_mask_0 = const()[name = tensor("prev_scale_9_squeeze_mask_0"), val = tensor([true, false])]; tensor prev_scale_9 = slice_by_index(begin = prev_scale_9_begin_0, end = prev_scale_9_end_0, end_mask = prev_scale_9_end_mask_0, squeeze_mask = prev_scale_9_squeeze_mask_0, x = dec_scale)[name = tensor("prev_scale_9")]; - tensor var_1082 = linear(bias = decoder_q_proj_0_bias, weight = decoder_q_proj_0_weight, x = x_29)[name = tensor("linear_38")]; - tensor var_1083 = const()[name = tensor("op_1083"), val = tensor([12, 3, 4, 64])]; - tensor var_1084 = reshape(shape = var_1083, x = var_1082)[name = tensor("op_1084")]; + tensor var_1105 = linear(bias = inner_decoder_q_proj_0_bias, weight = inner_decoder_q_proj_0_weight, x = x_29)[name = tensor("linear_38")]; + tensor var_1106 = const()[name = tensor("op_1106"), val = tensor([12, 3, 4, 64])]; + tensor var_1107 = reshape(shape = var_1106, x = var_1105)[name = tensor("op_1107")]; tensor q_9_perm_0 = const()[name = tensor("q_9_perm_0"), val = tensor([0, 2, 1, 3])]; - tensor var_1088 = linear(bias = decoder_k_proj_0_bias, weight = decoder_k_proj_0_weight, x = x_29)[name = tensor("linear_39")]; - tensor var_1089 = const()[name = tensor("op_1089"), val = tensor(0x1p-3)]; - tensor var_1090 = mul(x = var_1088, y = var_1089)[name = tensor("op_1090")]; - tensor var_1091 = const()[name = tensor("op_1091"), val = tensor([12, 3, 4, 64])]; - tensor var_1092 = reshape(shape = var_1091, x = var_1090)[name = tensor("op_1092")]; + tensor var_1111 = linear(bias = inner_decoder_k_proj_0_bias, weight = inner_decoder_k_proj_0_weight, x = x_29)[name = tensor("linear_39")]; + tensor var_1112 = const()[name = tensor("op_1112"), val = tensor(0x1p-3)]; + tensor var_1113 = mul(x = var_1111, y = var_1112)[name = tensor("op_1113")]; + tensor var_1114 = const()[name = tensor("op_1114"), val = tensor([12, 3, 4, 64])]; + tensor var_1115 = reshape(shape = var_1114, x = var_1113)[name = tensor("op_1115")]; tensor k_9_perm_0 = const()[name = tensor("k_9_perm_0"), val = tensor([0, 2, 1, 3])]; - tensor var_1096 = linear(bias = decoder_v_proj_0_bias, weight = decoder_v_proj_0_weight, x = x_29)[name = tensor("linear_40")]; - tensor var_1097 = const()[name = tensor("op_1097"), val = tensor([12, 3, 4, 64])]; - tensor var_1098 = reshape(shape = var_1097, x = var_1096)[name = tensor("op_1098")]; + tensor var_1119 = linear(bias = inner_decoder_v_proj_0_bias, weight = inner_decoder_v_proj_0_weight, x = x_29)[name = tensor("linear_40")]; + tensor var_1120 = const()[name = tensor("op_1120"), val = tensor([12, 3, 4, 64])]; + tensor var_1121 = reshape(shape = var_1120, x = var_1119)[name = tensor("op_1121")]; tensor v_9_perm_0 = const()[name = tensor("v_9_perm_0"), val = tensor([0, 2, 1, 3])]; - tensor input_169 = linear(bias = decoder_g_proj_0_bias, weight = decoder_g_proj_0_weight, x = x_29)[name = tensor("linear_41")]; + tensor input_171 = linear(bias = inner_decoder_g_proj_0_bias, weight = inner_decoder_g_proj_0_weight, x = x_29)[name = tensor("linear_41")]; tensor cumsum_mask_1_exclusive_0 = const()[name = tensor("cumsum_mask_1_exclusive_0"), val = tensor(false)]; tensor cumsum_mask_1_reverse_0 = const()[name = tensor("cumsum_mask_1_reverse_0"), val = tensor(false)]; - tensor cumsum_mask_1 = cumsum(axis = var_1000, exclusive = cumsum_mask_1_exclusive_0, reverse = cumsum_mask_1_reverse_0, x = valid_mask)[name = tensor("cumsum_mask_1")]; + tensor cumsum_mask_1 = cumsum(axis = var_59, exclusive = cumsum_mask_1_exclusive_0, reverse = cumsum_mask_1_reverse_0, x = valid_mask)[name = tensor("cumsum_mask_1")]; tensor sqrt_s0_9 = sqrt(x = prev_scale_9)[name = tensor("sqrt_s0_9")]; tensor s_t_9 = add(x = prev_scale_9, y = cumsum_mask_1)[name = tensor("s_t_9")]; tensor const_20 = const()[name = tensor("const_20"), val = tensor(0x1.fffffep+127)]; - tensor clip_1 = clip(alpha = var_990, beta = const_20, x = s_t_9)[name = tensor("clip_1")]; + tensor clip_1 = clip(alpha = var_49, beta = const_20, x = s_t_9)[name = tensor("clip_1")]; tensor sqrt_s_t_9 = sqrt(x = clip_1)[name = tensor("sqrt_s_t_9")]; tensor qk_9_transpose_x_1 = const()[name = tensor("qk_9_transpose_x_1"), val = tensor(false)]; tensor qk_9_transpose_y_1 = const()[name = tensor("qk_9_transpose_y_1"), val = tensor(true)]; - tensor k_9 = transpose(perm = k_9_perm_0, x = var_1092)[name = tensor("transpose_26")]; - tensor q_9 = transpose(perm = q_9_perm_0, x = var_1084)[name = tensor("transpose_27")]; + tensor k_9 = transpose(perm = k_9_perm_0, x = var_1115)[name = tensor("transpose_26")]; + tensor q_9 = transpose(perm = q_9_perm_0, x = var_1107)[name = tensor("transpose_27")]; tensor qk_9 = matmul(transpose_x = qk_9_transpose_x_1, transpose_y = qk_9_transpose_y_1, x = q_9, y = k_9)[name = tensor("qk_9")]; - tensor var_1110 = const()[name = tensor("op_1110"), val = tensor([1, 3])]; - tensor var_1111 = reshape(shape = var_1110, x = valid_mask)[name = tensor("op_1111")]; - tensor causal_with_valid_1 = mul(x = decoder__causal_mask, y = var_1111)[name = tensor("causal_with_valid_1")]; - tensor var_1113 = const()[name = tensor("op_1113"), val = tensor([3, 1])]; - tensor var_1114 = reshape(shape = var_1113, x = sqrt_s_t_9)[name = tensor("op_1114")]; - tensor M_9 = real_div(x = causal_with_valid_1, y = var_1114)[name = tensor("M_9")]; - tensor var_1116 = mul(x = qk_9, y = M_9)[name = tensor("op_1116")]; + tensor var_1133 = const()[name = tensor("op_1133"), val = tensor([1, 3])]; + tensor var_1134 = reshape(shape = var_1133, x = valid_mask)[name = tensor("op_1134")]; + tensor causal_with_valid_1 = mul(x = inner_decoder__causal_mask, y = var_1134)[name = tensor("causal_with_valid_1")]; + tensor var_1136 = const()[name = tensor("op_1136"), val = tensor([3, 1])]; + tensor var_1137 = reshape(shape = var_1136, x = sqrt_s_t_9)[name = tensor("op_1137")]; + tensor M_9 = real_div(x = causal_with_valid_1, y = var_1137)[name = tensor("M_9")]; + tensor var_1139 = mul(x = qk_9, y = M_9)[name = tensor("op_1139")]; tensor inner_9_transpose_x_0 = const()[name = tensor("inner_9_transpose_x_0"), val = tensor(false)]; tensor inner_9_transpose_y_0 = const()[name = tensor("inner_9_transpose_y_0"), val = tensor(false)]; - tensor v_9 = transpose(perm = v_9_perm_0, x = var_1098)[name = tensor("transpose_25")]; - tensor inner_9 = matmul(transpose_x = inner_9_transpose_x_0, transpose_y = inner_9_transpose_y_0, x = var_1116, y = v_9)[name = tensor("inner_9")]; - tensor var_1118_transpose_x_0 = const()[name = tensor("op_1118_transpose_x_0"), val = tensor(false)]; - tensor var_1118_transpose_y_0 = const()[name = tensor("op_1118_transpose_y_0"), val = tensor(false)]; - tensor var_1118 = matmul(transpose_x = var_1118_transpose_x_0, transpose_y = var_1118_transpose_y_0, x = q_9, y = prev_kv_9)[name = tensor("op_1118")]; - tensor var_1119 = real_div(x = sqrt_s0_9, y = sqrt_s_t_9)[name = tensor("op_1119")]; - tensor var_1120 = const()[name = tensor("op_1120"), val = tensor([1, 1, 3, 1])]; - tensor var_1121 = reshape(shape = var_1120, x = var_1119)[name = tensor("op_1121")]; - tensor cross_9 = mul(x = var_1118, y = var_1121)[name = tensor("cross_9")]; + tensor v_9 = transpose(perm = v_9_perm_0, x = var_1121)[name = tensor("transpose_25")]; + tensor inner_9 = matmul(transpose_x = inner_9_transpose_x_0, transpose_y = inner_9_transpose_y_0, x = var_1139, y = v_9)[name = tensor("inner_9")]; + tensor var_1141_transpose_x_0 = const()[name = tensor("op_1141_transpose_x_0"), val = tensor(false)]; + tensor var_1141_transpose_y_0 = const()[name = tensor("op_1141_transpose_y_0"), val = tensor(false)]; + tensor var_1141 = matmul(transpose_x = var_1141_transpose_x_0, transpose_y = var_1141_transpose_y_0, x = q_9, y = prev_kv_9)[name = tensor("op_1141")]; + tensor var_1142 = real_div(x = sqrt_s0_9, y = sqrt_s_t_9)[name = tensor("op_1142")]; + tensor var_1143 = const()[name = tensor("op_1143"), val = tensor([1, 1, 3, 1])]; + tensor var_1144 = reshape(shape = var_1143, x = var_1142)[name = tensor("op_1144")]; + tensor cross_9 = mul(x = var_1141, y = var_1144)[name = tensor("cross_9")]; tensor out_25 = add(x = inner_9, y = cross_9)[name = tensor("out_25")]; - tensor var_1124 = const()[name = tensor("op_1124"), val = tensor([1, 1, 3, 1])]; - tensor var_1125 = reshape(shape = var_1124, x = valid_mask)[name = tensor("op_1125")]; - tensor v_masked_1 = mul(x = v_9, y = var_1125)[name = tensor("v_masked_1")]; - tensor var_1127 = mul(x = prev_kv_9, y = sqrt_s0_9)[name = tensor("op_1127")]; - tensor var_1129_transpose_x_1 = const()[name = tensor("op_1129_transpose_x_1"), val = tensor(true)]; - tensor var_1129_transpose_y_1 = const()[name = tensor("op_1129_transpose_y_1"), val = tensor(false)]; - tensor var_1129 = matmul(transpose_x = var_1129_transpose_x_1, transpose_y = var_1129_transpose_y_1, x = k_9, y = v_masked_1)[name = tensor("op_1129")]; - tensor new_kv_unnorm_9 = add(x = var_1127, y = var_1129)[name = tensor("new_kv_unnorm_9")]; - tensor var_1131_keep_dims_0 = const()[name = tensor("op_1131_keep_dims_0"), val = tensor(false)]; - tensor var_1131 = reduce_sum(keep_dims = var_1131_keep_dims_0, x = valid_mask)[name = tensor("op_1131")]; - tensor var_1132 = const()[name = tensor("op_1132"), val = tensor([1])]; - tensor var_1133 = reshape(shape = var_1132, x = var_1131)[name = tensor("op_1133")]; - tensor new_scale_9 = add(x = prev_scale_9, y = var_1133)[name = tensor("new_scale_9")]; + tensor var_1147 = const()[name = tensor("op_1147"), val = tensor([1, 1, 3, 1])]; + tensor var_1148 = reshape(shape = var_1147, x = valid_mask)[name = tensor("op_1148")]; + tensor v_masked_1 = mul(x = v_9, y = var_1148)[name = tensor("v_masked_1")]; + tensor var_1150 = mul(x = prev_kv_9, y = sqrt_s0_9)[name = tensor("op_1150")]; + tensor var_1152_transpose_x_1 = const()[name = tensor("op_1152_transpose_x_1"), val = tensor(true)]; + tensor var_1152_transpose_y_1 = const()[name = tensor("op_1152_transpose_y_1"), val = tensor(false)]; + tensor var_1152 = matmul(transpose_x = var_1152_transpose_x_1, transpose_y = var_1152_transpose_y_1, x = k_9, y = v_masked_1)[name = tensor("op_1152")]; + tensor new_kv_unnorm_9 = add(x = var_1150, y = var_1152)[name = tensor("new_kv_unnorm_9")]; + tensor var_1154_keep_dims_0 = const()[name = tensor("op_1154_keep_dims_0"), val = tensor(false)]; + tensor var_1154 = reduce_sum(keep_dims = var_1154_keep_dims_0, x = valid_mask)[name = tensor("op_1154")]; + tensor var_1155 = const()[name = tensor("op_1155"), val = tensor([1])]; + tensor var_1156 = reshape(shape = var_1155, x = var_1154)[name = tensor("op_1156")]; + tensor new_scale_9 = add(x = prev_scale_9, y = var_1156)[name = tensor("new_scale_9")]; tensor const_21 = const()[name = tensor("const_21"), val = tensor(0x1.fffffep+127)]; - tensor clip_2 = clip(alpha = var_990, beta = const_21, x = new_scale_9)[name = tensor("clip_2")]; + tensor clip_2 = clip(alpha = var_49, beta = const_21, x = new_scale_9)[name = tensor("clip_2")]; tensor sqrt_new_scale_1 = sqrt(x = clip_2)[name = tensor("sqrt_new_scale_1")]; - tensor var_1137 = real_div(x = new_kv_unnorm_9, y = sqrt_new_scale_1)[name = tensor("op_1137")]; - tensor var_1138_perm_0 = const()[name = tensor("op_1138_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_1160 = real_div(x = new_kv_unnorm_9, y = sqrt_new_scale_1)[name = tensor("op_1160")]; + tensor var_1161_perm_0 = const()[name = tensor("op_1161_perm_0"), val = tensor([0, 2, 1, 3])]; tensor out_27_axes_0 = const()[name = tensor("out_27_axes_0"), val = tensor([-1])]; - tensor var_1138 = transpose(perm = var_1138_perm_0, x = out_25)[name = tensor("transpose_24")]; - tensor out_27 = layer_norm(axes = out_27_axes_0, epsilon = var_987, x = var_1138)[name = tensor("out_27")]; - tensor var_1142 = const()[name = tensor("op_1142"), val = tensor([12, 3, 256])]; - tensor out_29 = reshape(shape = var_1142, x = out_27)[name = tensor("out_29")]; - tensor var_1144 = silu(x = input_169)[name = tensor("op_1144")]; - tensor input_171 = mul(x = var_1144, y = out_29)[name = tensor("input_171")]; - tensor ret_out_9 = linear(bias = decoder_out_proj_0_bias, weight = decoder_out_proj_0_weight, x = input_171)[name = tensor("linear_42")]; - tensor input_173 = add(x = x_29, y = ret_out_9)[name = tensor("input_173")]; + tensor var_1161 = transpose(perm = var_1161_perm_0, x = out_25)[name = tensor("transpose_24")]; + tensor out_27 = layer_norm(axes = out_27_axes_0, epsilon = var_64, x = var_1161)[name = tensor("out_27")]; + tensor var_1165 = const()[name = tensor("op_1165"), val = tensor([12, 3, 256])]; + tensor out_29 = reshape(shape = var_1165, x = out_27)[name = tensor("out_29")]; + tensor var_1167 = silu(x = input_171)[name = tensor("op_1167")]; + tensor input_173 = mul(x = var_1167, y = out_29)[name = tensor("input_173")]; + tensor ret_out_9 = linear(bias = inner_decoder_out_proj_0_bias, weight = inner_decoder_out_proj_0_weight, x = input_173)[name = tensor("linear_42")]; + tensor input_175 = add(x = x_29, y = ret_out_9)[name = tensor("input_175")]; tensor xt_1_axes_0 = const()[name = tensor("xt_1_axes_0"), val = tensor([-1])]; - tensor xt_1 = layer_norm(axes = xt_1_axes_0, beta = decoder_norm11_0_bias, epsilon = var_985, gamma = decoder_norm11_0_weight, x = input_173)[name = tensor("xt_1")]; - tensor var_1154 = const()[name = tensor("op_1154"), val = tensor([1, 12, 3, 256])]; - tensor var_1155 = reshape(shape = var_1154, x = xt_1)[name = tensor("op_1155")]; - tensor var_1156_perm_0 = const()[name = tensor("op_1156_perm_0"), val = tensor([0, 2, 1, 3])]; - tensor var_1159 = const()[name = tensor("op_1159"), val = tensor([3, 12, 256])]; - tensor var_1156 = transpose(perm = var_1156_perm_0, x = var_1155)[name = tensor("transpose_23")]; - tensor query_1 = reshape(shape = var_1159, x = var_1156)[name = tensor("query_1")]; + tensor xt_1 = layer_norm(axes = xt_1_axes_0, beta = inner_decoder_norm11_0_bias, epsilon = var_56, gamma = inner_decoder_norm11_0_weight, x = input_175)[name = tensor("xt_1")]; + tensor var_1177 = const()[name = tensor("op_1177"), val = tensor([1, 12, 3, 256])]; + tensor var_1178 = reshape(shape = var_1177, x = xt_1)[name = tensor("op_1178")]; + tensor var_1179_perm_0 = const()[name = tensor("op_1179_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_1182 = const()[name = tensor("op_1182"), val = tensor([3, 12, 256])]; + tensor var_1179 = transpose(perm = var_1179_perm_0, x = var_1178)[name = tensor("transpose_23")]; + tensor query_1 = reshape(shape = var_1182, x = var_1179)[name = tensor("query_1")]; tensor query_3_perm_0 = const()[name = tensor("query_3_perm_0"), val = tensor([1, 0, 2])]; tensor query_3 = transpose(perm = query_3_perm_0, x = query_1)[name = tensor("transpose_22")]; - tensor var_1182 = linear(bias = decoder_self_attn2_0_in_proj_bias, weight = decoder_self_attn2_0_in_proj_weight, x = query_3)[name = tensor("linear_43")]; + tensor var_1205 = linear(bias = inner_decoder_self_attn2_0_in_proj_bias, weight = inner_decoder_self_attn2_0_in_proj_weight, x = query_3)[name = tensor("linear_43")]; tensor concat_1 = const()[name = tensor("concat_1"), val = tensor([12, 3, 3, 256])]; - tensor var_1184 = reshape(shape = concat_1, x = var_1182)[name = tensor("op_1184")]; - tensor var_1185_axes_0 = const()[name = tensor("op_1185_axes_0"), val = tensor([0])]; - tensor var_1185 = expand_dims(axes = var_1185_axes_0, x = var_1184)[name = tensor("op_1185")]; - tensor var_1186_perm_0 = const()[name = tensor("op_1186_perm_0"), val = tensor([-2, 1, 2, 0, 4])]; - tensor var_1187_axes_0 = const()[name = tensor("op_1187_axes_0"), val = tensor([-2])]; - tensor var_1186 = transpose(perm = var_1186_perm_0, x = var_1185)[name = tensor("transpose_21")]; - tensor var_1187 = squeeze(axes = var_1187_axes_0, x = var_1186)[name = tensor("op_1187")]; + tensor var_1207 = reshape(shape = concat_1, x = var_1205)[name = tensor("op_1207")]; + tensor var_1208_axes_0 = const()[name = tensor("op_1208_axes_0"), val = tensor([0])]; + tensor var_1208 = expand_dims(axes = var_1208_axes_0, x = var_1207)[name = tensor("op_1208")]; + tensor var_1209_perm_0 = const()[name = tensor("op_1209_perm_0"), val = tensor([-2, 1, 2, 0, 4])]; + tensor var_1210_axes_0 = const()[name = tensor("op_1210_axes_0"), val = tensor([-2])]; + tensor var_1209 = transpose(perm = var_1209_perm_0, x = var_1208)[name = tensor("transpose_21")]; + tensor var_1210 = squeeze(axes = var_1210_axes_0, x = var_1209)[name = tensor("op_1210")]; tensor q_11_begin_0 = const()[name = tensor("q_11_begin_0"), val = tensor([0, 0, 0, 0])]; tensor q_11_end_0 = const()[name = tensor("q_11_end_0"), val = tensor([1, 12, 3, 256])]; tensor q_11_end_mask_0 = const()[name = tensor("q_11_end_mask_0"), val = tensor([false, true, true, true])]; tensor q_11_squeeze_mask_0 = const()[name = tensor("q_11_squeeze_mask_0"), val = tensor([true, false, false, false])]; - tensor q_11 = slice_by_index(begin = q_11_begin_0, end = q_11_end_0, end_mask = q_11_end_mask_0, squeeze_mask = q_11_squeeze_mask_0, x = var_1187)[name = tensor("q_11")]; + tensor q_11 = slice_by_index(begin = q_11_begin_0, end = q_11_end_0, end_mask = q_11_end_mask_0, squeeze_mask = q_11_squeeze_mask_0, x = var_1210)[name = tensor("q_11")]; tensor k_11_begin_0 = const()[name = tensor("k_11_begin_0"), val = tensor([1, 0, 0, 0])]; tensor k_11_end_0 = const()[name = tensor("k_11_end_0"), val = tensor([2, 12, 3, 256])]; tensor k_11_end_mask_0 = const()[name = tensor("k_11_end_mask_0"), val = tensor([false, true, true, true])]; tensor k_11_squeeze_mask_0 = const()[name = tensor("k_11_squeeze_mask_0"), val = tensor([true, false, false, false])]; - tensor k_11 = slice_by_index(begin = k_11_begin_0, end = k_11_end_0, end_mask = k_11_end_mask_0, squeeze_mask = k_11_squeeze_mask_0, x = var_1187)[name = tensor("k_11")]; + tensor k_11 = slice_by_index(begin = k_11_begin_0, end = k_11_end_0, end_mask = k_11_end_mask_0, squeeze_mask = k_11_squeeze_mask_0, x = var_1210)[name = tensor("k_11")]; tensor v_11_begin_0 = const()[name = tensor("v_11_begin_0"), val = tensor([2, 0, 0, 0])]; tensor v_11_end_0 = const()[name = tensor("v_11_end_0"), val = tensor([3, 12, 3, 256])]; tensor v_11_end_mask_0 = const()[name = tensor("v_11_end_mask_0"), val = tensor([false, true, true, true])]; tensor v_11_squeeze_mask_0 = const()[name = tensor("v_11_squeeze_mask_0"), val = tensor([true, false, false, false])]; - tensor v_11 = slice_by_index(begin = v_11_begin_0, end = v_11_end_0, end_mask = v_11_end_mask_0, squeeze_mask = v_11_squeeze_mask_0, x = var_1187)[name = tensor("v_11")]; - tensor var_1195 = const()[name = tensor("op_1195"), val = tensor([12, 12, 64])]; - tensor var_1196 = reshape(shape = var_1195, x = q_11)[name = tensor("op_1196")]; + tensor v_11 = slice_by_index(begin = v_11_begin_0, end = v_11_end_0, end_mask = v_11_end_mask_0, squeeze_mask = v_11_squeeze_mask_0, x = var_1210)[name = tensor("v_11")]; + tensor var_1218 = const()[name = tensor("op_1218"), val = tensor([12, 12, 64])]; + tensor var_1219 = reshape(shape = var_1218, x = q_11)[name = tensor("op_1219")]; tensor q_13_perm_0 = const()[name = tensor("q_13_perm_0"), val = tensor([1, 0, 2])]; - tensor var_1202 = const()[name = tensor("op_1202"), val = tensor([12, 12, 64])]; - tensor var_1203 = reshape(shape = var_1202, x = k_11)[name = tensor("op_1203")]; + tensor var_1225 = const()[name = tensor("op_1225"), val = tensor([12, 12, 64])]; + tensor var_1226 = reshape(shape = var_1225, x = k_11)[name = tensor("op_1226")]; tensor k_13_perm_0 = const()[name = tensor("k_13_perm_0"), val = tensor([1, 0, 2])]; - tensor var_1209 = const()[name = tensor("op_1209"), val = tensor([12, 12, 64])]; - tensor var_1210 = reshape(shape = var_1209, x = v_11)[name = tensor("op_1210")]; + tensor var_1232 = const()[name = tensor("op_1232"), val = tensor([12, 12, 64])]; + tensor var_1233 = reshape(shape = var_1232, x = v_11)[name = tensor("op_1233")]; tensor v_13_perm_0 = const()[name = tensor("v_13_perm_0"), val = tensor([1, 0, 2])]; - tensor var_1213 = const()[name = tensor("op_1213"), val = tensor([3, 4, 12, 64])]; - tensor q_13 = transpose(perm = q_13_perm_0, x = var_1196)[name = tensor("transpose_20")]; - tensor q_15 = reshape(shape = var_1213, x = q_13)[name = tensor("q_15")]; - tensor var_1215 = const()[name = tensor("op_1215"), val = tensor([3, 4, 12, 64])]; - tensor k_13 = transpose(perm = k_13_perm_0, x = var_1203)[name = tensor("transpose_19")]; - tensor k_15 = reshape(shape = var_1215, x = k_13)[name = tensor("k_15")]; - tensor var_1217 = const()[name = tensor("op_1217"), val = tensor([3, 4, 12, 64])]; - tensor v_13 = transpose(perm = v_13_perm_0, x = var_1210)[name = tensor("transpose_18")]; - tensor v_15 = reshape(shape = var_1217, x = v_13)[name = tensor("v_15")]; + tensor var_1236 = const()[name = tensor("op_1236"), val = tensor([3, 4, 12, 64])]; + tensor q_13 = transpose(perm = q_13_perm_0, x = var_1219)[name = tensor("transpose_20")]; + tensor q_15 = reshape(shape = var_1236, x = q_13)[name = tensor("q_15")]; + tensor var_1238 = const()[name = tensor("op_1238"), val = tensor([3, 4, 12, 64])]; + tensor k_13 = transpose(perm = k_13_perm_0, x = var_1226)[name = tensor("transpose_19")]; + tensor k_15 = reshape(shape = var_1238, x = k_13)[name = tensor("k_15")]; + tensor var_1240 = const()[name = tensor("op_1240"), val = tensor([3, 4, 12, 64])]; + tensor v_13 = transpose(perm = v_13_perm_0, x = var_1233)[name = tensor("transpose_18")]; + tensor v_15 = reshape(shape = var_1240, x = v_13)[name = tensor("v_15")]; tensor mul_1_y_0 = const()[name = tensor("mul_1_y_0"), val = tensor(0x1p-3)]; tensor mul_1 = mul(x = q_15, y = mul_1_y_0)[name = tensor("mul_1")]; tensor matmul_0_transpose_y_0 = const()[name = tensor("matmul_0_transpose_y_0"), val = tensor(true)]; @@ -1079,30 +1090,30 @@ program(1.0) tensor attn_output_1_transpose_x_0 = const()[name = tensor("attn_output_1_transpose_x_0"), val = tensor(false)]; tensor attn_output_1_transpose_y_0 = const()[name = tensor("attn_output_1_transpose_y_0"), val = tensor(false)]; tensor attn_output_1 = matmul(transpose_x = attn_output_1_transpose_x_0, transpose_y = attn_output_1_transpose_y_0, x = softmax_0, y = v_15)[name = tensor("attn_output_1")]; - tensor var_1220 = const()[name = tensor("op_1220"), val = tensor([2, 0, 1, 3])]; - tensor var_1225 = const()[name = tensor("op_1225"), val = tensor([36, 256])]; - tensor var_1221 = transpose(perm = var_1220, x = attn_output_1)[name = tensor("transpose_17")]; - tensor attn_output_3 = reshape(shape = var_1225, x = var_1221)[name = tensor("attn_output_3")]; - tensor attn_output_5 = linear(bias = decoder_self_attn2_0_out_proj_bias, weight = decoder_self_attn2_0_out_proj_weight, x = attn_output_3)[name = tensor("linear_44")]; - tensor var_1229 = const()[name = tensor("op_1229"), val = tensor([12, 3, 256])]; - tensor attn_output_7 = reshape(shape = var_1229, x = attn_output_5)[name = tensor("attn_output_7")]; + tensor var_1243 = const()[name = tensor("op_1243"), val = tensor([2, 0, 1, 3])]; + tensor var_1248 = const()[name = tensor("op_1248"), val = tensor([36, 256])]; + tensor var_1244 = transpose(perm = var_1243, x = attn_output_1)[name = tensor("transpose_17")]; + tensor attn_output_3 = reshape(shape = var_1248, x = var_1244)[name = tensor("attn_output_3")]; + tensor attn_output_5 = linear(bias = inner_decoder_self_attn2_0_out_proj_bias, weight = inner_decoder_self_attn2_0_out_proj_weight, x = attn_output_3)[name = tensor("linear_44")]; + tensor var_1252 = const()[name = tensor("op_1252"), val = tensor([12, 3, 256])]; + tensor attn_output_7 = reshape(shape = var_1252, x = attn_output_5)[name = tensor("attn_output_7")]; tensor sa2_out_1_perm_0 = const()[name = tensor("sa2_out_1_perm_0"), val = tensor([1, 0, 2])]; tensor sa2_out_1 = transpose(perm = sa2_out_1_perm_0, x = attn_output_7)[name = tensor("transpose_16")]; - tensor input_175 = add(x = query_1, y = sa2_out_1)[name = tensor("input_175")]; - tensor input_177_axes_0 = const()[name = tensor("input_177_axes_0"), val = tensor([-1])]; - tensor input_177 = layer_norm(axes = input_177_axes_0, beta = decoder_norm21_0_bias, epsilon = var_985, gamma = decoder_norm21_0_weight, x = input_175)[name = tensor("input_177")]; - tensor input_179 = linear(bias = decoder_linear1_0_bias, weight = decoder_linear1_0_weight, x = input_177)[name = tensor("linear_45")]; - tensor input_181 = relu(x = input_179)[name = tensor("input_181")]; - tensor ffn_out_1 = linear(bias = decoder_linear2_0_bias, weight = decoder_linear2_0_weight, x = input_181)[name = tensor("linear_46")]; - tensor input_183 = add(x = input_177, y = ffn_out_1)[name = tensor("input_183")]; + tensor input_177 = add(x = query_1, y = sa2_out_1)[name = tensor("input_177")]; + tensor input_179_axes_0 = const()[name = tensor("input_179_axes_0"), val = tensor([-1])]; + tensor input_179 = layer_norm(axes = input_179_axes_0, beta = inner_decoder_norm21_0_bias, epsilon = var_56, gamma = inner_decoder_norm21_0_weight, x = input_177)[name = tensor("input_179")]; + tensor input_181 = linear(bias = inner_decoder_linear1_0_bias, weight = inner_decoder_linear1_0_weight, x = input_179)[name = tensor("linear_45")]; + tensor input_183 = relu(x = input_181)[name = tensor("input_183")]; + tensor ffn_out_1 = linear(bias = inner_decoder_linear2_0_bias, weight = inner_decoder_linear2_0_weight, x = input_183)[name = tensor("linear_46")]; + tensor input_185 = add(x = input_179, y = ffn_out_1)[name = tensor("input_185")]; tensor xt_3_axes_0 = const()[name = tensor("xt_3_axes_0"), val = tensor([-1])]; - tensor xt_3 = layer_norm(axes = xt_3_axes_0, beta = decoder_norm22_0_bias, epsilon = var_985, gamma = decoder_norm22_0_weight, x = input_183)[name = tensor("xt_3")]; - tensor var_1249 = const()[name = tensor("op_1249"), val = tensor([1, 3, 12, 256])]; - tensor x_31 = reshape(shape = var_1249, x = xt_3)[name = tensor("x_31")]; - tensor var_1251_perm_0 = const()[name = tensor("op_1251_perm_0"), val = tensor([0, 2, 1, 3])]; - tensor var_1255 = const()[name = tensor("op_1255"), val = tensor([12, 3, 256])]; - tensor var_1251 = transpose(perm = var_1251_perm_0, x = x_31)[name = tensor("transpose_15")]; - tensor x = reshape(shape = var_1255, x = var_1251)[name = tensor("x")]; + tensor xt_3 = layer_norm(axes = xt_3_axes_0, beta = inner_decoder_norm22_0_bias, epsilon = var_56, gamma = inner_decoder_norm22_0_weight, x = input_185)[name = tensor("xt_3")]; + tensor var_1272 = const()[name = tensor("op_1272"), val = tensor([1, 3, 12, 256])]; + tensor x_31 = reshape(shape = var_1272, x = xt_3)[name = tensor("x_31")]; + tensor var_1274_perm_0 = const()[name = tensor("op_1274_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_1278 = const()[name = tensor("op_1278"), val = tensor([12, 3, 256])]; + tensor var_1274 = transpose(perm = var_1274_perm_0, x = x_31)[name = tensor("transpose_15")]; + tensor x = reshape(shape = var_1278, x = var_1274)[name = tensor("x")]; tensor prev_kv_begin_0 = const()[name = tensor("prev_kv_begin_0"), val = tensor([1, 0, 0, 0, 0])]; tensor prev_kv_end_0 = const()[name = tensor("prev_kv_end_0"), val = tensor([2, 12, 4, 64, 64])]; tensor prev_kv_end_mask_0 = const()[name = tensor("prev_kv_end_mask_0"), val = tensor([false, true, true, true, true])]; @@ -1113,120 +1124,120 @@ program(1.0) tensor prev_scale_end_mask_0 = const()[name = tensor("prev_scale_end_mask_0"), val = tensor([false, true])]; tensor prev_scale_squeeze_mask_0 = const()[name = tensor("prev_scale_squeeze_mask_0"), val = tensor([true, false])]; tensor prev_scale = slice_by_index(begin = prev_scale_begin_0, end = prev_scale_end_0, end_mask = prev_scale_end_mask_0, squeeze_mask = prev_scale_squeeze_mask_0, x = dec_scale)[name = tensor("prev_scale")]; - tensor var_1263 = linear(bias = decoder_q_proj_1_bias, weight = decoder_q_proj_1_weight, x = x)[name = tensor("linear_47")]; - tensor var_1264 = const()[name = tensor("op_1264"), val = tensor([12, 3, 4, 64])]; - tensor var_1265 = reshape(shape = var_1264, x = var_1263)[name = tensor("op_1265")]; + tensor var_1286 = linear(bias = inner_decoder_q_proj_1_bias, weight = inner_decoder_q_proj_1_weight, x = x)[name = tensor("linear_47")]; + tensor var_1287 = const()[name = tensor("op_1287"), val = tensor([12, 3, 4, 64])]; + tensor var_1288 = reshape(shape = var_1287, x = var_1286)[name = tensor("op_1288")]; tensor q_17_perm_0 = const()[name = tensor("q_17_perm_0"), val = tensor([0, 2, 1, 3])]; - tensor var_1269 = linear(bias = decoder_k_proj_1_bias, weight = decoder_k_proj_1_weight, x = x)[name = tensor("linear_48")]; - tensor var_1270 = const()[name = tensor("op_1270"), val = tensor(0x1p-3)]; - tensor var_1271 = mul(x = var_1269, y = var_1270)[name = tensor("op_1271")]; - tensor var_1272 = const()[name = tensor("op_1272"), val = tensor([12, 3, 4, 64])]; - tensor var_1273 = reshape(shape = var_1272, x = var_1271)[name = tensor("op_1273")]; + tensor var_1292 = linear(bias = inner_decoder_k_proj_1_bias, weight = inner_decoder_k_proj_1_weight, x = x)[name = tensor("linear_48")]; + tensor var_1293 = const()[name = tensor("op_1293"), val = tensor(0x1p-3)]; + tensor var_1294 = mul(x = var_1292, y = var_1293)[name = tensor("op_1294")]; + tensor var_1295 = const()[name = tensor("op_1295"), val = tensor([12, 3, 4, 64])]; + tensor var_1296 = reshape(shape = var_1295, x = var_1294)[name = tensor("op_1296")]; tensor k_17_perm_0 = const()[name = tensor("k_17_perm_0"), val = tensor([0, 2, 1, 3])]; - tensor var_1277 = linear(bias = decoder_v_proj_1_bias, weight = decoder_v_proj_1_weight, x = x)[name = tensor("linear_49")]; - tensor var_1278 = const()[name = tensor("op_1278"), val = tensor([12, 3, 4, 64])]; - tensor var_1279 = reshape(shape = var_1278, x = var_1277)[name = tensor("op_1279")]; + tensor var_1300 = linear(bias = inner_decoder_v_proj_1_bias, weight = inner_decoder_v_proj_1_weight, x = x)[name = tensor("linear_49")]; + tensor var_1301 = const()[name = tensor("op_1301"), val = tensor([12, 3, 4, 64])]; + tensor var_1302 = reshape(shape = var_1301, x = var_1300)[name = tensor("op_1302")]; tensor v_17_perm_0 = const()[name = tensor("v_17_perm_0"), val = tensor([0, 2, 1, 3])]; - tensor input_187 = linear(bias = decoder_g_proj_1_bias, weight = decoder_g_proj_1_weight, x = x)[name = tensor("linear_50")]; + tensor input_189 = linear(bias = inner_decoder_g_proj_1_bias, weight = inner_decoder_g_proj_1_weight, x = x)[name = tensor("linear_50")]; tensor sqrt_s0 = sqrt(x = prev_scale)[name = tensor("sqrt_s0")]; tensor s_t = add(x = prev_scale, y = cumsum_mask_1)[name = tensor("s_t")]; tensor const_32 = const()[name = tensor("const_32"), val = tensor(0x1.fffffep+127)]; - tensor clip_3 = clip(alpha = var_990, beta = const_32, x = s_t)[name = tensor("clip_3")]; + tensor clip_3 = clip(alpha = var_49, beta = const_32, x = s_t)[name = tensor("clip_3")]; tensor sqrt_s_t = sqrt(x = clip_3)[name = tensor("sqrt_s_t")]; tensor qk_transpose_x_1 = const()[name = tensor("qk_transpose_x_1"), val = tensor(false)]; tensor qk_transpose_y_1 = const()[name = tensor("qk_transpose_y_1"), val = tensor(true)]; - tensor k_17 = transpose(perm = k_17_perm_0, x = var_1273)[name = tensor("transpose_13")]; - tensor q_17 = transpose(perm = q_17_perm_0, x = var_1265)[name = tensor("transpose_14")]; + tensor k_17 = transpose(perm = k_17_perm_0, x = var_1296)[name = tensor("transpose_13")]; + tensor q_17 = transpose(perm = q_17_perm_0, x = var_1288)[name = tensor("transpose_14")]; tensor qk = matmul(transpose_x = qk_transpose_x_1, transpose_y = qk_transpose_y_1, x = q_17, y = k_17)[name = tensor("qk")]; - tensor var_1294 = const()[name = tensor("op_1294"), val = tensor([3, 1])]; - tensor var_1295 = reshape(shape = var_1294, x = sqrt_s_t)[name = tensor("op_1295")]; - tensor M = real_div(x = causal_with_valid_1, y = var_1295)[name = tensor("M")]; - tensor var_1297 = mul(x = qk, y = M)[name = tensor("op_1297")]; - tensor inner_transpose_x_0 = const()[name = tensor("inner_transpose_x_0"), val = tensor(false)]; - tensor inner_transpose_y_0 = const()[name = tensor("inner_transpose_y_0"), val = tensor(false)]; - tensor v_17 = transpose(perm = v_17_perm_0, x = var_1279)[name = tensor("transpose_12")]; - tensor inner = matmul(transpose_x = inner_transpose_x_0, transpose_y = inner_transpose_y_0, x = var_1297, y = v_17)[name = tensor("inner")]; - tensor var_1299_transpose_x_0 = const()[name = tensor("op_1299_transpose_x_0"), val = tensor(false)]; - tensor var_1299_transpose_y_0 = const()[name = tensor("op_1299_transpose_y_0"), val = tensor(false)]; - tensor var_1299 = matmul(transpose_x = var_1299_transpose_x_0, transpose_y = var_1299_transpose_y_0, x = q_17, y = prev_kv)[name = tensor("op_1299")]; - tensor var_1300 = real_div(x = sqrt_s0, y = sqrt_s_t)[name = tensor("op_1300")]; - tensor var_1301 = const()[name = tensor("op_1301"), val = tensor([1, 1, 3, 1])]; - tensor var_1302 = reshape(shape = var_1301, x = var_1300)[name = tensor("op_1302")]; - tensor cross = mul(x = var_1299, y = var_1302)[name = tensor("cross")]; - tensor out_31 = add(x = inner, y = cross)[name = tensor("out_31")]; - tensor v_masked = mul(x = v_17, y = var_1125)[name = tensor("v_masked")]; - tensor var_1308 = mul(x = prev_kv, y = sqrt_s0)[name = tensor("op_1308")]; - tensor var_1310_transpose_x_1 = const()[name = tensor("op_1310_transpose_x_1"), val = tensor(true)]; - tensor var_1310_transpose_y_1 = const()[name = tensor("op_1310_transpose_y_1"), val = tensor(false)]; - tensor var_1310 = matmul(transpose_x = var_1310_transpose_x_1, transpose_y = var_1310_transpose_y_1, x = k_17, y = v_masked)[name = tensor("op_1310")]; - tensor new_kv_unnorm = add(x = var_1308, y = var_1310)[name = tensor("new_kv_unnorm")]; - tensor new_scale = add(x = prev_scale, y = var_1133)[name = tensor("new_scale")]; + tensor var_1317 = const()[name = tensor("op_1317"), val = tensor([3, 1])]; + tensor var_1318 = reshape(shape = var_1317, x = sqrt_s_t)[name = tensor("op_1318")]; + tensor M = real_div(x = causal_with_valid_1, y = var_1318)[name = tensor("M")]; + tensor var_1320 = mul(x = qk, y = M)[name = tensor("op_1320")]; + tensor inner_11_transpose_x_0 = const()[name = tensor("inner_11_transpose_x_0"), val = tensor(false)]; + tensor inner_11_transpose_y_0 = const()[name = tensor("inner_11_transpose_y_0"), val = tensor(false)]; + tensor v_17 = transpose(perm = v_17_perm_0, x = var_1302)[name = tensor("transpose_12")]; + tensor inner_11 = matmul(transpose_x = inner_11_transpose_x_0, transpose_y = inner_11_transpose_y_0, x = var_1320, y = v_17)[name = tensor("inner_11")]; + tensor var_1322_transpose_x_0 = const()[name = tensor("op_1322_transpose_x_0"), val = tensor(false)]; + tensor var_1322_transpose_y_0 = const()[name = tensor("op_1322_transpose_y_0"), val = tensor(false)]; + tensor var_1322 = matmul(transpose_x = var_1322_transpose_x_0, transpose_y = var_1322_transpose_y_0, x = q_17, y = prev_kv)[name = tensor("op_1322")]; + tensor var_1323 = real_div(x = sqrt_s0, y = sqrt_s_t)[name = tensor("op_1323")]; + tensor var_1324 = const()[name = tensor("op_1324"), val = tensor([1, 1, 3, 1])]; + tensor var_1325 = reshape(shape = var_1324, x = var_1323)[name = tensor("op_1325")]; + tensor cross = mul(x = var_1322, y = var_1325)[name = tensor("cross")]; + tensor out_31 = add(x = inner_11, y = cross)[name = tensor("out_31")]; + tensor v_masked = mul(x = v_17, y = var_1148)[name = tensor("v_masked")]; + tensor var_1331 = mul(x = prev_kv, y = sqrt_s0)[name = tensor("op_1331")]; + tensor var_1333_transpose_x_1 = const()[name = tensor("op_1333_transpose_x_1"), val = tensor(true)]; + tensor var_1333_transpose_y_1 = const()[name = tensor("op_1333_transpose_y_1"), val = tensor(false)]; + tensor var_1333 = matmul(transpose_x = var_1333_transpose_x_1, transpose_y = var_1333_transpose_y_1, x = k_17, y = v_masked)[name = tensor("op_1333")]; + tensor new_kv_unnorm = add(x = var_1331, y = var_1333)[name = tensor("new_kv_unnorm")]; + tensor new_scale = add(x = prev_scale, y = var_1156)[name = tensor("new_scale")]; tensor const_33 = const()[name = tensor("const_33"), val = tensor(0x1.fffffep+127)]; - tensor clip_4 = clip(alpha = var_990, beta = const_33, x = new_scale)[name = tensor("clip_4")]; + tensor clip_4 = clip(alpha = var_49, beta = const_33, x = new_scale)[name = tensor("clip_4")]; tensor sqrt_new_scale = sqrt(x = clip_4)[name = tensor("sqrt_new_scale")]; tensor nkv = real_div(x = new_kv_unnorm, y = sqrt_new_scale)[name = tensor("nkv")]; - tensor var_1319_perm_0 = const()[name = tensor("op_1319_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_1342_perm_0 = const()[name = tensor("op_1342_perm_0"), val = tensor([0, 2, 1, 3])]; tensor out_33_axes_0 = const()[name = tensor("out_33_axes_0"), val = tensor([-1])]; - tensor var_1319 = transpose(perm = var_1319_perm_0, x = out_31)[name = tensor("transpose_11")]; - tensor out_33 = layer_norm(axes = out_33_axes_0, epsilon = var_987, x = var_1319)[name = tensor("out_33")]; - tensor var_1323 = const()[name = tensor("op_1323"), val = tensor([12, 3, 256])]; - tensor out = reshape(shape = var_1323, x = out_33)[name = tensor("out")]; - tensor var_1325 = silu(x = input_187)[name = tensor("op_1325")]; - tensor input_189 = mul(x = var_1325, y = out)[name = tensor("input_189")]; - tensor ret_out = linear(bias = decoder_out_proj_1_bias, weight = decoder_out_proj_1_weight, x = input_189)[name = tensor("linear_51")]; - tensor input_191 = add(x = x, y = ret_out)[name = tensor("input_191")]; + tensor var_1342 = transpose(perm = var_1342_perm_0, x = out_31)[name = tensor("transpose_11")]; + tensor out_33 = layer_norm(axes = out_33_axes_0, epsilon = var_64, x = var_1342)[name = tensor("out_33")]; + tensor var_1346 = const()[name = tensor("op_1346"), val = tensor([12, 3, 256])]; + tensor out = reshape(shape = var_1346, x = out_33)[name = tensor("out")]; + tensor var_1348 = silu(x = input_189)[name = tensor("op_1348")]; + tensor input_191 = mul(x = var_1348, y = out)[name = tensor("input_191")]; + tensor ret_out = linear(bias = inner_decoder_out_proj_1_bias, weight = inner_decoder_out_proj_1_weight, x = input_191)[name = tensor("linear_51")]; + tensor input_193 = add(x = x, y = ret_out)[name = tensor("input_193")]; tensor xt_5_axes_0 = const()[name = tensor("xt_5_axes_0"), val = tensor([-1])]; - tensor xt_5 = layer_norm(axes = xt_5_axes_0, beta = decoder_norm11_1_bias, epsilon = var_985, gamma = decoder_norm11_1_weight, x = input_191)[name = tensor("xt_5")]; - tensor var_1335 = const()[name = tensor("op_1335"), val = tensor([1, 12, 3, 256])]; - tensor var_1336 = reshape(shape = var_1335, x = xt_5)[name = tensor("op_1336")]; - tensor var_1337_perm_0 = const()[name = tensor("op_1337_perm_0"), val = tensor([0, 2, 1, 3])]; - tensor var_1340 = const()[name = tensor("op_1340"), val = tensor([3, 12, 256])]; - tensor var_1337 = transpose(perm = var_1337_perm_0, x = var_1336)[name = tensor("transpose_10")]; - tensor query_5 = reshape(shape = var_1340, x = var_1337)[name = tensor("query_5")]; + tensor xt_5 = layer_norm(axes = xt_5_axes_0, beta = inner_decoder_norm11_1_bias, epsilon = var_56, gamma = inner_decoder_norm11_1_weight, x = input_193)[name = tensor("xt_5")]; + tensor var_1358 = const()[name = tensor("op_1358"), val = tensor([1, 12, 3, 256])]; + tensor var_1359 = reshape(shape = var_1358, x = xt_5)[name = tensor("op_1359")]; + tensor var_1360_perm_0 = const()[name = tensor("op_1360_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_1363 = const()[name = tensor("op_1363"), val = tensor([3, 12, 256])]; + tensor var_1360 = transpose(perm = var_1360_perm_0, x = var_1359)[name = tensor("transpose_10")]; + tensor query_5 = reshape(shape = var_1363, x = var_1360)[name = tensor("query_5")]; tensor query_perm_0 = const()[name = tensor("query_perm_0"), val = tensor([1, 0, 2])]; tensor query = transpose(perm = query_perm_0, x = query_5)[name = tensor("transpose_9")]; - tensor var_1363 = linear(bias = decoder_self_attn2_1_in_proj_bias, weight = decoder_self_attn2_1_in_proj_weight, x = query)[name = tensor("linear_52")]; + tensor var_1386 = linear(bias = inner_decoder_self_attn2_1_in_proj_bias, weight = inner_decoder_self_attn2_1_in_proj_weight, x = query)[name = tensor("linear_52")]; tensor concat_2 = const()[name = tensor("concat_2"), val = tensor([12, 3, 3, 256])]; - tensor var_1365 = reshape(shape = concat_2, x = var_1363)[name = tensor("op_1365")]; - tensor var_1366_axes_0 = const()[name = tensor("op_1366_axes_0"), val = tensor([0])]; - tensor var_1366 = expand_dims(axes = var_1366_axes_0, x = var_1365)[name = tensor("op_1366")]; - tensor var_1367_perm_0 = const()[name = tensor("op_1367_perm_0"), val = tensor([-2, 1, 2, 0, 4])]; - tensor var_1368_axes_0 = const()[name = tensor("op_1368_axes_0"), val = tensor([-2])]; - tensor var_1367 = transpose(perm = var_1367_perm_0, x = var_1366)[name = tensor("transpose_8")]; - tensor var_1368 = squeeze(axes = var_1368_axes_0, x = var_1367)[name = tensor("op_1368")]; + tensor var_1388 = reshape(shape = concat_2, x = var_1386)[name = tensor("op_1388")]; + tensor var_1389_axes_0 = const()[name = tensor("op_1389_axes_0"), val = tensor([0])]; + tensor var_1389 = expand_dims(axes = var_1389_axes_0, x = var_1388)[name = tensor("op_1389")]; + tensor var_1390_perm_0 = const()[name = tensor("op_1390_perm_0"), val = tensor([-2, 1, 2, 0, 4])]; + tensor var_1391_axes_0 = const()[name = tensor("op_1391_axes_0"), val = tensor([-2])]; + tensor var_1390 = transpose(perm = var_1390_perm_0, x = var_1389)[name = tensor("transpose_8")]; + tensor var_1391 = squeeze(axes = var_1391_axes_0, x = var_1390)[name = tensor("op_1391")]; tensor q_19_begin_0 = const()[name = tensor("q_19_begin_0"), val = tensor([0, 0, 0, 0])]; tensor q_19_end_0 = const()[name = tensor("q_19_end_0"), val = tensor([1, 12, 3, 256])]; tensor q_19_end_mask_0 = const()[name = tensor("q_19_end_mask_0"), val = tensor([false, true, true, true])]; tensor q_19_squeeze_mask_0 = const()[name = tensor("q_19_squeeze_mask_0"), val = tensor([true, false, false, false])]; - tensor q_19 = slice_by_index(begin = q_19_begin_0, end = q_19_end_0, end_mask = q_19_end_mask_0, squeeze_mask = q_19_squeeze_mask_0, x = var_1368)[name = tensor("q_19")]; + tensor q_19 = slice_by_index(begin = q_19_begin_0, end = q_19_end_0, end_mask = q_19_end_mask_0, squeeze_mask = q_19_squeeze_mask_0, x = var_1391)[name = tensor("q_19")]; tensor k_19_begin_0 = const()[name = tensor("k_19_begin_0"), val = tensor([1, 0, 0, 0])]; tensor k_19_end_0 = const()[name = tensor("k_19_end_0"), val = tensor([2, 12, 3, 256])]; tensor k_19_end_mask_0 = const()[name = tensor("k_19_end_mask_0"), val = tensor([false, true, true, true])]; tensor k_19_squeeze_mask_0 = const()[name = tensor("k_19_squeeze_mask_0"), val = tensor([true, false, false, false])]; - tensor k_19 = slice_by_index(begin = k_19_begin_0, end = k_19_end_0, end_mask = k_19_end_mask_0, squeeze_mask = k_19_squeeze_mask_0, x = var_1368)[name = tensor("k_19")]; + tensor k_19 = slice_by_index(begin = k_19_begin_0, end = k_19_end_0, end_mask = k_19_end_mask_0, squeeze_mask = k_19_squeeze_mask_0, x = var_1391)[name = tensor("k_19")]; tensor v_19_begin_0 = const()[name = tensor("v_19_begin_0"), val = tensor([2, 0, 0, 0])]; tensor v_19_end_0 = const()[name = tensor("v_19_end_0"), val = tensor([3, 12, 3, 256])]; tensor v_19_end_mask_0 = const()[name = tensor("v_19_end_mask_0"), val = tensor([false, true, true, true])]; tensor v_19_squeeze_mask_0 = const()[name = tensor("v_19_squeeze_mask_0"), val = tensor([true, false, false, false])]; - tensor v_19 = slice_by_index(begin = v_19_begin_0, end = v_19_end_0, end_mask = v_19_end_mask_0, squeeze_mask = v_19_squeeze_mask_0, x = var_1368)[name = tensor("v_19")]; - tensor var_1376 = const()[name = tensor("op_1376"), val = tensor([12, 12, 64])]; - tensor var_1377 = reshape(shape = var_1376, x = q_19)[name = tensor("op_1377")]; + tensor v_19 = slice_by_index(begin = v_19_begin_0, end = v_19_end_0, end_mask = v_19_end_mask_0, squeeze_mask = v_19_squeeze_mask_0, x = var_1391)[name = tensor("v_19")]; + tensor var_1399 = const()[name = tensor("op_1399"), val = tensor([12, 12, 64])]; + tensor var_1400 = reshape(shape = var_1399, x = q_19)[name = tensor("op_1400")]; tensor q_21_perm_0 = const()[name = tensor("q_21_perm_0"), val = tensor([1, 0, 2])]; - tensor var_1383 = const()[name = tensor("op_1383"), val = tensor([12, 12, 64])]; - tensor var_1384 = reshape(shape = var_1383, x = k_19)[name = tensor("op_1384")]; + tensor var_1406 = const()[name = tensor("op_1406"), val = tensor([12, 12, 64])]; + tensor var_1407 = reshape(shape = var_1406, x = k_19)[name = tensor("op_1407")]; tensor k_21_perm_0 = const()[name = tensor("k_21_perm_0"), val = tensor([1, 0, 2])]; - tensor var_1390 = const()[name = tensor("op_1390"), val = tensor([12, 12, 64])]; - tensor var_1391 = reshape(shape = var_1390, x = v_19)[name = tensor("op_1391")]; + tensor var_1413 = const()[name = tensor("op_1413"), val = tensor([12, 12, 64])]; + tensor var_1414 = reshape(shape = var_1413, x = v_19)[name = tensor("op_1414")]; tensor v_21_perm_0 = const()[name = tensor("v_21_perm_0"), val = tensor([1, 0, 2])]; - tensor var_1394 = const()[name = tensor("op_1394"), val = tensor([3, 4, 12, 64])]; - tensor q_21 = transpose(perm = q_21_perm_0, x = var_1377)[name = tensor("transpose_7")]; - tensor q = reshape(shape = var_1394, x = q_21)[name = tensor("q")]; - tensor var_1396 = const()[name = tensor("op_1396"), val = tensor([3, 4, 12, 64])]; - tensor k_21 = transpose(perm = k_21_perm_0, x = var_1384)[name = tensor("transpose_6")]; - tensor k = reshape(shape = var_1396, x = k_21)[name = tensor("k")]; - tensor var_1398 = const()[name = tensor("op_1398"), val = tensor([3, 4, 12, 64])]; - tensor v_21 = transpose(perm = v_21_perm_0, x = var_1391)[name = tensor("transpose_5")]; - tensor v = reshape(shape = var_1398, x = v_21)[name = tensor("v")]; + tensor var_1417 = const()[name = tensor("op_1417"), val = tensor([3, 4, 12, 64])]; + tensor q_21 = transpose(perm = q_21_perm_0, x = var_1400)[name = tensor("transpose_7")]; + tensor q = reshape(shape = var_1417, x = q_21)[name = tensor("q")]; + tensor var_1419 = const()[name = tensor("op_1419"), val = tensor([3, 4, 12, 64])]; + tensor k_21 = transpose(perm = k_21_perm_0, x = var_1407)[name = tensor("transpose_6")]; + tensor k = reshape(shape = var_1419, x = k_21)[name = tensor("k")]; + tensor var_1421 = const()[name = tensor("op_1421"), val = tensor([3, 4, 12, 64])]; + tensor v_21 = transpose(perm = v_21_perm_0, x = var_1414)[name = tensor("transpose_5")]; + tensor v = reshape(shape = var_1421, x = v_21)[name = tensor("v")]; tensor mul_3_y_0 = const()[name = tensor("mul_3_y_0"), val = tensor(0x1p-3)]; tensor mul_3 = mul(x = q, y = mul_3_y_0)[name = tensor("mul_3")]; tensor matmul_1_transpose_y_0 = const()[name = tensor("matmul_1_transpose_y_0"), val = tensor(true)]; @@ -1237,36 +1248,36 @@ program(1.0) tensor attn_output_9_transpose_x_0 = const()[name = tensor("attn_output_9_transpose_x_0"), val = tensor(false)]; tensor attn_output_9_transpose_y_0 = const()[name = tensor("attn_output_9_transpose_y_0"), val = tensor(false)]; tensor attn_output_9 = matmul(transpose_x = attn_output_9_transpose_x_0, transpose_y = attn_output_9_transpose_y_0, x = softmax_1, y = v)[name = tensor("attn_output_9")]; - tensor var_1401 = const()[name = tensor("op_1401"), val = tensor([2, 0, 1, 3])]; - tensor var_1406 = const()[name = tensor("op_1406"), val = tensor([36, 256])]; - tensor var_1402 = transpose(perm = var_1401, x = attn_output_9)[name = tensor("transpose_4")]; - tensor attn_output_11 = reshape(shape = var_1406, x = var_1402)[name = tensor("attn_output_11")]; - tensor attn_output_13 = linear(bias = decoder_self_attn2_1_out_proj_bias, weight = decoder_self_attn2_1_out_proj_weight, x = attn_output_11)[name = tensor("linear_53")]; - tensor var_1410 = const()[name = tensor("op_1410"), val = tensor([12, 3, 256])]; - tensor attn_output = reshape(shape = var_1410, x = attn_output_13)[name = tensor("attn_output")]; + tensor var_1424 = const()[name = tensor("op_1424"), val = tensor([2, 0, 1, 3])]; + tensor var_1429 = const()[name = tensor("op_1429"), val = tensor([36, 256])]; + tensor var_1425 = transpose(perm = var_1424, x = attn_output_9)[name = tensor("transpose_4")]; + tensor attn_output_11 = reshape(shape = var_1429, x = var_1425)[name = tensor("attn_output_11")]; + tensor attn_output_13 = linear(bias = inner_decoder_self_attn2_1_out_proj_bias, weight = inner_decoder_self_attn2_1_out_proj_weight, x = attn_output_11)[name = tensor("linear_53")]; + tensor var_1433 = const()[name = tensor("op_1433"), val = tensor([12, 3, 256])]; + tensor attn_output = reshape(shape = var_1433, x = attn_output_13)[name = tensor("attn_output")]; tensor sa2_out_perm_0 = const()[name = tensor("sa2_out_perm_0"), val = tensor([1, 0, 2])]; tensor sa2_out = transpose(perm = sa2_out_perm_0, x = attn_output)[name = tensor("transpose_3")]; - tensor input_193 = add(x = query_5, y = sa2_out)[name = tensor("input_193")]; - tensor input_195_axes_0 = const()[name = tensor("input_195_axes_0"), val = tensor([-1])]; - tensor input_195 = layer_norm(axes = input_195_axes_0, beta = decoder_norm21_1_bias, epsilon = var_985, gamma = decoder_norm21_1_weight, x = input_193)[name = tensor("input_195")]; - tensor input_197 = linear(bias = decoder_linear1_1_bias, weight = decoder_linear1_1_weight, x = input_195)[name = tensor("linear_54")]; - tensor input_199 = relu(x = input_197)[name = tensor("input_199")]; - tensor ffn_out = linear(bias = decoder_linear2_1_bias, weight = decoder_linear2_1_weight, x = input_199)[name = tensor("linear_55")]; - tensor input_201 = add(x = input_195, y = ffn_out)[name = tensor("input_201")]; + tensor input_195 = add(x = query_5, y = sa2_out)[name = tensor("input_195")]; + tensor input_197_axes_0 = const()[name = tensor("input_197_axes_0"), val = tensor([-1])]; + tensor input_197 = layer_norm(axes = input_197_axes_0, beta = inner_decoder_norm21_1_bias, epsilon = var_56, gamma = inner_decoder_norm21_1_weight, x = input_195)[name = tensor("input_197")]; + tensor input_199 = linear(bias = inner_decoder_linear1_1_bias, weight = inner_decoder_linear1_1_weight, x = input_197)[name = tensor("linear_54")]; + tensor input_201 = relu(x = input_199)[name = tensor("input_201")]; + tensor ffn_out = linear(bias = inner_decoder_linear2_1_bias, weight = inner_decoder_linear2_1_weight, x = input_201)[name = tensor("linear_55")]; + tensor input_203 = add(x = input_197, y = ffn_out)[name = tensor("input_203")]; tensor xt_axes_0 = const()[name = tensor("xt_axes_0"), val = tensor([-1])]; - tensor xt = layer_norm(axes = xt_axes_0, beta = decoder_norm22_1_bias, epsilon = var_985, gamma = decoder_norm22_1_weight, x = input_201)[name = tensor("xt")]; - tensor var_1430 = const()[name = tensor("op_1430"), val = tensor([1, 3, 12, 256])]; - tensor input = reshape(shape = var_1430, x = xt)[name = tensor("input")]; - tensor var_1432 = const()[name = tensor("op_1432"), val = tensor([-1])]; - tensor var_1433 = reduce_l2_norm(axes = var_1432, keep_dims = var_988, x = input)[name = tensor("op_1433")]; + tensor xt = layer_norm(axes = xt_axes_0, beta = inner_decoder_norm22_1_bias, epsilon = var_56, gamma = inner_decoder_norm22_1_weight, x = input_203)[name = tensor("xt")]; + tensor var_1453 = const()[name = tensor("op_1453"), val = tensor([1, 3, 12, 256])]; + tensor input = reshape(shape = var_1453, x = xt)[name = tensor("input")]; + tensor var_1455 = const()[name = tensor("op_1455"), val = tensor([-1])]; + tensor var_1456 = reduce_l2_norm(axes = var_1455, keep_dims = var_55, x = input)[name = tensor("op_1456")]; tensor const_42 = const()[name = tensor("const_42"), val = tensor(0x1.fffffep+127)]; - tensor clip_5 = clip(alpha = var_980, beta = const_42, x = var_1433)[name = tensor("clip_5")]; - tensor var_1435 = real_div(x = input, y = clip_5)[name = tensor("op_1435")]; + tensor clip_5 = clip(alpha = var_69, beta = const_42, x = var_1456)[name = tensor("clip_5")]; + tensor var_1458 = real_div(x = input, y = clip_5)[name = tensor("op_1458")]; tensor concat_6 = const()[name = tensor("concat_6"), val = tensor([3, 1, 256])]; tensor reshape_0 = reshape(shape = concat_6, x = emb)[name = tensor("reshape_0")]; tensor transpose_1_perm_0 = const()[name = tensor("transpose_1_perm_0"), val = tensor([0, 1, 3, 2])]; tensor concat_7 = const()[name = tensor("concat_7"), val = tensor([3, 256, 12])]; - tensor transpose_1 = transpose(perm = transpose_1_perm_0, x = var_1435)[name = tensor("transpose_2")]; + tensor transpose_1 = transpose(perm = transpose_1_perm_0, x = var_1458)[name = tensor("transpose_2")]; tensor reshape_1 = reshape(shape = concat_7, x = transpose_1)[name = tensor("reshape_1")]; tensor matmul_2_transpose_x_0 = const()[name = tensor("matmul_2_transpose_x_0"), val = tensor(false)]; tensor matmul_2_transpose_y_0 = const()[name = tensor("matmul_2_transpose_y_0"), val = tensor(false)]; @@ -1277,10 +1288,10 @@ program(1.0) tensor output_end_0 = const()[name = tensor("output_end_0"), val = tensor([1, 3, 11])]; tensor output_end_mask_0 = const()[name = tensor("output_end_mask_0"), val = tensor([true, true, false])]; tensor output = slice_by_index(begin = output_begin_0, end = output_end_0, end_mask = output_end_mask_0, x = reshape_2)[name = tensor("output")]; - tensor probs = sigmoid(x = output)[name = tensor("op_1439")]; - tensor var_1441_axis_0 = const()[name = tensor("op_1441_axis_0"), val = tensor(0)]; - tensor dec_kv_new = stack(axis = var_1441_axis_0, values = (var_1137, nkv))[name = tensor("op_1441")]; - tensor var_1443_axis_0 = const()[name = tensor("op_1443_axis_0"), val = tensor(0)]; - tensor dec_scale_new = stack(axis = var_1443_axis_0, values = (new_scale_9, new_scale))[name = tensor("op_1443")]; + tensor probs = sigmoid(x = output)[name = tensor("op_1462")]; + tensor var_1464_axis_0 = const()[name = tensor("op_1464_axis_0"), val = tensor(0)]; + tensor dec_kv_new = stack(axis = var_1464_axis_0, values = (var_1160, nkv))[name = tensor("op_1464")]; + tensor var_1466_axis_0 = const()[name = tensor("op_1466_axis_0"), val = tensor(0)]; + tensor dec_scale_new = stack(axis = var_1466_axis_0, values = (new_scale_9, new_scale))[name = tensor("op_1466")]; } -> (probs, enc_kv_new, enc_scale_new, enc_conv_cache_new, cnn_window_new, dec_kv_new, dec_scale_new); } \ No newline at end of file diff --git a/optimized/dih2/300ms/ls_eend_dih2_300ms.mlpackage/Data/com.apple.CoreML/model.mlmodel b/optimized/dih2/300ms/ls_eend_dih2_300ms.mlpackage/Data/com.apple.CoreML/model.mlmodel index 667d39182df060a6ae205f7f1936b532292aaa00..b82e8993338870af6f757ecc95fccb122327ae99 100644 --- a/optimized/dih2/300ms/ls_eend_dih2_300ms.mlpackage/Data/com.apple.CoreML/model.mlmodel +++ b/optimized/dih2/300ms/ls_eend_dih2_300ms.mlpackage/Data/com.apple.CoreML/model.mlmodel @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:9912c0d401a73181999617e3ae8cf8781300b6effc6ddd777a0d2d293c9dc351 -size 185467 +oid sha256:dce292a0f39072fef89d7e5722dcd50d690c6e70a70a8eab1fe6cf29832302fc +size 191012 diff --git a/optimized/dih2/300ms/ls_eend_dih2_300ms.mlpackage/Manifest.json b/optimized/dih2/300ms/ls_eend_dih2_300ms.mlpackage/Manifest.json index 94cff4c78a7ab58f03319cc612dc624390995764..5cf676d24266c54b61b52b18af6c04e45a45d021 100644 --- a/optimized/dih2/300ms/ls_eend_dih2_300ms.mlpackage/Manifest.json +++ b/optimized/dih2/300ms/ls_eend_dih2_300ms.mlpackage/Manifest.json @@ -1,18 +1,18 @@ { "fileFormatVersion": "1.0.0", "itemInfoEntries": { - "655B25EA-0632-4B1A-B464-F7DFCE1FB184": { + "C166D8CC-202F-4648-A5B5-459F6BCC257A": { "author": "com.apple.CoreML", "description": "CoreML Model Weights", "name": "weights", "path": "com.apple.CoreML/weights" }, - "E19F6AA3-6B6D-4C5E-8868-BE6F38C9CFE1": { + "D324280A-5B20-4ADB-9871-DFBBBF9AFC4C": { "author": "com.apple.CoreML", "description": "CoreML Model Specification", "name": "model.mlmodel", "path": "com.apple.CoreML/model.mlmodel" } }, - "rootModelIdentifier": "E19F6AA3-6B6D-4C5E-8868-BE6F38C9CFE1" + "rootModelIdentifier": "D324280A-5B20-4ADB-9871-DFBBBF9AFC4C" } diff --git a/optimized/dih2/400ms/ls_eend_dih2_400ms.mlmodelc/analytics/coremldata.bin b/optimized/dih2/400ms/ls_eend_dih2_400ms.mlmodelc/analytics/coremldata.bin index 60912e32f95f143258d8e51db95a3edea8749f6c..820140bf91e3fe099ecc6f42e7c99fcdf677d533 100644 --- a/optimized/dih2/400ms/ls_eend_dih2_400ms.mlmodelc/analytics/coremldata.bin +++ b/optimized/dih2/400ms/ls_eend_dih2_400ms.mlmodelc/analytics/coremldata.bin @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:4159dfa0fd403ab65353f56fd6e2f87e53062263e1e136c98c4fcc19b79ffab8 +oid sha256:910909e4d986721254e02d189e9a227e6ea0bd1d273f43d6dd5d416333399b8e size 243 diff --git a/optimized/dih2/400ms/ls_eend_dih2_400ms.mlmodelc/coremldata.bin b/optimized/dih2/400ms/ls_eend_dih2_400ms.mlmodelc/coremldata.bin index 410afba3c04fb8e61c00656e4d9170768e2bbe4a..4af4939a9045bcbd57ae5aca0a2b8ad2cc3dba00 100644 --- a/optimized/dih2/400ms/ls_eend_dih2_400ms.mlmodelc/coremldata.bin +++ b/optimized/dih2/400ms/ls_eend_dih2_400ms.mlmodelc/coremldata.bin @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:a4deedfd08d987353bb9e2e6585c389e10e71fd41bcbe7d5b3140ff4ff9bf985 -size 1308 +oid sha256:02454ef8dbe689c98b17a5dc454f28d1b26bd1c945e5ec74b62390d6d4fca1ae +size 1411 diff --git a/optimized/dih2/400ms/ls_eend_dih2_400ms.mlmodelc/metadata.json b/optimized/dih2/400ms/ls_eend_dih2_400ms.mlmodelc/metadata.json index c179a7eaf3aa7525369b7d58fdb519fbdc635371..607e5330f76f0f52016a51e5bcf2c00979aaa521 100644 --- a/optimized/dih2/400ms/ls_eend_dih2_400ms.mlmodelc/metadata.json +++ b/optimized/dih2/400ms/ls_eend_dih2_400ms.mlmodelc/metadata.json @@ -1,7 +1,7 @@ [ { "metadataOutputVersion" : "3.0", - "shortDescription" : "LS-EEND DIHARD II streaming diarizer (pipeline, T=4, max_speakers=10)", + "shortDescription" : "LS-EEND DIHARD II streaming diarizer (pipeline, T=4, max_speakers=10, layout=raw_mel, cu=all)", "outputSchema" : [ { "hasShapeFlexibility" : "0", @@ -81,7 +81,7 @@ "specificationVersion" : 8, "mlProgramOperationTypeHistogram" : { "Ios16.reduceL2Norm" : 2, - "Ios17.reshape" : 68, + "Ios17.reshape" : 69, "Ios16.softmax" : 2, "Ios17.matmul" : 29, "Ios17.transpose" : 57, @@ -89,7 +89,7 @@ "Ios17.expandDims" : 3, "Ios17.add" : 46, "Ios16.sigmoid" : 5, - "Ios17.sliceByIndex" : 64, + "Ios17.sliceByIndex" : 68, "Tile" : 1, "Ios16.reduceSum" : 1, "Ios17.squeeze" : 2, @@ -101,7 +101,7 @@ "Ios16.silu" : 18, "Ios17.realDiv" : 20, "Ios17.linear" : 56, - "Stack" : 5, + "Stack" : 6, "Ios17.concat" : 22, "Ios16.relu" : 2, "Ios16.cumsum" : 1, @@ -128,9 +128,9 @@ "hasShapeFlexibility" : "0", "isOptional" : "0", "dataType" : "Float32", - "formattedType" : "MultiArray (Float32 1 × 4 × 345)", + "formattedType" : "MultiArray (Float32 1 × 45 × 23)", "shortDescription" : "", - "shape" : "[1, 4, 345]", + "shape" : "[1, 45, 23]", "name" : "features", "type" : "MultiArray" }, @@ -206,8 +206,8 @@ } ], "userDefinedMetadata" : { - "com.github.apple.coremltools.conversion_date" : "2026-04-16", - "config" : "{\"model_name\": \"dih2\", \"model_label\": \"DIHARD II\", \"variant\": \"pipeline\", \"chunk_size\": 4, \"step_duration_ms\": 400, \"frame_duration_ms\": 100, \"frame_duration_seconds\": 0.1, \"max_speakers\": 10, \"max_nspks\": 12, \"n_units\": 256, \"n_heads\": 4, \"enc_n_layers\": 4, \"dec_n_layers\": 2, \"conv_kernel_size\": 16, \"conv_delay\": 9, \"sample_rate\": 8000, \"win_length\": 200, \"hop_length\": 80, \"n_mels\": 23, \"context_size\": 7, \"subsampling\": 10, \"feat_type\": \"logmel23_cummn\", \"pure_roll\": true}", + "com.github.apple.coremltools.conversion_date" : "2026-04-18", + "config" : "{\"model_name\": \"dih2\", \"model_label\": \"DIHARD II\", \"variant\": \"pipeline\", \"chunk_size\": 4, \"step_duration_ms\": 400, \"frame_duration_ms\": 100, \"frame_duration_seconds\": 0.1, \"max_speakers\": 10, \"max_nspks\": 12, \"n_units\": 256, \"n_heads\": 4, \"enc_n_layers\": 4, \"dec_n_layers\": 2, \"conv_kernel_size\": 16, \"conv_delay\": 9, \"sample_rate\": 8000, \"win_length\": 200, \"hop_length\": 80, \"n_mels\": 23, \"context_size\": 7, \"subsampling\": 10, \"feat_type\": \"logmel23_cummn\", \"pure_roll\": true, \"input_layout\": \"raw_mel\", \"compute_units_export\": \"all\", \"raw_mel_length\": 45}", "com.github.apple.coremltools.source" : "torch==2.6.0", "com.github.apple.coremltools.version" : "9.0", "com.github.apple.coremltools.source_dialect" : "TorchScript" diff --git a/optimized/dih2/400ms/ls_eend_dih2_400ms.mlmodelc/model.mil b/optimized/dih2/400ms/ls_eend_dih2_400ms.mlmodelc/model.mil index 7135f6328b831d1038e863b851b7563d75bba5c3..546c200804eb017fad873f0d353fbf1c38982ea7 100644 --- a/optimized/dih2/400ms/ls_eend_dih2_400ms.mlmodelc/model.mil +++ b/optimized/dih2/400ms/ls_eend_dih2_400ms.mlmodelc/model.mil @@ -1,234 +1,256 @@ program(1.0) [buildInfo = dict, tensor>({{"coremlc-component-MIL", "3510.2.1"}, {"coremlc-version", "3500.32.1"}, {"coremltools-component-torch", "2.6.0"}, {"coremltools-source-dialect", "TorchScript"}, {"coremltools-version", "9.0"}})] { - func main(tensor cnn_window, tensor dec_kv, tensor dec_scale, tensor enc_conv_cache, tensor enc_kv, tensor enc_scale, tensor features, tensor valid_mask) { - tensor encoder_cnn_bias = const()[name = tensor("encoder_cnn_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(64)))]; - tensor encoder_cnn_weight = const()[name = tensor("encoder_cnn_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1152)))]; - tensor encoder__causal_mask = const()[name = tensor("encoder__causal_mask"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4981952)))]; - tensor encoder__t_index = const()[name = tensor("encoder__t_index"), val = tensor([0x1p+0, 0x1p+1, 0x1.8p+1, 0x1p+2])]; - tensor encoder_input_projection_linear_bias = const()[name = tensor("encoder_input_projection_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4982080)))]; - tensor encoder_input_projection_linear_weight = const()[name = tensor("encoder_input_projection_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4983168)))]; - tensor encoder_pre_layer_norm_bias = const()[name = tensor("encoder_pre_layer_norm_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(5336512)))]; - tensor encoder_pre_layer_norm_weight = const()[name = tensor("encoder_pre_layer_norm_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(5337600)))]; - tensor encoder_ffn1_0_module_sequential_0_bias = const()[name = tensor("encoder_ffn1_0_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(5338688)))]; - tensor encoder_ffn1_0_module_sequential_0_weight = const()[name = tensor("encoder_ffn1_0_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(5339776)))]; - tensor encoder_ffn1_0_module_sequential_1_linear_bias = const()[name = tensor("encoder_ffn1_0_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(5340864)))]; - tensor encoder_ffn1_0_module_sequential_1_linear_weight = const()[name = tensor("encoder_ffn1_0_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(5345024)))]; - tensor encoder_ffn1_0_module_sequential_4_linear_bias = const()[name = tensor("encoder_ffn1_0_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(6393664)))]; - tensor encoder_ffn1_0_module_sequential_4_linear_weight = const()[name = tensor("encoder_ffn1_0_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(6394752)))]; - tensor encoder_ret_lns_0_bias = const()[name = tensor("encoder_ret_lns_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7443392)))]; - tensor encoder_ret_lns_0_weight = const()[name = tensor("encoder_ret_lns_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7444480)))]; - tensor encoder_q_proj_0_bias = const()[name = tensor("encoder_q_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7445568)))]; - tensor encoder_q_proj_0_weight = const()[name = tensor("encoder_q_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7446656)))]; - tensor encoder_k_proj_0_bias = const()[name = tensor("encoder_k_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7708864)))]; - tensor encoder_k_proj_0_weight = const()[name = tensor("encoder_k_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7709952)))]; - tensor encoder_v_proj_0_bias = const()[name = tensor("encoder_v_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7972160)))]; - tensor encoder_v_proj_0_weight = const()[name = tensor("encoder_v_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7973248)))]; - tensor encoder_g_proj_0_bias = const()[name = tensor("encoder_g_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8235456)))]; - tensor encoder_g_proj_0_weight = const()[name = tensor("encoder_g_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8236544)))]; - tensor encoder_out_proj_0_bias = const()[name = tensor("encoder_out_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8498752)))]; - tensor encoder_out_proj_0_weight = const()[name = tensor("encoder_out_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8499840)))]; - tensor encoder_conv_module_0_sequential_0_bias = const()[name = tensor("encoder_conv_module_0_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8762048)))]; - tensor encoder_conv_module_0_sequential_0_weight = const()[name = tensor("encoder_conv_module_0_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8763136)))]; - tensor encoder_conv_module_0_sequential_2_conv_bias = const()[name = tensor("encoder_conv_module_0_sequential_2_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8764224)))]; - tensor encoder_conv_module_0_sequential_2_conv_weight = const()[name = tensor("encoder_conv_module_0_sequential_2_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8766336)))]; - tensor encoder_conv_module_0_sequential_4_conv_weight = const()[name = tensor("encoder_conv_module_0_sequential_4_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9290688)))]; - tensor encoder_conv_module_0_sequential_5_running_var = const()[name = tensor("encoder_conv_module_0_sequential_5_running_var"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9307136)))]; - tensor encoder_conv_module_0_sequential_5_running_mean = const()[name = tensor("encoder_conv_module_0_sequential_5_running_mean"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9308224)))]; - tensor encoder_conv_module_0_sequential_5_bias = const()[name = tensor("encoder_conv_module_0_sequential_5_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9309312)))]; - tensor encoder_conv_module_0_sequential_5_weight = const()[name = tensor("encoder_conv_module_0_sequential_5_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9310400)))]; - tensor encoder_conv_module_0_sequential_7_conv_bias = const()[name = tensor("encoder_conv_module_0_sequential_7_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9311488)))]; - tensor encoder_conv_module_0_sequential_7_conv_weight = const()[name = tensor("encoder_conv_module_0_sequential_7_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9312576)))]; - tensor encoder_ffn2_0_module_sequential_0_bias = const()[name = tensor("encoder_ffn2_0_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9574784)))]; - tensor encoder_ffn2_0_module_sequential_0_weight = const()[name = tensor("encoder_ffn2_0_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9575872)))]; - tensor encoder_ffn2_0_module_sequential_1_linear_bias = const()[name = tensor("encoder_ffn2_0_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9576960)))]; - tensor encoder_ffn2_0_module_sequential_1_linear_weight = const()[name = tensor("encoder_ffn2_0_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9581120)))]; - tensor encoder_ffn2_0_module_sequential_4_linear_bias = const()[name = tensor("encoder_ffn2_0_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(10629760)))]; - tensor encoder_ffn2_0_module_sequential_4_linear_weight = const()[name = tensor("encoder_ffn2_0_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(10630848)))]; - tensor encoder_layer_norm_0_bias = const()[name = tensor("encoder_layer_norm_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11679488)))]; - tensor encoder_layer_norm_0_weight = const()[name = tensor("encoder_layer_norm_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11680576)))]; - tensor encoder_ffn1_1_module_sequential_0_bias = const()[name = tensor("encoder_ffn1_1_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11681664)))]; - tensor encoder_ffn1_1_module_sequential_0_weight = const()[name = tensor("encoder_ffn1_1_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11682752)))]; - tensor encoder_ffn1_1_module_sequential_1_linear_bias = const()[name = tensor("encoder_ffn1_1_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11683840)))]; - tensor encoder_ffn1_1_module_sequential_1_linear_weight = const()[name = tensor("encoder_ffn1_1_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11688000)))]; - tensor encoder_ffn1_1_module_sequential_4_linear_bias = const()[name = tensor("encoder_ffn1_1_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(12736640)))]; - tensor encoder_ffn1_1_module_sequential_4_linear_weight = const()[name = tensor("encoder_ffn1_1_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(12737728)))]; - tensor encoder_ret_lns_1_bias = const()[name = tensor("encoder_ret_lns_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(13786368)))]; - tensor encoder_ret_lns_1_weight = const()[name = tensor("encoder_ret_lns_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(13787456)))]; - tensor encoder_q_proj_1_bias = const()[name = tensor("encoder_q_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(13788544)))]; - tensor encoder_q_proj_1_weight = const()[name = tensor("encoder_q_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(13789632)))]; - tensor encoder_k_proj_1_bias = const()[name = tensor("encoder_k_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14051840)))]; - tensor encoder_k_proj_1_weight = const()[name = tensor("encoder_k_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14052928)))]; - tensor encoder_v_proj_1_bias = const()[name = tensor("encoder_v_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14315136)))]; - tensor encoder_v_proj_1_weight = const()[name = tensor("encoder_v_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14316224)))]; - tensor encoder_g_proj_1_bias = const()[name = tensor("encoder_g_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14578432)))]; - tensor encoder_g_proj_1_weight = const()[name = tensor("encoder_g_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14579520)))]; - tensor encoder_out_proj_1_bias = const()[name = tensor("encoder_out_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14841728)))]; - tensor encoder_out_proj_1_weight = const()[name = tensor("encoder_out_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14842816)))]; - tensor encoder_conv_module_1_sequential_0_bias = const()[name = tensor("encoder_conv_module_1_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15105024)))]; - tensor encoder_conv_module_1_sequential_0_weight = const()[name = tensor("encoder_conv_module_1_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15106112)))]; - tensor encoder_conv_module_1_sequential_2_conv_bias = const()[name = tensor("encoder_conv_module_1_sequential_2_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15107200)))]; - tensor encoder_conv_module_1_sequential_2_conv_weight = const()[name = tensor("encoder_conv_module_1_sequential_2_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15109312)))]; - tensor encoder_conv_module_1_sequential_4_conv_weight = const()[name = tensor("encoder_conv_module_1_sequential_4_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15633664)))]; - tensor encoder_conv_module_1_sequential_5_running_var = const()[name = tensor("encoder_conv_module_1_sequential_5_running_var"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15650112)))]; - tensor encoder_conv_module_1_sequential_5_running_mean = const()[name = tensor("encoder_conv_module_1_sequential_5_running_mean"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15651200)))]; - tensor encoder_conv_module_1_sequential_5_bias = const()[name = tensor("encoder_conv_module_1_sequential_5_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15652288)))]; - tensor encoder_conv_module_1_sequential_5_weight = const()[name = tensor("encoder_conv_module_1_sequential_5_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15653376)))]; - tensor encoder_conv_module_1_sequential_7_conv_bias = const()[name = tensor("encoder_conv_module_1_sequential_7_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15654464)))]; - tensor encoder_conv_module_1_sequential_7_conv_weight = const()[name = tensor("encoder_conv_module_1_sequential_7_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15655552)))]; - tensor encoder_ffn2_1_module_sequential_0_bias = const()[name = tensor("encoder_ffn2_1_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15917760)))]; - tensor encoder_ffn2_1_module_sequential_0_weight = const()[name = tensor("encoder_ffn2_1_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15918848)))]; - tensor encoder_ffn2_1_module_sequential_1_linear_bias = const()[name = tensor("encoder_ffn2_1_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15919936)))]; - tensor encoder_ffn2_1_module_sequential_1_linear_weight = const()[name = tensor("encoder_ffn2_1_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15924096)))]; - tensor encoder_ffn2_1_module_sequential_4_linear_bias = const()[name = tensor("encoder_ffn2_1_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(16972736)))]; - tensor encoder_ffn2_1_module_sequential_4_linear_weight = const()[name = tensor("encoder_ffn2_1_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(16973824)))]; - tensor encoder_layer_norm_1_bias = const()[name = tensor("encoder_layer_norm_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18022464)))]; - tensor encoder_layer_norm_1_weight = const()[name = tensor("encoder_layer_norm_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18023552)))]; - tensor encoder_ffn1_2_module_sequential_0_bias = const()[name = tensor("encoder_ffn1_2_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18024640)))]; - tensor encoder_ffn1_2_module_sequential_0_weight = const()[name = tensor("encoder_ffn1_2_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18025728)))]; - tensor encoder_ffn1_2_module_sequential_1_linear_bias = const()[name = tensor("encoder_ffn1_2_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18026816)))]; - tensor encoder_ffn1_2_module_sequential_1_linear_weight = const()[name = tensor("encoder_ffn1_2_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18030976)))]; - tensor encoder_ffn1_2_module_sequential_4_linear_bias = const()[name = tensor("encoder_ffn1_2_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(19079616)))]; - tensor encoder_ffn1_2_module_sequential_4_linear_weight = const()[name = tensor("encoder_ffn1_2_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(19080704)))]; - tensor encoder_ret_lns_2_bias = const()[name = tensor("encoder_ret_lns_2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20129344)))]; - tensor encoder_ret_lns_2_weight = const()[name = tensor("encoder_ret_lns_2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20130432)))]; - tensor encoder_q_proj_2_bias = const()[name = tensor("encoder_q_proj_2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20131520)))]; - tensor encoder_q_proj_2_weight = const()[name = tensor("encoder_q_proj_2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20132608)))]; - tensor encoder_k_proj_2_bias = const()[name = tensor("encoder_k_proj_2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20394816)))]; - tensor encoder_k_proj_2_weight = const()[name = tensor("encoder_k_proj_2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20395904)))]; - tensor encoder_v_proj_2_bias = const()[name = tensor("encoder_v_proj_2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20658112)))]; - tensor encoder_v_proj_2_weight = const()[name = tensor("encoder_v_proj_2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20659200)))]; - tensor encoder_g_proj_2_bias = const()[name = tensor("encoder_g_proj_2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20921408)))]; - tensor encoder_g_proj_2_weight = const()[name = tensor("encoder_g_proj_2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20922496)))]; - tensor encoder_out_proj_2_bias = const()[name = tensor("encoder_out_proj_2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21184704)))]; - tensor encoder_out_proj_2_weight = const()[name = tensor("encoder_out_proj_2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21185792)))]; - tensor encoder_conv_module_2_sequential_0_bias = const()[name = tensor("encoder_conv_module_2_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21448000)))]; - tensor encoder_conv_module_2_sequential_0_weight = const()[name = tensor("encoder_conv_module_2_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21449088)))]; - tensor encoder_conv_module_2_sequential_2_conv_bias = const()[name = tensor("encoder_conv_module_2_sequential_2_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21450176)))]; - tensor encoder_conv_module_2_sequential_2_conv_weight = const()[name = tensor("encoder_conv_module_2_sequential_2_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21452288)))]; - tensor encoder_conv_module_2_sequential_4_conv_weight = const()[name = tensor("encoder_conv_module_2_sequential_4_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21976640)))]; - tensor encoder_conv_module_2_sequential_5_running_var = const()[name = tensor("encoder_conv_module_2_sequential_5_running_var"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21993088)))]; - tensor encoder_conv_module_2_sequential_5_running_mean = const()[name = tensor("encoder_conv_module_2_sequential_5_running_mean"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21994176)))]; - tensor encoder_conv_module_2_sequential_5_bias = const()[name = tensor("encoder_conv_module_2_sequential_5_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21995264)))]; - tensor encoder_conv_module_2_sequential_5_weight = const()[name = tensor("encoder_conv_module_2_sequential_5_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21996352)))]; - tensor encoder_conv_module_2_sequential_7_conv_bias = const()[name = tensor("encoder_conv_module_2_sequential_7_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21997440)))]; - tensor encoder_conv_module_2_sequential_7_conv_weight = const()[name = tensor("encoder_conv_module_2_sequential_7_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21998528)))]; - tensor encoder_ffn2_2_module_sequential_0_bias = const()[name = tensor("encoder_ffn2_2_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(22260736)))]; - tensor encoder_ffn2_2_module_sequential_0_weight = const()[name = tensor("encoder_ffn2_2_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(22261824)))]; - tensor encoder_ffn2_2_module_sequential_1_linear_bias = const()[name = tensor("encoder_ffn2_2_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(22262912)))]; - tensor encoder_ffn2_2_module_sequential_1_linear_weight = const()[name = tensor("encoder_ffn2_2_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(22267072)))]; - tensor encoder_ffn2_2_module_sequential_4_linear_bias = const()[name = tensor("encoder_ffn2_2_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(23315712)))]; - tensor encoder_ffn2_2_module_sequential_4_linear_weight = const()[name = tensor("encoder_ffn2_2_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(23316800)))]; - tensor encoder_layer_norm_2_bias = const()[name = tensor("encoder_layer_norm_2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(24365440)))]; - tensor encoder_layer_norm_2_weight = const()[name = tensor("encoder_layer_norm_2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(24366528)))]; - tensor encoder_ffn1_3_module_sequential_0_bias = const()[name = tensor("encoder_ffn1_3_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(24367616)))]; - tensor encoder_ffn1_3_module_sequential_0_weight = const()[name = tensor("encoder_ffn1_3_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(24368704)))]; - tensor encoder_ffn1_3_module_sequential_1_linear_bias = const()[name = tensor("encoder_ffn1_3_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(24369792)))]; - tensor encoder_ffn1_3_module_sequential_1_linear_weight = const()[name = tensor("encoder_ffn1_3_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(24373952)))]; - tensor encoder_ffn1_3_module_sequential_4_linear_bias = const()[name = tensor("encoder_ffn1_3_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(25422592)))]; - tensor encoder_ffn1_3_module_sequential_4_linear_weight = const()[name = tensor("encoder_ffn1_3_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(25423680)))]; - tensor encoder_ret_lns_3_bias = const()[name = tensor("encoder_ret_lns_3_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(26472320)))]; - tensor encoder_ret_lns_3_weight = const()[name = tensor("encoder_ret_lns_3_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(26473408)))]; - tensor encoder_q_proj_3_bias = const()[name = tensor("encoder_q_proj_3_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(26474496)))]; - tensor encoder_q_proj_3_weight = const()[name = tensor("encoder_q_proj_3_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(26475584)))]; - tensor encoder_k_proj_3_bias = const()[name = tensor("encoder_k_proj_3_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(26737792)))]; - tensor encoder_k_proj_3_weight = const()[name = tensor("encoder_k_proj_3_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(26738880)))]; - tensor encoder_v_proj_3_bias = const()[name = tensor("encoder_v_proj_3_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27001088)))]; - tensor encoder_v_proj_3_weight = const()[name = tensor("encoder_v_proj_3_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27002176)))]; - tensor encoder_g_proj_3_bias = const()[name = tensor("encoder_g_proj_3_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27264384)))]; - tensor encoder_g_proj_3_weight = const()[name = tensor("encoder_g_proj_3_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27265472)))]; - tensor encoder_out_proj_3_bias = const()[name = tensor("encoder_out_proj_3_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27527680)))]; - tensor encoder_out_proj_3_weight = const()[name = tensor("encoder_out_proj_3_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27528768)))]; - tensor encoder_conv_module_3_sequential_0_bias = const()[name = tensor("encoder_conv_module_3_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27790976)))]; - tensor encoder_conv_module_3_sequential_0_weight = const()[name = tensor("encoder_conv_module_3_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27792064)))]; - tensor encoder_conv_module_3_sequential_2_conv_bias = const()[name = tensor("encoder_conv_module_3_sequential_2_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27793152)))]; - tensor encoder_conv_module_3_sequential_2_conv_weight = const()[name = tensor("encoder_conv_module_3_sequential_2_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27795264)))]; - tensor encoder_conv_module_3_sequential_4_conv_weight = const()[name = tensor("encoder_conv_module_3_sequential_4_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28319616)))]; - tensor encoder_conv_module_3_sequential_5_running_var = const()[name = tensor("encoder_conv_module_3_sequential_5_running_var"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28336064)))]; - tensor encoder_conv_module_3_sequential_5_running_mean = const()[name = tensor("encoder_conv_module_3_sequential_5_running_mean"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28337152)))]; - tensor encoder_conv_module_3_sequential_5_bias = const()[name = tensor("encoder_conv_module_3_sequential_5_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28338240)))]; - tensor encoder_conv_module_3_sequential_5_weight = const()[name = tensor("encoder_conv_module_3_sequential_5_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28339328)))]; - tensor encoder_conv_module_3_sequential_7_conv_bias = const()[name = tensor("encoder_conv_module_3_sequential_7_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28340416)))]; - tensor encoder_conv_module_3_sequential_7_conv_weight = const()[name = tensor("encoder_conv_module_3_sequential_7_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28341504)))]; - tensor encoder_ffn2_3_module_sequential_0_bias = const()[name = tensor("encoder_ffn2_3_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28603712)))]; - tensor encoder_ffn2_3_module_sequential_0_weight = const()[name = tensor("encoder_ffn2_3_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28604800)))]; - tensor encoder_ffn2_3_module_sequential_1_linear_bias = const()[name = tensor("encoder_ffn2_3_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28605888)))]; - tensor encoder_ffn2_3_module_sequential_1_linear_weight = const()[name = tensor("encoder_ffn2_3_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28610048)))]; - tensor encoder_ffn2_3_module_sequential_4_linear_bias = const()[name = tensor("encoder_ffn2_3_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(29658688)))]; - tensor encoder_ffn2_3_module_sequential_4_linear_weight = const()[name = tensor("encoder_ffn2_3_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(29659776)))]; - tensor encoder_layer_norm_3_bias = const()[name = tensor("encoder_layer_norm_3_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(30708416)))]; - tensor encoder_layer_norm_3_weight = const()[name = tensor("encoder_layer_norm_3_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(30709504)))]; - tensor decoder__causal_mask = const()[name = tensor("decoder__causal_mask"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(30710592)))]; - tensor decoder_convert_bias = const()[name = tensor("decoder_convert_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(30710720)))]; - tensor decoder_convert_weight = const()[name = tensor("decoder_convert_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(30711808)))]; - tensor decoder_q_proj_0_bias = const()[name = tensor("decoder_q_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(31236160)))]; - tensor decoder_q_proj_0_weight = const()[name = tensor("decoder_q_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(31237248)))]; - tensor decoder_k_proj_0_bias = const()[name = tensor("decoder_k_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(31499456)))]; - tensor decoder_k_proj_0_weight = const()[name = tensor("decoder_k_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(31500544)))]; - tensor decoder_v_proj_0_bias = const()[name = tensor("decoder_v_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(31762752)))]; - tensor decoder_v_proj_0_weight = const()[name = tensor("decoder_v_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(31763840)))]; - tensor decoder_g_proj_0_bias = const()[name = tensor("decoder_g_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32026048)))]; - tensor decoder_g_proj_0_weight = const()[name = tensor("decoder_g_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32027136)))]; - tensor decoder_out_proj_0_bias = const()[name = tensor("decoder_out_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32289344)))]; - tensor decoder_out_proj_0_weight = const()[name = tensor("decoder_out_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32290432)))]; - tensor decoder_norm11_0_bias = const()[name = tensor("decoder_norm11_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32552640)))]; - tensor decoder_norm11_0_weight = const()[name = tensor("decoder_norm11_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32553728)))]; - tensor decoder_self_attn2_0_out_proj_bias = const()[name = tensor("decoder_self_attn2_0_out_proj_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32554816)))]; - tensor decoder_self_attn2_0_out_proj_weight = const()[name = tensor("decoder_self_attn2_0_out_proj_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32555904)))]; - tensor decoder_self_attn2_0_in_proj_bias = const()[name = tensor("decoder_self_attn2_0_in_proj_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32818112)))]; - tensor decoder_self_attn2_0_in_proj_weight = const()[name = tensor("decoder_self_attn2_0_in_proj_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32821248)))]; - tensor decoder_norm21_0_bias = const()[name = tensor("decoder_norm21_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(33607744)))]; - tensor decoder_norm21_0_weight = const()[name = tensor("decoder_norm21_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(33608832)))]; - tensor decoder_linear1_0_bias = const()[name = tensor("decoder_linear1_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(33609920)))]; - tensor decoder_linear1_0_weight = const()[name = tensor("decoder_linear1_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(33618176)))]; - tensor decoder_linear2_0_bias = const()[name = tensor("decoder_linear2_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(35715392)))]; - tensor decoder_linear2_0_weight = const()[name = tensor("decoder_linear2_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(35716480)))]; - tensor decoder_norm22_0_bias = const()[name = tensor("decoder_norm22_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(37813696)))]; - tensor decoder_norm22_0_weight = const()[name = tensor("decoder_norm22_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(37814784)))]; - tensor decoder_q_proj_1_bias = const()[name = tensor("decoder_q_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(37815872)))]; - tensor decoder_q_proj_1_weight = const()[name = tensor("decoder_q_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(37816960)))]; - tensor decoder_k_proj_1_bias = const()[name = tensor("decoder_k_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38079168)))]; - tensor decoder_k_proj_1_weight = const()[name = tensor("decoder_k_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38080256)))]; - tensor decoder_v_proj_1_bias = const()[name = tensor("decoder_v_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38342464)))]; - tensor decoder_v_proj_1_weight = const()[name = tensor("decoder_v_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38343552)))]; - tensor decoder_g_proj_1_bias = const()[name = tensor("decoder_g_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38605760)))]; - tensor decoder_g_proj_1_weight = const()[name = tensor("decoder_g_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38606848)))]; - tensor decoder_out_proj_1_bias = const()[name = tensor("decoder_out_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38869056)))]; - tensor decoder_out_proj_1_weight = const()[name = tensor("decoder_out_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38870144)))]; - tensor decoder_norm11_1_bias = const()[name = tensor("decoder_norm11_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(39132352)))]; - tensor decoder_norm11_1_weight = const()[name = tensor("decoder_norm11_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(39133440)))]; - tensor decoder_self_attn2_1_out_proj_bias = const()[name = tensor("decoder_self_attn2_1_out_proj_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(39134528)))]; - tensor decoder_self_attn2_1_out_proj_weight = const()[name = tensor("decoder_self_attn2_1_out_proj_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(39135616)))]; - tensor decoder_self_attn2_1_in_proj_bias = const()[name = tensor("decoder_self_attn2_1_in_proj_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(39397824)))]; - tensor decoder_self_attn2_1_in_proj_weight = const()[name = tensor("decoder_self_attn2_1_in_proj_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(39400960)))]; - tensor decoder_norm21_1_bias = const()[name = tensor("decoder_norm21_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(40187456)))]; - tensor decoder_norm21_1_weight = const()[name = tensor("decoder_norm21_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(40188544)))]; - tensor decoder_linear1_1_bias = const()[name = tensor("decoder_linear1_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(40189632)))]; - tensor decoder_linear1_1_weight = const()[name = tensor("decoder_linear1_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(40197888)))]; - tensor decoder_linear2_1_bias = const()[name = tensor("decoder_linear2_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(42295104)))]; - tensor decoder_linear2_1_weight = const()[name = tensor("decoder_linear2_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(42296192)))]; - tensor decoder_norm22_1_bias = const()[name = tensor("decoder_norm22_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(44393408)))]; - tensor decoder_norm22_1_weight = const()[name = tensor("decoder_norm22_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(44394496)))]; - tensor var_11 = const()[name = tensor("op_11"), val = tensor(0x1.5798eep-27)]; - tensor var_17 = const()[name = tensor("op_17"), val = tensor(0x1.0c6f7ap-20)]; - tensor var_20 = const()[name = tensor("op_20"), val = tensor(2)]; - tensor var_23 = const()[name = tensor("op_23"), val = tensor(0)]; - tensor var_26 = const()[name = tensor("op_26"), val = tensor(1)]; - tensor var_28 = const()[name = tensor("op_28"), val = tensor(0x1.4f8b58p-17)]; - tensor var_29 = const()[name = tensor("op_29"), val = tensor(true)]; - tensor input_1 = linear(bias = encoder_input_projection_linear_bias, weight = encoder_input_projection_linear_weight, x = features)[name = tensor("linear_0")]; - tensor input_3_axes_0 = const()[name = tensor("input_3_axes_0"), val = tensor([-1])]; - tensor input_3 = layer_norm(axes = input_3_axes_0, beta = encoder_pre_layer_norm_bias, epsilon = var_28, gamma = encoder_pre_layer_norm_weight, x = input_1)[name = tensor("input_3")]; + func main(tensor cnn_window, tensor dec_kv, tensor dec_scale, tensor enc_conv_cache, tensor enc_kv, tensor enc_scale, tensor features, tensor valid_mask) { + tensor inner_encoder_cnn_bias = const()[name = tensor("inner_encoder_cnn_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(64)))]; + tensor inner_encoder_cnn_weight = const()[name = tensor("inner_encoder_cnn_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1152)))]; + tensor inner_encoder__causal_mask = const()[name = tensor("inner_encoder__causal_mask"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4981952)))]; + tensor inner_encoder__t_index = const()[name = tensor("inner_encoder__t_index"), val = tensor([0x1p+0, 0x1p+1, 0x1.8p+1, 0x1p+2])]; + tensor inner_encoder_input_projection_linear_bias = const()[name = tensor("inner_encoder_input_projection_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4982080)))]; + tensor inner_encoder_input_projection_linear_weight = const()[name = tensor("inner_encoder_input_projection_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4983168)))]; + tensor inner_encoder_pre_layer_norm_bias = const()[name = tensor("inner_encoder_pre_layer_norm_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(5336512)))]; + tensor inner_encoder_pre_layer_norm_weight = const()[name = tensor("inner_encoder_pre_layer_norm_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(5337600)))]; + tensor inner_encoder_ffn1_0_module_sequential_0_bias = const()[name = tensor("inner_encoder_ffn1_0_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(5338688)))]; + tensor inner_encoder_ffn1_0_module_sequential_0_weight = const()[name = tensor("inner_encoder_ffn1_0_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(5339776)))]; + tensor inner_encoder_ffn1_0_module_sequential_1_linear_bias = const()[name = tensor("inner_encoder_ffn1_0_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(5340864)))]; + tensor inner_encoder_ffn1_0_module_sequential_1_linear_weight = const()[name = tensor("inner_encoder_ffn1_0_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(5345024)))]; + tensor inner_encoder_ffn1_0_module_sequential_4_linear_bias = const()[name = tensor("inner_encoder_ffn1_0_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(6393664)))]; + tensor inner_encoder_ffn1_0_module_sequential_4_linear_weight = const()[name = tensor("inner_encoder_ffn1_0_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(6394752)))]; + tensor inner_encoder_ret_lns_0_bias = const()[name = tensor("inner_encoder_ret_lns_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7443392)))]; + tensor inner_encoder_ret_lns_0_weight = const()[name = tensor("inner_encoder_ret_lns_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7444480)))]; + tensor inner_encoder_q_proj_0_bias = const()[name = tensor("inner_encoder_q_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7445568)))]; + tensor inner_encoder_q_proj_0_weight = const()[name = tensor("inner_encoder_q_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7446656)))]; + tensor inner_encoder_k_proj_0_bias = const()[name = tensor("inner_encoder_k_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7708864)))]; + tensor inner_encoder_k_proj_0_weight = const()[name = tensor("inner_encoder_k_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7709952)))]; + tensor inner_encoder_v_proj_0_bias = const()[name = tensor("inner_encoder_v_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7972160)))]; + tensor inner_encoder_v_proj_0_weight = const()[name = tensor("inner_encoder_v_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7973248)))]; + tensor inner_encoder_g_proj_0_bias = const()[name = tensor("inner_encoder_g_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8235456)))]; + tensor inner_encoder_g_proj_0_weight = const()[name = tensor("inner_encoder_g_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8236544)))]; + tensor inner_encoder_out_proj_0_bias = const()[name = tensor("inner_encoder_out_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8498752)))]; + tensor inner_encoder_out_proj_0_weight = const()[name = tensor("inner_encoder_out_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8499840)))]; + tensor inner_encoder_conv_module_0_sequential_0_bias = const()[name = tensor("inner_encoder_conv_module_0_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8762048)))]; + tensor inner_encoder_conv_module_0_sequential_0_weight = const()[name = tensor("inner_encoder_conv_module_0_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8763136)))]; + tensor inner_encoder_conv_module_0_sequential_2_conv_bias = const()[name = tensor("inner_encoder_conv_module_0_sequential_2_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8764224)))]; + tensor inner_encoder_conv_module_0_sequential_2_conv_weight = const()[name = tensor("inner_encoder_conv_module_0_sequential_2_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8766336)))]; + tensor inner_encoder_conv_module_0_sequential_4_conv_weight = const()[name = tensor("inner_encoder_conv_module_0_sequential_4_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9290688)))]; + tensor inner_encoder_conv_module_0_sequential_5_running_var = const()[name = tensor("inner_encoder_conv_module_0_sequential_5_running_var"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9307136)))]; + tensor inner_encoder_conv_module_0_sequential_5_running_mean = const()[name = tensor("inner_encoder_conv_module_0_sequential_5_running_mean"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9308224)))]; + tensor inner_encoder_conv_module_0_sequential_5_bias = const()[name = tensor("inner_encoder_conv_module_0_sequential_5_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9309312)))]; + tensor inner_encoder_conv_module_0_sequential_5_weight = const()[name = tensor("inner_encoder_conv_module_0_sequential_5_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9310400)))]; + tensor inner_encoder_conv_module_0_sequential_7_conv_bias = const()[name = tensor("inner_encoder_conv_module_0_sequential_7_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9311488)))]; + tensor inner_encoder_conv_module_0_sequential_7_conv_weight = const()[name = tensor("inner_encoder_conv_module_0_sequential_7_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9312576)))]; + tensor inner_encoder_ffn2_0_module_sequential_0_bias = const()[name = tensor("inner_encoder_ffn2_0_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9574784)))]; + tensor inner_encoder_ffn2_0_module_sequential_0_weight = const()[name = tensor("inner_encoder_ffn2_0_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9575872)))]; + tensor inner_encoder_ffn2_0_module_sequential_1_linear_bias = const()[name = tensor("inner_encoder_ffn2_0_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9576960)))]; + tensor inner_encoder_ffn2_0_module_sequential_1_linear_weight = const()[name = tensor("inner_encoder_ffn2_0_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9581120)))]; + tensor inner_encoder_ffn2_0_module_sequential_4_linear_bias = const()[name = tensor("inner_encoder_ffn2_0_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(10629760)))]; + tensor inner_encoder_ffn2_0_module_sequential_4_linear_weight = const()[name = tensor("inner_encoder_ffn2_0_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(10630848)))]; + tensor inner_encoder_layer_norm_0_bias = const()[name = tensor("inner_encoder_layer_norm_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11679488)))]; + tensor inner_encoder_layer_norm_0_weight = const()[name = tensor("inner_encoder_layer_norm_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11680576)))]; + tensor inner_encoder_ffn1_1_module_sequential_0_bias = const()[name = tensor("inner_encoder_ffn1_1_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11681664)))]; + tensor inner_encoder_ffn1_1_module_sequential_0_weight = const()[name = tensor("inner_encoder_ffn1_1_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11682752)))]; + tensor inner_encoder_ffn1_1_module_sequential_1_linear_bias = const()[name = tensor("inner_encoder_ffn1_1_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11683840)))]; + tensor inner_encoder_ffn1_1_module_sequential_1_linear_weight = const()[name = tensor("inner_encoder_ffn1_1_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11688000)))]; + tensor inner_encoder_ffn1_1_module_sequential_4_linear_bias = const()[name = tensor("inner_encoder_ffn1_1_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(12736640)))]; + tensor inner_encoder_ffn1_1_module_sequential_4_linear_weight = const()[name = tensor("inner_encoder_ffn1_1_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(12737728)))]; + tensor inner_encoder_ret_lns_1_bias = const()[name = tensor("inner_encoder_ret_lns_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(13786368)))]; + tensor inner_encoder_ret_lns_1_weight = const()[name = tensor("inner_encoder_ret_lns_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(13787456)))]; + tensor inner_encoder_q_proj_1_bias = const()[name = tensor("inner_encoder_q_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(13788544)))]; + tensor inner_encoder_q_proj_1_weight = const()[name = tensor("inner_encoder_q_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(13789632)))]; + tensor inner_encoder_k_proj_1_bias = const()[name = tensor("inner_encoder_k_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14051840)))]; + tensor inner_encoder_k_proj_1_weight = const()[name = tensor("inner_encoder_k_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14052928)))]; + tensor inner_encoder_v_proj_1_bias = const()[name = tensor("inner_encoder_v_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14315136)))]; + tensor inner_encoder_v_proj_1_weight = const()[name = tensor("inner_encoder_v_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14316224)))]; + tensor inner_encoder_g_proj_1_bias = const()[name = tensor("inner_encoder_g_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14578432)))]; + tensor inner_encoder_g_proj_1_weight = const()[name = tensor("inner_encoder_g_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14579520)))]; + tensor inner_encoder_out_proj_1_bias = const()[name = tensor("inner_encoder_out_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14841728)))]; + tensor inner_encoder_out_proj_1_weight = const()[name = tensor("inner_encoder_out_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14842816)))]; + tensor inner_encoder_conv_module_1_sequential_0_bias = const()[name = tensor("inner_encoder_conv_module_1_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15105024)))]; + tensor inner_encoder_conv_module_1_sequential_0_weight = const()[name = tensor("inner_encoder_conv_module_1_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15106112)))]; + tensor inner_encoder_conv_module_1_sequential_2_conv_bias = const()[name = tensor("inner_encoder_conv_module_1_sequential_2_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15107200)))]; + tensor inner_encoder_conv_module_1_sequential_2_conv_weight = const()[name = tensor("inner_encoder_conv_module_1_sequential_2_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15109312)))]; + tensor inner_encoder_conv_module_1_sequential_4_conv_weight = const()[name = tensor("inner_encoder_conv_module_1_sequential_4_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15633664)))]; + tensor inner_encoder_conv_module_1_sequential_5_running_var = const()[name = tensor("inner_encoder_conv_module_1_sequential_5_running_var"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15650112)))]; + tensor inner_encoder_conv_module_1_sequential_5_running_mean = const()[name = tensor("inner_encoder_conv_module_1_sequential_5_running_mean"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15651200)))]; + tensor inner_encoder_conv_module_1_sequential_5_bias = const()[name = tensor("inner_encoder_conv_module_1_sequential_5_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15652288)))]; + tensor inner_encoder_conv_module_1_sequential_5_weight = const()[name = tensor("inner_encoder_conv_module_1_sequential_5_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15653376)))]; + tensor inner_encoder_conv_module_1_sequential_7_conv_bias = const()[name = tensor("inner_encoder_conv_module_1_sequential_7_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15654464)))]; + tensor inner_encoder_conv_module_1_sequential_7_conv_weight = const()[name = tensor("inner_encoder_conv_module_1_sequential_7_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15655552)))]; + tensor inner_encoder_ffn2_1_module_sequential_0_bias = const()[name = tensor("inner_encoder_ffn2_1_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15917760)))]; + tensor inner_encoder_ffn2_1_module_sequential_0_weight = const()[name = tensor("inner_encoder_ffn2_1_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15918848)))]; + tensor inner_encoder_ffn2_1_module_sequential_1_linear_bias = const()[name = tensor("inner_encoder_ffn2_1_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15919936)))]; + tensor inner_encoder_ffn2_1_module_sequential_1_linear_weight = const()[name = tensor("inner_encoder_ffn2_1_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15924096)))]; + tensor inner_encoder_ffn2_1_module_sequential_4_linear_bias = const()[name = tensor("inner_encoder_ffn2_1_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(16972736)))]; + tensor inner_encoder_ffn2_1_module_sequential_4_linear_weight = const()[name = tensor("inner_encoder_ffn2_1_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(16973824)))]; + tensor inner_encoder_layer_norm_1_bias = const()[name = tensor("inner_encoder_layer_norm_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18022464)))]; + tensor inner_encoder_layer_norm_1_weight = const()[name = tensor("inner_encoder_layer_norm_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18023552)))]; + tensor inner_encoder_ffn1_2_module_sequential_0_bias = const()[name = tensor("inner_encoder_ffn1_2_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18024640)))]; + tensor inner_encoder_ffn1_2_module_sequential_0_weight = const()[name = tensor("inner_encoder_ffn1_2_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18025728)))]; + tensor inner_encoder_ffn1_2_module_sequential_1_linear_bias = const()[name = tensor("inner_encoder_ffn1_2_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18026816)))]; + tensor inner_encoder_ffn1_2_module_sequential_1_linear_weight = const()[name = tensor("inner_encoder_ffn1_2_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18030976)))]; + tensor inner_encoder_ffn1_2_module_sequential_4_linear_bias = const()[name = tensor("inner_encoder_ffn1_2_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(19079616)))]; + tensor inner_encoder_ffn1_2_module_sequential_4_linear_weight = const()[name = tensor("inner_encoder_ffn1_2_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(19080704)))]; + tensor inner_encoder_ret_lns_2_bias = const()[name = tensor("inner_encoder_ret_lns_2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20129344)))]; + tensor inner_encoder_ret_lns_2_weight = const()[name = tensor("inner_encoder_ret_lns_2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20130432)))]; + tensor inner_encoder_q_proj_2_bias = const()[name = tensor("inner_encoder_q_proj_2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20131520)))]; + tensor inner_encoder_q_proj_2_weight = const()[name = tensor("inner_encoder_q_proj_2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20132608)))]; + tensor inner_encoder_k_proj_2_bias = const()[name = tensor("inner_encoder_k_proj_2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20394816)))]; + tensor inner_encoder_k_proj_2_weight = const()[name = tensor("inner_encoder_k_proj_2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20395904)))]; + tensor inner_encoder_v_proj_2_bias = const()[name = tensor("inner_encoder_v_proj_2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20658112)))]; + tensor inner_encoder_v_proj_2_weight = const()[name = tensor("inner_encoder_v_proj_2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20659200)))]; + tensor inner_encoder_g_proj_2_bias = const()[name = tensor("inner_encoder_g_proj_2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20921408)))]; + tensor inner_encoder_g_proj_2_weight = const()[name = tensor("inner_encoder_g_proj_2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20922496)))]; + tensor inner_encoder_out_proj_2_bias = const()[name = tensor("inner_encoder_out_proj_2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21184704)))]; + tensor inner_encoder_out_proj_2_weight = const()[name = tensor("inner_encoder_out_proj_2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21185792)))]; + tensor inner_encoder_conv_module_2_sequential_0_bias = const()[name = tensor("inner_encoder_conv_module_2_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21448000)))]; + tensor inner_encoder_conv_module_2_sequential_0_weight = const()[name = tensor("inner_encoder_conv_module_2_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21449088)))]; + tensor inner_encoder_conv_module_2_sequential_2_conv_bias = const()[name = tensor("inner_encoder_conv_module_2_sequential_2_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21450176)))]; + tensor inner_encoder_conv_module_2_sequential_2_conv_weight = const()[name = tensor("inner_encoder_conv_module_2_sequential_2_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21452288)))]; + tensor inner_encoder_conv_module_2_sequential_4_conv_weight = const()[name = tensor("inner_encoder_conv_module_2_sequential_4_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21976640)))]; + tensor inner_encoder_conv_module_2_sequential_5_running_var = const()[name = tensor("inner_encoder_conv_module_2_sequential_5_running_var"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21993088)))]; + tensor inner_encoder_conv_module_2_sequential_5_running_mean = const()[name = tensor("inner_encoder_conv_module_2_sequential_5_running_mean"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21994176)))]; + tensor inner_encoder_conv_module_2_sequential_5_bias = const()[name = tensor("inner_encoder_conv_module_2_sequential_5_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21995264)))]; + tensor inner_encoder_conv_module_2_sequential_5_weight = const()[name = tensor("inner_encoder_conv_module_2_sequential_5_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21996352)))]; + tensor inner_encoder_conv_module_2_sequential_7_conv_bias = const()[name = tensor("inner_encoder_conv_module_2_sequential_7_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21997440)))]; + tensor inner_encoder_conv_module_2_sequential_7_conv_weight = const()[name = tensor("inner_encoder_conv_module_2_sequential_7_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21998528)))]; + tensor inner_encoder_ffn2_2_module_sequential_0_bias = const()[name = tensor("inner_encoder_ffn2_2_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(22260736)))]; + tensor inner_encoder_ffn2_2_module_sequential_0_weight = const()[name = tensor("inner_encoder_ffn2_2_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(22261824)))]; + tensor inner_encoder_ffn2_2_module_sequential_1_linear_bias = const()[name = tensor("inner_encoder_ffn2_2_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(22262912)))]; + tensor inner_encoder_ffn2_2_module_sequential_1_linear_weight = const()[name = tensor("inner_encoder_ffn2_2_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(22267072)))]; + tensor inner_encoder_ffn2_2_module_sequential_4_linear_bias = const()[name = tensor("inner_encoder_ffn2_2_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(23315712)))]; + tensor inner_encoder_ffn2_2_module_sequential_4_linear_weight = const()[name = tensor("inner_encoder_ffn2_2_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(23316800)))]; + tensor inner_encoder_layer_norm_2_bias = const()[name = tensor("inner_encoder_layer_norm_2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(24365440)))]; + tensor inner_encoder_layer_norm_2_weight = const()[name = tensor("inner_encoder_layer_norm_2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(24366528)))]; + tensor inner_encoder_ffn1_3_module_sequential_0_bias = const()[name = tensor("inner_encoder_ffn1_3_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(24367616)))]; + tensor inner_encoder_ffn1_3_module_sequential_0_weight = const()[name = tensor("inner_encoder_ffn1_3_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(24368704)))]; + tensor inner_encoder_ffn1_3_module_sequential_1_linear_bias = const()[name = tensor("inner_encoder_ffn1_3_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(24369792)))]; + tensor inner_encoder_ffn1_3_module_sequential_1_linear_weight = const()[name = tensor("inner_encoder_ffn1_3_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(24373952)))]; + tensor inner_encoder_ffn1_3_module_sequential_4_linear_bias = const()[name = tensor("inner_encoder_ffn1_3_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(25422592)))]; + tensor inner_encoder_ffn1_3_module_sequential_4_linear_weight = const()[name = tensor("inner_encoder_ffn1_3_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(25423680)))]; + tensor inner_encoder_ret_lns_3_bias = const()[name = tensor("inner_encoder_ret_lns_3_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(26472320)))]; + tensor inner_encoder_ret_lns_3_weight = const()[name = tensor("inner_encoder_ret_lns_3_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(26473408)))]; + tensor inner_encoder_q_proj_3_bias = const()[name = tensor("inner_encoder_q_proj_3_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(26474496)))]; + tensor inner_encoder_q_proj_3_weight = const()[name = tensor("inner_encoder_q_proj_3_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(26475584)))]; + tensor inner_encoder_k_proj_3_bias = const()[name = tensor("inner_encoder_k_proj_3_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(26737792)))]; + tensor inner_encoder_k_proj_3_weight = const()[name = tensor("inner_encoder_k_proj_3_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(26738880)))]; + tensor inner_encoder_v_proj_3_bias = const()[name = tensor("inner_encoder_v_proj_3_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27001088)))]; + tensor inner_encoder_v_proj_3_weight = const()[name = tensor("inner_encoder_v_proj_3_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27002176)))]; + tensor inner_encoder_g_proj_3_bias = const()[name = tensor("inner_encoder_g_proj_3_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27264384)))]; + tensor inner_encoder_g_proj_3_weight = const()[name = tensor("inner_encoder_g_proj_3_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27265472)))]; + tensor inner_encoder_out_proj_3_bias = const()[name = tensor("inner_encoder_out_proj_3_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27527680)))]; + tensor inner_encoder_out_proj_3_weight = const()[name = tensor("inner_encoder_out_proj_3_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27528768)))]; + tensor inner_encoder_conv_module_3_sequential_0_bias = const()[name = tensor("inner_encoder_conv_module_3_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27790976)))]; + tensor inner_encoder_conv_module_3_sequential_0_weight = const()[name = tensor("inner_encoder_conv_module_3_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27792064)))]; + tensor inner_encoder_conv_module_3_sequential_2_conv_bias = const()[name = tensor("inner_encoder_conv_module_3_sequential_2_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27793152)))]; + tensor inner_encoder_conv_module_3_sequential_2_conv_weight = const()[name = tensor("inner_encoder_conv_module_3_sequential_2_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27795264)))]; + tensor inner_encoder_conv_module_3_sequential_4_conv_weight = const()[name = tensor("inner_encoder_conv_module_3_sequential_4_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28319616)))]; + tensor inner_encoder_conv_module_3_sequential_5_running_var = const()[name = tensor("inner_encoder_conv_module_3_sequential_5_running_var"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28336064)))]; + tensor inner_encoder_conv_module_3_sequential_5_running_mean = const()[name = tensor("inner_encoder_conv_module_3_sequential_5_running_mean"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28337152)))]; + tensor inner_encoder_conv_module_3_sequential_5_bias = const()[name = tensor("inner_encoder_conv_module_3_sequential_5_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28338240)))]; + tensor inner_encoder_conv_module_3_sequential_5_weight = const()[name = tensor("inner_encoder_conv_module_3_sequential_5_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28339328)))]; + tensor inner_encoder_conv_module_3_sequential_7_conv_bias = const()[name = tensor("inner_encoder_conv_module_3_sequential_7_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28340416)))]; + tensor inner_encoder_conv_module_3_sequential_7_conv_weight = const()[name = tensor("inner_encoder_conv_module_3_sequential_7_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28341504)))]; + tensor inner_encoder_ffn2_3_module_sequential_0_bias = const()[name = tensor("inner_encoder_ffn2_3_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28603712)))]; + tensor inner_encoder_ffn2_3_module_sequential_0_weight = const()[name = tensor("inner_encoder_ffn2_3_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28604800)))]; + tensor inner_encoder_ffn2_3_module_sequential_1_linear_bias = const()[name = tensor("inner_encoder_ffn2_3_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28605888)))]; + tensor inner_encoder_ffn2_3_module_sequential_1_linear_weight = const()[name = tensor("inner_encoder_ffn2_3_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28610048)))]; + tensor inner_encoder_ffn2_3_module_sequential_4_linear_bias = const()[name = tensor("inner_encoder_ffn2_3_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(29658688)))]; + tensor inner_encoder_ffn2_3_module_sequential_4_linear_weight = const()[name = tensor("inner_encoder_ffn2_3_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(29659776)))]; + tensor inner_encoder_layer_norm_3_bias = const()[name = tensor("inner_encoder_layer_norm_3_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(30708416)))]; + tensor inner_encoder_layer_norm_3_weight = const()[name = tensor("inner_encoder_layer_norm_3_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(30709504)))]; + tensor inner_decoder__causal_mask = const()[name = tensor("inner_decoder__causal_mask"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(30710592)))]; + tensor inner_decoder_convert_bias = const()[name = tensor("inner_decoder_convert_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(30710720)))]; + tensor inner_decoder_convert_weight = const()[name = tensor("inner_decoder_convert_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(30711808)))]; + tensor inner_decoder_q_proj_0_bias = const()[name = tensor("inner_decoder_q_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(31236160)))]; + tensor inner_decoder_q_proj_0_weight = const()[name = tensor("inner_decoder_q_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(31237248)))]; + tensor inner_decoder_k_proj_0_bias = const()[name = tensor("inner_decoder_k_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(31499456)))]; + tensor inner_decoder_k_proj_0_weight = const()[name = tensor("inner_decoder_k_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(31500544)))]; + tensor inner_decoder_v_proj_0_bias = const()[name = tensor("inner_decoder_v_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(31762752)))]; + tensor inner_decoder_v_proj_0_weight = const()[name = tensor("inner_decoder_v_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(31763840)))]; + tensor inner_decoder_g_proj_0_bias = const()[name = tensor("inner_decoder_g_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32026048)))]; + tensor inner_decoder_g_proj_0_weight = const()[name = tensor("inner_decoder_g_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32027136)))]; + tensor inner_decoder_out_proj_0_bias = const()[name = tensor("inner_decoder_out_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32289344)))]; + tensor inner_decoder_out_proj_0_weight = const()[name = tensor("inner_decoder_out_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32290432)))]; + tensor inner_decoder_norm11_0_bias = const()[name = tensor("inner_decoder_norm11_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32552640)))]; + tensor inner_decoder_norm11_0_weight = const()[name = tensor("inner_decoder_norm11_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32553728)))]; + tensor inner_decoder_self_attn2_0_out_proj_bias = const()[name = tensor("inner_decoder_self_attn2_0_out_proj_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32554816)))]; + tensor inner_decoder_self_attn2_0_out_proj_weight = const()[name = tensor("inner_decoder_self_attn2_0_out_proj_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32555904)))]; + tensor inner_decoder_self_attn2_0_in_proj_bias = const()[name = tensor("inner_decoder_self_attn2_0_in_proj_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32818112)))]; + tensor inner_decoder_self_attn2_0_in_proj_weight = const()[name = tensor("inner_decoder_self_attn2_0_in_proj_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32821248)))]; + tensor inner_decoder_norm21_0_bias = const()[name = tensor("inner_decoder_norm21_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(33607744)))]; + tensor inner_decoder_norm21_0_weight = const()[name = tensor("inner_decoder_norm21_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(33608832)))]; + tensor inner_decoder_linear1_0_bias = const()[name = tensor("inner_decoder_linear1_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(33609920)))]; + tensor inner_decoder_linear1_0_weight = const()[name = tensor("inner_decoder_linear1_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(33618176)))]; + tensor inner_decoder_linear2_0_bias = const()[name = tensor("inner_decoder_linear2_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(35715392)))]; + tensor inner_decoder_linear2_0_weight = const()[name = tensor("inner_decoder_linear2_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(35716480)))]; + tensor inner_decoder_norm22_0_bias = const()[name = tensor("inner_decoder_norm22_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(37813696)))]; + tensor inner_decoder_norm22_0_weight = const()[name = tensor("inner_decoder_norm22_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(37814784)))]; + tensor inner_decoder_q_proj_1_bias = const()[name = tensor("inner_decoder_q_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(37815872)))]; + tensor inner_decoder_q_proj_1_weight = const()[name = tensor("inner_decoder_q_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(37816960)))]; + tensor inner_decoder_k_proj_1_bias = const()[name = tensor("inner_decoder_k_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38079168)))]; + tensor inner_decoder_k_proj_1_weight = const()[name = tensor("inner_decoder_k_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38080256)))]; + tensor inner_decoder_v_proj_1_bias = const()[name = tensor("inner_decoder_v_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38342464)))]; + tensor inner_decoder_v_proj_1_weight = const()[name = tensor("inner_decoder_v_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38343552)))]; + tensor inner_decoder_g_proj_1_bias = const()[name = tensor("inner_decoder_g_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38605760)))]; + tensor inner_decoder_g_proj_1_weight = const()[name = tensor("inner_decoder_g_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38606848)))]; + tensor inner_decoder_out_proj_1_bias = const()[name = tensor("inner_decoder_out_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38869056)))]; + tensor inner_decoder_out_proj_1_weight = const()[name = tensor("inner_decoder_out_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38870144)))]; + tensor inner_decoder_norm11_1_bias = const()[name = tensor("inner_decoder_norm11_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(39132352)))]; + tensor inner_decoder_norm11_1_weight = const()[name = tensor("inner_decoder_norm11_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(39133440)))]; + tensor inner_decoder_self_attn2_1_out_proj_bias = const()[name = tensor("inner_decoder_self_attn2_1_out_proj_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(39134528)))]; + tensor inner_decoder_self_attn2_1_out_proj_weight = const()[name = tensor("inner_decoder_self_attn2_1_out_proj_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(39135616)))]; + tensor inner_decoder_self_attn2_1_in_proj_bias = const()[name = tensor("inner_decoder_self_attn2_1_in_proj_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(39397824)))]; + tensor inner_decoder_self_attn2_1_in_proj_weight = const()[name = tensor("inner_decoder_self_attn2_1_in_proj_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(39400960)))]; + tensor inner_decoder_norm21_1_bias = const()[name = tensor("inner_decoder_norm21_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(40187456)))]; + tensor inner_decoder_norm21_1_weight = const()[name = tensor("inner_decoder_norm21_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(40188544)))]; + tensor inner_decoder_linear1_1_bias = const()[name = tensor("inner_decoder_linear1_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(40189632)))]; + tensor inner_decoder_linear1_1_weight = const()[name = tensor("inner_decoder_linear1_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(40197888)))]; + tensor inner_decoder_linear2_1_bias = const()[name = tensor("inner_decoder_linear2_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(42295104)))]; + tensor inner_decoder_linear2_1_weight = const()[name = tensor("inner_decoder_linear2_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(42296192)))]; + tensor inner_decoder_norm22_1_bias = const()[name = tensor("inner_decoder_norm22_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(44393408)))]; + tensor inner_decoder_norm22_1_weight = const()[name = tensor("inner_decoder_norm22_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(44394496)))]; + tensor var_19_begin_0 = const()[name = tensor("op_19_begin_0"), val = tensor([0, 0, 0])]; + tensor var_19_end_0 = const()[name = tensor("op_19_end_0"), val = tensor([1, 15, 23])]; + tensor var_19_end_mask_0 = const()[name = tensor("op_19_end_mask_0"), val = tensor([true, false, true])]; + tensor var_19 = slice_by_index(begin = var_19_begin_0, end = var_19_end_0, end_mask = var_19_end_mask_0, x = features)[name = tensor("op_19")]; + tensor var_29_begin_0 = const()[name = tensor("op_29_begin_0"), val = tensor([0, 10, 0])]; + tensor var_29_end_0 = const()[name = tensor("op_29_end_0"), val = tensor([1, 25, 23])]; + tensor var_29_end_mask_0 = const()[name = tensor("op_29_end_mask_0"), val = tensor([true, false, true])]; + tensor var_29 = slice_by_index(begin = var_29_begin_0, end = var_29_end_0, end_mask = var_29_end_mask_0, x = features)[name = tensor("op_29")]; + tensor var_39_begin_0 = const()[name = tensor("op_39_begin_0"), val = tensor([0, 20, 0])]; + tensor var_39_end_0 = const()[name = tensor("op_39_end_0"), val = tensor([1, 35, 23])]; + tensor var_39_end_mask_0 = const()[name = tensor("op_39_end_mask_0"), val = tensor([true, false, true])]; + tensor var_39 = slice_by_index(begin = var_39_begin_0, end = var_39_end_0, end_mask = var_39_end_mask_0, x = features)[name = tensor("op_39")]; + tensor var_49_begin_0 = const()[name = tensor("op_49_begin_0"), val = tensor([0, 30, 0])]; + tensor var_49_end_0 = const()[name = tensor("op_49_end_0"), val = tensor([1, 1, 23])]; + tensor var_49_end_mask_0 = const()[name = tensor("op_49_end_mask_0"), val = tensor([true, true, true])]; + tensor var_49 = slice_by_index(begin = var_49_begin_0, end = var_49_end_0, end_mask = var_49_end_mask_0, x = features)[name = tensor("op_49")]; + tensor stacked_axis_0 = const()[name = tensor("stacked_axis_0"), val = tensor(1)]; + tensor stacked = stack(axis = stacked_axis_0, values = (var_19, var_29, var_39, var_49))[name = tensor("stacked")]; + tensor var_56 = const()[name = tensor("op_56"), val = tensor([1, 4, 345])]; + tensor input_1 = reshape(shape = var_56, x = stacked)[name = tensor("input_1")]; + tensor var_59 = const()[name = tensor("op_59"), val = tensor(0x1p+0)]; + tensor var_65 = const()[name = tensor("op_65"), val = tensor(true)]; + tensor var_66 = const()[name = tensor("op_66"), val = tensor(0x1.4f8b58p-17)]; + tensor var_69 = const()[name = tensor("op_69"), val = tensor(0)]; + tensor var_71 = const()[name = tensor("op_71"), val = tensor(2)]; + tensor var_72 = const()[name = tensor("op_72"), val = tensor(-1)]; + tensor var_74 = const()[name = tensor("op_74"), val = tensor(0x1.0c6f7ap-20)]; + tensor var_79 = const()[name = tensor("op_79"), val = tensor(0x1.5798eep-27)]; + tensor var_82 = const()[name = tensor("op_82"), val = tensor(1)]; + tensor input_3 = linear(bias = inner_encoder_input_projection_linear_bias, weight = inner_encoder_input_projection_linear_weight, x = input_1)[name = tensor("linear_0")]; tensor input_5_axes_0 = const()[name = tensor("input_5_axes_0"), val = tensor([-1])]; - tensor input_5 = layer_norm(axes = input_5_axes_0, beta = encoder_ffn1_0_module_sequential_0_bias, epsilon = var_28, gamma = encoder_ffn1_0_module_sequential_0_weight, x = input_3)[name = tensor("input_5")]; - tensor inputs_1 = linear(bias = encoder_ffn1_0_module_sequential_1_linear_bias, weight = encoder_ffn1_0_module_sequential_1_linear_weight, x = input_5)[name = tensor("linear_1")]; - tensor input_7 = silu(x = inputs_1)[name = tensor("input_7")]; - tensor input_11 = linear(bias = encoder_ffn1_0_module_sequential_4_linear_bias, weight = encoder_ffn1_0_module_sequential_4_linear_weight, x = input_7)[name = tensor("linear_2")]; - tensor var_147 = const()[name = tensor("op_147"), val = tensor(0x1p-1)]; - tensor var_148 = mul(x = input_11, y = var_147)[name = tensor("op_148")]; - tensor input_13 = add(x = var_148, y = input_3)[name = tensor("input_13")]; + tensor input_5 = layer_norm(axes = input_5_axes_0, beta = inner_encoder_pre_layer_norm_bias, epsilon = var_66, gamma = inner_encoder_pre_layer_norm_weight, x = input_3)[name = tensor("input_5")]; + tensor input_7_axes_0 = const()[name = tensor("input_7_axes_0"), val = tensor([-1])]; + tensor input_7 = layer_norm(axes = input_7_axes_0, beta = inner_encoder_ffn1_0_module_sequential_0_bias, epsilon = var_66, gamma = inner_encoder_ffn1_0_module_sequential_0_weight, x = input_5)[name = tensor("input_7")]; + tensor inputs_1 = linear(bias = inner_encoder_ffn1_0_module_sequential_1_linear_bias, weight = inner_encoder_ffn1_0_module_sequential_1_linear_weight, x = input_7)[name = tensor("linear_1")]; + tensor input_9 = silu(x = inputs_1)[name = tensor("input_9")]; + tensor input_13 = linear(bias = inner_encoder_ffn1_0_module_sequential_4_linear_bias, weight = inner_encoder_ffn1_0_module_sequential_4_linear_weight, x = input_9)[name = tensor("linear_2")]; + tensor var_203 = const()[name = tensor("op_203"), val = tensor(0x1p-1)]; + tensor var_204 = mul(x = input_13, y = var_203)[name = tensor("op_204")]; + tensor input_15 = add(x = var_204, y = input_5)[name = tensor("input_15")]; tensor x_1_axes_0 = const()[name = tensor("x_1_axes_0"), val = tensor([-1])]; - tensor x_1 = layer_norm(axes = x_1_axes_0, beta = encoder_ret_lns_0_bias, epsilon = var_28, gamma = encoder_ret_lns_0_weight, x = input_13)[name = tensor("x_1")]; + tensor x_1 = layer_norm(axes = x_1_axes_0, beta = inner_encoder_ret_lns_0_bias, epsilon = var_66, gamma = inner_encoder_ret_lns_0_weight, x = input_15)[name = tensor("x_1")]; tensor prev_kv_1_begin_0 = const()[name = tensor("prev_kv_1_begin_0"), val = tensor([0, 0, 0, 0, 0])]; tensor prev_kv_1_end_0 = const()[name = tensor("prev_kv_1_end_0"), val = tensor([1, 1, 4, 64, 64])]; tensor prev_kv_1_end_mask_0 = const()[name = tensor("prev_kv_1_end_mask_0"), val = tensor([false, true, true, true, true])]; @@ -239,173 +261,173 @@ program(1.0) tensor prev_scale_1_end_mask_0 = const()[name = tensor("prev_scale_1_end_mask_0"), val = tensor([false, true])]; tensor prev_scale_1_squeeze_mask_0 = const()[name = tensor("prev_scale_1_squeeze_mask_0"), val = tensor([true, false])]; tensor prev_scale_1 = slice_by_index(begin = prev_scale_1_begin_0, end = prev_scale_1_end_0, end_mask = prev_scale_1_end_mask_0, squeeze_mask = prev_scale_1_squeeze_mask_0, x = enc_scale)[name = tensor("prev_scale_1")]; - tensor var_162 = linear(bias = encoder_q_proj_0_bias, weight = encoder_q_proj_0_weight, x = x_1)[name = tensor("linear_3")]; - tensor var_163 = const()[name = tensor("op_163"), val = tensor([1, 4, 4, 64])]; - tensor var_164 = reshape(shape = var_163, x = var_162)[name = tensor("op_164")]; + tensor var_218 = linear(bias = inner_encoder_q_proj_0_bias, weight = inner_encoder_q_proj_0_weight, x = x_1)[name = tensor("linear_3")]; + tensor var_219 = const()[name = tensor("op_219"), val = tensor([1, 4, 4, 64])]; + tensor var_220 = reshape(shape = var_219, x = var_218)[name = tensor("op_220")]; tensor q_1_perm_0 = const()[name = tensor("q_1_perm_0"), val = tensor([0, 2, 1, 3])]; - tensor var_168 = linear(bias = encoder_k_proj_0_bias, weight = encoder_k_proj_0_weight, x = x_1)[name = tensor("linear_4")]; - tensor var_169 = const()[name = tensor("op_169"), val = tensor(0x1p-3)]; - tensor var_170 = mul(x = var_168, y = var_169)[name = tensor("op_170")]; - tensor var_171 = const()[name = tensor("op_171"), val = tensor([1, 4, 4, 64])]; - tensor var_172 = reshape(shape = var_171, x = var_170)[name = tensor("op_172")]; + tensor var_224 = linear(bias = inner_encoder_k_proj_0_bias, weight = inner_encoder_k_proj_0_weight, x = x_1)[name = tensor("linear_4")]; + tensor var_225 = const()[name = tensor("op_225"), val = tensor(0x1p-3)]; + tensor var_226 = mul(x = var_224, y = var_225)[name = tensor("op_226")]; + tensor var_227 = const()[name = tensor("op_227"), val = tensor([1, 4, 4, 64])]; + tensor var_228 = reshape(shape = var_227, x = var_226)[name = tensor("op_228")]; tensor k_1_perm_0 = const()[name = tensor("k_1_perm_0"), val = tensor([0, 2, 1, 3])]; - tensor var_176 = linear(bias = encoder_v_proj_0_bias, weight = encoder_v_proj_0_weight, x = x_1)[name = tensor("linear_5")]; - tensor var_177 = const()[name = tensor("op_177"), val = tensor([1, 4, 4, 64])]; - tensor var_178 = reshape(shape = var_177, x = var_176)[name = tensor("op_178")]; + tensor var_232 = linear(bias = inner_encoder_v_proj_0_bias, weight = inner_encoder_v_proj_0_weight, x = x_1)[name = tensor("linear_5")]; + tensor var_233 = const()[name = tensor("op_233"), val = tensor([1, 4, 4, 64])]; + tensor var_234 = reshape(shape = var_233, x = var_232)[name = tensor("op_234")]; tensor v_1_perm_0 = const()[name = tensor("v_1_perm_0"), val = tensor([0, 2, 1, 3])]; - tensor input_17 = linear(bias = encoder_g_proj_0_bias, weight = encoder_g_proj_0_weight, x = x_1)[name = tensor("linear_6")]; + tensor input_19 = linear(bias = inner_encoder_g_proj_0_bias, weight = inner_encoder_g_proj_0_weight, x = x_1)[name = tensor("linear_6")]; tensor sqrt_s0_1 = sqrt(x = prev_scale_1)[name = tensor("sqrt_s0_1")]; - tensor s_t_1 = add(x = prev_scale_1, y = encoder__t_index)[name = tensor("s_t_1")]; + tensor s_t_1 = add(x = prev_scale_1, y = inner_encoder__t_index)[name = tensor("s_t_1")]; tensor sqrt_s_t_1 = sqrt(x = s_t_1)[name = tensor("sqrt_s_t_1")]; tensor qk_1_transpose_x_1 = const()[name = tensor("qk_1_transpose_x_1"), val = tensor(false)]; tensor qk_1_transpose_y_1 = const()[name = tensor("qk_1_transpose_y_1"), val = tensor(true)]; - tensor k_1 = transpose(perm = k_1_perm_0, x = var_172)[name = tensor("transpose_57")]; - tensor q_1 = transpose(perm = q_1_perm_0, x = var_164)[name = tensor("transpose_58")]; + tensor k_1 = transpose(perm = k_1_perm_0, x = var_228)[name = tensor("transpose_57")]; + tensor q_1 = transpose(perm = q_1_perm_0, x = var_220)[name = tensor("transpose_58")]; tensor qk_1 = matmul(transpose_x = qk_1_transpose_x_1, transpose_y = qk_1_transpose_y_1, x = q_1, y = k_1)[name = tensor("qk_1")]; - tensor var_188 = const()[name = tensor("op_188"), val = tensor([4, 1])]; - tensor var_189 = reshape(shape = var_188, x = sqrt_s_t_1)[name = tensor("op_189")]; - tensor M_1 = real_div(x = encoder__causal_mask, y = var_189)[name = tensor("M_1")]; - tensor var_191 = mul(x = qk_1, y = M_1)[name = tensor("op_191")]; + tensor var_244 = const()[name = tensor("op_244"), val = tensor([4, 1])]; + tensor var_245 = reshape(shape = var_244, x = sqrt_s_t_1)[name = tensor("op_245")]; + tensor M_1 = real_div(x = inner_encoder__causal_mask, y = var_245)[name = tensor("M_1")]; + tensor var_247 = mul(x = qk_1, y = M_1)[name = tensor("op_247")]; tensor inner_1_transpose_x_0 = const()[name = tensor("inner_1_transpose_x_0"), val = tensor(false)]; tensor inner_1_transpose_y_0 = const()[name = tensor("inner_1_transpose_y_0"), val = tensor(false)]; - tensor v_1 = transpose(perm = v_1_perm_0, x = var_178)[name = tensor("transpose_56")]; - tensor inner_1 = matmul(transpose_x = inner_1_transpose_x_0, transpose_y = inner_1_transpose_y_0, x = var_191, y = v_1)[name = tensor("inner_1")]; - tensor var_193_transpose_x_0 = const()[name = tensor("op_193_transpose_x_0"), val = tensor(false)]; - tensor var_193_transpose_y_0 = const()[name = tensor("op_193_transpose_y_0"), val = tensor(false)]; - tensor var_193 = matmul(transpose_x = var_193_transpose_x_0, transpose_y = var_193_transpose_y_0, x = q_1, y = prev_kv_1)[name = tensor("op_193")]; - tensor var_194 = real_div(x = sqrt_s0_1, y = sqrt_s_t_1)[name = tensor("op_194")]; - tensor var_195 = const()[name = tensor("op_195"), val = tensor([1, 1, 4, 1])]; - tensor var_196 = reshape(shape = var_195, x = var_194)[name = tensor("op_196")]; - tensor cross_1 = mul(x = var_193, y = var_196)[name = tensor("cross_1")]; + tensor v_1 = transpose(perm = v_1_perm_0, x = var_234)[name = tensor("transpose_56")]; + tensor inner_1 = matmul(transpose_x = inner_1_transpose_x_0, transpose_y = inner_1_transpose_y_0, x = var_247, y = v_1)[name = tensor("inner_1")]; + tensor var_249_transpose_x_0 = const()[name = tensor("op_249_transpose_x_0"), val = tensor(false)]; + tensor var_249_transpose_y_0 = const()[name = tensor("op_249_transpose_y_0"), val = tensor(false)]; + tensor var_249 = matmul(transpose_x = var_249_transpose_x_0, transpose_y = var_249_transpose_y_0, x = q_1, y = prev_kv_1)[name = tensor("op_249")]; + tensor var_250 = real_div(x = sqrt_s0_1, y = sqrt_s_t_1)[name = tensor("op_250")]; + tensor var_251 = const()[name = tensor("op_251"), val = tensor([1, 1, 4, 1])]; + tensor var_252 = reshape(shape = var_251, x = var_250)[name = tensor("op_252")]; + tensor cross_1 = mul(x = var_249, y = var_252)[name = tensor("cross_1")]; tensor out_1 = add(x = inner_1, y = cross_1)[name = tensor("out_1")]; - tensor var_199 = mul(x = prev_kv_1, y = sqrt_s0_1)[name = tensor("op_199")]; - tensor var_201_transpose_x_1 = const()[name = tensor("op_201_transpose_x_1"), val = tensor(true)]; - tensor var_201_transpose_y_1 = const()[name = tensor("op_201_transpose_y_1"), val = tensor(false)]; - tensor var_201 = matmul(transpose_x = var_201_transpose_x_1, transpose_y = var_201_transpose_y_1, x = k_1, y = v_1)[name = tensor("op_201")]; - tensor new_kv_unnorm_1 = add(x = var_199, y = var_201)[name = tensor("new_kv_unnorm_1")]; - tensor var_203 = const()[name = tensor("op_203"), val = tensor(0x1p+2)]; - tensor new_scale_1 = add(x = prev_scale_1, y = var_203)[name = tensor("new_scale_1")]; - tensor var_205 = sqrt(x = new_scale_1)[name = tensor("op_205")]; - tensor var_206 = real_div(x = new_kv_unnorm_1, y = var_205)[name = tensor("op_206")]; - tensor var_207_perm_0 = const()[name = tensor("op_207_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_255 = mul(x = prev_kv_1, y = sqrt_s0_1)[name = tensor("op_255")]; + tensor var_257_transpose_x_1 = const()[name = tensor("op_257_transpose_x_1"), val = tensor(true)]; + tensor var_257_transpose_y_1 = const()[name = tensor("op_257_transpose_y_1"), val = tensor(false)]; + tensor var_257 = matmul(transpose_x = var_257_transpose_x_1, transpose_y = var_257_transpose_y_1, x = k_1, y = v_1)[name = tensor("op_257")]; + tensor new_kv_unnorm_1 = add(x = var_255, y = var_257)[name = tensor("new_kv_unnorm_1")]; + tensor var_259 = const()[name = tensor("op_259"), val = tensor(0x1p+2)]; + tensor new_scale_1 = add(x = prev_scale_1, y = var_259)[name = tensor("new_scale_1")]; + tensor var_261 = sqrt(x = new_scale_1)[name = tensor("op_261")]; + tensor var_262 = real_div(x = new_kv_unnorm_1, y = var_261)[name = tensor("op_262")]; + tensor var_263_perm_0 = const()[name = tensor("op_263_perm_0"), val = tensor([0, 2, 1, 3])]; tensor out_3_axes_0 = const()[name = tensor("out_3_axes_0"), val = tensor([-1])]; - tensor var_207 = transpose(perm = var_207_perm_0, x = out_1)[name = tensor("transpose_55")]; - tensor out_3 = layer_norm(axes = out_3_axes_0, epsilon = var_17, x = var_207)[name = tensor("out_3")]; - tensor var_211 = const()[name = tensor("op_211"), val = tensor([1, 4, 256])]; - tensor out_5 = reshape(shape = var_211, x = out_3)[name = tensor("out_5")]; - tensor var_213 = silu(x = input_17)[name = tensor("op_213")]; - tensor input_19 = mul(x = var_213, y = out_5)[name = tensor("input_19")]; - tensor ret_out_1 = linear(bias = encoder_out_proj_0_bias, weight = encoder_out_proj_0_weight, x = input_19)[name = tensor("linear_7")]; - tensor x_3 = add(x = input_13, y = ret_out_1)[name = tensor("x_3")]; + tensor var_263 = transpose(perm = var_263_perm_0, x = out_1)[name = tensor("transpose_55")]; + tensor out_3 = layer_norm(axes = out_3_axes_0, epsilon = var_74, x = var_263)[name = tensor("out_3")]; + tensor var_267 = const()[name = tensor("op_267"), val = tensor([1, 4, 256])]; + tensor out_5 = reshape(shape = var_267, x = out_3)[name = tensor("out_5")]; + tensor var_269 = silu(x = input_19)[name = tensor("op_269")]; + tensor input_21 = mul(x = var_269, y = out_5)[name = tensor("input_21")]; + tensor ret_out_1 = linear(bias = inner_encoder_out_proj_0_bias, weight = inner_encoder_out_proj_0_weight, x = input_21)[name = tensor("linear_7")]; + tensor x_3 = add(x = input_15, y = ret_out_1)[name = tensor("x_3")]; tensor window_1_begin_0 = const()[name = tensor("window_1_begin_0"), val = tensor([0, 0, 0, 0])]; tensor window_1_end_0 = const()[name = tensor("window_1_end_0"), val = tensor([1, 1, 16, 256])]; tensor window_1_end_mask_0 = const()[name = tensor("window_1_end_mask_0"), val = tensor([false, true, true, true])]; tensor window_1_squeeze_mask_0 = const()[name = tensor("window_1_squeeze_mask_0"), val = tensor([true, false, false, false])]; tensor window_1 = slice_by_index(begin = window_1_begin_0, end = window_1_end_0, end_mask = window_1_end_mask_0, squeeze_mask = window_1_squeeze_mask_0, x = enc_conv_cache)[name = tensor("window_1")]; - tensor var_221_begin_0 = const()[name = tensor("op_221_begin_0"), val = tensor([0, 0, 0])]; - tensor var_221_end_0 = const()[name = tensor("op_221_end_0"), val = tensor([1, 1, 256])]; - tensor var_221_end_mask_0 = const()[name = tensor("op_221_end_mask_0"), val = tensor([true, false, true])]; - tensor var_221 = slice_by_index(begin = var_221_begin_0, end = var_221_end_0, end_mask = var_221_end_mask_0, x = x_3)[name = tensor("op_221")]; - tensor var_224_begin_0 = const()[name = tensor("op_224_begin_0"), val = tensor([0, 1, 0])]; - tensor var_224_end_0 = const()[name = tensor("op_224_end_0"), val = tensor([1, 16, 256])]; - tensor var_224_end_mask_0 = const()[name = tensor("op_224_end_mask_0"), val = tensor([true, true, true])]; - tensor var_224 = slice_by_index(begin = var_224_begin_0, end = var_224_end_0, end_mask = var_224_end_mask_0, x = window_1)[name = tensor("op_224")]; + tensor var_277_begin_0 = const()[name = tensor("op_277_begin_0"), val = tensor([0, 0, 0])]; + tensor var_277_end_0 = const()[name = tensor("op_277_end_0"), val = tensor([1, 1, 256])]; + tensor var_277_end_mask_0 = const()[name = tensor("op_277_end_mask_0"), val = tensor([true, false, true])]; + tensor var_277 = slice_by_index(begin = var_277_begin_0, end = var_277_end_0, end_mask = var_277_end_mask_0, x = x_3)[name = tensor("op_277")]; + tensor var_280_begin_0 = const()[name = tensor("op_280_begin_0"), val = tensor([0, 1, 0])]; + tensor var_280_end_0 = const()[name = tensor("op_280_end_0"), val = tensor([1, 16, 256])]; + tensor var_280_end_mask_0 = const()[name = tensor("op_280_end_mask_0"), val = tensor([true, true, true])]; + tensor var_280 = slice_by_index(begin = var_280_begin_0, end = var_280_end_0, end_mask = var_280_end_mask_0, x = window_1)[name = tensor("op_280")]; tensor window_3_interleave_0 = const()[name = tensor("window_3_interleave_0"), val = tensor(false)]; - tensor window_3 = concat(axis = var_26, interleave = window_3_interleave_0, values = (var_224, var_221))[name = tensor("window_3")]; - tensor var_229_begin_0 = const()[name = tensor("op_229_begin_0"), val = tensor([0, 1, 0])]; - tensor var_229_end_0 = const()[name = tensor("op_229_end_0"), val = tensor([1, 2, 256])]; - tensor var_229_end_mask_0 = const()[name = tensor("op_229_end_mask_0"), val = tensor([true, false, true])]; - tensor var_229 = slice_by_index(begin = var_229_begin_0, end = var_229_end_0, end_mask = var_229_end_mask_0, x = x_3)[name = tensor("op_229")]; - tensor var_232_begin_0 = const()[name = tensor("op_232_begin_0"), val = tensor([0, 1, 0])]; - tensor var_232_end_0 = const()[name = tensor("op_232_end_0"), val = tensor([1, 16, 256])]; - tensor var_232_end_mask_0 = const()[name = tensor("op_232_end_mask_0"), val = tensor([true, true, true])]; - tensor var_232 = slice_by_index(begin = var_232_begin_0, end = var_232_end_0, end_mask = var_232_end_mask_0, x = window_3)[name = tensor("op_232")]; + tensor window_3 = concat(axis = var_82, interleave = window_3_interleave_0, values = (var_280, var_277))[name = tensor("window_3")]; + tensor var_285_begin_0 = const()[name = tensor("op_285_begin_0"), val = tensor([0, 1, 0])]; + tensor var_285_end_0 = const()[name = tensor("op_285_end_0"), val = tensor([1, 2, 256])]; + tensor var_285_end_mask_0 = const()[name = tensor("op_285_end_mask_0"), val = tensor([true, false, true])]; + tensor var_285 = slice_by_index(begin = var_285_begin_0, end = var_285_end_0, end_mask = var_285_end_mask_0, x = x_3)[name = tensor("op_285")]; + tensor var_288_begin_0 = const()[name = tensor("op_288_begin_0"), val = tensor([0, 1, 0])]; + tensor var_288_end_0 = const()[name = tensor("op_288_end_0"), val = tensor([1, 16, 256])]; + tensor var_288_end_mask_0 = const()[name = tensor("op_288_end_mask_0"), val = tensor([true, true, true])]; + tensor var_288 = slice_by_index(begin = var_288_begin_0, end = var_288_end_0, end_mask = var_288_end_mask_0, x = window_3)[name = tensor("op_288")]; tensor window_5_interleave_0 = const()[name = tensor("window_5_interleave_0"), val = tensor(false)]; - tensor window_5 = concat(axis = var_26, interleave = window_5_interleave_0, values = (var_232, var_229))[name = tensor("window_5")]; - tensor var_237_begin_0 = const()[name = tensor("op_237_begin_0"), val = tensor([0, 2, 0])]; - tensor var_237_end_0 = const()[name = tensor("op_237_end_0"), val = tensor([1, 3, 256])]; - tensor var_237_end_mask_0 = const()[name = tensor("op_237_end_mask_0"), val = tensor([true, false, true])]; - tensor var_237 = slice_by_index(begin = var_237_begin_0, end = var_237_end_0, end_mask = var_237_end_mask_0, x = x_3)[name = tensor("op_237")]; - tensor var_240_begin_0 = const()[name = tensor("op_240_begin_0"), val = tensor([0, 1, 0])]; - tensor var_240_end_0 = const()[name = tensor("op_240_end_0"), val = tensor([1, 16, 256])]; - tensor var_240_end_mask_0 = const()[name = tensor("op_240_end_mask_0"), val = tensor([true, true, true])]; - tensor var_240 = slice_by_index(begin = var_240_begin_0, end = var_240_end_0, end_mask = var_240_end_mask_0, x = window_5)[name = tensor("op_240")]; + tensor window_5 = concat(axis = var_82, interleave = window_5_interleave_0, values = (var_288, var_285))[name = tensor("window_5")]; + tensor var_293_begin_0 = const()[name = tensor("op_293_begin_0"), val = tensor([0, 2, 0])]; + tensor var_293_end_0 = const()[name = tensor("op_293_end_0"), val = tensor([1, 3, 256])]; + tensor var_293_end_mask_0 = const()[name = tensor("op_293_end_mask_0"), val = tensor([true, false, true])]; + tensor var_293 = slice_by_index(begin = var_293_begin_0, end = var_293_end_0, end_mask = var_293_end_mask_0, x = x_3)[name = tensor("op_293")]; + tensor var_296_begin_0 = const()[name = tensor("op_296_begin_0"), val = tensor([0, 1, 0])]; + tensor var_296_end_0 = const()[name = tensor("op_296_end_0"), val = tensor([1, 16, 256])]; + tensor var_296_end_mask_0 = const()[name = tensor("op_296_end_mask_0"), val = tensor([true, true, true])]; + tensor var_296 = slice_by_index(begin = var_296_begin_0, end = var_296_end_0, end_mask = var_296_end_mask_0, x = window_5)[name = tensor("op_296")]; tensor window_7_interleave_0 = const()[name = tensor("window_7_interleave_0"), val = tensor(false)]; - tensor window_7 = concat(axis = var_26, interleave = window_7_interleave_0, values = (var_240, var_237))[name = tensor("window_7")]; - tensor var_245_begin_0 = const()[name = tensor("op_245_begin_0"), val = tensor([0, 3, 0])]; - tensor var_245_end_0 = const()[name = tensor("op_245_end_0"), val = tensor([1, 1, 256])]; - tensor var_245_end_mask_0 = const()[name = tensor("op_245_end_mask_0"), val = tensor([true, true, true])]; - tensor var_245 = slice_by_index(begin = var_245_begin_0, end = var_245_end_0, end_mask = var_245_end_mask_0, x = x_3)[name = tensor("op_245")]; - tensor var_248_begin_0 = const()[name = tensor("op_248_begin_0"), val = tensor([0, 1, 0])]; - tensor var_248_end_0 = const()[name = tensor("op_248_end_0"), val = tensor([1, 16, 256])]; - tensor var_248_end_mask_0 = const()[name = tensor("op_248_end_mask_0"), val = tensor([true, true, true])]; - tensor var_248 = slice_by_index(begin = var_248_begin_0, end = var_248_end_0, end_mask = var_248_end_mask_0, x = window_7)[name = tensor("op_248")]; + tensor window_7 = concat(axis = var_82, interleave = window_7_interleave_0, values = (var_296, var_293))[name = tensor("window_7")]; + tensor var_301_begin_0 = const()[name = tensor("op_301_begin_0"), val = tensor([0, 3, 0])]; + tensor var_301_end_0 = const()[name = tensor("op_301_end_0"), val = tensor([1, 1, 256])]; + tensor var_301_end_mask_0 = const()[name = tensor("op_301_end_mask_0"), val = tensor([true, true, true])]; + tensor var_301 = slice_by_index(begin = var_301_begin_0, end = var_301_end_0, end_mask = var_301_end_mask_0, x = x_3)[name = tensor("op_301")]; + tensor var_304_begin_0 = const()[name = tensor("op_304_begin_0"), val = tensor([0, 1, 0])]; + tensor var_304_end_0 = const()[name = tensor("op_304_end_0"), val = tensor([1, 16, 256])]; + tensor var_304_end_mask_0 = const()[name = tensor("op_304_end_mask_0"), val = tensor([true, true, true])]; + tensor var_304 = slice_by_index(begin = var_304_begin_0, end = var_304_end_0, end_mask = var_304_end_mask_0, x = window_7)[name = tensor("op_304")]; tensor window_9_interleave_0 = const()[name = tensor("window_9_interleave_0"), val = tensor(false)]; - tensor window_9 = concat(axis = var_26, interleave = window_9_interleave_0, values = (var_248, var_245))[name = tensor("window_9")]; - tensor input_21_interleave_0 = const()[name = tensor("input_21_interleave_0"), val = tensor(false)]; - tensor input_21 = concat(axis = var_23, interleave = input_21_interleave_0, values = (window_3, window_5, window_7, window_9))[name = tensor("input_21")]; + tensor window_9 = concat(axis = var_82, interleave = window_9_interleave_0, values = (var_304, var_301))[name = tensor("window_9")]; + tensor input_23_interleave_0 = const()[name = tensor("input_23_interleave_0"), val = tensor(false)]; + tensor input_23 = concat(axis = var_69, interleave = input_23_interleave_0, values = (window_3, window_5, window_7, window_9))[name = tensor("input_23")]; tensor x_5_axes_0 = const()[name = tensor("x_5_axes_0"), val = tensor([-1])]; - tensor x_5 = layer_norm(axes = x_5_axes_0, beta = encoder_conv_module_0_sequential_0_bias, epsilon = var_28, gamma = encoder_conv_module_0_sequential_0_weight, x = input_21)[name = tensor("x_5")]; - tensor input_23_perm_0 = const()[name = tensor("input_23_perm_0"), val = tensor([0, 2, 1])]; + tensor x_5 = layer_norm(axes = x_5_axes_0, beta = inner_encoder_conv_module_0_sequential_0_bias, epsilon = var_66, gamma = inner_encoder_conv_module_0_sequential_0_weight, x = input_23)[name = tensor("x_5")]; + tensor input_25_perm_0 = const()[name = tensor("input_25_perm_0"), val = tensor([0, 2, 1])]; tensor inputs_3_pad_type_0 = const()[name = tensor("inputs_3_pad_type_0"), val = tensor("valid")]; tensor inputs_3_strides_0 = const()[name = tensor("inputs_3_strides_0"), val = tensor([1])]; tensor inputs_3_pad_0 = const()[name = tensor("inputs_3_pad_0"), val = tensor([0, 0])]; tensor inputs_3_dilations_0 = const()[name = tensor("inputs_3_dilations_0"), val = tensor([1])]; tensor inputs_3_groups_0 = const()[name = tensor("inputs_3_groups_0"), val = tensor(1)]; - tensor input_23 = transpose(perm = input_23_perm_0, x = x_5)[name = tensor("transpose_54")]; - tensor inputs_3 = conv(bias = encoder_conv_module_0_sequential_2_conv_bias, dilations = inputs_3_dilations_0, groups = inputs_3_groups_0, pad = inputs_3_pad_0, pad_type = inputs_3_pad_type_0, strides = inputs_3_strides_0, weight = encoder_conv_module_0_sequential_2_conv_weight, x = input_23)[name = tensor("inputs_3")]; - tensor var_273_split_sizes_0 = const()[name = tensor("op_273_split_sizes_0"), val = tensor([256, 256])]; - tensor var_273_axis_0 = const()[name = tensor("op_273_axis_0"), val = tensor(1)]; - tensor var_273_0, tensor var_273_1 = split(axis = var_273_axis_0, split_sizes = var_273_split_sizes_0, x = inputs_3)[name = tensor("op_273")]; - tensor var_275 = sigmoid(x = var_273_1)[name = tensor("op_275")]; - tensor inputs_5 = mul(x = var_273_0, y = var_275)[name = tensor("inputs_5")]; + tensor input_25 = transpose(perm = input_25_perm_0, x = x_5)[name = tensor("transpose_54")]; + tensor inputs_3 = conv(bias = inner_encoder_conv_module_0_sequential_2_conv_bias, dilations = inputs_3_dilations_0, groups = inputs_3_groups_0, pad = inputs_3_pad_0, pad_type = inputs_3_pad_type_0, strides = inputs_3_strides_0, weight = inner_encoder_conv_module_0_sequential_2_conv_weight, x = input_25)[name = tensor("inputs_3")]; + tensor var_329_split_sizes_0 = const()[name = tensor("op_329_split_sizes_0"), val = tensor([256, 256])]; + tensor var_329_axis_0 = const()[name = tensor("op_329_axis_0"), val = tensor(1)]; + tensor var_329_0, tensor var_329_1 = split(axis = var_329_axis_0, split_sizes = var_329_split_sizes_0, x = inputs_3)[name = tensor("op_329")]; + tensor var_331 = sigmoid(x = var_329_1)[name = tensor("op_331")]; + tensor inputs_5 = mul(x = var_329_0, y = var_331)[name = tensor("inputs_5")]; tensor outputs_aug_1_pad_type_0 = const()[name = tensor("outputs_aug_1_pad_type_0"), val = tensor("custom")]; tensor outputs_aug_1_pad_0 = const()[name = tensor("outputs_aug_1_pad_0"), val = tensor([15, 15])]; tensor outputs_aug_1_groups_0 = const()[name = tensor("outputs_aug_1_groups_0"), val = tensor(256)]; tensor outputs_aug_1_strides_0 = const()[name = tensor("outputs_aug_1_strides_0"), val = tensor([1])]; tensor outputs_aug_1_dilations_0 = const()[name = tensor("outputs_aug_1_dilations_0"), val = tensor([1])]; - tensor outputs_aug_1 = conv(dilations = outputs_aug_1_dilations_0, groups = outputs_aug_1_groups_0, pad = outputs_aug_1_pad_0, pad_type = outputs_aug_1_pad_type_0, strides = outputs_aug_1_strides_0, weight = encoder_conv_module_0_sequential_4_conv_weight, x = inputs_5)[name = tensor("outputs_aug_1")]; - tensor input_25_begin_0 = const()[name = tensor("input_25_begin_0"), val = tensor([0, 0, 0])]; - tensor input_25_end_0 = const()[name = tensor("input_25_end_0"), val = tensor([4, 256, 16])]; - tensor input_25_end_mask_0 = const()[name = tensor("input_25_end_mask_0"), val = tensor([true, true, false])]; - tensor input_25 = slice_by_index(begin = input_25_begin_0, end = input_25_end_0, end_mask = input_25_end_mask_0, x = outputs_aug_1)[name = tensor("input_25")]; - tensor inputs_7 = batch_norm(beta = encoder_conv_module_0_sequential_5_bias, epsilon = var_28, gamma = encoder_conv_module_0_sequential_5_weight, mean = encoder_conv_module_0_sequential_5_running_mean, variance = encoder_conv_module_0_sequential_5_running_var, x = input_25)[name = tensor("inputs_7")]; - tensor input_27 = silu(x = inputs_7)[name = tensor("input_27")]; - tensor input_29_pad_type_0 = const()[name = tensor("input_29_pad_type_0"), val = tensor("valid")]; - tensor input_29_strides_0 = const()[name = tensor("input_29_strides_0"), val = tensor([1])]; - tensor input_29_pad_0 = const()[name = tensor("input_29_pad_0"), val = tensor([0, 0])]; - tensor input_29_dilations_0 = const()[name = tensor("input_29_dilations_0"), val = tensor([1])]; - tensor input_29_groups_0 = const()[name = tensor("input_29_groups_0"), val = tensor(1)]; - tensor input_29 = conv(bias = encoder_conv_module_0_sequential_7_conv_bias, dilations = input_29_dilations_0, groups = input_29_groups_0, pad = input_29_pad_0, pad_type = input_29_pad_type_0, strides = input_29_strides_0, weight = encoder_conv_module_0_sequential_7_conv_weight, x = input_27)[name = tensor("input_29")]; + tensor outputs_aug_1 = conv(dilations = outputs_aug_1_dilations_0, groups = outputs_aug_1_groups_0, pad = outputs_aug_1_pad_0, pad_type = outputs_aug_1_pad_type_0, strides = outputs_aug_1_strides_0, weight = inner_encoder_conv_module_0_sequential_4_conv_weight, x = inputs_5)[name = tensor("outputs_aug_1")]; + tensor input_27_begin_0 = const()[name = tensor("input_27_begin_0"), val = tensor([0, 0, 0])]; + tensor input_27_end_0 = const()[name = tensor("input_27_end_0"), val = tensor([4, 256, 16])]; + tensor input_27_end_mask_0 = const()[name = tensor("input_27_end_mask_0"), val = tensor([true, true, false])]; + tensor input_27 = slice_by_index(begin = input_27_begin_0, end = input_27_end_0, end_mask = input_27_end_mask_0, x = outputs_aug_1)[name = tensor("input_27")]; + tensor inputs_7 = batch_norm(beta = inner_encoder_conv_module_0_sequential_5_bias, epsilon = var_66, gamma = inner_encoder_conv_module_0_sequential_5_weight, mean = inner_encoder_conv_module_0_sequential_5_running_mean, variance = inner_encoder_conv_module_0_sequential_5_running_var, x = input_27)[name = tensor("inputs_7")]; + tensor input_29 = silu(x = inputs_7)[name = tensor("input_29")]; + tensor input_31_pad_type_0 = const()[name = tensor("input_31_pad_type_0"), val = tensor("valid")]; + tensor input_31_strides_0 = const()[name = tensor("input_31_strides_0"), val = tensor([1])]; + tensor input_31_pad_0 = const()[name = tensor("input_31_pad_0"), val = tensor([0, 0])]; + tensor input_31_dilations_0 = const()[name = tensor("input_31_dilations_0"), val = tensor([1])]; + tensor input_31_groups_0 = const()[name = tensor("input_31_groups_0"), val = tensor(1)]; + tensor input_31 = conv(bias = inner_encoder_conv_module_0_sequential_7_conv_bias, dilations = input_31_dilations_0, groups = input_31_groups_0, pad = input_31_pad_0, pad_type = input_31_pad_type_0, strides = input_31_strides_0, weight = inner_encoder_conv_module_0_sequential_7_conv_weight, x = input_29)[name = tensor("input_31")]; tensor conv_out_1_perm_0 = const()[name = tensor("conv_out_1_perm_0"), val = tensor([0, 2, 1])]; - tensor var_306_begin_0 = const()[name = tensor("op_306_begin_0"), val = tensor([0, -1, 0])]; - tensor var_306_end_0 = const()[name = tensor("op_306_end_0"), val = tensor([4, 16, 256])]; - tensor var_306_end_mask_0 = const()[name = tensor("op_306_end_mask_0"), val = tensor([true, true, true])]; - tensor conv_out_1 = transpose(perm = conv_out_1_perm_0, x = input_29)[name = tensor("transpose_53")]; - tensor var_306 = slice_by_index(begin = var_306_begin_0, end = var_306_end_0, end_mask = var_306_end_mask_0, x = conv_out_1)[name = tensor("op_306")]; - tensor var_308_perm_0 = const()[name = tensor("op_308_perm_0"), val = tensor([1, 0, 2])]; - tensor var_308 = transpose(perm = var_308_perm_0, x = var_306)[name = tensor("transpose_52")]; - tensor input_31 = add(x = x_3, y = var_308)[name = tensor("input_31")]; - tensor input_33_axes_0 = const()[name = tensor("input_33_axes_0"), val = tensor([-1])]; - tensor input_33 = layer_norm(axes = input_33_axes_0, beta = encoder_ffn2_0_module_sequential_0_bias, epsilon = var_28, gamma = encoder_ffn2_0_module_sequential_0_weight, x = input_31)[name = tensor("input_33")]; - tensor inputs_9 = linear(bias = encoder_ffn2_0_module_sequential_1_linear_bias, weight = encoder_ffn2_0_module_sequential_1_linear_weight, x = input_33)[name = tensor("linear_8")]; - tensor input_35 = silu(x = inputs_9)[name = tensor("input_35")]; - tensor input_39 = linear(bias = encoder_ffn2_0_module_sequential_4_linear_bias, weight = encoder_ffn2_0_module_sequential_4_linear_weight, x = input_35)[name = tensor("linear_9")]; - tensor var_331 = const()[name = tensor("op_331"), val = tensor(0x1p-1)]; - tensor var_332 = mul(x = input_39, y = var_331)[name = tensor("op_332")]; - tensor input_41 = add(x = var_332, y = input_31)[name = tensor("input_41")]; - tensor input_43_axes_0 = const()[name = tensor("input_43_axes_0"), val = tensor([-1])]; - tensor input_43 = layer_norm(axes = input_43_axes_0, beta = encoder_layer_norm_0_bias, epsilon = var_28, gamma = encoder_layer_norm_0_weight, x = input_41)[name = tensor("input_43")]; + tensor var_362_begin_0 = const()[name = tensor("op_362_begin_0"), val = tensor([0, -1, 0])]; + tensor var_362_end_0 = const()[name = tensor("op_362_end_0"), val = tensor([4, 16, 256])]; + tensor var_362_end_mask_0 = const()[name = tensor("op_362_end_mask_0"), val = tensor([true, true, true])]; + tensor conv_out_1 = transpose(perm = conv_out_1_perm_0, x = input_31)[name = tensor("transpose_53")]; + tensor var_362 = slice_by_index(begin = var_362_begin_0, end = var_362_end_0, end_mask = var_362_end_mask_0, x = conv_out_1)[name = tensor("op_362")]; + tensor var_364_perm_0 = const()[name = tensor("op_364_perm_0"), val = tensor([1, 0, 2])]; + tensor var_364 = transpose(perm = var_364_perm_0, x = var_362)[name = tensor("transpose_52")]; + tensor input_33 = add(x = x_3, y = var_364)[name = tensor("input_33")]; + tensor input_35_axes_0 = const()[name = tensor("input_35_axes_0"), val = tensor([-1])]; + tensor input_35 = layer_norm(axes = input_35_axes_0, beta = inner_encoder_ffn2_0_module_sequential_0_bias, epsilon = var_66, gamma = inner_encoder_ffn2_0_module_sequential_0_weight, x = input_33)[name = tensor("input_35")]; + tensor inputs_9 = linear(bias = inner_encoder_ffn2_0_module_sequential_1_linear_bias, weight = inner_encoder_ffn2_0_module_sequential_1_linear_weight, x = input_35)[name = tensor("linear_8")]; + tensor input_37 = silu(x = inputs_9)[name = tensor("input_37")]; + tensor input_41 = linear(bias = inner_encoder_ffn2_0_module_sequential_4_linear_bias, weight = inner_encoder_ffn2_0_module_sequential_4_linear_weight, x = input_37)[name = tensor("linear_9")]; + tensor var_387 = const()[name = tensor("op_387"), val = tensor(0x1p-1)]; + tensor var_388 = mul(x = input_41, y = var_387)[name = tensor("op_388")]; + tensor input_43 = add(x = var_388, y = input_33)[name = tensor("input_43")]; tensor input_45_axes_0 = const()[name = tensor("input_45_axes_0"), val = tensor([-1])]; - tensor input_45 = layer_norm(axes = input_45_axes_0, beta = encoder_ffn1_1_module_sequential_0_bias, epsilon = var_28, gamma = encoder_ffn1_1_module_sequential_0_weight, x = input_43)[name = tensor("input_45")]; - tensor inputs_11 = linear(bias = encoder_ffn1_1_module_sequential_1_linear_bias, weight = encoder_ffn1_1_module_sequential_1_linear_weight, x = input_45)[name = tensor("linear_10")]; - tensor input_47 = silu(x = inputs_11)[name = tensor("input_47")]; - tensor input_51 = linear(bias = encoder_ffn1_1_module_sequential_4_linear_bias, weight = encoder_ffn1_1_module_sequential_4_linear_weight, x = input_47)[name = tensor("linear_11")]; - tensor var_361 = const()[name = tensor("op_361"), val = tensor(0x1p-1)]; - tensor var_362 = mul(x = input_51, y = var_361)[name = tensor("op_362")]; - tensor input_53 = add(x = var_362, y = input_43)[name = tensor("input_53")]; + tensor input_45 = layer_norm(axes = input_45_axes_0, beta = inner_encoder_layer_norm_0_bias, epsilon = var_66, gamma = inner_encoder_layer_norm_0_weight, x = input_43)[name = tensor("input_45")]; + tensor input_47_axes_0 = const()[name = tensor("input_47_axes_0"), val = tensor([-1])]; + tensor input_47 = layer_norm(axes = input_47_axes_0, beta = inner_encoder_ffn1_1_module_sequential_0_bias, epsilon = var_66, gamma = inner_encoder_ffn1_1_module_sequential_0_weight, x = input_45)[name = tensor("input_47")]; + tensor inputs_11 = linear(bias = inner_encoder_ffn1_1_module_sequential_1_linear_bias, weight = inner_encoder_ffn1_1_module_sequential_1_linear_weight, x = input_47)[name = tensor("linear_10")]; + tensor input_49 = silu(x = inputs_11)[name = tensor("input_49")]; + tensor input_53 = linear(bias = inner_encoder_ffn1_1_module_sequential_4_linear_bias, weight = inner_encoder_ffn1_1_module_sequential_4_linear_weight, x = input_49)[name = tensor("linear_11")]; + tensor var_417 = const()[name = tensor("op_417"), val = tensor(0x1p-1)]; + tensor var_418 = mul(x = input_53, y = var_417)[name = tensor("op_418")]; + tensor input_55 = add(x = var_418, y = input_45)[name = tensor("input_55")]; tensor x_7_axes_0 = const()[name = tensor("x_7_axes_0"), val = tensor([-1])]; - tensor x_7 = layer_norm(axes = x_7_axes_0, beta = encoder_ret_lns_1_bias, epsilon = var_28, gamma = encoder_ret_lns_1_weight, x = input_53)[name = tensor("x_7")]; + tensor x_7 = layer_norm(axes = x_7_axes_0, beta = inner_encoder_ret_lns_1_bias, epsilon = var_66, gamma = inner_encoder_ret_lns_1_weight, x = input_55)[name = tensor("x_7")]; tensor prev_kv_3_begin_0 = const()[name = tensor("prev_kv_3_begin_0"), val = tensor([1, 0, 0, 0, 0])]; tensor prev_kv_3_end_0 = const()[name = tensor("prev_kv_3_end_0"), val = tensor([2, 1, 4, 64, 64])]; tensor prev_kv_3_end_mask_0 = const()[name = tensor("prev_kv_3_end_mask_0"), val = tensor([false, true, true, true, true])]; @@ -416,173 +438,173 @@ program(1.0) tensor prev_scale_3_end_mask_0 = const()[name = tensor("prev_scale_3_end_mask_0"), val = tensor([false, true])]; tensor prev_scale_3_squeeze_mask_0 = const()[name = tensor("prev_scale_3_squeeze_mask_0"), val = tensor([true, false])]; tensor prev_scale_3 = slice_by_index(begin = prev_scale_3_begin_0, end = prev_scale_3_end_0, end_mask = prev_scale_3_end_mask_0, squeeze_mask = prev_scale_3_squeeze_mask_0, x = enc_scale)[name = tensor("prev_scale_3")]; - tensor var_376 = linear(bias = encoder_q_proj_1_bias, weight = encoder_q_proj_1_weight, x = x_7)[name = tensor("linear_12")]; - tensor var_377 = const()[name = tensor("op_377"), val = tensor([1, 4, 4, 64])]; - tensor var_378 = reshape(shape = var_377, x = var_376)[name = tensor("op_378")]; + tensor var_432 = linear(bias = inner_encoder_q_proj_1_bias, weight = inner_encoder_q_proj_1_weight, x = x_7)[name = tensor("linear_12")]; + tensor var_433 = const()[name = tensor("op_433"), val = tensor([1, 4, 4, 64])]; + tensor var_434 = reshape(shape = var_433, x = var_432)[name = tensor("op_434")]; tensor q_3_perm_0 = const()[name = tensor("q_3_perm_0"), val = tensor([0, 2, 1, 3])]; - tensor var_382 = linear(bias = encoder_k_proj_1_bias, weight = encoder_k_proj_1_weight, x = x_7)[name = tensor("linear_13")]; - tensor var_383 = const()[name = tensor("op_383"), val = tensor(0x1p-3)]; - tensor var_384 = mul(x = var_382, y = var_383)[name = tensor("op_384")]; - tensor var_385 = const()[name = tensor("op_385"), val = tensor([1, 4, 4, 64])]; - tensor var_386 = reshape(shape = var_385, x = var_384)[name = tensor("op_386")]; + tensor var_438 = linear(bias = inner_encoder_k_proj_1_bias, weight = inner_encoder_k_proj_1_weight, x = x_7)[name = tensor("linear_13")]; + tensor var_439 = const()[name = tensor("op_439"), val = tensor(0x1p-3)]; + tensor var_440 = mul(x = var_438, y = var_439)[name = tensor("op_440")]; + tensor var_441 = const()[name = tensor("op_441"), val = tensor([1, 4, 4, 64])]; + tensor var_442 = reshape(shape = var_441, x = var_440)[name = tensor("op_442")]; tensor k_3_perm_0 = const()[name = tensor("k_3_perm_0"), val = tensor([0, 2, 1, 3])]; - tensor var_390 = linear(bias = encoder_v_proj_1_bias, weight = encoder_v_proj_1_weight, x = x_7)[name = tensor("linear_14")]; - tensor var_391 = const()[name = tensor("op_391"), val = tensor([1, 4, 4, 64])]; - tensor var_392 = reshape(shape = var_391, x = var_390)[name = tensor("op_392")]; + tensor var_446 = linear(bias = inner_encoder_v_proj_1_bias, weight = inner_encoder_v_proj_1_weight, x = x_7)[name = tensor("linear_14")]; + tensor var_447 = const()[name = tensor("op_447"), val = tensor([1, 4, 4, 64])]; + tensor var_448 = reshape(shape = var_447, x = var_446)[name = tensor("op_448")]; tensor v_3_perm_0 = const()[name = tensor("v_3_perm_0"), val = tensor([0, 2, 1, 3])]; - tensor input_57 = linear(bias = encoder_g_proj_1_bias, weight = encoder_g_proj_1_weight, x = x_7)[name = tensor("linear_15")]; + tensor input_59 = linear(bias = inner_encoder_g_proj_1_bias, weight = inner_encoder_g_proj_1_weight, x = x_7)[name = tensor("linear_15")]; tensor sqrt_s0_3 = sqrt(x = prev_scale_3)[name = tensor("sqrt_s0_3")]; - tensor s_t_3 = add(x = prev_scale_3, y = encoder__t_index)[name = tensor("s_t_3")]; + tensor s_t_3 = add(x = prev_scale_3, y = inner_encoder__t_index)[name = tensor("s_t_3")]; tensor sqrt_s_t_3 = sqrt(x = s_t_3)[name = tensor("sqrt_s_t_3")]; tensor qk_3_transpose_x_1 = const()[name = tensor("qk_3_transpose_x_1"), val = tensor(false)]; tensor qk_3_transpose_y_1 = const()[name = tensor("qk_3_transpose_y_1"), val = tensor(true)]; - tensor k_3 = transpose(perm = k_3_perm_0, x = var_386)[name = tensor("transpose_50")]; - tensor q_3 = transpose(perm = q_3_perm_0, x = var_378)[name = tensor("transpose_51")]; + tensor k_3 = transpose(perm = k_3_perm_0, x = var_442)[name = tensor("transpose_50")]; + tensor q_3 = transpose(perm = q_3_perm_0, x = var_434)[name = tensor("transpose_51")]; tensor qk_3 = matmul(transpose_x = qk_3_transpose_x_1, transpose_y = qk_3_transpose_y_1, x = q_3, y = k_3)[name = tensor("qk_3")]; - tensor var_402 = const()[name = tensor("op_402"), val = tensor([4, 1])]; - tensor var_403 = reshape(shape = var_402, x = sqrt_s_t_3)[name = tensor("op_403")]; - tensor M_3 = real_div(x = encoder__causal_mask, y = var_403)[name = tensor("M_3")]; - tensor var_405 = mul(x = qk_3, y = M_3)[name = tensor("op_405")]; + tensor var_458 = const()[name = tensor("op_458"), val = tensor([4, 1])]; + tensor var_459 = reshape(shape = var_458, x = sqrt_s_t_3)[name = tensor("op_459")]; + tensor M_3 = real_div(x = inner_encoder__causal_mask, y = var_459)[name = tensor("M_3")]; + tensor var_461 = mul(x = qk_3, y = M_3)[name = tensor("op_461")]; tensor inner_3_transpose_x_0 = const()[name = tensor("inner_3_transpose_x_0"), val = tensor(false)]; tensor inner_3_transpose_y_0 = const()[name = tensor("inner_3_transpose_y_0"), val = tensor(false)]; - tensor v_3 = transpose(perm = v_3_perm_0, x = var_392)[name = tensor("transpose_49")]; - tensor inner_3 = matmul(transpose_x = inner_3_transpose_x_0, transpose_y = inner_3_transpose_y_0, x = var_405, y = v_3)[name = tensor("inner_3")]; - tensor var_407_transpose_x_0 = const()[name = tensor("op_407_transpose_x_0"), val = tensor(false)]; - tensor var_407_transpose_y_0 = const()[name = tensor("op_407_transpose_y_0"), val = tensor(false)]; - tensor var_407 = matmul(transpose_x = var_407_transpose_x_0, transpose_y = var_407_transpose_y_0, x = q_3, y = prev_kv_3)[name = tensor("op_407")]; - tensor var_408 = real_div(x = sqrt_s0_3, y = sqrt_s_t_3)[name = tensor("op_408")]; - tensor var_409 = const()[name = tensor("op_409"), val = tensor([1, 1, 4, 1])]; - tensor var_410 = reshape(shape = var_409, x = var_408)[name = tensor("op_410")]; - tensor cross_3 = mul(x = var_407, y = var_410)[name = tensor("cross_3")]; + tensor v_3 = transpose(perm = v_3_perm_0, x = var_448)[name = tensor("transpose_49")]; + tensor inner_3 = matmul(transpose_x = inner_3_transpose_x_0, transpose_y = inner_3_transpose_y_0, x = var_461, y = v_3)[name = tensor("inner_3")]; + tensor var_463_transpose_x_0 = const()[name = tensor("op_463_transpose_x_0"), val = tensor(false)]; + tensor var_463_transpose_y_0 = const()[name = tensor("op_463_transpose_y_0"), val = tensor(false)]; + tensor var_463 = matmul(transpose_x = var_463_transpose_x_0, transpose_y = var_463_transpose_y_0, x = q_3, y = prev_kv_3)[name = tensor("op_463")]; + tensor var_464 = real_div(x = sqrt_s0_3, y = sqrt_s_t_3)[name = tensor("op_464")]; + tensor var_465 = const()[name = tensor("op_465"), val = tensor([1, 1, 4, 1])]; + tensor var_466 = reshape(shape = var_465, x = var_464)[name = tensor("op_466")]; + tensor cross_3 = mul(x = var_463, y = var_466)[name = tensor("cross_3")]; tensor out_7 = add(x = inner_3, y = cross_3)[name = tensor("out_7")]; - tensor var_413 = mul(x = prev_kv_3, y = sqrt_s0_3)[name = tensor("op_413")]; - tensor var_415_transpose_x_1 = const()[name = tensor("op_415_transpose_x_1"), val = tensor(true)]; - tensor var_415_transpose_y_1 = const()[name = tensor("op_415_transpose_y_1"), val = tensor(false)]; - tensor var_415 = matmul(transpose_x = var_415_transpose_x_1, transpose_y = var_415_transpose_y_1, x = k_3, y = v_3)[name = tensor("op_415")]; - tensor new_kv_unnorm_3 = add(x = var_413, y = var_415)[name = tensor("new_kv_unnorm_3")]; - tensor var_417 = const()[name = tensor("op_417"), val = tensor(0x1p+2)]; - tensor new_scale_3 = add(x = prev_scale_3, y = var_417)[name = tensor("new_scale_3")]; - tensor var_419 = sqrt(x = new_scale_3)[name = tensor("op_419")]; - tensor var_420 = real_div(x = new_kv_unnorm_3, y = var_419)[name = tensor("op_420")]; - tensor var_421_perm_0 = const()[name = tensor("op_421_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_469 = mul(x = prev_kv_3, y = sqrt_s0_3)[name = tensor("op_469")]; + tensor var_471_transpose_x_1 = const()[name = tensor("op_471_transpose_x_1"), val = tensor(true)]; + tensor var_471_transpose_y_1 = const()[name = tensor("op_471_transpose_y_1"), val = tensor(false)]; + tensor var_471 = matmul(transpose_x = var_471_transpose_x_1, transpose_y = var_471_transpose_y_1, x = k_3, y = v_3)[name = tensor("op_471")]; + tensor new_kv_unnorm_3 = add(x = var_469, y = var_471)[name = tensor("new_kv_unnorm_3")]; + tensor var_473 = const()[name = tensor("op_473"), val = tensor(0x1p+2)]; + tensor new_scale_3 = add(x = prev_scale_3, y = var_473)[name = tensor("new_scale_3")]; + tensor var_475 = sqrt(x = new_scale_3)[name = tensor("op_475")]; + tensor var_476 = real_div(x = new_kv_unnorm_3, y = var_475)[name = tensor("op_476")]; + tensor var_477_perm_0 = const()[name = tensor("op_477_perm_0"), val = tensor([0, 2, 1, 3])]; tensor out_9_axes_0 = const()[name = tensor("out_9_axes_0"), val = tensor([-1])]; - tensor var_421 = transpose(perm = var_421_perm_0, x = out_7)[name = tensor("transpose_48")]; - tensor out_9 = layer_norm(axes = out_9_axes_0, epsilon = var_17, x = var_421)[name = tensor("out_9")]; - tensor var_425 = const()[name = tensor("op_425"), val = tensor([1, 4, 256])]; - tensor out_11 = reshape(shape = var_425, x = out_9)[name = tensor("out_11")]; - tensor var_427 = silu(x = input_57)[name = tensor("op_427")]; - tensor input_59 = mul(x = var_427, y = out_11)[name = tensor("input_59")]; - tensor ret_out_3 = linear(bias = encoder_out_proj_1_bias, weight = encoder_out_proj_1_weight, x = input_59)[name = tensor("linear_16")]; - tensor x_9 = add(x = input_53, y = ret_out_3)[name = tensor("x_9")]; + tensor var_477 = transpose(perm = var_477_perm_0, x = out_7)[name = tensor("transpose_48")]; + tensor out_9 = layer_norm(axes = out_9_axes_0, epsilon = var_74, x = var_477)[name = tensor("out_9")]; + tensor var_481 = const()[name = tensor("op_481"), val = tensor([1, 4, 256])]; + tensor out_11 = reshape(shape = var_481, x = out_9)[name = tensor("out_11")]; + tensor var_483 = silu(x = input_59)[name = tensor("op_483")]; + tensor input_61 = mul(x = var_483, y = out_11)[name = tensor("input_61")]; + tensor ret_out_3 = linear(bias = inner_encoder_out_proj_1_bias, weight = inner_encoder_out_proj_1_weight, x = input_61)[name = tensor("linear_16")]; + tensor x_9 = add(x = input_55, y = ret_out_3)[name = tensor("x_9")]; tensor window_11_begin_0 = const()[name = tensor("window_11_begin_0"), val = tensor([1, 0, 0, 0])]; tensor window_11_end_0 = const()[name = tensor("window_11_end_0"), val = tensor([2, 1, 16, 256])]; tensor window_11_end_mask_0 = const()[name = tensor("window_11_end_mask_0"), val = tensor([false, true, true, true])]; tensor window_11_squeeze_mask_0 = const()[name = tensor("window_11_squeeze_mask_0"), val = tensor([true, false, false, false])]; tensor window_11 = slice_by_index(begin = window_11_begin_0, end = window_11_end_0, end_mask = window_11_end_mask_0, squeeze_mask = window_11_squeeze_mask_0, x = enc_conv_cache)[name = tensor("window_11")]; - tensor var_435_begin_0 = const()[name = tensor("op_435_begin_0"), val = tensor([0, 0, 0])]; - tensor var_435_end_0 = const()[name = tensor("op_435_end_0"), val = tensor([1, 1, 256])]; - tensor var_435_end_mask_0 = const()[name = tensor("op_435_end_mask_0"), val = tensor([true, false, true])]; - tensor var_435 = slice_by_index(begin = var_435_begin_0, end = var_435_end_0, end_mask = var_435_end_mask_0, x = x_9)[name = tensor("op_435")]; - tensor var_438_begin_0 = const()[name = tensor("op_438_begin_0"), val = tensor([0, 1, 0])]; - tensor var_438_end_0 = const()[name = tensor("op_438_end_0"), val = tensor([1, 16, 256])]; - tensor var_438_end_mask_0 = const()[name = tensor("op_438_end_mask_0"), val = tensor([true, true, true])]; - tensor var_438 = slice_by_index(begin = var_438_begin_0, end = var_438_end_0, end_mask = var_438_end_mask_0, x = window_11)[name = tensor("op_438")]; + tensor var_491_begin_0 = const()[name = tensor("op_491_begin_0"), val = tensor([0, 0, 0])]; + tensor var_491_end_0 = const()[name = tensor("op_491_end_0"), val = tensor([1, 1, 256])]; + tensor var_491_end_mask_0 = const()[name = tensor("op_491_end_mask_0"), val = tensor([true, false, true])]; + tensor var_491 = slice_by_index(begin = var_491_begin_0, end = var_491_end_0, end_mask = var_491_end_mask_0, x = x_9)[name = tensor("op_491")]; + tensor var_494_begin_0 = const()[name = tensor("op_494_begin_0"), val = tensor([0, 1, 0])]; + tensor var_494_end_0 = const()[name = tensor("op_494_end_0"), val = tensor([1, 16, 256])]; + tensor var_494_end_mask_0 = const()[name = tensor("op_494_end_mask_0"), val = tensor([true, true, true])]; + tensor var_494 = slice_by_index(begin = var_494_begin_0, end = var_494_end_0, end_mask = var_494_end_mask_0, x = window_11)[name = tensor("op_494")]; tensor window_13_interleave_0 = const()[name = tensor("window_13_interleave_0"), val = tensor(false)]; - tensor window_13 = concat(axis = var_26, interleave = window_13_interleave_0, values = (var_438, var_435))[name = tensor("window_13")]; - tensor var_443_begin_0 = const()[name = tensor("op_443_begin_0"), val = tensor([0, 1, 0])]; - tensor var_443_end_0 = const()[name = tensor("op_443_end_0"), val = tensor([1, 2, 256])]; - tensor var_443_end_mask_0 = const()[name = tensor("op_443_end_mask_0"), val = tensor([true, false, true])]; - tensor var_443 = slice_by_index(begin = var_443_begin_0, end = var_443_end_0, end_mask = var_443_end_mask_0, x = x_9)[name = tensor("op_443")]; - tensor var_446_begin_0 = const()[name = tensor("op_446_begin_0"), val = tensor([0, 1, 0])]; - tensor var_446_end_0 = const()[name = tensor("op_446_end_0"), val = tensor([1, 16, 256])]; - tensor var_446_end_mask_0 = const()[name = tensor("op_446_end_mask_0"), val = tensor([true, true, true])]; - tensor var_446 = slice_by_index(begin = var_446_begin_0, end = var_446_end_0, end_mask = var_446_end_mask_0, x = window_13)[name = tensor("op_446")]; + tensor window_13 = concat(axis = var_82, interleave = window_13_interleave_0, values = (var_494, var_491))[name = tensor("window_13")]; + tensor var_499_begin_0 = const()[name = tensor("op_499_begin_0"), val = tensor([0, 1, 0])]; + tensor var_499_end_0 = const()[name = tensor("op_499_end_0"), val = tensor([1, 2, 256])]; + tensor var_499_end_mask_0 = const()[name = tensor("op_499_end_mask_0"), val = tensor([true, false, true])]; + tensor var_499 = slice_by_index(begin = var_499_begin_0, end = var_499_end_0, end_mask = var_499_end_mask_0, x = x_9)[name = tensor("op_499")]; + tensor var_502_begin_0 = const()[name = tensor("op_502_begin_0"), val = tensor([0, 1, 0])]; + tensor var_502_end_0 = const()[name = tensor("op_502_end_0"), val = tensor([1, 16, 256])]; + tensor var_502_end_mask_0 = const()[name = tensor("op_502_end_mask_0"), val = tensor([true, true, true])]; + tensor var_502 = slice_by_index(begin = var_502_begin_0, end = var_502_end_0, end_mask = var_502_end_mask_0, x = window_13)[name = tensor("op_502")]; tensor window_15_interleave_0 = const()[name = tensor("window_15_interleave_0"), val = tensor(false)]; - tensor window_15 = concat(axis = var_26, interleave = window_15_interleave_0, values = (var_446, var_443))[name = tensor("window_15")]; - tensor var_451_begin_0 = const()[name = tensor("op_451_begin_0"), val = tensor([0, 2, 0])]; - tensor var_451_end_0 = const()[name = tensor("op_451_end_0"), val = tensor([1, 3, 256])]; - tensor var_451_end_mask_0 = const()[name = tensor("op_451_end_mask_0"), val = tensor([true, false, true])]; - tensor var_451 = slice_by_index(begin = var_451_begin_0, end = var_451_end_0, end_mask = var_451_end_mask_0, x = x_9)[name = tensor("op_451")]; - tensor var_454_begin_0 = const()[name = tensor("op_454_begin_0"), val = tensor([0, 1, 0])]; - tensor var_454_end_0 = const()[name = tensor("op_454_end_0"), val = tensor([1, 16, 256])]; - tensor var_454_end_mask_0 = const()[name = tensor("op_454_end_mask_0"), val = tensor([true, true, true])]; - tensor var_454 = slice_by_index(begin = var_454_begin_0, end = var_454_end_0, end_mask = var_454_end_mask_0, x = window_15)[name = tensor("op_454")]; + tensor window_15 = concat(axis = var_82, interleave = window_15_interleave_0, values = (var_502, var_499))[name = tensor("window_15")]; + tensor var_507_begin_0 = const()[name = tensor("op_507_begin_0"), val = tensor([0, 2, 0])]; + tensor var_507_end_0 = const()[name = tensor("op_507_end_0"), val = tensor([1, 3, 256])]; + tensor var_507_end_mask_0 = const()[name = tensor("op_507_end_mask_0"), val = tensor([true, false, true])]; + tensor var_507 = slice_by_index(begin = var_507_begin_0, end = var_507_end_0, end_mask = var_507_end_mask_0, x = x_9)[name = tensor("op_507")]; + tensor var_510_begin_0 = const()[name = tensor("op_510_begin_0"), val = tensor([0, 1, 0])]; + tensor var_510_end_0 = const()[name = tensor("op_510_end_0"), val = tensor([1, 16, 256])]; + tensor var_510_end_mask_0 = const()[name = tensor("op_510_end_mask_0"), val = tensor([true, true, true])]; + tensor var_510 = slice_by_index(begin = var_510_begin_0, end = var_510_end_0, end_mask = var_510_end_mask_0, x = window_15)[name = tensor("op_510")]; tensor window_17_interleave_0 = const()[name = tensor("window_17_interleave_0"), val = tensor(false)]; - tensor window_17 = concat(axis = var_26, interleave = window_17_interleave_0, values = (var_454, var_451))[name = tensor("window_17")]; - tensor var_459_begin_0 = const()[name = tensor("op_459_begin_0"), val = tensor([0, 3, 0])]; - tensor var_459_end_0 = const()[name = tensor("op_459_end_0"), val = tensor([1, 1, 256])]; - tensor var_459_end_mask_0 = const()[name = tensor("op_459_end_mask_0"), val = tensor([true, true, true])]; - tensor var_459 = slice_by_index(begin = var_459_begin_0, end = var_459_end_0, end_mask = var_459_end_mask_0, x = x_9)[name = tensor("op_459")]; - tensor var_462_begin_0 = const()[name = tensor("op_462_begin_0"), val = tensor([0, 1, 0])]; - tensor var_462_end_0 = const()[name = tensor("op_462_end_0"), val = tensor([1, 16, 256])]; - tensor var_462_end_mask_0 = const()[name = tensor("op_462_end_mask_0"), val = tensor([true, true, true])]; - tensor var_462 = slice_by_index(begin = var_462_begin_0, end = var_462_end_0, end_mask = var_462_end_mask_0, x = window_17)[name = tensor("op_462")]; + tensor window_17 = concat(axis = var_82, interleave = window_17_interleave_0, values = (var_510, var_507))[name = tensor("window_17")]; + tensor var_515_begin_0 = const()[name = tensor("op_515_begin_0"), val = tensor([0, 3, 0])]; + tensor var_515_end_0 = const()[name = tensor("op_515_end_0"), val = tensor([1, 1, 256])]; + tensor var_515_end_mask_0 = const()[name = tensor("op_515_end_mask_0"), val = tensor([true, true, true])]; + tensor var_515 = slice_by_index(begin = var_515_begin_0, end = var_515_end_0, end_mask = var_515_end_mask_0, x = x_9)[name = tensor("op_515")]; + tensor var_518_begin_0 = const()[name = tensor("op_518_begin_0"), val = tensor([0, 1, 0])]; + tensor var_518_end_0 = const()[name = tensor("op_518_end_0"), val = tensor([1, 16, 256])]; + tensor var_518_end_mask_0 = const()[name = tensor("op_518_end_mask_0"), val = tensor([true, true, true])]; + tensor var_518 = slice_by_index(begin = var_518_begin_0, end = var_518_end_0, end_mask = var_518_end_mask_0, x = window_17)[name = tensor("op_518")]; tensor window_19_interleave_0 = const()[name = tensor("window_19_interleave_0"), val = tensor(false)]; - tensor window_19 = concat(axis = var_26, interleave = window_19_interleave_0, values = (var_462, var_459))[name = tensor("window_19")]; - tensor input_61_interleave_0 = const()[name = tensor("input_61_interleave_0"), val = tensor(false)]; - tensor input_61 = concat(axis = var_23, interleave = input_61_interleave_0, values = (window_13, window_15, window_17, window_19))[name = tensor("input_61")]; + tensor window_19 = concat(axis = var_82, interleave = window_19_interleave_0, values = (var_518, var_515))[name = tensor("window_19")]; + tensor input_63_interleave_0 = const()[name = tensor("input_63_interleave_0"), val = tensor(false)]; + tensor input_63 = concat(axis = var_69, interleave = input_63_interleave_0, values = (window_13, window_15, window_17, window_19))[name = tensor("input_63")]; tensor x_11_axes_0 = const()[name = tensor("x_11_axes_0"), val = tensor([-1])]; - tensor x_11 = layer_norm(axes = x_11_axes_0, beta = encoder_conv_module_1_sequential_0_bias, epsilon = var_28, gamma = encoder_conv_module_1_sequential_0_weight, x = input_61)[name = tensor("x_11")]; - tensor input_63_perm_0 = const()[name = tensor("input_63_perm_0"), val = tensor([0, 2, 1])]; + tensor x_11 = layer_norm(axes = x_11_axes_0, beta = inner_encoder_conv_module_1_sequential_0_bias, epsilon = var_66, gamma = inner_encoder_conv_module_1_sequential_0_weight, x = input_63)[name = tensor("x_11")]; + tensor input_65_perm_0 = const()[name = tensor("input_65_perm_0"), val = tensor([0, 2, 1])]; tensor inputs_13_pad_type_0 = const()[name = tensor("inputs_13_pad_type_0"), val = tensor("valid")]; tensor inputs_13_strides_0 = const()[name = tensor("inputs_13_strides_0"), val = tensor([1])]; tensor inputs_13_pad_0 = const()[name = tensor("inputs_13_pad_0"), val = tensor([0, 0])]; tensor inputs_13_dilations_0 = const()[name = tensor("inputs_13_dilations_0"), val = tensor([1])]; tensor inputs_13_groups_0 = const()[name = tensor("inputs_13_groups_0"), val = tensor(1)]; - tensor input_63 = transpose(perm = input_63_perm_0, x = x_11)[name = tensor("transpose_47")]; - tensor inputs_13 = conv(bias = encoder_conv_module_1_sequential_2_conv_bias, dilations = inputs_13_dilations_0, groups = inputs_13_groups_0, pad = inputs_13_pad_0, pad_type = inputs_13_pad_type_0, strides = inputs_13_strides_0, weight = encoder_conv_module_1_sequential_2_conv_weight, x = input_63)[name = tensor("inputs_13")]; - tensor var_487_split_sizes_0 = const()[name = tensor("op_487_split_sizes_0"), val = tensor([256, 256])]; - tensor var_487_axis_0 = const()[name = tensor("op_487_axis_0"), val = tensor(1)]; - tensor var_487_0, tensor var_487_1 = split(axis = var_487_axis_0, split_sizes = var_487_split_sizes_0, x = inputs_13)[name = tensor("op_487")]; - tensor var_489 = sigmoid(x = var_487_1)[name = tensor("op_489")]; - tensor inputs_15 = mul(x = var_487_0, y = var_489)[name = tensor("inputs_15")]; + tensor input_65 = transpose(perm = input_65_perm_0, x = x_11)[name = tensor("transpose_47")]; + tensor inputs_13 = conv(bias = inner_encoder_conv_module_1_sequential_2_conv_bias, dilations = inputs_13_dilations_0, groups = inputs_13_groups_0, pad = inputs_13_pad_0, pad_type = inputs_13_pad_type_0, strides = inputs_13_strides_0, weight = inner_encoder_conv_module_1_sequential_2_conv_weight, x = input_65)[name = tensor("inputs_13")]; + tensor var_543_split_sizes_0 = const()[name = tensor("op_543_split_sizes_0"), val = tensor([256, 256])]; + tensor var_543_axis_0 = const()[name = tensor("op_543_axis_0"), val = tensor(1)]; + tensor var_543_0, tensor var_543_1 = split(axis = var_543_axis_0, split_sizes = var_543_split_sizes_0, x = inputs_13)[name = tensor("op_543")]; + tensor var_545 = sigmoid(x = var_543_1)[name = tensor("op_545")]; + tensor inputs_15 = mul(x = var_543_0, y = var_545)[name = tensor("inputs_15")]; tensor outputs_aug_3_pad_type_0 = const()[name = tensor("outputs_aug_3_pad_type_0"), val = tensor("custom")]; tensor outputs_aug_3_pad_0 = const()[name = tensor("outputs_aug_3_pad_0"), val = tensor([15, 15])]; tensor outputs_aug_3_groups_0 = const()[name = tensor("outputs_aug_3_groups_0"), val = tensor(256)]; tensor outputs_aug_3_strides_0 = const()[name = tensor("outputs_aug_3_strides_0"), val = tensor([1])]; tensor outputs_aug_3_dilations_0 = const()[name = tensor("outputs_aug_3_dilations_0"), val = tensor([1])]; - tensor outputs_aug_3 = conv(dilations = outputs_aug_3_dilations_0, groups = outputs_aug_3_groups_0, pad = outputs_aug_3_pad_0, pad_type = outputs_aug_3_pad_type_0, strides = outputs_aug_3_strides_0, weight = encoder_conv_module_1_sequential_4_conv_weight, x = inputs_15)[name = tensor("outputs_aug_3")]; - tensor input_65_begin_0 = const()[name = tensor("input_65_begin_0"), val = tensor([0, 0, 0])]; - tensor input_65_end_0 = const()[name = tensor("input_65_end_0"), val = tensor([4, 256, 16])]; - tensor input_65_end_mask_0 = const()[name = tensor("input_65_end_mask_0"), val = tensor([true, true, false])]; - tensor input_65 = slice_by_index(begin = input_65_begin_0, end = input_65_end_0, end_mask = input_65_end_mask_0, x = outputs_aug_3)[name = tensor("input_65")]; - tensor inputs_17 = batch_norm(beta = encoder_conv_module_1_sequential_5_bias, epsilon = var_28, gamma = encoder_conv_module_1_sequential_5_weight, mean = encoder_conv_module_1_sequential_5_running_mean, variance = encoder_conv_module_1_sequential_5_running_var, x = input_65)[name = tensor("inputs_17")]; - tensor input_67 = silu(x = inputs_17)[name = tensor("input_67")]; - tensor input_69_pad_type_0 = const()[name = tensor("input_69_pad_type_0"), val = tensor("valid")]; - tensor input_69_strides_0 = const()[name = tensor("input_69_strides_0"), val = tensor([1])]; - tensor input_69_pad_0 = const()[name = tensor("input_69_pad_0"), val = tensor([0, 0])]; - tensor input_69_dilations_0 = const()[name = tensor("input_69_dilations_0"), val = tensor([1])]; - tensor input_69_groups_0 = const()[name = tensor("input_69_groups_0"), val = tensor(1)]; - tensor input_69 = conv(bias = encoder_conv_module_1_sequential_7_conv_bias, dilations = input_69_dilations_0, groups = input_69_groups_0, pad = input_69_pad_0, pad_type = input_69_pad_type_0, strides = input_69_strides_0, weight = encoder_conv_module_1_sequential_7_conv_weight, x = input_67)[name = tensor("input_69")]; + tensor outputs_aug_3 = conv(dilations = outputs_aug_3_dilations_0, groups = outputs_aug_3_groups_0, pad = outputs_aug_3_pad_0, pad_type = outputs_aug_3_pad_type_0, strides = outputs_aug_3_strides_0, weight = inner_encoder_conv_module_1_sequential_4_conv_weight, x = inputs_15)[name = tensor("outputs_aug_3")]; + tensor input_67_begin_0 = const()[name = tensor("input_67_begin_0"), val = tensor([0, 0, 0])]; + tensor input_67_end_0 = const()[name = tensor("input_67_end_0"), val = tensor([4, 256, 16])]; + tensor input_67_end_mask_0 = const()[name = tensor("input_67_end_mask_0"), val = tensor([true, true, false])]; + tensor input_67 = slice_by_index(begin = input_67_begin_0, end = input_67_end_0, end_mask = input_67_end_mask_0, x = outputs_aug_3)[name = tensor("input_67")]; + tensor inputs_17 = batch_norm(beta = inner_encoder_conv_module_1_sequential_5_bias, epsilon = var_66, gamma = inner_encoder_conv_module_1_sequential_5_weight, mean = inner_encoder_conv_module_1_sequential_5_running_mean, variance = inner_encoder_conv_module_1_sequential_5_running_var, x = input_67)[name = tensor("inputs_17")]; + tensor input_69 = silu(x = inputs_17)[name = tensor("input_69")]; + tensor input_71_pad_type_0 = const()[name = tensor("input_71_pad_type_0"), val = tensor("valid")]; + tensor input_71_strides_0 = const()[name = tensor("input_71_strides_0"), val = tensor([1])]; + tensor input_71_pad_0 = const()[name = tensor("input_71_pad_0"), val = tensor([0, 0])]; + tensor input_71_dilations_0 = const()[name = tensor("input_71_dilations_0"), val = tensor([1])]; + tensor input_71_groups_0 = const()[name = tensor("input_71_groups_0"), val = tensor(1)]; + tensor input_71 = conv(bias = inner_encoder_conv_module_1_sequential_7_conv_bias, dilations = input_71_dilations_0, groups = input_71_groups_0, pad = input_71_pad_0, pad_type = input_71_pad_type_0, strides = input_71_strides_0, weight = inner_encoder_conv_module_1_sequential_7_conv_weight, x = input_69)[name = tensor("input_71")]; tensor conv_out_3_perm_0 = const()[name = tensor("conv_out_3_perm_0"), val = tensor([0, 2, 1])]; - tensor var_520_begin_0 = const()[name = tensor("op_520_begin_0"), val = tensor([0, -1, 0])]; - tensor var_520_end_0 = const()[name = tensor("op_520_end_0"), val = tensor([4, 16, 256])]; - tensor var_520_end_mask_0 = const()[name = tensor("op_520_end_mask_0"), val = tensor([true, true, true])]; - tensor conv_out_3 = transpose(perm = conv_out_3_perm_0, x = input_69)[name = tensor("transpose_46")]; - tensor var_520 = slice_by_index(begin = var_520_begin_0, end = var_520_end_0, end_mask = var_520_end_mask_0, x = conv_out_3)[name = tensor("op_520")]; - tensor var_522_perm_0 = const()[name = tensor("op_522_perm_0"), val = tensor([1, 0, 2])]; - tensor var_522 = transpose(perm = var_522_perm_0, x = var_520)[name = tensor("transpose_45")]; - tensor input_71 = add(x = x_9, y = var_522)[name = tensor("input_71")]; - tensor input_73_axes_0 = const()[name = tensor("input_73_axes_0"), val = tensor([-1])]; - tensor input_73 = layer_norm(axes = input_73_axes_0, beta = encoder_ffn2_1_module_sequential_0_bias, epsilon = var_28, gamma = encoder_ffn2_1_module_sequential_0_weight, x = input_71)[name = tensor("input_73")]; - tensor inputs_19 = linear(bias = encoder_ffn2_1_module_sequential_1_linear_bias, weight = encoder_ffn2_1_module_sequential_1_linear_weight, x = input_73)[name = tensor("linear_17")]; - tensor input_75 = silu(x = inputs_19)[name = tensor("input_75")]; - tensor input_79 = linear(bias = encoder_ffn2_1_module_sequential_4_linear_bias, weight = encoder_ffn2_1_module_sequential_4_linear_weight, x = input_75)[name = tensor("linear_18")]; - tensor var_545 = const()[name = tensor("op_545"), val = tensor(0x1p-1)]; - tensor var_546 = mul(x = input_79, y = var_545)[name = tensor("op_546")]; - tensor input_81 = add(x = var_546, y = input_71)[name = tensor("input_81")]; - tensor input_83_axes_0 = const()[name = tensor("input_83_axes_0"), val = tensor([-1])]; - tensor input_83 = layer_norm(axes = input_83_axes_0, beta = encoder_layer_norm_1_bias, epsilon = var_28, gamma = encoder_layer_norm_1_weight, x = input_81)[name = tensor("input_83")]; + tensor var_576_begin_0 = const()[name = tensor("op_576_begin_0"), val = tensor([0, -1, 0])]; + tensor var_576_end_0 = const()[name = tensor("op_576_end_0"), val = tensor([4, 16, 256])]; + tensor var_576_end_mask_0 = const()[name = tensor("op_576_end_mask_0"), val = tensor([true, true, true])]; + tensor conv_out_3 = transpose(perm = conv_out_3_perm_0, x = input_71)[name = tensor("transpose_46")]; + tensor var_576 = slice_by_index(begin = var_576_begin_0, end = var_576_end_0, end_mask = var_576_end_mask_0, x = conv_out_3)[name = tensor("op_576")]; + tensor var_578_perm_0 = const()[name = tensor("op_578_perm_0"), val = tensor([1, 0, 2])]; + tensor var_578 = transpose(perm = var_578_perm_0, x = var_576)[name = tensor("transpose_45")]; + tensor input_73 = add(x = x_9, y = var_578)[name = tensor("input_73")]; + tensor input_75_axes_0 = const()[name = tensor("input_75_axes_0"), val = tensor([-1])]; + tensor input_75 = layer_norm(axes = input_75_axes_0, beta = inner_encoder_ffn2_1_module_sequential_0_bias, epsilon = var_66, gamma = inner_encoder_ffn2_1_module_sequential_0_weight, x = input_73)[name = tensor("input_75")]; + tensor inputs_19 = linear(bias = inner_encoder_ffn2_1_module_sequential_1_linear_bias, weight = inner_encoder_ffn2_1_module_sequential_1_linear_weight, x = input_75)[name = tensor("linear_17")]; + tensor input_77 = silu(x = inputs_19)[name = tensor("input_77")]; + tensor input_81 = linear(bias = inner_encoder_ffn2_1_module_sequential_4_linear_bias, weight = inner_encoder_ffn2_1_module_sequential_4_linear_weight, x = input_77)[name = tensor("linear_18")]; + tensor var_601 = const()[name = tensor("op_601"), val = tensor(0x1p-1)]; + tensor var_602 = mul(x = input_81, y = var_601)[name = tensor("op_602")]; + tensor input_83 = add(x = var_602, y = input_73)[name = tensor("input_83")]; tensor input_85_axes_0 = const()[name = tensor("input_85_axes_0"), val = tensor([-1])]; - tensor input_85 = layer_norm(axes = input_85_axes_0, beta = encoder_ffn1_2_module_sequential_0_bias, epsilon = var_28, gamma = encoder_ffn1_2_module_sequential_0_weight, x = input_83)[name = tensor("input_85")]; - tensor inputs_21 = linear(bias = encoder_ffn1_2_module_sequential_1_linear_bias, weight = encoder_ffn1_2_module_sequential_1_linear_weight, x = input_85)[name = tensor("linear_19")]; - tensor input_87 = silu(x = inputs_21)[name = tensor("input_87")]; - tensor input_91 = linear(bias = encoder_ffn1_2_module_sequential_4_linear_bias, weight = encoder_ffn1_2_module_sequential_4_linear_weight, x = input_87)[name = tensor("linear_20")]; - tensor var_575 = const()[name = tensor("op_575"), val = tensor(0x1p-1)]; - tensor var_576 = mul(x = input_91, y = var_575)[name = tensor("op_576")]; - tensor input_93 = add(x = var_576, y = input_83)[name = tensor("input_93")]; + tensor input_85 = layer_norm(axes = input_85_axes_0, beta = inner_encoder_layer_norm_1_bias, epsilon = var_66, gamma = inner_encoder_layer_norm_1_weight, x = input_83)[name = tensor("input_85")]; + tensor input_87_axes_0 = const()[name = tensor("input_87_axes_0"), val = tensor([-1])]; + tensor input_87 = layer_norm(axes = input_87_axes_0, beta = inner_encoder_ffn1_2_module_sequential_0_bias, epsilon = var_66, gamma = inner_encoder_ffn1_2_module_sequential_0_weight, x = input_85)[name = tensor("input_87")]; + tensor inputs_21 = linear(bias = inner_encoder_ffn1_2_module_sequential_1_linear_bias, weight = inner_encoder_ffn1_2_module_sequential_1_linear_weight, x = input_87)[name = tensor("linear_19")]; + tensor input_89 = silu(x = inputs_21)[name = tensor("input_89")]; + tensor input_93 = linear(bias = inner_encoder_ffn1_2_module_sequential_4_linear_bias, weight = inner_encoder_ffn1_2_module_sequential_4_linear_weight, x = input_89)[name = tensor("linear_20")]; + tensor var_631 = const()[name = tensor("op_631"), val = tensor(0x1p-1)]; + tensor var_632 = mul(x = input_93, y = var_631)[name = tensor("op_632")]; + tensor input_95 = add(x = var_632, y = input_85)[name = tensor("input_95")]; tensor x_13_axes_0 = const()[name = tensor("x_13_axes_0"), val = tensor([-1])]; - tensor x_13 = layer_norm(axes = x_13_axes_0, beta = encoder_ret_lns_2_bias, epsilon = var_28, gamma = encoder_ret_lns_2_weight, x = input_93)[name = tensor("x_13")]; + tensor x_13 = layer_norm(axes = x_13_axes_0, beta = inner_encoder_ret_lns_2_bias, epsilon = var_66, gamma = inner_encoder_ret_lns_2_weight, x = input_95)[name = tensor("x_13")]; tensor prev_kv_5_begin_0 = const()[name = tensor("prev_kv_5_begin_0"), val = tensor([2, 0, 0, 0, 0])]; tensor prev_kv_5_end_0 = const()[name = tensor("prev_kv_5_end_0"), val = tensor([3, 1, 4, 64, 64])]; tensor prev_kv_5_end_mask_0 = const()[name = tensor("prev_kv_5_end_mask_0"), val = tensor([false, true, true, true, true])]; @@ -593,173 +615,173 @@ program(1.0) tensor prev_scale_5_end_mask_0 = const()[name = tensor("prev_scale_5_end_mask_0"), val = tensor([false, true])]; tensor prev_scale_5_squeeze_mask_0 = const()[name = tensor("prev_scale_5_squeeze_mask_0"), val = tensor([true, false])]; tensor prev_scale_5 = slice_by_index(begin = prev_scale_5_begin_0, end = prev_scale_5_end_0, end_mask = prev_scale_5_end_mask_0, squeeze_mask = prev_scale_5_squeeze_mask_0, x = enc_scale)[name = tensor("prev_scale_5")]; - tensor var_590 = linear(bias = encoder_q_proj_2_bias, weight = encoder_q_proj_2_weight, x = x_13)[name = tensor("linear_21")]; - tensor var_591 = const()[name = tensor("op_591"), val = tensor([1, 4, 4, 64])]; - tensor var_592 = reshape(shape = var_591, x = var_590)[name = tensor("op_592")]; + tensor var_646 = linear(bias = inner_encoder_q_proj_2_bias, weight = inner_encoder_q_proj_2_weight, x = x_13)[name = tensor("linear_21")]; + tensor var_647 = const()[name = tensor("op_647"), val = tensor([1, 4, 4, 64])]; + tensor var_648 = reshape(shape = var_647, x = var_646)[name = tensor("op_648")]; tensor q_5_perm_0 = const()[name = tensor("q_5_perm_0"), val = tensor([0, 2, 1, 3])]; - tensor var_596 = linear(bias = encoder_k_proj_2_bias, weight = encoder_k_proj_2_weight, x = x_13)[name = tensor("linear_22")]; - tensor var_597 = const()[name = tensor("op_597"), val = tensor(0x1p-3)]; - tensor var_598 = mul(x = var_596, y = var_597)[name = tensor("op_598")]; - tensor var_599 = const()[name = tensor("op_599"), val = tensor([1, 4, 4, 64])]; - tensor var_600 = reshape(shape = var_599, x = var_598)[name = tensor("op_600")]; + tensor var_652 = linear(bias = inner_encoder_k_proj_2_bias, weight = inner_encoder_k_proj_2_weight, x = x_13)[name = tensor("linear_22")]; + tensor var_653 = const()[name = tensor("op_653"), val = tensor(0x1p-3)]; + tensor var_654 = mul(x = var_652, y = var_653)[name = tensor("op_654")]; + tensor var_655 = const()[name = tensor("op_655"), val = tensor([1, 4, 4, 64])]; + tensor var_656 = reshape(shape = var_655, x = var_654)[name = tensor("op_656")]; tensor k_5_perm_0 = const()[name = tensor("k_5_perm_0"), val = tensor([0, 2, 1, 3])]; - tensor var_604 = linear(bias = encoder_v_proj_2_bias, weight = encoder_v_proj_2_weight, x = x_13)[name = tensor("linear_23")]; - tensor var_605 = const()[name = tensor("op_605"), val = tensor([1, 4, 4, 64])]; - tensor var_606 = reshape(shape = var_605, x = var_604)[name = tensor("op_606")]; + tensor var_660 = linear(bias = inner_encoder_v_proj_2_bias, weight = inner_encoder_v_proj_2_weight, x = x_13)[name = tensor("linear_23")]; + tensor var_661 = const()[name = tensor("op_661"), val = tensor([1, 4, 4, 64])]; + tensor var_662 = reshape(shape = var_661, x = var_660)[name = tensor("op_662")]; tensor v_5_perm_0 = const()[name = tensor("v_5_perm_0"), val = tensor([0, 2, 1, 3])]; - tensor input_97 = linear(bias = encoder_g_proj_2_bias, weight = encoder_g_proj_2_weight, x = x_13)[name = tensor("linear_24")]; + tensor input_99 = linear(bias = inner_encoder_g_proj_2_bias, weight = inner_encoder_g_proj_2_weight, x = x_13)[name = tensor("linear_24")]; tensor sqrt_s0_5 = sqrt(x = prev_scale_5)[name = tensor("sqrt_s0_5")]; - tensor s_t_5 = add(x = prev_scale_5, y = encoder__t_index)[name = tensor("s_t_5")]; + tensor s_t_5 = add(x = prev_scale_5, y = inner_encoder__t_index)[name = tensor("s_t_5")]; tensor sqrt_s_t_5 = sqrt(x = s_t_5)[name = tensor("sqrt_s_t_5")]; tensor qk_5_transpose_x_1 = const()[name = tensor("qk_5_transpose_x_1"), val = tensor(false)]; tensor qk_5_transpose_y_1 = const()[name = tensor("qk_5_transpose_y_1"), val = tensor(true)]; - tensor k_5 = transpose(perm = k_5_perm_0, x = var_600)[name = tensor("transpose_43")]; - tensor q_5 = transpose(perm = q_5_perm_0, x = var_592)[name = tensor("transpose_44")]; + tensor k_5 = transpose(perm = k_5_perm_0, x = var_656)[name = tensor("transpose_43")]; + tensor q_5 = transpose(perm = q_5_perm_0, x = var_648)[name = tensor("transpose_44")]; tensor qk_5 = matmul(transpose_x = qk_5_transpose_x_1, transpose_y = qk_5_transpose_y_1, x = q_5, y = k_5)[name = tensor("qk_5")]; - tensor var_616 = const()[name = tensor("op_616"), val = tensor([4, 1])]; - tensor var_617 = reshape(shape = var_616, x = sqrt_s_t_5)[name = tensor("op_617")]; - tensor M_5 = real_div(x = encoder__causal_mask, y = var_617)[name = tensor("M_5")]; - tensor var_619 = mul(x = qk_5, y = M_5)[name = tensor("op_619")]; + tensor var_672 = const()[name = tensor("op_672"), val = tensor([4, 1])]; + tensor var_673 = reshape(shape = var_672, x = sqrt_s_t_5)[name = tensor("op_673")]; + tensor M_5 = real_div(x = inner_encoder__causal_mask, y = var_673)[name = tensor("M_5")]; + tensor var_675 = mul(x = qk_5, y = M_5)[name = tensor("op_675")]; tensor inner_5_transpose_x_0 = const()[name = tensor("inner_5_transpose_x_0"), val = tensor(false)]; tensor inner_5_transpose_y_0 = const()[name = tensor("inner_5_transpose_y_0"), val = tensor(false)]; - tensor v_5 = transpose(perm = v_5_perm_0, x = var_606)[name = tensor("transpose_42")]; - tensor inner_5 = matmul(transpose_x = inner_5_transpose_x_0, transpose_y = inner_5_transpose_y_0, x = var_619, y = v_5)[name = tensor("inner_5")]; - tensor var_621_transpose_x_0 = const()[name = tensor("op_621_transpose_x_0"), val = tensor(false)]; - tensor var_621_transpose_y_0 = const()[name = tensor("op_621_transpose_y_0"), val = tensor(false)]; - tensor var_621 = matmul(transpose_x = var_621_transpose_x_0, transpose_y = var_621_transpose_y_0, x = q_5, y = prev_kv_5)[name = tensor("op_621")]; - tensor var_622 = real_div(x = sqrt_s0_5, y = sqrt_s_t_5)[name = tensor("op_622")]; - tensor var_623 = const()[name = tensor("op_623"), val = tensor([1, 1, 4, 1])]; - tensor var_624 = reshape(shape = var_623, x = var_622)[name = tensor("op_624")]; - tensor cross_5 = mul(x = var_621, y = var_624)[name = tensor("cross_5")]; + tensor v_5 = transpose(perm = v_5_perm_0, x = var_662)[name = tensor("transpose_42")]; + tensor inner_5 = matmul(transpose_x = inner_5_transpose_x_0, transpose_y = inner_5_transpose_y_0, x = var_675, y = v_5)[name = tensor("inner_5")]; + tensor var_677_transpose_x_0 = const()[name = tensor("op_677_transpose_x_0"), val = tensor(false)]; + tensor var_677_transpose_y_0 = const()[name = tensor("op_677_transpose_y_0"), val = tensor(false)]; + tensor var_677 = matmul(transpose_x = var_677_transpose_x_0, transpose_y = var_677_transpose_y_0, x = q_5, y = prev_kv_5)[name = tensor("op_677")]; + tensor var_678 = real_div(x = sqrt_s0_5, y = sqrt_s_t_5)[name = tensor("op_678")]; + tensor var_679 = const()[name = tensor("op_679"), val = tensor([1, 1, 4, 1])]; + tensor var_680 = reshape(shape = var_679, x = var_678)[name = tensor("op_680")]; + tensor cross_5 = mul(x = var_677, y = var_680)[name = tensor("cross_5")]; tensor out_13 = add(x = inner_5, y = cross_5)[name = tensor("out_13")]; - tensor var_627 = mul(x = prev_kv_5, y = sqrt_s0_5)[name = tensor("op_627")]; - tensor var_629_transpose_x_1 = const()[name = tensor("op_629_transpose_x_1"), val = tensor(true)]; - tensor var_629_transpose_y_1 = const()[name = tensor("op_629_transpose_y_1"), val = tensor(false)]; - tensor var_629 = matmul(transpose_x = var_629_transpose_x_1, transpose_y = var_629_transpose_y_1, x = k_5, y = v_5)[name = tensor("op_629")]; - tensor new_kv_unnorm_5 = add(x = var_627, y = var_629)[name = tensor("new_kv_unnorm_5")]; - tensor var_631 = const()[name = tensor("op_631"), val = tensor(0x1p+2)]; - tensor new_scale_5 = add(x = prev_scale_5, y = var_631)[name = tensor("new_scale_5")]; - tensor var_633 = sqrt(x = new_scale_5)[name = tensor("op_633")]; - tensor var_634 = real_div(x = new_kv_unnorm_5, y = var_633)[name = tensor("op_634")]; - tensor var_635_perm_0 = const()[name = tensor("op_635_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_683 = mul(x = prev_kv_5, y = sqrt_s0_5)[name = tensor("op_683")]; + tensor var_685_transpose_x_1 = const()[name = tensor("op_685_transpose_x_1"), val = tensor(true)]; + tensor var_685_transpose_y_1 = const()[name = tensor("op_685_transpose_y_1"), val = tensor(false)]; + tensor var_685 = matmul(transpose_x = var_685_transpose_x_1, transpose_y = var_685_transpose_y_1, x = k_5, y = v_5)[name = tensor("op_685")]; + tensor new_kv_unnorm_5 = add(x = var_683, y = var_685)[name = tensor("new_kv_unnorm_5")]; + tensor var_687 = const()[name = tensor("op_687"), val = tensor(0x1p+2)]; + tensor new_scale_5 = add(x = prev_scale_5, y = var_687)[name = tensor("new_scale_5")]; + tensor var_689 = sqrt(x = new_scale_5)[name = tensor("op_689")]; + tensor var_690 = real_div(x = new_kv_unnorm_5, y = var_689)[name = tensor("op_690")]; + tensor var_691_perm_0 = const()[name = tensor("op_691_perm_0"), val = tensor([0, 2, 1, 3])]; tensor out_15_axes_0 = const()[name = tensor("out_15_axes_0"), val = tensor([-1])]; - tensor var_635 = transpose(perm = var_635_perm_0, x = out_13)[name = tensor("transpose_41")]; - tensor out_15 = layer_norm(axes = out_15_axes_0, epsilon = var_17, x = var_635)[name = tensor("out_15")]; - tensor var_639 = const()[name = tensor("op_639"), val = tensor([1, 4, 256])]; - tensor out_17 = reshape(shape = var_639, x = out_15)[name = tensor("out_17")]; - tensor var_641 = silu(x = input_97)[name = tensor("op_641")]; - tensor input_99 = mul(x = var_641, y = out_17)[name = tensor("input_99")]; - tensor ret_out_5 = linear(bias = encoder_out_proj_2_bias, weight = encoder_out_proj_2_weight, x = input_99)[name = tensor("linear_25")]; - tensor x_15 = add(x = input_93, y = ret_out_5)[name = tensor("x_15")]; + tensor var_691 = transpose(perm = var_691_perm_0, x = out_13)[name = tensor("transpose_41")]; + tensor out_15 = layer_norm(axes = out_15_axes_0, epsilon = var_74, x = var_691)[name = tensor("out_15")]; + tensor var_695 = const()[name = tensor("op_695"), val = tensor([1, 4, 256])]; + tensor out_17 = reshape(shape = var_695, x = out_15)[name = tensor("out_17")]; + tensor var_697 = silu(x = input_99)[name = tensor("op_697")]; + tensor input_101 = mul(x = var_697, y = out_17)[name = tensor("input_101")]; + tensor ret_out_5 = linear(bias = inner_encoder_out_proj_2_bias, weight = inner_encoder_out_proj_2_weight, x = input_101)[name = tensor("linear_25")]; + tensor x_15 = add(x = input_95, y = ret_out_5)[name = tensor("x_15")]; tensor window_21_begin_0 = const()[name = tensor("window_21_begin_0"), val = tensor([2, 0, 0, 0])]; tensor window_21_end_0 = const()[name = tensor("window_21_end_0"), val = tensor([3, 1, 16, 256])]; tensor window_21_end_mask_0 = const()[name = tensor("window_21_end_mask_0"), val = tensor([false, true, true, true])]; tensor window_21_squeeze_mask_0 = const()[name = tensor("window_21_squeeze_mask_0"), val = tensor([true, false, false, false])]; tensor window_21 = slice_by_index(begin = window_21_begin_0, end = window_21_end_0, end_mask = window_21_end_mask_0, squeeze_mask = window_21_squeeze_mask_0, x = enc_conv_cache)[name = tensor("window_21")]; - tensor var_649_begin_0 = const()[name = tensor("op_649_begin_0"), val = tensor([0, 0, 0])]; - tensor var_649_end_0 = const()[name = tensor("op_649_end_0"), val = tensor([1, 1, 256])]; - tensor var_649_end_mask_0 = const()[name = tensor("op_649_end_mask_0"), val = tensor([true, false, true])]; - tensor var_649 = slice_by_index(begin = var_649_begin_0, end = var_649_end_0, end_mask = var_649_end_mask_0, x = x_15)[name = tensor("op_649")]; - tensor var_652_begin_0 = const()[name = tensor("op_652_begin_0"), val = tensor([0, 1, 0])]; - tensor var_652_end_0 = const()[name = tensor("op_652_end_0"), val = tensor([1, 16, 256])]; - tensor var_652_end_mask_0 = const()[name = tensor("op_652_end_mask_0"), val = tensor([true, true, true])]; - tensor var_652 = slice_by_index(begin = var_652_begin_0, end = var_652_end_0, end_mask = var_652_end_mask_0, x = window_21)[name = tensor("op_652")]; + tensor var_705_begin_0 = const()[name = tensor("op_705_begin_0"), val = tensor([0, 0, 0])]; + tensor var_705_end_0 = const()[name = tensor("op_705_end_0"), val = tensor([1, 1, 256])]; + tensor var_705_end_mask_0 = const()[name = tensor("op_705_end_mask_0"), val = tensor([true, false, true])]; + tensor var_705 = slice_by_index(begin = var_705_begin_0, end = var_705_end_0, end_mask = var_705_end_mask_0, x = x_15)[name = tensor("op_705")]; + tensor var_708_begin_0 = const()[name = tensor("op_708_begin_0"), val = tensor([0, 1, 0])]; + tensor var_708_end_0 = const()[name = tensor("op_708_end_0"), val = tensor([1, 16, 256])]; + tensor var_708_end_mask_0 = const()[name = tensor("op_708_end_mask_0"), val = tensor([true, true, true])]; + tensor var_708 = slice_by_index(begin = var_708_begin_0, end = var_708_end_0, end_mask = var_708_end_mask_0, x = window_21)[name = tensor("op_708")]; tensor window_23_interleave_0 = const()[name = tensor("window_23_interleave_0"), val = tensor(false)]; - tensor window_23 = concat(axis = var_26, interleave = window_23_interleave_0, values = (var_652, var_649))[name = tensor("window_23")]; - tensor var_657_begin_0 = const()[name = tensor("op_657_begin_0"), val = tensor([0, 1, 0])]; - tensor var_657_end_0 = const()[name = tensor("op_657_end_0"), val = tensor([1, 2, 256])]; - tensor var_657_end_mask_0 = const()[name = tensor("op_657_end_mask_0"), val = tensor([true, false, true])]; - tensor var_657 = slice_by_index(begin = var_657_begin_0, end = var_657_end_0, end_mask = var_657_end_mask_0, x = x_15)[name = tensor("op_657")]; - tensor var_660_begin_0 = const()[name = tensor("op_660_begin_0"), val = tensor([0, 1, 0])]; - tensor var_660_end_0 = const()[name = tensor("op_660_end_0"), val = tensor([1, 16, 256])]; - tensor var_660_end_mask_0 = const()[name = tensor("op_660_end_mask_0"), val = tensor([true, true, true])]; - tensor var_660 = slice_by_index(begin = var_660_begin_0, end = var_660_end_0, end_mask = var_660_end_mask_0, x = window_23)[name = tensor("op_660")]; + tensor window_23 = concat(axis = var_82, interleave = window_23_interleave_0, values = (var_708, var_705))[name = tensor("window_23")]; + tensor var_713_begin_0 = const()[name = tensor("op_713_begin_0"), val = tensor([0, 1, 0])]; + tensor var_713_end_0 = const()[name = tensor("op_713_end_0"), val = tensor([1, 2, 256])]; + tensor var_713_end_mask_0 = const()[name = tensor("op_713_end_mask_0"), val = tensor([true, false, true])]; + tensor var_713 = slice_by_index(begin = var_713_begin_0, end = var_713_end_0, end_mask = var_713_end_mask_0, x = x_15)[name = tensor("op_713")]; + tensor var_716_begin_0 = const()[name = tensor("op_716_begin_0"), val = tensor([0, 1, 0])]; + tensor var_716_end_0 = const()[name = tensor("op_716_end_0"), val = tensor([1, 16, 256])]; + tensor var_716_end_mask_0 = const()[name = tensor("op_716_end_mask_0"), val = tensor([true, true, true])]; + tensor var_716 = slice_by_index(begin = var_716_begin_0, end = var_716_end_0, end_mask = var_716_end_mask_0, x = window_23)[name = tensor("op_716")]; tensor window_25_interleave_0 = const()[name = tensor("window_25_interleave_0"), val = tensor(false)]; - tensor window_25 = concat(axis = var_26, interleave = window_25_interleave_0, values = (var_660, var_657))[name = tensor("window_25")]; - tensor var_665_begin_0 = const()[name = tensor("op_665_begin_0"), val = tensor([0, 2, 0])]; - tensor var_665_end_0 = const()[name = tensor("op_665_end_0"), val = tensor([1, 3, 256])]; - tensor var_665_end_mask_0 = const()[name = tensor("op_665_end_mask_0"), val = tensor([true, false, true])]; - tensor var_665 = slice_by_index(begin = var_665_begin_0, end = var_665_end_0, end_mask = var_665_end_mask_0, x = x_15)[name = tensor("op_665")]; - tensor var_668_begin_0 = const()[name = tensor("op_668_begin_0"), val = tensor([0, 1, 0])]; - tensor var_668_end_0 = const()[name = tensor("op_668_end_0"), val = tensor([1, 16, 256])]; - tensor var_668_end_mask_0 = const()[name = tensor("op_668_end_mask_0"), val = tensor([true, true, true])]; - tensor var_668 = slice_by_index(begin = var_668_begin_0, end = var_668_end_0, end_mask = var_668_end_mask_0, x = window_25)[name = tensor("op_668")]; + tensor window_25 = concat(axis = var_82, interleave = window_25_interleave_0, values = (var_716, var_713))[name = tensor("window_25")]; + tensor var_721_begin_0 = const()[name = tensor("op_721_begin_0"), val = tensor([0, 2, 0])]; + tensor var_721_end_0 = const()[name = tensor("op_721_end_0"), val = tensor([1, 3, 256])]; + tensor var_721_end_mask_0 = const()[name = tensor("op_721_end_mask_0"), val = tensor([true, false, true])]; + tensor var_721 = slice_by_index(begin = var_721_begin_0, end = var_721_end_0, end_mask = var_721_end_mask_0, x = x_15)[name = tensor("op_721")]; + tensor var_724_begin_0 = const()[name = tensor("op_724_begin_0"), val = tensor([0, 1, 0])]; + tensor var_724_end_0 = const()[name = tensor("op_724_end_0"), val = tensor([1, 16, 256])]; + tensor var_724_end_mask_0 = const()[name = tensor("op_724_end_mask_0"), val = tensor([true, true, true])]; + tensor var_724 = slice_by_index(begin = var_724_begin_0, end = var_724_end_0, end_mask = var_724_end_mask_0, x = window_25)[name = tensor("op_724")]; tensor window_27_interleave_0 = const()[name = tensor("window_27_interleave_0"), val = tensor(false)]; - tensor window_27 = concat(axis = var_26, interleave = window_27_interleave_0, values = (var_668, var_665))[name = tensor("window_27")]; - tensor var_673_begin_0 = const()[name = tensor("op_673_begin_0"), val = tensor([0, 3, 0])]; - tensor var_673_end_0 = const()[name = tensor("op_673_end_0"), val = tensor([1, 1, 256])]; - tensor var_673_end_mask_0 = const()[name = tensor("op_673_end_mask_0"), val = tensor([true, true, true])]; - tensor var_673 = slice_by_index(begin = var_673_begin_0, end = var_673_end_0, end_mask = var_673_end_mask_0, x = x_15)[name = tensor("op_673")]; - tensor var_676_begin_0 = const()[name = tensor("op_676_begin_0"), val = tensor([0, 1, 0])]; - tensor var_676_end_0 = const()[name = tensor("op_676_end_0"), val = tensor([1, 16, 256])]; - tensor var_676_end_mask_0 = const()[name = tensor("op_676_end_mask_0"), val = tensor([true, true, true])]; - tensor var_676 = slice_by_index(begin = var_676_begin_0, end = var_676_end_0, end_mask = var_676_end_mask_0, x = window_27)[name = tensor("op_676")]; + tensor window_27 = concat(axis = var_82, interleave = window_27_interleave_0, values = (var_724, var_721))[name = tensor("window_27")]; + tensor var_729_begin_0 = const()[name = tensor("op_729_begin_0"), val = tensor([0, 3, 0])]; + tensor var_729_end_0 = const()[name = tensor("op_729_end_0"), val = tensor([1, 1, 256])]; + tensor var_729_end_mask_0 = const()[name = tensor("op_729_end_mask_0"), val = tensor([true, true, true])]; + tensor var_729 = slice_by_index(begin = var_729_begin_0, end = var_729_end_0, end_mask = var_729_end_mask_0, x = x_15)[name = tensor("op_729")]; + tensor var_732_begin_0 = const()[name = tensor("op_732_begin_0"), val = tensor([0, 1, 0])]; + tensor var_732_end_0 = const()[name = tensor("op_732_end_0"), val = tensor([1, 16, 256])]; + tensor var_732_end_mask_0 = const()[name = tensor("op_732_end_mask_0"), val = tensor([true, true, true])]; + tensor var_732 = slice_by_index(begin = var_732_begin_0, end = var_732_end_0, end_mask = var_732_end_mask_0, x = window_27)[name = tensor("op_732")]; tensor window_29_interleave_0 = const()[name = tensor("window_29_interleave_0"), val = tensor(false)]; - tensor window_29 = concat(axis = var_26, interleave = window_29_interleave_0, values = (var_676, var_673))[name = tensor("window_29")]; - tensor input_101_interleave_0 = const()[name = tensor("input_101_interleave_0"), val = tensor(false)]; - tensor input_101 = concat(axis = var_23, interleave = input_101_interleave_0, values = (window_23, window_25, window_27, window_29))[name = tensor("input_101")]; + tensor window_29 = concat(axis = var_82, interleave = window_29_interleave_0, values = (var_732, var_729))[name = tensor("window_29")]; + tensor input_103_interleave_0 = const()[name = tensor("input_103_interleave_0"), val = tensor(false)]; + tensor input_103 = concat(axis = var_69, interleave = input_103_interleave_0, values = (window_23, window_25, window_27, window_29))[name = tensor("input_103")]; tensor x_17_axes_0 = const()[name = tensor("x_17_axes_0"), val = tensor([-1])]; - tensor x_17 = layer_norm(axes = x_17_axes_0, beta = encoder_conv_module_2_sequential_0_bias, epsilon = var_28, gamma = encoder_conv_module_2_sequential_0_weight, x = input_101)[name = tensor("x_17")]; - tensor input_103_perm_0 = const()[name = tensor("input_103_perm_0"), val = tensor([0, 2, 1])]; + tensor x_17 = layer_norm(axes = x_17_axes_0, beta = inner_encoder_conv_module_2_sequential_0_bias, epsilon = var_66, gamma = inner_encoder_conv_module_2_sequential_0_weight, x = input_103)[name = tensor("x_17")]; + tensor input_105_perm_0 = const()[name = tensor("input_105_perm_0"), val = tensor([0, 2, 1])]; tensor inputs_23_pad_type_0 = const()[name = tensor("inputs_23_pad_type_0"), val = tensor("valid")]; tensor inputs_23_strides_0 = const()[name = tensor("inputs_23_strides_0"), val = tensor([1])]; tensor inputs_23_pad_0 = const()[name = tensor("inputs_23_pad_0"), val = tensor([0, 0])]; tensor inputs_23_dilations_0 = const()[name = tensor("inputs_23_dilations_0"), val = tensor([1])]; tensor inputs_23_groups_0 = const()[name = tensor("inputs_23_groups_0"), val = tensor(1)]; - tensor input_103 = transpose(perm = input_103_perm_0, x = x_17)[name = tensor("transpose_40")]; - tensor inputs_23 = conv(bias = encoder_conv_module_2_sequential_2_conv_bias, dilations = inputs_23_dilations_0, groups = inputs_23_groups_0, pad = inputs_23_pad_0, pad_type = inputs_23_pad_type_0, strides = inputs_23_strides_0, weight = encoder_conv_module_2_sequential_2_conv_weight, x = input_103)[name = tensor("inputs_23")]; - tensor var_701_split_sizes_0 = const()[name = tensor("op_701_split_sizes_0"), val = tensor([256, 256])]; - tensor var_701_axis_0 = const()[name = tensor("op_701_axis_0"), val = tensor(1)]; - tensor var_701_0, tensor var_701_1 = split(axis = var_701_axis_0, split_sizes = var_701_split_sizes_0, x = inputs_23)[name = tensor("op_701")]; - tensor var_703 = sigmoid(x = var_701_1)[name = tensor("op_703")]; - tensor inputs_25 = mul(x = var_701_0, y = var_703)[name = tensor("inputs_25")]; + tensor input_105 = transpose(perm = input_105_perm_0, x = x_17)[name = tensor("transpose_40")]; + tensor inputs_23 = conv(bias = inner_encoder_conv_module_2_sequential_2_conv_bias, dilations = inputs_23_dilations_0, groups = inputs_23_groups_0, pad = inputs_23_pad_0, pad_type = inputs_23_pad_type_0, strides = inputs_23_strides_0, weight = inner_encoder_conv_module_2_sequential_2_conv_weight, x = input_105)[name = tensor("inputs_23")]; + tensor var_757_split_sizes_0 = const()[name = tensor("op_757_split_sizes_0"), val = tensor([256, 256])]; + tensor var_757_axis_0 = const()[name = tensor("op_757_axis_0"), val = tensor(1)]; + tensor var_757_0, tensor var_757_1 = split(axis = var_757_axis_0, split_sizes = var_757_split_sizes_0, x = inputs_23)[name = tensor("op_757")]; + tensor var_759 = sigmoid(x = var_757_1)[name = tensor("op_759")]; + tensor inputs_25 = mul(x = var_757_0, y = var_759)[name = tensor("inputs_25")]; tensor outputs_aug_5_pad_type_0 = const()[name = tensor("outputs_aug_5_pad_type_0"), val = tensor("custom")]; tensor outputs_aug_5_pad_0 = const()[name = tensor("outputs_aug_5_pad_0"), val = tensor([15, 15])]; tensor outputs_aug_5_groups_0 = const()[name = tensor("outputs_aug_5_groups_0"), val = tensor(256)]; tensor outputs_aug_5_strides_0 = const()[name = tensor("outputs_aug_5_strides_0"), val = tensor([1])]; tensor outputs_aug_5_dilations_0 = const()[name = tensor("outputs_aug_5_dilations_0"), val = tensor([1])]; - tensor outputs_aug_5 = conv(dilations = outputs_aug_5_dilations_0, groups = outputs_aug_5_groups_0, pad = outputs_aug_5_pad_0, pad_type = outputs_aug_5_pad_type_0, strides = outputs_aug_5_strides_0, weight = encoder_conv_module_2_sequential_4_conv_weight, x = inputs_25)[name = tensor("outputs_aug_5")]; - tensor input_105_begin_0 = const()[name = tensor("input_105_begin_0"), val = tensor([0, 0, 0])]; - tensor input_105_end_0 = const()[name = tensor("input_105_end_0"), val = tensor([4, 256, 16])]; - tensor input_105_end_mask_0 = const()[name = tensor("input_105_end_mask_0"), val = tensor([true, true, false])]; - tensor input_105 = slice_by_index(begin = input_105_begin_0, end = input_105_end_0, end_mask = input_105_end_mask_0, x = outputs_aug_5)[name = tensor("input_105")]; - tensor inputs_27 = batch_norm(beta = encoder_conv_module_2_sequential_5_bias, epsilon = var_28, gamma = encoder_conv_module_2_sequential_5_weight, mean = encoder_conv_module_2_sequential_5_running_mean, variance = encoder_conv_module_2_sequential_5_running_var, x = input_105)[name = tensor("inputs_27")]; - tensor input_107 = silu(x = inputs_27)[name = tensor("input_107")]; - tensor input_109_pad_type_0 = const()[name = tensor("input_109_pad_type_0"), val = tensor("valid")]; - tensor input_109_strides_0 = const()[name = tensor("input_109_strides_0"), val = tensor([1])]; - tensor input_109_pad_0 = const()[name = tensor("input_109_pad_0"), val = tensor([0, 0])]; - tensor input_109_dilations_0 = const()[name = tensor("input_109_dilations_0"), val = tensor([1])]; - tensor input_109_groups_0 = const()[name = tensor("input_109_groups_0"), val = tensor(1)]; - tensor input_109 = conv(bias = encoder_conv_module_2_sequential_7_conv_bias, dilations = input_109_dilations_0, groups = input_109_groups_0, pad = input_109_pad_0, pad_type = input_109_pad_type_0, strides = input_109_strides_0, weight = encoder_conv_module_2_sequential_7_conv_weight, x = input_107)[name = tensor("input_109")]; + tensor outputs_aug_5 = conv(dilations = outputs_aug_5_dilations_0, groups = outputs_aug_5_groups_0, pad = outputs_aug_5_pad_0, pad_type = outputs_aug_5_pad_type_0, strides = outputs_aug_5_strides_0, weight = inner_encoder_conv_module_2_sequential_4_conv_weight, x = inputs_25)[name = tensor("outputs_aug_5")]; + tensor input_107_begin_0 = const()[name = tensor("input_107_begin_0"), val = tensor([0, 0, 0])]; + tensor input_107_end_0 = const()[name = tensor("input_107_end_0"), val = tensor([4, 256, 16])]; + tensor input_107_end_mask_0 = const()[name = tensor("input_107_end_mask_0"), val = tensor([true, true, false])]; + tensor input_107 = slice_by_index(begin = input_107_begin_0, end = input_107_end_0, end_mask = input_107_end_mask_0, x = outputs_aug_5)[name = tensor("input_107")]; + tensor inputs_27 = batch_norm(beta = inner_encoder_conv_module_2_sequential_5_bias, epsilon = var_66, gamma = inner_encoder_conv_module_2_sequential_5_weight, mean = inner_encoder_conv_module_2_sequential_5_running_mean, variance = inner_encoder_conv_module_2_sequential_5_running_var, x = input_107)[name = tensor("inputs_27")]; + tensor input_109 = silu(x = inputs_27)[name = tensor("input_109")]; + tensor input_111_pad_type_0 = const()[name = tensor("input_111_pad_type_0"), val = tensor("valid")]; + tensor input_111_strides_0 = const()[name = tensor("input_111_strides_0"), val = tensor([1])]; + tensor input_111_pad_0 = const()[name = tensor("input_111_pad_0"), val = tensor([0, 0])]; + tensor input_111_dilations_0 = const()[name = tensor("input_111_dilations_0"), val = tensor([1])]; + tensor input_111_groups_0 = const()[name = tensor("input_111_groups_0"), val = tensor(1)]; + tensor input_111 = conv(bias = inner_encoder_conv_module_2_sequential_7_conv_bias, dilations = input_111_dilations_0, groups = input_111_groups_0, pad = input_111_pad_0, pad_type = input_111_pad_type_0, strides = input_111_strides_0, weight = inner_encoder_conv_module_2_sequential_7_conv_weight, x = input_109)[name = tensor("input_111")]; tensor conv_out_5_perm_0 = const()[name = tensor("conv_out_5_perm_0"), val = tensor([0, 2, 1])]; - tensor var_734_begin_0 = const()[name = tensor("op_734_begin_0"), val = tensor([0, -1, 0])]; - tensor var_734_end_0 = const()[name = tensor("op_734_end_0"), val = tensor([4, 16, 256])]; - tensor var_734_end_mask_0 = const()[name = tensor("op_734_end_mask_0"), val = tensor([true, true, true])]; - tensor conv_out_5 = transpose(perm = conv_out_5_perm_0, x = input_109)[name = tensor("transpose_39")]; - tensor var_734 = slice_by_index(begin = var_734_begin_0, end = var_734_end_0, end_mask = var_734_end_mask_0, x = conv_out_5)[name = tensor("op_734")]; - tensor var_736_perm_0 = const()[name = tensor("op_736_perm_0"), val = tensor([1, 0, 2])]; - tensor var_736 = transpose(perm = var_736_perm_0, x = var_734)[name = tensor("transpose_38")]; - tensor input_111 = add(x = x_15, y = var_736)[name = tensor("input_111")]; - tensor input_113_axes_0 = const()[name = tensor("input_113_axes_0"), val = tensor([-1])]; - tensor input_113 = layer_norm(axes = input_113_axes_0, beta = encoder_ffn2_2_module_sequential_0_bias, epsilon = var_28, gamma = encoder_ffn2_2_module_sequential_0_weight, x = input_111)[name = tensor("input_113")]; - tensor inputs_29 = linear(bias = encoder_ffn2_2_module_sequential_1_linear_bias, weight = encoder_ffn2_2_module_sequential_1_linear_weight, x = input_113)[name = tensor("linear_26")]; - tensor input_115 = silu(x = inputs_29)[name = tensor("input_115")]; - tensor input_119 = linear(bias = encoder_ffn2_2_module_sequential_4_linear_bias, weight = encoder_ffn2_2_module_sequential_4_linear_weight, x = input_115)[name = tensor("linear_27")]; - tensor var_759 = const()[name = tensor("op_759"), val = tensor(0x1p-1)]; - tensor var_760 = mul(x = input_119, y = var_759)[name = tensor("op_760")]; - tensor input_121 = add(x = var_760, y = input_111)[name = tensor("input_121")]; - tensor input_123_axes_0 = const()[name = tensor("input_123_axes_0"), val = tensor([-1])]; - tensor input_123 = layer_norm(axes = input_123_axes_0, beta = encoder_layer_norm_2_bias, epsilon = var_28, gamma = encoder_layer_norm_2_weight, x = input_121)[name = tensor("input_123")]; + tensor var_790_begin_0 = const()[name = tensor("op_790_begin_0"), val = tensor([0, -1, 0])]; + tensor var_790_end_0 = const()[name = tensor("op_790_end_0"), val = tensor([4, 16, 256])]; + tensor var_790_end_mask_0 = const()[name = tensor("op_790_end_mask_0"), val = tensor([true, true, true])]; + tensor conv_out_5 = transpose(perm = conv_out_5_perm_0, x = input_111)[name = tensor("transpose_39")]; + tensor var_790 = slice_by_index(begin = var_790_begin_0, end = var_790_end_0, end_mask = var_790_end_mask_0, x = conv_out_5)[name = tensor("op_790")]; + tensor var_792_perm_0 = const()[name = tensor("op_792_perm_0"), val = tensor([1, 0, 2])]; + tensor var_792 = transpose(perm = var_792_perm_0, x = var_790)[name = tensor("transpose_38")]; + tensor input_113 = add(x = x_15, y = var_792)[name = tensor("input_113")]; + tensor input_115_axes_0 = const()[name = tensor("input_115_axes_0"), val = tensor([-1])]; + tensor input_115 = layer_norm(axes = input_115_axes_0, beta = inner_encoder_ffn2_2_module_sequential_0_bias, epsilon = var_66, gamma = inner_encoder_ffn2_2_module_sequential_0_weight, x = input_113)[name = tensor("input_115")]; + tensor inputs_29 = linear(bias = inner_encoder_ffn2_2_module_sequential_1_linear_bias, weight = inner_encoder_ffn2_2_module_sequential_1_linear_weight, x = input_115)[name = tensor("linear_26")]; + tensor input_117 = silu(x = inputs_29)[name = tensor("input_117")]; + tensor input_121 = linear(bias = inner_encoder_ffn2_2_module_sequential_4_linear_bias, weight = inner_encoder_ffn2_2_module_sequential_4_linear_weight, x = input_117)[name = tensor("linear_27")]; + tensor var_815 = const()[name = tensor("op_815"), val = tensor(0x1p-1)]; + tensor var_816 = mul(x = input_121, y = var_815)[name = tensor("op_816")]; + tensor input_123 = add(x = var_816, y = input_113)[name = tensor("input_123")]; tensor input_125_axes_0 = const()[name = tensor("input_125_axes_0"), val = tensor([-1])]; - tensor input_125 = layer_norm(axes = input_125_axes_0, beta = encoder_ffn1_3_module_sequential_0_bias, epsilon = var_28, gamma = encoder_ffn1_3_module_sequential_0_weight, x = input_123)[name = tensor("input_125")]; - tensor inputs_31 = linear(bias = encoder_ffn1_3_module_sequential_1_linear_bias, weight = encoder_ffn1_3_module_sequential_1_linear_weight, x = input_125)[name = tensor("linear_28")]; - tensor input_127 = silu(x = inputs_31)[name = tensor("input_127")]; - tensor input_131 = linear(bias = encoder_ffn1_3_module_sequential_4_linear_bias, weight = encoder_ffn1_3_module_sequential_4_linear_weight, x = input_127)[name = tensor("linear_29")]; - tensor var_789 = const()[name = tensor("op_789"), val = tensor(0x1p-1)]; - tensor var_790 = mul(x = input_131, y = var_789)[name = tensor("op_790")]; - tensor input_133 = add(x = var_790, y = input_123)[name = tensor("input_133")]; + tensor input_125 = layer_norm(axes = input_125_axes_0, beta = inner_encoder_layer_norm_2_bias, epsilon = var_66, gamma = inner_encoder_layer_norm_2_weight, x = input_123)[name = tensor("input_125")]; + tensor input_127_axes_0 = const()[name = tensor("input_127_axes_0"), val = tensor([-1])]; + tensor input_127 = layer_norm(axes = input_127_axes_0, beta = inner_encoder_ffn1_3_module_sequential_0_bias, epsilon = var_66, gamma = inner_encoder_ffn1_3_module_sequential_0_weight, x = input_125)[name = tensor("input_127")]; + tensor inputs_31 = linear(bias = inner_encoder_ffn1_3_module_sequential_1_linear_bias, weight = inner_encoder_ffn1_3_module_sequential_1_linear_weight, x = input_127)[name = tensor("linear_28")]; + tensor input_129 = silu(x = inputs_31)[name = tensor("input_129")]; + tensor input_133 = linear(bias = inner_encoder_ffn1_3_module_sequential_4_linear_bias, weight = inner_encoder_ffn1_3_module_sequential_4_linear_weight, x = input_129)[name = tensor("linear_29")]; + tensor var_845 = const()[name = tensor("op_845"), val = tensor(0x1p-1)]; + tensor var_846 = mul(x = input_133, y = var_845)[name = tensor("op_846")]; + tensor input_135 = add(x = var_846, y = input_125)[name = tensor("input_135")]; tensor x_19_axes_0 = const()[name = tensor("x_19_axes_0"), val = tensor([-1])]; - tensor x_19 = layer_norm(axes = x_19_axes_0, beta = encoder_ret_lns_3_bias, epsilon = var_28, gamma = encoder_ret_lns_3_weight, x = input_133)[name = tensor("x_19")]; + tensor x_19 = layer_norm(axes = x_19_axes_0, beta = inner_encoder_ret_lns_3_bias, epsilon = var_66, gamma = inner_encoder_ret_lns_3_weight, x = input_135)[name = tensor("x_19")]; tensor prev_kv_7_begin_0 = const()[name = tensor("prev_kv_7_begin_0"), val = tensor([3, 0, 0, 0, 0])]; tensor prev_kv_7_end_0 = const()[name = tensor("prev_kv_7_end_0"), val = tensor([4, 1, 4, 64, 64])]; tensor prev_kv_7_end_mask_0 = const()[name = tensor("prev_kv_7_end_mask_0"), val = tensor([false, true, true, true, true])]; @@ -770,209 +792,202 @@ program(1.0) tensor prev_scale_7_end_mask_0 = const()[name = tensor("prev_scale_7_end_mask_0"), val = tensor([false, true])]; tensor prev_scale_7_squeeze_mask_0 = const()[name = tensor("prev_scale_7_squeeze_mask_0"), val = tensor([true, false])]; tensor prev_scale_7 = slice_by_index(begin = prev_scale_7_begin_0, end = prev_scale_7_end_0, end_mask = prev_scale_7_end_mask_0, squeeze_mask = prev_scale_7_squeeze_mask_0, x = enc_scale)[name = tensor("prev_scale_7")]; - tensor var_804 = linear(bias = encoder_q_proj_3_bias, weight = encoder_q_proj_3_weight, x = x_19)[name = tensor("linear_30")]; - tensor var_805 = const()[name = tensor("op_805"), val = tensor([1, 4, 4, 64])]; - tensor var_806 = reshape(shape = var_805, x = var_804)[name = tensor("op_806")]; + tensor var_860 = linear(bias = inner_encoder_q_proj_3_bias, weight = inner_encoder_q_proj_3_weight, x = x_19)[name = tensor("linear_30")]; + tensor var_861 = const()[name = tensor("op_861"), val = tensor([1, 4, 4, 64])]; + tensor var_862 = reshape(shape = var_861, x = var_860)[name = tensor("op_862")]; tensor q_7_perm_0 = const()[name = tensor("q_7_perm_0"), val = tensor([0, 2, 1, 3])]; - tensor var_810 = linear(bias = encoder_k_proj_3_bias, weight = encoder_k_proj_3_weight, x = x_19)[name = tensor("linear_31")]; - tensor var_811 = const()[name = tensor("op_811"), val = tensor(0x1p-3)]; - tensor var_812 = mul(x = var_810, y = var_811)[name = tensor("op_812")]; - tensor var_813 = const()[name = tensor("op_813"), val = tensor([1, 4, 4, 64])]; - tensor var_814 = reshape(shape = var_813, x = var_812)[name = tensor("op_814")]; + tensor var_866 = linear(bias = inner_encoder_k_proj_3_bias, weight = inner_encoder_k_proj_3_weight, x = x_19)[name = tensor("linear_31")]; + tensor var_867 = const()[name = tensor("op_867"), val = tensor(0x1p-3)]; + tensor var_868 = mul(x = var_866, y = var_867)[name = tensor("op_868")]; + tensor var_869 = const()[name = tensor("op_869"), val = tensor([1, 4, 4, 64])]; + tensor var_870 = reshape(shape = var_869, x = var_868)[name = tensor("op_870")]; tensor k_7_perm_0 = const()[name = tensor("k_7_perm_0"), val = tensor([0, 2, 1, 3])]; - tensor var_818 = linear(bias = encoder_v_proj_3_bias, weight = encoder_v_proj_3_weight, x = x_19)[name = tensor("linear_32")]; - tensor var_819 = const()[name = tensor("op_819"), val = tensor([1, 4, 4, 64])]; - tensor var_820 = reshape(shape = var_819, x = var_818)[name = tensor("op_820")]; + tensor var_874 = linear(bias = inner_encoder_v_proj_3_bias, weight = inner_encoder_v_proj_3_weight, x = x_19)[name = tensor("linear_32")]; + tensor var_875 = const()[name = tensor("op_875"), val = tensor([1, 4, 4, 64])]; + tensor var_876 = reshape(shape = var_875, x = var_874)[name = tensor("op_876")]; tensor v_7_perm_0 = const()[name = tensor("v_7_perm_0"), val = tensor([0, 2, 1, 3])]; - tensor input_137 = linear(bias = encoder_g_proj_3_bias, weight = encoder_g_proj_3_weight, x = x_19)[name = tensor("linear_33")]; + tensor input_139 = linear(bias = inner_encoder_g_proj_3_bias, weight = inner_encoder_g_proj_3_weight, x = x_19)[name = tensor("linear_33")]; tensor sqrt_s0_7 = sqrt(x = prev_scale_7)[name = tensor("sqrt_s0_7")]; - tensor s_t_7 = add(x = prev_scale_7, y = encoder__t_index)[name = tensor("s_t_7")]; + tensor s_t_7 = add(x = prev_scale_7, y = inner_encoder__t_index)[name = tensor("s_t_7")]; tensor sqrt_s_t_7 = sqrt(x = s_t_7)[name = tensor("sqrt_s_t_7")]; tensor qk_7_transpose_x_1 = const()[name = tensor("qk_7_transpose_x_1"), val = tensor(false)]; tensor qk_7_transpose_y_1 = const()[name = tensor("qk_7_transpose_y_1"), val = tensor(true)]; - tensor k_7 = transpose(perm = k_7_perm_0, x = var_814)[name = tensor("transpose_36")]; - tensor q_7 = transpose(perm = q_7_perm_0, x = var_806)[name = tensor("transpose_37")]; + tensor k_7 = transpose(perm = k_7_perm_0, x = var_870)[name = tensor("transpose_36")]; + tensor q_7 = transpose(perm = q_7_perm_0, x = var_862)[name = tensor("transpose_37")]; tensor qk_7 = matmul(transpose_x = qk_7_transpose_x_1, transpose_y = qk_7_transpose_y_1, x = q_7, y = k_7)[name = tensor("qk_7")]; - tensor var_830 = const()[name = tensor("op_830"), val = tensor([4, 1])]; - tensor var_831 = reshape(shape = var_830, x = sqrt_s_t_7)[name = tensor("op_831")]; - tensor M_7 = real_div(x = encoder__causal_mask, y = var_831)[name = tensor("M_7")]; - tensor var_833 = mul(x = qk_7, y = M_7)[name = tensor("op_833")]; + tensor var_886 = const()[name = tensor("op_886"), val = tensor([4, 1])]; + tensor var_887 = reshape(shape = var_886, x = sqrt_s_t_7)[name = tensor("op_887")]; + tensor M_7 = real_div(x = inner_encoder__causal_mask, y = var_887)[name = tensor("M_7")]; + tensor var_889 = mul(x = qk_7, y = M_7)[name = tensor("op_889")]; tensor inner_7_transpose_x_0 = const()[name = tensor("inner_7_transpose_x_0"), val = tensor(false)]; tensor inner_7_transpose_y_0 = const()[name = tensor("inner_7_transpose_y_0"), val = tensor(false)]; - tensor v_7 = transpose(perm = v_7_perm_0, x = var_820)[name = tensor("transpose_35")]; - tensor inner_7 = matmul(transpose_x = inner_7_transpose_x_0, transpose_y = inner_7_transpose_y_0, x = var_833, y = v_7)[name = tensor("inner_7")]; - tensor var_835_transpose_x_0 = const()[name = tensor("op_835_transpose_x_0"), val = tensor(false)]; - tensor var_835_transpose_y_0 = const()[name = tensor("op_835_transpose_y_0"), val = tensor(false)]; - tensor var_835 = matmul(transpose_x = var_835_transpose_x_0, transpose_y = var_835_transpose_y_0, x = q_7, y = prev_kv_7)[name = tensor("op_835")]; - tensor var_836 = real_div(x = sqrt_s0_7, y = sqrt_s_t_7)[name = tensor("op_836")]; - tensor var_837 = const()[name = tensor("op_837"), val = tensor([1, 1, 4, 1])]; - tensor var_838 = reshape(shape = var_837, x = var_836)[name = tensor("op_838")]; - tensor cross_7 = mul(x = var_835, y = var_838)[name = tensor("cross_7")]; + tensor v_7 = transpose(perm = v_7_perm_0, x = var_876)[name = tensor("transpose_35")]; + tensor inner_7 = matmul(transpose_x = inner_7_transpose_x_0, transpose_y = inner_7_transpose_y_0, x = var_889, y = v_7)[name = tensor("inner_7")]; + tensor var_891_transpose_x_0 = const()[name = tensor("op_891_transpose_x_0"), val = tensor(false)]; + tensor var_891_transpose_y_0 = const()[name = tensor("op_891_transpose_y_0"), val = tensor(false)]; + tensor var_891 = matmul(transpose_x = var_891_transpose_x_0, transpose_y = var_891_transpose_y_0, x = q_7, y = prev_kv_7)[name = tensor("op_891")]; + tensor var_892 = real_div(x = sqrt_s0_7, y = sqrt_s_t_7)[name = tensor("op_892")]; + tensor var_893 = const()[name = tensor("op_893"), val = tensor([1, 1, 4, 1])]; + tensor var_894 = reshape(shape = var_893, x = var_892)[name = tensor("op_894")]; + tensor cross_7 = mul(x = var_891, y = var_894)[name = tensor("cross_7")]; tensor out_19 = add(x = inner_7, y = cross_7)[name = tensor("out_19")]; - tensor var_841 = mul(x = prev_kv_7, y = sqrt_s0_7)[name = tensor("op_841")]; - tensor var_843_transpose_x_1 = const()[name = tensor("op_843_transpose_x_1"), val = tensor(true)]; - tensor var_843_transpose_y_1 = const()[name = tensor("op_843_transpose_y_1"), val = tensor(false)]; - tensor var_843 = matmul(transpose_x = var_843_transpose_x_1, transpose_y = var_843_transpose_y_1, x = k_7, y = v_7)[name = tensor("op_843")]; - tensor new_kv_unnorm_7 = add(x = var_841, y = var_843)[name = tensor("new_kv_unnorm_7")]; - tensor var_845 = const()[name = tensor("op_845"), val = tensor(0x1p+2)]; - tensor new_scale_7 = add(x = prev_scale_7, y = var_845)[name = tensor("new_scale_7")]; - tensor var_847 = sqrt(x = new_scale_7)[name = tensor("op_847")]; - tensor nkv_1 = real_div(x = new_kv_unnorm_7, y = var_847)[name = tensor("nkv_1")]; - tensor var_849_perm_0 = const()[name = tensor("op_849_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_897 = mul(x = prev_kv_7, y = sqrt_s0_7)[name = tensor("op_897")]; + tensor var_899_transpose_x_1 = const()[name = tensor("op_899_transpose_x_1"), val = tensor(true)]; + tensor var_899_transpose_y_1 = const()[name = tensor("op_899_transpose_y_1"), val = tensor(false)]; + tensor var_899 = matmul(transpose_x = var_899_transpose_x_1, transpose_y = var_899_transpose_y_1, x = k_7, y = v_7)[name = tensor("op_899")]; + tensor new_kv_unnorm_7 = add(x = var_897, y = var_899)[name = tensor("new_kv_unnorm_7")]; + tensor var_901 = const()[name = tensor("op_901"), val = tensor(0x1p+2)]; + tensor new_scale_7 = add(x = prev_scale_7, y = var_901)[name = tensor("new_scale_7")]; + tensor var_903 = sqrt(x = new_scale_7)[name = tensor("op_903")]; + tensor nkv_1 = real_div(x = new_kv_unnorm_7, y = var_903)[name = tensor("nkv_1")]; + tensor var_905_perm_0 = const()[name = tensor("op_905_perm_0"), val = tensor([0, 2, 1, 3])]; tensor out_21_axes_0 = const()[name = tensor("out_21_axes_0"), val = tensor([-1])]; - tensor var_849 = transpose(perm = var_849_perm_0, x = out_19)[name = tensor("transpose_34")]; - tensor out_21 = layer_norm(axes = out_21_axes_0, epsilon = var_17, x = var_849)[name = tensor("out_21")]; - tensor var_853 = const()[name = tensor("op_853"), val = tensor([1, 4, 256])]; - tensor out_23 = reshape(shape = var_853, x = out_21)[name = tensor("out_23")]; - tensor var_855 = silu(x = input_137)[name = tensor("op_855")]; - tensor input_139 = mul(x = var_855, y = out_23)[name = tensor("input_139")]; - tensor ret_out_7 = linear(bias = encoder_out_proj_3_bias, weight = encoder_out_proj_3_weight, x = input_139)[name = tensor("linear_34")]; - tensor x_21 = add(x = input_133, y = ret_out_7)[name = tensor("x_21")]; + tensor var_905 = transpose(perm = var_905_perm_0, x = out_19)[name = tensor("transpose_34")]; + tensor out_21 = layer_norm(axes = out_21_axes_0, epsilon = var_74, x = var_905)[name = tensor("out_21")]; + tensor var_909 = const()[name = tensor("op_909"), val = tensor([1, 4, 256])]; + tensor out_23 = reshape(shape = var_909, x = out_21)[name = tensor("out_23")]; + tensor var_911 = silu(x = input_139)[name = tensor("op_911")]; + tensor input_141 = mul(x = var_911, y = out_23)[name = tensor("input_141")]; + tensor ret_out_7 = linear(bias = inner_encoder_out_proj_3_bias, weight = inner_encoder_out_proj_3_weight, x = input_141)[name = tensor("linear_34")]; + tensor x_21 = add(x = input_135, y = ret_out_7)[name = tensor("x_21")]; tensor window_31_begin_0 = const()[name = tensor("window_31_begin_0"), val = tensor([3, 0, 0, 0])]; tensor window_31_end_0 = const()[name = tensor("window_31_end_0"), val = tensor([4, 1, 16, 256])]; tensor window_31_end_mask_0 = const()[name = tensor("window_31_end_mask_0"), val = tensor([false, true, true, true])]; tensor window_31_squeeze_mask_0 = const()[name = tensor("window_31_squeeze_mask_0"), val = tensor([true, false, false, false])]; tensor window_31 = slice_by_index(begin = window_31_begin_0, end = window_31_end_0, end_mask = window_31_end_mask_0, squeeze_mask = window_31_squeeze_mask_0, x = enc_conv_cache)[name = tensor("window_31")]; - tensor var_863_begin_0 = const()[name = tensor("op_863_begin_0"), val = tensor([0, 0, 0])]; - tensor var_863_end_0 = const()[name = tensor("op_863_end_0"), val = tensor([1, 1, 256])]; - tensor var_863_end_mask_0 = const()[name = tensor("op_863_end_mask_0"), val = tensor([true, false, true])]; - tensor var_863 = slice_by_index(begin = var_863_begin_0, end = var_863_end_0, end_mask = var_863_end_mask_0, x = x_21)[name = tensor("op_863")]; - tensor var_866_begin_0 = const()[name = tensor("op_866_begin_0"), val = tensor([0, 1, 0])]; - tensor var_866_end_0 = const()[name = tensor("op_866_end_0"), val = tensor([1, 16, 256])]; - tensor var_866_end_mask_0 = const()[name = tensor("op_866_end_mask_0"), val = tensor([true, true, true])]; - tensor var_866 = slice_by_index(begin = var_866_begin_0, end = var_866_end_0, end_mask = var_866_end_mask_0, x = window_31)[name = tensor("op_866")]; + tensor var_919_begin_0 = const()[name = tensor("op_919_begin_0"), val = tensor([0, 0, 0])]; + tensor var_919_end_0 = const()[name = tensor("op_919_end_0"), val = tensor([1, 1, 256])]; + tensor var_919_end_mask_0 = const()[name = tensor("op_919_end_mask_0"), val = tensor([true, false, true])]; + tensor var_919 = slice_by_index(begin = var_919_begin_0, end = var_919_end_0, end_mask = var_919_end_mask_0, x = x_21)[name = tensor("op_919")]; + tensor var_922_begin_0 = const()[name = tensor("op_922_begin_0"), val = tensor([0, 1, 0])]; + tensor var_922_end_0 = const()[name = tensor("op_922_end_0"), val = tensor([1, 16, 256])]; + tensor var_922_end_mask_0 = const()[name = tensor("op_922_end_mask_0"), val = tensor([true, true, true])]; + tensor var_922 = slice_by_index(begin = var_922_begin_0, end = var_922_end_0, end_mask = var_922_end_mask_0, x = window_31)[name = tensor("op_922")]; tensor window_33_interleave_0 = const()[name = tensor("window_33_interleave_0"), val = tensor(false)]; - tensor window_33 = concat(axis = var_26, interleave = window_33_interleave_0, values = (var_866, var_863))[name = tensor("window_33")]; - tensor var_871_begin_0 = const()[name = tensor("op_871_begin_0"), val = tensor([0, 1, 0])]; - tensor var_871_end_0 = const()[name = tensor("op_871_end_0"), val = tensor([1, 2, 256])]; - tensor var_871_end_mask_0 = const()[name = tensor("op_871_end_mask_0"), val = tensor([true, false, true])]; - tensor var_871 = slice_by_index(begin = var_871_begin_0, end = var_871_end_0, end_mask = var_871_end_mask_0, x = x_21)[name = tensor("op_871")]; - tensor var_874_begin_0 = const()[name = tensor("op_874_begin_0"), val = tensor([0, 1, 0])]; - tensor var_874_end_0 = const()[name = tensor("op_874_end_0"), val = tensor([1, 16, 256])]; - tensor var_874_end_mask_0 = const()[name = tensor("op_874_end_mask_0"), val = tensor([true, true, true])]; - tensor var_874 = slice_by_index(begin = var_874_begin_0, end = var_874_end_0, end_mask = var_874_end_mask_0, x = window_33)[name = tensor("op_874")]; + tensor window_33 = concat(axis = var_82, interleave = window_33_interleave_0, values = (var_922, var_919))[name = tensor("window_33")]; + tensor var_927_begin_0 = const()[name = tensor("op_927_begin_0"), val = tensor([0, 1, 0])]; + tensor var_927_end_0 = const()[name = tensor("op_927_end_0"), val = tensor([1, 2, 256])]; + tensor var_927_end_mask_0 = const()[name = tensor("op_927_end_mask_0"), val = tensor([true, false, true])]; + tensor var_927 = slice_by_index(begin = var_927_begin_0, end = var_927_end_0, end_mask = var_927_end_mask_0, x = x_21)[name = tensor("op_927")]; + tensor var_930_begin_0 = const()[name = tensor("op_930_begin_0"), val = tensor([0, 1, 0])]; + tensor var_930_end_0 = const()[name = tensor("op_930_end_0"), val = tensor([1, 16, 256])]; + tensor var_930_end_mask_0 = const()[name = tensor("op_930_end_mask_0"), val = tensor([true, true, true])]; + tensor var_930 = slice_by_index(begin = var_930_begin_0, end = var_930_end_0, end_mask = var_930_end_mask_0, x = window_33)[name = tensor("op_930")]; tensor window_35_interleave_0 = const()[name = tensor("window_35_interleave_0"), val = tensor(false)]; - tensor window_35 = concat(axis = var_26, interleave = window_35_interleave_0, values = (var_874, var_871))[name = tensor("window_35")]; - tensor var_879_begin_0 = const()[name = tensor("op_879_begin_0"), val = tensor([0, 2, 0])]; - tensor var_879_end_0 = const()[name = tensor("op_879_end_0"), val = tensor([1, 3, 256])]; - tensor var_879_end_mask_0 = const()[name = tensor("op_879_end_mask_0"), val = tensor([true, false, true])]; - tensor var_879 = slice_by_index(begin = var_879_begin_0, end = var_879_end_0, end_mask = var_879_end_mask_0, x = x_21)[name = tensor("op_879")]; - tensor var_882_begin_0 = const()[name = tensor("op_882_begin_0"), val = tensor([0, 1, 0])]; - tensor var_882_end_0 = const()[name = tensor("op_882_end_0"), val = tensor([1, 16, 256])]; - tensor var_882_end_mask_0 = const()[name = tensor("op_882_end_mask_0"), val = tensor([true, true, true])]; - tensor var_882 = slice_by_index(begin = var_882_begin_0, end = var_882_end_0, end_mask = var_882_end_mask_0, x = window_35)[name = tensor("op_882")]; + tensor window_35 = concat(axis = var_82, interleave = window_35_interleave_0, values = (var_930, var_927))[name = tensor("window_35")]; + tensor var_935_begin_0 = const()[name = tensor("op_935_begin_0"), val = tensor([0, 2, 0])]; + tensor var_935_end_0 = const()[name = tensor("op_935_end_0"), val = tensor([1, 3, 256])]; + tensor var_935_end_mask_0 = const()[name = tensor("op_935_end_mask_0"), val = tensor([true, false, true])]; + tensor var_935 = slice_by_index(begin = var_935_begin_0, end = var_935_end_0, end_mask = var_935_end_mask_0, x = x_21)[name = tensor("op_935")]; + tensor var_938_begin_0 = const()[name = tensor("op_938_begin_0"), val = tensor([0, 1, 0])]; + tensor var_938_end_0 = const()[name = tensor("op_938_end_0"), val = tensor([1, 16, 256])]; + tensor var_938_end_mask_0 = const()[name = tensor("op_938_end_mask_0"), val = tensor([true, true, true])]; + tensor var_938 = slice_by_index(begin = var_938_begin_0, end = var_938_end_0, end_mask = var_938_end_mask_0, x = window_35)[name = tensor("op_938")]; tensor window_37_interleave_0 = const()[name = tensor("window_37_interleave_0"), val = tensor(false)]; - tensor window_37 = concat(axis = var_26, interleave = window_37_interleave_0, values = (var_882, var_879))[name = tensor("window_37")]; - tensor var_887_begin_0 = const()[name = tensor("op_887_begin_0"), val = tensor([0, 3, 0])]; - tensor var_887_end_0 = const()[name = tensor("op_887_end_0"), val = tensor([1, 1, 256])]; - tensor var_887_end_mask_0 = const()[name = tensor("op_887_end_mask_0"), val = tensor([true, true, true])]; - tensor var_887 = slice_by_index(begin = var_887_begin_0, end = var_887_end_0, end_mask = var_887_end_mask_0, x = x_21)[name = tensor("op_887")]; - tensor var_890_begin_0 = const()[name = tensor("op_890_begin_0"), val = tensor([0, 1, 0])]; - tensor var_890_end_0 = const()[name = tensor("op_890_end_0"), val = tensor([1, 16, 256])]; - tensor var_890_end_mask_0 = const()[name = tensor("op_890_end_mask_0"), val = tensor([true, true, true])]; - tensor var_890 = slice_by_index(begin = var_890_begin_0, end = var_890_end_0, end_mask = var_890_end_mask_0, x = window_37)[name = tensor("op_890")]; + tensor window_37 = concat(axis = var_82, interleave = window_37_interleave_0, values = (var_938, var_935))[name = tensor("window_37")]; + tensor var_943_begin_0 = const()[name = tensor("op_943_begin_0"), val = tensor([0, 3, 0])]; + tensor var_943_end_0 = const()[name = tensor("op_943_end_0"), val = tensor([1, 1, 256])]; + tensor var_943_end_mask_0 = const()[name = tensor("op_943_end_mask_0"), val = tensor([true, true, true])]; + tensor var_943 = slice_by_index(begin = var_943_begin_0, end = var_943_end_0, end_mask = var_943_end_mask_0, x = x_21)[name = tensor("op_943")]; + tensor var_946_begin_0 = const()[name = tensor("op_946_begin_0"), val = tensor([0, 1, 0])]; + tensor var_946_end_0 = const()[name = tensor("op_946_end_0"), val = tensor([1, 16, 256])]; + tensor var_946_end_mask_0 = const()[name = tensor("op_946_end_mask_0"), val = tensor([true, true, true])]; + tensor var_946 = slice_by_index(begin = var_946_begin_0, end = var_946_end_0, end_mask = var_946_end_mask_0, x = window_37)[name = tensor("op_946")]; tensor window_interleave_0 = const()[name = tensor("window_interleave_0"), val = tensor(false)]; - tensor window = concat(axis = var_26, interleave = window_interleave_0, values = (var_890, var_887))[name = tensor("window")]; - tensor input_141_interleave_0 = const()[name = tensor("input_141_interleave_0"), val = tensor(false)]; - tensor input_141 = concat(axis = var_23, interleave = input_141_interleave_0, values = (window_33, window_35, window_37, window))[name = tensor("input_141")]; + tensor window = concat(axis = var_82, interleave = window_interleave_0, values = (var_946, var_943))[name = tensor("window")]; + tensor input_143_interleave_0 = const()[name = tensor("input_143_interleave_0"), val = tensor(false)]; + tensor input_143 = concat(axis = var_69, interleave = input_143_interleave_0, values = (window_33, window_35, window_37, window))[name = tensor("input_143")]; tensor x_23_axes_0 = const()[name = tensor("x_23_axes_0"), val = tensor([-1])]; - tensor x_23 = layer_norm(axes = x_23_axes_0, beta = encoder_conv_module_3_sequential_0_bias, epsilon = var_28, gamma = encoder_conv_module_3_sequential_0_weight, x = input_141)[name = tensor("x_23")]; - tensor input_143_perm_0 = const()[name = tensor("input_143_perm_0"), val = tensor([0, 2, 1])]; + tensor x_23 = layer_norm(axes = x_23_axes_0, beta = inner_encoder_conv_module_3_sequential_0_bias, epsilon = var_66, gamma = inner_encoder_conv_module_3_sequential_0_weight, x = input_143)[name = tensor("x_23")]; + tensor input_145_perm_0 = const()[name = tensor("input_145_perm_0"), val = tensor([0, 2, 1])]; tensor inputs_33_pad_type_0 = const()[name = tensor("inputs_33_pad_type_0"), val = tensor("valid")]; tensor inputs_33_strides_0 = const()[name = tensor("inputs_33_strides_0"), val = tensor([1])]; tensor inputs_33_pad_0 = const()[name = tensor("inputs_33_pad_0"), val = tensor([0, 0])]; tensor inputs_33_dilations_0 = const()[name = tensor("inputs_33_dilations_0"), val = tensor([1])]; tensor inputs_33_groups_0 = const()[name = tensor("inputs_33_groups_0"), val = tensor(1)]; - tensor input_143 = transpose(perm = input_143_perm_0, x = x_23)[name = tensor("transpose_33")]; - tensor inputs_33 = conv(bias = encoder_conv_module_3_sequential_2_conv_bias, dilations = inputs_33_dilations_0, groups = inputs_33_groups_0, pad = inputs_33_pad_0, pad_type = inputs_33_pad_type_0, strides = inputs_33_strides_0, weight = encoder_conv_module_3_sequential_2_conv_weight, x = input_143)[name = tensor("inputs_33")]; - tensor var_915_split_sizes_0 = const()[name = tensor("op_915_split_sizes_0"), val = tensor([256, 256])]; - tensor var_915_axis_0 = const()[name = tensor("op_915_axis_0"), val = tensor(1)]; - tensor var_915_0, tensor var_915_1 = split(axis = var_915_axis_0, split_sizes = var_915_split_sizes_0, x = inputs_33)[name = tensor("op_915")]; - tensor var_917 = sigmoid(x = var_915_1)[name = tensor("op_917")]; - tensor inputs_35 = mul(x = var_915_0, y = var_917)[name = tensor("inputs_35")]; + tensor input_145 = transpose(perm = input_145_perm_0, x = x_23)[name = tensor("transpose_33")]; + tensor inputs_33 = conv(bias = inner_encoder_conv_module_3_sequential_2_conv_bias, dilations = inputs_33_dilations_0, groups = inputs_33_groups_0, pad = inputs_33_pad_0, pad_type = inputs_33_pad_type_0, strides = inputs_33_strides_0, weight = inner_encoder_conv_module_3_sequential_2_conv_weight, x = input_145)[name = tensor("inputs_33")]; + tensor var_971_split_sizes_0 = const()[name = tensor("op_971_split_sizes_0"), val = tensor([256, 256])]; + tensor var_971_axis_0 = const()[name = tensor("op_971_axis_0"), val = tensor(1)]; + tensor var_971_0, tensor var_971_1 = split(axis = var_971_axis_0, split_sizes = var_971_split_sizes_0, x = inputs_33)[name = tensor("op_971")]; + tensor var_973 = sigmoid(x = var_971_1)[name = tensor("op_973")]; + tensor inputs_35 = mul(x = var_971_0, y = var_973)[name = tensor("inputs_35")]; tensor outputs_aug_pad_type_0 = const()[name = tensor("outputs_aug_pad_type_0"), val = tensor("custom")]; tensor outputs_aug_pad_0 = const()[name = tensor("outputs_aug_pad_0"), val = tensor([15, 15])]; tensor outputs_aug_groups_0 = const()[name = tensor("outputs_aug_groups_0"), val = tensor(256)]; tensor outputs_aug_strides_0 = const()[name = tensor("outputs_aug_strides_0"), val = tensor([1])]; tensor outputs_aug_dilations_0 = const()[name = tensor("outputs_aug_dilations_0"), val = tensor([1])]; - tensor outputs_aug = conv(dilations = outputs_aug_dilations_0, groups = outputs_aug_groups_0, pad = outputs_aug_pad_0, pad_type = outputs_aug_pad_type_0, strides = outputs_aug_strides_0, weight = encoder_conv_module_3_sequential_4_conv_weight, x = inputs_35)[name = tensor("outputs_aug")]; - tensor input_145_begin_0 = const()[name = tensor("input_145_begin_0"), val = tensor([0, 0, 0])]; - tensor input_145_end_0 = const()[name = tensor("input_145_end_0"), val = tensor([4, 256, 16])]; - tensor input_145_end_mask_0 = const()[name = tensor("input_145_end_mask_0"), val = tensor([true, true, false])]; - tensor input_145 = slice_by_index(begin = input_145_begin_0, end = input_145_end_0, end_mask = input_145_end_mask_0, x = outputs_aug)[name = tensor("input_145")]; - tensor inputs_37 = batch_norm(beta = encoder_conv_module_3_sequential_5_bias, epsilon = var_28, gamma = encoder_conv_module_3_sequential_5_weight, mean = encoder_conv_module_3_sequential_5_running_mean, variance = encoder_conv_module_3_sequential_5_running_var, x = input_145)[name = tensor("inputs_37")]; - tensor input_147 = silu(x = inputs_37)[name = tensor("input_147")]; - tensor input_149_pad_type_0 = const()[name = tensor("input_149_pad_type_0"), val = tensor("valid")]; - tensor input_149_strides_0 = const()[name = tensor("input_149_strides_0"), val = tensor([1])]; - tensor input_149_pad_0 = const()[name = tensor("input_149_pad_0"), val = tensor([0, 0])]; - tensor input_149_dilations_0 = const()[name = tensor("input_149_dilations_0"), val = tensor([1])]; - tensor input_149_groups_0 = const()[name = tensor("input_149_groups_0"), val = tensor(1)]; - tensor input_149 = conv(bias = encoder_conv_module_3_sequential_7_conv_bias, dilations = input_149_dilations_0, groups = input_149_groups_0, pad = input_149_pad_0, pad_type = input_149_pad_type_0, strides = input_149_strides_0, weight = encoder_conv_module_3_sequential_7_conv_weight, x = input_147)[name = tensor("input_149")]; + tensor outputs_aug = conv(dilations = outputs_aug_dilations_0, groups = outputs_aug_groups_0, pad = outputs_aug_pad_0, pad_type = outputs_aug_pad_type_0, strides = outputs_aug_strides_0, weight = inner_encoder_conv_module_3_sequential_4_conv_weight, x = inputs_35)[name = tensor("outputs_aug")]; + tensor input_147_begin_0 = const()[name = tensor("input_147_begin_0"), val = tensor([0, 0, 0])]; + tensor input_147_end_0 = const()[name = tensor("input_147_end_0"), val = tensor([4, 256, 16])]; + tensor input_147_end_mask_0 = const()[name = tensor("input_147_end_mask_0"), val = tensor([true, true, false])]; + tensor input_147 = slice_by_index(begin = input_147_begin_0, end = input_147_end_0, end_mask = input_147_end_mask_0, x = outputs_aug)[name = tensor("input_147")]; + tensor inputs_37 = batch_norm(beta = inner_encoder_conv_module_3_sequential_5_bias, epsilon = var_66, gamma = inner_encoder_conv_module_3_sequential_5_weight, mean = inner_encoder_conv_module_3_sequential_5_running_mean, variance = inner_encoder_conv_module_3_sequential_5_running_var, x = input_147)[name = tensor("inputs_37")]; + tensor input_149 = silu(x = inputs_37)[name = tensor("input_149")]; + tensor input_151_pad_type_0 = const()[name = tensor("input_151_pad_type_0"), val = tensor("valid")]; + tensor input_151_strides_0 = const()[name = tensor("input_151_strides_0"), val = tensor([1])]; + tensor input_151_pad_0 = const()[name = tensor("input_151_pad_0"), val = tensor([0, 0])]; + tensor input_151_dilations_0 = const()[name = tensor("input_151_dilations_0"), val = tensor([1])]; + tensor input_151_groups_0 = const()[name = tensor("input_151_groups_0"), val = tensor(1)]; + tensor input_151 = conv(bias = inner_encoder_conv_module_3_sequential_7_conv_bias, dilations = input_151_dilations_0, groups = input_151_groups_0, pad = input_151_pad_0, pad_type = input_151_pad_type_0, strides = input_151_strides_0, weight = inner_encoder_conv_module_3_sequential_7_conv_weight, x = input_149)[name = tensor("input_151")]; tensor conv_out_7_perm_0 = const()[name = tensor("conv_out_7_perm_0"), val = tensor([0, 2, 1])]; - tensor var_948_begin_0 = const()[name = tensor("op_948_begin_0"), val = tensor([0, -1, 0])]; - tensor var_948_end_0 = const()[name = tensor("op_948_end_0"), val = tensor([4, 16, 256])]; - tensor var_948_end_mask_0 = const()[name = tensor("op_948_end_mask_0"), val = tensor([true, true, true])]; - tensor conv_out_7 = transpose(perm = conv_out_7_perm_0, x = input_149)[name = tensor("transpose_32")]; - tensor var_948 = slice_by_index(begin = var_948_begin_0, end = var_948_end_0, end_mask = var_948_end_mask_0, x = conv_out_7)[name = tensor("op_948")]; - tensor var_950_perm_0 = const()[name = tensor("op_950_perm_0"), val = tensor([1, 0, 2])]; - tensor var_950 = transpose(perm = var_950_perm_0, x = var_948)[name = tensor("transpose_31")]; - tensor input_151 = add(x = x_21, y = var_950)[name = tensor("input_151")]; - tensor input_153_axes_0 = const()[name = tensor("input_153_axes_0"), val = tensor([-1])]; - tensor input_153 = layer_norm(axes = input_153_axes_0, beta = encoder_ffn2_3_module_sequential_0_bias, epsilon = var_28, gamma = encoder_ffn2_3_module_sequential_0_weight, x = input_151)[name = tensor("input_153")]; - tensor inputs = linear(bias = encoder_ffn2_3_module_sequential_1_linear_bias, weight = encoder_ffn2_3_module_sequential_1_linear_weight, x = input_153)[name = tensor("linear_35")]; - tensor input_155 = silu(x = inputs)[name = tensor("input_155")]; - tensor input_159 = linear(bias = encoder_ffn2_3_module_sequential_4_linear_bias, weight = encoder_ffn2_3_module_sequential_4_linear_weight, x = input_155)[name = tensor("linear_36")]; - tensor var_973 = const()[name = tensor("op_973"), val = tensor(0x1p-1)]; - tensor var_974 = mul(x = input_159, y = var_973)[name = tensor("op_974")]; - tensor input_161 = add(x = var_974, y = input_151)[name = tensor("input_161")]; + tensor var_1004_begin_0 = const()[name = tensor("op_1004_begin_0"), val = tensor([0, -1, 0])]; + tensor var_1004_end_0 = const()[name = tensor("op_1004_end_0"), val = tensor([4, 16, 256])]; + tensor var_1004_end_mask_0 = const()[name = tensor("op_1004_end_mask_0"), val = tensor([true, true, true])]; + tensor conv_out_7 = transpose(perm = conv_out_7_perm_0, x = input_151)[name = tensor("transpose_32")]; + tensor var_1004 = slice_by_index(begin = var_1004_begin_0, end = var_1004_end_0, end_mask = var_1004_end_mask_0, x = conv_out_7)[name = tensor("op_1004")]; + tensor var_1006_perm_0 = const()[name = tensor("op_1006_perm_0"), val = tensor([1, 0, 2])]; + tensor var_1006 = transpose(perm = var_1006_perm_0, x = var_1004)[name = tensor("transpose_31")]; + tensor input_153 = add(x = x_21, y = var_1006)[name = tensor("input_153")]; + tensor input_155_axes_0 = const()[name = tensor("input_155_axes_0"), val = tensor([-1])]; + tensor input_155 = layer_norm(axes = input_155_axes_0, beta = inner_encoder_ffn2_3_module_sequential_0_bias, epsilon = var_66, gamma = inner_encoder_ffn2_3_module_sequential_0_weight, x = input_153)[name = tensor("input_155")]; + tensor inputs = linear(bias = inner_encoder_ffn2_3_module_sequential_1_linear_bias, weight = inner_encoder_ffn2_3_module_sequential_1_linear_weight, x = input_155)[name = tensor("linear_35")]; + tensor input_157 = silu(x = inputs)[name = tensor("input_157")]; + tensor input_161 = linear(bias = inner_encoder_ffn2_3_module_sequential_4_linear_bias, weight = inner_encoder_ffn2_3_module_sequential_4_linear_weight, x = input_157)[name = tensor("linear_36")]; + tensor var_1029 = const()[name = tensor("op_1029"), val = tensor(0x1p-1)]; + tensor var_1030 = mul(x = input_161, y = var_1029)[name = tensor("op_1030")]; + tensor input_163 = add(x = var_1030, y = input_153)[name = tensor("input_163")]; tensor x_25_axes_0 = const()[name = tensor("x_25_axes_0"), val = tensor([-1])]; - tensor x_25 = layer_norm(axes = x_25_axes_0, beta = encoder_layer_norm_3_bias, epsilon = var_28, gamma = encoder_layer_norm_3_weight, x = input_161)[name = tensor("x_25")]; + tensor x_25 = layer_norm(axes = x_25_axes_0, beta = inner_encoder_layer_norm_3_bias, epsilon = var_66, gamma = inner_encoder_layer_norm_3_weight, x = input_163)[name = tensor("x_25")]; tensor x_ct_perm_0 = const()[name = tensor("x_ct_perm_0"), val = tensor([0, 2, 1])]; tensor cat_interleave_0 = const()[name = tensor("cat_interleave_0"), val = tensor(false)]; tensor x_ct = transpose(perm = x_ct_perm_0, x = x_25)[name = tensor("transpose_30")]; - tensor cat = concat(axis = var_20, interleave = cat_interleave_0, values = (cnn_window, x_ct))[name = tensor("cat")]; + tensor cat = concat(axis = var_71, interleave = cat_interleave_0, values = (cnn_window, x_ct))[name = tensor("cat")]; tensor conv_out_pad_type_0 = const()[name = tensor("conv_out_pad_type_0"), val = tensor("valid")]; tensor conv_out_strides_0 = const()[name = tensor("conv_out_strides_0"), val = tensor([1])]; tensor conv_out_pad_0 = const()[name = tensor("conv_out_pad_0"), val = tensor([0, 0])]; tensor conv_out_dilations_0 = const()[name = tensor("conv_out_dilations_0"), val = tensor([1])]; tensor conv_out_groups_0 = const()[name = tensor("conv_out_groups_0"), val = tensor(1)]; - tensor conv_out = conv(bias = encoder_cnn_bias, dilations = conv_out_dilations_0, groups = conv_out_groups_0, pad = conv_out_pad_0, pad_type = conv_out_pad_type_0, strides = conv_out_strides_0, weight = encoder_cnn_weight, x = cat)[name = tensor("conv_out")]; - tensor var_992_begin_0 = const()[name = tensor("op_992_begin_0"), val = tensor([0, 0, 4])]; - tensor var_992_end_0 = const()[name = tensor("op_992_end_0"), val = tensor([1, 256, 22])]; - tensor var_992_end_mask_0 = const()[name = tensor("op_992_end_mask_0"), val = tensor([true, true, true])]; - tensor cnn_window_new = slice_by_index(begin = var_992_begin_0, end = var_992_end_0, end_mask = var_992_end_mask_0, x = cat)[name = tensor("op_992")]; - tensor input_163_perm_0 = const()[name = tensor("input_163_perm_0"), val = tensor([0, 2, 1])]; - tensor var_994 = const()[name = tensor("op_994"), val = tensor([-1])]; - tensor input_163 = transpose(perm = input_163_perm_0, x = conv_out)[name = tensor("transpose_29")]; - tensor var_995 = reduce_l2_norm(axes = var_994, keep_dims = var_29, x = input_163)[name = tensor("op_995")]; + tensor conv_out = conv(bias = inner_encoder_cnn_bias, dilations = conv_out_dilations_0, groups = conv_out_groups_0, pad = conv_out_pad_0, pad_type = conv_out_pad_type_0, strides = conv_out_strides_0, weight = inner_encoder_cnn_weight, x = cat)[name = tensor("conv_out")]; + tensor var_1048_begin_0 = const()[name = tensor("op_1048_begin_0"), val = tensor([0, 0, 4])]; + tensor var_1048_end_0 = const()[name = tensor("op_1048_end_0"), val = tensor([1, 256, 22])]; + tensor var_1048_end_mask_0 = const()[name = tensor("op_1048_end_mask_0"), val = tensor([true, true, true])]; + tensor cnn_window_new = slice_by_index(begin = var_1048_begin_0, end = var_1048_end_0, end_mask = var_1048_end_mask_0, x = cat)[name = tensor("op_1048")]; + tensor input_165_perm_0 = const()[name = tensor("input_165_perm_0"), val = tensor([0, 2, 1])]; + tensor var_1050 = const()[name = tensor("op_1050"), val = tensor([-1])]; + tensor input_165 = transpose(perm = input_165_perm_0, x = conv_out)[name = tensor("transpose_29")]; + tensor var_1051 = reduce_l2_norm(axes = var_1050, keep_dims = var_65, x = input_165)[name = tensor("op_1051")]; tensor const_12 = const()[name = tensor("const_12"), val = tensor(0x1.fffffep+127)]; - tensor clip_0 = clip(alpha = var_11, beta = const_12, x = var_995)[name = tensor("clip_0")]; - tensor emb = real_div(x = input_163, y = clip_0)[name = tensor("emb")]; - tensor var_999_axis_0 = const()[name = tensor("op_999_axis_0"), val = tensor(0)]; - tensor enc_kv_new = stack(axis = var_999_axis_0, values = (var_206, var_420, var_634, nkv_1))[name = tensor("op_999")]; - tensor var_1001_axis_0 = const()[name = tensor("op_1001_axis_0"), val = tensor(0)]; - tensor enc_scale_new = stack(axis = var_1001_axis_0, values = (new_scale_1, new_scale_3, new_scale_5, new_scale_7))[name = tensor("op_1001")]; - tensor var_1003_axis_0 = const()[name = tensor("op_1003_axis_0"), val = tensor(0)]; - tensor enc_conv_cache_new = stack(axis = var_1003_axis_0, values = (window_9, window_19, window_29, window))[name = tensor("op_1003")]; - tensor var_1012 = const()[name = tensor("op_1012"), val = tensor(0x1.5798eep-27)]; - tensor var_1017 = const()[name = tensor("op_1017"), val = tensor(0x1.4f8b58p-17)]; - tensor var_1019 = const()[name = tensor("op_1019"), val = tensor(0x1.0c6f7ap-20)]; - tensor var_1020 = const()[name = tensor("op_1020"), val = tensor(true)]; - tensor var_1022 = const()[name = tensor("op_1022"), val = tensor(0x1p+0)]; - tensor var_1026 = const()[name = tensor("op_1026"), val = tensor(-1)]; - tensor var_1032 = const()[name = tensor("op_1032"), val = tensor(0)]; + tensor clip_0 = clip(alpha = var_79, beta = const_12, x = var_1051)[name = tensor("clip_0")]; + tensor emb = real_div(x = input_165, y = clip_0)[name = tensor("emb")]; + tensor var_1055_axis_0 = const()[name = tensor("op_1055_axis_0"), val = tensor(0)]; + tensor enc_kv_new = stack(axis = var_1055_axis_0, values = (var_262, var_476, var_690, nkv_1))[name = tensor("op_1055")]; + tensor var_1057_axis_0 = const()[name = tensor("op_1057_axis_0"), val = tensor(0)]; + tensor enc_scale_new = stack(axis = var_1057_axis_0, values = (new_scale_1, new_scale_3, new_scale_5, new_scale_7))[name = tensor("op_1057")]; + tensor var_1059_axis_0 = const()[name = tensor("op_1059_axis_0"), val = tensor(0)]; + tensor enc_conv_cache_new = stack(axis = var_1059_axis_0, values = (window_9, window_19, window_29, window))[name = tensor("op_1059")]; tensor pos = const()[name = tensor("pos"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(44395584)))]; - tensor var_1094_axes_0 = const()[name = tensor("op_1094_axes_0"), val = tensor([2])]; - tensor var_1094 = expand_dims(axes = var_1094_axes_0, x = emb)[name = tensor("op_1094")]; + tensor var_1127_axes_0 = const()[name = tensor("op_1127_axes_0"), val = tensor([2])]; + tensor var_1127 = expand_dims(axes = var_1127_axes_0, x = emb)[name = tensor("op_1127")]; tensor emb_exp_reps_0 = const()[name = tensor("emb_exp_reps_0"), val = tensor([1, 1, 12, 1])]; - tensor emb_exp = tile(reps = emb_exp_reps_0, x = var_1094)[name = tensor("emb_exp")]; - tensor input_165_interleave_0 = const()[name = tensor("input_165_interleave_0"), val = tensor(false)]; - tensor input_165 = concat(axis = var_1026, interleave = input_165_interleave_0, values = (emb_exp, pos))[name = tensor("input_165")]; - tensor x_27 = linear(bias = decoder_convert_bias, weight = decoder_convert_weight, x = input_165)[name = tensor("linear_37")]; - tensor var_1102_perm_0 = const()[name = tensor("op_1102_perm_0"), val = tensor([0, 2, 1, 3])]; - tensor var_1106 = const()[name = tensor("op_1106"), val = tensor([12, 4, 256])]; - tensor var_1102 = transpose(perm = var_1102_perm_0, x = x_27)[name = tensor("transpose_28")]; - tensor x_29 = reshape(shape = var_1106, x = var_1102)[name = tensor("x_29")]; + tensor emb_exp = tile(reps = emb_exp_reps_0, x = var_1127)[name = tensor("emb_exp")]; + tensor input_167_interleave_0 = const()[name = tensor("input_167_interleave_0"), val = tensor(false)]; + tensor input_167 = concat(axis = var_72, interleave = input_167_interleave_0, values = (emb_exp, pos))[name = tensor("input_167")]; + tensor x_27 = linear(bias = inner_decoder_convert_bias, weight = inner_decoder_convert_weight, x = input_167)[name = tensor("linear_37")]; + tensor var_1135_perm_0 = const()[name = tensor("op_1135_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_1139 = const()[name = tensor("op_1139"), val = tensor([12, 4, 256])]; + tensor var_1135 = transpose(perm = var_1135_perm_0, x = x_27)[name = tensor("transpose_28")]; + tensor x_29 = reshape(shape = var_1139, x = var_1135)[name = tensor("x_29")]; tensor prev_kv_9_begin_0 = const()[name = tensor("prev_kv_9_begin_0"), val = tensor([0, 0, 0, 0, 0])]; tensor prev_kv_9_end_0 = const()[name = tensor("prev_kv_9_end_0"), val = tensor([1, 12, 4, 64, 64])]; tensor prev_kv_9_end_mask_0 = const()[name = tensor("prev_kv_9_end_mask_0"), val = tensor([false, true, true, true, true])]; @@ -983,132 +998,132 @@ program(1.0) tensor prev_scale_9_end_mask_0 = const()[name = tensor("prev_scale_9_end_mask_0"), val = tensor([false, true])]; tensor prev_scale_9_squeeze_mask_0 = const()[name = tensor("prev_scale_9_squeeze_mask_0"), val = tensor([true, false])]; tensor prev_scale_9 = slice_by_index(begin = prev_scale_9_begin_0, end = prev_scale_9_end_0, end_mask = prev_scale_9_end_mask_0, squeeze_mask = prev_scale_9_squeeze_mask_0, x = dec_scale)[name = tensor("prev_scale_9")]; - tensor var_1114 = linear(bias = decoder_q_proj_0_bias, weight = decoder_q_proj_0_weight, x = x_29)[name = tensor("linear_38")]; - tensor var_1115 = const()[name = tensor("op_1115"), val = tensor([12, 4, 4, 64])]; - tensor var_1116 = reshape(shape = var_1115, x = var_1114)[name = tensor("op_1116")]; + tensor var_1147 = linear(bias = inner_decoder_q_proj_0_bias, weight = inner_decoder_q_proj_0_weight, x = x_29)[name = tensor("linear_38")]; + tensor var_1148 = const()[name = tensor("op_1148"), val = tensor([12, 4, 4, 64])]; + tensor var_1149 = reshape(shape = var_1148, x = var_1147)[name = tensor("op_1149")]; tensor q_9_perm_0 = const()[name = tensor("q_9_perm_0"), val = tensor([0, 2, 1, 3])]; - tensor var_1120 = linear(bias = decoder_k_proj_0_bias, weight = decoder_k_proj_0_weight, x = x_29)[name = tensor("linear_39")]; - tensor var_1121 = const()[name = tensor("op_1121"), val = tensor(0x1p-3)]; - tensor var_1122 = mul(x = var_1120, y = var_1121)[name = tensor("op_1122")]; - tensor var_1123 = const()[name = tensor("op_1123"), val = tensor([12, 4, 4, 64])]; - tensor var_1124 = reshape(shape = var_1123, x = var_1122)[name = tensor("op_1124")]; + tensor var_1153 = linear(bias = inner_decoder_k_proj_0_bias, weight = inner_decoder_k_proj_0_weight, x = x_29)[name = tensor("linear_39")]; + tensor var_1154 = const()[name = tensor("op_1154"), val = tensor(0x1p-3)]; + tensor var_1155 = mul(x = var_1153, y = var_1154)[name = tensor("op_1155")]; + tensor var_1156 = const()[name = tensor("op_1156"), val = tensor([12, 4, 4, 64])]; + tensor var_1157 = reshape(shape = var_1156, x = var_1155)[name = tensor("op_1157")]; tensor k_9_perm_0 = const()[name = tensor("k_9_perm_0"), val = tensor([0, 2, 1, 3])]; - tensor var_1128 = linear(bias = decoder_v_proj_0_bias, weight = decoder_v_proj_0_weight, x = x_29)[name = tensor("linear_40")]; - tensor var_1129 = const()[name = tensor("op_1129"), val = tensor([12, 4, 4, 64])]; - tensor var_1130 = reshape(shape = var_1129, x = var_1128)[name = tensor("op_1130")]; + tensor var_1161 = linear(bias = inner_decoder_v_proj_0_bias, weight = inner_decoder_v_proj_0_weight, x = x_29)[name = tensor("linear_40")]; + tensor var_1162 = const()[name = tensor("op_1162"), val = tensor([12, 4, 4, 64])]; + tensor var_1163 = reshape(shape = var_1162, x = var_1161)[name = tensor("op_1163")]; tensor v_9_perm_0 = const()[name = tensor("v_9_perm_0"), val = tensor([0, 2, 1, 3])]; - tensor input_169 = linear(bias = decoder_g_proj_0_bias, weight = decoder_g_proj_0_weight, x = x_29)[name = tensor("linear_41")]; + tensor input_171 = linear(bias = inner_decoder_g_proj_0_bias, weight = inner_decoder_g_proj_0_weight, x = x_29)[name = tensor("linear_41")]; tensor cumsum_mask_1_exclusive_0 = const()[name = tensor("cumsum_mask_1_exclusive_0"), val = tensor(false)]; tensor cumsum_mask_1_reverse_0 = const()[name = tensor("cumsum_mask_1_reverse_0"), val = tensor(false)]; - tensor cumsum_mask_1 = cumsum(axis = var_1032, exclusive = cumsum_mask_1_exclusive_0, reverse = cumsum_mask_1_reverse_0, x = valid_mask)[name = tensor("cumsum_mask_1")]; + tensor cumsum_mask_1 = cumsum(axis = var_69, exclusive = cumsum_mask_1_exclusive_0, reverse = cumsum_mask_1_reverse_0, x = valid_mask)[name = tensor("cumsum_mask_1")]; tensor sqrt_s0_9 = sqrt(x = prev_scale_9)[name = tensor("sqrt_s0_9")]; tensor s_t_9 = add(x = prev_scale_9, y = cumsum_mask_1)[name = tensor("s_t_9")]; tensor const_20 = const()[name = tensor("const_20"), val = tensor(0x1.fffffep+127)]; - tensor clip_1 = clip(alpha = var_1022, beta = const_20, x = s_t_9)[name = tensor("clip_1")]; + tensor clip_1 = clip(alpha = var_59, beta = const_20, x = s_t_9)[name = tensor("clip_1")]; tensor sqrt_s_t_9 = sqrt(x = clip_1)[name = tensor("sqrt_s_t_9")]; tensor qk_9_transpose_x_1 = const()[name = tensor("qk_9_transpose_x_1"), val = tensor(false)]; tensor qk_9_transpose_y_1 = const()[name = tensor("qk_9_transpose_y_1"), val = tensor(true)]; - tensor k_9 = transpose(perm = k_9_perm_0, x = var_1124)[name = tensor("transpose_26")]; - tensor q_9 = transpose(perm = q_9_perm_0, x = var_1116)[name = tensor("transpose_27")]; + tensor k_9 = transpose(perm = k_9_perm_0, x = var_1157)[name = tensor("transpose_26")]; + tensor q_9 = transpose(perm = q_9_perm_0, x = var_1149)[name = tensor("transpose_27")]; tensor qk_9 = matmul(transpose_x = qk_9_transpose_x_1, transpose_y = qk_9_transpose_y_1, x = q_9, y = k_9)[name = tensor("qk_9")]; - tensor var_1142 = const()[name = tensor("op_1142"), val = tensor([1, 4])]; - tensor var_1143 = reshape(shape = var_1142, x = valid_mask)[name = tensor("op_1143")]; - tensor causal_with_valid_1 = mul(x = decoder__causal_mask, y = var_1143)[name = tensor("causal_with_valid_1")]; - tensor var_1145 = const()[name = tensor("op_1145"), val = tensor([4, 1])]; - tensor var_1146 = reshape(shape = var_1145, x = sqrt_s_t_9)[name = tensor("op_1146")]; - tensor M_9 = real_div(x = causal_with_valid_1, y = var_1146)[name = tensor("M_9")]; - tensor var_1148 = mul(x = qk_9, y = M_9)[name = tensor("op_1148")]; + tensor var_1175 = const()[name = tensor("op_1175"), val = tensor([1, 4])]; + tensor var_1176 = reshape(shape = var_1175, x = valid_mask)[name = tensor("op_1176")]; + tensor causal_with_valid_1 = mul(x = inner_decoder__causal_mask, y = var_1176)[name = tensor("causal_with_valid_1")]; + tensor var_1178 = const()[name = tensor("op_1178"), val = tensor([4, 1])]; + tensor var_1179 = reshape(shape = var_1178, x = sqrt_s_t_9)[name = tensor("op_1179")]; + tensor M_9 = real_div(x = causal_with_valid_1, y = var_1179)[name = tensor("M_9")]; + tensor var_1181 = mul(x = qk_9, y = M_9)[name = tensor("op_1181")]; tensor inner_9_transpose_x_0 = const()[name = tensor("inner_9_transpose_x_0"), val = tensor(false)]; tensor inner_9_transpose_y_0 = const()[name = tensor("inner_9_transpose_y_0"), val = tensor(false)]; - tensor v_9 = transpose(perm = v_9_perm_0, x = var_1130)[name = tensor("transpose_25")]; - tensor inner_9 = matmul(transpose_x = inner_9_transpose_x_0, transpose_y = inner_9_transpose_y_0, x = var_1148, y = v_9)[name = tensor("inner_9")]; - tensor var_1150_transpose_x_0 = const()[name = tensor("op_1150_transpose_x_0"), val = tensor(false)]; - tensor var_1150_transpose_y_0 = const()[name = tensor("op_1150_transpose_y_0"), val = tensor(false)]; - tensor var_1150 = matmul(transpose_x = var_1150_transpose_x_0, transpose_y = var_1150_transpose_y_0, x = q_9, y = prev_kv_9)[name = tensor("op_1150")]; - tensor var_1151 = real_div(x = sqrt_s0_9, y = sqrt_s_t_9)[name = tensor("op_1151")]; - tensor var_1152 = const()[name = tensor("op_1152"), val = tensor([1, 1, 4, 1])]; - tensor var_1153 = reshape(shape = var_1152, x = var_1151)[name = tensor("op_1153")]; - tensor cross_9 = mul(x = var_1150, y = var_1153)[name = tensor("cross_9")]; + tensor v_9 = transpose(perm = v_9_perm_0, x = var_1163)[name = tensor("transpose_25")]; + tensor inner_9 = matmul(transpose_x = inner_9_transpose_x_0, transpose_y = inner_9_transpose_y_0, x = var_1181, y = v_9)[name = tensor("inner_9")]; + tensor var_1183_transpose_x_0 = const()[name = tensor("op_1183_transpose_x_0"), val = tensor(false)]; + tensor var_1183_transpose_y_0 = const()[name = tensor("op_1183_transpose_y_0"), val = tensor(false)]; + tensor var_1183 = matmul(transpose_x = var_1183_transpose_x_0, transpose_y = var_1183_transpose_y_0, x = q_9, y = prev_kv_9)[name = tensor("op_1183")]; + tensor var_1184 = real_div(x = sqrt_s0_9, y = sqrt_s_t_9)[name = tensor("op_1184")]; + tensor var_1185 = const()[name = tensor("op_1185"), val = tensor([1, 1, 4, 1])]; + tensor var_1186 = reshape(shape = var_1185, x = var_1184)[name = tensor("op_1186")]; + tensor cross_9 = mul(x = var_1183, y = var_1186)[name = tensor("cross_9")]; tensor out_25 = add(x = inner_9, y = cross_9)[name = tensor("out_25")]; - tensor var_1156 = const()[name = tensor("op_1156"), val = tensor([1, 1, 4, 1])]; - tensor var_1157 = reshape(shape = var_1156, x = valid_mask)[name = tensor("op_1157")]; - tensor v_masked_1 = mul(x = v_9, y = var_1157)[name = tensor("v_masked_1")]; - tensor var_1159 = mul(x = prev_kv_9, y = sqrt_s0_9)[name = tensor("op_1159")]; - tensor var_1161_transpose_x_1 = const()[name = tensor("op_1161_transpose_x_1"), val = tensor(true)]; - tensor var_1161_transpose_y_1 = const()[name = tensor("op_1161_transpose_y_1"), val = tensor(false)]; - tensor var_1161 = matmul(transpose_x = var_1161_transpose_x_1, transpose_y = var_1161_transpose_y_1, x = k_9, y = v_masked_1)[name = tensor("op_1161")]; - tensor new_kv_unnorm_9 = add(x = var_1159, y = var_1161)[name = tensor("new_kv_unnorm_9")]; - tensor var_1163_keep_dims_0 = const()[name = tensor("op_1163_keep_dims_0"), val = tensor(false)]; - tensor var_1163 = reduce_sum(keep_dims = var_1163_keep_dims_0, x = valid_mask)[name = tensor("op_1163")]; - tensor var_1164 = const()[name = tensor("op_1164"), val = tensor([1])]; - tensor var_1165 = reshape(shape = var_1164, x = var_1163)[name = tensor("op_1165")]; - tensor new_scale_9 = add(x = prev_scale_9, y = var_1165)[name = tensor("new_scale_9")]; + tensor var_1189 = const()[name = tensor("op_1189"), val = tensor([1, 1, 4, 1])]; + tensor var_1190 = reshape(shape = var_1189, x = valid_mask)[name = tensor("op_1190")]; + tensor v_masked_1 = mul(x = v_9, y = var_1190)[name = tensor("v_masked_1")]; + tensor var_1192 = mul(x = prev_kv_9, y = sqrt_s0_9)[name = tensor("op_1192")]; + tensor var_1194_transpose_x_1 = const()[name = tensor("op_1194_transpose_x_1"), val = tensor(true)]; + tensor var_1194_transpose_y_1 = const()[name = tensor("op_1194_transpose_y_1"), val = tensor(false)]; + tensor var_1194 = matmul(transpose_x = var_1194_transpose_x_1, transpose_y = var_1194_transpose_y_1, x = k_9, y = v_masked_1)[name = tensor("op_1194")]; + tensor new_kv_unnorm_9 = add(x = var_1192, y = var_1194)[name = tensor("new_kv_unnorm_9")]; + tensor var_1196_keep_dims_0 = const()[name = tensor("op_1196_keep_dims_0"), val = tensor(false)]; + tensor var_1196 = reduce_sum(keep_dims = var_1196_keep_dims_0, x = valid_mask)[name = tensor("op_1196")]; + tensor var_1197 = const()[name = tensor("op_1197"), val = tensor([1])]; + tensor var_1198 = reshape(shape = var_1197, x = var_1196)[name = tensor("op_1198")]; + tensor new_scale_9 = add(x = prev_scale_9, y = var_1198)[name = tensor("new_scale_9")]; tensor const_21 = const()[name = tensor("const_21"), val = tensor(0x1.fffffep+127)]; - tensor clip_2 = clip(alpha = var_1022, beta = const_21, x = new_scale_9)[name = tensor("clip_2")]; + tensor clip_2 = clip(alpha = var_59, beta = const_21, x = new_scale_9)[name = tensor("clip_2")]; tensor sqrt_new_scale_1 = sqrt(x = clip_2)[name = tensor("sqrt_new_scale_1")]; - tensor var_1169 = real_div(x = new_kv_unnorm_9, y = sqrt_new_scale_1)[name = tensor("op_1169")]; - tensor var_1170_perm_0 = const()[name = tensor("op_1170_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_1202 = real_div(x = new_kv_unnorm_9, y = sqrt_new_scale_1)[name = tensor("op_1202")]; + tensor var_1203_perm_0 = const()[name = tensor("op_1203_perm_0"), val = tensor([0, 2, 1, 3])]; tensor out_27_axes_0 = const()[name = tensor("out_27_axes_0"), val = tensor([-1])]; - tensor var_1170 = transpose(perm = var_1170_perm_0, x = out_25)[name = tensor("transpose_24")]; - tensor out_27 = layer_norm(axes = out_27_axes_0, epsilon = var_1019, x = var_1170)[name = tensor("out_27")]; - tensor var_1174 = const()[name = tensor("op_1174"), val = tensor([12, 4, 256])]; - tensor out_29 = reshape(shape = var_1174, x = out_27)[name = tensor("out_29")]; - tensor var_1176 = silu(x = input_169)[name = tensor("op_1176")]; - tensor input_171 = mul(x = var_1176, y = out_29)[name = tensor("input_171")]; - tensor ret_out_9 = linear(bias = decoder_out_proj_0_bias, weight = decoder_out_proj_0_weight, x = input_171)[name = tensor("linear_42")]; - tensor input_173 = add(x = x_29, y = ret_out_9)[name = tensor("input_173")]; + tensor var_1203 = transpose(perm = var_1203_perm_0, x = out_25)[name = tensor("transpose_24")]; + tensor out_27 = layer_norm(axes = out_27_axes_0, epsilon = var_74, x = var_1203)[name = tensor("out_27")]; + tensor var_1207 = const()[name = tensor("op_1207"), val = tensor([12, 4, 256])]; + tensor out_29 = reshape(shape = var_1207, x = out_27)[name = tensor("out_29")]; + tensor var_1209 = silu(x = input_171)[name = tensor("op_1209")]; + tensor input_173 = mul(x = var_1209, y = out_29)[name = tensor("input_173")]; + tensor ret_out_9 = linear(bias = inner_decoder_out_proj_0_bias, weight = inner_decoder_out_proj_0_weight, x = input_173)[name = tensor("linear_42")]; + tensor input_175 = add(x = x_29, y = ret_out_9)[name = tensor("input_175")]; tensor xt_1_axes_0 = const()[name = tensor("xt_1_axes_0"), val = tensor([-1])]; - tensor xt_1 = layer_norm(axes = xt_1_axes_0, beta = decoder_norm11_0_bias, epsilon = var_1017, gamma = decoder_norm11_0_weight, x = input_173)[name = tensor("xt_1")]; - tensor var_1186 = const()[name = tensor("op_1186"), val = tensor([1, 12, 4, 256])]; - tensor var_1187 = reshape(shape = var_1186, x = xt_1)[name = tensor("op_1187")]; - tensor var_1188_perm_0 = const()[name = tensor("op_1188_perm_0"), val = tensor([0, 2, 1, 3])]; - tensor var_1191 = const()[name = tensor("op_1191"), val = tensor([4, 12, 256])]; - tensor var_1188 = transpose(perm = var_1188_perm_0, x = var_1187)[name = tensor("transpose_23")]; - tensor query_1 = reshape(shape = var_1191, x = var_1188)[name = tensor("query_1")]; + tensor xt_1 = layer_norm(axes = xt_1_axes_0, beta = inner_decoder_norm11_0_bias, epsilon = var_66, gamma = inner_decoder_norm11_0_weight, x = input_175)[name = tensor("xt_1")]; + tensor var_1219 = const()[name = tensor("op_1219"), val = tensor([1, 12, 4, 256])]; + tensor var_1220 = reshape(shape = var_1219, x = xt_1)[name = tensor("op_1220")]; + tensor var_1221_perm_0 = const()[name = tensor("op_1221_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_1224 = const()[name = tensor("op_1224"), val = tensor([4, 12, 256])]; + tensor var_1221 = transpose(perm = var_1221_perm_0, x = var_1220)[name = tensor("transpose_23")]; + tensor query_1 = reshape(shape = var_1224, x = var_1221)[name = tensor("query_1")]; tensor query_3_perm_0 = const()[name = tensor("query_3_perm_0"), val = tensor([1, 0, 2])]; tensor query_3 = transpose(perm = query_3_perm_0, x = query_1)[name = tensor("transpose_22")]; - tensor var_1214 = linear(bias = decoder_self_attn2_0_in_proj_bias, weight = decoder_self_attn2_0_in_proj_weight, x = query_3)[name = tensor("linear_43")]; + tensor var_1247 = linear(bias = inner_decoder_self_attn2_0_in_proj_bias, weight = inner_decoder_self_attn2_0_in_proj_weight, x = query_3)[name = tensor("linear_43")]; tensor concat_1 = const()[name = tensor("concat_1"), val = tensor([12, 4, 3, 256])]; - tensor var_1216 = reshape(shape = concat_1, x = var_1214)[name = tensor("op_1216")]; - tensor var_1217_axes_0 = const()[name = tensor("op_1217_axes_0"), val = tensor([0])]; - tensor var_1217 = expand_dims(axes = var_1217_axes_0, x = var_1216)[name = tensor("op_1217")]; - tensor var_1218_perm_0 = const()[name = tensor("op_1218_perm_0"), val = tensor([-2, 1, 2, 0, 4])]; - tensor var_1219_axes_0 = const()[name = tensor("op_1219_axes_0"), val = tensor([-2])]; - tensor var_1218 = transpose(perm = var_1218_perm_0, x = var_1217)[name = tensor("transpose_21")]; - tensor var_1219 = squeeze(axes = var_1219_axes_0, x = var_1218)[name = tensor("op_1219")]; + tensor var_1249 = reshape(shape = concat_1, x = var_1247)[name = tensor("op_1249")]; + tensor var_1250_axes_0 = const()[name = tensor("op_1250_axes_0"), val = tensor([0])]; + tensor var_1250 = expand_dims(axes = var_1250_axes_0, x = var_1249)[name = tensor("op_1250")]; + tensor var_1251_perm_0 = const()[name = tensor("op_1251_perm_0"), val = tensor([-2, 1, 2, 0, 4])]; + tensor var_1252_axes_0 = const()[name = tensor("op_1252_axes_0"), val = tensor([-2])]; + tensor var_1251 = transpose(perm = var_1251_perm_0, x = var_1250)[name = tensor("transpose_21")]; + tensor var_1252 = squeeze(axes = var_1252_axes_0, x = var_1251)[name = tensor("op_1252")]; tensor q_11_begin_0 = const()[name = tensor("q_11_begin_0"), val = tensor([0, 0, 0, 0])]; tensor q_11_end_0 = const()[name = tensor("q_11_end_0"), val = tensor([1, 12, 4, 256])]; tensor q_11_end_mask_0 = const()[name = tensor("q_11_end_mask_0"), val = tensor([false, true, true, true])]; tensor q_11_squeeze_mask_0 = const()[name = tensor("q_11_squeeze_mask_0"), val = tensor([true, false, false, false])]; - tensor q_11 = slice_by_index(begin = q_11_begin_0, end = q_11_end_0, end_mask = q_11_end_mask_0, squeeze_mask = q_11_squeeze_mask_0, x = var_1219)[name = tensor("q_11")]; + tensor q_11 = slice_by_index(begin = q_11_begin_0, end = q_11_end_0, end_mask = q_11_end_mask_0, squeeze_mask = q_11_squeeze_mask_0, x = var_1252)[name = tensor("q_11")]; tensor k_11_begin_0 = const()[name = tensor("k_11_begin_0"), val = tensor([1, 0, 0, 0])]; tensor k_11_end_0 = const()[name = tensor("k_11_end_0"), val = tensor([2, 12, 4, 256])]; tensor k_11_end_mask_0 = const()[name = tensor("k_11_end_mask_0"), val = tensor([false, true, true, true])]; tensor k_11_squeeze_mask_0 = const()[name = tensor("k_11_squeeze_mask_0"), val = tensor([true, false, false, false])]; - tensor k_11 = slice_by_index(begin = k_11_begin_0, end = k_11_end_0, end_mask = k_11_end_mask_0, squeeze_mask = k_11_squeeze_mask_0, x = var_1219)[name = tensor("k_11")]; + tensor k_11 = slice_by_index(begin = k_11_begin_0, end = k_11_end_0, end_mask = k_11_end_mask_0, squeeze_mask = k_11_squeeze_mask_0, x = var_1252)[name = tensor("k_11")]; tensor v_11_begin_0 = const()[name = tensor("v_11_begin_0"), val = tensor([2, 0, 0, 0])]; tensor v_11_end_0 = const()[name = tensor("v_11_end_0"), val = tensor([3, 12, 4, 256])]; tensor v_11_end_mask_0 = const()[name = tensor("v_11_end_mask_0"), val = tensor([false, true, true, true])]; tensor v_11_squeeze_mask_0 = const()[name = tensor("v_11_squeeze_mask_0"), val = tensor([true, false, false, false])]; - tensor v_11 = slice_by_index(begin = v_11_begin_0, end = v_11_end_0, end_mask = v_11_end_mask_0, squeeze_mask = v_11_squeeze_mask_0, x = var_1219)[name = tensor("v_11")]; - tensor var_1227 = const()[name = tensor("op_1227"), val = tensor([12, 16, 64])]; - tensor var_1228 = reshape(shape = var_1227, x = q_11)[name = tensor("op_1228")]; + tensor v_11 = slice_by_index(begin = v_11_begin_0, end = v_11_end_0, end_mask = v_11_end_mask_0, squeeze_mask = v_11_squeeze_mask_0, x = var_1252)[name = tensor("v_11")]; + tensor var_1260 = const()[name = tensor("op_1260"), val = tensor([12, 16, 64])]; + tensor var_1261 = reshape(shape = var_1260, x = q_11)[name = tensor("op_1261")]; tensor q_13_perm_0 = const()[name = tensor("q_13_perm_0"), val = tensor([1, 0, 2])]; - tensor var_1234 = const()[name = tensor("op_1234"), val = tensor([12, 16, 64])]; - tensor var_1235 = reshape(shape = var_1234, x = k_11)[name = tensor("op_1235")]; + tensor var_1267 = const()[name = tensor("op_1267"), val = tensor([12, 16, 64])]; + tensor var_1268 = reshape(shape = var_1267, x = k_11)[name = tensor("op_1268")]; tensor k_13_perm_0 = const()[name = tensor("k_13_perm_0"), val = tensor([1, 0, 2])]; - tensor var_1241 = const()[name = tensor("op_1241"), val = tensor([12, 16, 64])]; - tensor var_1242 = reshape(shape = var_1241, x = v_11)[name = tensor("op_1242")]; + tensor var_1274 = const()[name = tensor("op_1274"), val = tensor([12, 16, 64])]; + tensor var_1275 = reshape(shape = var_1274, x = v_11)[name = tensor("op_1275")]; tensor v_13_perm_0 = const()[name = tensor("v_13_perm_0"), val = tensor([1, 0, 2])]; - tensor var_1245 = const()[name = tensor("op_1245"), val = tensor([4, 4, 12, 64])]; - tensor q_13 = transpose(perm = q_13_perm_0, x = var_1228)[name = tensor("transpose_20")]; - tensor q_15 = reshape(shape = var_1245, x = q_13)[name = tensor("q_15")]; - tensor var_1247 = const()[name = tensor("op_1247"), val = tensor([4, 4, 12, 64])]; - tensor k_13 = transpose(perm = k_13_perm_0, x = var_1235)[name = tensor("transpose_19")]; - tensor k_15 = reshape(shape = var_1247, x = k_13)[name = tensor("k_15")]; - tensor var_1249 = const()[name = tensor("op_1249"), val = tensor([4, 4, 12, 64])]; - tensor v_13 = transpose(perm = v_13_perm_0, x = var_1242)[name = tensor("transpose_18")]; - tensor v_15 = reshape(shape = var_1249, x = v_13)[name = tensor("v_15")]; + tensor var_1278 = const()[name = tensor("op_1278"), val = tensor([4, 4, 12, 64])]; + tensor q_13 = transpose(perm = q_13_perm_0, x = var_1261)[name = tensor("transpose_20")]; + tensor q_15 = reshape(shape = var_1278, x = q_13)[name = tensor("q_15")]; + tensor var_1280 = const()[name = tensor("op_1280"), val = tensor([4, 4, 12, 64])]; + tensor k_13 = transpose(perm = k_13_perm_0, x = var_1268)[name = tensor("transpose_19")]; + tensor k_15 = reshape(shape = var_1280, x = k_13)[name = tensor("k_15")]; + tensor var_1282 = const()[name = tensor("op_1282"), val = tensor([4, 4, 12, 64])]; + tensor v_13 = transpose(perm = v_13_perm_0, x = var_1275)[name = tensor("transpose_18")]; + tensor v_15 = reshape(shape = var_1282, x = v_13)[name = tensor("v_15")]; tensor mul_1_y_0 = const()[name = tensor("mul_1_y_0"), val = tensor(0x1p-3)]; tensor mul_1 = mul(x = q_15, y = mul_1_y_0)[name = tensor("mul_1")]; tensor matmul_0_transpose_y_0 = const()[name = tensor("matmul_0_transpose_y_0"), val = tensor(true)]; @@ -1119,30 +1134,30 @@ program(1.0) tensor attn_output_1_transpose_x_0 = const()[name = tensor("attn_output_1_transpose_x_0"), val = tensor(false)]; tensor attn_output_1_transpose_y_0 = const()[name = tensor("attn_output_1_transpose_y_0"), val = tensor(false)]; tensor attn_output_1 = matmul(transpose_x = attn_output_1_transpose_x_0, transpose_y = attn_output_1_transpose_y_0, x = softmax_0, y = v_15)[name = tensor("attn_output_1")]; - tensor var_1252 = const()[name = tensor("op_1252"), val = tensor([2, 0, 1, 3])]; - tensor var_1257 = const()[name = tensor("op_1257"), val = tensor([48, 256])]; - tensor var_1253 = transpose(perm = var_1252, x = attn_output_1)[name = tensor("transpose_17")]; - tensor attn_output_3 = reshape(shape = var_1257, x = var_1253)[name = tensor("attn_output_3")]; - tensor attn_output_5 = linear(bias = decoder_self_attn2_0_out_proj_bias, weight = decoder_self_attn2_0_out_proj_weight, x = attn_output_3)[name = tensor("linear_44")]; - tensor var_1261 = const()[name = tensor("op_1261"), val = tensor([12, 4, 256])]; - tensor attn_output_7 = reshape(shape = var_1261, x = attn_output_5)[name = tensor("attn_output_7")]; + tensor var_1285 = const()[name = tensor("op_1285"), val = tensor([2, 0, 1, 3])]; + tensor var_1290 = const()[name = tensor("op_1290"), val = tensor([48, 256])]; + tensor var_1286 = transpose(perm = var_1285, x = attn_output_1)[name = tensor("transpose_17")]; + tensor attn_output_3 = reshape(shape = var_1290, x = var_1286)[name = tensor("attn_output_3")]; + tensor attn_output_5 = linear(bias = inner_decoder_self_attn2_0_out_proj_bias, weight = inner_decoder_self_attn2_0_out_proj_weight, x = attn_output_3)[name = tensor("linear_44")]; + tensor var_1294 = const()[name = tensor("op_1294"), val = tensor([12, 4, 256])]; + tensor attn_output_7 = reshape(shape = var_1294, x = attn_output_5)[name = tensor("attn_output_7")]; tensor sa2_out_1_perm_0 = const()[name = tensor("sa2_out_1_perm_0"), val = tensor([1, 0, 2])]; tensor sa2_out_1 = transpose(perm = sa2_out_1_perm_0, x = attn_output_7)[name = tensor("transpose_16")]; - tensor input_175 = add(x = query_1, y = sa2_out_1)[name = tensor("input_175")]; - tensor input_177_axes_0 = const()[name = tensor("input_177_axes_0"), val = tensor([-1])]; - tensor input_177 = layer_norm(axes = input_177_axes_0, beta = decoder_norm21_0_bias, epsilon = var_1017, gamma = decoder_norm21_0_weight, x = input_175)[name = tensor("input_177")]; - tensor input_179 = linear(bias = decoder_linear1_0_bias, weight = decoder_linear1_0_weight, x = input_177)[name = tensor("linear_45")]; - tensor input_181 = relu(x = input_179)[name = tensor("input_181")]; - tensor ffn_out_1 = linear(bias = decoder_linear2_0_bias, weight = decoder_linear2_0_weight, x = input_181)[name = tensor("linear_46")]; - tensor input_183 = add(x = input_177, y = ffn_out_1)[name = tensor("input_183")]; + tensor input_177 = add(x = query_1, y = sa2_out_1)[name = tensor("input_177")]; + tensor input_179_axes_0 = const()[name = tensor("input_179_axes_0"), val = tensor([-1])]; + tensor input_179 = layer_norm(axes = input_179_axes_0, beta = inner_decoder_norm21_0_bias, epsilon = var_66, gamma = inner_decoder_norm21_0_weight, x = input_177)[name = tensor("input_179")]; + tensor input_181 = linear(bias = inner_decoder_linear1_0_bias, weight = inner_decoder_linear1_0_weight, x = input_179)[name = tensor("linear_45")]; + tensor input_183 = relu(x = input_181)[name = tensor("input_183")]; + tensor ffn_out_1 = linear(bias = inner_decoder_linear2_0_bias, weight = inner_decoder_linear2_0_weight, x = input_183)[name = tensor("linear_46")]; + tensor input_185 = add(x = input_179, y = ffn_out_1)[name = tensor("input_185")]; tensor xt_3_axes_0 = const()[name = tensor("xt_3_axes_0"), val = tensor([-1])]; - tensor xt_3 = layer_norm(axes = xt_3_axes_0, beta = decoder_norm22_0_bias, epsilon = var_1017, gamma = decoder_norm22_0_weight, x = input_183)[name = tensor("xt_3")]; - tensor var_1281 = const()[name = tensor("op_1281"), val = tensor([1, 4, 12, 256])]; - tensor x_31 = reshape(shape = var_1281, x = xt_3)[name = tensor("x_31")]; - tensor var_1283_perm_0 = const()[name = tensor("op_1283_perm_0"), val = tensor([0, 2, 1, 3])]; - tensor var_1287 = const()[name = tensor("op_1287"), val = tensor([12, 4, 256])]; - tensor var_1283 = transpose(perm = var_1283_perm_0, x = x_31)[name = tensor("transpose_15")]; - tensor x = reshape(shape = var_1287, x = var_1283)[name = tensor("x")]; + tensor xt_3 = layer_norm(axes = xt_3_axes_0, beta = inner_decoder_norm22_0_bias, epsilon = var_66, gamma = inner_decoder_norm22_0_weight, x = input_185)[name = tensor("xt_3")]; + tensor var_1314 = const()[name = tensor("op_1314"), val = tensor([1, 4, 12, 256])]; + tensor x_31 = reshape(shape = var_1314, x = xt_3)[name = tensor("x_31")]; + tensor var_1316_perm_0 = const()[name = tensor("op_1316_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_1320 = const()[name = tensor("op_1320"), val = tensor([12, 4, 256])]; + tensor var_1316 = transpose(perm = var_1316_perm_0, x = x_31)[name = tensor("transpose_15")]; + tensor x = reshape(shape = var_1320, x = var_1316)[name = tensor("x")]; tensor prev_kv_begin_0 = const()[name = tensor("prev_kv_begin_0"), val = tensor([1, 0, 0, 0, 0])]; tensor prev_kv_end_0 = const()[name = tensor("prev_kv_end_0"), val = tensor([2, 12, 4, 64, 64])]; tensor prev_kv_end_mask_0 = const()[name = tensor("prev_kv_end_mask_0"), val = tensor([false, true, true, true, true])]; @@ -1153,120 +1168,120 @@ program(1.0) tensor prev_scale_end_mask_0 = const()[name = tensor("prev_scale_end_mask_0"), val = tensor([false, true])]; tensor prev_scale_squeeze_mask_0 = const()[name = tensor("prev_scale_squeeze_mask_0"), val = tensor([true, false])]; tensor prev_scale = slice_by_index(begin = prev_scale_begin_0, end = prev_scale_end_0, end_mask = prev_scale_end_mask_0, squeeze_mask = prev_scale_squeeze_mask_0, x = dec_scale)[name = tensor("prev_scale")]; - tensor var_1295 = linear(bias = decoder_q_proj_1_bias, weight = decoder_q_proj_1_weight, x = x)[name = tensor("linear_47")]; - tensor var_1296 = const()[name = tensor("op_1296"), val = tensor([12, 4, 4, 64])]; - tensor var_1297 = reshape(shape = var_1296, x = var_1295)[name = tensor("op_1297")]; + tensor var_1328 = linear(bias = inner_decoder_q_proj_1_bias, weight = inner_decoder_q_proj_1_weight, x = x)[name = tensor("linear_47")]; + tensor var_1329 = const()[name = tensor("op_1329"), val = tensor([12, 4, 4, 64])]; + tensor var_1330 = reshape(shape = var_1329, x = var_1328)[name = tensor("op_1330")]; tensor q_17_perm_0 = const()[name = tensor("q_17_perm_0"), val = tensor([0, 2, 1, 3])]; - tensor var_1301 = linear(bias = decoder_k_proj_1_bias, weight = decoder_k_proj_1_weight, x = x)[name = tensor("linear_48")]; - tensor var_1302 = const()[name = tensor("op_1302"), val = tensor(0x1p-3)]; - tensor var_1303 = mul(x = var_1301, y = var_1302)[name = tensor("op_1303")]; - tensor var_1304 = const()[name = tensor("op_1304"), val = tensor([12, 4, 4, 64])]; - tensor var_1305 = reshape(shape = var_1304, x = var_1303)[name = tensor("op_1305")]; + tensor var_1334 = linear(bias = inner_decoder_k_proj_1_bias, weight = inner_decoder_k_proj_1_weight, x = x)[name = tensor("linear_48")]; + tensor var_1335 = const()[name = tensor("op_1335"), val = tensor(0x1p-3)]; + tensor var_1336 = mul(x = var_1334, y = var_1335)[name = tensor("op_1336")]; + tensor var_1337 = const()[name = tensor("op_1337"), val = tensor([12, 4, 4, 64])]; + tensor var_1338 = reshape(shape = var_1337, x = var_1336)[name = tensor("op_1338")]; tensor k_17_perm_0 = const()[name = tensor("k_17_perm_0"), val = tensor([0, 2, 1, 3])]; - tensor var_1309 = linear(bias = decoder_v_proj_1_bias, weight = decoder_v_proj_1_weight, x = x)[name = tensor("linear_49")]; - tensor var_1310 = const()[name = tensor("op_1310"), val = tensor([12, 4, 4, 64])]; - tensor var_1311 = reshape(shape = var_1310, x = var_1309)[name = tensor("op_1311")]; + tensor var_1342 = linear(bias = inner_decoder_v_proj_1_bias, weight = inner_decoder_v_proj_1_weight, x = x)[name = tensor("linear_49")]; + tensor var_1343 = const()[name = tensor("op_1343"), val = tensor([12, 4, 4, 64])]; + tensor var_1344 = reshape(shape = var_1343, x = var_1342)[name = tensor("op_1344")]; tensor v_17_perm_0 = const()[name = tensor("v_17_perm_0"), val = tensor([0, 2, 1, 3])]; - tensor input_187 = linear(bias = decoder_g_proj_1_bias, weight = decoder_g_proj_1_weight, x = x)[name = tensor("linear_50")]; + tensor input_189 = linear(bias = inner_decoder_g_proj_1_bias, weight = inner_decoder_g_proj_1_weight, x = x)[name = tensor("linear_50")]; tensor sqrt_s0 = sqrt(x = prev_scale)[name = tensor("sqrt_s0")]; tensor s_t = add(x = prev_scale, y = cumsum_mask_1)[name = tensor("s_t")]; tensor const_32 = const()[name = tensor("const_32"), val = tensor(0x1.fffffep+127)]; - tensor clip_3 = clip(alpha = var_1022, beta = const_32, x = s_t)[name = tensor("clip_3")]; + tensor clip_3 = clip(alpha = var_59, beta = const_32, x = s_t)[name = tensor("clip_3")]; tensor sqrt_s_t = sqrt(x = clip_3)[name = tensor("sqrt_s_t")]; tensor qk_transpose_x_1 = const()[name = tensor("qk_transpose_x_1"), val = tensor(false)]; tensor qk_transpose_y_1 = const()[name = tensor("qk_transpose_y_1"), val = tensor(true)]; - tensor k_17 = transpose(perm = k_17_perm_0, x = var_1305)[name = tensor("transpose_13")]; - tensor q_17 = transpose(perm = q_17_perm_0, x = var_1297)[name = tensor("transpose_14")]; + tensor k_17 = transpose(perm = k_17_perm_0, x = var_1338)[name = tensor("transpose_13")]; + tensor q_17 = transpose(perm = q_17_perm_0, x = var_1330)[name = tensor("transpose_14")]; tensor qk = matmul(transpose_x = qk_transpose_x_1, transpose_y = qk_transpose_y_1, x = q_17, y = k_17)[name = tensor("qk")]; - tensor var_1326 = const()[name = tensor("op_1326"), val = tensor([4, 1])]; - tensor var_1327 = reshape(shape = var_1326, x = sqrt_s_t)[name = tensor("op_1327")]; - tensor M = real_div(x = causal_with_valid_1, y = var_1327)[name = tensor("M")]; - tensor var_1329 = mul(x = qk, y = M)[name = tensor("op_1329")]; - tensor inner_transpose_x_0 = const()[name = tensor("inner_transpose_x_0"), val = tensor(false)]; - tensor inner_transpose_y_0 = const()[name = tensor("inner_transpose_y_0"), val = tensor(false)]; - tensor v_17 = transpose(perm = v_17_perm_0, x = var_1311)[name = tensor("transpose_12")]; - tensor inner = matmul(transpose_x = inner_transpose_x_0, transpose_y = inner_transpose_y_0, x = var_1329, y = v_17)[name = tensor("inner")]; - tensor var_1331_transpose_x_0 = const()[name = tensor("op_1331_transpose_x_0"), val = tensor(false)]; - tensor var_1331_transpose_y_0 = const()[name = tensor("op_1331_transpose_y_0"), val = tensor(false)]; - tensor var_1331 = matmul(transpose_x = var_1331_transpose_x_0, transpose_y = var_1331_transpose_y_0, x = q_17, y = prev_kv)[name = tensor("op_1331")]; - tensor var_1332 = real_div(x = sqrt_s0, y = sqrt_s_t)[name = tensor("op_1332")]; - tensor var_1333 = const()[name = tensor("op_1333"), val = tensor([1, 1, 4, 1])]; - tensor var_1334 = reshape(shape = var_1333, x = var_1332)[name = tensor("op_1334")]; - tensor cross = mul(x = var_1331, y = var_1334)[name = tensor("cross")]; - tensor out_31 = add(x = inner, y = cross)[name = tensor("out_31")]; - tensor v_masked = mul(x = v_17, y = var_1157)[name = tensor("v_masked")]; - tensor var_1340 = mul(x = prev_kv, y = sqrt_s0)[name = tensor("op_1340")]; - tensor var_1342_transpose_x_1 = const()[name = tensor("op_1342_transpose_x_1"), val = tensor(true)]; - tensor var_1342_transpose_y_1 = const()[name = tensor("op_1342_transpose_y_1"), val = tensor(false)]; - tensor var_1342 = matmul(transpose_x = var_1342_transpose_x_1, transpose_y = var_1342_transpose_y_1, x = k_17, y = v_masked)[name = tensor("op_1342")]; - tensor new_kv_unnorm = add(x = var_1340, y = var_1342)[name = tensor("new_kv_unnorm")]; - tensor new_scale = add(x = prev_scale, y = var_1165)[name = tensor("new_scale")]; + tensor var_1359 = const()[name = tensor("op_1359"), val = tensor([4, 1])]; + tensor var_1360 = reshape(shape = var_1359, x = sqrt_s_t)[name = tensor("op_1360")]; + tensor M = real_div(x = causal_with_valid_1, y = var_1360)[name = tensor("M")]; + tensor var_1362 = mul(x = qk, y = M)[name = tensor("op_1362")]; + tensor inner_11_transpose_x_0 = const()[name = tensor("inner_11_transpose_x_0"), val = tensor(false)]; + tensor inner_11_transpose_y_0 = const()[name = tensor("inner_11_transpose_y_0"), val = tensor(false)]; + tensor v_17 = transpose(perm = v_17_perm_0, x = var_1344)[name = tensor("transpose_12")]; + tensor inner_11 = matmul(transpose_x = inner_11_transpose_x_0, transpose_y = inner_11_transpose_y_0, x = var_1362, y = v_17)[name = tensor("inner_11")]; + tensor var_1364_transpose_x_0 = const()[name = tensor("op_1364_transpose_x_0"), val = tensor(false)]; + tensor var_1364_transpose_y_0 = const()[name = tensor("op_1364_transpose_y_0"), val = tensor(false)]; + tensor var_1364 = matmul(transpose_x = var_1364_transpose_x_0, transpose_y = var_1364_transpose_y_0, x = q_17, y = prev_kv)[name = tensor("op_1364")]; + tensor var_1365 = real_div(x = sqrt_s0, y = sqrt_s_t)[name = tensor("op_1365")]; + tensor var_1366 = const()[name = tensor("op_1366"), val = tensor([1, 1, 4, 1])]; + tensor var_1367 = reshape(shape = var_1366, x = var_1365)[name = tensor("op_1367")]; + tensor cross = mul(x = var_1364, y = var_1367)[name = tensor("cross")]; + tensor out_31 = add(x = inner_11, y = cross)[name = tensor("out_31")]; + tensor v_masked = mul(x = v_17, y = var_1190)[name = tensor("v_masked")]; + tensor var_1373 = mul(x = prev_kv, y = sqrt_s0)[name = tensor("op_1373")]; + tensor var_1375_transpose_x_1 = const()[name = tensor("op_1375_transpose_x_1"), val = tensor(true)]; + tensor var_1375_transpose_y_1 = const()[name = tensor("op_1375_transpose_y_1"), val = tensor(false)]; + tensor var_1375 = matmul(transpose_x = var_1375_transpose_x_1, transpose_y = var_1375_transpose_y_1, x = k_17, y = v_masked)[name = tensor("op_1375")]; + tensor new_kv_unnorm = add(x = var_1373, y = var_1375)[name = tensor("new_kv_unnorm")]; + tensor new_scale = add(x = prev_scale, y = var_1198)[name = tensor("new_scale")]; tensor const_33 = const()[name = tensor("const_33"), val = tensor(0x1.fffffep+127)]; - tensor clip_4 = clip(alpha = var_1022, beta = const_33, x = new_scale)[name = tensor("clip_4")]; + tensor clip_4 = clip(alpha = var_59, beta = const_33, x = new_scale)[name = tensor("clip_4")]; tensor sqrt_new_scale = sqrt(x = clip_4)[name = tensor("sqrt_new_scale")]; tensor nkv = real_div(x = new_kv_unnorm, y = sqrt_new_scale)[name = tensor("nkv")]; - tensor var_1351_perm_0 = const()[name = tensor("op_1351_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_1384_perm_0 = const()[name = tensor("op_1384_perm_0"), val = tensor([0, 2, 1, 3])]; tensor out_33_axes_0 = const()[name = tensor("out_33_axes_0"), val = tensor([-1])]; - tensor var_1351 = transpose(perm = var_1351_perm_0, x = out_31)[name = tensor("transpose_11")]; - tensor out_33 = layer_norm(axes = out_33_axes_0, epsilon = var_1019, x = var_1351)[name = tensor("out_33")]; - tensor var_1355 = const()[name = tensor("op_1355"), val = tensor([12, 4, 256])]; - tensor out = reshape(shape = var_1355, x = out_33)[name = tensor("out")]; - tensor var_1357 = silu(x = input_187)[name = tensor("op_1357")]; - tensor input_189 = mul(x = var_1357, y = out)[name = tensor("input_189")]; - tensor ret_out = linear(bias = decoder_out_proj_1_bias, weight = decoder_out_proj_1_weight, x = input_189)[name = tensor("linear_51")]; - tensor input_191 = add(x = x, y = ret_out)[name = tensor("input_191")]; + tensor var_1384 = transpose(perm = var_1384_perm_0, x = out_31)[name = tensor("transpose_11")]; + tensor out_33 = layer_norm(axes = out_33_axes_0, epsilon = var_74, x = var_1384)[name = tensor("out_33")]; + tensor var_1388 = const()[name = tensor("op_1388"), val = tensor([12, 4, 256])]; + tensor out = reshape(shape = var_1388, x = out_33)[name = tensor("out")]; + tensor var_1390 = silu(x = input_189)[name = tensor("op_1390")]; + tensor input_191 = mul(x = var_1390, y = out)[name = tensor("input_191")]; + tensor ret_out = linear(bias = inner_decoder_out_proj_1_bias, weight = inner_decoder_out_proj_1_weight, x = input_191)[name = tensor("linear_51")]; + tensor input_193 = add(x = x, y = ret_out)[name = tensor("input_193")]; tensor xt_5_axes_0 = const()[name = tensor("xt_5_axes_0"), val = tensor([-1])]; - tensor xt_5 = layer_norm(axes = xt_5_axes_0, beta = decoder_norm11_1_bias, epsilon = var_1017, gamma = decoder_norm11_1_weight, x = input_191)[name = tensor("xt_5")]; - tensor var_1367 = const()[name = tensor("op_1367"), val = tensor([1, 12, 4, 256])]; - tensor var_1368 = reshape(shape = var_1367, x = xt_5)[name = tensor("op_1368")]; - tensor var_1369_perm_0 = const()[name = tensor("op_1369_perm_0"), val = tensor([0, 2, 1, 3])]; - tensor var_1372 = const()[name = tensor("op_1372"), val = tensor([4, 12, 256])]; - tensor var_1369 = transpose(perm = var_1369_perm_0, x = var_1368)[name = tensor("transpose_10")]; - tensor query_5 = reshape(shape = var_1372, x = var_1369)[name = tensor("query_5")]; + tensor xt_5 = layer_norm(axes = xt_5_axes_0, beta = inner_decoder_norm11_1_bias, epsilon = var_66, gamma = inner_decoder_norm11_1_weight, x = input_193)[name = tensor("xt_5")]; + tensor var_1400 = const()[name = tensor("op_1400"), val = tensor([1, 12, 4, 256])]; + tensor var_1401 = reshape(shape = var_1400, x = xt_5)[name = tensor("op_1401")]; + tensor var_1402_perm_0 = const()[name = tensor("op_1402_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_1405 = const()[name = tensor("op_1405"), val = tensor([4, 12, 256])]; + tensor var_1402 = transpose(perm = var_1402_perm_0, x = var_1401)[name = tensor("transpose_10")]; + tensor query_5 = reshape(shape = var_1405, x = var_1402)[name = tensor("query_5")]; tensor query_perm_0 = const()[name = tensor("query_perm_0"), val = tensor([1, 0, 2])]; tensor query = transpose(perm = query_perm_0, x = query_5)[name = tensor("transpose_9")]; - tensor var_1395 = linear(bias = decoder_self_attn2_1_in_proj_bias, weight = decoder_self_attn2_1_in_proj_weight, x = query)[name = tensor("linear_52")]; + tensor var_1428 = linear(bias = inner_decoder_self_attn2_1_in_proj_bias, weight = inner_decoder_self_attn2_1_in_proj_weight, x = query)[name = tensor("linear_52")]; tensor concat_2 = const()[name = tensor("concat_2"), val = tensor([12, 4, 3, 256])]; - tensor var_1397 = reshape(shape = concat_2, x = var_1395)[name = tensor("op_1397")]; - tensor var_1398_axes_0 = const()[name = tensor("op_1398_axes_0"), val = tensor([0])]; - tensor var_1398 = expand_dims(axes = var_1398_axes_0, x = var_1397)[name = tensor("op_1398")]; - tensor var_1399_perm_0 = const()[name = tensor("op_1399_perm_0"), val = tensor([-2, 1, 2, 0, 4])]; - tensor var_1400_axes_0 = const()[name = tensor("op_1400_axes_0"), val = tensor([-2])]; - tensor var_1399 = transpose(perm = var_1399_perm_0, x = var_1398)[name = tensor("transpose_8")]; - tensor var_1400 = squeeze(axes = var_1400_axes_0, x = var_1399)[name = tensor("op_1400")]; + tensor var_1430 = reshape(shape = concat_2, x = var_1428)[name = tensor("op_1430")]; + tensor var_1431_axes_0 = const()[name = tensor("op_1431_axes_0"), val = tensor([0])]; + tensor var_1431 = expand_dims(axes = var_1431_axes_0, x = var_1430)[name = tensor("op_1431")]; + tensor var_1432_perm_0 = const()[name = tensor("op_1432_perm_0"), val = tensor([-2, 1, 2, 0, 4])]; + tensor var_1433_axes_0 = const()[name = tensor("op_1433_axes_0"), val = tensor([-2])]; + tensor var_1432 = transpose(perm = var_1432_perm_0, x = var_1431)[name = tensor("transpose_8")]; + tensor var_1433 = squeeze(axes = var_1433_axes_0, x = var_1432)[name = tensor("op_1433")]; tensor q_19_begin_0 = const()[name = tensor("q_19_begin_0"), val = tensor([0, 0, 0, 0])]; tensor q_19_end_0 = const()[name = tensor("q_19_end_0"), val = tensor([1, 12, 4, 256])]; tensor q_19_end_mask_0 = const()[name = tensor("q_19_end_mask_0"), val = tensor([false, true, true, true])]; tensor q_19_squeeze_mask_0 = const()[name = tensor("q_19_squeeze_mask_0"), val = tensor([true, false, false, false])]; - tensor q_19 = slice_by_index(begin = q_19_begin_0, end = q_19_end_0, end_mask = q_19_end_mask_0, squeeze_mask = q_19_squeeze_mask_0, x = var_1400)[name = tensor("q_19")]; + tensor q_19 = slice_by_index(begin = q_19_begin_0, end = q_19_end_0, end_mask = q_19_end_mask_0, squeeze_mask = q_19_squeeze_mask_0, x = var_1433)[name = tensor("q_19")]; tensor k_19_begin_0 = const()[name = tensor("k_19_begin_0"), val = tensor([1, 0, 0, 0])]; tensor k_19_end_0 = const()[name = tensor("k_19_end_0"), val = tensor([2, 12, 4, 256])]; tensor k_19_end_mask_0 = const()[name = tensor("k_19_end_mask_0"), val = tensor([false, true, true, true])]; tensor k_19_squeeze_mask_0 = const()[name = tensor("k_19_squeeze_mask_0"), val = tensor([true, false, false, false])]; - tensor k_19 = slice_by_index(begin = k_19_begin_0, end = k_19_end_0, end_mask = k_19_end_mask_0, squeeze_mask = k_19_squeeze_mask_0, x = var_1400)[name = tensor("k_19")]; + tensor k_19 = slice_by_index(begin = k_19_begin_0, end = k_19_end_0, end_mask = k_19_end_mask_0, squeeze_mask = k_19_squeeze_mask_0, x = var_1433)[name = tensor("k_19")]; tensor v_19_begin_0 = const()[name = tensor("v_19_begin_0"), val = tensor([2, 0, 0, 0])]; tensor v_19_end_0 = const()[name = tensor("v_19_end_0"), val = tensor([3, 12, 4, 256])]; tensor v_19_end_mask_0 = const()[name = tensor("v_19_end_mask_0"), val = tensor([false, true, true, true])]; tensor v_19_squeeze_mask_0 = const()[name = tensor("v_19_squeeze_mask_0"), val = tensor([true, false, false, false])]; - tensor v_19 = slice_by_index(begin = v_19_begin_0, end = v_19_end_0, end_mask = v_19_end_mask_0, squeeze_mask = v_19_squeeze_mask_0, x = var_1400)[name = tensor("v_19")]; - tensor var_1408 = const()[name = tensor("op_1408"), val = tensor([12, 16, 64])]; - tensor var_1409 = reshape(shape = var_1408, x = q_19)[name = tensor("op_1409")]; + tensor v_19 = slice_by_index(begin = v_19_begin_0, end = v_19_end_0, end_mask = v_19_end_mask_0, squeeze_mask = v_19_squeeze_mask_0, x = var_1433)[name = tensor("v_19")]; + tensor var_1441 = const()[name = tensor("op_1441"), val = tensor([12, 16, 64])]; + tensor var_1442 = reshape(shape = var_1441, x = q_19)[name = tensor("op_1442")]; tensor q_21_perm_0 = const()[name = tensor("q_21_perm_0"), val = tensor([1, 0, 2])]; - tensor var_1415 = const()[name = tensor("op_1415"), val = tensor([12, 16, 64])]; - tensor var_1416 = reshape(shape = var_1415, x = k_19)[name = tensor("op_1416")]; + tensor var_1448 = const()[name = tensor("op_1448"), val = tensor([12, 16, 64])]; + tensor var_1449 = reshape(shape = var_1448, x = k_19)[name = tensor("op_1449")]; tensor k_21_perm_0 = const()[name = tensor("k_21_perm_0"), val = tensor([1, 0, 2])]; - tensor var_1422 = const()[name = tensor("op_1422"), val = tensor([12, 16, 64])]; - tensor var_1423 = reshape(shape = var_1422, x = v_19)[name = tensor("op_1423")]; + tensor var_1455 = const()[name = tensor("op_1455"), val = tensor([12, 16, 64])]; + tensor var_1456 = reshape(shape = var_1455, x = v_19)[name = tensor("op_1456")]; tensor v_21_perm_0 = const()[name = tensor("v_21_perm_0"), val = tensor([1, 0, 2])]; - tensor var_1426 = const()[name = tensor("op_1426"), val = tensor([4, 4, 12, 64])]; - tensor q_21 = transpose(perm = q_21_perm_0, x = var_1409)[name = tensor("transpose_7")]; - tensor q = reshape(shape = var_1426, x = q_21)[name = tensor("q")]; - tensor var_1428 = const()[name = tensor("op_1428"), val = tensor([4, 4, 12, 64])]; - tensor k_21 = transpose(perm = k_21_perm_0, x = var_1416)[name = tensor("transpose_6")]; - tensor k = reshape(shape = var_1428, x = k_21)[name = tensor("k")]; - tensor var_1430 = const()[name = tensor("op_1430"), val = tensor([4, 4, 12, 64])]; - tensor v_21 = transpose(perm = v_21_perm_0, x = var_1423)[name = tensor("transpose_5")]; - tensor v = reshape(shape = var_1430, x = v_21)[name = tensor("v")]; + tensor var_1459 = const()[name = tensor("op_1459"), val = tensor([4, 4, 12, 64])]; + tensor q_21 = transpose(perm = q_21_perm_0, x = var_1442)[name = tensor("transpose_7")]; + tensor q = reshape(shape = var_1459, x = q_21)[name = tensor("q")]; + tensor var_1461 = const()[name = tensor("op_1461"), val = tensor([4, 4, 12, 64])]; + tensor k_21 = transpose(perm = k_21_perm_0, x = var_1449)[name = tensor("transpose_6")]; + tensor k = reshape(shape = var_1461, x = k_21)[name = tensor("k")]; + tensor var_1463 = const()[name = tensor("op_1463"), val = tensor([4, 4, 12, 64])]; + tensor v_21 = transpose(perm = v_21_perm_0, x = var_1456)[name = tensor("transpose_5")]; + tensor v = reshape(shape = var_1463, x = v_21)[name = tensor("v")]; tensor mul_3_y_0 = const()[name = tensor("mul_3_y_0"), val = tensor(0x1p-3)]; tensor mul_3 = mul(x = q, y = mul_3_y_0)[name = tensor("mul_3")]; tensor matmul_1_transpose_y_0 = const()[name = tensor("matmul_1_transpose_y_0"), val = tensor(true)]; @@ -1277,36 +1292,36 @@ program(1.0) tensor attn_output_9_transpose_x_0 = const()[name = tensor("attn_output_9_transpose_x_0"), val = tensor(false)]; tensor attn_output_9_transpose_y_0 = const()[name = tensor("attn_output_9_transpose_y_0"), val = tensor(false)]; tensor attn_output_9 = matmul(transpose_x = attn_output_9_transpose_x_0, transpose_y = attn_output_9_transpose_y_0, x = softmax_1, y = v)[name = tensor("attn_output_9")]; - tensor var_1433 = const()[name = tensor("op_1433"), val = tensor([2, 0, 1, 3])]; - tensor var_1438 = const()[name = tensor("op_1438"), val = tensor([48, 256])]; - tensor var_1434 = transpose(perm = var_1433, x = attn_output_9)[name = tensor("transpose_4")]; - tensor attn_output_11 = reshape(shape = var_1438, x = var_1434)[name = tensor("attn_output_11")]; - tensor attn_output_13 = linear(bias = decoder_self_attn2_1_out_proj_bias, weight = decoder_self_attn2_1_out_proj_weight, x = attn_output_11)[name = tensor("linear_53")]; - tensor var_1442 = const()[name = tensor("op_1442"), val = tensor([12, 4, 256])]; - tensor attn_output = reshape(shape = var_1442, x = attn_output_13)[name = tensor("attn_output")]; + tensor var_1466 = const()[name = tensor("op_1466"), val = tensor([2, 0, 1, 3])]; + tensor var_1471 = const()[name = tensor("op_1471"), val = tensor([48, 256])]; + tensor var_1467 = transpose(perm = var_1466, x = attn_output_9)[name = tensor("transpose_4")]; + tensor attn_output_11 = reshape(shape = var_1471, x = var_1467)[name = tensor("attn_output_11")]; + tensor attn_output_13 = linear(bias = inner_decoder_self_attn2_1_out_proj_bias, weight = inner_decoder_self_attn2_1_out_proj_weight, x = attn_output_11)[name = tensor("linear_53")]; + tensor var_1475 = const()[name = tensor("op_1475"), val = tensor([12, 4, 256])]; + tensor attn_output = reshape(shape = var_1475, x = attn_output_13)[name = tensor("attn_output")]; tensor sa2_out_perm_0 = const()[name = tensor("sa2_out_perm_0"), val = tensor([1, 0, 2])]; tensor sa2_out = transpose(perm = sa2_out_perm_0, x = attn_output)[name = tensor("transpose_3")]; - tensor input_193 = add(x = query_5, y = sa2_out)[name = tensor("input_193")]; - tensor input_195_axes_0 = const()[name = tensor("input_195_axes_0"), val = tensor([-1])]; - tensor input_195 = layer_norm(axes = input_195_axes_0, beta = decoder_norm21_1_bias, epsilon = var_1017, gamma = decoder_norm21_1_weight, x = input_193)[name = tensor("input_195")]; - tensor input_197 = linear(bias = decoder_linear1_1_bias, weight = decoder_linear1_1_weight, x = input_195)[name = tensor("linear_54")]; - tensor input_199 = relu(x = input_197)[name = tensor("input_199")]; - tensor ffn_out = linear(bias = decoder_linear2_1_bias, weight = decoder_linear2_1_weight, x = input_199)[name = tensor("linear_55")]; - tensor input_201 = add(x = input_195, y = ffn_out)[name = tensor("input_201")]; + tensor input_195 = add(x = query_5, y = sa2_out)[name = tensor("input_195")]; + tensor input_197_axes_0 = const()[name = tensor("input_197_axes_0"), val = tensor([-1])]; + tensor input_197 = layer_norm(axes = input_197_axes_0, beta = inner_decoder_norm21_1_bias, epsilon = var_66, gamma = inner_decoder_norm21_1_weight, x = input_195)[name = tensor("input_197")]; + tensor input_199 = linear(bias = inner_decoder_linear1_1_bias, weight = inner_decoder_linear1_1_weight, x = input_197)[name = tensor("linear_54")]; + tensor input_201 = relu(x = input_199)[name = tensor("input_201")]; + tensor ffn_out = linear(bias = inner_decoder_linear2_1_bias, weight = inner_decoder_linear2_1_weight, x = input_201)[name = tensor("linear_55")]; + tensor input_203 = add(x = input_197, y = ffn_out)[name = tensor("input_203")]; tensor xt_axes_0 = const()[name = tensor("xt_axes_0"), val = tensor([-1])]; - tensor xt = layer_norm(axes = xt_axes_0, beta = decoder_norm22_1_bias, epsilon = var_1017, gamma = decoder_norm22_1_weight, x = input_201)[name = tensor("xt")]; - tensor var_1462 = const()[name = tensor("op_1462"), val = tensor([1, 4, 12, 256])]; - tensor input = reshape(shape = var_1462, x = xt)[name = tensor("input")]; - tensor var_1464 = const()[name = tensor("op_1464"), val = tensor([-1])]; - tensor var_1465 = reduce_l2_norm(axes = var_1464, keep_dims = var_1020, x = input)[name = tensor("op_1465")]; + tensor xt = layer_norm(axes = xt_axes_0, beta = inner_decoder_norm22_1_bias, epsilon = var_66, gamma = inner_decoder_norm22_1_weight, x = input_203)[name = tensor("xt")]; + tensor var_1495 = const()[name = tensor("op_1495"), val = tensor([1, 4, 12, 256])]; + tensor input = reshape(shape = var_1495, x = xt)[name = tensor("input")]; + tensor var_1497 = const()[name = tensor("op_1497"), val = tensor([-1])]; + tensor var_1498 = reduce_l2_norm(axes = var_1497, keep_dims = var_65, x = input)[name = tensor("op_1498")]; tensor const_42 = const()[name = tensor("const_42"), val = tensor(0x1.fffffep+127)]; - tensor clip_5 = clip(alpha = var_1012, beta = const_42, x = var_1465)[name = tensor("clip_5")]; - tensor var_1467 = real_div(x = input, y = clip_5)[name = tensor("op_1467")]; + tensor clip_5 = clip(alpha = var_79, beta = const_42, x = var_1498)[name = tensor("clip_5")]; + tensor var_1500 = real_div(x = input, y = clip_5)[name = tensor("op_1500")]; tensor concat_6 = const()[name = tensor("concat_6"), val = tensor([4, 1, 256])]; tensor reshape_0 = reshape(shape = concat_6, x = emb)[name = tensor("reshape_0")]; tensor transpose_1_perm_0 = const()[name = tensor("transpose_1_perm_0"), val = tensor([0, 1, 3, 2])]; tensor concat_7 = const()[name = tensor("concat_7"), val = tensor([4, 256, 12])]; - tensor transpose_1 = transpose(perm = transpose_1_perm_0, x = var_1467)[name = tensor("transpose_2")]; + tensor transpose_1 = transpose(perm = transpose_1_perm_0, x = var_1500)[name = tensor("transpose_2")]; tensor reshape_1 = reshape(shape = concat_7, x = transpose_1)[name = tensor("reshape_1")]; tensor matmul_2_transpose_x_0 = const()[name = tensor("matmul_2_transpose_x_0"), val = tensor(false)]; tensor matmul_2_transpose_y_0 = const()[name = tensor("matmul_2_transpose_y_0"), val = tensor(false)]; @@ -1317,10 +1332,10 @@ program(1.0) tensor output_end_0 = const()[name = tensor("output_end_0"), val = tensor([1, 4, 11])]; tensor output_end_mask_0 = const()[name = tensor("output_end_mask_0"), val = tensor([true, true, false])]; tensor output = slice_by_index(begin = output_begin_0, end = output_end_0, end_mask = output_end_mask_0, x = reshape_2)[name = tensor("output")]; - tensor probs = sigmoid(x = output)[name = tensor("op_1471")]; - tensor var_1473_axis_0 = const()[name = tensor("op_1473_axis_0"), val = tensor(0)]; - tensor dec_kv_new = stack(axis = var_1473_axis_0, values = (var_1169, nkv))[name = tensor("op_1473")]; - tensor var_1475_axis_0 = const()[name = tensor("op_1475_axis_0"), val = tensor(0)]; - tensor dec_scale_new = stack(axis = var_1475_axis_0, values = (new_scale_9, new_scale))[name = tensor("op_1475")]; + tensor probs = sigmoid(x = output)[name = tensor("op_1504")]; + tensor var_1506_axis_0 = const()[name = tensor("op_1506_axis_0"), val = tensor(0)]; + tensor dec_kv_new = stack(axis = var_1506_axis_0, values = (var_1202, nkv))[name = tensor("op_1506")]; + tensor var_1508_axis_0 = const()[name = tensor("op_1508_axis_0"), val = tensor(0)]; + tensor dec_scale_new = stack(axis = var_1508_axis_0, values = (new_scale_9, new_scale))[name = tensor("op_1508")]; } -> (probs, enc_kv_new, enc_scale_new, enc_conv_cache_new, cnn_window_new, dec_kv_new, dec_scale_new); } \ No newline at end of file diff --git a/optimized/dih2/400ms/ls_eend_dih2_400ms.mlpackage/Data/com.apple.CoreML/model.mlmodel b/optimized/dih2/400ms/ls_eend_dih2_400ms.mlpackage/Data/com.apple.CoreML/model.mlmodel index 8c1ca861fb9558eb088d1aeb653c17845f7087b0..ce2691fc8e6729b80850a421f04c1ba265fe63e3 100644 --- a/optimized/dih2/400ms/ls_eend_dih2_400ms.mlpackage/Data/com.apple.CoreML/model.mlmodel +++ b/optimized/dih2/400ms/ls_eend_dih2_400ms.mlpackage/Data/com.apple.CoreML/model.mlmodel @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:512460ea937729458d31e1520a4866d8d4204db686d08c065f181bcfa9f13c79 -size 191051 +oid sha256:8eca0c1cf1c295154af76c01fc161e535e1e1cea55de94bee0645f33ba0761c8 +size 197123 diff --git a/optimized/dih2/400ms/ls_eend_dih2_400ms.mlpackage/Manifest.json b/optimized/dih2/400ms/ls_eend_dih2_400ms.mlpackage/Manifest.json index 1f2e5792778d89ff02951ec657bbd30348216ac5..7be2773e83f16979af8b07eeeb27521e182fb490 100644 --- a/optimized/dih2/400ms/ls_eend_dih2_400ms.mlpackage/Manifest.json +++ b/optimized/dih2/400ms/ls_eend_dih2_400ms.mlpackage/Manifest.json @@ -1,18 +1,18 @@ { "fileFormatVersion": "1.0.0", "itemInfoEntries": { - "8AAD4EFD-A1AE-4676-AFDF-014CF1A0D1F3": { - "author": "com.apple.CoreML", - "description": "CoreML Model Specification", - "name": "model.mlmodel", - "path": "com.apple.CoreML/model.mlmodel" - }, - "F3FC4090-5689-4260-9E54-5F399C74FF28": { + "817CC62A-ED39-420D-B527-1DBFEF63AF42": { "author": "com.apple.CoreML", "description": "CoreML Model Weights", "name": "weights", "path": "com.apple.CoreML/weights" + }, + "945611BB-429A-4DA9-B76F-AE8F78395321": { + "author": "com.apple.CoreML", + "description": "CoreML Model Specification", + "name": "model.mlmodel", + "path": "com.apple.CoreML/model.mlmodel" } }, - "rootModelIdentifier": "8AAD4EFD-A1AE-4676-AFDF-014CF1A0D1F3" + "rootModelIdentifier": "945611BB-429A-4DA9-B76F-AE8F78395321" } diff --git a/optimized/dih2/500ms/ls_eend_dih2_500ms.mlmodelc/analytics/coremldata.bin b/optimized/dih2/500ms/ls_eend_dih2_500ms.mlmodelc/analytics/coremldata.bin index 71b915acdd9abe5171728dadc38af3b7263e0247..f9423badbe72fa7ddbb043591f6219c56acdc2d5 100644 --- a/optimized/dih2/500ms/ls_eend_dih2_500ms.mlmodelc/analytics/coremldata.bin +++ b/optimized/dih2/500ms/ls_eend_dih2_500ms.mlmodelc/analytics/coremldata.bin @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:f3ab2696bac4217bb0bfc9f74776b079447b45424191c9523d5d87f9a001c761 +oid sha256:4c27ede5b72b63110ea75bc15260db0ccfac78e1baae1d7690710f20d3cccce0 size 243 diff --git a/optimized/dih2/500ms/ls_eend_dih2_500ms.mlmodelc/coremldata.bin b/optimized/dih2/500ms/ls_eend_dih2_500ms.mlmodelc/coremldata.bin index f1fe815e519cef7d8e571c8c3bb4fc3c51bf16bd..d89904087ef63365cfe5fa100c8bd5edc7c0c119 100644 --- a/optimized/dih2/500ms/ls_eend_dih2_500ms.mlmodelc/coremldata.bin +++ b/optimized/dih2/500ms/ls_eend_dih2_500ms.mlmodelc/coremldata.bin @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:f8aa47228de8673fd3089ab1dcf43fd425952e0eeaf3be2efeb7db0bdae626a6 -size 1308 +oid sha256:bee4de6a09887fa76315702a01b13f6d86ea41194abc0846c004f7052395b97d +size 1411 diff --git a/optimized/dih2/500ms/ls_eend_dih2_500ms.mlmodelc/metadata.json b/optimized/dih2/500ms/ls_eend_dih2_500ms.mlmodelc/metadata.json index 118719eec347c4d50bae0ae8b401d770b4bf1ac8..b2071cb30a9295712fc171bbed673c20ff18db8a 100644 --- a/optimized/dih2/500ms/ls_eend_dih2_500ms.mlmodelc/metadata.json +++ b/optimized/dih2/500ms/ls_eend_dih2_500ms.mlmodelc/metadata.json @@ -1,7 +1,7 @@ [ { "metadataOutputVersion" : "3.0", - "shortDescription" : "LS-EEND DIHARD II streaming diarizer (pipeline, T=5, max_speakers=10)", + "shortDescription" : "LS-EEND DIHARD II streaming diarizer (pipeline, T=5, max_speakers=10, layout=raw_mel, cu=all)", "outputSchema" : [ { "hasShapeFlexibility" : "0", @@ -81,7 +81,7 @@ "specificationVersion" : 8, "mlProgramOperationTypeHistogram" : { "Ios16.reduceL2Norm" : 2, - "Ios17.reshape" : 68, + "Ios17.reshape" : 69, "Ios16.softmax" : 2, "Ios17.matmul" : 29, "Ios17.transpose" : 57, @@ -89,7 +89,7 @@ "Ios17.expandDims" : 3, "Ios17.add" : 46, "Ios16.sigmoid" : 5, - "Ios17.sliceByIndex" : 72, + "Ios17.sliceByIndex" : 77, "Tile" : 1, "Ios16.reduceSum" : 1, "Ios17.squeeze" : 2, @@ -101,7 +101,7 @@ "Ios16.silu" : 18, "Ios17.realDiv" : 20, "Ios17.linear" : 56, - "Stack" : 5, + "Stack" : 6, "Ios17.concat" : 26, "Ios16.relu" : 2, "Ios16.cumsum" : 1, @@ -128,9 +128,9 @@ "hasShapeFlexibility" : "0", "isOptional" : "0", "dataType" : "Float32", - "formattedType" : "MultiArray (Float32 1 × 5 × 345)", + "formattedType" : "MultiArray (Float32 1 × 55 × 23)", "shortDescription" : "", - "shape" : "[1, 5, 345]", + "shape" : "[1, 55, 23]", "name" : "features", "type" : "MultiArray" }, @@ -206,8 +206,8 @@ } ], "userDefinedMetadata" : { - "com.github.apple.coremltools.conversion_date" : "2026-04-16", - "config" : "{\"model_name\": \"dih2\", \"model_label\": \"DIHARD II\", \"variant\": \"pipeline\", \"chunk_size\": 5, \"step_duration_ms\": 500, \"frame_duration_ms\": 100, \"frame_duration_seconds\": 0.1, \"max_speakers\": 10, \"max_nspks\": 12, \"n_units\": 256, \"n_heads\": 4, \"enc_n_layers\": 4, \"dec_n_layers\": 2, \"conv_kernel_size\": 16, \"conv_delay\": 9, \"sample_rate\": 8000, \"win_length\": 200, \"hop_length\": 80, \"n_mels\": 23, \"context_size\": 7, \"subsampling\": 10, \"feat_type\": \"logmel23_cummn\", \"pure_roll\": true}", + "com.github.apple.coremltools.conversion_date" : "2026-04-18", + "config" : "{\"model_name\": \"dih2\", \"model_label\": \"DIHARD II\", \"variant\": \"pipeline\", \"chunk_size\": 5, \"step_duration_ms\": 500, \"frame_duration_ms\": 100, \"frame_duration_seconds\": 0.1, \"max_speakers\": 10, \"max_nspks\": 12, \"n_units\": 256, \"n_heads\": 4, \"enc_n_layers\": 4, \"dec_n_layers\": 2, \"conv_kernel_size\": 16, \"conv_delay\": 9, \"sample_rate\": 8000, \"win_length\": 200, \"hop_length\": 80, \"n_mels\": 23, \"context_size\": 7, \"subsampling\": 10, \"feat_type\": \"logmel23_cummn\", \"pure_roll\": true, \"input_layout\": \"raw_mel\", \"compute_units_export\": \"all\", \"raw_mel_length\": 55}", "com.github.apple.coremltools.source" : "torch==2.6.0", "com.github.apple.coremltools.version" : "9.0", "com.github.apple.coremltools.source_dialect" : "TorchScript" diff --git a/optimized/dih2/500ms/ls_eend_dih2_500ms.mlmodelc/model.mil b/optimized/dih2/500ms/ls_eend_dih2_500ms.mlmodelc/model.mil index d06d8cb5807c4c0b54c3e07f38ba5655c925422a..ab59896474305668d95443aae5b48ae6a2df960c 100644 --- a/optimized/dih2/500ms/ls_eend_dih2_500ms.mlmodelc/model.mil +++ b/optimized/dih2/500ms/ls_eend_dih2_500ms.mlmodelc/model.mil @@ -1,234 +1,260 @@ program(1.0) [buildInfo = dict, tensor>({{"coremlc-component-MIL", "3510.2.1"}, {"coremlc-version", "3500.32.1"}, {"coremltools-component-torch", "2.6.0"}, {"coremltools-source-dialect", "TorchScript"}, {"coremltools-version", "9.0"}})] { - func main(tensor cnn_window, tensor dec_kv, tensor dec_scale, tensor enc_conv_cache, tensor enc_kv, tensor enc_scale, tensor features, tensor valid_mask) { - tensor encoder_cnn_bias = const()[name = tensor("encoder_cnn_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(64)))]; - tensor encoder_cnn_weight = const()[name = tensor("encoder_cnn_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1152)))]; - tensor encoder__causal_mask = const()[name = tensor("encoder__causal_mask"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4981952)))]; - tensor encoder__t_index = const()[name = tensor("encoder__t_index"), val = tensor([0x1p+0, 0x1p+1, 0x1.8p+1, 0x1p+2, 0x1.4p+2])]; - tensor encoder_input_projection_linear_bias = const()[name = tensor("encoder_input_projection_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4982144)))]; - tensor encoder_input_projection_linear_weight = const()[name = tensor("encoder_input_projection_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4983232)))]; - tensor encoder_pre_layer_norm_bias = const()[name = tensor("encoder_pre_layer_norm_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(5336576)))]; - tensor encoder_pre_layer_norm_weight = const()[name = tensor("encoder_pre_layer_norm_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(5337664)))]; - tensor encoder_ffn1_0_module_sequential_0_bias = const()[name = tensor("encoder_ffn1_0_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(5338752)))]; - tensor encoder_ffn1_0_module_sequential_0_weight = const()[name = tensor("encoder_ffn1_0_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(5339840)))]; - tensor encoder_ffn1_0_module_sequential_1_linear_bias = const()[name = tensor("encoder_ffn1_0_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(5340928)))]; - tensor encoder_ffn1_0_module_sequential_1_linear_weight = const()[name = tensor("encoder_ffn1_0_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(5345088)))]; - tensor encoder_ffn1_0_module_sequential_4_linear_bias = const()[name = tensor("encoder_ffn1_0_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(6393728)))]; - tensor encoder_ffn1_0_module_sequential_4_linear_weight = const()[name = tensor("encoder_ffn1_0_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(6394816)))]; - tensor encoder_ret_lns_0_bias = const()[name = tensor("encoder_ret_lns_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7443456)))]; - tensor encoder_ret_lns_0_weight = const()[name = tensor("encoder_ret_lns_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7444544)))]; - tensor encoder_q_proj_0_bias = const()[name = tensor("encoder_q_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7445632)))]; - tensor encoder_q_proj_0_weight = const()[name = tensor("encoder_q_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7446720)))]; - tensor encoder_k_proj_0_bias = const()[name = tensor("encoder_k_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7708928)))]; - tensor encoder_k_proj_0_weight = const()[name = tensor("encoder_k_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7710016)))]; - tensor encoder_v_proj_0_bias = const()[name = tensor("encoder_v_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7972224)))]; - tensor encoder_v_proj_0_weight = const()[name = tensor("encoder_v_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7973312)))]; - tensor encoder_g_proj_0_bias = const()[name = tensor("encoder_g_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8235520)))]; - tensor encoder_g_proj_0_weight = const()[name = tensor("encoder_g_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8236608)))]; - tensor encoder_out_proj_0_bias = const()[name = tensor("encoder_out_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8498816)))]; - tensor encoder_out_proj_0_weight = const()[name = tensor("encoder_out_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8499904)))]; - tensor encoder_conv_module_0_sequential_0_bias = const()[name = tensor("encoder_conv_module_0_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8762112)))]; - tensor encoder_conv_module_0_sequential_0_weight = const()[name = tensor("encoder_conv_module_0_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8763200)))]; - tensor encoder_conv_module_0_sequential_2_conv_bias = const()[name = tensor("encoder_conv_module_0_sequential_2_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8764288)))]; - tensor encoder_conv_module_0_sequential_2_conv_weight = const()[name = tensor("encoder_conv_module_0_sequential_2_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8766400)))]; - tensor encoder_conv_module_0_sequential_4_conv_weight = const()[name = tensor("encoder_conv_module_0_sequential_4_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9290752)))]; - tensor encoder_conv_module_0_sequential_5_running_var = const()[name = tensor("encoder_conv_module_0_sequential_5_running_var"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9307200)))]; - tensor encoder_conv_module_0_sequential_5_running_mean = const()[name = tensor("encoder_conv_module_0_sequential_5_running_mean"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9308288)))]; - tensor encoder_conv_module_0_sequential_5_bias = const()[name = tensor("encoder_conv_module_0_sequential_5_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9309376)))]; - tensor encoder_conv_module_0_sequential_5_weight = const()[name = tensor("encoder_conv_module_0_sequential_5_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9310464)))]; - tensor encoder_conv_module_0_sequential_7_conv_bias = const()[name = tensor("encoder_conv_module_0_sequential_7_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9311552)))]; - tensor encoder_conv_module_0_sequential_7_conv_weight = const()[name = tensor("encoder_conv_module_0_sequential_7_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9312640)))]; - tensor encoder_ffn2_0_module_sequential_0_bias = const()[name = tensor("encoder_ffn2_0_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9574848)))]; - tensor encoder_ffn2_0_module_sequential_0_weight = const()[name = tensor("encoder_ffn2_0_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9575936)))]; - tensor encoder_ffn2_0_module_sequential_1_linear_bias = const()[name = tensor("encoder_ffn2_0_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9577024)))]; - tensor encoder_ffn2_0_module_sequential_1_linear_weight = const()[name = tensor("encoder_ffn2_0_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9581184)))]; - tensor encoder_ffn2_0_module_sequential_4_linear_bias = const()[name = tensor("encoder_ffn2_0_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(10629824)))]; - tensor encoder_ffn2_0_module_sequential_4_linear_weight = const()[name = tensor("encoder_ffn2_0_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(10630912)))]; - tensor encoder_layer_norm_0_bias = const()[name = tensor("encoder_layer_norm_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11679552)))]; - tensor encoder_layer_norm_0_weight = const()[name = tensor("encoder_layer_norm_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11680640)))]; - tensor encoder_ffn1_1_module_sequential_0_bias = const()[name = tensor("encoder_ffn1_1_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11681728)))]; - tensor encoder_ffn1_1_module_sequential_0_weight = const()[name = tensor("encoder_ffn1_1_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11682816)))]; - tensor encoder_ffn1_1_module_sequential_1_linear_bias = const()[name = tensor("encoder_ffn1_1_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11683904)))]; - tensor encoder_ffn1_1_module_sequential_1_linear_weight = const()[name = tensor("encoder_ffn1_1_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11688064)))]; - tensor encoder_ffn1_1_module_sequential_4_linear_bias = const()[name = tensor("encoder_ffn1_1_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(12736704)))]; - tensor encoder_ffn1_1_module_sequential_4_linear_weight = const()[name = tensor("encoder_ffn1_1_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(12737792)))]; - tensor encoder_ret_lns_1_bias = const()[name = tensor("encoder_ret_lns_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(13786432)))]; - tensor encoder_ret_lns_1_weight = const()[name = tensor("encoder_ret_lns_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(13787520)))]; - tensor encoder_q_proj_1_bias = const()[name = tensor("encoder_q_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(13788608)))]; - tensor encoder_q_proj_1_weight = const()[name = tensor("encoder_q_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(13789696)))]; - tensor encoder_k_proj_1_bias = const()[name = tensor("encoder_k_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14051904)))]; - tensor encoder_k_proj_1_weight = const()[name = tensor("encoder_k_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14052992)))]; - tensor encoder_v_proj_1_bias = const()[name = tensor("encoder_v_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14315200)))]; - tensor encoder_v_proj_1_weight = const()[name = tensor("encoder_v_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14316288)))]; - tensor encoder_g_proj_1_bias = const()[name = tensor("encoder_g_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14578496)))]; - tensor encoder_g_proj_1_weight = const()[name = tensor("encoder_g_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14579584)))]; - tensor encoder_out_proj_1_bias = const()[name = tensor("encoder_out_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14841792)))]; - tensor encoder_out_proj_1_weight = const()[name = tensor("encoder_out_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14842880)))]; - tensor encoder_conv_module_1_sequential_0_bias = const()[name = tensor("encoder_conv_module_1_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15105088)))]; - tensor encoder_conv_module_1_sequential_0_weight = const()[name = tensor("encoder_conv_module_1_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15106176)))]; - tensor encoder_conv_module_1_sequential_2_conv_bias = const()[name = tensor("encoder_conv_module_1_sequential_2_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15107264)))]; - tensor encoder_conv_module_1_sequential_2_conv_weight = const()[name = tensor("encoder_conv_module_1_sequential_2_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15109376)))]; - tensor encoder_conv_module_1_sequential_4_conv_weight = const()[name = tensor("encoder_conv_module_1_sequential_4_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15633728)))]; - tensor encoder_conv_module_1_sequential_5_running_var = const()[name = tensor("encoder_conv_module_1_sequential_5_running_var"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15650176)))]; - tensor encoder_conv_module_1_sequential_5_running_mean = const()[name = tensor("encoder_conv_module_1_sequential_5_running_mean"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15651264)))]; - tensor encoder_conv_module_1_sequential_5_bias = const()[name = tensor("encoder_conv_module_1_sequential_5_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15652352)))]; - tensor encoder_conv_module_1_sequential_5_weight = const()[name = tensor("encoder_conv_module_1_sequential_5_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15653440)))]; - tensor encoder_conv_module_1_sequential_7_conv_bias = const()[name = tensor("encoder_conv_module_1_sequential_7_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15654528)))]; - tensor encoder_conv_module_1_sequential_7_conv_weight = const()[name = tensor("encoder_conv_module_1_sequential_7_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15655616)))]; - tensor encoder_ffn2_1_module_sequential_0_bias = const()[name = tensor("encoder_ffn2_1_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15917824)))]; - tensor encoder_ffn2_1_module_sequential_0_weight = const()[name = tensor("encoder_ffn2_1_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15918912)))]; - tensor encoder_ffn2_1_module_sequential_1_linear_bias = const()[name = tensor("encoder_ffn2_1_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15920000)))]; - tensor encoder_ffn2_1_module_sequential_1_linear_weight = const()[name = tensor("encoder_ffn2_1_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15924160)))]; - tensor encoder_ffn2_1_module_sequential_4_linear_bias = const()[name = tensor("encoder_ffn2_1_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(16972800)))]; - tensor encoder_ffn2_1_module_sequential_4_linear_weight = const()[name = tensor("encoder_ffn2_1_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(16973888)))]; - tensor encoder_layer_norm_1_bias = const()[name = tensor("encoder_layer_norm_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18022528)))]; - tensor encoder_layer_norm_1_weight = const()[name = tensor("encoder_layer_norm_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18023616)))]; - tensor encoder_ffn1_2_module_sequential_0_bias = const()[name = tensor("encoder_ffn1_2_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18024704)))]; - tensor encoder_ffn1_2_module_sequential_0_weight = const()[name = tensor("encoder_ffn1_2_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18025792)))]; - tensor encoder_ffn1_2_module_sequential_1_linear_bias = const()[name = tensor("encoder_ffn1_2_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18026880)))]; - tensor encoder_ffn1_2_module_sequential_1_linear_weight = const()[name = tensor("encoder_ffn1_2_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18031040)))]; - tensor encoder_ffn1_2_module_sequential_4_linear_bias = const()[name = tensor("encoder_ffn1_2_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(19079680)))]; - tensor encoder_ffn1_2_module_sequential_4_linear_weight = const()[name = tensor("encoder_ffn1_2_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(19080768)))]; - tensor encoder_ret_lns_2_bias = const()[name = tensor("encoder_ret_lns_2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20129408)))]; - tensor encoder_ret_lns_2_weight = const()[name = tensor("encoder_ret_lns_2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20130496)))]; - tensor encoder_q_proj_2_bias = const()[name = tensor("encoder_q_proj_2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20131584)))]; - tensor encoder_q_proj_2_weight = const()[name = tensor("encoder_q_proj_2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20132672)))]; - tensor encoder_k_proj_2_bias = const()[name = tensor("encoder_k_proj_2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20394880)))]; - tensor encoder_k_proj_2_weight = const()[name = tensor("encoder_k_proj_2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20395968)))]; - tensor encoder_v_proj_2_bias = const()[name = tensor("encoder_v_proj_2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20658176)))]; - tensor encoder_v_proj_2_weight = const()[name = tensor("encoder_v_proj_2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20659264)))]; - tensor encoder_g_proj_2_bias = const()[name = tensor("encoder_g_proj_2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20921472)))]; - tensor encoder_g_proj_2_weight = const()[name = tensor("encoder_g_proj_2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20922560)))]; - tensor encoder_out_proj_2_bias = const()[name = tensor("encoder_out_proj_2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21184768)))]; - tensor encoder_out_proj_2_weight = const()[name = tensor("encoder_out_proj_2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21185856)))]; - tensor encoder_conv_module_2_sequential_0_bias = const()[name = tensor("encoder_conv_module_2_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21448064)))]; - tensor encoder_conv_module_2_sequential_0_weight = const()[name = tensor("encoder_conv_module_2_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21449152)))]; - tensor encoder_conv_module_2_sequential_2_conv_bias = const()[name = tensor("encoder_conv_module_2_sequential_2_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21450240)))]; - tensor encoder_conv_module_2_sequential_2_conv_weight = const()[name = tensor("encoder_conv_module_2_sequential_2_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21452352)))]; - tensor encoder_conv_module_2_sequential_4_conv_weight = const()[name = tensor("encoder_conv_module_2_sequential_4_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21976704)))]; - tensor encoder_conv_module_2_sequential_5_running_var = const()[name = tensor("encoder_conv_module_2_sequential_5_running_var"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21993152)))]; - tensor encoder_conv_module_2_sequential_5_running_mean = const()[name = tensor("encoder_conv_module_2_sequential_5_running_mean"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21994240)))]; - tensor encoder_conv_module_2_sequential_5_bias = const()[name = tensor("encoder_conv_module_2_sequential_5_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21995328)))]; - tensor encoder_conv_module_2_sequential_5_weight = const()[name = tensor("encoder_conv_module_2_sequential_5_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21996416)))]; - tensor encoder_conv_module_2_sequential_7_conv_bias = const()[name = tensor("encoder_conv_module_2_sequential_7_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21997504)))]; - tensor encoder_conv_module_2_sequential_7_conv_weight = const()[name = tensor("encoder_conv_module_2_sequential_7_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21998592)))]; - tensor encoder_ffn2_2_module_sequential_0_bias = const()[name = tensor("encoder_ffn2_2_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(22260800)))]; - tensor encoder_ffn2_2_module_sequential_0_weight = const()[name = tensor("encoder_ffn2_2_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(22261888)))]; - tensor encoder_ffn2_2_module_sequential_1_linear_bias = const()[name = tensor("encoder_ffn2_2_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(22262976)))]; - tensor encoder_ffn2_2_module_sequential_1_linear_weight = const()[name = tensor("encoder_ffn2_2_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(22267136)))]; - tensor encoder_ffn2_2_module_sequential_4_linear_bias = const()[name = tensor("encoder_ffn2_2_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(23315776)))]; - tensor encoder_ffn2_2_module_sequential_4_linear_weight = const()[name = tensor("encoder_ffn2_2_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(23316864)))]; - tensor encoder_layer_norm_2_bias = const()[name = tensor("encoder_layer_norm_2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(24365504)))]; - tensor encoder_layer_norm_2_weight = const()[name = tensor("encoder_layer_norm_2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(24366592)))]; - tensor encoder_ffn1_3_module_sequential_0_bias = const()[name = tensor("encoder_ffn1_3_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(24367680)))]; - tensor encoder_ffn1_3_module_sequential_0_weight = const()[name = tensor("encoder_ffn1_3_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(24368768)))]; - tensor encoder_ffn1_3_module_sequential_1_linear_bias = const()[name = tensor("encoder_ffn1_3_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(24369856)))]; - tensor encoder_ffn1_3_module_sequential_1_linear_weight = const()[name = tensor("encoder_ffn1_3_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(24374016)))]; - tensor encoder_ffn1_3_module_sequential_4_linear_bias = const()[name = tensor("encoder_ffn1_3_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(25422656)))]; - tensor encoder_ffn1_3_module_sequential_4_linear_weight = const()[name = tensor("encoder_ffn1_3_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(25423744)))]; - tensor encoder_ret_lns_3_bias = const()[name = tensor("encoder_ret_lns_3_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(26472384)))]; - tensor encoder_ret_lns_3_weight = const()[name = tensor("encoder_ret_lns_3_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(26473472)))]; - tensor encoder_q_proj_3_bias = const()[name = tensor("encoder_q_proj_3_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(26474560)))]; - tensor encoder_q_proj_3_weight = const()[name = tensor("encoder_q_proj_3_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(26475648)))]; - tensor encoder_k_proj_3_bias = const()[name = tensor("encoder_k_proj_3_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(26737856)))]; - tensor encoder_k_proj_3_weight = const()[name = tensor("encoder_k_proj_3_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(26738944)))]; - tensor encoder_v_proj_3_bias = const()[name = tensor("encoder_v_proj_3_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27001152)))]; - tensor encoder_v_proj_3_weight = const()[name = tensor("encoder_v_proj_3_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27002240)))]; - tensor encoder_g_proj_3_bias = const()[name = tensor("encoder_g_proj_3_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27264448)))]; - tensor encoder_g_proj_3_weight = const()[name = tensor("encoder_g_proj_3_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27265536)))]; - tensor encoder_out_proj_3_bias = const()[name = tensor("encoder_out_proj_3_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27527744)))]; - tensor encoder_out_proj_3_weight = const()[name = tensor("encoder_out_proj_3_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27528832)))]; - tensor encoder_conv_module_3_sequential_0_bias = const()[name = tensor("encoder_conv_module_3_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27791040)))]; - tensor encoder_conv_module_3_sequential_0_weight = const()[name = tensor("encoder_conv_module_3_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27792128)))]; - tensor encoder_conv_module_3_sequential_2_conv_bias = const()[name = tensor("encoder_conv_module_3_sequential_2_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27793216)))]; - tensor encoder_conv_module_3_sequential_2_conv_weight = const()[name = tensor("encoder_conv_module_3_sequential_2_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27795328)))]; - tensor encoder_conv_module_3_sequential_4_conv_weight = const()[name = tensor("encoder_conv_module_3_sequential_4_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28319680)))]; - tensor encoder_conv_module_3_sequential_5_running_var = const()[name = tensor("encoder_conv_module_3_sequential_5_running_var"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28336128)))]; - tensor encoder_conv_module_3_sequential_5_running_mean = const()[name = tensor("encoder_conv_module_3_sequential_5_running_mean"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28337216)))]; - tensor encoder_conv_module_3_sequential_5_bias = const()[name = tensor("encoder_conv_module_3_sequential_5_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28338304)))]; - tensor encoder_conv_module_3_sequential_5_weight = const()[name = tensor("encoder_conv_module_3_sequential_5_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28339392)))]; - tensor encoder_conv_module_3_sequential_7_conv_bias = const()[name = tensor("encoder_conv_module_3_sequential_7_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28340480)))]; - tensor encoder_conv_module_3_sequential_7_conv_weight = const()[name = tensor("encoder_conv_module_3_sequential_7_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28341568)))]; - tensor encoder_ffn2_3_module_sequential_0_bias = const()[name = tensor("encoder_ffn2_3_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28603776)))]; - tensor encoder_ffn2_3_module_sequential_0_weight = const()[name = tensor("encoder_ffn2_3_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28604864)))]; - tensor encoder_ffn2_3_module_sequential_1_linear_bias = const()[name = tensor("encoder_ffn2_3_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28605952)))]; - tensor encoder_ffn2_3_module_sequential_1_linear_weight = const()[name = tensor("encoder_ffn2_3_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28610112)))]; - tensor encoder_ffn2_3_module_sequential_4_linear_bias = const()[name = tensor("encoder_ffn2_3_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(29658752)))]; - tensor encoder_ffn2_3_module_sequential_4_linear_weight = const()[name = tensor("encoder_ffn2_3_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(29659840)))]; - tensor encoder_layer_norm_3_bias = const()[name = tensor("encoder_layer_norm_3_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(30708480)))]; - tensor encoder_layer_norm_3_weight = const()[name = tensor("encoder_layer_norm_3_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(30709568)))]; - tensor decoder__causal_mask = const()[name = tensor("decoder__causal_mask"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(30710656)))]; - tensor decoder_convert_bias = const()[name = tensor("decoder_convert_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(30710848)))]; - tensor decoder_convert_weight = const()[name = tensor("decoder_convert_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(30711936)))]; - tensor decoder_q_proj_0_bias = const()[name = tensor("decoder_q_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(31236288)))]; - tensor decoder_q_proj_0_weight = const()[name = tensor("decoder_q_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(31237376)))]; - tensor decoder_k_proj_0_bias = const()[name = tensor("decoder_k_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(31499584)))]; - tensor decoder_k_proj_0_weight = const()[name = tensor("decoder_k_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(31500672)))]; - tensor decoder_v_proj_0_bias = const()[name = tensor("decoder_v_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(31762880)))]; - tensor decoder_v_proj_0_weight = const()[name = tensor("decoder_v_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(31763968)))]; - tensor decoder_g_proj_0_bias = const()[name = tensor("decoder_g_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32026176)))]; - tensor decoder_g_proj_0_weight = const()[name = tensor("decoder_g_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32027264)))]; - tensor decoder_out_proj_0_bias = const()[name = tensor("decoder_out_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32289472)))]; - tensor decoder_out_proj_0_weight = const()[name = tensor("decoder_out_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32290560)))]; - tensor decoder_norm11_0_bias = const()[name = tensor("decoder_norm11_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32552768)))]; - tensor decoder_norm11_0_weight = const()[name = tensor("decoder_norm11_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32553856)))]; - tensor decoder_self_attn2_0_out_proj_bias = const()[name = tensor("decoder_self_attn2_0_out_proj_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32554944)))]; - tensor decoder_self_attn2_0_out_proj_weight = const()[name = tensor("decoder_self_attn2_0_out_proj_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32556032)))]; - tensor decoder_self_attn2_0_in_proj_bias = const()[name = tensor("decoder_self_attn2_0_in_proj_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32818240)))]; - tensor decoder_self_attn2_0_in_proj_weight = const()[name = tensor("decoder_self_attn2_0_in_proj_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32821376)))]; - tensor decoder_norm21_0_bias = const()[name = tensor("decoder_norm21_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(33607872)))]; - tensor decoder_norm21_0_weight = const()[name = tensor("decoder_norm21_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(33608960)))]; - tensor decoder_linear1_0_bias = const()[name = tensor("decoder_linear1_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(33610048)))]; - tensor decoder_linear1_0_weight = const()[name = tensor("decoder_linear1_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(33618304)))]; - tensor decoder_linear2_0_bias = const()[name = tensor("decoder_linear2_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(35715520)))]; - tensor decoder_linear2_0_weight = const()[name = tensor("decoder_linear2_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(35716608)))]; - tensor decoder_norm22_0_bias = const()[name = tensor("decoder_norm22_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(37813824)))]; - tensor decoder_norm22_0_weight = const()[name = tensor("decoder_norm22_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(37814912)))]; - tensor decoder_q_proj_1_bias = const()[name = tensor("decoder_q_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(37816000)))]; - tensor decoder_q_proj_1_weight = const()[name = tensor("decoder_q_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(37817088)))]; - tensor decoder_k_proj_1_bias = const()[name = tensor("decoder_k_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38079296)))]; - tensor decoder_k_proj_1_weight = const()[name = tensor("decoder_k_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38080384)))]; - tensor decoder_v_proj_1_bias = const()[name = tensor("decoder_v_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38342592)))]; - tensor decoder_v_proj_1_weight = const()[name = tensor("decoder_v_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38343680)))]; - tensor decoder_g_proj_1_bias = const()[name = tensor("decoder_g_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38605888)))]; - tensor decoder_g_proj_1_weight = const()[name = tensor("decoder_g_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38606976)))]; - tensor decoder_out_proj_1_bias = const()[name = tensor("decoder_out_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38869184)))]; - tensor decoder_out_proj_1_weight = const()[name = tensor("decoder_out_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38870272)))]; - tensor decoder_norm11_1_bias = const()[name = tensor("decoder_norm11_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(39132480)))]; - tensor decoder_norm11_1_weight = const()[name = tensor("decoder_norm11_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(39133568)))]; - tensor decoder_self_attn2_1_out_proj_bias = const()[name = tensor("decoder_self_attn2_1_out_proj_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(39134656)))]; - tensor decoder_self_attn2_1_out_proj_weight = const()[name = tensor("decoder_self_attn2_1_out_proj_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(39135744)))]; - tensor decoder_self_attn2_1_in_proj_bias = const()[name = tensor("decoder_self_attn2_1_in_proj_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(39397952)))]; - tensor decoder_self_attn2_1_in_proj_weight = const()[name = tensor("decoder_self_attn2_1_in_proj_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(39401088)))]; - tensor decoder_norm21_1_bias = const()[name = tensor("decoder_norm21_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(40187584)))]; - tensor decoder_norm21_1_weight = const()[name = tensor("decoder_norm21_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(40188672)))]; - tensor decoder_linear1_1_bias = const()[name = tensor("decoder_linear1_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(40189760)))]; - tensor decoder_linear1_1_weight = const()[name = tensor("decoder_linear1_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(40198016)))]; - tensor decoder_linear2_1_bias = const()[name = tensor("decoder_linear2_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(42295232)))]; - tensor decoder_linear2_1_weight = const()[name = tensor("decoder_linear2_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(42296320)))]; - tensor decoder_norm22_1_bias = const()[name = tensor("decoder_norm22_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(44393536)))]; - tensor decoder_norm22_1_weight = const()[name = tensor("decoder_norm22_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(44394624)))]; - tensor var_11 = const()[name = tensor("op_11"), val = tensor(0x1.5798eep-27)]; - tensor var_18 = const()[name = tensor("op_18"), val = tensor(0x1.0c6f7ap-20)]; - tensor var_21 = const()[name = tensor("op_21"), val = tensor(2)]; - tensor var_24 = const()[name = tensor("op_24"), val = tensor(0)]; - tensor var_27 = const()[name = tensor("op_27"), val = tensor(1)]; - tensor var_29 = const()[name = tensor("op_29"), val = tensor(0x1.4f8b58p-17)]; - tensor var_30 = const()[name = tensor("op_30"), val = tensor(true)]; - tensor input_1 = linear(bias = encoder_input_projection_linear_bias, weight = encoder_input_projection_linear_weight, x = features)[name = tensor("linear_0")]; - tensor input_3_axes_0 = const()[name = tensor("input_3_axes_0"), val = tensor([-1])]; - tensor input_3 = layer_norm(axes = input_3_axes_0, beta = encoder_pre_layer_norm_bias, epsilon = var_29, gamma = encoder_pre_layer_norm_weight, x = input_1)[name = tensor("input_3")]; + func main(tensor cnn_window, tensor dec_kv, tensor dec_scale, tensor enc_conv_cache, tensor enc_kv, tensor enc_scale, tensor features, tensor valid_mask) { + tensor inner_encoder_cnn_bias = const()[name = tensor("inner_encoder_cnn_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(64)))]; + tensor inner_encoder_cnn_weight = const()[name = tensor("inner_encoder_cnn_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1152)))]; + tensor inner_encoder__causal_mask = const()[name = tensor("inner_encoder__causal_mask"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4981952)))]; + tensor inner_encoder__t_index = const()[name = tensor("inner_encoder__t_index"), val = tensor([0x1p+0, 0x1p+1, 0x1.8p+1, 0x1p+2, 0x1.4p+2])]; + tensor inner_encoder_input_projection_linear_bias = const()[name = tensor("inner_encoder_input_projection_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4982144)))]; + tensor inner_encoder_input_projection_linear_weight = const()[name = tensor("inner_encoder_input_projection_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4983232)))]; + tensor inner_encoder_pre_layer_norm_bias = const()[name = tensor("inner_encoder_pre_layer_norm_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(5336576)))]; + tensor inner_encoder_pre_layer_norm_weight = const()[name = tensor("inner_encoder_pre_layer_norm_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(5337664)))]; + tensor inner_encoder_ffn1_0_module_sequential_0_bias = const()[name = tensor("inner_encoder_ffn1_0_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(5338752)))]; + tensor inner_encoder_ffn1_0_module_sequential_0_weight = const()[name = tensor("inner_encoder_ffn1_0_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(5339840)))]; + tensor inner_encoder_ffn1_0_module_sequential_1_linear_bias = const()[name = tensor("inner_encoder_ffn1_0_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(5340928)))]; + tensor inner_encoder_ffn1_0_module_sequential_1_linear_weight = const()[name = tensor("inner_encoder_ffn1_0_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(5345088)))]; + tensor inner_encoder_ffn1_0_module_sequential_4_linear_bias = const()[name = tensor("inner_encoder_ffn1_0_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(6393728)))]; + tensor inner_encoder_ffn1_0_module_sequential_4_linear_weight = const()[name = tensor("inner_encoder_ffn1_0_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(6394816)))]; + tensor inner_encoder_ret_lns_0_bias = const()[name = tensor("inner_encoder_ret_lns_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7443456)))]; + tensor inner_encoder_ret_lns_0_weight = const()[name = tensor("inner_encoder_ret_lns_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7444544)))]; + tensor inner_encoder_q_proj_0_bias = const()[name = tensor("inner_encoder_q_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7445632)))]; + tensor inner_encoder_q_proj_0_weight = const()[name = tensor("inner_encoder_q_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7446720)))]; + tensor inner_encoder_k_proj_0_bias = const()[name = tensor("inner_encoder_k_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7708928)))]; + tensor inner_encoder_k_proj_0_weight = const()[name = tensor("inner_encoder_k_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7710016)))]; + tensor inner_encoder_v_proj_0_bias = const()[name = tensor("inner_encoder_v_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7972224)))]; + tensor inner_encoder_v_proj_0_weight = const()[name = tensor("inner_encoder_v_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7973312)))]; + tensor inner_encoder_g_proj_0_bias = const()[name = tensor("inner_encoder_g_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8235520)))]; + tensor inner_encoder_g_proj_0_weight = const()[name = tensor("inner_encoder_g_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8236608)))]; + tensor inner_encoder_out_proj_0_bias = const()[name = tensor("inner_encoder_out_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8498816)))]; + tensor inner_encoder_out_proj_0_weight = const()[name = tensor("inner_encoder_out_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8499904)))]; + tensor inner_encoder_conv_module_0_sequential_0_bias = const()[name = tensor("inner_encoder_conv_module_0_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8762112)))]; + tensor inner_encoder_conv_module_0_sequential_0_weight = const()[name = tensor("inner_encoder_conv_module_0_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8763200)))]; + tensor inner_encoder_conv_module_0_sequential_2_conv_bias = const()[name = tensor("inner_encoder_conv_module_0_sequential_2_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8764288)))]; + tensor inner_encoder_conv_module_0_sequential_2_conv_weight = const()[name = tensor("inner_encoder_conv_module_0_sequential_2_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8766400)))]; + tensor inner_encoder_conv_module_0_sequential_4_conv_weight = const()[name = tensor("inner_encoder_conv_module_0_sequential_4_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9290752)))]; + tensor inner_encoder_conv_module_0_sequential_5_running_var = const()[name = tensor("inner_encoder_conv_module_0_sequential_5_running_var"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9307200)))]; + tensor inner_encoder_conv_module_0_sequential_5_running_mean = const()[name = tensor("inner_encoder_conv_module_0_sequential_5_running_mean"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9308288)))]; + tensor inner_encoder_conv_module_0_sequential_5_bias = const()[name = tensor("inner_encoder_conv_module_0_sequential_5_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9309376)))]; + tensor inner_encoder_conv_module_0_sequential_5_weight = const()[name = tensor("inner_encoder_conv_module_0_sequential_5_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9310464)))]; + tensor inner_encoder_conv_module_0_sequential_7_conv_bias = const()[name = tensor("inner_encoder_conv_module_0_sequential_7_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9311552)))]; + tensor inner_encoder_conv_module_0_sequential_7_conv_weight = const()[name = tensor("inner_encoder_conv_module_0_sequential_7_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9312640)))]; + tensor inner_encoder_ffn2_0_module_sequential_0_bias = const()[name = tensor("inner_encoder_ffn2_0_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9574848)))]; + tensor inner_encoder_ffn2_0_module_sequential_0_weight = const()[name = tensor("inner_encoder_ffn2_0_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9575936)))]; + tensor inner_encoder_ffn2_0_module_sequential_1_linear_bias = const()[name = tensor("inner_encoder_ffn2_0_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9577024)))]; + tensor inner_encoder_ffn2_0_module_sequential_1_linear_weight = const()[name = tensor("inner_encoder_ffn2_0_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9581184)))]; + tensor inner_encoder_ffn2_0_module_sequential_4_linear_bias = const()[name = tensor("inner_encoder_ffn2_0_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(10629824)))]; + tensor inner_encoder_ffn2_0_module_sequential_4_linear_weight = const()[name = tensor("inner_encoder_ffn2_0_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(10630912)))]; + tensor inner_encoder_layer_norm_0_bias = const()[name = tensor("inner_encoder_layer_norm_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11679552)))]; + tensor inner_encoder_layer_norm_0_weight = const()[name = tensor("inner_encoder_layer_norm_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11680640)))]; + tensor inner_encoder_ffn1_1_module_sequential_0_bias = const()[name = tensor("inner_encoder_ffn1_1_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11681728)))]; + tensor inner_encoder_ffn1_1_module_sequential_0_weight = const()[name = tensor("inner_encoder_ffn1_1_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11682816)))]; + tensor inner_encoder_ffn1_1_module_sequential_1_linear_bias = const()[name = tensor("inner_encoder_ffn1_1_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11683904)))]; + tensor inner_encoder_ffn1_1_module_sequential_1_linear_weight = const()[name = tensor("inner_encoder_ffn1_1_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11688064)))]; + tensor inner_encoder_ffn1_1_module_sequential_4_linear_bias = const()[name = tensor("inner_encoder_ffn1_1_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(12736704)))]; + tensor inner_encoder_ffn1_1_module_sequential_4_linear_weight = const()[name = tensor("inner_encoder_ffn1_1_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(12737792)))]; + tensor inner_encoder_ret_lns_1_bias = const()[name = tensor("inner_encoder_ret_lns_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(13786432)))]; + tensor inner_encoder_ret_lns_1_weight = const()[name = tensor("inner_encoder_ret_lns_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(13787520)))]; + tensor inner_encoder_q_proj_1_bias = const()[name = tensor("inner_encoder_q_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(13788608)))]; + tensor inner_encoder_q_proj_1_weight = const()[name = tensor("inner_encoder_q_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(13789696)))]; + tensor inner_encoder_k_proj_1_bias = const()[name = tensor("inner_encoder_k_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14051904)))]; + tensor inner_encoder_k_proj_1_weight = const()[name = tensor("inner_encoder_k_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14052992)))]; + tensor inner_encoder_v_proj_1_bias = const()[name = tensor("inner_encoder_v_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14315200)))]; + tensor inner_encoder_v_proj_1_weight = const()[name = tensor("inner_encoder_v_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14316288)))]; + tensor inner_encoder_g_proj_1_bias = const()[name = tensor("inner_encoder_g_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14578496)))]; + tensor inner_encoder_g_proj_1_weight = const()[name = tensor("inner_encoder_g_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14579584)))]; + tensor inner_encoder_out_proj_1_bias = const()[name = tensor("inner_encoder_out_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14841792)))]; + tensor inner_encoder_out_proj_1_weight = const()[name = tensor("inner_encoder_out_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14842880)))]; + tensor inner_encoder_conv_module_1_sequential_0_bias = const()[name = tensor("inner_encoder_conv_module_1_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15105088)))]; + tensor inner_encoder_conv_module_1_sequential_0_weight = const()[name = tensor("inner_encoder_conv_module_1_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15106176)))]; + tensor inner_encoder_conv_module_1_sequential_2_conv_bias = const()[name = tensor("inner_encoder_conv_module_1_sequential_2_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15107264)))]; + tensor inner_encoder_conv_module_1_sequential_2_conv_weight = const()[name = tensor("inner_encoder_conv_module_1_sequential_2_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15109376)))]; + tensor inner_encoder_conv_module_1_sequential_4_conv_weight = const()[name = tensor("inner_encoder_conv_module_1_sequential_4_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15633728)))]; + tensor inner_encoder_conv_module_1_sequential_5_running_var = const()[name = tensor("inner_encoder_conv_module_1_sequential_5_running_var"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15650176)))]; + tensor inner_encoder_conv_module_1_sequential_5_running_mean = const()[name = tensor("inner_encoder_conv_module_1_sequential_5_running_mean"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15651264)))]; + tensor inner_encoder_conv_module_1_sequential_5_bias = const()[name = tensor("inner_encoder_conv_module_1_sequential_5_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15652352)))]; + tensor inner_encoder_conv_module_1_sequential_5_weight = const()[name = tensor("inner_encoder_conv_module_1_sequential_5_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15653440)))]; + tensor inner_encoder_conv_module_1_sequential_7_conv_bias = const()[name = tensor("inner_encoder_conv_module_1_sequential_7_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15654528)))]; + tensor inner_encoder_conv_module_1_sequential_7_conv_weight = const()[name = tensor("inner_encoder_conv_module_1_sequential_7_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15655616)))]; + tensor inner_encoder_ffn2_1_module_sequential_0_bias = const()[name = tensor("inner_encoder_ffn2_1_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15917824)))]; + tensor inner_encoder_ffn2_1_module_sequential_0_weight = const()[name = tensor("inner_encoder_ffn2_1_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15918912)))]; + tensor inner_encoder_ffn2_1_module_sequential_1_linear_bias = const()[name = tensor("inner_encoder_ffn2_1_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15920000)))]; + tensor inner_encoder_ffn2_1_module_sequential_1_linear_weight = const()[name = tensor("inner_encoder_ffn2_1_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15924160)))]; + tensor inner_encoder_ffn2_1_module_sequential_4_linear_bias = const()[name = tensor("inner_encoder_ffn2_1_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(16972800)))]; + tensor inner_encoder_ffn2_1_module_sequential_4_linear_weight = const()[name = tensor("inner_encoder_ffn2_1_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(16973888)))]; + tensor inner_encoder_layer_norm_1_bias = const()[name = tensor("inner_encoder_layer_norm_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18022528)))]; + tensor inner_encoder_layer_norm_1_weight = const()[name = tensor("inner_encoder_layer_norm_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18023616)))]; + tensor inner_encoder_ffn1_2_module_sequential_0_bias = const()[name = tensor("inner_encoder_ffn1_2_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18024704)))]; + tensor inner_encoder_ffn1_2_module_sequential_0_weight = const()[name = tensor("inner_encoder_ffn1_2_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18025792)))]; + tensor inner_encoder_ffn1_2_module_sequential_1_linear_bias = const()[name = tensor("inner_encoder_ffn1_2_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18026880)))]; + tensor inner_encoder_ffn1_2_module_sequential_1_linear_weight = const()[name = tensor("inner_encoder_ffn1_2_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18031040)))]; + tensor inner_encoder_ffn1_2_module_sequential_4_linear_bias = const()[name = tensor("inner_encoder_ffn1_2_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(19079680)))]; + tensor inner_encoder_ffn1_2_module_sequential_4_linear_weight = const()[name = tensor("inner_encoder_ffn1_2_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(19080768)))]; + tensor inner_encoder_ret_lns_2_bias = const()[name = tensor("inner_encoder_ret_lns_2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20129408)))]; + tensor inner_encoder_ret_lns_2_weight = const()[name = tensor("inner_encoder_ret_lns_2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20130496)))]; + tensor inner_encoder_q_proj_2_bias = const()[name = tensor("inner_encoder_q_proj_2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20131584)))]; + tensor inner_encoder_q_proj_2_weight = const()[name = tensor("inner_encoder_q_proj_2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20132672)))]; + tensor inner_encoder_k_proj_2_bias = const()[name = tensor("inner_encoder_k_proj_2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20394880)))]; + tensor inner_encoder_k_proj_2_weight = const()[name = tensor("inner_encoder_k_proj_2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20395968)))]; + tensor inner_encoder_v_proj_2_bias = const()[name = tensor("inner_encoder_v_proj_2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20658176)))]; + tensor inner_encoder_v_proj_2_weight = const()[name = tensor("inner_encoder_v_proj_2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20659264)))]; + tensor inner_encoder_g_proj_2_bias = const()[name = tensor("inner_encoder_g_proj_2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20921472)))]; + tensor inner_encoder_g_proj_2_weight = const()[name = tensor("inner_encoder_g_proj_2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20922560)))]; + tensor inner_encoder_out_proj_2_bias = const()[name = tensor("inner_encoder_out_proj_2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21184768)))]; + tensor inner_encoder_out_proj_2_weight = const()[name = tensor("inner_encoder_out_proj_2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21185856)))]; + tensor inner_encoder_conv_module_2_sequential_0_bias = const()[name = tensor("inner_encoder_conv_module_2_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21448064)))]; + tensor inner_encoder_conv_module_2_sequential_0_weight = const()[name = tensor("inner_encoder_conv_module_2_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21449152)))]; + tensor inner_encoder_conv_module_2_sequential_2_conv_bias = const()[name = tensor("inner_encoder_conv_module_2_sequential_2_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21450240)))]; + tensor inner_encoder_conv_module_2_sequential_2_conv_weight = const()[name = tensor("inner_encoder_conv_module_2_sequential_2_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21452352)))]; + tensor inner_encoder_conv_module_2_sequential_4_conv_weight = const()[name = tensor("inner_encoder_conv_module_2_sequential_4_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21976704)))]; + tensor inner_encoder_conv_module_2_sequential_5_running_var = const()[name = tensor("inner_encoder_conv_module_2_sequential_5_running_var"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21993152)))]; + tensor inner_encoder_conv_module_2_sequential_5_running_mean = const()[name = tensor("inner_encoder_conv_module_2_sequential_5_running_mean"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21994240)))]; + tensor inner_encoder_conv_module_2_sequential_5_bias = const()[name = tensor("inner_encoder_conv_module_2_sequential_5_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21995328)))]; + tensor inner_encoder_conv_module_2_sequential_5_weight = const()[name = tensor("inner_encoder_conv_module_2_sequential_5_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21996416)))]; + tensor inner_encoder_conv_module_2_sequential_7_conv_bias = const()[name = tensor("inner_encoder_conv_module_2_sequential_7_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21997504)))]; + tensor inner_encoder_conv_module_2_sequential_7_conv_weight = const()[name = tensor("inner_encoder_conv_module_2_sequential_7_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21998592)))]; + tensor inner_encoder_ffn2_2_module_sequential_0_bias = const()[name = tensor("inner_encoder_ffn2_2_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(22260800)))]; + tensor inner_encoder_ffn2_2_module_sequential_0_weight = const()[name = tensor("inner_encoder_ffn2_2_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(22261888)))]; + tensor inner_encoder_ffn2_2_module_sequential_1_linear_bias = const()[name = tensor("inner_encoder_ffn2_2_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(22262976)))]; + tensor inner_encoder_ffn2_2_module_sequential_1_linear_weight = const()[name = tensor("inner_encoder_ffn2_2_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(22267136)))]; + tensor inner_encoder_ffn2_2_module_sequential_4_linear_bias = const()[name = tensor("inner_encoder_ffn2_2_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(23315776)))]; + tensor inner_encoder_ffn2_2_module_sequential_4_linear_weight = const()[name = tensor("inner_encoder_ffn2_2_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(23316864)))]; + tensor inner_encoder_layer_norm_2_bias = const()[name = tensor("inner_encoder_layer_norm_2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(24365504)))]; + tensor inner_encoder_layer_norm_2_weight = const()[name = tensor("inner_encoder_layer_norm_2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(24366592)))]; + tensor inner_encoder_ffn1_3_module_sequential_0_bias = const()[name = tensor("inner_encoder_ffn1_3_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(24367680)))]; + tensor inner_encoder_ffn1_3_module_sequential_0_weight = const()[name = tensor("inner_encoder_ffn1_3_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(24368768)))]; + tensor inner_encoder_ffn1_3_module_sequential_1_linear_bias = const()[name = tensor("inner_encoder_ffn1_3_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(24369856)))]; + tensor inner_encoder_ffn1_3_module_sequential_1_linear_weight = const()[name = tensor("inner_encoder_ffn1_3_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(24374016)))]; + tensor inner_encoder_ffn1_3_module_sequential_4_linear_bias = const()[name = tensor("inner_encoder_ffn1_3_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(25422656)))]; + tensor inner_encoder_ffn1_3_module_sequential_4_linear_weight = const()[name = tensor("inner_encoder_ffn1_3_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(25423744)))]; + tensor inner_encoder_ret_lns_3_bias = const()[name = tensor("inner_encoder_ret_lns_3_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(26472384)))]; + tensor inner_encoder_ret_lns_3_weight = const()[name = tensor("inner_encoder_ret_lns_3_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(26473472)))]; + tensor inner_encoder_q_proj_3_bias = const()[name = tensor("inner_encoder_q_proj_3_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(26474560)))]; + tensor inner_encoder_q_proj_3_weight = const()[name = tensor("inner_encoder_q_proj_3_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(26475648)))]; + tensor inner_encoder_k_proj_3_bias = const()[name = tensor("inner_encoder_k_proj_3_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(26737856)))]; + tensor inner_encoder_k_proj_3_weight = const()[name = tensor("inner_encoder_k_proj_3_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(26738944)))]; + tensor inner_encoder_v_proj_3_bias = const()[name = tensor("inner_encoder_v_proj_3_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27001152)))]; + tensor inner_encoder_v_proj_3_weight = const()[name = tensor("inner_encoder_v_proj_3_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27002240)))]; + tensor inner_encoder_g_proj_3_bias = const()[name = tensor("inner_encoder_g_proj_3_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27264448)))]; + tensor inner_encoder_g_proj_3_weight = const()[name = tensor("inner_encoder_g_proj_3_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27265536)))]; + tensor inner_encoder_out_proj_3_bias = const()[name = tensor("inner_encoder_out_proj_3_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27527744)))]; + tensor inner_encoder_out_proj_3_weight = const()[name = tensor("inner_encoder_out_proj_3_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27528832)))]; + tensor inner_encoder_conv_module_3_sequential_0_bias = const()[name = tensor("inner_encoder_conv_module_3_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27791040)))]; + tensor inner_encoder_conv_module_3_sequential_0_weight = const()[name = tensor("inner_encoder_conv_module_3_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27792128)))]; + tensor inner_encoder_conv_module_3_sequential_2_conv_bias = const()[name = tensor("inner_encoder_conv_module_3_sequential_2_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27793216)))]; + tensor inner_encoder_conv_module_3_sequential_2_conv_weight = const()[name = tensor("inner_encoder_conv_module_3_sequential_2_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27795328)))]; + tensor inner_encoder_conv_module_3_sequential_4_conv_weight = const()[name = tensor("inner_encoder_conv_module_3_sequential_4_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28319680)))]; + tensor inner_encoder_conv_module_3_sequential_5_running_var = const()[name = tensor("inner_encoder_conv_module_3_sequential_5_running_var"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28336128)))]; + tensor inner_encoder_conv_module_3_sequential_5_running_mean = const()[name = tensor("inner_encoder_conv_module_3_sequential_5_running_mean"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28337216)))]; + tensor inner_encoder_conv_module_3_sequential_5_bias = const()[name = tensor("inner_encoder_conv_module_3_sequential_5_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28338304)))]; + tensor inner_encoder_conv_module_3_sequential_5_weight = const()[name = tensor("inner_encoder_conv_module_3_sequential_5_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28339392)))]; + tensor inner_encoder_conv_module_3_sequential_7_conv_bias = const()[name = tensor("inner_encoder_conv_module_3_sequential_7_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28340480)))]; + tensor inner_encoder_conv_module_3_sequential_7_conv_weight = const()[name = tensor("inner_encoder_conv_module_3_sequential_7_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28341568)))]; + tensor inner_encoder_ffn2_3_module_sequential_0_bias = const()[name = tensor("inner_encoder_ffn2_3_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28603776)))]; + tensor inner_encoder_ffn2_3_module_sequential_0_weight = const()[name = tensor("inner_encoder_ffn2_3_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28604864)))]; + tensor inner_encoder_ffn2_3_module_sequential_1_linear_bias = const()[name = tensor("inner_encoder_ffn2_3_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28605952)))]; + tensor inner_encoder_ffn2_3_module_sequential_1_linear_weight = const()[name = tensor("inner_encoder_ffn2_3_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28610112)))]; + tensor inner_encoder_ffn2_3_module_sequential_4_linear_bias = const()[name = tensor("inner_encoder_ffn2_3_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(29658752)))]; + tensor inner_encoder_ffn2_3_module_sequential_4_linear_weight = const()[name = tensor("inner_encoder_ffn2_3_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(29659840)))]; + tensor inner_encoder_layer_norm_3_bias = const()[name = tensor("inner_encoder_layer_norm_3_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(30708480)))]; + tensor inner_encoder_layer_norm_3_weight = const()[name = tensor("inner_encoder_layer_norm_3_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(30709568)))]; + tensor inner_decoder__causal_mask = const()[name = tensor("inner_decoder__causal_mask"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(30710656)))]; + tensor inner_decoder_convert_bias = const()[name = tensor("inner_decoder_convert_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(30710848)))]; + tensor inner_decoder_convert_weight = const()[name = tensor("inner_decoder_convert_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(30711936)))]; + tensor inner_decoder_q_proj_0_bias = const()[name = tensor("inner_decoder_q_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(31236288)))]; + tensor inner_decoder_q_proj_0_weight = const()[name = tensor("inner_decoder_q_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(31237376)))]; + tensor inner_decoder_k_proj_0_bias = const()[name = tensor("inner_decoder_k_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(31499584)))]; + tensor inner_decoder_k_proj_0_weight = const()[name = tensor("inner_decoder_k_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(31500672)))]; + tensor inner_decoder_v_proj_0_bias = const()[name = tensor("inner_decoder_v_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(31762880)))]; + tensor inner_decoder_v_proj_0_weight = const()[name = tensor("inner_decoder_v_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(31763968)))]; + tensor inner_decoder_g_proj_0_bias = const()[name = tensor("inner_decoder_g_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32026176)))]; + tensor inner_decoder_g_proj_0_weight = const()[name = tensor("inner_decoder_g_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32027264)))]; + tensor inner_decoder_out_proj_0_bias = const()[name = tensor("inner_decoder_out_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32289472)))]; + tensor inner_decoder_out_proj_0_weight = const()[name = tensor("inner_decoder_out_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32290560)))]; + tensor inner_decoder_norm11_0_bias = const()[name = tensor("inner_decoder_norm11_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32552768)))]; + tensor inner_decoder_norm11_0_weight = const()[name = tensor("inner_decoder_norm11_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32553856)))]; + tensor inner_decoder_self_attn2_0_out_proj_bias = const()[name = tensor("inner_decoder_self_attn2_0_out_proj_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32554944)))]; + tensor inner_decoder_self_attn2_0_out_proj_weight = const()[name = tensor("inner_decoder_self_attn2_0_out_proj_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32556032)))]; + tensor inner_decoder_self_attn2_0_in_proj_bias = const()[name = tensor("inner_decoder_self_attn2_0_in_proj_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32818240)))]; + tensor inner_decoder_self_attn2_0_in_proj_weight = const()[name = tensor("inner_decoder_self_attn2_0_in_proj_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32821376)))]; + tensor inner_decoder_norm21_0_bias = const()[name = tensor("inner_decoder_norm21_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(33607872)))]; + tensor inner_decoder_norm21_0_weight = const()[name = tensor("inner_decoder_norm21_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(33608960)))]; + tensor inner_decoder_linear1_0_bias = const()[name = tensor("inner_decoder_linear1_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(33610048)))]; + tensor inner_decoder_linear1_0_weight = const()[name = tensor("inner_decoder_linear1_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(33618304)))]; + tensor inner_decoder_linear2_0_bias = const()[name = tensor("inner_decoder_linear2_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(35715520)))]; + tensor inner_decoder_linear2_0_weight = const()[name = tensor("inner_decoder_linear2_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(35716608)))]; + tensor inner_decoder_norm22_0_bias = const()[name = tensor("inner_decoder_norm22_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(37813824)))]; + tensor inner_decoder_norm22_0_weight = const()[name = tensor("inner_decoder_norm22_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(37814912)))]; + tensor inner_decoder_q_proj_1_bias = const()[name = tensor("inner_decoder_q_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(37816000)))]; + tensor inner_decoder_q_proj_1_weight = const()[name = tensor("inner_decoder_q_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(37817088)))]; + tensor inner_decoder_k_proj_1_bias = const()[name = tensor("inner_decoder_k_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38079296)))]; + tensor inner_decoder_k_proj_1_weight = const()[name = tensor("inner_decoder_k_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38080384)))]; + tensor inner_decoder_v_proj_1_bias = const()[name = tensor("inner_decoder_v_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38342592)))]; + tensor inner_decoder_v_proj_1_weight = const()[name = tensor("inner_decoder_v_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38343680)))]; + tensor inner_decoder_g_proj_1_bias = const()[name = tensor("inner_decoder_g_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38605888)))]; + tensor inner_decoder_g_proj_1_weight = const()[name = tensor("inner_decoder_g_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38606976)))]; + tensor inner_decoder_out_proj_1_bias = const()[name = tensor("inner_decoder_out_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38869184)))]; + tensor inner_decoder_out_proj_1_weight = const()[name = tensor("inner_decoder_out_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38870272)))]; + tensor inner_decoder_norm11_1_bias = const()[name = tensor("inner_decoder_norm11_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(39132480)))]; + tensor inner_decoder_norm11_1_weight = const()[name = tensor("inner_decoder_norm11_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(39133568)))]; + tensor inner_decoder_self_attn2_1_out_proj_bias = const()[name = tensor("inner_decoder_self_attn2_1_out_proj_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(39134656)))]; + tensor inner_decoder_self_attn2_1_out_proj_weight = const()[name = tensor("inner_decoder_self_attn2_1_out_proj_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(39135744)))]; + tensor inner_decoder_self_attn2_1_in_proj_bias = const()[name = tensor("inner_decoder_self_attn2_1_in_proj_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(39397952)))]; + tensor inner_decoder_self_attn2_1_in_proj_weight = const()[name = tensor("inner_decoder_self_attn2_1_in_proj_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(39401088)))]; + tensor inner_decoder_norm21_1_bias = const()[name = tensor("inner_decoder_norm21_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(40187584)))]; + tensor inner_decoder_norm21_1_weight = const()[name = tensor("inner_decoder_norm21_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(40188672)))]; + tensor inner_decoder_linear1_1_bias = const()[name = tensor("inner_decoder_linear1_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(40189760)))]; + tensor inner_decoder_linear1_1_weight = const()[name = tensor("inner_decoder_linear1_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(40198016)))]; + tensor inner_decoder_linear2_1_bias = const()[name = tensor("inner_decoder_linear2_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(42295232)))]; + tensor inner_decoder_linear2_1_weight = const()[name = tensor("inner_decoder_linear2_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(42296320)))]; + tensor inner_decoder_norm22_1_bias = const()[name = tensor("inner_decoder_norm22_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(44393536)))]; + tensor inner_decoder_norm22_1_weight = const()[name = tensor("inner_decoder_norm22_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(44394624)))]; + tensor var_19_begin_0 = const()[name = tensor("op_19_begin_0"), val = tensor([0, 0, 0])]; + tensor var_19_end_0 = const()[name = tensor("op_19_end_0"), val = tensor([1, 15, 23])]; + tensor var_19_end_mask_0 = const()[name = tensor("op_19_end_mask_0"), val = tensor([true, false, true])]; + tensor var_19 = slice_by_index(begin = var_19_begin_0, end = var_19_end_0, end_mask = var_19_end_mask_0, x = features)[name = tensor("op_19")]; + tensor var_29_begin_0 = const()[name = tensor("op_29_begin_0"), val = tensor([0, 10, 0])]; + tensor var_29_end_0 = const()[name = tensor("op_29_end_0"), val = tensor([1, 25, 23])]; + tensor var_29_end_mask_0 = const()[name = tensor("op_29_end_mask_0"), val = tensor([true, false, true])]; + tensor var_29 = slice_by_index(begin = var_29_begin_0, end = var_29_end_0, end_mask = var_29_end_mask_0, x = features)[name = tensor("op_29")]; + tensor var_39_begin_0 = const()[name = tensor("op_39_begin_0"), val = tensor([0, 20, 0])]; + tensor var_39_end_0 = const()[name = tensor("op_39_end_0"), val = tensor([1, 35, 23])]; + tensor var_39_end_mask_0 = const()[name = tensor("op_39_end_mask_0"), val = tensor([true, false, true])]; + tensor var_39 = slice_by_index(begin = var_39_begin_0, end = var_39_end_0, end_mask = var_39_end_mask_0, x = features)[name = tensor("op_39")]; + tensor var_49_begin_0 = const()[name = tensor("op_49_begin_0"), val = tensor([0, 30, 0])]; + tensor var_49_end_0 = const()[name = tensor("op_49_end_0"), val = tensor([1, 45, 23])]; + tensor var_49_end_mask_0 = const()[name = tensor("op_49_end_mask_0"), val = tensor([true, false, true])]; + tensor var_49 = slice_by_index(begin = var_49_begin_0, end = var_49_end_0, end_mask = var_49_end_mask_0, x = features)[name = tensor("op_49")]; + tensor var_59_begin_0 = const()[name = tensor("op_59_begin_0"), val = tensor([0, 40, 0])]; + tensor var_59_end_0 = const()[name = tensor("op_59_end_0"), val = tensor([1, 1, 23])]; + tensor var_59_end_mask_0 = const()[name = tensor("op_59_end_mask_0"), val = tensor([true, true, true])]; + tensor var_59 = slice_by_index(begin = var_59_begin_0, end = var_59_end_0, end_mask = var_59_end_mask_0, x = features)[name = tensor("op_59")]; + tensor stacked_axis_0 = const()[name = tensor("stacked_axis_0"), val = tensor(1)]; + tensor stacked = stack(axis = stacked_axis_0, values = (var_19, var_29, var_39, var_49, var_59))[name = tensor("stacked")]; + tensor var_66 = const()[name = tensor("op_66"), val = tensor([1, 5, 345])]; + tensor input_1 = reshape(shape = var_66, x = stacked)[name = tensor("input_1")]; + tensor var_69 = const()[name = tensor("op_69"), val = tensor(0x1p+0)]; + tensor var_75 = const()[name = tensor("op_75"), val = tensor(true)]; + tensor var_76 = const()[name = tensor("op_76"), val = tensor(0x1.4f8b58p-17)]; + tensor var_79 = const()[name = tensor("op_79"), val = tensor(0)]; + tensor var_81 = const()[name = tensor("op_81"), val = tensor(2)]; + tensor var_82 = const()[name = tensor("op_82"), val = tensor(-1)]; + tensor var_84 = const()[name = tensor("op_84"), val = tensor(0x1.0c6f7ap-20)]; + tensor var_90 = const()[name = tensor("op_90"), val = tensor(0x1.5798eep-27)]; + tensor var_93 = const()[name = tensor("op_93"), val = tensor(1)]; + tensor input_3 = linear(bias = inner_encoder_input_projection_linear_bias, weight = inner_encoder_input_projection_linear_weight, x = input_1)[name = tensor("linear_0")]; tensor input_5_axes_0 = const()[name = tensor("input_5_axes_0"), val = tensor([-1])]; - tensor input_5 = layer_norm(axes = input_5_axes_0, beta = encoder_ffn1_0_module_sequential_0_bias, epsilon = var_29, gamma = encoder_ffn1_0_module_sequential_0_weight, x = input_3)[name = tensor("input_5")]; - tensor inputs_1 = linear(bias = encoder_ffn1_0_module_sequential_1_linear_bias, weight = encoder_ffn1_0_module_sequential_1_linear_weight, x = input_5)[name = tensor("linear_1")]; - tensor input_7 = silu(x = inputs_1)[name = tensor("input_7")]; - tensor input_11 = linear(bias = encoder_ffn1_0_module_sequential_4_linear_bias, weight = encoder_ffn1_0_module_sequential_4_linear_weight, x = input_7)[name = tensor("linear_2")]; - tensor var_148 = const()[name = tensor("op_148"), val = tensor(0x1p-1)]; - tensor var_149 = mul(x = input_11, y = var_148)[name = tensor("op_149")]; - tensor input_13 = add(x = var_149, y = input_3)[name = tensor("input_13")]; + tensor input_5 = layer_norm(axes = input_5_axes_0, beta = inner_encoder_pre_layer_norm_bias, epsilon = var_76, gamma = inner_encoder_pre_layer_norm_weight, x = input_3)[name = tensor("input_5")]; + tensor input_7_axes_0 = const()[name = tensor("input_7_axes_0"), val = tensor([-1])]; + tensor input_7 = layer_norm(axes = input_7_axes_0, beta = inner_encoder_ffn1_0_module_sequential_0_bias, epsilon = var_76, gamma = inner_encoder_ffn1_0_module_sequential_0_weight, x = input_5)[name = tensor("input_7")]; + tensor inputs_1 = linear(bias = inner_encoder_ffn1_0_module_sequential_1_linear_bias, weight = inner_encoder_ffn1_0_module_sequential_1_linear_weight, x = input_7)[name = tensor("linear_1")]; + tensor input_9 = silu(x = inputs_1)[name = tensor("input_9")]; + tensor input_13 = linear(bias = inner_encoder_ffn1_0_module_sequential_4_linear_bias, weight = inner_encoder_ffn1_0_module_sequential_4_linear_weight, x = input_9)[name = tensor("linear_2")]; + tensor var_214 = const()[name = tensor("op_214"), val = tensor(0x1p-1)]; + tensor var_215 = mul(x = input_13, y = var_214)[name = tensor("op_215")]; + tensor input_15 = add(x = var_215, y = input_5)[name = tensor("input_15")]; tensor x_1_axes_0 = const()[name = tensor("x_1_axes_0"), val = tensor([-1])]; - tensor x_1 = layer_norm(axes = x_1_axes_0, beta = encoder_ret_lns_0_bias, epsilon = var_29, gamma = encoder_ret_lns_0_weight, x = input_13)[name = tensor("x_1")]; + tensor x_1 = layer_norm(axes = x_1_axes_0, beta = inner_encoder_ret_lns_0_bias, epsilon = var_76, gamma = inner_encoder_ret_lns_0_weight, x = input_15)[name = tensor("x_1")]; tensor prev_kv_1_begin_0 = const()[name = tensor("prev_kv_1_begin_0"), val = tensor([0, 0, 0, 0, 0])]; tensor prev_kv_1_end_0 = const()[name = tensor("prev_kv_1_end_0"), val = tensor([1, 1, 4, 64, 64])]; tensor prev_kv_1_end_mask_0 = const()[name = tensor("prev_kv_1_end_mask_0"), val = tensor([false, true, true, true, true])]; @@ -239,183 +265,183 @@ program(1.0) tensor prev_scale_1_end_mask_0 = const()[name = tensor("prev_scale_1_end_mask_0"), val = tensor([false, true])]; tensor prev_scale_1_squeeze_mask_0 = const()[name = tensor("prev_scale_1_squeeze_mask_0"), val = tensor([true, false])]; tensor prev_scale_1 = slice_by_index(begin = prev_scale_1_begin_0, end = prev_scale_1_end_0, end_mask = prev_scale_1_end_mask_0, squeeze_mask = prev_scale_1_squeeze_mask_0, x = enc_scale)[name = tensor("prev_scale_1")]; - tensor var_163 = linear(bias = encoder_q_proj_0_bias, weight = encoder_q_proj_0_weight, x = x_1)[name = tensor("linear_3")]; - tensor var_164 = const()[name = tensor("op_164"), val = tensor([1, 5, 4, 64])]; - tensor var_165 = reshape(shape = var_164, x = var_163)[name = tensor("op_165")]; + tensor var_229 = linear(bias = inner_encoder_q_proj_0_bias, weight = inner_encoder_q_proj_0_weight, x = x_1)[name = tensor("linear_3")]; + tensor var_230 = const()[name = tensor("op_230"), val = tensor([1, 5, 4, 64])]; + tensor var_231 = reshape(shape = var_230, x = var_229)[name = tensor("op_231")]; tensor q_1_perm_0 = const()[name = tensor("q_1_perm_0"), val = tensor([0, 2, 1, 3])]; - tensor var_169 = linear(bias = encoder_k_proj_0_bias, weight = encoder_k_proj_0_weight, x = x_1)[name = tensor("linear_4")]; - tensor var_170 = const()[name = tensor("op_170"), val = tensor(0x1p-3)]; - tensor var_171 = mul(x = var_169, y = var_170)[name = tensor("op_171")]; - tensor var_172 = const()[name = tensor("op_172"), val = tensor([1, 5, 4, 64])]; - tensor var_173 = reshape(shape = var_172, x = var_171)[name = tensor("op_173")]; + tensor var_235 = linear(bias = inner_encoder_k_proj_0_bias, weight = inner_encoder_k_proj_0_weight, x = x_1)[name = tensor("linear_4")]; + tensor var_236 = const()[name = tensor("op_236"), val = tensor(0x1p-3)]; + tensor var_237 = mul(x = var_235, y = var_236)[name = tensor("op_237")]; + tensor var_238 = const()[name = tensor("op_238"), val = tensor([1, 5, 4, 64])]; + tensor var_239 = reshape(shape = var_238, x = var_237)[name = tensor("op_239")]; tensor k_1_perm_0 = const()[name = tensor("k_1_perm_0"), val = tensor([0, 2, 1, 3])]; - tensor var_177 = linear(bias = encoder_v_proj_0_bias, weight = encoder_v_proj_0_weight, x = x_1)[name = tensor("linear_5")]; - tensor var_178 = const()[name = tensor("op_178"), val = tensor([1, 5, 4, 64])]; - tensor var_179 = reshape(shape = var_178, x = var_177)[name = tensor("op_179")]; + tensor var_243 = linear(bias = inner_encoder_v_proj_0_bias, weight = inner_encoder_v_proj_0_weight, x = x_1)[name = tensor("linear_5")]; + tensor var_244 = const()[name = tensor("op_244"), val = tensor([1, 5, 4, 64])]; + tensor var_245 = reshape(shape = var_244, x = var_243)[name = tensor("op_245")]; tensor v_1_perm_0 = const()[name = tensor("v_1_perm_0"), val = tensor([0, 2, 1, 3])]; - tensor input_17 = linear(bias = encoder_g_proj_0_bias, weight = encoder_g_proj_0_weight, x = x_1)[name = tensor("linear_6")]; + tensor input_19 = linear(bias = inner_encoder_g_proj_0_bias, weight = inner_encoder_g_proj_0_weight, x = x_1)[name = tensor("linear_6")]; tensor sqrt_s0_1 = sqrt(x = prev_scale_1)[name = tensor("sqrt_s0_1")]; - tensor s_t_1 = add(x = prev_scale_1, y = encoder__t_index)[name = tensor("s_t_1")]; + tensor s_t_1 = add(x = prev_scale_1, y = inner_encoder__t_index)[name = tensor("s_t_1")]; tensor sqrt_s_t_1 = sqrt(x = s_t_1)[name = tensor("sqrt_s_t_1")]; tensor qk_1_transpose_x_1 = const()[name = tensor("qk_1_transpose_x_1"), val = tensor(false)]; tensor qk_1_transpose_y_1 = const()[name = tensor("qk_1_transpose_y_1"), val = tensor(true)]; - tensor k_1 = transpose(perm = k_1_perm_0, x = var_173)[name = tensor("transpose_57")]; - tensor q_1 = transpose(perm = q_1_perm_0, x = var_165)[name = tensor("transpose_58")]; + tensor k_1 = transpose(perm = k_1_perm_0, x = var_239)[name = tensor("transpose_57")]; + tensor q_1 = transpose(perm = q_1_perm_0, x = var_231)[name = tensor("transpose_58")]; tensor qk_1 = matmul(transpose_x = qk_1_transpose_x_1, transpose_y = qk_1_transpose_y_1, x = q_1, y = k_1)[name = tensor("qk_1")]; - tensor var_189 = const()[name = tensor("op_189"), val = tensor([5, 1])]; - tensor var_190 = reshape(shape = var_189, x = sqrt_s_t_1)[name = tensor("op_190")]; - tensor M_1 = real_div(x = encoder__causal_mask, y = var_190)[name = tensor("M_1")]; - tensor var_192 = mul(x = qk_1, y = M_1)[name = tensor("op_192")]; + tensor var_255 = const()[name = tensor("op_255"), val = tensor([5, 1])]; + tensor var_256 = reshape(shape = var_255, x = sqrt_s_t_1)[name = tensor("op_256")]; + tensor M_1 = real_div(x = inner_encoder__causal_mask, y = var_256)[name = tensor("M_1")]; + tensor var_258 = mul(x = qk_1, y = M_1)[name = tensor("op_258")]; tensor inner_1_transpose_x_0 = const()[name = tensor("inner_1_transpose_x_0"), val = tensor(false)]; tensor inner_1_transpose_y_0 = const()[name = tensor("inner_1_transpose_y_0"), val = tensor(false)]; - tensor v_1 = transpose(perm = v_1_perm_0, x = var_179)[name = tensor("transpose_56")]; - tensor inner_1 = matmul(transpose_x = inner_1_transpose_x_0, transpose_y = inner_1_transpose_y_0, x = var_192, y = v_1)[name = tensor("inner_1")]; - tensor var_194_transpose_x_0 = const()[name = tensor("op_194_transpose_x_0"), val = tensor(false)]; - tensor var_194_transpose_y_0 = const()[name = tensor("op_194_transpose_y_0"), val = tensor(false)]; - tensor var_194 = matmul(transpose_x = var_194_transpose_x_0, transpose_y = var_194_transpose_y_0, x = q_1, y = prev_kv_1)[name = tensor("op_194")]; - tensor var_195 = real_div(x = sqrt_s0_1, y = sqrt_s_t_1)[name = tensor("op_195")]; - tensor var_196 = const()[name = tensor("op_196"), val = tensor([1, 1, 5, 1])]; - tensor var_197 = reshape(shape = var_196, x = var_195)[name = tensor("op_197")]; - tensor cross_1 = mul(x = var_194, y = var_197)[name = tensor("cross_1")]; + tensor v_1 = transpose(perm = v_1_perm_0, x = var_245)[name = tensor("transpose_56")]; + tensor inner_1 = matmul(transpose_x = inner_1_transpose_x_0, transpose_y = inner_1_transpose_y_0, x = var_258, y = v_1)[name = tensor("inner_1")]; + tensor var_260_transpose_x_0 = const()[name = tensor("op_260_transpose_x_0"), val = tensor(false)]; + tensor var_260_transpose_y_0 = const()[name = tensor("op_260_transpose_y_0"), val = tensor(false)]; + tensor var_260 = matmul(transpose_x = var_260_transpose_x_0, transpose_y = var_260_transpose_y_0, x = q_1, y = prev_kv_1)[name = tensor("op_260")]; + tensor var_261 = real_div(x = sqrt_s0_1, y = sqrt_s_t_1)[name = tensor("op_261")]; + tensor var_262 = const()[name = tensor("op_262"), val = tensor([1, 1, 5, 1])]; + tensor var_263 = reshape(shape = var_262, x = var_261)[name = tensor("op_263")]; + tensor cross_1 = mul(x = var_260, y = var_263)[name = tensor("cross_1")]; tensor out_1 = add(x = inner_1, y = cross_1)[name = tensor("out_1")]; - tensor var_200 = mul(x = prev_kv_1, y = sqrt_s0_1)[name = tensor("op_200")]; - tensor var_202_transpose_x_1 = const()[name = tensor("op_202_transpose_x_1"), val = tensor(true)]; - tensor var_202_transpose_y_1 = const()[name = tensor("op_202_transpose_y_1"), val = tensor(false)]; - tensor var_202 = matmul(transpose_x = var_202_transpose_x_1, transpose_y = var_202_transpose_y_1, x = k_1, y = v_1)[name = tensor("op_202")]; - tensor new_kv_unnorm_1 = add(x = var_200, y = var_202)[name = tensor("new_kv_unnorm_1")]; - tensor var_204 = const()[name = tensor("op_204"), val = tensor(0x1.4p+2)]; - tensor new_scale_1 = add(x = prev_scale_1, y = var_204)[name = tensor("new_scale_1")]; - tensor var_206 = sqrt(x = new_scale_1)[name = tensor("op_206")]; - tensor var_207 = real_div(x = new_kv_unnorm_1, y = var_206)[name = tensor("op_207")]; - tensor var_208_perm_0 = const()[name = tensor("op_208_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_266 = mul(x = prev_kv_1, y = sqrt_s0_1)[name = tensor("op_266")]; + tensor var_268_transpose_x_1 = const()[name = tensor("op_268_transpose_x_1"), val = tensor(true)]; + tensor var_268_transpose_y_1 = const()[name = tensor("op_268_transpose_y_1"), val = tensor(false)]; + tensor var_268 = matmul(transpose_x = var_268_transpose_x_1, transpose_y = var_268_transpose_y_1, x = k_1, y = v_1)[name = tensor("op_268")]; + tensor new_kv_unnorm_1 = add(x = var_266, y = var_268)[name = tensor("new_kv_unnorm_1")]; + tensor var_270 = const()[name = tensor("op_270"), val = tensor(0x1.4p+2)]; + tensor new_scale_1 = add(x = prev_scale_1, y = var_270)[name = tensor("new_scale_1")]; + tensor var_272 = sqrt(x = new_scale_1)[name = tensor("op_272")]; + tensor var_273 = real_div(x = new_kv_unnorm_1, y = var_272)[name = tensor("op_273")]; + tensor var_274_perm_0 = const()[name = tensor("op_274_perm_0"), val = tensor([0, 2, 1, 3])]; tensor out_3_axes_0 = const()[name = tensor("out_3_axes_0"), val = tensor([-1])]; - tensor var_208 = transpose(perm = var_208_perm_0, x = out_1)[name = tensor("transpose_55")]; - tensor out_3 = layer_norm(axes = out_3_axes_0, epsilon = var_18, x = var_208)[name = tensor("out_3")]; - tensor var_212 = const()[name = tensor("op_212"), val = tensor([1, 5, 256])]; - tensor out_5 = reshape(shape = var_212, x = out_3)[name = tensor("out_5")]; - tensor var_214 = silu(x = input_17)[name = tensor("op_214")]; - tensor input_19 = mul(x = var_214, y = out_5)[name = tensor("input_19")]; - tensor ret_out_1 = linear(bias = encoder_out_proj_0_bias, weight = encoder_out_proj_0_weight, x = input_19)[name = tensor("linear_7")]; - tensor x_3 = add(x = input_13, y = ret_out_1)[name = tensor("x_3")]; + tensor var_274 = transpose(perm = var_274_perm_0, x = out_1)[name = tensor("transpose_55")]; + tensor out_3 = layer_norm(axes = out_3_axes_0, epsilon = var_84, x = var_274)[name = tensor("out_3")]; + tensor var_278 = const()[name = tensor("op_278"), val = tensor([1, 5, 256])]; + tensor out_5 = reshape(shape = var_278, x = out_3)[name = tensor("out_5")]; + tensor var_280 = silu(x = input_19)[name = tensor("op_280")]; + tensor input_21 = mul(x = var_280, y = out_5)[name = tensor("input_21")]; + tensor ret_out_1 = linear(bias = inner_encoder_out_proj_0_bias, weight = inner_encoder_out_proj_0_weight, x = input_21)[name = tensor("linear_7")]; + tensor x_3 = add(x = input_15, y = ret_out_1)[name = tensor("x_3")]; tensor window_1_begin_0 = const()[name = tensor("window_1_begin_0"), val = tensor([0, 0, 0, 0])]; tensor window_1_end_0 = const()[name = tensor("window_1_end_0"), val = tensor([1, 1, 16, 256])]; tensor window_1_end_mask_0 = const()[name = tensor("window_1_end_mask_0"), val = tensor([false, true, true, true])]; tensor window_1_squeeze_mask_0 = const()[name = tensor("window_1_squeeze_mask_0"), val = tensor([true, false, false, false])]; tensor window_1 = slice_by_index(begin = window_1_begin_0, end = window_1_end_0, end_mask = window_1_end_mask_0, squeeze_mask = window_1_squeeze_mask_0, x = enc_conv_cache)[name = tensor("window_1")]; - tensor var_222_begin_0 = const()[name = tensor("op_222_begin_0"), val = tensor([0, 0, 0])]; - tensor var_222_end_0 = const()[name = tensor("op_222_end_0"), val = tensor([1, 1, 256])]; - tensor var_222_end_mask_0 = const()[name = tensor("op_222_end_mask_0"), val = tensor([true, false, true])]; - tensor var_222 = slice_by_index(begin = var_222_begin_0, end = var_222_end_0, end_mask = var_222_end_mask_0, x = x_3)[name = tensor("op_222")]; - tensor var_225_begin_0 = const()[name = tensor("op_225_begin_0"), val = tensor([0, 1, 0])]; - tensor var_225_end_0 = const()[name = tensor("op_225_end_0"), val = tensor([1, 16, 256])]; - tensor var_225_end_mask_0 = const()[name = tensor("op_225_end_mask_0"), val = tensor([true, true, true])]; - tensor var_225 = slice_by_index(begin = var_225_begin_0, end = var_225_end_0, end_mask = var_225_end_mask_0, x = window_1)[name = tensor("op_225")]; + tensor var_288_begin_0 = const()[name = tensor("op_288_begin_0"), val = tensor([0, 0, 0])]; + tensor var_288_end_0 = const()[name = tensor("op_288_end_0"), val = tensor([1, 1, 256])]; + tensor var_288_end_mask_0 = const()[name = tensor("op_288_end_mask_0"), val = tensor([true, false, true])]; + tensor var_288 = slice_by_index(begin = var_288_begin_0, end = var_288_end_0, end_mask = var_288_end_mask_0, x = x_3)[name = tensor("op_288")]; + tensor var_291_begin_0 = const()[name = tensor("op_291_begin_0"), val = tensor([0, 1, 0])]; + tensor var_291_end_0 = const()[name = tensor("op_291_end_0"), val = tensor([1, 16, 256])]; + tensor var_291_end_mask_0 = const()[name = tensor("op_291_end_mask_0"), val = tensor([true, true, true])]; + tensor var_291 = slice_by_index(begin = var_291_begin_0, end = var_291_end_0, end_mask = var_291_end_mask_0, x = window_1)[name = tensor("op_291")]; tensor window_3_interleave_0 = const()[name = tensor("window_3_interleave_0"), val = tensor(false)]; - tensor window_3 = concat(axis = var_27, interleave = window_3_interleave_0, values = (var_225, var_222))[name = tensor("window_3")]; - tensor var_230_begin_0 = const()[name = tensor("op_230_begin_0"), val = tensor([0, 1, 0])]; - tensor var_230_end_0 = const()[name = tensor("op_230_end_0"), val = tensor([1, 2, 256])]; - tensor var_230_end_mask_0 = const()[name = tensor("op_230_end_mask_0"), val = tensor([true, false, true])]; - tensor var_230 = slice_by_index(begin = var_230_begin_0, end = var_230_end_0, end_mask = var_230_end_mask_0, x = x_3)[name = tensor("op_230")]; - tensor var_233_begin_0 = const()[name = tensor("op_233_begin_0"), val = tensor([0, 1, 0])]; - tensor var_233_end_0 = const()[name = tensor("op_233_end_0"), val = tensor([1, 16, 256])]; - tensor var_233_end_mask_0 = const()[name = tensor("op_233_end_mask_0"), val = tensor([true, true, true])]; - tensor var_233 = slice_by_index(begin = var_233_begin_0, end = var_233_end_0, end_mask = var_233_end_mask_0, x = window_3)[name = tensor("op_233")]; + tensor window_3 = concat(axis = var_93, interleave = window_3_interleave_0, values = (var_291, var_288))[name = tensor("window_3")]; + tensor var_296_begin_0 = const()[name = tensor("op_296_begin_0"), val = tensor([0, 1, 0])]; + tensor var_296_end_0 = const()[name = tensor("op_296_end_0"), val = tensor([1, 2, 256])]; + tensor var_296_end_mask_0 = const()[name = tensor("op_296_end_mask_0"), val = tensor([true, false, true])]; + tensor var_296 = slice_by_index(begin = var_296_begin_0, end = var_296_end_0, end_mask = var_296_end_mask_0, x = x_3)[name = tensor("op_296")]; + tensor var_299_begin_0 = const()[name = tensor("op_299_begin_0"), val = tensor([0, 1, 0])]; + tensor var_299_end_0 = const()[name = tensor("op_299_end_0"), val = tensor([1, 16, 256])]; + tensor var_299_end_mask_0 = const()[name = tensor("op_299_end_mask_0"), val = tensor([true, true, true])]; + tensor var_299 = slice_by_index(begin = var_299_begin_0, end = var_299_end_0, end_mask = var_299_end_mask_0, x = window_3)[name = tensor("op_299")]; tensor window_5_interleave_0 = const()[name = tensor("window_5_interleave_0"), val = tensor(false)]; - tensor window_5 = concat(axis = var_27, interleave = window_5_interleave_0, values = (var_233, var_230))[name = tensor("window_5")]; - tensor var_238_begin_0 = const()[name = tensor("op_238_begin_0"), val = tensor([0, 2, 0])]; - tensor var_238_end_0 = const()[name = tensor("op_238_end_0"), val = tensor([1, 3, 256])]; - tensor var_238_end_mask_0 = const()[name = tensor("op_238_end_mask_0"), val = tensor([true, false, true])]; - tensor var_238 = slice_by_index(begin = var_238_begin_0, end = var_238_end_0, end_mask = var_238_end_mask_0, x = x_3)[name = tensor("op_238")]; - tensor var_241_begin_0 = const()[name = tensor("op_241_begin_0"), val = tensor([0, 1, 0])]; - tensor var_241_end_0 = const()[name = tensor("op_241_end_0"), val = tensor([1, 16, 256])]; - tensor var_241_end_mask_0 = const()[name = tensor("op_241_end_mask_0"), val = tensor([true, true, true])]; - tensor var_241 = slice_by_index(begin = var_241_begin_0, end = var_241_end_0, end_mask = var_241_end_mask_0, x = window_5)[name = tensor("op_241")]; + tensor window_5 = concat(axis = var_93, interleave = window_5_interleave_0, values = (var_299, var_296))[name = tensor("window_5")]; + tensor var_304_begin_0 = const()[name = tensor("op_304_begin_0"), val = tensor([0, 2, 0])]; + tensor var_304_end_0 = const()[name = tensor("op_304_end_0"), val = tensor([1, 3, 256])]; + tensor var_304_end_mask_0 = const()[name = tensor("op_304_end_mask_0"), val = tensor([true, false, true])]; + tensor var_304 = slice_by_index(begin = var_304_begin_0, end = var_304_end_0, end_mask = var_304_end_mask_0, x = x_3)[name = tensor("op_304")]; + tensor var_307_begin_0 = const()[name = tensor("op_307_begin_0"), val = tensor([0, 1, 0])]; + tensor var_307_end_0 = const()[name = tensor("op_307_end_0"), val = tensor([1, 16, 256])]; + tensor var_307_end_mask_0 = const()[name = tensor("op_307_end_mask_0"), val = tensor([true, true, true])]; + tensor var_307 = slice_by_index(begin = var_307_begin_0, end = var_307_end_0, end_mask = var_307_end_mask_0, x = window_5)[name = tensor("op_307")]; tensor window_7_interleave_0 = const()[name = tensor("window_7_interleave_0"), val = tensor(false)]; - tensor window_7 = concat(axis = var_27, interleave = window_7_interleave_0, values = (var_241, var_238))[name = tensor("window_7")]; - tensor var_246_begin_0 = const()[name = tensor("op_246_begin_0"), val = tensor([0, 3, 0])]; - tensor var_246_end_0 = const()[name = tensor("op_246_end_0"), val = tensor([1, 4, 256])]; - tensor var_246_end_mask_0 = const()[name = tensor("op_246_end_mask_0"), val = tensor([true, false, true])]; - tensor var_246 = slice_by_index(begin = var_246_begin_0, end = var_246_end_0, end_mask = var_246_end_mask_0, x = x_3)[name = tensor("op_246")]; - tensor var_249_begin_0 = const()[name = tensor("op_249_begin_0"), val = tensor([0, 1, 0])]; - tensor var_249_end_0 = const()[name = tensor("op_249_end_0"), val = tensor([1, 16, 256])]; - tensor var_249_end_mask_0 = const()[name = tensor("op_249_end_mask_0"), val = tensor([true, true, true])]; - tensor var_249 = slice_by_index(begin = var_249_begin_0, end = var_249_end_0, end_mask = var_249_end_mask_0, x = window_7)[name = tensor("op_249")]; + tensor window_7 = concat(axis = var_93, interleave = window_7_interleave_0, values = (var_307, var_304))[name = tensor("window_7")]; + tensor var_312_begin_0 = const()[name = tensor("op_312_begin_0"), val = tensor([0, 3, 0])]; + tensor var_312_end_0 = const()[name = tensor("op_312_end_0"), val = tensor([1, 4, 256])]; + tensor var_312_end_mask_0 = const()[name = tensor("op_312_end_mask_0"), val = tensor([true, false, true])]; + tensor var_312 = slice_by_index(begin = var_312_begin_0, end = var_312_end_0, end_mask = var_312_end_mask_0, x = x_3)[name = tensor("op_312")]; + tensor var_315_begin_0 = const()[name = tensor("op_315_begin_0"), val = tensor([0, 1, 0])]; + tensor var_315_end_0 = const()[name = tensor("op_315_end_0"), val = tensor([1, 16, 256])]; + tensor var_315_end_mask_0 = const()[name = tensor("op_315_end_mask_0"), val = tensor([true, true, true])]; + tensor var_315 = slice_by_index(begin = var_315_begin_0, end = var_315_end_0, end_mask = var_315_end_mask_0, x = window_7)[name = tensor("op_315")]; tensor window_9_interleave_0 = const()[name = tensor("window_9_interleave_0"), val = tensor(false)]; - tensor window_9 = concat(axis = var_27, interleave = window_9_interleave_0, values = (var_249, var_246))[name = tensor("window_9")]; - tensor var_254_begin_0 = const()[name = tensor("op_254_begin_0"), val = tensor([0, 4, 0])]; - tensor var_254_end_0 = const()[name = tensor("op_254_end_0"), val = tensor([1, 1, 256])]; - tensor var_254_end_mask_0 = const()[name = tensor("op_254_end_mask_0"), val = tensor([true, true, true])]; - tensor var_254 = slice_by_index(begin = var_254_begin_0, end = var_254_end_0, end_mask = var_254_end_mask_0, x = x_3)[name = tensor("op_254")]; - tensor var_257_begin_0 = const()[name = tensor("op_257_begin_0"), val = tensor([0, 1, 0])]; - tensor var_257_end_0 = const()[name = tensor("op_257_end_0"), val = tensor([1, 16, 256])]; - tensor var_257_end_mask_0 = const()[name = tensor("op_257_end_mask_0"), val = tensor([true, true, true])]; - tensor var_257 = slice_by_index(begin = var_257_begin_0, end = var_257_end_0, end_mask = var_257_end_mask_0, x = window_9)[name = tensor("op_257")]; + tensor window_9 = concat(axis = var_93, interleave = window_9_interleave_0, values = (var_315, var_312))[name = tensor("window_9")]; + tensor var_320_begin_0 = const()[name = tensor("op_320_begin_0"), val = tensor([0, 4, 0])]; + tensor var_320_end_0 = const()[name = tensor("op_320_end_0"), val = tensor([1, 1, 256])]; + tensor var_320_end_mask_0 = const()[name = tensor("op_320_end_mask_0"), val = tensor([true, true, true])]; + tensor var_320 = slice_by_index(begin = var_320_begin_0, end = var_320_end_0, end_mask = var_320_end_mask_0, x = x_3)[name = tensor("op_320")]; + tensor var_323_begin_0 = const()[name = tensor("op_323_begin_0"), val = tensor([0, 1, 0])]; + tensor var_323_end_0 = const()[name = tensor("op_323_end_0"), val = tensor([1, 16, 256])]; + tensor var_323_end_mask_0 = const()[name = tensor("op_323_end_mask_0"), val = tensor([true, true, true])]; + tensor var_323 = slice_by_index(begin = var_323_begin_0, end = var_323_end_0, end_mask = var_323_end_mask_0, x = window_9)[name = tensor("op_323")]; tensor window_11_interleave_0 = const()[name = tensor("window_11_interleave_0"), val = tensor(false)]; - tensor window_11 = concat(axis = var_27, interleave = window_11_interleave_0, values = (var_257, var_254))[name = tensor("window_11")]; - tensor input_21_interleave_0 = const()[name = tensor("input_21_interleave_0"), val = tensor(false)]; - tensor input_21 = concat(axis = var_24, interleave = input_21_interleave_0, values = (window_3, window_5, window_7, window_9, window_11))[name = tensor("input_21")]; + tensor window_11 = concat(axis = var_93, interleave = window_11_interleave_0, values = (var_323, var_320))[name = tensor("window_11")]; + tensor input_23_interleave_0 = const()[name = tensor("input_23_interleave_0"), val = tensor(false)]; + tensor input_23 = concat(axis = var_79, interleave = input_23_interleave_0, values = (window_3, window_5, window_7, window_9, window_11))[name = tensor("input_23")]; tensor x_5_axes_0 = const()[name = tensor("x_5_axes_0"), val = tensor([-1])]; - tensor x_5 = layer_norm(axes = x_5_axes_0, beta = encoder_conv_module_0_sequential_0_bias, epsilon = var_29, gamma = encoder_conv_module_0_sequential_0_weight, x = input_21)[name = tensor("x_5")]; - tensor input_23_perm_0 = const()[name = tensor("input_23_perm_0"), val = tensor([0, 2, 1])]; + tensor x_5 = layer_norm(axes = x_5_axes_0, beta = inner_encoder_conv_module_0_sequential_0_bias, epsilon = var_76, gamma = inner_encoder_conv_module_0_sequential_0_weight, x = input_23)[name = tensor("x_5")]; + tensor input_25_perm_0 = const()[name = tensor("input_25_perm_0"), val = tensor([0, 2, 1])]; tensor inputs_3_pad_type_0 = const()[name = tensor("inputs_3_pad_type_0"), val = tensor("valid")]; tensor inputs_3_strides_0 = const()[name = tensor("inputs_3_strides_0"), val = tensor([1])]; tensor inputs_3_pad_0 = const()[name = tensor("inputs_3_pad_0"), val = tensor([0, 0])]; tensor inputs_3_dilations_0 = const()[name = tensor("inputs_3_dilations_0"), val = tensor([1])]; tensor inputs_3_groups_0 = const()[name = tensor("inputs_3_groups_0"), val = tensor(1)]; - tensor input_23 = transpose(perm = input_23_perm_0, x = x_5)[name = tensor("transpose_54")]; - tensor inputs_3 = conv(bias = encoder_conv_module_0_sequential_2_conv_bias, dilations = inputs_3_dilations_0, groups = inputs_3_groups_0, pad = inputs_3_pad_0, pad_type = inputs_3_pad_type_0, strides = inputs_3_strides_0, weight = encoder_conv_module_0_sequential_2_conv_weight, x = input_23)[name = tensor("inputs_3")]; - tensor var_282_split_sizes_0 = const()[name = tensor("op_282_split_sizes_0"), val = tensor([256, 256])]; - tensor var_282_axis_0 = const()[name = tensor("op_282_axis_0"), val = tensor(1)]; - tensor var_282_0, tensor var_282_1 = split(axis = var_282_axis_0, split_sizes = var_282_split_sizes_0, x = inputs_3)[name = tensor("op_282")]; - tensor var_284 = sigmoid(x = var_282_1)[name = tensor("op_284")]; - tensor inputs_5 = mul(x = var_282_0, y = var_284)[name = tensor("inputs_5")]; + tensor input_25 = transpose(perm = input_25_perm_0, x = x_5)[name = tensor("transpose_54")]; + tensor inputs_3 = conv(bias = inner_encoder_conv_module_0_sequential_2_conv_bias, dilations = inputs_3_dilations_0, groups = inputs_3_groups_0, pad = inputs_3_pad_0, pad_type = inputs_3_pad_type_0, strides = inputs_3_strides_0, weight = inner_encoder_conv_module_0_sequential_2_conv_weight, x = input_25)[name = tensor("inputs_3")]; + tensor var_348_split_sizes_0 = const()[name = tensor("op_348_split_sizes_0"), val = tensor([256, 256])]; + tensor var_348_axis_0 = const()[name = tensor("op_348_axis_0"), val = tensor(1)]; + tensor var_348_0, tensor var_348_1 = split(axis = var_348_axis_0, split_sizes = var_348_split_sizes_0, x = inputs_3)[name = tensor("op_348")]; + tensor var_350 = sigmoid(x = var_348_1)[name = tensor("op_350")]; + tensor inputs_5 = mul(x = var_348_0, y = var_350)[name = tensor("inputs_5")]; tensor outputs_aug_1_pad_type_0 = const()[name = tensor("outputs_aug_1_pad_type_0"), val = tensor("custom")]; tensor outputs_aug_1_pad_0 = const()[name = tensor("outputs_aug_1_pad_0"), val = tensor([15, 15])]; tensor outputs_aug_1_groups_0 = const()[name = tensor("outputs_aug_1_groups_0"), val = tensor(256)]; tensor outputs_aug_1_strides_0 = const()[name = tensor("outputs_aug_1_strides_0"), val = tensor([1])]; tensor outputs_aug_1_dilations_0 = const()[name = tensor("outputs_aug_1_dilations_0"), val = tensor([1])]; - tensor outputs_aug_1 = conv(dilations = outputs_aug_1_dilations_0, groups = outputs_aug_1_groups_0, pad = outputs_aug_1_pad_0, pad_type = outputs_aug_1_pad_type_0, strides = outputs_aug_1_strides_0, weight = encoder_conv_module_0_sequential_4_conv_weight, x = inputs_5)[name = tensor("outputs_aug_1")]; - tensor input_25_begin_0 = const()[name = tensor("input_25_begin_0"), val = tensor([0, 0, 0])]; - tensor input_25_end_0 = const()[name = tensor("input_25_end_0"), val = tensor([5, 256, 16])]; - tensor input_25_end_mask_0 = const()[name = tensor("input_25_end_mask_0"), val = tensor([true, true, false])]; - tensor input_25 = slice_by_index(begin = input_25_begin_0, end = input_25_end_0, end_mask = input_25_end_mask_0, x = outputs_aug_1)[name = tensor("input_25")]; - tensor inputs_7 = batch_norm(beta = encoder_conv_module_0_sequential_5_bias, epsilon = var_29, gamma = encoder_conv_module_0_sequential_5_weight, mean = encoder_conv_module_0_sequential_5_running_mean, variance = encoder_conv_module_0_sequential_5_running_var, x = input_25)[name = tensor("inputs_7")]; - tensor input_27 = silu(x = inputs_7)[name = tensor("input_27")]; - tensor input_29_pad_type_0 = const()[name = tensor("input_29_pad_type_0"), val = tensor("valid")]; - tensor input_29_strides_0 = const()[name = tensor("input_29_strides_0"), val = tensor([1])]; - tensor input_29_pad_0 = const()[name = tensor("input_29_pad_0"), val = tensor([0, 0])]; - tensor input_29_dilations_0 = const()[name = tensor("input_29_dilations_0"), val = tensor([1])]; - tensor input_29_groups_0 = const()[name = tensor("input_29_groups_0"), val = tensor(1)]; - tensor input_29 = conv(bias = encoder_conv_module_0_sequential_7_conv_bias, dilations = input_29_dilations_0, groups = input_29_groups_0, pad = input_29_pad_0, pad_type = input_29_pad_type_0, strides = input_29_strides_0, weight = encoder_conv_module_0_sequential_7_conv_weight, x = input_27)[name = tensor("input_29")]; + tensor outputs_aug_1 = conv(dilations = outputs_aug_1_dilations_0, groups = outputs_aug_1_groups_0, pad = outputs_aug_1_pad_0, pad_type = outputs_aug_1_pad_type_0, strides = outputs_aug_1_strides_0, weight = inner_encoder_conv_module_0_sequential_4_conv_weight, x = inputs_5)[name = tensor("outputs_aug_1")]; + tensor input_27_begin_0 = const()[name = tensor("input_27_begin_0"), val = tensor([0, 0, 0])]; + tensor input_27_end_0 = const()[name = tensor("input_27_end_0"), val = tensor([5, 256, 16])]; + tensor input_27_end_mask_0 = const()[name = tensor("input_27_end_mask_0"), val = tensor([true, true, false])]; + tensor input_27 = slice_by_index(begin = input_27_begin_0, end = input_27_end_0, end_mask = input_27_end_mask_0, x = outputs_aug_1)[name = tensor("input_27")]; + tensor inputs_7 = batch_norm(beta = inner_encoder_conv_module_0_sequential_5_bias, epsilon = var_76, gamma = inner_encoder_conv_module_0_sequential_5_weight, mean = inner_encoder_conv_module_0_sequential_5_running_mean, variance = inner_encoder_conv_module_0_sequential_5_running_var, x = input_27)[name = tensor("inputs_7")]; + tensor input_29 = silu(x = inputs_7)[name = tensor("input_29")]; + tensor input_31_pad_type_0 = const()[name = tensor("input_31_pad_type_0"), val = tensor("valid")]; + tensor input_31_strides_0 = const()[name = tensor("input_31_strides_0"), val = tensor([1])]; + tensor input_31_pad_0 = const()[name = tensor("input_31_pad_0"), val = tensor([0, 0])]; + tensor input_31_dilations_0 = const()[name = tensor("input_31_dilations_0"), val = tensor([1])]; + tensor input_31_groups_0 = const()[name = tensor("input_31_groups_0"), val = tensor(1)]; + tensor input_31 = conv(bias = inner_encoder_conv_module_0_sequential_7_conv_bias, dilations = input_31_dilations_0, groups = input_31_groups_0, pad = input_31_pad_0, pad_type = input_31_pad_type_0, strides = input_31_strides_0, weight = inner_encoder_conv_module_0_sequential_7_conv_weight, x = input_29)[name = tensor("input_31")]; tensor conv_out_1_perm_0 = const()[name = tensor("conv_out_1_perm_0"), val = tensor([0, 2, 1])]; - tensor var_315_begin_0 = const()[name = tensor("op_315_begin_0"), val = tensor([0, -1, 0])]; - tensor var_315_end_0 = const()[name = tensor("op_315_end_0"), val = tensor([5, 16, 256])]; - tensor var_315_end_mask_0 = const()[name = tensor("op_315_end_mask_0"), val = tensor([true, true, true])]; - tensor conv_out_1 = transpose(perm = conv_out_1_perm_0, x = input_29)[name = tensor("transpose_53")]; - tensor var_315 = slice_by_index(begin = var_315_begin_0, end = var_315_end_0, end_mask = var_315_end_mask_0, x = conv_out_1)[name = tensor("op_315")]; - tensor var_317_perm_0 = const()[name = tensor("op_317_perm_0"), val = tensor([1, 0, 2])]; - tensor var_317 = transpose(perm = var_317_perm_0, x = var_315)[name = tensor("transpose_52")]; - tensor input_31 = add(x = x_3, y = var_317)[name = tensor("input_31")]; - tensor input_33_axes_0 = const()[name = tensor("input_33_axes_0"), val = tensor([-1])]; - tensor input_33 = layer_norm(axes = input_33_axes_0, beta = encoder_ffn2_0_module_sequential_0_bias, epsilon = var_29, gamma = encoder_ffn2_0_module_sequential_0_weight, x = input_31)[name = tensor("input_33")]; - tensor inputs_9 = linear(bias = encoder_ffn2_0_module_sequential_1_linear_bias, weight = encoder_ffn2_0_module_sequential_1_linear_weight, x = input_33)[name = tensor("linear_8")]; - tensor input_35 = silu(x = inputs_9)[name = tensor("input_35")]; - tensor input_39 = linear(bias = encoder_ffn2_0_module_sequential_4_linear_bias, weight = encoder_ffn2_0_module_sequential_4_linear_weight, x = input_35)[name = tensor("linear_9")]; - tensor var_340 = const()[name = tensor("op_340"), val = tensor(0x1p-1)]; - tensor var_341 = mul(x = input_39, y = var_340)[name = tensor("op_341")]; - tensor input_41 = add(x = var_341, y = input_31)[name = tensor("input_41")]; - tensor input_43_axes_0 = const()[name = tensor("input_43_axes_0"), val = tensor([-1])]; - tensor input_43 = layer_norm(axes = input_43_axes_0, beta = encoder_layer_norm_0_bias, epsilon = var_29, gamma = encoder_layer_norm_0_weight, x = input_41)[name = tensor("input_43")]; + tensor var_381_begin_0 = const()[name = tensor("op_381_begin_0"), val = tensor([0, -1, 0])]; + tensor var_381_end_0 = const()[name = tensor("op_381_end_0"), val = tensor([5, 16, 256])]; + tensor var_381_end_mask_0 = const()[name = tensor("op_381_end_mask_0"), val = tensor([true, true, true])]; + tensor conv_out_1 = transpose(perm = conv_out_1_perm_0, x = input_31)[name = tensor("transpose_53")]; + tensor var_381 = slice_by_index(begin = var_381_begin_0, end = var_381_end_0, end_mask = var_381_end_mask_0, x = conv_out_1)[name = tensor("op_381")]; + tensor var_383_perm_0 = const()[name = tensor("op_383_perm_0"), val = tensor([1, 0, 2])]; + tensor var_383 = transpose(perm = var_383_perm_0, x = var_381)[name = tensor("transpose_52")]; + tensor input_33 = add(x = x_3, y = var_383)[name = tensor("input_33")]; + tensor input_35_axes_0 = const()[name = tensor("input_35_axes_0"), val = tensor([-1])]; + tensor input_35 = layer_norm(axes = input_35_axes_0, beta = inner_encoder_ffn2_0_module_sequential_0_bias, epsilon = var_76, gamma = inner_encoder_ffn2_0_module_sequential_0_weight, x = input_33)[name = tensor("input_35")]; + tensor inputs_9 = linear(bias = inner_encoder_ffn2_0_module_sequential_1_linear_bias, weight = inner_encoder_ffn2_0_module_sequential_1_linear_weight, x = input_35)[name = tensor("linear_8")]; + tensor input_37 = silu(x = inputs_9)[name = tensor("input_37")]; + tensor input_41 = linear(bias = inner_encoder_ffn2_0_module_sequential_4_linear_bias, weight = inner_encoder_ffn2_0_module_sequential_4_linear_weight, x = input_37)[name = tensor("linear_9")]; + tensor var_406 = const()[name = tensor("op_406"), val = tensor(0x1p-1)]; + tensor var_407 = mul(x = input_41, y = var_406)[name = tensor("op_407")]; + tensor input_43 = add(x = var_407, y = input_33)[name = tensor("input_43")]; tensor input_45_axes_0 = const()[name = tensor("input_45_axes_0"), val = tensor([-1])]; - tensor input_45 = layer_norm(axes = input_45_axes_0, beta = encoder_ffn1_1_module_sequential_0_bias, epsilon = var_29, gamma = encoder_ffn1_1_module_sequential_0_weight, x = input_43)[name = tensor("input_45")]; - tensor inputs_11 = linear(bias = encoder_ffn1_1_module_sequential_1_linear_bias, weight = encoder_ffn1_1_module_sequential_1_linear_weight, x = input_45)[name = tensor("linear_10")]; - tensor input_47 = silu(x = inputs_11)[name = tensor("input_47")]; - tensor input_51 = linear(bias = encoder_ffn1_1_module_sequential_4_linear_bias, weight = encoder_ffn1_1_module_sequential_4_linear_weight, x = input_47)[name = tensor("linear_11")]; - tensor var_370 = const()[name = tensor("op_370"), val = tensor(0x1p-1)]; - tensor var_371 = mul(x = input_51, y = var_370)[name = tensor("op_371")]; - tensor input_53 = add(x = var_371, y = input_43)[name = tensor("input_53")]; + tensor input_45 = layer_norm(axes = input_45_axes_0, beta = inner_encoder_layer_norm_0_bias, epsilon = var_76, gamma = inner_encoder_layer_norm_0_weight, x = input_43)[name = tensor("input_45")]; + tensor input_47_axes_0 = const()[name = tensor("input_47_axes_0"), val = tensor([-1])]; + tensor input_47 = layer_norm(axes = input_47_axes_0, beta = inner_encoder_ffn1_1_module_sequential_0_bias, epsilon = var_76, gamma = inner_encoder_ffn1_1_module_sequential_0_weight, x = input_45)[name = tensor("input_47")]; + tensor inputs_11 = linear(bias = inner_encoder_ffn1_1_module_sequential_1_linear_bias, weight = inner_encoder_ffn1_1_module_sequential_1_linear_weight, x = input_47)[name = tensor("linear_10")]; + tensor input_49 = silu(x = inputs_11)[name = tensor("input_49")]; + tensor input_53 = linear(bias = inner_encoder_ffn1_1_module_sequential_4_linear_bias, weight = inner_encoder_ffn1_1_module_sequential_4_linear_weight, x = input_49)[name = tensor("linear_11")]; + tensor var_436 = const()[name = tensor("op_436"), val = tensor(0x1p-1)]; + tensor var_437 = mul(x = input_53, y = var_436)[name = tensor("op_437")]; + tensor input_55 = add(x = var_437, y = input_45)[name = tensor("input_55")]; tensor x_7_axes_0 = const()[name = tensor("x_7_axes_0"), val = tensor([-1])]; - tensor x_7 = layer_norm(axes = x_7_axes_0, beta = encoder_ret_lns_1_bias, epsilon = var_29, gamma = encoder_ret_lns_1_weight, x = input_53)[name = tensor("x_7")]; + tensor x_7 = layer_norm(axes = x_7_axes_0, beta = inner_encoder_ret_lns_1_bias, epsilon = var_76, gamma = inner_encoder_ret_lns_1_weight, x = input_55)[name = tensor("x_7")]; tensor prev_kv_3_begin_0 = const()[name = tensor("prev_kv_3_begin_0"), val = tensor([1, 0, 0, 0, 0])]; tensor prev_kv_3_end_0 = const()[name = tensor("prev_kv_3_end_0"), val = tensor([2, 1, 4, 64, 64])]; tensor prev_kv_3_end_mask_0 = const()[name = tensor("prev_kv_3_end_mask_0"), val = tensor([false, true, true, true, true])]; @@ -426,183 +452,183 @@ program(1.0) tensor prev_scale_3_end_mask_0 = const()[name = tensor("prev_scale_3_end_mask_0"), val = tensor([false, true])]; tensor prev_scale_3_squeeze_mask_0 = const()[name = tensor("prev_scale_3_squeeze_mask_0"), val = tensor([true, false])]; tensor prev_scale_3 = slice_by_index(begin = prev_scale_3_begin_0, end = prev_scale_3_end_0, end_mask = prev_scale_3_end_mask_0, squeeze_mask = prev_scale_3_squeeze_mask_0, x = enc_scale)[name = tensor("prev_scale_3")]; - tensor var_385 = linear(bias = encoder_q_proj_1_bias, weight = encoder_q_proj_1_weight, x = x_7)[name = tensor("linear_12")]; - tensor var_386 = const()[name = tensor("op_386"), val = tensor([1, 5, 4, 64])]; - tensor var_387 = reshape(shape = var_386, x = var_385)[name = tensor("op_387")]; + tensor var_451 = linear(bias = inner_encoder_q_proj_1_bias, weight = inner_encoder_q_proj_1_weight, x = x_7)[name = tensor("linear_12")]; + tensor var_452 = const()[name = tensor("op_452"), val = tensor([1, 5, 4, 64])]; + tensor var_453 = reshape(shape = var_452, x = var_451)[name = tensor("op_453")]; tensor q_3_perm_0 = const()[name = tensor("q_3_perm_0"), val = tensor([0, 2, 1, 3])]; - tensor var_391 = linear(bias = encoder_k_proj_1_bias, weight = encoder_k_proj_1_weight, x = x_7)[name = tensor("linear_13")]; - tensor var_392 = const()[name = tensor("op_392"), val = tensor(0x1p-3)]; - tensor var_393 = mul(x = var_391, y = var_392)[name = tensor("op_393")]; - tensor var_394 = const()[name = tensor("op_394"), val = tensor([1, 5, 4, 64])]; - tensor var_395 = reshape(shape = var_394, x = var_393)[name = tensor("op_395")]; + tensor var_457 = linear(bias = inner_encoder_k_proj_1_bias, weight = inner_encoder_k_proj_1_weight, x = x_7)[name = tensor("linear_13")]; + tensor var_458 = const()[name = tensor("op_458"), val = tensor(0x1p-3)]; + tensor var_459 = mul(x = var_457, y = var_458)[name = tensor("op_459")]; + tensor var_460 = const()[name = tensor("op_460"), val = tensor([1, 5, 4, 64])]; + tensor var_461 = reshape(shape = var_460, x = var_459)[name = tensor("op_461")]; tensor k_3_perm_0 = const()[name = tensor("k_3_perm_0"), val = tensor([0, 2, 1, 3])]; - tensor var_399 = linear(bias = encoder_v_proj_1_bias, weight = encoder_v_proj_1_weight, x = x_7)[name = tensor("linear_14")]; - tensor var_400 = const()[name = tensor("op_400"), val = tensor([1, 5, 4, 64])]; - tensor var_401 = reshape(shape = var_400, x = var_399)[name = tensor("op_401")]; + tensor var_465 = linear(bias = inner_encoder_v_proj_1_bias, weight = inner_encoder_v_proj_1_weight, x = x_7)[name = tensor("linear_14")]; + tensor var_466 = const()[name = tensor("op_466"), val = tensor([1, 5, 4, 64])]; + tensor var_467 = reshape(shape = var_466, x = var_465)[name = tensor("op_467")]; tensor v_3_perm_0 = const()[name = tensor("v_3_perm_0"), val = tensor([0, 2, 1, 3])]; - tensor input_57 = linear(bias = encoder_g_proj_1_bias, weight = encoder_g_proj_1_weight, x = x_7)[name = tensor("linear_15")]; + tensor input_59 = linear(bias = inner_encoder_g_proj_1_bias, weight = inner_encoder_g_proj_1_weight, x = x_7)[name = tensor("linear_15")]; tensor sqrt_s0_3 = sqrt(x = prev_scale_3)[name = tensor("sqrt_s0_3")]; - tensor s_t_3 = add(x = prev_scale_3, y = encoder__t_index)[name = tensor("s_t_3")]; + tensor s_t_3 = add(x = prev_scale_3, y = inner_encoder__t_index)[name = tensor("s_t_3")]; tensor sqrt_s_t_3 = sqrt(x = s_t_3)[name = tensor("sqrt_s_t_3")]; tensor qk_3_transpose_x_1 = const()[name = tensor("qk_3_transpose_x_1"), val = tensor(false)]; tensor qk_3_transpose_y_1 = const()[name = tensor("qk_3_transpose_y_1"), val = tensor(true)]; - tensor k_3 = transpose(perm = k_3_perm_0, x = var_395)[name = tensor("transpose_50")]; - tensor q_3 = transpose(perm = q_3_perm_0, x = var_387)[name = tensor("transpose_51")]; + tensor k_3 = transpose(perm = k_3_perm_0, x = var_461)[name = tensor("transpose_50")]; + tensor q_3 = transpose(perm = q_3_perm_0, x = var_453)[name = tensor("transpose_51")]; tensor qk_3 = matmul(transpose_x = qk_3_transpose_x_1, transpose_y = qk_3_transpose_y_1, x = q_3, y = k_3)[name = tensor("qk_3")]; - tensor var_411 = const()[name = tensor("op_411"), val = tensor([5, 1])]; - tensor var_412 = reshape(shape = var_411, x = sqrt_s_t_3)[name = tensor("op_412")]; - tensor M_3 = real_div(x = encoder__causal_mask, y = var_412)[name = tensor("M_3")]; - tensor var_414 = mul(x = qk_3, y = M_3)[name = tensor("op_414")]; + tensor var_477 = const()[name = tensor("op_477"), val = tensor([5, 1])]; + tensor var_478 = reshape(shape = var_477, x = sqrt_s_t_3)[name = tensor("op_478")]; + tensor M_3 = real_div(x = inner_encoder__causal_mask, y = var_478)[name = tensor("M_3")]; + tensor var_480 = mul(x = qk_3, y = M_3)[name = tensor("op_480")]; tensor inner_3_transpose_x_0 = const()[name = tensor("inner_3_transpose_x_0"), val = tensor(false)]; tensor inner_3_transpose_y_0 = const()[name = tensor("inner_3_transpose_y_0"), val = tensor(false)]; - tensor v_3 = transpose(perm = v_3_perm_0, x = var_401)[name = tensor("transpose_49")]; - tensor inner_3 = matmul(transpose_x = inner_3_transpose_x_0, transpose_y = inner_3_transpose_y_0, x = var_414, y = v_3)[name = tensor("inner_3")]; - tensor var_416_transpose_x_0 = const()[name = tensor("op_416_transpose_x_0"), val = tensor(false)]; - tensor var_416_transpose_y_0 = const()[name = tensor("op_416_transpose_y_0"), val = tensor(false)]; - tensor var_416 = matmul(transpose_x = var_416_transpose_x_0, transpose_y = var_416_transpose_y_0, x = q_3, y = prev_kv_3)[name = tensor("op_416")]; - tensor var_417 = real_div(x = sqrt_s0_3, y = sqrt_s_t_3)[name = tensor("op_417")]; - tensor var_418 = const()[name = tensor("op_418"), val = tensor([1, 1, 5, 1])]; - tensor var_419 = reshape(shape = var_418, x = var_417)[name = tensor("op_419")]; - tensor cross_3 = mul(x = var_416, y = var_419)[name = tensor("cross_3")]; + tensor v_3 = transpose(perm = v_3_perm_0, x = var_467)[name = tensor("transpose_49")]; + tensor inner_3 = matmul(transpose_x = inner_3_transpose_x_0, transpose_y = inner_3_transpose_y_0, x = var_480, y = v_3)[name = tensor("inner_3")]; + tensor var_482_transpose_x_0 = const()[name = tensor("op_482_transpose_x_0"), val = tensor(false)]; + tensor var_482_transpose_y_0 = const()[name = tensor("op_482_transpose_y_0"), val = tensor(false)]; + tensor var_482 = matmul(transpose_x = var_482_transpose_x_0, transpose_y = var_482_transpose_y_0, x = q_3, y = prev_kv_3)[name = tensor("op_482")]; + tensor var_483 = real_div(x = sqrt_s0_3, y = sqrt_s_t_3)[name = tensor("op_483")]; + tensor var_484 = const()[name = tensor("op_484"), val = tensor([1, 1, 5, 1])]; + tensor var_485 = reshape(shape = var_484, x = var_483)[name = tensor("op_485")]; + tensor cross_3 = mul(x = var_482, y = var_485)[name = tensor("cross_3")]; tensor out_7 = add(x = inner_3, y = cross_3)[name = tensor("out_7")]; - tensor var_422 = mul(x = prev_kv_3, y = sqrt_s0_3)[name = tensor("op_422")]; - tensor var_424_transpose_x_1 = const()[name = tensor("op_424_transpose_x_1"), val = tensor(true)]; - tensor var_424_transpose_y_1 = const()[name = tensor("op_424_transpose_y_1"), val = tensor(false)]; - tensor var_424 = matmul(transpose_x = var_424_transpose_x_1, transpose_y = var_424_transpose_y_1, x = k_3, y = v_3)[name = tensor("op_424")]; - tensor new_kv_unnorm_3 = add(x = var_422, y = var_424)[name = tensor("new_kv_unnorm_3")]; - tensor var_426 = const()[name = tensor("op_426"), val = tensor(0x1.4p+2)]; - tensor new_scale_3 = add(x = prev_scale_3, y = var_426)[name = tensor("new_scale_3")]; - tensor var_428 = sqrt(x = new_scale_3)[name = tensor("op_428")]; - tensor var_429 = real_div(x = new_kv_unnorm_3, y = var_428)[name = tensor("op_429")]; - tensor var_430_perm_0 = const()[name = tensor("op_430_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_488 = mul(x = prev_kv_3, y = sqrt_s0_3)[name = tensor("op_488")]; + tensor var_490_transpose_x_1 = const()[name = tensor("op_490_transpose_x_1"), val = tensor(true)]; + tensor var_490_transpose_y_1 = const()[name = tensor("op_490_transpose_y_1"), val = tensor(false)]; + tensor var_490 = matmul(transpose_x = var_490_transpose_x_1, transpose_y = var_490_transpose_y_1, x = k_3, y = v_3)[name = tensor("op_490")]; + tensor new_kv_unnorm_3 = add(x = var_488, y = var_490)[name = tensor("new_kv_unnorm_3")]; + tensor var_492 = const()[name = tensor("op_492"), val = tensor(0x1.4p+2)]; + tensor new_scale_3 = add(x = prev_scale_3, y = var_492)[name = tensor("new_scale_3")]; + tensor var_494 = sqrt(x = new_scale_3)[name = tensor("op_494")]; + tensor var_495 = real_div(x = new_kv_unnorm_3, y = var_494)[name = tensor("op_495")]; + tensor var_496_perm_0 = const()[name = tensor("op_496_perm_0"), val = tensor([0, 2, 1, 3])]; tensor out_9_axes_0 = const()[name = tensor("out_9_axes_0"), val = tensor([-1])]; - tensor var_430 = transpose(perm = var_430_perm_0, x = out_7)[name = tensor("transpose_48")]; - tensor out_9 = layer_norm(axes = out_9_axes_0, epsilon = var_18, x = var_430)[name = tensor("out_9")]; - tensor var_434 = const()[name = tensor("op_434"), val = tensor([1, 5, 256])]; - tensor out_11 = reshape(shape = var_434, x = out_9)[name = tensor("out_11")]; - tensor var_436 = silu(x = input_57)[name = tensor("op_436")]; - tensor input_59 = mul(x = var_436, y = out_11)[name = tensor("input_59")]; - tensor ret_out_3 = linear(bias = encoder_out_proj_1_bias, weight = encoder_out_proj_1_weight, x = input_59)[name = tensor("linear_16")]; - tensor x_9 = add(x = input_53, y = ret_out_3)[name = tensor("x_9")]; + tensor var_496 = transpose(perm = var_496_perm_0, x = out_7)[name = tensor("transpose_48")]; + tensor out_9 = layer_norm(axes = out_9_axes_0, epsilon = var_84, x = var_496)[name = tensor("out_9")]; + tensor var_500 = const()[name = tensor("op_500"), val = tensor([1, 5, 256])]; + tensor out_11 = reshape(shape = var_500, x = out_9)[name = tensor("out_11")]; + tensor var_502 = silu(x = input_59)[name = tensor("op_502")]; + tensor input_61 = mul(x = var_502, y = out_11)[name = tensor("input_61")]; + tensor ret_out_3 = linear(bias = inner_encoder_out_proj_1_bias, weight = inner_encoder_out_proj_1_weight, x = input_61)[name = tensor("linear_16")]; + tensor x_9 = add(x = input_55, y = ret_out_3)[name = tensor("x_9")]; tensor window_13_begin_0 = const()[name = tensor("window_13_begin_0"), val = tensor([1, 0, 0, 0])]; tensor window_13_end_0 = const()[name = tensor("window_13_end_0"), val = tensor([2, 1, 16, 256])]; tensor window_13_end_mask_0 = const()[name = tensor("window_13_end_mask_0"), val = tensor([false, true, true, true])]; tensor window_13_squeeze_mask_0 = const()[name = tensor("window_13_squeeze_mask_0"), val = tensor([true, false, false, false])]; tensor window_13 = slice_by_index(begin = window_13_begin_0, end = window_13_end_0, end_mask = window_13_end_mask_0, squeeze_mask = window_13_squeeze_mask_0, x = enc_conv_cache)[name = tensor("window_13")]; - tensor var_444_begin_0 = const()[name = tensor("op_444_begin_0"), val = tensor([0, 0, 0])]; - tensor var_444_end_0 = const()[name = tensor("op_444_end_0"), val = tensor([1, 1, 256])]; - tensor var_444_end_mask_0 = const()[name = tensor("op_444_end_mask_0"), val = tensor([true, false, true])]; - tensor var_444 = slice_by_index(begin = var_444_begin_0, end = var_444_end_0, end_mask = var_444_end_mask_0, x = x_9)[name = tensor("op_444")]; - tensor var_447_begin_0 = const()[name = tensor("op_447_begin_0"), val = tensor([0, 1, 0])]; - tensor var_447_end_0 = const()[name = tensor("op_447_end_0"), val = tensor([1, 16, 256])]; - tensor var_447_end_mask_0 = const()[name = tensor("op_447_end_mask_0"), val = tensor([true, true, true])]; - tensor var_447 = slice_by_index(begin = var_447_begin_0, end = var_447_end_0, end_mask = var_447_end_mask_0, x = window_13)[name = tensor("op_447")]; + tensor var_510_begin_0 = const()[name = tensor("op_510_begin_0"), val = tensor([0, 0, 0])]; + tensor var_510_end_0 = const()[name = tensor("op_510_end_0"), val = tensor([1, 1, 256])]; + tensor var_510_end_mask_0 = const()[name = tensor("op_510_end_mask_0"), val = tensor([true, false, true])]; + tensor var_510 = slice_by_index(begin = var_510_begin_0, end = var_510_end_0, end_mask = var_510_end_mask_0, x = x_9)[name = tensor("op_510")]; + tensor var_513_begin_0 = const()[name = tensor("op_513_begin_0"), val = tensor([0, 1, 0])]; + tensor var_513_end_0 = const()[name = tensor("op_513_end_0"), val = tensor([1, 16, 256])]; + tensor var_513_end_mask_0 = const()[name = tensor("op_513_end_mask_0"), val = tensor([true, true, true])]; + tensor var_513 = slice_by_index(begin = var_513_begin_0, end = var_513_end_0, end_mask = var_513_end_mask_0, x = window_13)[name = tensor("op_513")]; tensor window_15_interleave_0 = const()[name = tensor("window_15_interleave_0"), val = tensor(false)]; - tensor window_15 = concat(axis = var_27, interleave = window_15_interleave_0, values = (var_447, var_444))[name = tensor("window_15")]; - tensor var_452_begin_0 = const()[name = tensor("op_452_begin_0"), val = tensor([0, 1, 0])]; - tensor var_452_end_0 = const()[name = tensor("op_452_end_0"), val = tensor([1, 2, 256])]; - tensor var_452_end_mask_0 = const()[name = tensor("op_452_end_mask_0"), val = tensor([true, false, true])]; - tensor var_452 = slice_by_index(begin = var_452_begin_0, end = var_452_end_0, end_mask = var_452_end_mask_0, x = x_9)[name = tensor("op_452")]; - tensor var_455_begin_0 = const()[name = tensor("op_455_begin_0"), val = tensor([0, 1, 0])]; - tensor var_455_end_0 = const()[name = tensor("op_455_end_0"), val = tensor([1, 16, 256])]; - tensor var_455_end_mask_0 = const()[name = tensor("op_455_end_mask_0"), val = tensor([true, true, true])]; - tensor var_455 = slice_by_index(begin = var_455_begin_0, end = var_455_end_0, end_mask = var_455_end_mask_0, x = window_15)[name = tensor("op_455")]; + tensor window_15 = concat(axis = var_93, interleave = window_15_interleave_0, values = (var_513, var_510))[name = tensor("window_15")]; + tensor var_518_begin_0 = const()[name = tensor("op_518_begin_0"), val = tensor([0, 1, 0])]; + tensor var_518_end_0 = const()[name = tensor("op_518_end_0"), val = tensor([1, 2, 256])]; + tensor var_518_end_mask_0 = const()[name = tensor("op_518_end_mask_0"), val = tensor([true, false, true])]; + tensor var_518 = slice_by_index(begin = var_518_begin_0, end = var_518_end_0, end_mask = var_518_end_mask_0, x = x_9)[name = tensor("op_518")]; + tensor var_521_begin_0 = const()[name = tensor("op_521_begin_0"), val = tensor([0, 1, 0])]; + tensor var_521_end_0 = const()[name = tensor("op_521_end_0"), val = tensor([1, 16, 256])]; + tensor var_521_end_mask_0 = const()[name = tensor("op_521_end_mask_0"), val = tensor([true, true, true])]; + tensor var_521 = slice_by_index(begin = var_521_begin_0, end = var_521_end_0, end_mask = var_521_end_mask_0, x = window_15)[name = tensor("op_521")]; tensor window_17_interleave_0 = const()[name = tensor("window_17_interleave_0"), val = tensor(false)]; - tensor window_17 = concat(axis = var_27, interleave = window_17_interleave_0, values = (var_455, var_452))[name = tensor("window_17")]; - tensor var_460_begin_0 = const()[name = tensor("op_460_begin_0"), val = tensor([0, 2, 0])]; - tensor var_460_end_0 = const()[name = tensor("op_460_end_0"), val = tensor([1, 3, 256])]; - tensor var_460_end_mask_0 = const()[name = tensor("op_460_end_mask_0"), val = tensor([true, false, true])]; - tensor var_460 = slice_by_index(begin = var_460_begin_0, end = var_460_end_0, end_mask = var_460_end_mask_0, x = x_9)[name = tensor("op_460")]; - tensor var_463_begin_0 = const()[name = tensor("op_463_begin_0"), val = tensor([0, 1, 0])]; - tensor var_463_end_0 = const()[name = tensor("op_463_end_0"), val = tensor([1, 16, 256])]; - tensor var_463_end_mask_0 = const()[name = tensor("op_463_end_mask_0"), val = tensor([true, true, true])]; - tensor var_463 = slice_by_index(begin = var_463_begin_0, end = var_463_end_0, end_mask = var_463_end_mask_0, x = window_17)[name = tensor("op_463")]; + tensor window_17 = concat(axis = var_93, interleave = window_17_interleave_0, values = (var_521, var_518))[name = tensor("window_17")]; + tensor var_526_begin_0 = const()[name = tensor("op_526_begin_0"), val = tensor([0, 2, 0])]; + tensor var_526_end_0 = const()[name = tensor("op_526_end_0"), val = tensor([1, 3, 256])]; + tensor var_526_end_mask_0 = const()[name = tensor("op_526_end_mask_0"), val = tensor([true, false, true])]; + tensor var_526 = slice_by_index(begin = var_526_begin_0, end = var_526_end_0, end_mask = var_526_end_mask_0, x = x_9)[name = tensor("op_526")]; + tensor var_529_begin_0 = const()[name = tensor("op_529_begin_0"), val = tensor([0, 1, 0])]; + tensor var_529_end_0 = const()[name = tensor("op_529_end_0"), val = tensor([1, 16, 256])]; + tensor var_529_end_mask_0 = const()[name = tensor("op_529_end_mask_0"), val = tensor([true, true, true])]; + tensor var_529 = slice_by_index(begin = var_529_begin_0, end = var_529_end_0, end_mask = var_529_end_mask_0, x = window_17)[name = tensor("op_529")]; tensor window_19_interleave_0 = const()[name = tensor("window_19_interleave_0"), val = tensor(false)]; - tensor window_19 = concat(axis = var_27, interleave = window_19_interleave_0, values = (var_463, var_460))[name = tensor("window_19")]; - tensor var_468_begin_0 = const()[name = tensor("op_468_begin_0"), val = tensor([0, 3, 0])]; - tensor var_468_end_0 = const()[name = tensor("op_468_end_0"), val = tensor([1, 4, 256])]; - tensor var_468_end_mask_0 = const()[name = tensor("op_468_end_mask_0"), val = tensor([true, false, true])]; - tensor var_468 = slice_by_index(begin = var_468_begin_0, end = var_468_end_0, end_mask = var_468_end_mask_0, x = x_9)[name = tensor("op_468")]; - tensor var_471_begin_0 = const()[name = tensor("op_471_begin_0"), val = tensor([0, 1, 0])]; - tensor var_471_end_0 = const()[name = tensor("op_471_end_0"), val = tensor([1, 16, 256])]; - tensor var_471_end_mask_0 = const()[name = tensor("op_471_end_mask_0"), val = tensor([true, true, true])]; - tensor var_471 = slice_by_index(begin = var_471_begin_0, end = var_471_end_0, end_mask = var_471_end_mask_0, x = window_19)[name = tensor("op_471")]; + tensor window_19 = concat(axis = var_93, interleave = window_19_interleave_0, values = (var_529, var_526))[name = tensor("window_19")]; + tensor var_534_begin_0 = const()[name = tensor("op_534_begin_0"), val = tensor([0, 3, 0])]; + tensor var_534_end_0 = const()[name = tensor("op_534_end_0"), val = tensor([1, 4, 256])]; + tensor var_534_end_mask_0 = const()[name = tensor("op_534_end_mask_0"), val = tensor([true, false, true])]; + tensor var_534 = slice_by_index(begin = var_534_begin_0, end = var_534_end_0, end_mask = var_534_end_mask_0, x = x_9)[name = tensor("op_534")]; + tensor var_537_begin_0 = const()[name = tensor("op_537_begin_0"), val = tensor([0, 1, 0])]; + tensor var_537_end_0 = const()[name = tensor("op_537_end_0"), val = tensor([1, 16, 256])]; + tensor var_537_end_mask_0 = const()[name = tensor("op_537_end_mask_0"), val = tensor([true, true, true])]; + tensor var_537 = slice_by_index(begin = var_537_begin_0, end = var_537_end_0, end_mask = var_537_end_mask_0, x = window_19)[name = tensor("op_537")]; tensor window_21_interleave_0 = const()[name = tensor("window_21_interleave_0"), val = tensor(false)]; - tensor window_21 = concat(axis = var_27, interleave = window_21_interleave_0, values = (var_471, var_468))[name = tensor("window_21")]; - tensor var_476_begin_0 = const()[name = tensor("op_476_begin_0"), val = tensor([0, 4, 0])]; - tensor var_476_end_0 = const()[name = tensor("op_476_end_0"), val = tensor([1, 1, 256])]; - tensor var_476_end_mask_0 = const()[name = tensor("op_476_end_mask_0"), val = tensor([true, true, true])]; - tensor var_476 = slice_by_index(begin = var_476_begin_0, end = var_476_end_0, end_mask = var_476_end_mask_0, x = x_9)[name = tensor("op_476")]; - tensor var_479_begin_0 = const()[name = tensor("op_479_begin_0"), val = tensor([0, 1, 0])]; - tensor var_479_end_0 = const()[name = tensor("op_479_end_0"), val = tensor([1, 16, 256])]; - tensor var_479_end_mask_0 = const()[name = tensor("op_479_end_mask_0"), val = tensor([true, true, true])]; - tensor var_479 = slice_by_index(begin = var_479_begin_0, end = var_479_end_0, end_mask = var_479_end_mask_0, x = window_21)[name = tensor("op_479")]; + tensor window_21 = concat(axis = var_93, interleave = window_21_interleave_0, values = (var_537, var_534))[name = tensor("window_21")]; + tensor var_542_begin_0 = const()[name = tensor("op_542_begin_0"), val = tensor([0, 4, 0])]; + tensor var_542_end_0 = const()[name = tensor("op_542_end_0"), val = tensor([1, 1, 256])]; + tensor var_542_end_mask_0 = const()[name = tensor("op_542_end_mask_0"), val = tensor([true, true, true])]; + tensor var_542 = slice_by_index(begin = var_542_begin_0, end = var_542_end_0, end_mask = var_542_end_mask_0, x = x_9)[name = tensor("op_542")]; + tensor var_545_begin_0 = const()[name = tensor("op_545_begin_0"), val = tensor([0, 1, 0])]; + tensor var_545_end_0 = const()[name = tensor("op_545_end_0"), val = tensor([1, 16, 256])]; + tensor var_545_end_mask_0 = const()[name = tensor("op_545_end_mask_0"), val = tensor([true, true, true])]; + tensor var_545 = slice_by_index(begin = var_545_begin_0, end = var_545_end_0, end_mask = var_545_end_mask_0, x = window_21)[name = tensor("op_545")]; tensor window_23_interleave_0 = const()[name = tensor("window_23_interleave_0"), val = tensor(false)]; - tensor window_23 = concat(axis = var_27, interleave = window_23_interleave_0, values = (var_479, var_476))[name = tensor("window_23")]; - tensor input_61_interleave_0 = const()[name = tensor("input_61_interleave_0"), val = tensor(false)]; - tensor input_61 = concat(axis = var_24, interleave = input_61_interleave_0, values = (window_15, window_17, window_19, window_21, window_23))[name = tensor("input_61")]; + tensor window_23 = concat(axis = var_93, interleave = window_23_interleave_0, values = (var_545, var_542))[name = tensor("window_23")]; + tensor input_63_interleave_0 = const()[name = tensor("input_63_interleave_0"), val = tensor(false)]; + tensor input_63 = concat(axis = var_79, interleave = input_63_interleave_0, values = (window_15, window_17, window_19, window_21, window_23))[name = tensor("input_63")]; tensor x_11_axes_0 = const()[name = tensor("x_11_axes_0"), val = tensor([-1])]; - tensor x_11 = layer_norm(axes = x_11_axes_0, beta = encoder_conv_module_1_sequential_0_bias, epsilon = var_29, gamma = encoder_conv_module_1_sequential_0_weight, x = input_61)[name = tensor("x_11")]; - tensor input_63_perm_0 = const()[name = tensor("input_63_perm_0"), val = tensor([0, 2, 1])]; + tensor x_11 = layer_norm(axes = x_11_axes_0, beta = inner_encoder_conv_module_1_sequential_0_bias, epsilon = var_76, gamma = inner_encoder_conv_module_1_sequential_0_weight, x = input_63)[name = tensor("x_11")]; + tensor input_65_perm_0 = const()[name = tensor("input_65_perm_0"), val = tensor([0, 2, 1])]; tensor inputs_13_pad_type_0 = const()[name = tensor("inputs_13_pad_type_0"), val = tensor("valid")]; tensor inputs_13_strides_0 = const()[name = tensor("inputs_13_strides_0"), val = tensor([1])]; tensor inputs_13_pad_0 = const()[name = tensor("inputs_13_pad_0"), val = tensor([0, 0])]; tensor inputs_13_dilations_0 = const()[name = tensor("inputs_13_dilations_0"), val = tensor([1])]; tensor inputs_13_groups_0 = const()[name = tensor("inputs_13_groups_0"), val = tensor(1)]; - tensor input_63 = transpose(perm = input_63_perm_0, x = x_11)[name = tensor("transpose_47")]; - tensor inputs_13 = conv(bias = encoder_conv_module_1_sequential_2_conv_bias, dilations = inputs_13_dilations_0, groups = inputs_13_groups_0, pad = inputs_13_pad_0, pad_type = inputs_13_pad_type_0, strides = inputs_13_strides_0, weight = encoder_conv_module_1_sequential_2_conv_weight, x = input_63)[name = tensor("inputs_13")]; - tensor var_504_split_sizes_0 = const()[name = tensor("op_504_split_sizes_0"), val = tensor([256, 256])]; - tensor var_504_axis_0 = const()[name = tensor("op_504_axis_0"), val = tensor(1)]; - tensor var_504_0, tensor var_504_1 = split(axis = var_504_axis_0, split_sizes = var_504_split_sizes_0, x = inputs_13)[name = tensor("op_504")]; - tensor var_506 = sigmoid(x = var_504_1)[name = tensor("op_506")]; - tensor inputs_15 = mul(x = var_504_0, y = var_506)[name = tensor("inputs_15")]; + tensor input_65 = transpose(perm = input_65_perm_0, x = x_11)[name = tensor("transpose_47")]; + tensor inputs_13 = conv(bias = inner_encoder_conv_module_1_sequential_2_conv_bias, dilations = inputs_13_dilations_0, groups = inputs_13_groups_0, pad = inputs_13_pad_0, pad_type = inputs_13_pad_type_0, strides = inputs_13_strides_0, weight = inner_encoder_conv_module_1_sequential_2_conv_weight, x = input_65)[name = tensor("inputs_13")]; + tensor var_570_split_sizes_0 = const()[name = tensor("op_570_split_sizes_0"), val = tensor([256, 256])]; + tensor var_570_axis_0 = const()[name = tensor("op_570_axis_0"), val = tensor(1)]; + tensor var_570_0, tensor var_570_1 = split(axis = var_570_axis_0, split_sizes = var_570_split_sizes_0, x = inputs_13)[name = tensor("op_570")]; + tensor var_572 = sigmoid(x = var_570_1)[name = tensor("op_572")]; + tensor inputs_15 = mul(x = var_570_0, y = var_572)[name = tensor("inputs_15")]; tensor outputs_aug_3_pad_type_0 = const()[name = tensor("outputs_aug_3_pad_type_0"), val = tensor("custom")]; tensor outputs_aug_3_pad_0 = const()[name = tensor("outputs_aug_3_pad_0"), val = tensor([15, 15])]; tensor outputs_aug_3_groups_0 = const()[name = tensor("outputs_aug_3_groups_0"), val = tensor(256)]; tensor outputs_aug_3_strides_0 = const()[name = tensor("outputs_aug_3_strides_0"), val = tensor([1])]; tensor outputs_aug_3_dilations_0 = const()[name = tensor("outputs_aug_3_dilations_0"), val = tensor([1])]; - tensor outputs_aug_3 = conv(dilations = outputs_aug_3_dilations_0, groups = outputs_aug_3_groups_0, pad = outputs_aug_3_pad_0, pad_type = outputs_aug_3_pad_type_0, strides = outputs_aug_3_strides_0, weight = encoder_conv_module_1_sequential_4_conv_weight, x = inputs_15)[name = tensor("outputs_aug_3")]; - tensor input_65_begin_0 = const()[name = tensor("input_65_begin_0"), val = tensor([0, 0, 0])]; - tensor input_65_end_0 = const()[name = tensor("input_65_end_0"), val = tensor([5, 256, 16])]; - tensor input_65_end_mask_0 = const()[name = tensor("input_65_end_mask_0"), val = tensor([true, true, false])]; - tensor input_65 = slice_by_index(begin = input_65_begin_0, end = input_65_end_0, end_mask = input_65_end_mask_0, x = outputs_aug_3)[name = tensor("input_65")]; - tensor inputs_17 = batch_norm(beta = encoder_conv_module_1_sequential_5_bias, epsilon = var_29, gamma = encoder_conv_module_1_sequential_5_weight, mean = encoder_conv_module_1_sequential_5_running_mean, variance = encoder_conv_module_1_sequential_5_running_var, x = input_65)[name = tensor("inputs_17")]; - tensor input_67 = silu(x = inputs_17)[name = tensor("input_67")]; - tensor input_69_pad_type_0 = const()[name = tensor("input_69_pad_type_0"), val = tensor("valid")]; - tensor input_69_strides_0 = const()[name = tensor("input_69_strides_0"), val = tensor([1])]; - tensor input_69_pad_0 = const()[name = tensor("input_69_pad_0"), val = tensor([0, 0])]; - tensor input_69_dilations_0 = const()[name = tensor("input_69_dilations_0"), val = tensor([1])]; - tensor input_69_groups_0 = const()[name = tensor("input_69_groups_0"), val = tensor(1)]; - tensor input_69 = conv(bias = encoder_conv_module_1_sequential_7_conv_bias, dilations = input_69_dilations_0, groups = input_69_groups_0, pad = input_69_pad_0, pad_type = input_69_pad_type_0, strides = input_69_strides_0, weight = encoder_conv_module_1_sequential_7_conv_weight, x = input_67)[name = tensor("input_69")]; + tensor outputs_aug_3 = conv(dilations = outputs_aug_3_dilations_0, groups = outputs_aug_3_groups_0, pad = outputs_aug_3_pad_0, pad_type = outputs_aug_3_pad_type_0, strides = outputs_aug_3_strides_0, weight = inner_encoder_conv_module_1_sequential_4_conv_weight, x = inputs_15)[name = tensor("outputs_aug_3")]; + tensor input_67_begin_0 = const()[name = tensor("input_67_begin_0"), val = tensor([0, 0, 0])]; + tensor input_67_end_0 = const()[name = tensor("input_67_end_0"), val = tensor([5, 256, 16])]; + tensor input_67_end_mask_0 = const()[name = tensor("input_67_end_mask_0"), val = tensor([true, true, false])]; + tensor input_67 = slice_by_index(begin = input_67_begin_0, end = input_67_end_0, end_mask = input_67_end_mask_0, x = outputs_aug_3)[name = tensor("input_67")]; + tensor inputs_17 = batch_norm(beta = inner_encoder_conv_module_1_sequential_5_bias, epsilon = var_76, gamma = inner_encoder_conv_module_1_sequential_5_weight, mean = inner_encoder_conv_module_1_sequential_5_running_mean, variance = inner_encoder_conv_module_1_sequential_5_running_var, x = input_67)[name = tensor("inputs_17")]; + tensor input_69 = silu(x = inputs_17)[name = tensor("input_69")]; + tensor input_71_pad_type_0 = const()[name = tensor("input_71_pad_type_0"), val = tensor("valid")]; + tensor input_71_strides_0 = const()[name = tensor("input_71_strides_0"), val = tensor([1])]; + tensor input_71_pad_0 = const()[name = tensor("input_71_pad_0"), val = tensor([0, 0])]; + tensor input_71_dilations_0 = const()[name = tensor("input_71_dilations_0"), val = tensor([1])]; + tensor input_71_groups_0 = const()[name = tensor("input_71_groups_0"), val = tensor(1)]; + tensor input_71 = conv(bias = inner_encoder_conv_module_1_sequential_7_conv_bias, dilations = input_71_dilations_0, groups = input_71_groups_0, pad = input_71_pad_0, pad_type = input_71_pad_type_0, strides = input_71_strides_0, weight = inner_encoder_conv_module_1_sequential_7_conv_weight, x = input_69)[name = tensor("input_71")]; tensor conv_out_3_perm_0 = const()[name = tensor("conv_out_3_perm_0"), val = tensor([0, 2, 1])]; - tensor var_537_begin_0 = const()[name = tensor("op_537_begin_0"), val = tensor([0, -1, 0])]; - tensor var_537_end_0 = const()[name = tensor("op_537_end_0"), val = tensor([5, 16, 256])]; - tensor var_537_end_mask_0 = const()[name = tensor("op_537_end_mask_0"), val = tensor([true, true, true])]; - tensor conv_out_3 = transpose(perm = conv_out_3_perm_0, x = input_69)[name = tensor("transpose_46")]; - tensor var_537 = slice_by_index(begin = var_537_begin_0, end = var_537_end_0, end_mask = var_537_end_mask_0, x = conv_out_3)[name = tensor("op_537")]; - tensor var_539_perm_0 = const()[name = tensor("op_539_perm_0"), val = tensor([1, 0, 2])]; - tensor var_539 = transpose(perm = var_539_perm_0, x = var_537)[name = tensor("transpose_45")]; - tensor input_71 = add(x = x_9, y = var_539)[name = tensor("input_71")]; - tensor input_73_axes_0 = const()[name = tensor("input_73_axes_0"), val = tensor([-1])]; - tensor input_73 = layer_norm(axes = input_73_axes_0, beta = encoder_ffn2_1_module_sequential_0_bias, epsilon = var_29, gamma = encoder_ffn2_1_module_sequential_0_weight, x = input_71)[name = tensor("input_73")]; - tensor inputs_19 = linear(bias = encoder_ffn2_1_module_sequential_1_linear_bias, weight = encoder_ffn2_1_module_sequential_1_linear_weight, x = input_73)[name = tensor("linear_17")]; - tensor input_75 = silu(x = inputs_19)[name = tensor("input_75")]; - tensor input_79 = linear(bias = encoder_ffn2_1_module_sequential_4_linear_bias, weight = encoder_ffn2_1_module_sequential_4_linear_weight, x = input_75)[name = tensor("linear_18")]; - tensor var_562 = const()[name = tensor("op_562"), val = tensor(0x1p-1)]; - tensor var_563 = mul(x = input_79, y = var_562)[name = tensor("op_563")]; - tensor input_81 = add(x = var_563, y = input_71)[name = tensor("input_81")]; - tensor input_83_axes_0 = const()[name = tensor("input_83_axes_0"), val = tensor([-1])]; - tensor input_83 = layer_norm(axes = input_83_axes_0, beta = encoder_layer_norm_1_bias, epsilon = var_29, gamma = encoder_layer_norm_1_weight, x = input_81)[name = tensor("input_83")]; + tensor var_603_begin_0 = const()[name = tensor("op_603_begin_0"), val = tensor([0, -1, 0])]; + tensor var_603_end_0 = const()[name = tensor("op_603_end_0"), val = tensor([5, 16, 256])]; + tensor var_603_end_mask_0 = const()[name = tensor("op_603_end_mask_0"), val = tensor([true, true, true])]; + tensor conv_out_3 = transpose(perm = conv_out_3_perm_0, x = input_71)[name = tensor("transpose_46")]; + tensor var_603 = slice_by_index(begin = var_603_begin_0, end = var_603_end_0, end_mask = var_603_end_mask_0, x = conv_out_3)[name = tensor("op_603")]; + tensor var_605_perm_0 = const()[name = tensor("op_605_perm_0"), val = tensor([1, 0, 2])]; + tensor var_605 = transpose(perm = var_605_perm_0, x = var_603)[name = tensor("transpose_45")]; + tensor input_73 = add(x = x_9, y = var_605)[name = tensor("input_73")]; + tensor input_75_axes_0 = const()[name = tensor("input_75_axes_0"), val = tensor([-1])]; + tensor input_75 = layer_norm(axes = input_75_axes_0, beta = inner_encoder_ffn2_1_module_sequential_0_bias, epsilon = var_76, gamma = inner_encoder_ffn2_1_module_sequential_0_weight, x = input_73)[name = tensor("input_75")]; + tensor inputs_19 = linear(bias = inner_encoder_ffn2_1_module_sequential_1_linear_bias, weight = inner_encoder_ffn2_1_module_sequential_1_linear_weight, x = input_75)[name = tensor("linear_17")]; + tensor input_77 = silu(x = inputs_19)[name = tensor("input_77")]; + tensor input_81 = linear(bias = inner_encoder_ffn2_1_module_sequential_4_linear_bias, weight = inner_encoder_ffn2_1_module_sequential_4_linear_weight, x = input_77)[name = tensor("linear_18")]; + tensor var_628 = const()[name = tensor("op_628"), val = tensor(0x1p-1)]; + tensor var_629 = mul(x = input_81, y = var_628)[name = tensor("op_629")]; + tensor input_83 = add(x = var_629, y = input_73)[name = tensor("input_83")]; tensor input_85_axes_0 = const()[name = tensor("input_85_axes_0"), val = tensor([-1])]; - tensor input_85 = layer_norm(axes = input_85_axes_0, beta = encoder_ffn1_2_module_sequential_0_bias, epsilon = var_29, gamma = encoder_ffn1_2_module_sequential_0_weight, x = input_83)[name = tensor("input_85")]; - tensor inputs_21 = linear(bias = encoder_ffn1_2_module_sequential_1_linear_bias, weight = encoder_ffn1_2_module_sequential_1_linear_weight, x = input_85)[name = tensor("linear_19")]; - tensor input_87 = silu(x = inputs_21)[name = tensor("input_87")]; - tensor input_91 = linear(bias = encoder_ffn1_2_module_sequential_4_linear_bias, weight = encoder_ffn1_2_module_sequential_4_linear_weight, x = input_87)[name = tensor("linear_20")]; - tensor var_592 = const()[name = tensor("op_592"), val = tensor(0x1p-1)]; - tensor var_593 = mul(x = input_91, y = var_592)[name = tensor("op_593")]; - tensor input_93 = add(x = var_593, y = input_83)[name = tensor("input_93")]; + tensor input_85 = layer_norm(axes = input_85_axes_0, beta = inner_encoder_layer_norm_1_bias, epsilon = var_76, gamma = inner_encoder_layer_norm_1_weight, x = input_83)[name = tensor("input_85")]; + tensor input_87_axes_0 = const()[name = tensor("input_87_axes_0"), val = tensor([-1])]; + tensor input_87 = layer_norm(axes = input_87_axes_0, beta = inner_encoder_ffn1_2_module_sequential_0_bias, epsilon = var_76, gamma = inner_encoder_ffn1_2_module_sequential_0_weight, x = input_85)[name = tensor("input_87")]; + tensor inputs_21 = linear(bias = inner_encoder_ffn1_2_module_sequential_1_linear_bias, weight = inner_encoder_ffn1_2_module_sequential_1_linear_weight, x = input_87)[name = tensor("linear_19")]; + tensor input_89 = silu(x = inputs_21)[name = tensor("input_89")]; + tensor input_93 = linear(bias = inner_encoder_ffn1_2_module_sequential_4_linear_bias, weight = inner_encoder_ffn1_2_module_sequential_4_linear_weight, x = input_89)[name = tensor("linear_20")]; + tensor var_658 = const()[name = tensor("op_658"), val = tensor(0x1p-1)]; + tensor var_659 = mul(x = input_93, y = var_658)[name = tensor("op_659")]; + tensor input_95 = add(x = var_659, y = input_85)[name = tensor("input_95")]; tensor x_13_axes_0 = const()[name = tensor("x_13_axes_0"), val = tensor([-1])]; - tensor x_13 = layer_norm(axes = x_13_axes_0, beta = encoder_ret_lns_2_bias, epsilon = var_29, gamma = encoder_ret_lns_2_weight, x = input_93)[name = tensor("x_13")]; + tensor x_13 = layer_norm(axes = x_13_axes_0, beta = inner_encoder_ret_lns_2_bias, epsilon = var_76, gamma = inner_encoder_ret_lns_2_weight, x = input_95)[name = tensor("x_13")]; tensor prev_kv_5_begin_0 = const()[name = tensor("prev_kv_5_begin_0"), val = tensor([2, 0, 0, 0, 0])]; tensor prev_kv_5_end_0 = const()[name = tensor("prev_kv_5_end_0"), val = tensor([3, 1, 4, 64, 64])]; tensor prev_kv_5_end_mask_0 = const()[name = tensor("prev_kv_5_end_mask_0"), val = tensor([false, true, true, true, true])]; @@ -613,183 +639,183 @@ program(1.0) tensor prev_scale_5_end_mask_0 = const()[name = tensor("prev_scale_5_end_mask_0"), val = tensor([false, true])]; tensor prev_scale_5_squeeze_mask_0 = const()[name = tensor("prev_scale_5_squeeze_mask_0"), val = tensor([true, false])]; tensor prev_scale_5 = slice_by_index(begin = prev_scale_5_begin_0, end = prev_scale_5_end_0, end_mask = prev_scale_5_end_mask_0, squeeze_mask = prev_scale_5_squeeze_mask_0, x = enc_scale)[name = tensor("prev_scale_5")]; - tensor var_607 = linear(bias = encoder_q_proj_2_bias, weight = encoder_q_proj_2_weight, x = x_13)[name = tensor("linear_21")]; - tensor var_608 = const()[name = tensor("op_608"), val = tensor([1, 5, 4, 64])]; - tensor var_609 = reshape(shape = var_608, x = var_607)[name = tensor("op_609")]; + tensor var_673 = linear(bias = inner_encoder_q_proj_2_bias, weight = inner_encoder_q_proj_2_weight, x = x_13)[name = tensor("linear_21")]; + tensor var_674 = const()[name = tensor("op_674"), val = tensor([1, 5, 4, 64])]; + tensor var_675 = reshape(shape = var_674, x = var_673)[name = tensor("op_675")]; tensor q_5_perm_0 = const()[name = tensor("q_5_perm_0"), val = tensor([0, 2, 1, 3])]; - tensor var_613 = linear(bias = encoder_k_proj_2_bias, weight = encoder_k_proj_2_weight, x = x_13)[name = tensor("linear_22")]; - tensor var_614 = const()[name = tensor("op_614"), val = tensor(0x1p-3)]; - tensor var_615 = mul(x = var_613, y = var_614)[name = tensor("op_615")]; - tensor var_616 = const()[name = tensor("op_616"), val = tensor([1, 5, 4, 64])]; - tensor var_617 = reshape(shape = var_616, x = var_615)[name = tensor("op_617")]; + tensor var_679 = linear(bias = inner_encoder_k_proj_2_bias, weight = inner_encoder_k_proj_2_weight, x = x_13)[name = tensor("linear_22")]; + tensor var_680 = const()[name = tensor("op_680"), val = tensor(0x1p-3)]; + tensor var_681 = mul(x = var_679, y = var_680)[name = tensor("op_681")]; + tensor var_682 = const()[name = tensor("op_682"), val = tensor([1, 5, 4, 64])]; + tensor var_683 = reshape(shape = var_682, x = var_681)[name = tensor("op_683")]; tensor k_5_perm_0 = const()[name = tensor("k_5_perm_0"), val = tensor([0, 2, 1, 3])]; - tensor var_621 = linear(bias = encoder_v_proj_2_bias, weight = encoder_v_proj_2_weight, x = x_13)[name = tensor("linear_23")]; - tensor var_622 = const()[name = tensor("op_622"), val = tensor([1, 5, 4, 64])]; - tensor var_623 = reshape(shape = var_622, x = var_621)[name = tensor("op_623")]; + tensor var_687 = linear(bias = inner_encoder_v_proj_2_bias, weight = inner_encoder_v_proj_2_weight, x = x_13)[name = tensor("linear_23")]; + tensor var_688 = const()[name = tensor("op_688"), val = tensor([1, 5, 4, 64])]; + tensor var_689 = reshape(shape = var_688, x = var_687)[name = tensor("op_689")]; tensor v_5_perm_0 = const()[name = tensor("v_5_perm_0"), val = tensor([0, 2, 1, 3])]; - tensor input_97 = linear(bias = encoder_g_proj_2_bias, weight = encoder_g_proj_2_weight, x = x_13)[name = tensor("linear_24")]; + tensor input_99 = linear(bias = inner_encoder_g_proj_2_bias, weight = inner_encoder_g_proj_2_weight, x = x_13)[name = tensor("linear_24")]; tensor sqrt_s0_5 = sqrt(x = prev_scale_5)[name = tensor("sqrt_s0_5")]; - tensor s_t_5 = add(x = prev_scale_5, y = encoder__t_index)[name = tensor("s_t_5")]; + tensor s_t_5 = add(x = prev_scale_5, y = inner_encoder__t_index)[name = tensor("s_t_5")]; tensor sqrt_s_t_5 = sqrt(x = s_t_5)[name = tensor("sqrt_s_t_5")]; tensor qk_5_transpose_x_1 = const()[name = tensor("qk_5_transpose_x_1"), val = tensor(false)]; tensor qk_5_transpose_y_1 = const()[name = tensor("qk_5_transpose_y_1"), val = tensor(true)]; - tensor k_5 = transpose(perm = k_5_perm_0, x = var_617)[name = tensor("transpose_43")]; - tensor q_5 = transpose(perm = q_5_perm_0, x = var_609)[name = tensor("transpose_44")]; + tensor k_5 = transpose(perm = k_5_perm_0, x = var_683)[name = tensor("transpose_43")]; + tensor q_5 = transpose(perm = q_5_perm_0, x = var_675)[name = tensor("transpose_44")]; tensor qk_5 = matmul(transpose_x = qk_5_transpose_x_1, transpose_y = qk_5_transpose_y_1, x = q_5, y = k_5)[name = tensor("qk_5")]; - tensor var_633 = const()[name = tensor("op_633"), val = tensor([5, 1])]; - tensor var_634 = reshape(shape = var_633, x = sqrt_s_t_5)[name = tensor("op_634")]; - tensor M_5 = real_div(x = encoder__causal_mask, y = var_634)[name = tensor("M_5")]; - tensor var_636 = mul(x = qk_5, y = M_5)[name = tensor("op_636")]; + tensor var_699 = const()[name = tensor("op_699"), val = tensor([5, 1])]; + tensor var_700 = reshape(shape = var_699, x = sqrt_s_t_5)[name = tensor("op_700")]; + tensor M_5 = real_div(x = inner_encoder__causal_mask, y = var_700)[name = tensor("M_5")]; + tensor var_702 = mul(x = qk_5, y = M_5)[name = tensor("op_702")]; tensor inner_5_transpose_x_0 = const()[name = tensor("inner_5_transpose_x_0"), val = tensor(false)]; tensor inner_5_transpose_y_0 = const()[name = tensor("inner_5_transpose_y_0"), val = tensor(false)]; - tensor v_5 = transpose(perm = v_5_perm_0, x = var_623)[name = tensor("transpose_42")]; - tensor inner_5 = matmul(transpose_x = inner_5_transpose_x_0, transpose_y = inner_5_transpose_y_0, x = var_636, y = v_5)[name = tensor("inner_5")]; - tensor var_638_transpose_x_0 = const()[name = tensor("op_638_transpose_x_0"), val = tensor(false)]; - tensor var_638_transpose_y_0 = const()[name = tensor("op_638_transpose_y_0"), val = tensor(false)]; - tensor var_638 = matmul(transpose_x = var_638_transpose_x_0, transpose_y = var_638_transpose_y_0, x = q_5, y = prev_kv_5)[name = tensor("op_638")]; - tensor var_639 = real_div(x = sqrt_s0_5, y = sqrt_s_t_5)[name = tensor("op_639")]; - tensor var_640 = const()[name = tensor("op_640"), val = tensor([1, 1, 5, 1])]; - tensor var_641 = reshape(shape = var_640, x = var_639)[name = tensor("op_641")]; - tensor cross_5 = mul(x = var_638, y = var_641)[name = tensor("cross_5")]; + tensor v_5 = transpose(perm = v_5_perm_0, x = var_689)[name = tensor("transpose_42")]; + tensor inner_5 = matmul(transpose_x = inner_5_transpose_x_0, transpose_y = inner_5_transpose_y_0, x = var_702, y = v_5)[name = tensor("inner_5")]; + tensor var_704_transpose_x_0 = const()[name = tensor("op_704_transpose_x_0"), val = tensor(false)]; + tensor var_704_transpose_y_0 = const()[name = tensor("op_704_transpose_y_0"), val = tensor(false)]; + tensor var_704 = matmul(transpose_x = var_704_transpose_x_0, transpose_y = var_704_transpose_y_0, x = q_5, y = prev_kv_5)[name = tensor("op_704")]; + tensor var_705 = real_div(x = sqrt_s0_5, y = sqrt_s_t_5)[name = tensor("op_705")]; + tensor var_706 = const()[name = tensor("op_706"), val = tensor([1, 1, 5, 1])]; + tensor var_707 = reshape(shape = var_706, x = var_705)[name = tensor("op_707")]; + tensor cross_5 = mul(x = var_704, y = var_707)[name = tensor("cross_5")]; tensor out_13 = add(x = inner_5, y = cross_5)[name = tensor("out_13")]; - tensor var_644 = mul(x = prev_kv_5, y = sqrt_s0_5)[name = tensor("op_644")]; - tensor var_646_transpose_x_1 = const()[name = tensor("op_646_transpose_x_1"), val = tensor(true)]; - tensor var_646_transpose_y_1 = const()[name = tensor("op_646_transpose_y_1"), val = tensor(false)]; - tensor var_646 = matmul(transpose_x = var_646_transpose_x_1, transpose_y = var_646_transpose_y_1, x = k_5, y = v_5)[name = tensor("op_646")]; - tensor new_kv_unnorm_5 = add(x = var_644, y = var_646)[name = tensor("new_kv_unnorm_5")]; - tensor var_648 = const()[name = tensor("op_648"), val = tensor(0x1.4p+2)]; - tensor new_scale_5 = add(x = prev_scale_5, y = var_648)[name = tensor("new_scale_5")]; - tensor var_650 = sqrt(x = new_scale_5)[name = tensor("op_650")]; - tensor var_651 = real_div(x = new_kv_unnorm_5, y = var_650)[name = tensor("op_651")]; - tensor var_652_perm_0 = const()[name = tensor("op_652_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_710 = mul(x = prev_kv_5, y = sqrt_s0_5)[name = tensor("op_710")]; + tensor var_712_transpose_x_1 = const()[name = tensor("op_712_transpose_x_1"), val = tensor(true)]; + tensor var_712_transpose_y_1 = const()[name = tensor("op_712_transpose_y_1"), val = tensor(false)]; + tensor var_712 = matmul(transpose_x = var_712_transpose_x_1, transpose_y = var_712_transpose_y_1, x = k_5, y = v_5)[name = tensor("op_712")]; + tensor new_kv_unnorm_5 = add(x = var_710, y = var_712)[name = tensor("new_kv_unnorm_5")]; + tensor var_714 = const()[name = tensor("op_714"), val = tensor(0x1.4p+2)]; + tensor new_scale_5 = add(x = prev_scale_5, y = var_714)[name = tensor("new_scale_5")]; + tensor var_716 = sqrt(x = new_scale_5)[name = tensor("op_716")]; + tensor var_717 = real_div(x = new_kv_unnorm_5, y = var_716)[name = tensor("op_717")]; + tensor var_718_perm_0 = const()[name = tensor("op_718_perm_0"), val = tensor([0, 2, 1, 3])]; tensor out_15_axes_0 = const()[name = tensor("out_15_axes_0"), val = tensor([-1])]; - tensor var_652 = transpose(perm = var_652_perm_0, x = out_13)[name = tensor("transpose_41")]; - tensor out_15 = layer_norm(axes = out_15_axes_0, epsilon = var_18, x = var_652)[name = tensor("out_15")]; - tensor var_656 = const()[name = tensor("op_656"), val = tensor([1, 5, 256])]; - tensor out_17 = reshape(shape = var_656, x = out_15)[name = tensor("out_17")]; - tensor var_658 = silu(x = input_97)[name = tensor("op_658")]; - tensor input_99 = mul(x = var_658, y = out_17)[name = tensor("input_99")]; - tensor ret_out_5 = linear(bias = encoder_out_proj_2_bias, weight = encoder_out_proj_2_weight, x = input_99)[name = tensor("linear_25")]; - tensor x_15 = add(x = input_93, y = ret_out_5)[name = tensor("x_15")]; + tensor var_718 = transpose(perm = var_718_perm_0, x = out_13)[name = tensor("transpose_41")]; + tensor out_15 = layer_norm(axes = out_15_axes_0, epsilon = var_84, x = var_718)[name = tensor("out_15")]; + tensor var_722 = const()[name = tensor("op_722"), val = tensor([1, 5, 256])]; + tensor out_17 = reshape(shape = var_722, x = out_15)[name = tensor("out_17")]; + tensor var_724 = silu(x = input_99)[name = tensor("op_724")]; + tensor input_101 = mul(x = var_724, y = out_17)[name = tensor("input_101")]; + tensor ret_out_5 = linear(bias = inner_encoder_out_proj_2_bias, weight = inner_encoder_out_proj_2_weight, x = input_101)[name = tensor("linear_25")]; + tensor x_15 = add(x = input_95, y = ret_out_5)[name = tensor("x_15")]; tensor window_25_begin_0 = const()[name = tensor("window_25_begin_0"), val = tensor([2, 0, 0, 0])]; tensor window_25_end_0 = const()[name = tensor("window_25_end_0"), val = tensor([3, 1, 16, 256])]; tensor window_25_end_mask_0 = const()[name = tensor("window_25_end_mask_0"), val = tensor([false, true, true, true])]; tensor window_25_squeeze_mask_0 = const()[name = tensor("window_25_squeeze_mask_0"), val = tensor([true, false, false, false])]; tensor window_25 = slice_by_index(begin = window_25_begin_0, end = window_25_end_0, end_mask = window_25_end_mask_0, squeeze_mask = window_25_squeeze_mask_0, x = enc_conv_cache)[name = tensor("window_25")]; - tensor var_666_begin_0 = const()[name = tensor("op_666_begin_0"), val = tensor([0, 0, 0])]; - tensor var_666_end_0 = const()[name = tensor("op_666_end_0"), val = tensor([1, 1, 256])]; - tensor var_666_end_mask_0 = const()[name = tensor("op_666_end_mask_0"), val = tensor([true, false, true])]; - tensor var_666 = slice_by_index(begin = var_666_begin_0, end = var_666_end_0, end_mask = var_666_end_mask_0, x = x_15)[name = tensor("op_666")]; - tensor var_669_begin_0 = const()[name = tensor("op_669_begin_0"), val = tensor([0, 1, 0])]; - tensor var_669_end_0 = const()[name = tensor("op_669_end_0"), val = tensor([1, 16, 256])]; - tensor var_669_end_mask_0 = const()[name = tensor("op_669_end_mask_0"), val = tensor([true, true, true])]; - tensor var_669 = slice_by_index(begin = var_669_begin_0, end = var_669_end_0, end_mask = var_669_end_mask_0, x = window_25)[name = tensor("op_669")]; + tensor var_732_begin_0 = const()[name = tensor("op_732_begin_0"), val = tensor([0, 0, 0])]; + tensor var_732_end_0 = const()[name = tensor("op_732_end_0"), val = tensor([1, 1, 256])]; + tensor var_732_end_mask_0 = const()[name = tensor("op_732_end_mask_0"), val = tensor([true, false, true])]; + tensor var_732 = slice_by_index(begin = var_732_begin_0, end = var_732_end_0, end_mask = var_732_end_mask_0, x = x_15)[name = tensor("op_732")]; + tensor var_735_begin_0 = const()[name = tensor("op_735_begin_0"), val = tensor([0, 1, 0])]; + tensor var_735_end_0 = const()[name = tensor("op_735_end_0"), val = tensor([1, 16, 256])]; + tensor var_735_end_mask_0 = const()[name = tensor("op_735_end_mask_0"), val = tensor([true, true, true])]; + tensor var_735 = slice_by_index(begin = var_735_begin_0, end = var_735_end_0, end_mask = var_735_end_mask_0, x = window_25)[name = tensor("op_735")]; tensor window_27_interleave_0 = const()[name = tensor("window_27_interleave_0"), val = tensor(false)]; - tensor window_27 = concat(axis = var_27, interleave = window_27_interleave_0, values = (var_669, var_666))[name = tensor("window_27")]; - tensor var_674_begin_0 = const()[name = tensor("op_674_begin_0"), val = tensor([0, 1, 0])]; - tensor var_674_end_0 = const()[name = tensor("op_674_end_0"), val = tensor([1, 2, 256])]; - tensor var_674_end_mask_0 = const()[name = tensor("op_674_end_mask_0"), val = tensor([true, false, true])]; - tensor var_674 = slice_by_index(begin = var_674_begin_0, end = var_674_end_0, end_mask = var_674_end_mask_0, x = x_15)[name = tensor("op_674")]; - tensor var_677_begin_0 = const()[name = tensor("op_677_begin_0"), val = tensor([0, 1, 0])]; - tensor var_677_end_0 = const()[name = tensor("op_677_end_0"), val = tensor([1, 16, 256])]; - tensor var_677_end_mask_0 = const()[name = tensor("op_677_end_mask_0"), val = tensor([true, true, true])]; - tensor var_677 = slice_by_index(begin = var_677_begin_0, end = var_677_end_0, end_mask = var_677_end_mask_0, x = window_27)[name = tensor("op_677")]; + tensor window_27 = concat(axis = var_93, interleave = window_27_interleave_0, values = (var_735, var_732))[name = tensor("window_27")]; + tensor var_740_begin_0 = const()[name = tensor("op_740_begin_0"), val = tensor([0, 1, 0])]; + tensor var_740_end_0 = const()[name = tensor("op_740_end_0"), val = tensor([1, 2, 256])]; + tensor var_740_end_mask_0 = const()[name = tensor("op_740_end_mask_0"), val = tensor([true, false, true])]; + tensor var_740 = slice_by_index(begin = var_740_begin_0, end = var_740_end_0, end_mask = var_740_end_mask_0, x = x_15)[name = tensor("op_740")]; + tensor var_743_begin_0 = const()[name = tensor("op_743_begin_0"), val = tensor([0, 1, 0])]; + tensor var_743_end_0 = const()[name = tensor("op_743_end_0"), val = tensor([1, 16, 256])]; + tensor var_743_end_mask_0 = const()[name = tensor("op_743_end_mask_0"), val = tensor([true, true, true])]; + tensor var_743 = slice_by_index(begin = var_743_begin_0, end = var_743_end_0, end_mask = var_743_end_mask_0, x = window_27)[name = tensor("op_743")]; tensor window_29_interleave_0 = const()[name = tensor("window_29_interleave_0"), val = tensor(false)]; - tensor window_29 = concat(axis = var_27, interleave = window_29_interleave_0, values = (var_677, var_674))[name = tensor("window_29")]; - tensor var_682_begin_0 = const()[name = tensor("op_682_begin_0"), val = tensor([0, 2, 0])]; - tensor var_682_end_0 = const()[name = tensor("op_682_end_0"), val = tensor([1, 3, 256])]; - tensor var_682_end_mask_0 = const()[name = tensor("op_682_end_mask_0"), val = tensor([true, false, true])]; - tensor var_682 = slice_by_index(begin = var_682_begin_0, end = var_682_end_0, end_mask = var_682_end_mask_0, x = x_15)[name = tensor("op_682")]; - tensor var_685_begin_0 = const()[name = tensor("op_685_begin_0"), val = tensor([0, 1, 0])]; - tensor var_685_end_0 = const()[name = tensor("op_685_end_0"), val = tensor([1, 16, 256])]; - tensor var_685_end_mask_0 = const()[name = tensor("op_685_end_mask_0"), val = tensor([true, true, true])]; - tensor var_685 = slice_by_index(begin = var_685_begin_0, end = var_685_end_0, end_mask = var_685_end_mask_0, x = window_29)[name = tensor("op_685")]; + tensor window_29 = concat(axis = var_93, interleave = window_29_interleave_0, values = (var_743, var_740))[name = tensor("window_29")]; + tensor var_748_begin_0 = const()[name = tensor("op_748_begin_0"), val = tensor([0, 2, 0])]; + tensor var_748_end_0 = const()[name = tensor("op_748_end_0"), val = tensor([1, 3, 256])]; + tensor var_748_end_mask_0 = const()[name = tensor("op_748_end_mask_0"), val = tensor([true, false, true])]; + tensor var_748 = slice_by_index(begin = var_748_begin_0, end = var_748_end_0, end_mask = var_748_end_mask_0, x = x_15)[name = tensor("op_748")]; + tensor var_751_begin_0 = const()[name = tensor("op_751_begin_0"), val = tensor([0, 1, 0])]; + tensor var_751_end_0 = const()[name = tensor("op_751_end_0"), val = tensor([1, 16, 256])]; + tensor var_751_end_mask_0 = const()[name = tensor("op_751_end_mask_0"), val = tensor([true, true, true])]; + tensor var_751 = slice_by_index(begin = var_751_begin_0, end = var_751_end_0, end_mask = var_751_end_mask_0, x = window_29)[name = tensor("op_751")]; tensor window_31_interleave_0 = const()[name = tensor("window_31_interleave_0"), val = tensor(false)]; - tensor window_31 = concat(axis = var_27, interleave = window_31_interleave_0, values = (var_685, var_682))[name = tensor("window_31")]; - tensor var_690_begin_0 = const()[name = tensor("op_690_begin_0"), val = tensor([0, 3, 0])]; - tensor var_690_end_0 = const()[name = tensor("op_690_end_0"), val = tensor([1, 4, 256])]; - tensor var_690_end_mask_0 = const()[name = tensor("op_690_end_mask_0"), val = tensor([true, false, true])]; - tensor var_690 = slice_by_index(begin = var_690_begin_0, end = var_690_end_0, end_mask = var_690_end_mask_0, x = x_15)[name = tensor("op_690")]; - tensor var_693_begin_0 = const()[name = tensor("op_693_begin_0"), val = tensor([0, 1, 0])]; - tensor var_693_end_0 = const()[name = tensor("op_693_end_0"), val = tensor([1, 16, 256])]; - tensor var_693_end_mask_0 = const()[name = tensor("op_693_end_mask_0"), val = tensor([true, true, true])]; - tensor var_693 = slice_by_index(begin = var_693_begin_0, end = var_693_end_0, end_mask = var_693_end_mask_0, x = window_31)[name = tensor("op_693")]; + tensor window_31 = concat(axis = var_93, interleave = window_31_interleave_0, values = (var_751, var_748))[name = tensor("window_31")]; + tensor var_756_begin_0 = const()[name = tensor("op_756_begin_0"), val = tensor([0, 3, 0])]; + tensor var_756_end_0 = const()[name = tensor("op_756_end_0"), val = tensor([1, 4, 256])]; + tensor var_756_end_mask_0 = const()[name = tensor("op_756_end_mask_0"), val = tensor([true, false, true])]; + tensor var_756 = slice_by_index(begin = var_756_begin_0, end = var_756_end_0, end_mask = var_756_end_mask_0, x = x_15)[name = tensor("op_756")]; + tensor var_759_begin_0 = const()[name = tensor("op_759_begin_0"), val = tensor([0, 1, 0])]; + tensor var_759_end_0 = const()[name = tensor("op_759_end_0"), val = tensor([1, 16, 256])]; + tensor var_759_end_mask_0 = const()[name = tensor("op_759_end_mask_0"), val = tensor([true, true, true])]; + tensor var_759 = slice_by_index(begin = var_759_begin_0, end = var_759_end_0, end_mask = var_759_end_mask_0, x = window_31)[name = tensor("op_759")]; tensor window_33_interleave_0 = const()[name = tensor("window_33_interleave_0"), val = tensor(false)]; - tensor window_33 = concat(axis = var_27, interleave = window_33_interleave_0, values = (var_693, var_690))[name = tensor("window_33")]; - tensor var_698_begin_0 = const()[name = tensor("op_698_begin_0"), val = tensor([0, 4, 0])]; - tensor var_698_end_0 = const()[name = tensor("op_698_end_0"), val = tensor([1, 1, 256])]; - tensor var_698_end_mask_0 = const()[name = tensor("op_698_end_mask_0"), val = tensor([true, true, true])]; - tensor var_698 = slice_by_index(begin = var_698_begin_0, end = var_698_end_0, end_mask = var_698_end_mask_0, x = x_15)[name = tensor("op_698")]; - tensor var_701_begin_0 = const()[name = tensor("op_701_begin_0"), val = tensor([0, 1, 0])]; - tensor var_701_end_0 = const()[name = tensor("op_701_end_0"), val = tensor([1, 16, 256])]; - tensor var_701_end_mask_0 = const()[name = tensor("op_701_end_mask_0"), val = tensor([true, true, true])]; - tensor var_701 = slice_by_index(begin = var_701_begin_0, end = var_701_end_0, end_mask = var_701_end_mask_0, x = window_33)[name = tensor("op_701")]; + tensor window_33 = concat(axis = var_93, interleave = window_33_interleave_0, values = (var_759, var_756))[name = tensor("window_33")]; + tensor var_764_begin_0 = const()[name = tensor("op_764_begin_0"), val = tensor([0, 4, 0])]; + tensor var_764_end_0 = const()[name = tensor("op_764_end_0"), val = tensor([1, 1, 256])]; + tensor var_764_end_mask_0 = const()[name = tensor("op_764_end_mask_0"), val = tensor([true, true, true])]; + tensor var_764 = slice_by_index(begin = var_764_begin_0, end = var_764_end_0, end_mask = var_764_end_mask_0, x = x_15)[name = tensor("op_764")]; + tensor var_767_begin_0 = const()[name = tensor("op_767_begin_0"), val = tensor([0, 1, 0])]; + tensor var_767_end_0 = const()[name = tensor("op_767_end_0"), val = tensor([1, 16, 256])]; + tensor var_767_end_mask_0 = const()[name = tensor("op_767_end_mask_0"), val = tensor([true, true, true])]; + tensor var_767 = slice_by_index(begin = var_767_begin_0, end = var_767_end_0, end_mask = var_767_end_mask_0, x = window_33)[name = tensor("op_767")]; tensor window_35_interleave_0 = const()[name = tensor("window_35_interleave_0"), val = tensor(false)]; - tensor window_35 = concat(axis = var_27, interleave = window_35_interleave_0, values = (var_701, var_698))[name = tensor("window_35")]; - tensor input_101_interleave_0 = const()[name = tensor("input_101_interleave_0"), val = tensor(false)]; - tensor input_101 = concat(axis = var_24, interleave = input_101_interleave_0, values = (window_27, window_29, window_31, window_33, window_35))[name = tensor("input_101")]; + tensor window_35 = concat(axis = var_93, interleave = window_35_interleave_0, values = (var_767, var_764))[name = tensor("window_35")]; + tensor input_103_interleave_0 = const()[name = tensor("input_103_interleave_0"), val = tensor(false)]; + tensor input_103 = concat(axis = var_79, interleave = input_103_interleave_0, values = (window_27, window_29, window_31, window_33, window_35))[name = tensor("input_103")]; tensor x_17_axes_0 = const()[name = tensor("x_17_axes_0"), val = tensor([-1])]; - tensor x_17 = layer_norm(axes = x_17_axes_0, beta = encoder_conv_module_2_sequential_0_bias, epsilon = var_29, gamma = encoder_conv_module_2_sequential_0_weight, x = input_101)[name = tensor("x_17")]; - tensor input_103_perm_0 = const()[name = tensor("input_103_perm_0"), val = tensor([0, 2, 1])]; + tensor x_17 = layer_norm(axes = x_17_axes_0, beta = inner_encoder_conv_module_2_sequential_0_bias, epsilon = var_76, gamma = inner_encoder_conv_module_2_sequential_0_weight, x = input_103)[name = tensor("x_17")]; + tensor input_105_perm_0 = const()[name = tensor("input_105_perm_0"), val = tensor([0, 2, 1])]; tensor inputs_23_pad_type_0 = const()[name = tensor("inputs_23_pad_type_0"), val = tensor("valid")]; tensor inputs_23_strides_0 = const()[name = tensor("inputs_23_strides_0"), val = tensor([1])]; tensor inputs_23_pad_0 = const()[name = tensor("inputs_23_pad_0"), val = tensor([0, 0])]; tensor inputs_23_dilations_0 = const()[name = tensor("inputs_23_dilations_0"), val = tensor([1])]; tensor inputs_23_groups_0 = const()[name = tensor("inputs_23_groups_0"), val = tensor(1)]; - tensor input_103 = transpose(perm = input_103_perm_0, x = x_17)[name = tensor("transpose_40")]; - tensor inputs_23 = conv(bias = encoder_conv_module_2_sequential_2_conv_bias, dilations = inputs_23_dilations_0, groups = inputs_23_groups_0, pad = inputs_23_pad_0, pad_type = inputs_23_pad_type_0, strides = inputs_23_strides_0, weight = encoder_conv_module_2_sequential_2_conv_weight, x = input_103)[name = tensor("inputs_23")]; - tensor var_726_split_sizes_0 = const()[name = tensor("op_726_split_sizes_0"), val = tensor([256, 256])]; - tensor var_726_axis_0 = const()[name = tensor("op_726_axis_0"), val = tensor(1)]; - tensor var_726_0, tensor var_726_1 = split(axis = var_726_axis_0, split_sizes = var_726_split_sizes_0, x = inputs_23)[name = tensor("op_726")]; - tensor var_728 = sigmoid(x = var_726_1)[name = tensor("op_728")]; - tensor inputs_25 = mul(x = var_726_0, y = var_728)[name = tensor("inputs_25")]; + tensor input_105 = transpose(perm = input_105_perm_0, x = x_17)[name = tensor("transpose_40")]; + tensor inputs_23 = conv(bias = inner_encoder_conv_module_2_sequential_2_conv_bias, dilations = inputs_23_dilations_0, groups = inputs_23_groups_0, pad = inputs_23_pad_0, pad_type = inputs_23_pad_type_0, strides = inputs_23_strides_0, weight = inner_encoder_conv_module_2_sequential_2_conv_weight, x = input_105)[name = tensor("inputs_23")]; + tensor var_792_split_sizes_0 = const()[name = tensor("op_792_split_sizes_0"), val = tensor([256, 256])]; + tensor var_792_axis_0 = const()[name = tensor("op_792_axis_0"), val = tensor(1)]; + tensor var_792_0, tensor var_792_1 = split(axis = var_792_axis_0, split_sizes = var_792_split_sizes_0, x = inputs_23)[name = tensor("op_792")]; + tensor var_794 = sigmoid(x = var_792_1)[name = tensor("op_794")]; + tensor inputs_25 = mul(x = var_792_0, y = var_794)[name = tensor("inputs_25")]; tensor outputs_aug_5_pad_type_0 = const()[name = tensor("outputs_aug_5_pad_type_0"), val = tensor("custom")]; tensor outputs_aug_5_pad_0 = const()[name = tensor("outputs_aug_5_pad_0"), val = tensor([15, 15])]; tensor outputs_aug_5_groups_0 = const()[name = tensor("outputs_aug_5_groups_0"), val = tensor(256)]; tensor outputs_aug_5_strides_0 = const()[name = tensor("outputs_aug_5_strides_0"), val = tensor([1])]; tensor outputs_aug_5_dilations_0 = const()[name = tensor("outputs_aug_5_dilations_0"), val = tensor([1])]; - tensor outputs_aug_5 = conv(dilations = outputs_aug_5_dilations_0, groups = outputs_aug_5_groups_0, pad = outputs_aug_5_pad_0, pad_type = outputs_aug_5_pad_type_0, strides = outputs_aug_5_strides_0, weight = encoder_conv_module_2_sequential_4_conv_weight, x = inputs_25)[name = tensor("outputs_aug_5")]; - tensor input_105_begin_0 = const()[name = tensor("input_105_begin_0"), val = tensor([0, 0, 0])]; - tensor input_105_end_0 = const()[name = tensor("input_105_end_0"), val = tensor([5, 256, 16])]; - tensor input_105_end_mask_0 = const()[name = tensor("input_105_end_mask_0"), val = tensor([true, true, false])]; - tensor input_105 = slice_by_index(begin = input_105_begin_0, end = input_105_end_0, end_mask = input_105_end_mask_0, x = outputs_aug_5)[name = tensor("input_105")]; - tensor inputs_27 = batch_norm(beta = encoder_conv_module_2_sequential_5_bias, epsilon = var_29, gamma = encoder_conv_module_2_sequential_5_weight, mean = encoder_conv_module_2_sequential_5_running_mean, variance = encoder_conv_module_2_sequential_5_running_var, x = input_105)[name = tensor("inputs_27")]; - tensor input_107 = silu(x = inputs_27)[name = tensor("input_107")]; - tensor input_109_pad_type_0 = const()[name = tensor("input_109_pad_type_0"), val = tensor("valid")]; - tensor input_109_strides_0 = const()[name = tensor("input_109_strides_0"), val = tensor([1])]; - tensor input_109_pad_0 = const()[name = tensor("input_109_pad_0"), val = tensor([0, 0])]; - tensor input_109_dilations_0 = const()[name = tensor("input_109_dilations_0"), val = tensor([1])]; - tensor input_109_groups_0 = const()[name = tensor("input_109_groups_0"), val = tensor(1)]; - tensor input_109 = conv(bias = encoder_conv_module_2_sequential_7_conv_bias, dilations = input_109_dilations_0, groups = input_109_groups_0, pad = input_109_pad_0, pad_type = input_109_pad_type_0, strides = input_109_strides_0, weight = encoder_conv_module_2_sequential_7_conv_weight, x = input_107)[name = tensor("input_109")]; + tensor outputs_aug_5 = conv(dilations = outputs_aug_5_dilations_0, groups = outputs_aug_5_groups_0, pad = outputs_aug_5_pad_0, pad_type = outputs_aug_5_pad_type_0, strides = outputs_aug_5_strides_0, weight = inner_encoder_conv_module_2_sequential_4_conv_weight, x = inputs_25)[name = tensor("outputs_aug_5")]; + tensor input_107_begin_0 = const()[name = tensor("input_107_begin_0"), val = tensor([0, 0, 0])]; + tensor input_107_end_0 = const()[name = tensor("input_107_end_0"), val = tensor([5, 256, 16])]; + tensor input_107_end_mask_0 = const()[name = tensor("input_107_end_mask_0"), val = tensor([true, true, false])]; + tensor input_107 = slice_by_index(begin = input_107_begin_0, end = input_107_end_0, end_mask = input_107_end_mask_0, x = outputs_aug_5)[name = tensor("input_107")]; + tensor inputs_27 = batch_norm(beta = inner_encoder_conv_module_2_sequential_5_bias, epsilon = var_76, gamma = inner_encoder_conv_module_2_sequential_5_weight, mean = inner_encoder_conv_module_2_sequential_5_running_mean, variance = inner_encoder_conv_module_2_sequential_5_running_var, x = input_107)[name = tensor("inputs_27")]; + tensor input_109 = silu(x = inputs_27)[name = tensor("input_109")]; + tensor input_111_pad_type_0 = const()[name = tensor("input_111_pad_type_0"), val = tensor("valid")]; + tensor input_111_strides_0 = const()[name = tensor("input_111_strides_0"), val = tensor([1])]; + tensor input_111_pad_0 = const()[name = tensor("input_111_pad_0"), val = tensor([0, 0])]; + tensor input_111_dilations_0 = const()[name = tensor("input_111_dilations_0"), val = tensor([1])]; + tensor input_111_groups_0 = const()[name = tensor("input_111_groups_0"), val = tensor(1)]; + tensor input_111 = conv(bias = inner_encoder_conv_module_2_sequential_7_conv_bias, dilations = input_111_dilations_0, groups = input_111_groups_0, pad = input_111_pad_0, pad_type = input_111_pad_type_0, strides = input_111_strides_0, weight = inner_encoder_conv_module_2_sequential_7_conv_weight, x = input_109)[name = tensor("input_111")]; tensor conv_out_5_perm_0 = const()[name = tensor("conv_out_5_perm_0"), val = tensor([0, 2, 1])]; - tensor var_759_begin_0 = const()[name = tensor("op_759_begin_0"), val = tensor([0, -1, 0])]; - tensor var_759_end_0 = const()[name = tensor("op_759_end_0"), val = tensor([5, 16, 256])]; - tensor var_759_end_mask_0 = const()[name = tensor("op_759_end_mask_0"), val = tensor([true, true, true])]; - tensor conv_out_5 = transpose(perm = conv_out_5_perm_0, x = input_109)[name = tensor("transpose_39")]; - tensor var_759 = slice_by_index(begin = var_759_begin_0, end = var_759_end_0, end_mask = var_759_end_mask_0, x = conv_out_5)[name = tensor("op_759")]; - tensor var_761_perm_0 = const()[name = tensor("op_761_perm_0"), val = tensor([1, 0, 2])]; - tensor var_761 = transpose(perm = var_761_perm_0, x = var_759)[name = tensor("transpose_38")]; - tensor input_111 = add(x = x_15, y = var_761)[name = tensor("input_111")]; - tensor input_113_axes_0 = const()[name = tensor("input_113_axes_0"), val = tensor([-1])]; - tensor input_113 = layer_norm(axes = input_113_axes_0, beta = encoder_ffn2_2_module_sequential_0_bias, epsilon = var_29, gamma = encoder_ffn2_2_module_sequential_0_weight, x = input_111)[name = tensor("input_113")]; - tensor inputs_29 = linear(bias = encoder_ffn2_2_module_sequential_1_linear_bias, weight = encoder_ffn2_2_module_sequential_1_linear_weight, x = input_113)[name = tensor("linear_26")]; - tensor input_115 = silu(x = inputs_29)[name = tensor("input_115")]; - tensor input_119 = linear(bias = encoder_ffn2_2_module_sequential_4_linear_bias, weight = encoder_ffn2_2_module_sequential_4_linear_weight, x = input_115)[name = tensor("linear_27")]; - tensor var_784 = const()[name = tensor("op_784"), val = tensor(0x1p-1)]; - tensor var_785 = mul(x = input_119, y = var_784)[name = tensor("op_785")]; - tensor input_121 = add(x = var_785, y = input_111)[name = tensor("input_121")]; - tensor input_123_axes_0 = const()[name = tensor("input_123_axes_0"), val = tensor([-1])]; - tensor input_123 = layer_norm(axes = input_123_axes_0, beta = encoder_layer_norm_2_bias, epsilon = var_29, gamma = encoder_layer_norm_2_weight, x = input_121)[name = tensor("input_123")]; + tensor var_825_begin_0 = const()[name = tensor("op_825_begin_0"), val = tensor([0, -1, 0])]; + tensor var_825_end_0 = const()[name = tensor("op_825_end_0"), val = tensor([5, 16, 256])]; + tensor var_825_end_mask_0 = const()[name = tensor("op_825_end_mask_0"), val = tensor([true, true, true])]; + tensor conv_out_5 = transpose(perm = conv_out_5_perm_0, x = input_111)[name = tensor("transpose_39")]; + tensor var_825 = slice_by_index(begin = var_825_begin_0, end = var_825_end_0, end_mask = var_825_end_mask_0, x = conv_out_5)[name = tensor("op_825")]; + tensor var_827_perm_0 = const()[name = tensor("op_827_perm_0"), val = tensor([1, 0, 2])]; + tensor var_827 = transpose(perm = var_827_perm_0, x = var_825)[name = tensor("transpose_38")]; + tensor input_113 = add(x = x_15, y = var_827)[name = tensor("input_113")]; + tensor input_115_axes_0 = const()[name = tensor("input_115_axes_0"), val = tensor([-1])]; + tensor input_115 = layer_norm(axes = input_115_axes_0, beta = inner_encoder_ffn2_2_module_sequential_0_bias, epsilon = var_76, gamma = inner_encoder_ffn2_2_module_sequential_0_weight, x = input_113)[name = tensor("input_115")]; + tensor inputs_29 = linear(bias = inner_encoder_ffn2_2_module_sequential_1_linear_bias, weight = inner_encoder_ffn2_2_module_sequential_1_linear_weight, x = input_115)[name = tensor("linear_26")]; + tensor input_117 = silu(x = inputs_29)[name = tensor("input_117")]; + tensor input_121 = linear(bias = inner_encoder_ffn2_2_module_sequential_4_linear_bias, weight = inner_encoder_ffn2_2_module_sequential_4_linear_weight, x = input_117)[name = tensor("linear_27")]; + tensor var_850 = const()[name = tensor("op_850"), val = tensor(0x1p-1)]; + tensor var_851 = mul(x = input_121, y = var_850)[name = tensor("op_851")]; + tensor input_123 = add(x = var_851, y = input_113)[name = tensor("input_123")]; tensor input_125_axes_0 = const()[name = tensor("input_125_axes_0"), val = tensor([-1])]; - tensor input_125 = layer_norm(axes = input_125_axes_0, beta = encoder_ffn1_3_module_sequential_0_bias, epsilon = var_29, gamma = encoder_ffn1_3_module_sequential_0_weight, x = input_123)[name = tensor("input_125")]; - tensor inputs_31 = linear(bias = encoder_ffn1_3_module_sequential_1_linear_bias, weight = encoder_ffn1_3_module_sequential_1_linear_weight, x = input_125)[name = tensor("linear_28")]; - tensor input_127 = silu(x = inputs_31)[name = tensor("input_127")]; - tensor input_131 = linear(bias = encoder_ffn1_3_module_sequential_4_linear_bias, weight = encoder_ffn1_3_module_sequential_4_linear_weight, x = input_127)[name = tensor("linear_29")]; - tensor var_814 = const()[name = tensor("op_814"), val = tensor(0x1p-1)]; - tensor var_815 = mul(x = input_131, y = var_814)[name = tensor("op_815")]; - tensor input_133 = add(x = var_815, y = input_123)[name = tensor("input_133")]; + tensor input_125 = layer_norm(axes = input_125_axes_0, beta = inner_encoder_layer_norm_2_bias, epsilon = var_76, gamma = inner_encoder_layer_norm_2_weight, x = input_123)[name = tensor("input_125")]; + tensor input_127_axes_0 = const()[name = tensor("input_127_axes_0"), val = tensor([-1])]; + tensor input_127 = layer_norm(axes = input_127_axes_0, beta = inner_encoder_ffn1_3_module_sequential_0_bias, epsilon = var_76, gamma = inner_encoder_ffn1_3_module_sequential_0_weight, x = input_125)[name = tensor("input_127")]; + tensor inputs_31 = linear(bias = inner_encoder_ffn1_3_module_sequential_1_linear_bias, weight = inner_encoder_ffn1_3_module_sequential_1_linear_weight, x = input_127)[name = tensor("linear_28")]; + tensor input_129 = silu(x = inputs_31)[name = tensor("input_129")]; + tensor input_133 = linear(bias = inner_encoder_ffn1_3_module_sequential_4_linear_bias, weight = inner_encoder_ffn1_3_module_sequential_4_linear_weight, x = input_129)[name = tensor("linear_29")]; + tensor var_880 = const()[name = tensor("op_880"), val = tensor(0x1p-1)]; + tensor var_881 = mul(x = input_133, y = var_880)[name = tensor("op_881")]; + tensor input_135 = add(x = var_881, y = input_125)[name = tensor("input_135")]; tensor x_19_axes_0 = const()[name = tensor("x_19_axes_0"), val = tensor([-1])]; - tensor x_19 = layer_norm(axes = x_19_axes_0, beta = encoder_ret_lns_3_bias, epsilon = var_29, gamma = encoder_ret_lns_3_weight, x = input_133)[name = tensor("x_19")]; + tensor x_19 = layer_norm(axes = x_19_axes_0, beta = inner_encoder_ret_lns_3_bias, epsilon = var_76, gamma = inner_encoder_ret_lns_3_weight, x = input_135)[name = tensor("x_19")]; tensor prev_kv_7_begin_0 = const()[name = tensor("prev_kv_7_begin_0"), val = tensor([3, 0, 0, 0, 0])]; tensor prev_kv_7_end_0 = const()[name = tensor("prev_kv_7_end_0"), val = tensor([4, 1, 4, 64, 64])]; tensor prev_kv_7_end_mask_0 = const()[name = tensor("prev_kv_7_end_mask_0"), val = tensor([false, true, true, true, true])]; @@ -800,219 +826,212 @@ program(1.0) tensor prev_scale_7_end_mask_0 = const()[name = tensor("prev_scale_7_end_mask_0"), val = tensor([false, true])]; tensor prev_scale_7_squeeze_mask_0 = const()[name = tensor("prev_scale_7_squeeze_mask_0"), val = tensor([true, false])]; tensor prev_scale_7 = slice_by_index(begin = prev_scale_7_begin_0, end = prev_scale_7_end_0, end_mask = prev_scale_7_end_mask_0, squeeze_mask = prev_scale_7_squeeze_mask_0, x = enc_scale)[name = tensor("prev_scale_7")]; - tensor var_829 = linear(bias = encoder_q_proj_3_bias, weight = encoder_q_proj_3_weight, x = x_19)[name = tensor("linear_30")]; - tensor var_830 = const()[name = tensor("op_830"), val = tensor([1, 5, 4, 64])]; - tensor var_831 = reshape(shape = var_830, x = var_829)[name = tensor("op_831")]; + tensor var_895 = linear(bias = inner_encoder_q_proj_3_bias, weight = inner_encoder_q_proj_3_weight, x = x_19)[name = tensor("linear_30")]; + tensor var_896 = const()[name = tensor("op_896"), val = tensor([1, 5, 4, 64])]; + tensor var_897 = reshape(shape = var_896, x = var_895)[name = tensor("op_897")]; tensor q_7_perm_0 = const()[name = tensor("q_7_perm_0"), val = tensor([0, 2, 1, 3])]; - tensor var_835 = linear(bias = encoder_k_proj_3_bias, weight = encoder_k_proj_3_weight, x = x_19)[name = tensor("linear_31")]; - tensor var_836 = const()[name = tensor("op_836"), val = tensor(0x1p-3)]; - tensor var_837 = mul(x = var_835, y = var_836)[name = tensor("op_837")]; - tensor var_838 = const()[name = tensor("op_838"), val = tensor([1, 5, 4, 64])]; - tensor var_839 = reshape(shape = var_838, x = var_837)[name = tensor("op_839")]; + tensor var_901 = linear(bias = inner_encoder_k_proj_3_bias, weight = inner_encoder_k_proj_3_weight, x = x_19)[name = tensor("linear_31")]; + tensor var_902 = const()[name = tensor("op_902"), val = tensor(0x1p-3)]; + tensor var_903 = mul(x = var_901, y = var_902)[name = tensor("op_903")]; + tensor var_904 = const()[name = tensor("op_904"), val = tensor([1, 5, 4, 64])]; + tensor var_905 = reshape(shape = var_904, x = var_903)[name = tensor("op_905")]; tensor k_7_perm_0 = const()[name = tensor("k_7_perm_0"), val = tensor([0, 2, 1, 3])]; - tensor var_843 = linear(bias = encoder_v_proj_3_bias, weight = encoder_v_proj_3_weight, x = x_19)[name = tensor("linear_32")]; - tensor var_844 = const()[name = tensor("op_844"), val = tensor([1, 5, 4, 64])]; - tensor var_845 = reshape(shape = var_844, x = var_843)[name = tensor("op_845")]; + tensor var_909 = linear(bias = inner_encoder_v_proj_3_bias, weight = inner_encoder_v_proj_3_weight, x = x_19)[name = tensor("linear_32")]; + tensor var_910 = const()[name = tensor("op_910"), val = tensor([1, 5, 4, 64])]; + tensor var_911 = reshape(shape = var_910, x = var_909)[name = tensor("op_911")]; tensor v_7_perm_0 = const()[name = tensor("v_7_perm_0"), val = tensor([0, 2, 1, 3])]; - tensor input_137 = linear(bias = encoder_g_proj_3_bias, weight = encoder_g_proj_3_weight, x = x_19)[name = tensor("linear_33")]; + tensor input_139 = linear(bias = inner_encoder_g_proj_3_bias, weight = inner_encoder_g_proj_3_weight, x = x_19)[name = tensor("linear_33")]; tensor sqrt_s0_7 = sqrt(x = prev_scale_7)[name = tensor("sqrt_s0_7")]; - tensor s_t_7 = add(x = prev_scale_7, y = encoder__t_index)[name = tensor("s_t_7")]; + tensor s_t_7 = add(x = prev_scale_7, y = inner_encoder__t_index)[name = tensor("s_t_7")]; tensor sqrt_s_t_7 = sqrt(x = s_t_7)[name = tensor("sqrt_s_t_7")]; tensor qk_7_transpose_x_1 = const()[name = tensor("qk_7_transpose_x_1"), val = tensor(false)]; tensor qk_7_transpose_y_1 = const()[name = tensor("qk_7_transpose_y_1"), val = tensor(true)]; - tensor k_7 = transpose(perm = k_7_perm_0, x = var_839)[name = tensor("transpose_36")]; - tensor q_7 = transpose(perm = q_7_perm_0, x = var_831)[name = tensor("transpose_37")]; + tensor k_7 = transpose(perm = k_7_perm_0, x = var_905)[name = tensor("transpose_36")]; + tensor q_7 = transpose(perm = q_7_perm_0, x = var_897)[name = tensor("transpose_37")]; tensor qk_7 = matmul(transpose_x = qk_7_transpose_x_1, transpose_y = qk_7_transpose_y_1, x = q_7, y = k_7)[name = tensor("qk_7")]; - tensor var_855 = const()[name = tensor("op_855"), val = tensor([5, 1])]; - tensor var_856 = reshape(shape = var_855, x = sqrt_s_t_7)[name = tensor("op_856")]; - tensor M_7 = real_div(x = encoder__causal_mask, y = var_856)[name = tensor("M_7")]; - tensor var_858 = mul(x = qk_7, y = M_7)[name = tensor("op_858")]; + tensor var_921 = const()[name = tensor("op_921"), val = tensor([5, 1])]; + tensor var_922 = reshape(shape = var_921, x = sqrt_s_t_7)[name = tensor("op_922")]; + tensor M_7 = real_div(x = inner_encoder__causal_mask, y = var_922)[name = tensor("M_7")]; + tensor var_924 = mul(x = qk_7, y = M_7)[name = tensor("op_924")]; tensor inner_7_transpose_x_0 = const()[name = tensor("inner_7_transpose_x_0"), val = tensor(false)]; tensor inner_7_transpose_y_0 = const()[name = tensor("inner_7_transpose_y_0"), val = tensor(false)]; - tensor v_7 = transpose(perm = v_7_perm_0, x = var_845)[name = tensor("transpose_35")]; - tensor inner_7 = matmul(transpose_x = inner_7_transpose_x_0, transpose_y = inner_7_transpose_y_0, x = var_858, y = v_7)[name = tensor("inner_7")]; - tensor var_860_transpose_x_0 = const()[name = tensor("op_860_transpose_x_0"), val = tensor(false)]; - tensor var_860_transpose_y_0 = const()[name = tensor("op_860_transpose_y_0"), val = tensor(false)]; - tensor var_860 = matmul(transpose_x = var_860_transpose_x_0, transpose_y = var_860_transpose_y_0, x = q_7, y = prev_kv_7)[name = tensor("op_860")]; - tensor var_861 = real_div(x = sqrt_s0_7, y = sqrt_s_t_7)[name = tensor("op_861")]; - tensor var_862 = const()[name = tensor("op_862"), val = tensor([1, 1, 5, 1])]; - tensor var_863 = reshape(shape = var_862, x = var_861)[name = tensor("op_863")]; - tensor cross_7 = mul(x = var_860, y = var_863)[name = tensor("cross_7")]; + tensor v_7 = transpose(perm = v_7_perm_0, x = var_911)[name = tensor("transpose_35")]; + tensor inner_7 = matmul(transpose_x = inner_7_transpose_x_0, transpose_y = inner_7_transpose_y_0, x = var_924, y = v_7)[name = tensor("inner_7")]; + tensor var_926_transpose_x_0 = const()[name = tensor("op_926_transpose_x_0"), val = tensor(false)]; + tensor var_926_transpose_y_0 = const()[name = tensor("op_926_transpose_y_0"), val = tensor(false)]; + tensor var_926 = matmul(transpose_x = var_926_transpose_x_0, transpose_y = var_926_transpose_y_0, x = q_7, y = prev_kv_7)[name = tensor("op_926")]; + tensor var_927 = real_div(x = sqrt_s0_7, y = sqrt_s_t_7)[name = tensor("op_927")]; + tensor var_928 = const()[name = tensor("op_928"), val = tensor([1, 1, 5, 1])]; + tensor var_929 = reshape(shape = var_928, x = var_927)[name = tensor("op_929")]; + tensor cross_7 = mul(x = var_926, y = var_929)[name = tensor("cross_7")]; tensor out_19 = add(x = inner_7, y = cross_7)[name = tensor("out_19")]; - tensor var_866 = mul(x = prev_kv_7, y = sqrt_s0_7)[name = tensor("op_866")]; - tensor var_868_transpose_x_1 = const()[name = tensor("op_868_transpose_x_1"), val = tensor(true)]; - tensor var_868_transpose_y_1 = const()[name = tensor("op_868_transpose_y_1"), val = tensor(false)]; - tensor var_868 = matmul(transpose_x = var_868_transpose_x_1, transpose_y = var_868_transpose_y_1, x = k_7, y = v_7)[name = tensor("op_868")]; - tensor new_kv_unnorm_7 = add(x = var_866, y = var_868)[name = tensor("new_kv_unnorm_7")]; - tensor var_870 = const()[name = tensor("op_870"), val = tensor(0x1.4p+2)]; - tensor new_scale_7 = add(x = prev_scale_7, y = var_870)[name = tensor("new_scale_7")]; - tensor var_872 = sqrt(x = new_scale_7)[name = tensor("op_872")]; - tensor nkv_1 = real_div(x = new_kv_unnorm_7, y = var_872)[name = tensor("nkv_1")]; - tensor var_874_perm_0 = const()[name = tensor("op_874_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_932 = mul(x = prev_kv_7, y = sqrt_s0_7)[name = tensor("op_932")]; + tensor var_934_transpose_x_1 = const()[name = tensor("op_934_transpose_x_1"), val = tensor(true)]; + tensor var_934_transpose_y_1 = const()[name = tensor("op_934_transpose_y_1"), val = tensor(false)]; + tensor var_934 = matmul(transpose_x = var_934_transpose_x_1, transpose_y = var_934_transpose_y_1, x = k_7, y = v_7)[name = tensor("op_934")]; + tensor new_kv_unnorm_7 = add(x = var_932, y = var_934)[name = tensor("new_kv_unnorm_7")]; + tensor var_936 = const()[name = tensor("op_936"), val = tensor(0x1.4p+2)]; + tensor new_scale_7 = add(x = prev_scale_7, y = var_936)[name = tensor("new_scale_7")]; + tensor var_938 = sqrt(x = new_scale_7)[name = tensor("op_938")]; + tensor nkv_1 = real_div(x = new_kv_unnorm_7, y = var_938)[name = tensor("nkv_1")]; + tensor var_940_perm_0 = const()[name = tensor("op_940_perm_0"), val = tensor([0, 2, 1, 3])]; tensor out_21_axes_0 = const()[name = tensor("out_21_axes_0"), val = tensor([-1])]; - tensor var_874 = transpose(perm = var_874_perm_0, x = out_19)[name = tensor("transpose_34")]; - tensor out_21 = layer_norm(axes = out_21_axes_0, epsilon = var_18, x = var_874)[name = tensor("out_21")]; - tensor var_878 = const()[name = tensor("op_878"), val = tensor([1, 5, 256])]; - tensor out_23 = reshape(shape = var_878, x = out_21)[name = tensor("out_23")]; - tensor var_880 = silu(x = input_137)[name = tensor("op_880")]; - tensor input_139 = mul(x = var_880, y = out_23)[name = tensor("input_139")]; - tensor ret_out_7 = linear(bias = encoder_out_proj_3_bias, weight = encoder_out_proj_3_weight, x = input_139)[name = tensor("linear_34")]; - tensor x_21 = add(x = input_133, y = ret_out_7)[name = tensor("x_21")]; + tensor var_940 = transpose(perm = var_940_perm_0, x = out_19)[name = tensor("transpose_34")]; + tensor out_21 = layer_norm(axes = out_21_axes_0, epsilon = var_84, x = var_940)[name = tensor("out_21")]; + tensor var_944 = const()[name = tensor("op_944"), val = tensor([1, 5, 256])]; + tensor out_23 = reshape(shape = var_944, x = out_21)[name = tensor("out_23")]; + tensor var_946 = silu(x = input_139)[name = tensor("op_946")]; + tensor input_141 = mul(x = var_946, y = out_23)[name = tensor("input_141")]; + tensor ret_out_7 = linear(bias = inner_encoder_out_proj_3_bias, weight = inner_encoder_out_proj_3_weight, x = input_141)[name = tensor("linear_34")]; + tensor x_21 = add(x = input_135, y = ret_out_7)[name = tensor("x_21")]; tensor window_37_begin_0 = const()[name = tensor("window_37_begin_0"), val = tensor([3, 0, 0, 0])]; tensor window_37_end_0 = const()[name = tensor("window_37_end_0"), val = tensor([4, 1, 16, 256])]; tensor window_37_end_mask_0 = const()[name = tensor("window_37_end_mask_0"), val = tensor([false, true, true, true])]; tensor window_37_squeeze_mask_0 = const()[name = tensor("window_37_squeeze_mask_0"), val = tensor([true, false, false, false])]; tensor window_37 = slice_by_index(begin = window_37_begin_0, end = window_37_end_0, end_mask = window_37_end_mask_0, squeeze_mask = window_37_squeeze_mask_0, x = enc_conv_cache)[name = tensor("window_37")]; - tensor var_888_begin_0 = const()[name = tensor("op_888_begin_0"), val = tensor([0, 0, 0])]; - tensor var_888_end_0 = const()[name = tensor("op_888_end_0"), val = tensor([1, 1, 256])]; - tensor var_888_end_mask_0 = const()[name = tensor("op_888_end_mask_0"), val = tensor([true, false, true])]; - tensor var_888 = slice_by_index(begin = var_888_begin_0, end = var_888_end_0, end_mask = var_888_end_mask_0, x = x_21)[name = tensor("op_888")]; - tensor var_891_begin_0 = const()[name = tensor("op_891_begin_0"), val = tensor([0, 1, 0])]; - tensor var_891_end_0 = const()[name = tensor("op_891_end_0"), val = tensor([1, 16, 256])]; - tensor var_891_end_mask_0 = const()[name = tensor("op_891_end_mask_0"), val = tensor([true, true, true])]; - tensor var_891 = slice_by_index(begin = var_891_begin_0, end = var_891_end_0, end_mask = var_891_end_mask_0, x = window_37)[name = tensor("op_891")]; + tensor var_954_begin_0 = const()[name = tensor("op_954_begin_0"), val = tensor([0, 0, 0])]; + tensor var_954_end_0 = const()[name = tensor("op_954_end_0"), val = tensor([1, 1, 256])]; + tensor var_954_end_mask_0 = const()[name = tensor("op_954_end_mask_0"), val = tensor([true, false, true])]; + tensor var_954 = slice_by_index(begin = var_954_begin_0, end = var_954_end_0, end_mask = var_954_end_mask_0, x = x_21)[name = tensor("op_954")]; + tensor var_957_begin_0 = const()[name = tensor("op_957_begin_0"), val = tensor([0, 1, 0])]; + tensor var_957_end_0 = const()[name = tensor("op_957_end_0"), val = tensor([1, 16, 256])]; + tensor var_957_end_mask_0 = const()[name = tensor("op_957_end_mask_0"), val = tensor([true, true, true])]; + tensor var_957 = slice_by_index(begin = var_957_begin_0, end = var_957_end_0, end_mask = var_957_end_mask_0, x = window_37)[name = tensor("op_957")]; tensor window_39_interleave_0 = const()[name = tensor("window_39_interleave_0"), val = tensor(false)]; - tensor window_39 = concat(axis = var_27, interleave = window_39_interleave_0, values = (var_891, var_888))[name = tensor("window_39")]; - tensor var_896_begin_0 = const()[name = tensor("op_896_begin_0"), val = tensor([0, 1, 0])]; - tensor var_896_end_0 = const()[name = tensor("op_896_end_0"), val = tensor([1, 2, 256])]; - tensor var_896_end_mask_0 = const()[name = tensor("op_896_end_mask_0"), val = tensor([true, false, true])]; - tensor var_896 = slice_by_index(begin = var_896_begin_0, end = var_896_end_0, end_mask = var_896_end_mask_0, x = x_21)[name = tensor("op_896")]; - tensor var_899_begin_0 = const()[name = tensor("op_899_begin_0"), val = tensor([0, 1, 0])]; - tensor var_899_end_0 = const()[name = tensor("op_899_end_0"), val = tensor([1, 16, 256])]; - tensor var_899_end_mask_0 = const()[name = tensor("op_899_end_mask_0"), val = tensor([true, true, true])]; - tensor var_899 = slice_by_index(begin = var_899_begin_0, end = var_899_end_0, end_mask = var_899_end_mask_0, x = window_39)[name = tensor("op_899")]; + tensor window_39 = concat(axis = var_93, interleave = window_39_interleave_0, values = (var_957, var_954))[name = tensor("window_39")]; + tensor var_962_begin_0 = const()[name = tensor("op_962_begin_0"), val = tensor([0, 1, 0])]; + tensor var_962_end_0 = const()[name = tensor("op_962_end_0"), val = tensor([1, 2, 256])]; + tensor var_962_end_mask_0 = const()[name = tensor("op_962_end_mask_0"), val = tensor([true, false, true])]; + tensor var_962 = slice_by_index(begin = var_962_begin_0, end = var_962_end_0, end_mask = var_962_end_mask_0, x = x_21)[name = tensor("op_962")]; + tensor var_965_begin_0 = const()[name = tensor("op_965_begin_0"), val = tensor([0, 1, 0])]; + tensor var_965_end_0 = const()[name = tensor("op_965_end_0"), val = tensor([1, 16, 256])]; + tensor var_965_end_mask_0 = const()[name = tensor("op_965_end_mask_0"), val = tensor([true, true, true])]; + tensor var_965 = slice_by_index(begin = var_965_begin_0, end = var_965_end_0, end_mask = var_965_end_mask_0, x = window_39)[name = tensor("op_965")]; tensor window_41_interleave_0 = const()[name = tensor("window_41_interleave_0"), val = tensor(false)]; - tensor window_41 = concat(axis = var_27, interleave = window_41_interleave_0, values = (var_899, var_896))[name = tensor("window_41")]; - tensor var_904_begin_0 = const()[name = tensor("op_904_begin_0"), val = tensor([0, 2, 0])]; - tensor var_904_end_0 = const()[name = tensor("op_904_end_0"), val = tensor([1, 3, 256])]; - tensor var_904_end_mask_0 = const()[name = tensor("op_904_end_mask_0"), val = tensor([true, false, true])]; - tensor var_904 = slice_by_index(begin = var_904_begin_0, end = var_904_end_0, end_mask = var_904_end_mask_0, x = x_21)[name = tensor("op_904")]; - tensor var_907_begin_0 = const()[name = tensor("op_907_begin_0"), val = tensor([0, 1, 0])]; - tensor var_907_end_0 = const()[name = tensor("op_907_end_0"), val = tensor([1, 16, 256])]; - tensor var_907_end_mask_0 = const()[name = tensor("op_907_end_mask_0"), val = tensor([true, true, true])]; - tensor var_907 = slice_by_index(begin = var_907_begin_0, end = var_907_end_0, end_mask = var_907_end_mask_0, x = window_41)[name = tensor("op_907")]; + tensor window_41 = concat(axis = var_93, interleave = window_41_interleave_0, values = (var_965, var_962))[name = tensor("window_41")]; + tensor var_970_begin_0 = const()[name = tensor("op_970_begin_0"), val = tensor([0, 2, 0])]; + tensor var_970_end_0 = const()[name = tensor("op_970_end_0"), val = tensor([1, 3, 256])]; + tensor var_970_end_mask_0 = const()[name = tensor("op_970_end_mask_0"), val = tensor([true, false, true])]; + tensor var_970 = slice_by_index(begin = var_970_begin_0, end = var_970_end_0, end_mask = var_970_end_mask_0, x = x_21)[name = tensor("op_970")]; + tensor var_973_begin_0 = const()[name = tensor("op_973_begin_0"), val = tensor([0, 1, 0])]; + tensor var_973_end_0 = const()[name = tensor("op_973_end_0"), val = tensor([1, 16, 256])]; + tensor var_973_end_mask_0 = const()[name = tensor("op_973_end_mask_0"), val = tensor([true, true, true])]; + tensor var_973 = slice_by_index(begin = var_973_begin_0, end = var_973_end_0, end_mask = var_973_end_mask_0, x = window_41)[name = tensor("op_973")]; tensor window_43_interleave_0 = const()[name = tensor("window_43_interleave_0"), val = tensor(false)]; - tensor window_43 = concat(axis = var_27, interleave = window_43_interleave_0, values = (var_907, var_904))[name = tensor("window_43")]; - tensor var_912_begin_0 = const()[name = tensor("op_912_begin_0"), val = tensor([0, 3, 0])]; - tensor var_912_end_0 = const()[name = tensor("op_912_end_0"), val = tensor([1, 4, 256])]; - tensor var_912_end_mask_0 = const()[name = tensor("op_912_end_mask_0"), val = tensor([true, false, true])]; - tensor var_912 = slice_by_index(begin = var_912_begin_0, end = var_912_end_0, end_mask = var_912_end_mask_0, x = x_21)[name = tensor("op_912")]; - tensor var_915_begin_0 = const()[name = tensor("op_915_begin_0"), val = tensor([0, 1, 0])]; - tensor var_915_end_0 = const()[name = tensor("op_915_end_0"), val = tensor([1, 16, 256])]; - tensor var_915_end_mask_0 = const()[name = tensor("op_915_end_mask_0"), val = tensor([true, true, true])]; - tensor var_915 = slice_by_index(begin = var_915_begin_0, end = var_915_end_0, end_mask = var_915_end_mask_0, x = window_43)[name = tensor("op_915")]; + tensor window_43 = concat(axis = var_93, interleave = window_43_interleave_0, values = (var_973, var_970))[name = tensor("window_43")]; + tensor var_978_begin_0 = const()[name = tensor("op_978_begin_0"), val = tensor([0, 3, 0])]; + tensor var_978_end_0 = const()[name = tensor("op_978_end_0"), val = tensor([1, 4, 256])]; + tensor var_978_end_mask_0 = const()[name = tensor("op_978_end_mask_0"), val = tensor([true, false, true])]; + tensor var_978 = slice_by_index(begin = var_978_begin_0, end = var_978_end_0, end_mask = var_978_end_mask_0, x = x_21)[name = tensor("op_978")]; + tensor var_981_begin_0 = const()[name = tensor("op_981_begin_0"), val = tensor([0, 1, 0])]; + tensor var_981_end_0 = const()[name = tensor("op_981_end_0"), val = tensor([1, 16, 256])]; + tensor var_981_end_mask_0 = const()[name = tensor("op_981_end_mask_0"), val = tensor([true, true, true])]; + tensor var_981 = slice_by_index(begin = var_981_begin_0, end = var_981_end_0, end_mask = var_981_end_mask_0, x = window_43)[name = tensor("op_981")]; tensor window_45_interleave_0 = const()[name = tensor("window_45_interleave_0"), val = tensor(false)]; - tensor window_45 = concat(axis = var_27, interleave = window_45_interleave_0, values = (var_915, var_912))[name = tensor("window_45")]; - tensor var_920_begin_0 = const()[name = tensor("op_920_begin_0"), val = tensor([0, 4, 0])]; - tensor var_920_end_0 = const()[name = tensor("op_920_end_0"), val = tensor([1, 1, 256])]; - tensor var_920_end_mask_0 = const()[name = tensor("op_920_end_mask_0"), val = tensor([true, true, true])]; - tensor var_920 = slice_by_index(begin = var_920_begin_0, end = var_920_end_0, end_mask = var_920_end_mask_0, x = x_21)[name = tensor("op_920")]; - tensor var_923_begin_0 = const()[name = tensor("op_923_begin_0"), val = tensor([0, 1, 0])]; - tensor var_923_end_0 = const()[name = tensor("op_923_end_0"), val = tensor([1, 16, 256])]; - tensor var_923_end_mask_0 = const()[name = tensor("op_923_end_mask_0"), val = tensor([true, true, true])]; - tensor var_923 = slice_by_index(begin = var_923_begin_0, end = var_923_end_0, end_mask = var_923_end_mask_0, x = window_45)[name = tensor("op_923")]; + tensor window_45 = concat(axis = var_93, interleave = window_45_interleave_0, values = (var_981, var_978))[name = tensor("window_45")]; + tensor var_986_begin_0 = const()[name = tensor("op_986_begin_0"), val = tensor([0, 4, 0])]; + tensor var_986_end_0 = const()[name = tensor("op_986_end_0"), val = tensor([1, 1, 256])]; + tensor var_986_end_mask_0 = const()[name = tensor("op_986_end_mask_0"), val = tensor([true, true, true])]; + tensor var_986 = slice_by_index(begin = var_986_begin_0, end = var_986_end_0, end_mask = var_986_end_mask_0, x = x_21)[name = tensor("op_986")]; + tensor var_989_begin_0 = const()[name = tensor("op_989_begin_0"), val = tensor([0, 1, 0])]; + tensor var_989_end_0 = const()[name = tensor("op_989_end_0"), val = tensor([1, 16, 256])]; + tensor var_989_end_mask_0 = const()[name = tensor("op_989_end_mask_0"), val = tensor([true, true, true])]; + tensor var_989 = slice_by_index(begin = var_989_begin_0, end = var_989_end_0, end_mask = var_989_end_mask_0, x = window_45)[name = tensor("op_989")]; tensor window_interleave_0 = const()[name = tensor("window_interleave_0"), val = tensor(false)]; - tensor window = concat(axis = var_27, interleave = window_interleave_0, values = (var_923, var_920))[name = tensor("window")]; - tensor input_141_interleave_0 = const()[name = tensor("input_141_interleave_0"), val = tensor(false)]; - tensor input_141 = concat(axis = var_24, interleave = input_141_interleave_0, values = (window_39, window_41, window_43, window_45, window))[name = tensor("input_141")]; + tensor window = concat(axis = var_93, interleave = window_interleave_0, values = (var_989, var_986))[name = tensor("window")]; + tensor input_143_interleave_0 = const()[name = tensor("input_143_interleave_0"), val = tensor(false)]; + tensor input_143 = concat(axis = var_79, interleave = input_143_interleave_0, values = (window_39, window_41, window_43, window_45, window))[name = tensor("input_143")]; tensor x_23_axes_0 = const()[name = tensor("x_23_axes_0"), val = tensor([-1])]; - tensor x_23 = layer_norm(axes = x_23_axes_0, beta = encoder_conv_module_3_sequential_0_bias, epsilon = var_29, gamma = encoder_conv_module_3_sequential_0_weight, x = input_141)[name = tensor("x_23")]; - tensor input_143_perm_0 = const()[name = tensor("input_143_perm_0"), val = tensor([0, 2, 1])]; + tensor x_23 = layer_norm(axes = x_23_axes_0, beta = inner_encoder_conv_module_3_sequential_0_bias, epsilon = var_76, gamma = inner_encoder_conv_module_3_sequential_0_weight, x = input_143)[name = tensor("x_23")]; + tensor input_145_perm_0 = const()[name = tensor("input_145_perm_0"), val = tensor([0, 2, 1])]; tensor inputs_33_pad_type_0 = const()[name = tensor("inputs_33_pad_type_0"), val = tensor("valid")]; tensor inputs_33_strides_0 = const()[name = tensor("inputs_33_strides_0"), val = tensor([1])]; tensor inputs_33_pad_0 = const()[name = tensor("inputs_33_pad_0"), val = tensor([0, 0])]; tensor inputs_33_dilations_0 = const()[name = tensor("inputs_33_dilations_0"), val = tensor([1])]; tensor inputs_33_groups_0 = const()[name = tensor("inputs_33_groups_0"), val = tensor(1)]; - tensor input_143 = transpose(perm = input_143_perm_0, x = x_23)[name = tensor("transpose_33")]; - tensor inputs_33 = conv(bias = encoder_conv_module_3_sequential_2_conv_bias, dilations = inputs_33_dilations_0, groups = inputs_33_groups_0, pad = inputs_33_pad_0, pad_type = inputs_33_pad_type_0, strides = inputs_33_strides_0, weight = encoder_conv_module_3_sequential_2_conv_weight, x = input_143)[name = tensor("inputs_33")]; - tensor var_948_split_sizes_0 = const()[name = tensor("op_948_split_sizes_0"), val = tensor([256, 256])]; - tensor var_948_axis_0 = const()[name = tensor("op_948_axis_0"), val = tensor(1)]; - tensor var_948_0, tensor var_948_1 = split(axis = var_948_axis_0, split_sizes = var_948_split_sizes_0, x = inputs_33)[name = tensor("op_948")]; - tensor var_950 = sigmoid(x = var_948_1)[name = tensor("op_950")]; - tensor inputs_35 = mul(x = var_948_0, y = var_950)[name = tensor("inputs_35")]; + tensor input_145 = transpose(perm = input_145_perm_0, x = x_23)[name = tensor("transpose_33")]; + tensor inputs_33 = conv(bias = inner_encoder_conv_module_3_sequential_2_conv_bias, dilations = inputs_33_dilations_0, groups = inputs_33_groups_0, pad = inputs_33_pad_0, pad_type = inputs_33_pad_type_0, strides = inputs_33_strides_0, weight = inner_encoder_conv_module_3_sequential_2_conv_weight, x = input_145)[name = tensor("inputs_33")]; + tensor var_1014_split_sizes_0 = const()[name = tensor("op_1014_split_sizes_0"), val = tensor([256, 256])]; + tensor var_1014_axis_0 = const()[name = tensor("op_1014_axis_0"), val = tensor(1)]; + tensor var_1014_0, tensor var_1014_1 = split(axis = var_1014_axis_0, split_sizes = var_1014_split_sizes_0, x = inputs_33)[name = tensor("op_1014")]; + tensor var_1016 = sigmoid(x = var_1014_1)[name = tensor("op_1016")]; + tensor inputs_35 = mul(x = var_1014_0, y = var_1016)[name = tensor("inputs_35")]; tensor outputs_aug_pad_type_0 = const()[name = tensor("outputs_aug_pad_type_0"), val = tensor("custom")]; tensor outputs_aug_pad_0 = const()[name = tensor("outputs_aug_pad_0"), val = tensor([15, 15])]; tensor outputs_aug_groups_0 = const()[name = tensor("outputs_aug_groups_0"), val = tensor(256)]; tensor outputs_aug_strides_0 = const()[name = tensor("outputs_aug_strides_0"), val = tensor([1])]; tensor outputs_aug_dilations_0 = const()[name = tensor("outputs_aug_dilations_0"), val = tensor([1])]; - tensor outputs_aug = conv(dilations = outputs_aug_dilations_0, groups = outputs_aug_groups_0, pad = outputs_aug_pad_0, pad_type = outputs_aug_pad_type_0, strides = outputs_aug_strides_0, weight = encoder_conv_module_3_sequential_4_conv_weight, x = inputs_35)[name = tensor("outputs_aug")]; - tensor input_145_begin_0 = const()[name = tensor("input_145_begin_0"), val = tensor([0, 0, 0])]; - tensor input_145_end_0 = const()[name = tensor("input_145_end_0"), val = tensor([5, 256, 16])]; - tensor input_145_end_mask_0 = const()[name = tensor("input_145_end_mask_0"), val = tensor([true, true, false])]; - tensor input_145 = slice_by_index(begin = input_145_begin_0, end = input_145_end_0, end_mask = input_145_end_mask_0, x = outputs_aug)[name = tensor("input_145")]; - tensor inputs_37 = batch_norm(beta = encoder_conv_module_3_sequential_5_bias, epsilon = var_29, gamma = encoder_conv_module_3_sequential_5_weight, mean = encoder_conv_module_3_sequential_5_running_mean, variance = encoder_conv_module_3_sequential_5_running_var, x = input_145)[name = tensor("inputs_37")]; - tensor input_147 = silu(x = inputs_37)[name = tensor("input_147")]; - tensor input_149_pad_type_0 = const()[name = tensor("input_149_pad_type_0"), val = tensor("valid")]; - tensor input_149_strides_0 = const()[name = tensor("input_149_strides_0"), val = tensor([1])]; - tensor input_149_pad_0 = const()[name = tensor("input_149_pad_0"), val = tensor([0, 0])]; - tensor input_149_dilations_0 = const()[name = tensor("input_149_dilations_0"), val = tensor([1])]; - tensor input_149_groups_0 = const()[name = tensor("input_149_groups_0"), val = tensor(1)]; - tensor input_149 = conv(bias = encoder_conv_module_3_sequential_7_conv_bias, dilations = input_149_dilations_0, groups = input_149_groups_0, pad = input_149_pad_0, pad_type = input_149_pad_type_0, strides = input_149_strides_0, weight = encoder_conv_module_3_sequential_7_conv_weight, x = input_147)[name = tensor("input_149")]; + tensor outputs_aug = conv(dilations = outputs_aug_dilations_0, groups = outputs_aug_groups_0, pad = outputs_aug_pad_0, pad_type = outputs_aug_pad_type_0, strides = outputs_aug_strides_0, weight = inner_encoder_conv_module_3_sequential_4_conv_weight, x = inputs_35)[name = tensor("outputs_aug")]; + tensor input_147_begin_0 = const()[name = tensor("input_147_begin_0"), val = tensor([0, 0, 0])]; + tensor input_147_end_0 = const()[name = tensor("input_147_end_0"), val = tensor([5, 256, 16])]; + tensor input_147_end_mask_0 = const()[name = tensor("input_147_end_mask_0"), val = tensor([true, true, false])]; + tensor input_147 = slice_by_index(begin = input_147_begin_0, end = input_147_end_0, end_mask = input_147_end_mask_0, x = outputs_aug)[name = tensor("input_147")]; + tensor inputs_37 = batch_norm(beta = inner_encoder_conv_module_3_sequential_5_bias, epsilon = var_76, gamma = inner_encoder_conv_module_3_sequential_5_weight, mean = inner_encoder_conv_module_3_sequential_5_running_mean, variance = inner_encoder_conv_module_3_sequential_5_running_var, x = input_147)[name = tensor("inputs_37")]; + tensor input_149 = silu(x = inputs_37)[name = tensor("input_149")]; + tensor input_151_pad_type_0 = const()[name = tensor("input_151_pad_type_0"), val = tensor("valid")]; + tensor input_151_strides_0 = const()[name = tensor("input_151_strides_0"), val = tensor([1])]; + tensor input_151_pad_0 = const()[name = tensor("input_151_pad_0"), val = tensor([0, 0])]; + tensor input_151_dilations_0 = const()[name = tensor("input_151_dilations_0"), val = tensor([1])]; + tensor input_151_groups_0 = const()[name = tensor("input_151_groups_0"), val = tensor(1)]; + tensor input_151 = conv(bias = inner_encoder_conv_module_3_sequential_7_conv_bias, dilations = input_151_dilations_0, groups = input_151_groups_0, pad = input_151_pad_0, pad_type = input_151_pad_type_0, strides = input_151_strides_0, weight = inner_encoder_conv_module_3_sequential_7_conv_weight, x = input_149)[name = tensor("input_151")]; tensor conv_out_7_perm_0 = const()[name = tensor("conv_out_7_perm_0"), val = tensor([0, 2, 1])]; - tensor var_981_begin_0 = const()[name = tensor("op_981_begin_0"), val = tensor([0, -1, 0])]; - tensor var_981_end_0 = const()[name = tensor("op_981_end_0"), val = tensor([5, 16, 256])]; - tensor var_981_end_mask_0 = const()[name = tensor("op_981_end_mask_0"), val = tensor([true, true, true])]; - tensor conv_out_7 = transpose(perm = conv_out_7_perm_0, x = input_149)[name = tensor("transpose_32")]; - tensor var_981 = slice_by_index(begin = var_981_begin_0, end = var_981_end_0, end_mask = var_981_end_mask_0, x = conv_out_7)[name = tensor("op_981")]; - tensor var_983_perm_0 = const()[name = tensor("op_983_perm_0"), val = tensor([1, 0, 2])]; - tensor var_983 = transpose(perm = var_983_perm_0, x = var_981)[name = tensor("transpose_31")]; - tensor input_151 = add(x = x_21, y = var_983)[name = tensor("input_151")]; - tensor input_153_axes_0 = const()[name = tensor("input_153_axes_0"), val = tensor([-1])]; - tensor input_153 = layer_norm(axes = input_153_axes_0, beta = encoder_ffn2_3_module_sequential_0_bias, epsilon = var_29, gamma = encoder_ffn2_3_module_sequential_0_weight, x = input_151)[name = tensor("input_153")]; - tensor inputs = linear(bias = encoder_ffn2_3_module_sequential_1_linear_bias, weight = encoder_ffn2_3_module_sequential_1_linear_weight, x = input_153)[name = tensor("linear_35")]; - tensor input_155 = silu(x = inputs)[name = tensor("input_155")]; - tensor input_159 = linear(bias = encoder_ffn2_3_module_sequential_4_linear_bias, weight = encoder_ffn2_3_module_sequential_4_linear_weight, x = input_155)[name = tensor("linear_36")]; - tensor var_1006 = const()[name = tensor("op_1006"), val = tensor(0x1p-1)]; - tensor var_1007 = mul(x = input_159, y = var_1006)[name = tensor("op_1007")]; - tensor input_161 = add(x = var_1007, y = input_151)[name = tensor("input_161")]; + tensor var_1047_begin_0 = const()[name = tensor("op_1047_begin_0"), val = tensor([0, -1, 0])]; + tensor var_1047_end_0 = const()[name = tensor("op_1047_end_0"), val = tensor([5, 16, 256])]; + tensor var_1047_end_mask_0 = const()[name = tensor("op_1047_end_mask_0"), val = tensor([true, true, true])]; + tensor conv_out_7 = transpose(perm = conv_out_7_perm_0, x = input_151)[name = tensor("transpose_32")]; + tensor var_1047 = slice_by_index(begin = var_1047_begin_0, end = var_1047_end_0, end_mask = var_1047_end_mask_0, x = conv_out_7)[name = tensor("op_1047")]; + tensor var_1049_perm_0 = const()[name = tensor("op_1049_perm_0"), val = tensor([1, 0, 2])]; + tensor var_1049 = transpose(perm = var_1049_perm_0, x = var_1047)[name = tensor("transpose_31")]; + tensor input_153 = add(x = x_21, y = var_1049)[name = tensor("input_153")]; + tensor input_155_axes_0 = const()[name = tensor("input_155_axes_0"), val = tensor([-1])]; + tensor input_155 = layer_norm(axes = input_155_axes_0, beta = inner_encoder_ffn2_3_module_sequential_0_bias, epsilon = var_76, gamma = inner_encoder_ffn2_3_module_sequential_0_weight, x = input_153)[name = tensor("input_155")]; + tensor inputs = linear(bias = inner_encoder_ffn2_3_module_sequential_1_linear_bias, weight = inner_encoder_ffn2_3_module_sequential_1_linear_weight, x = input_155)[name = tensor("linear_35")]; + tensor input_157 = silu(x = inputs)[name = tensor("input_157")]; + tensor input_161 = linear(bias = inner_encoder_ffn2_3_module_sequential_4_linear_bias, weight = inner_encoder_ffn2_3_module_sequential_4_linear_weight, x = input_157)[name = tensor("linear_36")]; + tensor var_1072 = const()[name = tensor("op_1072"), val = tensor(0x1p-1)]; + tensor var_1073 = mul(x = input_161, y = var_1072)[name = tensor("op_1073")]; + tensor input_163 = add(x = var_1073, y = input_153)[name = tensor("input_163")]; tensor x_25_axes_0 = const()[name = tensor("x_25_axes_0"), val = tensor([-1])]; - tensor x_25 = layer_norm(axes = x_25_axes_0, beta = encoder_layer_norm_3_bias, epsilon = var_29, gamma = encoder_layer_norm_3_weight, x = input_161)[name = tensor("x_25")]; + tensor x_25 = layer_norm(axes = x_25_axes_0, beta = inner_encoder_layer_norm_3_bias, epsilon = var_76, gamma = inner_encoder_layer_norm_3_weight, x = input_163)[name = tensor("x_25")]; tensor x_ct_perm_0 = const()[name = tensor("x_ct_perm_0"), val = tensor([0, 2, 1])]; tensor cat_interleave_0 = const()[name = tensor("cat_interleave_0"), val = tensor(false)]; tensor x_ct = transpose(perm = x_ct_perm_0, x = x_25)[name = tensor("transpose_30")]; - tensor cat = concat(axis = var_21, interleave = cat_interleave_0, values = (cnn_window, x_ct))[name = tensor("cat")]; + tensor cat = concat(axis = var_81, interleave = cat_interleave_0, values = (cnn_window, x_ct))[name = tensor("cat")]; tensor conv_out_pad_type_0 = const()[name = tensor("conv_out_pad_type_0"), val = tensor("valid")]; tensor conv_out_strides_0 = const()[name = tensor("conv_out_strides_0"), val = tensor([1])]; tensor conv_out_pad_0 = const()[name = tensor("conv_out_pad_0"), val = tensor([0, 0])]; tensor conv_out_dilations_0 = const()[name = tensor("conv_out_dilations_0"), val = tensor([1])]; tensor conv_out_groups_0 = const()[name = tensor("conv_out_groups_0"), val = tensor(1)]; - tensor conv_out = conv(bias = encoder_cnn_bias, dilations = conv_out_dilations_0, groups = conv_out_groups_0, pad = conv_out_pad_0, pad_type = conv_out_pad_type_0, strides = conv_out_strides_0, weight = encoder_cnn_weight, x = cat)[name = tensor("conv_out")]; - tensor var_1025_begin_0 = const()[name = tensor("op_1025_begin_0"), val = tensor([0, 0, 5])]; - tensor var_1025_end_0 = const()[name = tensor("op_1025_end_0"), val = tensor([1, 256, 23])]; - tensor var_1025_end_mask_0 = const()[name = tensor("op_1025_end_mask_0"), val = tensor([true, true, true])]; - tensor cnn_window_new = slice_by_index(begin = var_1025_begin_0, end = var_1025_end_0, end_mask = var_1025_end_mask_0, x = cat)[name = tensor("op_1025")]; - tensor input_163_perm_0 = const()[name = tensor("input_163_perm_0"), val = tensor([0, 2, 1])]; - tensor var_1027 = const()[name = tensor("op_1027"), val = tensor([-1])]; - tensor input_163 = transpose(perm = input_163_perm_0, x = conv_out)[name = tensor("transpose_29")]; - tensor var_1028 = reduce_l2_norm(axes = var_1027, keep_dims = var_30, x = input_163)[name = tensor("op_1028")]; + tensor conv_out = conv(bias = inner_encoder_cnn_bias, dilations = conv_out_dilations_0, groups = conv_out_groups_0, pad = conv_out_pad_0, pad_type = conv_out_pad_type_0, strides = conv_out_strides_0, weight = inner_encoder_cnn_weight, x = cat)[name = tensor("conv_out")]; + tensor var_1091_begin_0 = const()[name = tensor("op_1091_begin_0"), val = tensor([0, 0, 5])]; + tensor var_1091_end_0 = const()[name = tensor("op_1091_end_0"), val = tensor([1, 256, 23])]; + tensor var_1091_end_mask_0 = const()[name = tensor("op_1091_end_mask_0"), val = tensor([true, true, true])]; + tensor cnn_window_new = slice_by_index(begin = var_1091_begin_0, end = var_1091_end_0, end_mask = var_1091_end_mask_0, x = cat)[name = tensor("op_1091")]; + tensor input_165_perm_0 = const()[name = tensor("input_165_perm_0"), val = tensor([0, 2, 1])]; + tensor var_1093 = const()[name = tensor("op_1093"), val = tensor([-1])]; + tensor input_165 = transpose(perm = input_165_perm_0, x = conv_out)[name = tensor("transpose_29")]; + tensor var_1094 = reduce_l2_norm(axes = var_1093, keep_dims = var_75, x = input_165)[name = tensor("op_1094")]; tensor const_12 = const()[name = tensor("const_12"), val = tensor(0x1.fffffep+127)]; - tensor clip_0 = clip(alpha = var_11, beta = const_12, x = var_1028)[name = tensor("clip_0")]; - tensor emb = real_div(x = input_163, y = clip_0)[name = tensor("emb")]; - tensor var_1032_axis_0 = const()[name = tensor("op_1032_axis_0"), val = tensor(0)]; - tensor enc_kv_new = stack(axis = var_1032_axis_0, values = (var_207, var_429, var_651, nkv_1))[name = tensor("op_1032")]; - tensor var_1034_axis_0 = const()[name = tensor("op_1034_axis_0"), val = tensor(0)]; - tensor enc_scale_new = stack(axis = var_1034_axis_0, values = (new_scale_1, new_scale_3, new_scale_5, new_scale_7))[name = tensor("op_1034")]; - tensor var_1036_axis_0 = const()[name = tensor("op_1036_axis_0"), val = tensor(0)]; - tensor enc_conv_cache_new = stack(axis = var_1036_axis_0, values = (window_11, window_23, window_35, window))[name = tensor("op_1036")]; - tensor var_1045 = const()[name = tensor("op_1045"), val = tensor(0x1.5798eep-27)]; - tensor var_1050 = const()[name = tensor("op_1050"), val = tensor(0x1.4f8b58p-17)]; - tensor var_1052 = const()[name = tensor("op_1052"), val = tensor(0x1.0c6f7ap-20)]; - tensor var_1053 = const()[name = tensor("op_1053"), val = tensor(true)]; - tensor var_1055 = const()[name = tensor("op_1055"), val = tensor(0x1p+0)]; - tensor var_1059 = const()[name = tensor("op_1059"), val = tensor(-1)]; - tensor var_1065 = const()[name = tensor("op_1065"), val = tensor(0)]; + tensor clip_0 = clip(alpha = var_90, beta = const_12, x = var_1094)[name = tensor("clip_0")]; + tensor emb = real_div(x = input_165, y = clip_0)[name = tensor("emb")]; + tensor var_1098_axis_0 = const()[name = tensor("op_1098_axis_0"), val = tensor(0)]; + tensor enc_kv_new = stack(axis = var_1098_axis_0, values = (var_273, var_495, var_717, nkv_1))[name = tensor("op_1098")]; + tensor var_1100_axis_0 = const()[name = tensor("op_1100_axis_0"), val = tensor(0)]; + tensor enc_scale_new = stack(axis = var_1100_axis_0, values = (new_scale_1, new_scale_3, new_scale_5, new_scale_7))[name = tensor("op_1100")]; + tensor var_1102_axis_0 = const()[name = tensor("op_1102_axis_0"), val = tensor(0)]; + tensor enc_conv_cache_new = stack(axis = var_1102_axis_0, values = (window_11, window_23, window_35, window))[name = tensor("op_1102")]; tensor pos = const()[name = tensor("pos"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(44395712)))]; - tensor var_1127_axes_0 = const()[name = tensor("op_1127_axes_0"), val = tensor([2])]; - tensor var_1127 = expand_dims(axes = var_1127_axes_0, x = emb)[name = tensor("op_1127")]; + tensor var_1170_axes_0 = const()[name = tensor("op_1170_axes_0"), val = tensor([2])]; + tensor var_1170 = expand_dims(axes = var_1170_axes_0, x = emb)[name = tensor("op_1170")]; tensor emb_exp_reps_0 = const()[name = tensor("emb_exp_reps_0"), val = tensor([1, 1, 12, 1])]; - tensor emb_exp = tile(reps = emb_exp_reps_0, x = var_1127)[name = tensor("emb_exp")]; - tensor input_165_interleave_0 = const()[name = tensor("input_165_interleave_0"), val = tensor(false)]; - tensor input_165 = concat(axis = var_1059, interleave = input_165_interleave_0, values = (emb_exp, pos))[name = tensor("input_165")]; - tensor x_27 = linear(bias = decoder_convert_bias, weight = decoder_convert_weight, x = input_165)[name = tensor("linear_37")]; - tensor var_1135_perm_0 = const()[name = tensor("op_1135_perm_0"), val = tensor([0, 2, 1, 3])]; - tensor var_1139 = const()[name = tensor("op_1139"), val = tensor([12, 5, 256])]; - tensor var_1135 = transpose(perm = var_1135_perm_0, x = x_27)[name = tensor("transpose_28")]; - tensor x_29 = reshape(shape = var_1139, x = var_1135)[name = tensor("x_29")]; + tensor emb_exp = tile(reps = emb_exp_reps_0, x = var_1170)[name = tensor("emb_exp")]; + tensor input_167_interleave_0 = const()[name = tensor("input_167_interleave_0"), val = tensor(false)]; + tensor input_167 = concat(axis = var_82, interleave = input_167_interleave_0, values = (emb_exp, pos))[name = tensor("input_167")]; + tensor x_27 = linear(bias = inner_decoder_convert_bias, weight = inner_decoder_convert_weight, x = input_167)[name = tensor("linear_37")]; + tensor var_1178_perm_0 = const()[name = tensor("op_1178_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_1182 = const()[name = tensor("op_1182"), val = tensor([12, 5, 256])]; + tensor var_1178 = transpose(perm = var_1178_perm_0, x = x_27)[name = tensor("transpose_28")]; + tensor x_29 = reshape(shape = var_1182, x = var_1178)[name = tensor("x_29")]; tensor prev_kv_9_begin_0 = const()[name = tensor("prev_kv_9_begin_0"), val = tensor([0, 0, 0, 0, 0])]; tensor prev_kv_9_end_0 = const()[name = tensor("prev_kv_9_end_0"), val = tensor([1, 12, 4, 64, 64])]; tensor prev_kv_9_end_mask_0 = const()[name = tensor("prev_kv_9_end_mask_0"), val = tensor([false, true, true, true, true])]; @@ -1023,132 +1042,132 @@ program(1.0) tensor prev_scale_9_end_mask_0 = const()[name = tensor("prev_scale_9_end_mask_0"), val = tensor([false, true])]; tensor prev_scale_9_squeeze_mask_0 = const()[name = tensor("prev_scale_9_squeeze_mask_0"), val = tensor([true, false])]; tensor prev_scale_9 = slice_by_index(begin = prev_scale_9_begin_0, end = prev_scale_9_end_0, end_mask = prev_scale_9_end_mask_0, squeeze_mask = prev_scale_9_squeeze_mask_0, x = dec_scale)[name = tensor("prev_scale_9")]; - tensor var_1147 = linear(bias = decoder_q_proj_0_bias, weight = decoder_q_proj_0_weight, x = x_29)[name = tensor("linear_38")]; - tensor var_1148 = const()[name = tensor("op_1148"), val = tensor([12, 5, 4, 64])]; - tensor var_1149 = reshape(shape = var_1148, x = var_1147)[name = tensor("op_1149")]; + tensor var_1190 = linear(bias = inner_decoder_q_proj_0_bias, weight = inner_decoder_q_proj_0_weight, x = x_29)[name = tensor("linear_38")]; + tensor var_1191 = const()[name = tensor("op_1191"), val = tensor([12, 5, 4, 64])]; + tensor var_1192 = reshape(shape = var_1191, x = var_1190)[name = tensor("op_1192")]; tensor q_9_perm_0 = const()[name = tensor("q_9_perm_0"), val = tensor([0, 2, 1, 3])]; - tensor var_1153 = linear(bias = decoder_k_proj_0_bias, weight = decoder_k_proj_0_weight, x = x_29)[name = tensor("linear_39")]; - tensor var_1154 = const()[name = tensor("op_1154"), val = tensor(0x1p-3)]; - tensor var_1155 = mul(x = var_1153, y = var_1154)[name = tensor("op_1155")]; - tensor var_1156 = const()[name = tensor("op_1156"), val = tensor([12, 5, 4, 64])]; - tensor var_1157 = reshape(shape = var_1156, x = var_1155)[name = tensor("op_1157")]; + tensor var_1196 = linear(bias = inner_decoder_k_proj_0_bias, weight = inner_decoder_k_proj_0_weight, x = x_29)[name = tensor("linear_39")]; + tensor var_1197 = const()[name = tensor("op_1197"), val = tensor(0x1p-3)]; + tensor var_1198 = mul(x = var_1196, y = var_1197)[name = tensor("op_1198")]; + tensor var_1199 = const()[name = tensor("op_1199"), val = tensor([12, 5, 4, 64])]; + tensor var_1200 = reshape(shape = var_1199, x = var_1198)[name = tensor("op_1200")]; tensor k_9_perm_0 = const()[name = tensor("k_9_perm_0"), val = tensor([0, 2, 1, 3])]; - tensor var_1161 = linear(bias = decoder_v_proj_0_bias, weight = decoder_v_proj_0_weight, x = x_29)[name = tensor("linear_40")]; - tensor var_1162 = const()[name = tensor("op_1162"), val = tensor([12, 5, 4, 64])]; - tensor var_1163 = reshape(shape = var_1162, x = var_1161)[name = tensor("op_1163")]; + tensor var_1204 = linear(bias = inner_decoder_v_proj_0_bias, weight = inner_decoder_v_proj_0_weight, x = x_29)[name = tensor("linear_40")]; + tensor var_1205 = const()[name = tensor("op_1205"), val = tensor([12, 5, 4, 64])]; + tensor var_1206 = reshape(shape = var_1205, x = var_1204)[name = tensor("op_1206")]; tensor v_9_perm_0 = const()[name = tensor("v_9_perm_0"), val = tensor([0, 2, 1, 3])]; - tensor input_169 = linear(bias = decoder_g_proj_0_bias, weight = decoder_g_proj_0_weight, x = x_29)[name = tensor("linear_41")]; + tensor input_171 = linear(bias = inner_decoder_g_proj_0_bias, weight = inner_decoder_g_proj_0_weight, x = x_29)[name = tensor("linear_41")]; tensor cumsum_mask_1_exclusive_0 = const()[name = tensor("cumsum_mask_1_exclusive_0"), val = tensor(false)]; tensor cumsum_mask_1_reverse_0 = const()[name = tensor("cumsum_mask_1_reverse_0"), val = tensor(false)]; - tensor cumsum_mask_1 = cumsum(axis = var_1065, exclusive = cumsum_mask_1_exclusive_0, reverse = cumsum_mask_1_reverse_0, x = valid_mask)[name = tensor("cumsum_mask_1")]; + tensor cumsum_mask_1 = cumsum(axis = var_79, exclusive = cumsum_mask_1_exclusive_0, reverse = cumsum_mask_1_reverse_0, x = valid_mask)[name = tensor("cumsum_mask_1")]; tensor sqrt_s0_9 = sqrt(x = prev_scale_9)[name = tensor("sqrt_s0_9")]; tensor s_t_9 = add(x = prev_scale_9, y = cumsum_mask_1)[name = tensor("s_t_9")]; tensor const_20 = const()[name = tensor("const_20"), val = tensor(0x1.fffffep+127)]; - tensor clip_1 = clip(alpha = var_1055, beta = const_20, x = s_t_9)[name = tensor("clip_1")]; + tensor clip_1 = clip(alpha = var_69, beta = const_20, x = s_t_9)[name = tensor("clip_1")]; tensor sqrt_s_t_9 = sqrt(x = clip_1)[name = tensor("sqrt_s_t_9")]; tensor qk_9_transpose_x_1 = const()[name = tensor("qk_9_transpose_x_1"), val = tensor(false)]; tensor qk_9_transpose_y_1 = const()[name = tensor("qk_9_transpose_y_1"), val = tensor(true)]; - tensor k_9 = transpose(perm = k_9_perm_0, x = var_1157)[name = tensor("transpose_26")]; - tensor q_9 = transpose(perm = q_9_perm_0, x = var_1149)[name = tensor("transpose_27")]; + tensor k_9 = transpose(perm = k_9_perm_0, x = var_1200)[name = tensor("transpose_26")]; + tensor q_9 = transpose(perm = q_9_perm_0, x = var_1192)[name = tensor("transpose_27")]; tensor qk_9 = matmul(transpose_x = qk_9_transpose_x_1, transpose_y = qk_9_transpose_y_1, x = q_9, y = k_9)[name = tensor("qk_9")]; - tensor var_1175 = const()[name = tensor("op_1175"), val = tensor([1, 5])]; - tensor var_1176 = reshape(shape = var_1175, x = valid_mask)[name = tensor("op_1176")]; - tensor causal_with_valid_1 = mul(x = decoder__causal_mask, y = var_1176)[name = tensor("causal_with_valid_1")]; - tensor var_1178 = const()[name = tensor("op_1178"), val = tensor([5, 1])]; - tensor var_1179 = reshape(shape = var_1178, x = sqrt_s_t_9)[name = tensor("op_1179")]; - tensor M_9 = real_div(x = causal_with_valid_1, y = var_1179)[name = tensor("M_9")]; - tensor var_1181 = mul(x = qk_9, y = M_9)[name = tensor("op_1181")]; + tensor var_1218 = const()[name = tensor("op_1218"), val = tensor([1, 5])]; + tensor var_1219 = reshape(shape = var_1218, x = valid_mask)[name = tensor("op_1219")]; + tensor causal_with_valid_1 = mul(x = inner_decoder__causal_mask, y = var_1219)[name = tensor("causal_with_valid_1")]; + tensor var_1221 = const()[name = tensor("op_1221"), val = tensor([5, 1])]; + tensor var_1222 = reshape(shape = var_1221, x = sqrt_s_t_9)[name = tensor("op_1222")]; + tensor M_9 = real_div(x = causal_with_valid_1, y = var_1222)[name = tensor("M_9")]; + tensor var_1224 = mul(x = qk_9, y = M_9)[name = tensor("op_1224")]; tensor inner_9_transpose_x_0 = const()[name = tensor("inner_9_transpose_x_0"), val = tensor(false)]; tensor inner_9_transpose_y_0 = const()[name = tensor("inner_9_transpose_y_0"), val = tensor(false)]; - tensor v_9 = transpose(perm = v_9_perm_0, x = var_1163)[name = tensor("transpose_25")]; - tensor inner_9 = matmul(transpose_x = inner_9_transpose_x_0, transpose_y = inner_9_transpose_y_0, x = var_1181, y = v_9)[name = tensor("inner_9")]; - tensor var_1183_transpose_x_0 = const()[name = tensor("op_1183_transpose_x_0"), val = tensor(false)]; - tensor var_1183_transpose_y_0 = const()[name = tensor("op_1183_transpose_y_0"), val = tensor(false)]; - tensor var_1183 = matmul(transpose_x = var_1183_transpose_x_0, transpose_y = var_1183_transpose_y_0, x = q_9, y = prev_kv_9)[name = tensor("op_1183")]; - tensor var_1184 = real_div(x = sqrt_s0_9, y = sqrt_s_t_9)[name = tensor("op_1184")]; - tensor var_1185 = const()[name = tensor("op_1185"), val = tensor([1, 1, 5, 1])]; - tensor var_1186 = reshape(shape = var_1185, x = var_1184)[name = tensor("op_1186")]; - tensor cross_9 = mul(x = var_1183, y = var_1186)[name = tensor("cross_9")]; + tensor v_9 = transpose(perm = v_9_perm_0, x = var_1206)[name = tensor("transpose_25")]; + tensor inner_9 = matmul(transpose_x = inner_9_transpose_x_0, transpose_y = inner_9_transpose_y_0, x = var_1224, y = v_9)[name = tensor("inner_9")]; + tensor var_1226_transpose_x_0 = const()[name = tensor("op_1226_transpose_x_0"), val = tensor(false)]; + tensor var_1226_transpose_y_0 = const()[name = tensor("op_1226_transpose_y_0"), val = tensor(false)]; + tensor var_1226 = matmul(transpose_x = var_1226_transpose_x_0, transpose_y = var_1226_transpose_y_0, x = q_9, y = prev_kv_9)[name = tensor("op_1226")]; + tensor var_1227 = real_div(x = sqrt_s0_9, y = sqrt_s_t_9)[name = tensor("op_1227")]; + tensor var_1228 = const()[name = tensor("op_1228"), val = tensor([1, 1, 5, 1])]; + tensor var_1229 = reshape(shape = var_1228, x = var_1227)[name = tensor("op_1229")]; + tensor cross_9 = mul(x = var_1226, y = var_1229)[name = tensor("cross_9")]; tensor out_25 = add(x = inner_9, y = cross_9)[name = tensor("out_25")]; - tensor var_1189 = const()[name = tensor("op_1189"), val = tensor([1, 1, 5, 1])]; - tensor var_1190 = reshape(shape = var_1189, x = valid_mask)[name = tensor("op_1190")]; - tensor v_masked_1 = mul(x = v_9, y = var_1190)[name = tensor("v_masked_1")]; - tensor var_1192 = mul(x = prev_kv_9, y = sqrt_s0_9)[name = tensor("op_1192")]; - tensor var_1194_transpose_x_1 = const()[name = tensor("op_1194_transpose_x_1"), val = tensor(true)]; - tensor var_1194_transpose_y_1 = const()[name = tensor("op_1194_transpose_y_1"), val = tensor(false)]; - tensor var_1194 = matmul(transpose_x = var_1194_transpose_x_1, transpose_y = var_1194_transpose_y_1, x = k_9, y = v_masked_1)[name = tensor("op_1194")]; - tensor new_kv_unnorm_9 = add(x = var_1192, y = var_1194)[name = tensor("new_kv_unnorm_9")]; - tensor var_1196_keep_dims_0 = const()[name = tensor("op_1196_keep_dims_0"), val = tensor(false)]; - tensor var_1196 = reduce_sum(keep_dims = var_1196_keep_dims_0, x = valid_mask)[name = tensor("op_1196")]; - tensor var_1197 = const()[name = tensor("op_1197"), val = tensor([1])]; - tensor var_1198 = reshape(shape = var_1197, x = var_1196)[name = tensor("op_1198")]; - tensor new_scale_9 = add(x = prev_scale_9, y = var_1198)[name = tensor("new_scale_9")]; + tensor var_1232 = const()[name = tensor("op_1232"), val = tensor([1, 1, 5, 1])]; + tensor var_1233 = reshape(shape = var_1232, x = valid_mask)[name = tensor("op_1233")]; + tensor v_masked_1 = mul(x = v_9, y = var_1233)[name = tensor("v_masked_1")]; + tensor var_1235 = mul(x = prev_kv_9, y = sqrt_s0_9)[name = tensor("op_1235")]; + tensor var_1237_transpose_x_1 = const()[name = tensor("op_1237_transpose_x_1"), val = tensor(true)]; + tensor var_1237_transpose_y_1 = const()[name = tensor("op_1237_transpose_y_1"), val = tensor(false)]; + tensor var_1237 = matmul(transpose_x = var_1237_transpose_x_1, transpose_y = var_1237_transpose_y_1, x = k_9, y = v_masked_1)[name = tensor("op_1237")]; + tensor new_kv_unnorm_9 = add(x = var_1235, y = var_1237)[name = tensor("new_kv_unnorm_9")]; + tensor var_1239_keep_dims_0 = const()[name = tensor("op_1239_keep_dims_0"), val = tensor(false)]; + tensor var_1239 = reduce_sum(keep_dims = var_1239_keep_dims_0, x = valid_mask)[name = tensor("op_1239")]; + tensor var_1240 = const()[name = tensor("op_1240"), val = tensor([1])]; + tensor var_1241 = reshape(shape = var_1240, x = var_1239)[name = tensor("op_1241")]; + tensor new_scale_9 = add(x = prev_scale_9, y = var_1241)[name = tensor("new_scale_9")]; tensor const_21 = const()[name = tensor("const_21"), val = tensor(0x1.fffffep+127)]; - tensor clip_2 = clip(alpha = var_1055, beta = const_21, x = new_scale_9)[name = tensor("clip_2")]; + tensor clip_2 = clip(alpha = var_69, beta = const_21, x = new_scale_9)[name = tensor("clip_2")]; tensor sqrt_new_scale_1 = sqrt(x = clip_2)[name = tensor("sqrt_new_scale_1")]; - tensor var_1202 = real_div(x = new_kv_unnorm_9, y = sqrt_new_scale_1)[name = tensor("op_1202")]; - tensor var_1203_perm_0 = const()[name = tensor("op_1203_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_1245 = real_div(x = new_kv_unnorm_9, y = sqrt_new_scale_1)[name = tensor("op_1245")]; + tensor var_1246_perm_0 = const()[name = tensor("op_1246_perm_0"), val = tensor([0, 2, 1, 3])]; tensor out_27_axes_0 = const()[name = tensor("out_27_axes_0"), val = tensor([-1])]; - tensor var_1203 = transpose(perm = var_1203_perm_0, x = out_25)[name = tensor("transpose_24")]; - tensor out_27 = layer_norm(axes = out_27_axes_0, epsilon = var_1052, x = var_1203)[name = tensor("out_27")]; - tensor var_1207 = const()[name = tensor("op_1207"), val = tensor([12, 5, 256])]; - tensor out_29 = reshape(shape = var_1207, x = out_27)[name = tensor("out_29")]; - tensor var_1209 = silu(x = input_169)[name = tensor("op_1209")]; - tensor input_171 = mul(x = var_1209, y = out_29)[name = tensor("input_171")]; - tensor ret_out_9 = linear(bias = decoder_out_proj_0_bias, weight = decoder_out_proj_0_weight, x = input_171)[name = tensor("linear_42")]; - tensor input_173 = add(x = x_29, y = ret_out_9)[name = tensor("input_173")]; + tensor var_1246 = transpose(perm = var_1246_perm_0, x = out_25)[name = tensor("transpose_24")]; + tensor out_27 = layer_norm(axes = out_27_axes_0, epsilon = var_84, x = var_1246)[name = tensor("out_27")]; + tensor var_1250 = const()[name = tensor("op_1250"), val = tensor([12, 5, 256])]; + tensor out_29 = reshape(shape = var_1250, x = out_27)[name = tensor("out_29")]; + tensor var_1252 = silu(x = input_171)[name = tensor("op_1252")]; + tensor input_173 = mul(x = var_1252, y = out_29)[name = tensor("input_173")]; + tensor ret_out_9 = linear(bias = inner_decoder_out_proj_0_bias, weight = inner_decoder_out_proj_0_weight, x = input_173)[name = tensor("linear_42")]; + tensor input_175 = add(x = x_29, y = ret_out_9)[name = tensor("input_175")]; tensor xt_1_axes_0 = const()[name = tensor("xt_1_axes_0"), val = tensor([-1])]; - tensor xt_1 = layer_norm(axes = xt_1_axes_0, beta = decoder_norm11_0_bias, epsilon = var_1050, gamma = decoder_norm11_0_weight, x = input_173)[name = tensor("xt_1")]; - tensor var_1219 = const()[name = tensor("op_1219"), val = tensor([1, 12, 5, 256])]; - tensor var_1220 = reshape(shape = var_1219, x = xt_1)[name = tensor("op_1220")]; - tensor var_1221_perm_0 = const()[name = tensor("op_1221_perm_0"), val = tensor([0, 2, 1, 3])]; - tensor var_1224 = const()[name = tensor("op_1224"), val = tensor([5, 12, 256])]; - tensor var_1221 = transpose(perm = var_1221_perm_0, x = var_1220)[name = tensor("transpose_23")]; - tensor query_1 = reshape(shape = var_1224, x = var_1221)[name = tensor("query_1")]; + tensor xt_1 = layer_norm(axes = xt_1_axes_0, beta = inner_decoder_norm11_0_bias, epsilon = var_76, gamma = inner_decoder_norm11_0_weight, x = input_175)[name = tensor("xt_1")]; + tensor var_1262 = const()[name = tensor("op_1262"), val = tensor([1, 12, 5, 256])]; + tensor var_1263 = reshape(shape = var_1262, x = xt_1)[name = tensor("op_1263")]; + tensor var_1264_perm_0 = const()[name = tensor("op_1264_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_1267 = const()[name = tensor("op_1267"), val = tensor([5, 12, 256])]; + tensor var_1264 = transpose(perm = var_1264_perm_0, x = var_1263)[name = tensor("transpose_23")]; + tensor query_1 = reshape(shape = var_1267, x = var_1264)[name = tensor("query_1")]; tensor query_3_perm_0 = const()[name = tensor("query_3_perm_0"), val = tensor([1, 0, 2])]; tensor query_3 = transpose(perm = query_3_perm_0, x = query_1)[name = tensor("transpose_22")]; - tensor var_1247 = linear(bias = decoder_self_attn2_0_in_proj_bias, weight = decoder_self_attn2_0_in_proj_weight, x = query_3)[name = tensor("linear_43")]; + tensor var_1290 = linear(bias = inner_decoder_self_attn2_0_in_proj_bias, weight = inner_decoder_self_attn2_0_in_proj_weight, x = query_3)[name = tensor("linear_43")]; tensor concat_1 = const()[name = tensor("concat_1"), val = tensor([12, 5, 3, 256])]; - tensor var_1249 = reshape(shape = concat_1, x = var_1247)[name = tensor("op_1249")]; - tensor var_1250_axes_0 = const()[name = tensor("op_1250_axes_0"), val = tensor([0])]; - tensor var_1250 = expand_dims(axes = var_1250_axes_0, x = var_1249)[name = tensor("op_1250")]; - tensor var_1251_perm_0 = const()[name = tensor("op_1251_perm_0"), val = tensor([-2, 1, 2, 0, 4])]; - tensor var_1252_axes_0 = const()[name = tensor("op_1252_axes_0"), val = tensor([-2])]; - tensor var_1251 = transpose(perm = var_1251_perm_0, x = var_1250)[name = tensor("transpose_21")]; - tensor var_1252 = squeeze(axes = var_1252_axes_0, x = var_1251)[name = tensor("op_1252")]; + tensor var_1292 = reshape(shape = concat_1, x = var_1290)[name = tensor("op_1292")]; + tensor var_1293_axes_0 = const()[name = tensor("op_1293_axes_0"), val = tensor([0])]; + tensor var_1293 = expand_dims(axes = var_1293_axes_0, x = var_1292)[name = tensor("op_1293")]; + tensor var_1294_perm_0 = const()[name = tensor("op_1294_perm_0"), val = tensor([-2, 1, 2, 0, 4])]; + tensor var_1295_axes_0 = const()[name = tensor("op_1295_axes_0"), val = tensor([-2])]; + tensor var_1294 = transpose(perm = var_1294_perm_0, x = var_1293)[name = tensor("transpose_21")]; + tensor var_1295 = squeeze(axes = var_1295_axes_0, x = var_1294)[name = tensor("op_1295")]; tensor q_11_begin_0 = const()[name = tensor("q_11_begin_0"), val = tensor([0, 0, 0, 0])]; tensor q_11_end_0 = const()[name = tensor("q_11_end_0"), val = tensor([1, 12, 5, 256])]; tensor q_11_end_mask_0 = const()[name = tensor("q_11_end_mask_0"), val = tensor([false, true, true, true])]; tensor q_11_squeeze_mask_0 = const()[name = tensor("q_11_squeeze_mask_0"), val = tensor([true, false, false, false])]; - tensor q_11 = slice_by_index(begin = q_11_begin_0, end = q_11_end_0, end_mask = q_11_end_mask_0, squeeze_mask = q_11_squeeze_mask_0, x = var_1252)[name = tensor("q_11")]; + tensor q_11 = slice_by_index(begin = q_11_begin_0, end = q_11_end_0, end_mask = q_11_end_mask_0, squeeze_mask = q_11_squeeze_mask_0, x = var_1295)[name = tensor("q_11")]; tensor k_11_begin_0 = const()[name = tensor("k_11_begin_0"), val = tensor([1, 0, 0, 0])]; tensor k_11_end_0 = const()[name = tensor("k_11_end_0"), val = tensor([2, 12, 5, 256])]; tensor k_11_end_mask_0 = const()[name = tensor("k_11_end_mask_0"), val = tensor([false, true, true, true])]; tensor k_11_squeeze_mask_0 = const()[name = tensor("k_11_squeeze_mask_0"), val = tensor([true, false, false, false])]; - tensor k_11 = slice_by_index(begin = k_11_begin_0, end = k_11_end_0, end_mask = k_11_end_mask_0, squeeze_mask = k_11_squeeze_mask_0, x = var_1252)[name = tensor("k_11")]; + tensor k_11 = slice_by_index(begin = k_11_begin_0, end = k_11_end_0, end_mask = k_11_end_mask_0, squeeze_mask = k_11_squeeze_mask_0, x = var_1295)[name = tensor("k_11")]; tensor v_11_begin_0 = const()[name = tensor("v_11_begin_0"), val = tensor([2, 0, 0, 0])]; tensor v_11_end_0 = const()[name = tensor("v_11_end_0"), val = tensor([3, 12, 5, 256])]; tensor v_11_end_mask_0 = const()[name = tensor("v_11_end_mask_0"), val = tensor([false, true, true, true])]; tensor v_11_squeeze_mask_0 = const()[name = tensor("v_11_squeeze_mask_0"), val = tensor([true, false, false, false])]; - tensor v_11 = slice_by_index(begin = v_11_begin_0, end = v_11_end_0, end_mask = v_11_end_mask_0, squeeze_mask = v_11_squeeze_mask_0, x = var_1252)[name = tensor("v_11")]; - tensor var_1260 = const()[name = tensor("op_1260"), val = tensor([12, 20, 64])]; - tensor var_1261 = reshape(shape = var_1260, x = q_11)[name = tensor("op_1261")]; + tensor v_11 = slice_by_index(begin = v_11_begin_0, end = v_11_end_0, end_mask = v_11_end_mask_0, squeeze_mask = v_11_squeeze_mask_0, x = var_1295)[name = tensor("v_11")]; + tensor var_1303 = const()[name = tensor("op_1303"), val = tensor([12, 20, 64])]; + tensor var_1304 = reshape(shape = var_1303, x = q_11)[name = tensor("op_1304")]; tensor q_13_perm_0 = const()[name = tensor("q_13_perm_0"), val = tensor([1, 0, 2])]; - tensor var_1267 = const()[name = tensor("op_1267"), val = tensor([12, 20, 64])]; - tensor var_1268 = reshape(shape = var_1267, x = k_11)[name = tensor("op_1268")]; + tensor var_1310 = const()[name = tensor("op_1310"), val = tensor([12, 20, 64])]; + tensor var_1311 = reshape(shape = var_1310, x = k_11)[name = tensor("op_1311")]; tensor k_13_perm_0 = const()[name = tensor("k_13_perm_0"), val = tensor([1, 0, 2])]; - tensor var_1274 = const()[name = tensor("op_1274"), val = tensor([12, 20, 64])]; - tensor var_1275 = reshape(shape = var_1274, x = v_11)[name = tensor("op_1275")]; + tensor var_1317 = const()[name = tensor("op_1317"), val = tensor([12, 20, 64])]; + tensor var_1318 = reshape(shape = var_1317, x = v_11)[name = tensor("op_1318")]; tensor v_13_perm_0 = const()[name = tensor("v_13_perm_0"), val = tensor([1, 0, 2])]; - tensor var_1278 = const()[name = tensor("op_1278"), val = tensor([5, 4, 12, 64])]; - tensor q_13 = transpose(perm = q_13_perm_0, x = var_1261)[name = tensor("transpose_20")]; - tensor q_15 = reshape(shape = var_1278, x = q_13)[name = tensor("q_15")]; - tensor var_1280 = const()[name = tensor("op_1280"), val = tensor([5, 4, 12, 64])]; - tensor k_13 = transpose(perm = k_13_perm_0, x = var_1268)[name = tensor("transpose_19")]; - tensor k_15 = reshape(shape = var_1280, x = k_13)[name = tensor("k_15")]; - tensor var_1282 = const()[name = tensor("op_1282"), val = tensor([5, 4, 12, 64])]; - tensor v_13 = transpose(perm = v_13_perm_0, x = var_1275)[name = tensor("transpose_18")]; - tensor v_15 = reshape(shape = var_1282, x = v_13)[name = tensor("v_15")]; + tensor var_1321 = const()[name = tensor("op_1321"), val = tensor([5, 4, 12, 64])]; + tensor q_13 = transpose(perm = q_13_perm_0, x = var_1304)[name = tensor("transpose_20")]; + tensor q_15 = reshape(shape = var_1321, x = q_13)[name = tensor("q_15")]; + tensor var_1323 = const()[name = tensor("op_1323"), val = tensor([5, 4, 12, 64])]; + tensor k_13 = transpose(perm = k_13_perm_0, x = var_1311)[name = tensor("transpose_19")]; + tensor k_15 = reshape(shape = var_1323, x = k_13)[name = tensor("k_15")]; + tensor var_1325 = const()[name = tensor("op_1325"), val = tensor([5, 4, 12, 64])]; + tensor v_13 = transpose(perm = v_13_perm_0, x = var_1318)[name = tensor("transpose_18")]; + tensor v_15 = reshape(shape = var_1325, x = v_13)[name = tensor("v_15")]; tensor mul_1_y_0 = const()[name = tensor("mul_1_y_0"), val = tensor(0x1p-3)]; tensor mul_1 = mul(x = q_15, y = mul_1_y_0)[name = tensor("mul_1")]; tensor matmul_0_transpose_y_0 = const()[name = tensor("matmul_0_transpose_y_0"), val = tensor(true)]; @@ -1159,30 +1178,30 @@ program(1.0) tensor attn_output_1_transpose_x_0 = const()[name = tensor("attn_output_1_transpose_x_0"), val = tensor(false)]; tensor attn_output_1_transpose_y_0 = const()[name = tensor("attn_output_1_transpose_y_0"), val = tensor(false)]; tensor attn_output_1 = matmul(transpose_x = attn_output_1_transpose_x_0, transpose_y = attn_output_1_transpose_y_0, x = softmax_0, y = v_15)[name = tensor("attn_output_1")]; - tensor var_1285 = const()[name = tensor("op_1285"), val = tensor([2, 0, 1, 3])]; - tensor var_1290 = const()[name = tensor("op_1290"), val = tensor([60, 256])]; - tensor var_1286 = transpose(perm = var_1285, x = attn_output_1)[name = tensor("transpose_17")]; - tensor attn_output_3 = reshape(shape = var_1290, x = var_1286)[name = tensor("attn_output_3")]; - tensor attn_output_5 = linear(bias = decoder_self_attn2_0_out_proj_bias, weight = decoder_self_attn2_0_out_proj_weight, x = attn_output_3)[name = tensor("linear_44")]; - tensor var_1294 = const()[name = tensor("op_1294"), val = tensor([12, 5, 256])]; - tensor attn_output_7 = reshape(shape = var_1294, x = attn_output_5)[name = tensor("attn_output_7")]; + tensor var_1328 = const()[name = tensor("op_1328"), val = tensor([2, 0, 1, 3])]; + tensor var_1333 = const()[name = tensor("op_1333"), val = tensor([60, 256])]; + tensor var_1329 = transpose(perm = var_1328, x = attn_output_1)[name = tensor("transpose_17")]; + tensor attn_output_3 = reshape(shape = var_1333, x = var_1329)[name = tensor("attn_output_3")]; + tensor attn_output_5 = linear(bias = inner_decoder_self_attn2_0_out_proj_bias, weight = inner_decoder_self_attn2_0_out_proj_weight, x = attn_output_3)[name = tensor("linear_44")]; + tensor var_1337 = const()[name = tensor("op_1337"), val = tensor([12, 5, 256])]; + tensor attn_output_7 = reshape(shape = var_1337, x = attn_output_5)[name = tensor("attn_output_7")]; tensor sa2_out_1_perm_0 = const()[name = tensor("sa2_out_1_perm_0"), val = tensor([1, 0, 2])]; tensor sa2_out_1 = transpose(perm = sa2_out_1_perm_0, x = attn_output_7)[name = tensor("transpose_16")]; - tensor input_175 = add(x = query_1, y = sa2_out_1)[name = tensor("input_175")]; - tensor input_177_axes_0 = const()[name = tensor("input_177_axes_0"), val = tensor([-1])]; - tensor input_177 = layer_norm(axes = input_177_axes_0, beta = decoder_norm21_0_bias, epsilon = var_1050, gamma = decoder_norm21_0_weight, x = input_175)[name = tensor("input_177")]; - tensor input_179 = linear(bias = decoder_linear1_0_bias, weight = decoder_linear1_0_weight, x = input_177)[name = tensor("linear_45")]; - tensor input_181 = relu(x = input_179)[name = tensor("input_181")]; - tensor ffn_out_1 = linear(bias = decoder_linear2_0_bias, weight = decoder_linear2_0_weight, x = input_181)[name = tensor("linear_46")]; - tensor input_183 = add(x = input_177, y = ffn_out_1)[name = tensor("input_183")]; + tensor input_177 = add(x = query_1, y = sa2_out_1)[name = tensor("input_177")]; + tensor input_179_axes_0 = const()[name = tensor("input_179_axes_0"), val = tensor([-1])]; + tensor input_179 = layer_norm(axes = input_179_axes_0, beta = inner_decoder_norm21_0_bias, epsilon = var_76, gamma = inner_decoder_norm21_0_weight, x = input_177)[name = tensor("input_179")]; + tensor input_181 = linear(bias = inner_decoder_linear1_0_bias, weight = inner_decoder_linear1_0_weight, x = input_179)[name = tensor("linear_45")]; + tensor input_183 = relu(x = input_181)[name = tensor("input_183")]; + tensor ffn_out_1 = linear(bias = inner_decoder_linear2_0_bias, weight = inner_decoder_linear2_0_weight, x = input_183)[name = tensor("linear_46")]; + tensor input_185 = add(x = input_179, y = ffn_out_1)[name = tensor("input_185")]; tensor xt_3_axes_0 = const()[name = tensor("xt_3_axes_0"), val = tensor([-1])]; - tensor xt_3 = layer_norm(axes = xt_3_axes_0, beta = decoder_norm22_0_bias, epsilon = var_1050, gamma = decoder_norm22_0_weight, x = input_183)[name = tensor("xt_3")]; - tensor var_1314 = const()[name = tensor("op_1314"), val = tensor([1, 5, 12, 256])]; - tensor x_31 = reshape(shape = var_1314, x = xt_3)[name = tensor("x_31")]; - tensor var_1316_perm_0 = const()[name = tensor("op_1316_perm_0"), val = tensor([0, 2, 1, 3])]; - tensor var_1320 = const()[name = tensor("op_1320"), val = tensor([12, 5, 256])]; - tensor var_1316 = transpose(perm = var_1316_perm_0, x = x_31)[name = tensor("transpose_15")]; - tensor x = reshape(shape = var_1320, x = var_1316)[name = tensor("x")]; + tensor xt_3 = layer_norm(axes = xt_3_axes_0, beta = inner_decoder_norm22_0_bias, epsilon = var_76, gamma = inner_decoder_norm22_0_weight, x = input_185)[name = tensor("xt_3")]; + tensor var_1357 = const()[name = tensor("op_1357"), val = tensor([1, 5, 12, 256])]; + tensor x_31 = reshape(shape = var_1357, x = xt_3)[name = tensor("x_31")]; + tensor var_1359_perm_0 = const()[name = tensor("op_1359_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_1363 = const()[name = tensor("op_1363"), val = tensor([12, 5, 256])]; + tensor var_1359 = transpose(perm = var_1359_perm_0, x = x_31)[name = tensor("transpose_15")]; + tensor x = reshape(shape = var_1363, x = var_1359)[name = tensor("x")]; tensor prev_kv_begin_0 = const()[name = tensor("prev_kv_begin_0"), val = tensor([1, 0, 0, 0, 0])]; tensor prev_kv_end_0 = const()[name = tensor("prev_kv_end_0"), val = tensor([2, 12, 4, 64, 64])]; tensor prev_kv_end_mask_0 = const()[name = tensor("prev_kv_end_mask_0"), val = tensor([false, true, true, true, true])]; @@ -1193,120 +1212,120 @@ program(1.0) tensor prev_scale_end_mask_0 = const()[name = tensor("prev_scale_end_mask_0"), val = tensor([false, true])]; tensor prev_scale_squeeze_mask_0 = const()[name = tensor("prev_scale_squeeze_mask_0"), val = tensor([true, false])]; tensor prev_scale = slice_by_index(begin = prev_scale_begin_0, end = prev_scale_end_0, end_mask = prev_scale_end_mask_0, squeeze_mask = prev_scale_squeeze_mask_0, x = dec_scale)[name = tensor("prev_scale")]; - tensor var_1328 = linear(bias = decoder_q_proj_1_bias, weight = decoder_q_proj_1_weight, x = x)[name = tensor("linear_47")]; - tensor var_1329 = const()[name = tensor("op_1329"), val = tensor([12, 5, 4, 64])]; - tensor var_1330 = reshape(shape = var_1329, x = var_1328)[name = tensor("op_1330")]; + tensor var_1371 = linear(bias = inner_decoder_q_proj_1_bias, weight = inner_decoder_q_proj_1_weight, x = x)[name = tensor("linear_47")]; + tensor var_1372 = const()[name = tensor("op_1372"), val = tensor([12, 5, 4, 64])]; + tensor var_1373 = reshape(shape = var_1372, x = var_1371)[name = tensor("op_1373")]; tensor q_17_perm_0 = const()[name = tensor("q_17_perm_0"), val = tensor([0, 2, 1, 3])]; - tensor var_1334 = linear(bias = decoder_k_proj_1_bias, weight = decoder_k_proj_1_weight, x = x)[name = tensor("linear_48")]; - tensor var_1335 = const()[name = tensor("op_1335"), val = tensor(0x1p-3)]; - tensor var_1336 = mul(x = var_1334, y = var_1335)[name = tensor("op_1336")]; - tensor var_1337 = const()[name = tensor("op_1337"), val = tensor([12, 5, 4, 64])]; - tensor var_1338 = reshape(shape = var_1337, x = var_1336)[name = tensor("op_1338")]; + tensor var_1377 = linear(bias = inner_decoder_k_proj_1_bias, weight = inner_decoder_k_proj_1_weight, x = x)[name = tensor("linear_48")]; + tensor var_1378 = const()[name = tensor("op_1378"), val = tensor(0x1p-3)]; + tensor var_1379 = mul(x = var_1377, y = var_1378)[name = tensor("op_1379")]; + tensor var_1380 = const()[name = tensor("op_1380"), val = tensor([12, 5, 4, 64])]; + tensor var_1381 = reshape(shape = var_1380, x = var_1379)[name = tensor("op_1381")]; tensor k_17_perm_0 = const()[name = tensor("k_17_perm_0"), val = tensor([0, 2, 1, 3])]; - tensor var_1342 = linear(bias = decoder_v_proj_1_bias, weight = decoder_v_proj_1_weight, x = x)[name = tensor("linear_49")]; - tensor var_1343 = const()[name = tensor("op_1343"), val = tensor([12, 5, 4, 64])]; - tensor var_1344 = reshape(shape = var_1343, x = var_1342)[name = tensor("op_1344")]; + tensor var_1385 = linear(bias = inner_decoder_v_proj_1_bias, weight = inner_decoder_v_proj_1_weight, x = x)[name = tensor("linear_49")]; + tensor var_1386 = const()[name = tensor("op_1386"), val = tensor([12, 5, 4, 64])]; + tensor var_1387 = reshape(shape = var_1386, x = var_1385)[name = tensor("op_1387")]; tensor v_17_perm_0 = const()[name = tensor("v_17_perm_0"), val = tensor([0, 2, 1, 3])]; - tensor input_187 = linear(bias = decoder_g_proj_1_bias, weight = decoder_g_proj_1_weight, x = x)[name = tensor("linear_50")]; + tensor input_189 = linear(bias = inner_decoder_g_proj_1_bias, weight = inner_decoder_g_proj_1_weight, x = x)[name = tensor("linear_50")]; tensor sqrt_s0 = sqrt(x = prev_scale)[name = tensor("sqrt_s0")]; tensor s_t = add(x = prev_scale, y = cumsum_mask_1)[name = tensor("s_t")]; tensor const_32 = const()[name = tensor("const_32"), val = tensor(0x1.fffffep+127)]; - tensor clip_3 = clip(alpha = var_1055, beta = const_32, x = s_t)[name = tensor("clip_3")]; + tensor clip_3 = clip(alpha = var_69, beta = const_32, x = s_t)[name = tensor("clip_3")]; tensor sqrt_s_t = sqrt(x = clip_3)[name = tensor("sqrt_s_t")]; tensor qk_transpose_x_1 = const()[name = tensor("qk_transpose_x_1"), val = tensor(false)]; tensor qk_transpose_y_1 = const()[name = tensor("qk_transpose_y_1"), val = tensor(true)]; - tensor k_17 = transpose(perm = k_17_perm_0, x = var_1338)[name = tensor("transpose_13")]; - tensor q_17 = transpose(perm = q_17_perm_0, x = var_1330)[name = tensor("transpose_14")]; + tensor k_17 = transpose(perm = k_17_perm_0, x = var_1381)[name = tensor("transpose_13")]; + tensor q_17 = transpose(perm = q_17_perm_0, x = var_1373)[name = tensor("transpose_14")]; tensor qk = matmul(transpose_x = qk_transpose_x_1, transpose_y = qk_transpose_y_1, x = q_17, y = k_17)[name = tensor("qk")]; - tensor var_1359 = const()[name = tensor("op_1359"), val = tensor([5, 1])]; - tensor var_1360 = reshape(shape = var_1359, x = sqrt_s_t)[name = tensor("op_1360")]; - tensor M = real_div(x = causal_with_valid_1, y = var_1360)[name = tensor("M")]; - tensor var_1362 = mul(x = qk, y = M)[name = tensor("op_1362")]; - tensor inner_transpose_x_0 = const()[name = tensor("inner_transpose_x_0"), val = tensor(false)]; - tensor inner_transpose_y_0 = const()[name = tensor("inner_transpose_y_0"), val = tensor(false)]; - tensor v_17 = transpose(perm = v_17_perm_0, x = var_1344)[name = tensor("transpose_12")]; - tensor inner = matmul(transpose_x = inner_transpose_x_0, transpose_y = inner_transpose_y_0, x = var_1362, y = v_17)[name = tensor("inner")]; - tensor var_1364_transpose_x_0 = const()[name = tensor("op_1364_transpose_x_0"), val = tensor(false)]; - tensor var_1364_transpose_y_0 = const()[name = tensor("op_1364_transpose_y_0"), val = tensor(false)]; - tensor var_1364 = matmul(transpose_x = var_1364_transpose_x_0, transpose_y = var_1364_transpose_y_0, x = q_17, y = prev_kv)[name = tensor("op_1364")]; - tensor var_1365 = real_div(x = sqrt_s0, y = sqrt_s_t)[name = tensor("op_1365")]; - tensor var_1366 = const()[name = tensor("op_1366"), val = tensor([1, 1, 5, 1])]; - tensor var_1367 = reshape(shape = var_1366, x = var_1365)[name = tensor("op_1367")]; - tensor cross = mul(x = var_1364, y = var_1367)[name = tensor("cross")]; - tensor out_31 = add(x = inner, y = cross)[name = tensor("out_31")]; - tensor v_masked = mul(x = v_17, y = var_1190)[name = tensor("v_masked")]; - tensor var_1373 = mul(x = prev_kv, y = sqrt_s0)[name = tensor("op_1373")]; - tensor var_1375_transpose_x_1 = const()[name = tensor("op_1375_transpose_x_1"), val = tensor(true)]; - tensor var_1375_transpose_y_1 = const()[name = tensor("op_1375_transpose_y_1"), val = tensor(false)]; - tensor var_1375 = matmul(transpose_x = var_1375_transpose_x_1, transpose_y = var_1375_transpose_y_1, x = k_17, y = v_masked)[name = tensor("op_1375")]; - tensor new_kv_unnorm = add(x = var_1373, y = var_1375)[name = tensor("new_kv_unnorm")]; - tensor new_scale = add(x = prev_scale, y = var_1198)[name = tensor("new_scale")]; + tensor var_1402 = const()[name = tensor("op_1402"), val = tensor([5, 1])]; + tensor var_1403 = reshape(shape = var_1402, x = sqrt_s_t)[name = tensor("op_1403")]; + tensor M = real_div(x = causal_with_valid_1, y = var_1403)[name = tensor("M")]; + tensor var_1405 = mul(x = qk, y = M)[name = tensor("op_1405")]; + tensor inner_11_transpose_x_0 = const()[name = tensor("inner_11_transpose_x_0"), val = tensor(false)]; + tensor inner_11_transpose_y_0 = const()[name = tensor("inner_11_transpose_y_0"), val = tensor(false)]; + tensor v_17 = transpose(perm = v_17_perm_0, x = var_1387)[name = tensor("transpose_12")]; + tensor inner_11 = matmul(transpose_x = inner_11_transpose_x_0, transpose_y = inner_11_transpose_y_0, x = var_1405, y = v_17)[name = tensor("inner_11")]; + tensor var_1407_transpose_x_0 = const()[name = tensor("op_1407_transpose_x_0"), val = tensor(false)]; + tensor var_1407_transpose_y_0 = const()[name = tensor("op_1407_transpose_y_0"), val = tensor(false)]; + tensor var_1407 = matmul(transpose_x = var_1407_transpose_x_0, transpose_y = var_1407_transpose_y_0, x = q_17, y = prev_kv)[name = tensor("op_1407")]; + tensor var_1408 = real_div(x = sqrt_s0, y = sqrt_s_t)[name = tensor("op_1408")]; + tensor var_1409 = const()[name = tensor("op_1409"), val = tensor([1, 1, 5, 1])]; + tensor var_1410 = reshape(shape = var_1409, x = var_1408)[name = tensor("op_1410")]; + tensor cross = mul(x = var_1407, y = var_1410)[name = tensor("cross")]; + tensor out_31 = add(x = inner_11, y = cross)[name = tensor("out_31")]; + tensor v_masked = mul(x = v_17, y = var_1233)[name = tensor("v_masked")]; + tensor var_1416 = mul(x = prev_kv, y = sqrt_s0)[name = tensor("op_1416")]; + tensor var_1418_transpose_x_1 = const()[name = tensor("op_1418_transpose_x_1"), val = tensor(true)]; + tensor var_1418_transpose_y_1 = const()[name = tensor("op_1418_transpose_y_1"), val = tensor(false)]; + tensor var_1418 = matmul(transpose_x = var_1418_transpose_x_1, transpose_y = var_1418_transpose_y_1, x = k_17, y = v_masked)[name = tensor("op_1418")]; + tensor new_kv_unnorm = add(x = var_1416, y = var_1418)[name = tensor("new_kv_unnorm")]; + tensor new_scale = add(x = prev_scale, y = var_1241)[name = tensor("new_scale")]; tensor const_33 = const()[name = tensor("const_33"), val = tensor(0x1.fffffep+127)]; - tensor clip_4 = clip(alpha = var_1055, beta = const_33, x = new_scale)[name = tensor("clip_4")]; + tensor clip_4 = clip(alpha = var_69, beta = const_33, x = new_scale)[name = tensor("clip_4")]; tensor sqrt_new_scale = sqrt(x = clip_4)[name = tensor("sqrt_new_scale")]; tensor nkv = real_div(x = new_kv_unnorm, y = sqrt_new_scale)[name = tensor("nkv")]; - tensor var_1384_perm_0 = const()[name = tensor("op_1384_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_1427_perm_0 = const()[name = tensor("op_1427_perm_0"), val = tensor([0, 2, 1, 3])]; tensor out_33_axes_0 = const()[name = tensor("out_33_axes_0"), val = tensor([-1])]; - tensor var_1384 = transpose(perm = var_1384_perm_0, x = out_31)[name = tensor("transpose_11")]; - tensor out_33 = layer_norm(axes = out_33_axes_0, epsilon = var_1052, x = var_1384)[name = tensor("out_33")]; - tensor var_1388 = const()[name = tensor("op_1388"), val = tensor([12, 5, 256])]; - tensor out = reshape(shape = var_1388, x = out_33)[name = tensor("out")]; - tensor var_1390 = silu(x = input_187)[name = tensor("op_1390")]; - tensor input_189 = mul(x = var_1390, y = out)[name = tensor("input_189")]; - tensor ret_out = linear(bias = decoder_out_proj_1_bias, weight = decoder_out_proj_1_weight, x = input_189)[name = tensor("linear_51")]; - tensor input_191 = add(x = x, y = ret_out)[name = tensor("input_191")]; + tensor var_1427 = transpose(perm = var_1427_perm_0, x = out_31)[name = tensor("transpose_11")]; + tensor out_33 = layer_norm(axes = out_33_axes_0, epsilon = var_84, x = var_1427)[name = tensor("out_33")]; + tensor var_1431 = const()[name = tensor("op_1431"), val = tensor([12, 5, 256])]; + tensor out = reshape(shape = var_1431, x = out_33)[name = tensor("out")]; + tensor var_1433 = silu(x = input_189)[name = tensor("op_1433")]; + tensor input_191 = mul(x = var_1433, y = out)[name = tensor("input_191")]; + tensor ret_out = linear(bias = inner_decoder_out_proj_1_bias, weight = inner_decoder_out_proj_1_weight, x = input_191)[name = tensor("linear_51")]; + tensor input_193 = add(x = x, y = ret_out)[name = tensor("input_193")]; tensor xt_5_axes_0 = const()[name = tensor("xt_5_axes_0"), val = tensor([-1])]; - tensor xt_5 = layer_norm(axes = xt_5_axes_0, beta = decoder_norm11_1_bias, epsilon = var_1050, gamma = decoder_norm11_1_weight, x = input_191)[name = tensor("xt_5")]; - tensor var_1400 = const()[name = tensor("op_1400"), val = tensor([1, 12, 5, 256])]; - tensor var_1401 = reshape(shape = var_1400, x = xt_5)[name = tensor("op_1401")]; - tensor var_1402_perm_0 = const()[name = tensor("op_1402_perm_0"), val = tensor([0, 2, 1, 3])]; - tensor var_1405 = const()[name = tensor("op_1405"), val = tensor([5, 12, 256])]; - tensor var_1402 = transpose(perm = var_1402_perm_0, x = var_1401)[name = tensor("transpose_10")]; - tensor query_5 = reshape(shape = var_1405, x = var_1402)[name = tensor("query_5")]; + tensor xt_5 = layer_norm(axes = xt_5_axes_0, beta = inner_decoder_norm11_1_bias, epsilon = var_76, gamma = inner_decoder_norm11_1_weight, x = input_193)[name = tensor("xt_5")]; + tensor var_1443 = const()[name = tensor("op_1443"), val = tensor([1, 12, 5, 256])]; + tensor var_1444 = reshape(shape = var_1443, x = xt_5)[name = tensor("op_1444")]; + tensor var_1445_perm_0 = const()[name = tensor("op_1445_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_1448 = const()[name = tensor("op_1448"), val = tensor([5, 12, 256])]; + tensor var_1445 = transpose(perm = var_1445_perm_0, x = var_1444)[name = tensor("transpose_10")]; + tensor query_5 = reshape(shape = var_1448, x = var_1445)[name = tensor("query_5")]; tensor query_perm_0 = const()[name = tensor("query_perm_0"), val = tensor([1, 0, 2])]; tensor query = transpose(perm = query_perm_0, x = query_5)[name = tensor("transpose_9")]; - tensor var_1428 = linear(bias = decoder_self_attn2_1_in_proj_bias, weight = decoder_self_attn2_1_in_proj_weight, x = query)[name = tensor("linear_52")]; + tensor var_1471 = linear(bias = inner_decoder_self_attn2_1_in_proj_bias, weight = inner_decoder_self_attn2_1_in_proj_weight, x = query)[name = tensor("linear_52")]; tensor concat_2 = const()[name = tensor("concat_2"), val = tensor([12, 5, 3, 256])]; - tensor var_1430 = reshape(shape = concat_2, x = var_1428)[name = tensor("op_1430")]; - tensor var_1431_axes_0 = const()[name = tensor("op_1431_axes_0"), val = tensor([0])]; - tensor var_1431 = expand_dims(axes = var_1431_axes_0, x = var_1430)[name = tensor("op_1431")]; - tensor var_1432_perm_0 = const()[name = tensor("op_1432_perm_0"), val = tensor([-2, 1, 2, 0, 4])]; - tensor var_1433_axes_0 = const()[name = tensor("op_1433_axes_0"), val = tensor([-2])]; - tensor var_1432 = transpose(perm = var_1432_perm_0, x = var_1431)[name = tensor("transpose_8")]; - tensor var_1433 = squeeze(axes = var_1433_axes_0, x = var_1432)[name = tensor("op_1433")]; + tensor var_1473 = reshape(shape = concat_2, x = var_1471)[name = tensor("op_1473")]; + tensor var_1474_axes_0 = const()[name = tensor("op_1474_axes_0"), val = tensor([0])]; + tensor var_1474 = expand_dims(axes = var_1474_axes_0, x = var_1473)[name = tensor("op_1474")]; + tensor var_1475_perm_0 = const()[name = tensor("op_1475_perm_0"), val = tensor([-2, 1, 2, 0, 4])]; + tensor var_1476_axes_0 = const()[name = tensor("op_1476_axes_0"), val = tensor([-2])]; + tensor var_1475 = transpose(perm = var_1475_perm_0, x = var_1474)[name = tensor("transpose_8")]; + tensor var_1476 = squeeze(axes = var_1476_axes_0, x = var_1475)[name = tensor("op_1476")]; tensor q_19_begin_0 = const()[name = tensor("q_19_begin_0"), val = tensor([0, 0, 0, 0])]; tensor q_19_end_0 = const()[name = tensor("q_19_end_0"), val = tensor([1, 12, 5, 256])]; tensor q_19_end_mask_0 = const()[name = tensor("q_19_end_mask_0"), val = tensor([false, true, true, true])]; tensor q_19_squeeze_mask_0 = const()[name = tensor("q_19_squeeze_mask_0"), val = tensor([true, false, false, false])]; - tensor q_19 = slice_by_index(begin = q_19_begin_0, end = q_19_end_0, end_mask = q_19_end_mask_0, squeeze_mask = q_19_squeeze_mask_0, x = var_1433)[name = tensor("q_19")]; + tensor q_19 = slice_by_index(begin = q_19_begin_0, end = q_19_end_0, end_mask = q_19_end_mask_0, squeeze_mask = q_19_squeeze_mask_0, x = var_1476)[name = tensor("q_19")]; tensor k_19_begin_0 = const()[name = tensor("k_19_begin_0"), val = tensor([1, 0, 0, 0])]; tensor k_19_end_0 = const()[name = tensor("k_19_end_0"), val = tensor([2, 12, 5, 256])]; tensor k_19_end_mask_0 = const()[name = tensor("k_19_end_mask_0"), val = tensor([false, true, true, true])]; tensor k_19_squeeze_mask_0 = const()[name = tensor("k_19_squeeze_mask_0"), val = tensor([true, false, false, false])]; - tensor k_19 = slice_by_index(begin = k_19_begin_0, end = k_19_end_0, end_mask = k_19_end_mask_0, squeeze_mask = k_19_squeeze_mask_0, x = var_1433)[name = tensor("k_19")]; + tensor k_19 = slice_by_index(begin = k_19_begin_0, end = k_19_end_0, end_mask = k_19_end_mask_0, squeeze_mask = k_19_squeeze_mask_0, x = var_1476)[name = tensor("k_19")]; tensor v_19_begin_0 = const()[name = tensor("v_19_begin_0"), val = tensor([2, 0, 0, 0])]; tensor v_19_end_0 = const()[name = tensor("v_19_end_0"), val = tensor([3, 12, 5, 256])]; tensor v_19_end_mask_0 = const()[name = tensor("v_19_end_mask_0"), val = tensor([false, true, true, true])]; tensor v_19_squeeze_mask_0 = const()[name = tensor("v_19_squeeze_mask_0"), val = tensor([true, false, false, false])]; - tensor v_19 = slice_by_index(begin = v_19_begin_0, end = v_19_end_0, end_mask = v_19_end_mask_0, squeeze_mask = v_19_squeeze_mask_0, x = var_1433)[name = tensor("v_19")]; - tensor var_1441 = const()[name = tensor("op_1441"), val = tensor([12, 20, 64])]; - tensor var_1442 = reshape(shape = var_1441, x = q_19)[name = tensor("op_1442")]; + tensor v_19 = slice_by_index(begin = v_19_begin_0, end = v_19_end_0, end_mask = v_19_end_mask_0, squeeze_mask = v_19_squeeze_mask_0, x = var_1476)[name = tensor("v_19")]; + tensor var_1484 = const()[name = tensor("op_1484"), val = tensor([12, 20, 64])]; + tensor var_1485 = reshape(shape = var_1484, x = q_19)[name = tensor("op_1485")]; tensor q_21_perm_0 = const()[name = tensor("q_21_perm_0"), val = tensor([1, 0, 2])]; - tensor var_1448 = const()[name = tensor("op_1448"), val = tensor([12, 20, 64])]; - tensor var_1449 = reshape(shape = var_1448, x = k_19)[name = tensor("op_1449")]; + tensor var_1491 = const()[name = tensor("op_1491"), val = tensor([12, 20, 64])]; + tensor var_1492 = reshape(shape = var_1491, x = k_19)[name = tensor("op_1492")]; tensor k_21_perm_0 = const()[name = tensor("k_21_perm_0"), val = tensor([1, 0, 2])]; - tensor var_1455 = const()[name = tensor("op_1455"), val = tensor([12, 20, 64])]; - tensor var_1456 = reshape(shape = var_1455, x = v_19)[name = tensor("op_1456")]; + tensor var_1498 = const()[name = tensor("op_1498"), val = tensor([12, 20, 64])]; + tensor var_1499 = reshape(shape = var_1498, x = v_19)[name = tensor("op_1499")]; tensor v_21_perm_0 = const()[name = tensor("v_21_perm_0"), val = tensor([1, 0, 2])]; - tensor var_1459 = const()[name = tensor("op_1459"), val = tensor([5, 4, 12, 64])]; - tensor q_21 = transpose(perm = q_21_perm_0, x = var_1442)[name = tensor("transpose_7")]; - tensor q = reshape(shape = var_1459, x = q_21)[name = tensor("q")]; - tensor var_1461 = const()[name = tensor("op_1461"), val = tensor([5, 4, 12, 64])]; - tensor k_21 = transpose(perm = k_21_perm_0, x = var_1449)[name = tensor("transpose_6")]; - tensor k = reshape(shape = var_1461, x = k_21)[name = tensor("k")]; - tensor var_1463 = const()[name = tensor("op_1463"), val = tensor([5, 4, 12, 64])]; - tensor v_21 = transpose(perm = v_21_perm_0, x = var_1456)[name = tensor("transpose_5")]; - tensor v = reshape(shape = var_1463, x = v_21)[name = tensor("v")]; + tensor var_1502 = const()[name = tensor("op_1502"), val = tensor([5, 4, 12, 64])]; + tensor q_21 = transpose(perm = q_21_perm_0, x = var_1485)[name = tensor("transpose_7")]; + tensor q = reshape(shape = var_1502, x = q_21)[name = tensor("q")]; + tensor var_1504 = const()[name = tensor("op_1504"), val = tensor([5, 4, 12, 64])]; + tensor k_21 = transpose(perm = k_21_perm_0, x = var_1492)[name = tensor("transpose_6")]; + tensor k = reshape(shape = var_1504, x = k_21)[name = tensor("k")]; + tensor var_1506 = const()[name = tensor("op_1506"), val = tensor([5, 4, 12, 64])]; + tensor v_21 = transpose(perm = v_21_perm_0, x = var_1499)[name = tensor("transpose_5")]; + tensor v = reshape(shape = var_1506, x = v_21)[name = tensor("v")]; tensor mul_3_y_0 = const()[name = tensor("mul_3_y_0"), val = tensor(0x1p-3)]; tensor mul_3 = mul(x = q, y = mul_3_y_0)[name = tensor("mul_3")]; tensor matmul_1_transpose_y_0 = const()[name = tensor("matmul_1_transpose_y_0"), val = tensor(true)]; @@ -1317,36 +1336,36 @@ program(1.0) tensor attn_output_9_transpose_x_0 = const()[name = tensor("attn_output_9_transpose_x_0"), val = tensor(false)]; tensor attn_output_9_transpose_y_0 = const()[name = tensor("attn_output_9_transpose_y_0"), val = tensor(false)]; tensor attn_output_9 = matmul(transpose_x = attn_output_9_transpose_x_0, transpose_y = attn_output_9_transpose_y_0, x = softmax_1, y = v)[name = tensor("attn_output_9")]; - tensor var_1466 = const()[name = tensor("op_1466"), val = tensor([2, 0, 1, 3])]; - tensor var_1471 = const()[name = tensor("op_1471"), val = tensor([60, 256])]; - tensor var_1467 = transpose(perm = var_1466, x = attn_output_9)[name = tensor("transpose_4")]; - tensor attn_output_11 = reshape(shape = var_1471, x = var_1467)[name = tensor("attn_output_11")]; - tensor attn_output_13 = linear(bias = decoder_self_attn2_1_out_proj_bias, weight = decoder_self_attn2_1_out_proj_weight, x = attn_output_11)[name = tensor("linear_53")]; - tensor var_1475 = const()[name = tensor("op_1475"), val = tensor([12, 5, 256])]; - tensor attn_output = reshape(shape = var_1475, x = attn_output_13)[name = tensor("attn_output")]; + tensor var_1509 = const()[name = tensor("op_1509"), val = tensor([2, 0, 1, 3])]; + tensor var_1514 = const()[name = tensor("op_1514"), val = tensor([60, 256])]; + tensor var_1510 = transpose(perm = var_1509, x = attn_output_9)[name = tensor("transpose_4")]; + tensor attn_output_11 = reshape(shape = var_1514, x = var_1510)[name = tensor("attn_output_11")]; + tensor attn_output_13 = linear(bias = inner_decoder_self_attn2_1_out_proj_bias, weight = inner_decoder_self_attn2_1_out_proj_weight, x = attn_output_11)[name = tensor("linear_53")]; + tensor var_1518 = const()[name = tensor("op_1518"), val = tensor([12, 5, 256])]; + tensor attn_output = reshape(shape = var_1518, x = attn_output_13)[name = tensor("attn_output")]; tensor sa2_out_perm_0 = const()[name = tensor("sa2_out_perm_0"), val = tensor([1, 0, 2])]; tensor sa2_out = transpose(perm = sa2_out_perm_0, x = attn_output)[name = tensor("transpose_3")]; - tensor input_193 = add(x = query_5, y = sa2_out)[name = tensor("input_193")]; - tensor input_195_axes_0 = const()[name = tensor("input_195_axes_0"), val = tensor([-1])]; - tensor input_195 = layer_norm(axes = input_195_axes_0, beta = decoder_norm21_1_bias, epsilon = var_1050, gamma = decoder_norm21_1_weight, x = input_193)[name = tensor("input_195")]; - tensor input_197 = linear(bias = decoder_linear1_1_bias, weight = decoder_linear1_1_weight, x = input_195)[name = tensor("linear_54")]; - tensor input_199 = relu(x = input_197)[name = tensor("input_199")]; - tensor ffn_out = linear(bias = decoder_linear2_1_bias, weight = decoder_linear2_1_weight, x = input_199)[name = tensor("linear_55")]; - tensor input_201 = add(x = input_195, y = ffn_out)[name = tensor("input_201")]; + tensor input_195 = add(x = query_5, y = sa2_out)[name = tensor("input_195")]; + tensor input_197_axes_0 = const()[name = tensor("input_197_axes_0"), val = tensor([-1])]; + tensor input_197 = layer_norm(axes = input_197_axes_0, beta = inner_decoder_norm21_1_bias, epsilon = var_76, gamma = inner_decoder_norm21_1_weight, x = input_195)[name = tensor("input_197")]; + tensor input_199 = linear(bias = inner_decoder_linear1_1_bias, weight = inner_decoder_linear1_1_weight, x = input_197)[name = tensor("linear_54")]; + tensor input_201 = relu(x = input_199)[name = tensor("input_201")]; + tensor ffn_out = linear(bias = inner_decoder_linear2_1_bias, weight = inner_decoder_linear2_1_weight, x = input_201)[name = tensor("linear_55")]; + tensor input_203 = add(x = input_197, y = ffn_out)[name = tensor("input_203")]; tensor xt_axes_0 = const()[name = tensor("xt_axes_0"), val = tensor([-1])]; - tensor xt = layer_norm(axes = xt_axes_0, beta = decoder_norm22_1_bias, epsilon = var_1050, gamma = decoder_norm22_1_weight, x = input_201)[name = tensor("xt")]; - tensor var_1495 = const()[name = tensor("op_1495"), val = tensor([1, 5, 12, 256])]; - tensor input = reshape(shape = var_1495, x = xt)[name = tensor("input")]; - tensor var_1497 = const()[name = tensor("op_1497"), val = tensor([-1])]; - tensor var_1498 = reduce_l2_norm(axes = var_1497, keep_dims = var_1053, x = input)[name = tensor("op_1498")]; + tensor xt = layer_norm(axes = xt_axes_0, beta = inner_decoder_norm22_1_bias, epsilon = var_76, gamma = inner_decoder_norm22_1_weight, x = input_203)[name = tensor("xt")]; + tensor var_1538 = const()[name = tensor("op_1538"), val = tensor([1, 5, 12, 256])]; + tensor input = reshape(shape = var_1538, x = xt)[name = tensor("input")]; + tensor var_1540 = const()[name = tensor("op_1540"), val = tensor([-1])]; + tensor var_1541 = reduce_l2_norm(axes = var_1540, keep_dims = var_75, x = input)[name = tensor("op_1541")]; tensor const_42 = const()[name = tensor("const_42"), val = tensor(0x1.fffffep+127)]; - tensor clip_5 = clip(alpha = var_1045, beta = const_42, x = var_1498)[name = tensor("clip_5")]; - tensor var_1500 = real_div(x = input, y = clip_5)[name = tensor("op_1500")]; + tensor clip_5 = clip(alpha = var_90, beta = const_42, x = var_1541)[name = tensor("clip_5")]; + tensor var_1543 = real_div(x = input, y = clip_5)[name = tensor("op_1543")]; tensor concat_6 = const()[name = tensor("concat_6"), val = tensor([5, 1, 256])]; tensor reshape_0 = reshape(shape = concat_6, x = emb)[name = tensor("reshape_0")]; tensor transpose_1_perm_0 = const()[name = tensor("transpose_1_perm_0"), val = tensor([0, 1, 3, 2])]; tensor concat_7 = const()[name = tensor("concat_7"), val = tensor([5, 256, 12])]; - tensor transpose_1 = transpose(perm = transpose_1_perm_0, x = var_1500)[name = tensor("transpose_2")]; + tensor transpose_1 = transpose(perm = transpose_1_perm_0, x = var_1543)[name = tensor("transpose_2")]; tensor reshape_1 = reshape(shape = concat_7, x = transpose_1)[name = tensor("reshape_1")]; tensor matmul_2_transpose_x_0 = const()[name = tensor("matmul_2_transpose_x_0"), val = tensor(false)]; tensor matmul_2_transpose_y_0 = const()[name = tensor("matmul_2_transpose_y_0"), val = tensor(false)]; @@ -1357,10 +1376,10 @@ program(1.0) tensor output_end_0 = const()[name = tensor("output_end_0"), val = tensor([1, 5, 11])]; tensor output_end_mask_0 = const()[name = tensor("output_end_mask_0"), val = tensor([true, true, false])]; tensor output = slice_by_index(begin = output_begin_0, end = output_end_0, end_mask = output_end_mask_0, x = reshape_2)[name = tensor("output")]; - tensor probs = sigmoid(x = output)[name = tensor("op_1504")]; - tensor var_1506_axis_0 = const()[name = tensor("op_1506_axis_0"), val = tensor(0)]; - tensor dec_kv_new = stack(axis = var_1506_axis_0, values = (var_1202, nkv))[name = tensor("op_1506")]; - tensor var_1508_axis_0 = const()[name = tensor("op_1508_axis_0"), val = tensor(0)]; - tensor dec_scale_new = stack(axis = var_1508_axis_0, values = (new_scale_9, new_scale))[name = tensor("op_1508")]; + tensor probs = sigmoid(x = output)[name = tensor("op_1547")]; + tensor var_1549_axis_0 = const()[name = tensor("op_1549_axis_0"), val = tensor(0)]; + tensor dec_kv_new = stack(axis = var_1549_axis_0, values = (var_1245, nkv))[name = tensor("op_1549")]; + tensor var_1551_axis_0 = const()[name = tensor("op_1551_axis_0"), val = tensor(0)]; + tensor dec_scale_new = stack(axis = var_1551_axis_0, values = (new_scale_9, new_scale))[name = tensor("op_1551")]; } -> (probs, enc_kv_new, enc_scale_new, enc_conv_cache_new, cnn_window_new, dec_kv_new, dec_scale_new); } \ No newline at end of file diff --git a/optimized/dih2/500ms/ls_eend_dih2_500ms.mlpackage/Data/com.apple.CoreML/model.mlmodel b/optimized/dih2/500ms/ls_eend_dih2_500ms.mlpackage/Data/com.apple.CoreML/model.mlmodel index ee2f5685e43d7adb38c9c1a4e1b73168d29549bc..8d7849f201b9a56c345f9d633052694a7179a893 100644 --- a/optimized/dih2/500ms/ls_eend_dih2_500ms.mlpackage/Data/com.apple.CoreML/model.mlmodel +++ b/optimized/dih2/500ms/ls_eend_dih2_500ms.mlpackage/Data/com.apple.CoreML/model.mlmodel @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:c600d793bc84760911328536382b994a486396bf5f73c044403146818f2ba2f1 -size 196627 +oid sha256:a2c020978aa574f4e94c5c4eb17591bdfd5a7c47b821971904d55db85631b0dc +size 203227 diff --git a/optimized/dih2/500ms/ls_eend_dih2_500ms.mlpackage/Manifest.json b/optimized/dih2/500ms/ls_eend_dih2_500ms.mlpackage/Manifest.json index 9b05f7f5a11617791e1ce810a13372eb75f0afe2..be0e3204e5fd671e217172b8c6e0f13f9de50722 100644 --- a/optimized/dih2/500ms/ls_eend_dih2_500ms.mlpackage/Manifest.json +++ b/optimized/dih2/500ms/ls_eend_dih2_500ms.mlpackage/Manifest.json @@ -1,18 +1,18 @@ { "fileFormatVersion": "1.0.0", "itemInfoEntries": { - "4275B100-4E00-4F6A-9426-B138B856397B": { - "author": "com.apple.CoreML", - "description": "CoreML Model Specification", - "name": "model.mlmodel", - "path": "com.apple.CoreML/model.mlmodel" - }, - "DBD9F036-EA3D-4076-9E45-DF3A8F67CEED": { + "7AB5BB1E-EBD5-49AC-8966-C82191C767B6": { "author": "com.apple.CoreML", "description": "CoreML Model Weights", "name": "weights", "path": "com.apple.CoreML/weights" + }, + "F7EB0522-7F8B-4B11-955A-E1CF6D14C11C": { + "author": "com.apple.CoreML", + "description": "CoreML Model Specification", + "name": "model.mlmodel", + "path": "com.apple.CoreML/model.mlmodel" } }, - "rootModelIdentifier": "4275B100-4E00-4F6A-9426-B138B856397B" + "rootModelIdentifier": "F7EB0522-7F8B-4B11-955A-E1CF6D14C11C" } diff --git a/optimized/dih3/100ms/ls_eend_dih3_100ms.mlmodelc/analytics/coremldata.bin b/optimized/dih3/100ms/ls_eend_dih3_100ms.mlmodelc/analytics/coremldata.bin index 009646e23619a556a10590ffca6efa31fb7f6eeb..86bc9045c3e712049d0ae18400c2da6a70347117 100644 --- a/optimized/dih3/100ms/ls_eend_dih3_100ms.mlmodelc/analytics/coremldata.bin +++ b/optimized/dih3/100ms/ls_eend_dih3_100ms.mlmodelc/analytics/coremldata.bin @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:23998be748af73a6dec0a0dced0634a6a32bf25edbeb9b6f24938033cbb5bcbf +oid sha256:eae41bbb03511ff0e04bb217b278b49e67a9744724c28ac2c3b1ecbb6a719544 size 243 diff --git a/optimized/dih3/100ms/ls_eend_dih3_100ms.mlmodelc/coremldata.bin b/optimized/dih3/100ms/ls_eend_dih3_100ms.mlmodelc/coremldata.bin index c569cfb25eae428b97226bb2c29697a5083561fb..a0a5315c87eb589f6b299c4b786fe45a45a95486 100644 --- a/optimized/dih3/100ms/ls_eend_dih3_100ms.mlmodelc/coremldata.bin +++ b/optimized/dih3/100ms/ls_eend_dih3_100ms.mlmodelc/coremldata.bin @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:868aa7ddffd7f41be3a3dc214b15d2e85acbbceb685c24d452f5c4629170b995 -size 1310 +oid sha256:ca07a132288cc2acbaee7e034d18c3a9d66bd5be22617b52e9e812d160268ccf +size 1413 diff --git a/optimized/dih3/100ms/ls_eend_dih3_100ms.mlmodelc/metadata.json b/optimized/dih3/100ms/ls_eend_dih3_100ms.mlmodelc/metadata.json index 74e418759f83a521fd3e5c06595da4efc184dc64..55ea7f0524227f02cd802073192368a90c729142 100644 --- a/optimized/dih3/100ms/ls_eend_dih3_100ms.mlmodelc/metadata.json +++ b/optimized/dih3/100ms/ls_eend_dih3_100ms.mlmodelc/metadata.json @@ -1,7 +1,7 @@ [ { "metadataOutputVersion" : "3.0", - "shortDescription" : "LS-EEND DIHARD III streaming diarizer (pipeline, T=1, max_speakers=10)", + "shortDescription" : "LS-EEND DIHARD III streaming diarizer (pipeline, T=1, max_speakers=10, layout=raw_mel, cu=all)", "outputSchema" : [ { "hasShapeFlexibility" : "0", @@ -81,12 +81,12 @@ "specificationVersion" : 8, "mlProgramOperationTypeHistogram" : { "Ios16.reduceL2Norm" : 2, - "Ios17.reshape" : 66, + "Ios17.reshape" : 67, "Ios16.softmax" : 2, "Ios17.matmul" : 29, "Ios17.transpose" : 57, "Split" : 4, - "Ios17.expandDims" : 3, + "Ios17.expandDims" : 4, "Ios17.add" : 46, "Ios16.sigmoid" : 5, "Ios17.sliceByIndex" : 36, @@ -128,9 +128,9 @@ "hasShapeFlexibility" : "0", "isOptional" : "0", "dataType" : "Float32", - "formattedType" : "MultiArray (Float32 1 × 1 × 345)", + "formattedType" : "MultiArray (Float32 1 × 15 × 23)", "shortDescription" : "", - "shape" : "[1, 1, 345]", + "shape" : "[1, 15, 23]", "name" : "features", "type" : "MultiArray" }, @@ -206,8 +206,8 @@ } ], "userDefinedMetadata" : { - "com.github.apple.coremltools.conversion_date" : "2026-04-16", - "config" : "{\"model_name\": \"dih3\", \"model_label\": \"DIHARD III\", \"variant\": \"pipeline\", \"chunk_size\": 1, \"step_duration_ms\": 100, \"frame_duration_ms\": 100, \"frame_duration_seconds\": 0.1, \"max_speakers\": 10, \"max_nspks\": 12, \"n_units\": 256, \"n_heads\": 4, \"enc_n_layers\": 4, \"dec_n_layers\": 2, \"conv_kernel_size\": 16, \"conv_delay\": 9, \"sample_rate\": 8000, \"win_length\": 200, \"hop_length\": 80, \"n_mels\": 23, \"context_size\": 7, \"subsampling\": 10, \"feat_type\": \"logmel23_cummn\", \"pure_roll\": true}", + "com.github.apple.coremltools.conversion_date" : "2026-04-18", + "config" : "{\"model_name\": \"dih3\", \"model_label\": \"DIHARD III\", \"variant\": \"pipeline\", \"chunk_size\": 1, \"step_duration_ms\": 100, \"frame_duration_ms\": 100, \"frame_duration_seconds\": 0.1, \"max_speakers\": 10, \"max_nspks\": 12, \"n_units\": 256, \"n_heads\": 4, \"enc_n_layers\": 4, \"dec_n_layers\": 2, \"conv_kernel_size\": 16, \"conv_delay\": 9, \"sample_rate\": 8000, \"win_length\": 200, \"hop_length\": 80, \"n_mels\": 23, \"context_size\": 7, \"subsampling\": 10, \"feat_type\": \"logmel23_cummn\", \"pure_roll\": true, \"input_layout\": \"raw_mel\", \"compute_units_export\": \"all\", \"raw_mel_length\": 15}", "com.github.apple.coremltools.source" : "torch==2.6.0", "com.github.apple.coremltools.version" : "9.0", "com.github.apple.coremltools.source_dialect" : "TorchScript" diff --git a/optimized/dih3/100ms/ls_eend_dih3_100ms.mlmodelc/model.mil b/optimized/dih3/100ms/ls_eend_dih3_100ms.mlmodelc/model.mil index 182f88767b6a7156fba8d5e17cc71543704b1b0a..4bc2f36465c8a6b77c6ec5021d7ddf40dec1643b 100644 --- a/optimized/dih3/100ms/ls_eend_dih3_100ms.mlmodelc/model.mil +++ b/optimized/dih3/100ms/ls_eend_dih3_100ms.mlmodelc/model.mil @@ -1,233 +1,239 @@ program(1.0) [buildInfo = dict, tensor>({{"coremlc-component-MIL", "3510.2.1"}, {"coremlc-version", "3500.32.1"}, {"coremltools-component-torch", "2.6.0"}, {"coremltools-source-dialect", "TorchScript"}, {"coremltools-version", "9.0"}})] { - func main(tensor cnn_window, tensor dec_kv, tensor dec_scale, tensor enc_conv_cache, tensor enc_kv, tensor enc_scale, tensor features, tensor valid_mask) { - tensor encoder_cnn_bias = const()[name = tensor("encoder_cnn_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(64)))]; - tensor encoder_cnn_weight = const()[name = tensor("encoder_cnn_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1152)))]; - tensor encoder__causal_mask = const()[name = tensor("encoder__causal_mask"), val = tensor([[0x1p+0]])]; - tensor encoder__t_index = const()[name = tensor("encoder__t_index"), val = tensor([0x1p+0])]; - tensor encoder_input_projection_linear_bias = const()[name = tensor("encoder_input_projection_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4981952)))]; - tensor encoder_input_projection_linear_weight = const()[name = tensor("encoder_input_projection_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4983040)))]; - tensor encoder_pre_layer_norm_bias = const()[name = tensor("encoder_pre_layer_norm_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(5336384)))]; - tensor encoder_pre_layer_norm_weight = const()[name = tensor("encoder_pre_layer_norm_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(5337472)))]; - tensor encoder_ffn1_0_module_sequential_0_bias = const()[name = tensor("encoder_ffn1_0_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(5338560)))]; - tensor encoder_ffn1_0_module_sequential_0_weight = const()[name = tensor("encoder_ffn1_0_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(5339648)))]; - tensor encoder_ffn1_0_module_sequential_1_linear_bias = const()[name = tensor("encoder_ffn1_0_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(5340736)))]; - tensor encoder_ffn1_0_module_sequential_1_linear_weight = const()[name = tensor("encoder_ffn1_0_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(5344896)))]; - tensor encoder_ffn1_0_module_sequential_4_linear_bias = const()[name = tensor("encoder_ffn1_0_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(6393536)))]; - tensor encoder_ffn1_0_module_sequential_4_linear_weight = const()[name = tensor("encoder_ffn1_0_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(6394624)))]; - tensor encoder_ret_lns_0_bias = const()[name = tensor("encoder_ret_lns_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7443264)))]; - tensor encoder_ret_lns_0_weight = const()[name = tensor("encoder_ret_lns_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7444352)))]; - tensor encoder_q_proj_0_bias = const()[name = tensor("encoder_q_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7445440)))]; - tensor encoder_q_proj_0_weight = const()[name = tensor("encoder_q_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7446528)))]; - tensor encoder_k_proj_0_bias = const()[name = tensor("encoder_k_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7708736)))]; - tensor encoder_k_proj_0_weight = const()[name = tensor("encoder_k_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7709824)))]; - tensor encoder_v_proj_0_bias = const()[name = tensor("encoder_v_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7972032)))]; - tensor encoder_v_proj_0_weight = const()[name = tensor("encoder_v_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7973120)))]; - tensor encoder_g_proj_0_bias = const()[name = tensor("encoder_g_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8235328)))]; - tensor encoder_g_proj_0_weight = const()[name = tensor("encoder_g_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8236416)))]; - tensor encoder_out_proj_0_bias = const()[name = tensor("encoder_out_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8498624)))]; - tensor encoder_out_proj_0_weight = const()[name = tensor("encoder_out_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8499712)))]; - tensor encoder_conv_module_0_sequential_0_bias = const()[name = tensor("encoder_conv_module_0_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8761920)))]; - tensor encoder_conv_module_0_sequential_0_weight = const()[name = tensor("encoder_conv_module_0_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8763008)))]; - tensor encoder_conv_module_0_sequential_2_conv_bias = const()[name = tensor("encoder_conv_module_0_sequential_2_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8764096)))]; - tensor encoder_conv_module_0_sequential_2_conv_weight = const()[name = tensor("encoder_conv_module_0_sequential_2_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8766208)))]; - tensor encoder_conv_module_0_sequential_4_conv_weight = const()[name = tensor("encoder_conv_module_0_sequential_4_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9290560)))]; - tensor encoder_conv_module_0_sequential_5_running_var = const()[name = tensor("encoder_conv_module_0_sequential_5_running_var"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9307008)))]; - tensor encoder_conv_module_0_sequential_5_running_mean = const()[name = tensor("encoder_conv_module_0_sequential_5_running_mean"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9308096)))]; - tensor encoder_conv_module_0_sequential_5_bias = const()[name = tensor("encoder_conv_module_0_sequential_5_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9309184)))]; - tensor encoder_conv_module_0_sequential_5_weight = const()[name = tensor("encoder_conv_module_0_sequential_5_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9310272)))]; - tensor encoder_conv_module_0_sequential_7_conv_bias = const()[name = tensor("encoder_conv_module_0_sequential_7_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9311360)))]; - tensor encoder_conv_module_0_sequential_7_conv_weight = const()[name = tensor("encoder_conv_module_0_sequential_7_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9312448)))]; - tensor encoder_ffn2_0_module_sequential_0_bias = const()[name = tensor("encoder_ffn2_0_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9574656)))]; - tensor encoder_ffn2_0_module_sequential_0_weight = const()[name = tensor("encoder_ffn2_0_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9575744)))]; - tensor encoder_ffn2_0_module_sequential_1_linear_bias = const()[name = tensor("encoder_ffn2_0_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9576832)))]; - tensor encoder_ffn2_0_module_sequential_1_linear_weight = const()[name = tensor("encoder_ffn2_0_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9580992)))]; - tensor encoder_ffn2_0_module_sequential_4_linear_bias = const()[name = tensor("encoder_ffn2_0_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(10629632)))]; - tensor encoder_ffn2_0_module_sequential_4_linear_weight = const()[name = tensor("encoder_ffn2_0_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(10630720)))]; - tensor encoder_layer_norm_0_bias = const()[name = tensor("encoder_layer_norm_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11679360)))]; - tensor encoder_layer_norm_0_weight = const()[name = tensor("encoder_layer_norm_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11680448)))]; - tensor encoder_ffn1_1_module_sequential_0_bias = const()[name = tensor("encoder_ffn1_1_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11681536)))]; - tensor encoder_ffn1_1_module_sequential_0_weight = const()[name = tensor("encoder_ffn1_1_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11682624)))]; - tensor encoder_ffn1_1_module_sequential_1_linear_bias = const()[name = tensor("encoder_ffn1_1_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11683712)))]; - tensor encoder_ffn1_1_module_sequential_1_linear_weight = const()[name = tensor("encoder_ffn1_1_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11687872)))]; - tensor encoder_ffn1_1_module_sequential_4_linear_bias = const()[name = tensor("encoder_ffn1_1_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(12736512)))]; - tensor encoder_ffn1_1_module_sequential_4_linear_weight = const()[name = tensor("encoder_ffn1_1_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(12737600)))]; - tensor encoder_ret_lns_1_bias = const()[name = tensor("encoder_ret_lns_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(13786240)))]; - tensor encoder_ret_lns_1_weight = const()[name = tensor("encoder_ret_lns_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(13787328)))]; - tensor encoder_q_proj_1_bias = const()[name = tensor("encoder_q_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(13788416)))]; - tensor encoder_q_proj_1_weight = const()[name = tensor("encoder_q_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(13789504)))]; - tensor encoder_k_proj_1_bias = const()[name = tensor("encoder_k_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14051712)))]; - tensor encoder_k_proj_1_weight = const()[name = tensor("encoder_k_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14052800)))]; - tensor encoder_v_proj_1_bias = const()[name = tensor("encoder_v_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14315008)))]; - tensor encoder_v_proj_1_weight = const()[name = tensor("encoder_v_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14316096)))]; - tensor encoder_g_proj_1_bias = const()[name = tensor("encoder_g_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14578304)))]; - tensor encoder_g_proj_1_weight = const()[name = tensor("encoder_g_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14579392)))]; - tensor encoder_out_proj_1_bias = const()[name = tensor("encoder_out_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14841600)))]; - tensor encoder_out_proj_1_weight = const()[name = tensor("encoder_out_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14842688)))]; - tensor encoder_conv_module_1_sequential_0_bias = const()[name = tensor("encoder_conv_module_1_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15104896)))]; - tensor encoder_conv_module_1_sequential_0_weight = const()[name = tensor("encoder_conv_module_1_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15105984)))]; - tensor encoder_conv_module_1_sequential_2_conv_bias = const()[name = tensor("encoder_conv_module_1_sequential_2_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15107072)))]; - tensor encoder_conv_module_1_sequential_2_conv_weight = const()[name = tensor("encoder_conv_module_1_sequential_2_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15109184)))]; - tensor encoder_conv_module_1_sequential_4_conv_weight = const()[name = tensor("encoder_conv_module_1_sequential_4_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15633536)))]; - tensor encoder_conv_module_1_sequential_5_running_var = const()[name = tensor("encoder_conv_module_1_sequential_5_running_var"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15649984)))]; - tensor encoder_conv_module_1_sequential_5_running_mean = const()[name = tensor("encoder_conv_module_1_sequential_5_running_mean"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15651072)))]; - tensor encoder_conv_module_1_sequential_5_bias = const()[name = tensor("encoder_conv_module_1_sequential_5_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15652160)))]; - tensor encoder_conv_module_1_sequential_5_weight = const()[name = tensor("encoder_conv_module_1_sequential_5_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15653248)))]; - tensor encoder_conv_module_1_sequential_7_conv_bias = const()[name = tensor("encoder_conv_module_1_sequential_7_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15654336)))]; - tensor encoder_conv_module_1_sequential_7_conv_weight = const()[name = tensor("encoder_conv_module_1_sequential_7_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15655424)))]; - tensor encoder_ffn2_1_module_sequential_0_bias = const()[name = tensor("encoder_ffn2_1_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15917632)))]; - tensor encoder_ffn2_1_module_sequential_0_weight = const()[name = tensor("encoder_ffn2_1_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15918720)))]; - tensor encoder_ffn2_1_module_sequential_1_linear_bias = const()[name = tensor("encoder_ffn2_1_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15919808)))]; - tensor encoder_ffn2_1_module_sequential_1_linear_weight = const()[name = tensor("encoder_ffn2_1_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15923968)))]; - tensor encoder_ffn2_1_module_sequential_4_linear_bias = const()[name = tensor("encoder_ffn2_1_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(16972608)))]; - tensor encoder_ffn2_1_module_sequential_4_linear_weight = const()[name = tensor("encoder_ffn2_1_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(16973696)))]; - tensor encoder_layer_norm_1_bias = const()[name = tensor("encoder_layer_norm_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18022336)))]; - tensor encoder_layer_norm_1_weight = const()[name = tensor("encoder_layer_norm_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18023424)))]; - tensor encoder_ffn1_2_module_sequential_0_bias = const()[name = tensor("encoder_ffn1_2_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18024512)))]; - tensor encoder_ffn1_2_module_sequential_0_weight = const()[name = tensor("encoder_ffn1_2_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18025600)))]; - tensor encoder_ffn1_2_module_sequential_1_linear_bias = const()[name = tensor("encoder_ffn1_2_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18026688)))]; - tensor encoder_ffn1_2_module_sequential_1_linear_weight = const()[name = tensor("encoder_ffn1_2_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18030848)))]; - tensor encoder_ffn1_2_module_sequential_4_linear_bias = const()[name = tensor("encoder_ffn1_2_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(19079488)))]; - tensor encoder_ffn1_2_module_sequential_4_linear_weight = const()[name = tensor("encoder_ffn1_2_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(19080576)))]; - tensor encoder_ret_lns_2_bias = const()[name = tensor("encoder_ret_lns_2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20129216)))]; - tensor encoder_ret_lns_2_weight = const()[name = tensor("encoder_ret_lns_2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20130304)))]; - tensor encoder_q_proj_2_bias = const()[name = tensor("encoder_q_proj_2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20131392)))]; - tensor encoder_q_proj_2_weight = const()[name = tensor("encoder_q_proj_2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20132480)))]; - tensor encoder_k_proj_2_bias = const()[name = tensor("encoder_k_proj_2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20394688)))]; - tensor encoder_k_proj_2_weight = const()[name = tensor("encoder_k_proj_2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20395776)))]; - tensor encoder_v_proj_2_bias = const()[name = tensor("encoder_v_proj_2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20657984)))]; - tensor encoder_v_proj_2_weight = const()[name = tensor("encoder_v_proj_2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20659072)))]; - tensor encoder_g_proj_2_bias = const()[name = tensor("encoder_g_proj_2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20921280)))]; - tensor encoder_g_proj_2_weight = const()[name = tensor("encoder_g_proj_2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20922368)))]; - tensor encoder_out_proj_2_bias = const()[name = tensor("encoder_out_proj_2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21184576)))]; - tensor encoder_out_proj_2_weight = const()[name = tensor("encoder_out_proj_2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21185664)))]; - tensor encoder_conv_module_2_sequential_0_bias = const()[name = tensor("encoder_conv_module_2_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21447872)))]; - tensor encoder_conv_module_2_sequential_0_weight = const()[name = tensor("encoder_conv_module_2_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21448960)))]; - tensor encoder_conv_module_2_sequential_2_conv_bias = const()[name = tensor("encoder_conv_module_2_sequential_2_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21450048)))]; - tensor encoder_conv_module_2_sequential_2_conv_weight = const()[name = tensor("encoder_conv_module_2_sequential_2_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21452160)))]; - tensor encoder_conv_module_2_sequential_4_conv_weight = const()[name = tensor("encoder_conv_module_2_sequential_4_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21976512)))]; - tensor encoder_conv_module_2_sequential_5_running_var = const()[name = tensor("encoder_conv_module_2_sequential_5_running_var"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21992960)))]; - tensor encoder_conv_module_2_sequential_5_running_mean = const()[name = tensor("encoder_conv_module_2_sequential_5_running_mean"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21994048)))]; - tensor encoder_conv_module_2_sequential_5_bias = const()[name = tensor("encoder_conv_module_2_sequential_5_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21995136)))]; - tensor encoder_conv_module_2_sequential_5_weight = const()[name = tensor("encoder_conv_module_2_sequential_5_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21996224)))]; - tensor encoder_conv_module_2_sequential_7_conv_bias = const()[name = tensor("encoder_conv_module_2_sequential_7_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21997312)))]; - tensor encoder_conv_module_2_sequential_7_conv_weight = const()[name = tensor("encoder_conv_module_2_sequential_7_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21998400)))]; - tensor encoder_ffn2_2_module_sequential_0_bias = const()[name = tensor("encoder_ffn2_2_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(22260608)))]; - tensor encoder_ffn2_2_module_sequential_0_weight = const()[name = tensor("encoder_ffn2_2_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(22261696)))]; - tensor encoder_ffn2_2_module_sequential_1_linear_bias = const()[name = tensor("encoder_ffn2_2_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(22262784)))]; - tensor encoder_ffn2_2_module_sequential_1_linear_weight = const()[name = tensor("encoder_ffn2_2_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(22266944)))]; - tensor encoder_ffn2_2_module_sequential_4_linear_bias = const()[name = tensor("encoder_ffn2_2_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(23315584)))]; - tensor encoder_ffn2_2_module_sequential_4_linear_weight = const()[name = tensor("encoder_ffn2_2_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(23316672)))]; - tensor encoder_layer_norm_2_bias = const()[name = tensor("encoder_layer_norm_2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(24365312)))]; - tensor encoder_layer_norm_2_weight = const()[name = tensor("encoder_layer_norm_2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(24366400)))]; - tensor encoder_ffn1_3_module_sequential_0_bias = const()[name = tensor("encoder_ffn1_3_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(24367488)))]; - tensor encoder_ffn1_3_module_sequential_0_weight = const()[name = tensor("encoder_ffn1_3_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(24368576)))]; - tensor encoder_ffn1_3_module_sequential_1_linear_bias = const()[name = tensor("encoder_ffn1_3_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(24369664)))]; - tensor encoder_ffn1_3_module_sequential_1_linear_weight = const()[name = tensor("encoder_ffn1_3_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(24373824)))]; - tensor encoder_ffn1_3_module_sequential_4_linear_bias = const()[name = tensor("encoder_ffn1_3_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(25422464)))]; - tensor encoder_ffn1_3_module_sequential_4_linear_weight = const()[name = tensor("encoder_ffn1_3_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(25423552)))]; - tensor encoder_ret_lns_3_bias = const()[name = tensor("encoder_ret_lns_3_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(26472192)))]; - tensor encoder_ret_lns_3_weight = const()[name = tensor("encoder_ret_lns_3_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(26473280)))]; - tensor encoder_q_proj_3_bias = const()[name = tensor("encoder_q_proj_3_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(26474368)))]; - tensor encoder_q_proj_3_weight = const()[name = tensor("encoder_q_proj_3_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(26475456)))]; - tensor encoder_k_proj_3_bias = const()[name = tensor("encoder_k_proj_3_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(26737664)))]; - tensor encoder_k_proj_3_weight = const()[name = tensor("encoder_k_proj_3_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(26738752)))]; - tensor encoder_v_proj_3_bias = const()[name = tensor("encoder_v_proj_3_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27000960)))]; - tensor encoder_v_proj_3_weight = const()[name = tensor("encoder_v_proj_3_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27002048)))]; - tensor encoder_g_proj_3_bias = const()[name = tensor("encoder_g_proj_3_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27264256)))]; - tensor encoder_g_proj_3_weight = const()[name = tensor("encoder_g_proj_3_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27265344)))]; - tensor encoder_out_proj_3_bias = const()[name = tensor("encoder_out_proj_3_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27527552)))]; - tensor encoder_out_proj_3_weight = const()[name = tensor("encoder_out_proj_3_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27528640)))]; - tensor encoder_conv_module_3_sequential_0_bias = const()[name = tensor("encoder_conv_module_3_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27790848)))]; - tensor encoder_conv_module_3_sequential_0_weight = const()[name = tensor("encoder_conv_module_3_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27791936)))]; - tensor encoder_conv_module_3_sequential_2_conv_bias = const()[name = tensor("encoder_conv_module_3_sequential_2_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27793024)))]; - tensor encoder_conv_module_3_sequential_2_conv_weight = const()[name = tensor("encoder_conv_module_3_sequential_2_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27795136)))]; - tensor encoder_conv_module_3_sequential_4_conv_weight = const()[name = tensor("encoder_conv_module_3_sequential_4_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28319488)))]; - tensor encoder_conv_module_3_sequential_5_running_var = const()[name = tensor("encoder_conv_module_3_sequential_5_running_var"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28335936)))]; - tensor encoder_conv_module_3_sequential_5_running_mean = const()[name = tensor("encoder_conv_module_3_sequential_5_running_mean"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28337024)))]; - tensor encoder_conv_module_3_sequential_5_bias = const()[name = tensor("encoder_conv_module_3_sequential_5_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28338112)))]; - tensor encoder_conv_module_3_sequential_5_weight = const()[name = tensor("encoder_conv_module_3_sequential_5_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28339200)))]; - tensor encoder_conv_module_3_sequential_7_conv_bias = const()[name = tensor("encoder_conv_module_3_sequential_7_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28340288)))]; - tensor encoder_conv_module_3_sequential_7_conv_weight = const()[name = tensor("encoder_conv_module_3_sequential_7_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28341376)))]; - tensor encoder_ffn2_3_module_sequential_0_bias = const()[name = tensor("encoder_ffn2_3_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28603584)))]; - tensor encoder_ffn2_3_module_sequential_0_weight = const()[name = tensor("encoder_ffn2_3_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28604672)))]; - tensor encoder_ffn2_3_module_sequential_1_linear_bias = const()[name = tensor("encoder_ffn2_3_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28605760)))]; - tensor encoder_ffn2_3_module_sequential_1_linear_weight = const()[name = tensor("encoder_ffn2_3_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28609920)))]; - tensor encoder_ffn2_3_module_sequential_4_linear_bias = const()[name = tensor("encoder_ffn2_3_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(29658560)))]; - tensor encoder_ffn2_3_module_sequential_4_linear_weight = const()[name = tensor("encoder_ffn2_3_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(29659648)))]; - tensor encoder_layer_norm_3_bias = const()[name = tensor("encoder_layer_norm_3_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(30708288)))]; - tensor encoder_layer_norm_3_weight = const()[name = tensor("encoder_layer_norm_3_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(30709376)))]; - tensor decoder_convert_bias = const()[name = tensor("decoder_convert_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(30710464)))]; - tensor decoder_convert_weight = const()[name = tensor("decoder_convert_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(30711552)))]; - tensor decoder_q_proj_0_bias = const()[name = tensor("decoder_q_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(31235904)))]; - tensor decoder_q_proj_0_weight = const()[name = tensor("decoder_q_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(31236992)))]; - tensor decoder_k_proj_0_bias = const()[name = tensor("decoder_k_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(31499200)))]; - tensor decoder_k_proj_0_weight = const()[name = tensor("decoder_k_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(31500288)))]; - tensor decoder_v_proj_0_bias = const()[name = tensor("decoder_v_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(31762496)))]; - tensor decoder_v_proj_0_weight = const()[name = tensor("decoder_v_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(31763584)))]; - tensor decoder_g_proj_0_bias = const()[name = tensor("decoder_g_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32025792)))]; - tensor decoder_g_proj_0_weight = const()[name = tensor("decoder_g_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32026880)))]; - tensor decoder_out_proj_0_bias = const()[name = tensor("decoder_out_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32289088)))]; - tensor decoder_out_proj_0_weight = const()[name = tensor("decoder_out_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32290176)))]; - tensor decoder_norm11_0_bias = const()[name = tensor("decoder_norm11_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32552384)))]; - tensor decoder_norm11_0_weight = const()[name = tensor("decoder_norm11_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32553472)))]; - tensor decoder_self_attn2_0_out_proj_bias = const()[name = tensor("decoder_self_attn2_0_out_proj_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32554560)))]; - tensor decoder_self_attn2_0_out_proj_weight = const()[name = tensor("decoder_self_attn2_0_out_proj_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32555648)))]; - tensor decoder_self_attn2_0_in_proj_bias = const()[name = tensor("decoder_self_attn2_0_in_proj_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32817856)))]; - tensor decoder_self_attn2_0_in_proj_weight = const()[name = tensor("decoder_self_attn2_0_in_proj_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32820992)))]; - tensor decoder_norm21_0_bias = const()[name = tensor("decoder_norm21_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(33607488)))]; - tensor decoder_norm21_0_weight = const()[name = tensor("decoder_norm21_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(33608576)))]; - tensor decoder_linear1_0_bias = const()[name = tensor("decoder_linear1_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(33609664)))]; - tensor decoder_linear1_0_weight = const()[name = tensor("decoder_linear1_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(33617920)))]; - tensor decoder_linear2_0_bias = const()[name = tensor("decoder_linear2_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(35715136)))]; - tensor decoder_linear2_0_weight = const()[name = tensor("decoder_linear2_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(35716224)))]; - tensor decoder_norm22_0_bias = const()[name = tensor("decoder_norm22_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(37813440)))]; - tensor decoder_norm22_0_weight = const()[name = tensor("decoder_norm22_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(37814528)))]; - tensor decoder_q_proj_1_bias = const()[name = tensor("decoder_q_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(37815616)))]; - tensor decoder_q_proj_1_weight = const()[name = tensor("decoder_q_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(37816704)))]; - tensor decoder_k_proj_1_bias = const()[name = tensor("decoder_k_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38078912)))]; - tensor decoder_k_proj_1_weight = const()[name = tensor("decoder_k_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38080000)))]; - tensor decoder_v_proj_1_bias = const()[name = tensor("decoder_v_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38342208)))]; - tensor decoder_v_proj_1_weight = const()[name = tensor("decoder_v_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38343296)))]; - tensor decoder_g_proj_1_bias = const()[name = tensor("decoder_g_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38605504)))]; - tensor decoder_g_proj_1_weight = const()[name = tensor("decoder_g_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38606592)))]; - tensor decoder_out_proj_1_bias = const()[name = tensor("decoder_out_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38868800)))]; - tensor decoder_out_proj_1_weight = const()[name = tensor("decoder_out_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38869888)))]; - tensor decoder_norm11_1_bias = const()[name = tensor("decoder_norm11_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(39132096)))]; - tensor decoder_norm11_1_weight = const()[name = tensor("decoder_norm11_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(39133184)))]; - tensor decoder_self_attn2_1_out_proj_bias = const()[name = tensor("decoder_self_attn2_1_out_proj_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(39134272)))]; - tensor decoder_self_attn2_1_out_proj_weight = const()[name = tensor("decoder_self_attn2_1_out_proj_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(39135360)))]; - tensor decoder_self_attn2_1_in_proj_bias = const()[name = tensor("decoder_self_attn2_1_in_proj_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(39397568)))]; - tensor decoder_self_attn2_1_in_proj_weight = const()[name = tensor("decoder_self_attn2_1_in_proj_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(39400704)))]; - tensor decoder_norm21_1_bias = const()[name = tensor("decoder_norm21_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(40187200)))]; - tensor decoder_norm21_1_weight = const()[name = tensor("decoder_norm21_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(40188288)))]; - tensor decoder_linear1_1_bias = const()[name = tensor("decoder_linear1_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(40189376)))]; - tensor decoder_linear1_1_weight = const()[name = tensor("decoder_linear1_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(40197632)))]; - tensor decoder_linear2_1_bias = const()[name = tensor("decoder_linear2_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(42294848)))]; - tensor decoder_linear2_1_weight = const()[name = tensor("decoder_linear2_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(42295936)))]; - tensor decoder_norm22_1_bias = const()[name = tensor("decoder_norm22_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(44393152)))]; - tensor decoder_norm22_1_weight = const()[name = tensor("decoder_norm22_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(44394240)))]; - tensor var_11 = const()[name = tensor("op_11"), val = tensor(0x1.5798eep-27)]; - tensor var_17 = const()[name = tensor("op_17"), val = tensor(0x1.0c6f7ap-20)]; - tensor var_20 = const()[name = tensor("op_20"), val = tensor(2)]; - tensor var_23 = const()[name = tensor("op_23"), val = tensor(0)]; - tensor var_26 = const()[name = tensor("op_26"), val = tensor(1)]; - tensor var_28 = const()[name = tensor("op_28"), val = tensor(0x1.4f8b58p-17)]; - tensor var_29 = const()[name = tensor("op_29"), val = tensor(true)]; - tensor input_1 = linear(bias = encoder_input_projection_linear_bias, weight = encoder_input_projection_linear_weight, x = features)[name = tensor("linear_0")]; - tensor input_3_axes_0 = const()[name = tensor("input_3_axes_0"), val = tensor([-1])]; - tensor input_3 = layer_norm(axes = input_3_axes_0, beta = encoder_pre_layer_norm_bias, epsilon = var_28, gamma = encoder_pre_layer_norm_weight, x = input_1)[name = tensor("input_3")]; + func main(tensor cnn_window, tensor dec_kv, tensor dec_scale, tensor enc_conv_cache, tensor enc_kv, tensor enc_scale, tensor features, tensor valid_mask) { + tensor inner_encoder_cnn_bias = const()[name = tensor("inner_encoder_cnn_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(64)))]; + tensor inner_encoder_cnn_weight = const()[name = tensor("inner_encoder_cnn_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1152)))]; + tensor inner_encoder__causal_mask = const()[name = tensor("inner_encoder__causal_mask"), val = tensor([[0x1p+0]])]; + tensor inner_encoder__t_index = const()[name = tensor("inner_encoder__t_index"), val = tensor([0x1p+0])]; + tensor inner_encoder_input_projection_linear_bias = const()[name = tensor("inner_encoder_input_projection_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4981952)))]; + tensor inner_encoder_input_projection_linear_weight = const()[name = tensor("inner_encoder_input_projection_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4983040)))]; + tensor inner_encoder_pre_layer_norm_bias = const()[name = tensor("inner_encoder_pre_layer_norm_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(5336384)))]; + tensor inner_encoder_pre_layer_norm_weight = const()[name = tensor("inner_encoder_pre_layer_norm_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(5337472)))]; + tensor inner_encoder_ffn1_0_module_sequential_0_bias = const()[name = tensor("inner_encoder_ffn1_0_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(5338560)))]; + tensor inner_encoder_ffn1_0_module_sequential_0_weight = const()[name = tensor("inner_encoder_ffn1_0_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(5339648)))]; + tensor inner_encoder_ffn1_0_module_sequential_1_linear_bias = const()[name = tensor("inner_encoder_ffn1_0_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(5340736)))]; + tensor inner_encoder_ffn1_0_module_sequential_1_linear_weight = const()[name = tensor("inner_encoder_ffn1_0_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(5344896)))]; + tensor inner_encoder_ffn1_0_module_sequential_4_linear_bias = const()[name = tensor("inner_encoder_ffn1_0_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(6393536)))]; + tensor inner_encoder_ffn1_0_module_sequential_4_linear_weight = const()[name = tensor("inner_encoder_ffn1_0_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(6394624)))]; + tensor inner_encoder_ret_lns_0_bias = const()[name = tensor("inner_encoder_ret_lns_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7443264)))]; + tensor inner_encoder_ret_lns_0_weight = const()[name = tensor("inner_encoder_ret_lns_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7444352)))]; + tensor inner_encoder_q_proj_0_bias = const()[name = tensor("inner_encoder_q_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7445440)))]; + tensor inner_encoder_q_proj_0_weight = const()[name = tensor("inner_encoder_q_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7446528)))]; + tensor inner_encoder_k_proj_0_bias = const()[name = tensor("inner_encoder_k_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7708736)))]; + tensor inner_encoder_k_proj_0_weight = const()[name = tensor("inner_encoder_k_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7709824)))]; + tensor inner_encoder_v_proj_0_bias = const()[name = tensor("inner_encoder_v_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7972032)))]; + tensor inner_encoder_v_proj_0_weight = const()[name = tensor("inner_encoder_v_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7973120)))]; + tensor inner_encoder_g_proj_0_bias = const()[name = tensor("inner_encoder_g_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8235328)))]; + tensor inner_encoder_g_proj_0_weight = const()[name = tensor("inner_encoder_g_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8236416)))]; + tensor inner_encoder_out_proj_0_bias = const()[name = tensor("inner_encoder_out_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8498624)))]; + tensor inner_encoder_out_proj_0_weight = const()[name = tensor("inner_encoder_out_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8499712)))]; + tensor inner_encoder_conv_module_0_sequential_0_bias = const()[name = tensor("inner_encoder_conv_module_0_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8761920)))]; + tensor inner_encoder_conv_module_0_sequential_0_weight = const()[name = tensor("inner_encoder_conv_module_0_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8763008)))]; + tensor inner_encoder_conv_module_0_sequential_2_conv_bias = const()[name = tensor("inner_encoder_conv_module_0_sequential_2_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8764096)))]; + tensor inner_encoder_conv_module_0_sequential_2_conv_weight = const()[name = tensor("inner_encoder_conv_module_0_sequential_2_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8766208)))]; + tensor inner_encoder_conv_module_0_sequential_4_conv_weight = const()[name = tensor("inner_encoder_conv_module_0_sequential_4_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9290560)))]; + tensor inner_encoder_conv_module_0_sequential_5_running_var = const()[name = tensor("inner_encoder_conv_module_0_sequential_5_running_var"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9307008)))]; + tensor inner_encoder_conv_module_0_sequential_5_running_mean = const()[name = tensor("inner_encoder_conv_module_0_sequential_5_running_mean"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9308096)))]; + tensor inner_encoder_conv_module_0_sequential_5_bias = const()[name = tensor("inner_encoder_conv_module_0_sequential_5_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9309184)))]; + tensor inner_encoder_conv_module_0_sequential_5_weight = const()[name = tensor("inner_encoder_conv_module_0_sequential_5_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9310272)))]; + tensor inner_encoder_conv_module_0_sequential_7_conv_bias = const()[name = tensor("inner_encoder_conv_module_0_sequential_7_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9311360)))]; + tensor inner_encoder_conv_module_0_sequential_7_conv_weight = const()[name = tensor("inner_encoder_conv_module_0_sequential_7_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9312448)))]; + tensor inner_encoder_ffn2_0_module_sequential_0_bias = const()[name = tensor("inner_encoder_ffn2_0_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9574656)))]; + tensor inner_encoder_ffn2_0_module_sequential_0_weight = const()[name = tensor("inner_encoder_ffn2_0_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9575744)))]; + tensor inner_encoder_ffn2_0_module_sequential_1_linear_bias = const()[name = tensor("inner_encoder_ffn2_0_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9576832)))]; + tensor inner_encoder_ffn2_0_module_sequential_1_linear_weight = const()[name = tensor("inner_encoder_ffn2_0_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9580992)))]; + tensor inner_encoder_ffn2_0_module_sequential_4_linear_bias = const()[name = tensor("inner_encoder_ffn2_0_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(10629632)))]; + tensor inner_encoder_ffn2_0_module_sequential_4_linear_weight = const()[name = tensor("inner_encoder_ffn2_0_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(10630720)))]; + tensor inner_encoder_layer_norm_0_bias = const()[name = tensor("inner_encoder_layer_norm_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11679360)))]; + tensor inner_encoder_layer_norm_0_weight = const()[name = tensor("inner_encoder_layer_norm_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11680448)))]; + tensor inner_encoder_ffn1_1_module_sequential_0_bias = const()[name = tensor("inner_encoder_ffn1_1_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11681536)))]; + tensor inner_encoder_ffn1_1_module_sequential_0_weight = const()[name = tensor("inner_encoder_ffn1_1_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11682624)))]; + tensor inner_encoder_ffn1_1_module_sequential_1_linear_bias = const()[name = tensor("inner_encoder_ffn1_1_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11683712)))]; + tensor inner_encoder_ffn1_1_module_sequential_1_linear_weight = const()[name = tensor("inner_encoder_ffn1_1_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11687872)))]; + tensor inner_encoder_ffn1_1_module_sequential_4_linear_bias = const()[name = tensor("inner_encoder_ffn1_1_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(12736512)))]; + tensor inner_encoder_ffn1_1_module_sequential_4_linear_weight = const()[name = tensor("inner_encoder_ffn1_1_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(12737600)))]; + tensor inner_encoder_ret_lns_1_bias = const()[name = tensor("inner_encoder_ret_lns_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(13786240)))]; + tensor inner_encoder_ret_lns_1_weight = const()[name = tensor("inner_encoder_ret_lns_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(13787328)))]; + tensor inner_encoder_q_proj_1_bias = const()[name = tensor("inner_encoder_q_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(13788416)))]; + tensor inner_encoder_q_proj_1_weight = const()[name = tensor("inner_encoder_q_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(13789504)))]; + tensor inner_encoder_k_proj_1_bias = const()[name = tensor("inner_encoder_k_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14051712)))]; + tensor inner_encoder_k_proj_1_weight = const()[name = tensor("inner_encoder_k_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14052800)))]; + tensor inner_encoder_v_proj_1_bias = const()[name = tensor("inner_encoder_v_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14315008)))]; + tensor inner_encoder_v_proj_1_weight = const()[name = tensor("inner_encoder_v_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14316096)))]; + tensor inner_encoder_g_proj_1_bias = const()[name = tensor("inner_encoder_g_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14578304)))]; + tensor inner_encoder_g_proj_1_weight = const()[name = tensor("inner_encoder_g_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14579392)))]; + tensor inner_encoder_out_proj_1_bias = const()[name = tensor("inner_encoder_out_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14841600)))]; + tensor inner_encoder_out_proj_1_weight = const()[name = tensor("inner_encoder_out_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14842688)))]; + tensor inner_encoder_conv_module_1_sequential_0_bias = const()[name = tensor("inner_encoder_conv_module_1_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15104896)))]; + tensor inner_encoder_conv_module_1_sequential_0_weight = const()[name = tensor("inner_encoder_conv_module_1_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15105984)))]; + tensor inner_encoder_conv_module_1_sequential_2_conv_bias = const()[name = tensor("inner_encoder_conv_module_1_sequential_2_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15107072)))]; + tensor inner_encoder_conv_module_1_sequential_2_conv_weight = const()[name = tensor("inner_encoder_conv_module_1_sequential_2_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15109184)))]; + tensor inner_encoder_conv_module_1_sequential_4_conv_weight = const()[name = tensor("inner_encoder_conv_module_1_sequential_4_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15633536)))]; + tensor inner_encoder_conv_module_1_sequential_5_running_var = const()[name = tensor("inner_encoder_conv_module_1_sequential_5_running_var"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15649984)))]; + tensor inner_encoder_conv_module_1_sequential_5_running_mean = const()[name = tensor("inner_encoder_conv_module_1_sequential_5_running_mean"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15651072)))]; + tensor inner_encoder_conv_module_1_sequential_5_bias = const()[name = tensor("inner_encoder_conv_module_1_sequential_5_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15652160)))]; + tensor inner_encoder_conv_module_1_sequential_5_weight = const()[name = tensor("inner_encoder_conv_module_1_sequential_5_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15653248)))]; + tensor inner_encoder_conv_module_1_sequential_7_conv_bias = const()[name = tensor("inner_encoder_conv_module_1_sequential_7_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15654336)))]; + tensor inner_encoder_conv_module_1_sequential_7_conv_weight = const()[name = tensor("inner_encoder_conv_module_1_sequential_7_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15655424)))]; + tensor inner_encoder_ffn2_1_module_sequential_0_bias = const()[name = tensor("inner_encoder_ffn2_1_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15917632)))]; + tensor inner_encoder_ffn2_1_module_sequential_0_weight = const()[name = tensor("inner_encoder_ffn2_1_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15918720)))]; + tensor inner_encoder_ffn2_1_module_sequential_1_linear_bias = const()[name = tensor("inner_encoder_ffn2_1_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15919808)))]; + tensor inner_encoder_ffn2_1_module_sequential_1_linear_weight = const()[name = tensor("inner_encoder_ffn2_1_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15923968)))]; + tensor inner_encoder_ffn2_1_module_sequential_4_linear_bias = const()[name = tensor("inner_encoder_ffn2_1_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(16972608)))]; + tensor inner_encoder_ffn2_1_module_sequential_4_linear_weight = const()[name = tensor("inner_encoder_ffn2_1_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(16973696)))]; + tensor inner_encoder_layer_norm_1_bias = const()[name = tensor("inner_encoder_layer_norm_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18022336)))]; + tensor inner_encoder_layer_norm_1_weight = const()[name = tensor("inner_encoder_layer_norm_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18023424)))]; + tensor inner_encoder_ffn1_2_module_sequential_0_bias = const()[name = tensor("inner_encoder_ffn1_2_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18024512)))]; + tensor inner_encoder_ffn1_2_module_sequential_0_weight = const()[name = tensor("inner_encoder_ffn1_2_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18025600)))]; + tensor inner_encoder_ffn1_2_module_sequential_1_linear_bias = const()[name = tensor("inner_encoder_ffn1_2_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18026688)))]; + tensor inner_encoder_ffn1_2_module_sequential_1_linear_weight = const()[name = tensor("inner_encoder_ffn1_2_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18030848)))]; + tensor inner_encoder_ffn1_2_module_sequential_4_linear_bias = const()[name = tensor("inner_encoder_ffn1_2_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(19079488)))]; + tensor inner_encoder_ffn1_2_module_sequential_4_linear_weight = const()[name = tensor("inner_encoder_ffn1_2_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(19080576)))]; + tensor inner_encoder_ret_lns_2_bias = const()[name = tensor("inner_encoder_ret_lns_2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20129216)))]; + tensor inner_encoder_ret_lns_2_weight = const()[name = tensor("inner_encoder_ret_lns_2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20130304)))]; + tensor inner_encoder_q_proj_2_bias = const()[name = tensor("inner_encoder_q_proj_2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20131392)))]; + tensor inner_encoder_q_proj_2_weight = const()[name = tensor("inner_encoder_q_proj_2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20132480)))]; + tensor inner_encoder_k_proj_2_bias = const()[name = tensor("inner_encoder_k_proj_2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20394688)))]; + tensor inner_encoder_k_proj_2_weight = const()[name = tensor("inner_encoder_k_proj_2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20395776)))]; + tensor inner_encoder_v_proj_2_bias = const()[name = tensor("inner_encoder_v_proj_2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20657984)))]; + tensor inner_encoder_v_proj_2_weight = const()[name = tensor("inner_encoder_v_proj_2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20659072)))]; + tensor inner_encoder_g_proj_2_bias = const()[name = tensor("inner_encoder_g_proj_2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20921280)))]; + tensor inner_encoder_g_proj_2_weight = const()[name = tensor("inner_encoder_g_proj_2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20922368)))]; + tensor inner_encoder_out_proj_2_bias = const()[name = tensor("inner_encoder_out_proj_2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21184576)))]; + tensor inner_encoder_out_proj_2_weight = const()[name = tensor("inner_encoder_out_proj_2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21185664)))]; + tensor inner_encoder_conv_module_2_sequential_0_bias = const()[name = tensor("inner_encoder_conv_module_2_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21447872)))]; + tensor inner_encoder_conv_module_2_sequential_0_weight = const()[name = tensor("inner_encoder_conv_module_2_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21448960)))]; + tensor inner_encoder_conv_module_2_sequential_2_conv_bias = const()[name = tensor("inner_encoder_conv_module_2_sequential_2_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21450048)))]; + tensor inner_encoder_conv_module_2_sequential_2_conv_weight = const()[name = tensor("inner_encoder_conv_module_2_sequential_2_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21452160)))]; + tensor inner_encoder_conv_module_2_sequential_4_conv_weight = const()[name = tensor("inner_encoder_conv_module_2_sequential_4_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21976512)))]; + tensor inner_encoder_conv_module_2_sequential_5_running_var = const()[name = tensor("inner_encoder_conv_module_2_sequential_5_running_var"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21992960)))]; + tensor inner_encoder_conv_module_2_sequential_5_running_mean = const()[name = tensor("inner_encoder_conv_module_2_sequential_5_running_mean"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21994048)))]; + tensor inner_encoder_conv_module_2_sequential_5_bias = const()[name = tensor("inner_encoder_conv_module_2_sequential_5_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21995136)))]; + tensor inner_encoder_conv_module_2_sequential_5_weight = const()[name = tensor("inner_encoder_conv_module_2_sequential_5_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21996224)))]; + tensor inner_encoder_conv_module_2_sequential_7_conv_bias = const()[name = tensor("inner_encoder_conv_module_2_sequential_7_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21997312)))]; + tensor inner_encoder_conv_module_2_sequential_7_conv_weight = const()[name = tensor("inner_encoder_conv_module_2_sequential_7_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21998400)))]; + tensor inner_encoder_ffn2_2_module_sequential_0_bias = const()[name = tensor("inner_encoder_ffn2_2_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(22260608)))]; + tensor inner_encoder_ffn2_2_module_sequential_0_weight = const()[name = tensor("inner_encoder_ffn2_2_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(22261696)))]; + tensor inner_encoder_ffn2_2_module_sequential_1_linear_bias = const()[name = tensor("inner_encoder_ffn2_2_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(22262784)))]; + tensor inner_encoder_ffn2_2_module_sequential_1_linear_weight = const()[name = tensor("inner_encoder_ffn2_2_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(22266944)))]; + tensor inner_encoder_ffn2_2_module_sequential_4_linear_bias = const()[name = tensor("inner_encoder_ffn2_2_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(23315584)))]; + tensor inner_encoder_ffn2_2_module_sequential_4_linear_weight = const()[name = tensor("inner_encoder_ffn2_2_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(23316672)))]; + tensor inner_encoder_layer_norm_2_bias = const()[name = tensor("inner_encoder_layer_norm_2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(24365312)))]; + tensor inner_encoder_layer_norm_2_weight = const()[name = tensor("inner_encoder_layer_norm_2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(24366400)))]; + tensor inner_encoder_ffn1_3_module_sequential_0_bias = const()[name = tensor("inner_encoder_ffn1_3_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(24367488)))]; + tensor inner_encoder_ffn1_3_module_sequential_0_weight = const()[name = tensor("inner_encoder_ffn1_3_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(24368576)))]; + tensor inner_encoder_ffn1_3_module_sequential_1_linear_bias = const()[name = tensor("inner_encoder_ffn1_3_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(24369664)))]; + tensor inner_encoder_ffn1_3_module_sequential_1_linear_weight = const()[name = tensor("inner_encoder_ffn1_3_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(24373824)))]; + tensor inner_encoder_ffn1_3_module_sequential_4_linear_bias = const()[name = tensor("inner_encoder_ffn1_3_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(25422464)))]; + tensor inner_encoder_ffn1_3_module_sequential_4_linear_weight = const()[name = tensor("inner_encoder_ffn1_3_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(25423552)))]; + tensor inner_encoder_ret_lns_3_bias = const()[name = tensor("inner_encoder_ret_lns_3_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(26472192)))]; + tensor inner_encoder_ret_lns_3_weight = const()[name = tensor("inner_encoder_ret_lns_3_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(26473280)))]; + tensor inner_encoder_q_proj_3_bias = const()[name = tensor("inner_encoder_q_proj_3_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(26474368)))]; + tensor inner_encoder_q_proj_3_weight = const()[name = tensor("inner_encoder_q_proj_3_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(26475456)))]; + tensor inner_encoder_k_proj_3_bias = const()[name = tensor("inner_encoder_k_proj_3_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(26737664)))]; + tensor inner_encoder_k_proj_3_weight = const()[name = tensor("inner_encoder_k_proj_3_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(26738752)))]; + tensor inner_encoder_v_proj_3_bias = const()[name = tensor("inner_encoder_v_proj_3_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27000960)))]; + tensor inner_encoder_v_proj_3_weight = const()[name = tensor("inner_encoder_v_proj_3_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27002048)))]; + tensor inner_encoder_g_proj_3_bias = const()[name = tensor("inner_encoder_g_proj_3_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27264256)))]; + tensor inner_encoder_g_proj_3_weight = const()[name = tensor("inner_encoder_g_proj_3_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27265344)))]; + tensor inner_encoder_out_proj_3_bias = const()[name = tensor("inner_encoder_out_proj_3_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27527552)))]; + tensor inner_encoder_out_proj_3_weight = const()[name = tensor("inner_encoder_out_proj_3_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27528640)))]; + tensor inner_encoder_conv_module_3_sequential_0_bias = const()[name = tensor("inner_encoder_conv_module_3_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27790848)))]; + tensor inner_encoder_conv_module_3_sequential_0_weight = const()[name = tensor("inner_encoder_conv_module_3_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27791936)))]; + tensor inner_encoder_conv_module_3_sequential_2_conv_bias = const()[name = tensor("inner_encoder_conv_module_3_sequential_2_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27793024)))]; + tensor inner_encoder_conv_module_3_sequential_2_conv_weight = const()[name = tensor("inner_encoder_conv_module_3_sequential_2_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27795136)))]; + tensor inner_encoder_conv_module_3_sequential_4_conv_weight = const()[name = tensor("inner_encoder_conv_module_3_sequential_4_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28319488)))]; + tensor inner_encoder_conv_module_3_sequential_5_running_var = const()[name = tensor("inner_encoder_conv_module_3_sequential_5_running_var"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28335936)))]; + tensor inner_encoder_conv_module_3_sequential_5_running_mean = const()[name = tensor("inner_encoder_conv_module_3_sequential_5_running_mean"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28337024)))]; + tensor inner_encoder_conv_module_3_sequential_5_bias = const()[name = tensor("inner_encoder_conv_module_3_sequential_5_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28338112)))]; + tensor inner_encoder_conv_module_3_sequential_5_weight = const()[name = tensor("inner_encoder_conv_module_3_sequential_5_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28339200)))]; + tensor inner_encoder_conv_module_3_sequential_7_conv_bias = const()[name = tensor("inner_encoder_conv_module_3_sequential_7_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28340288)))]; + tensor inner_encoder_conv_module_3_sequential_7_conv_weight = const()[name = tensor("inner_encoder_conv_module_3_sequential_7_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28341376)))]; + tensor inner_encoder_ffn2_3_module_sequential_0_bias = const()[name = tensor("inner_encoder_ffn2_3_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28603584)))]; + tensor inner_encoder_ffn2_3_module_sequential_0_weight = const()[name = tensor("inner_encoder_ffn2_3_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28604672)))]; + tensor inner_encoder_ffn2_3_module_sequential_1_linear_bias = const()[name = tensor("inner_encoder_ffn2_3_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28605760)))]; + tensor inner_encoder_ffn2_3_module_sequential_1_linear_weight = const()[name = tensor("inner_encoder_ffn2_3_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28609920)))]; + tensor inner_encoder_ffn2_3_module_sequential_4_linear_bias = const()[name = tensor("inner_encoder_ffn2_3_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(29658560)))]; + tensor inner_encoder_ffn2_3_module_sequential_4_linear_weight = const()[name = tensor("inner_encoder_ffn2_3_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(29659648)))]; + tensor inner_encoder_layer_norm_3_bias = const()[name = tensor("inner_encoder_layer_norm_3_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(30708288)))]; + tensor inner_encoder_layer_norm_3_weight = const()[name = tensor("inner_encoder_layer_norm_3_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(30709376)))]; + tensor inner_decoder_convert_bias = const()[name = tensor("inner_decoder_convert_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(30710464)))]; + tensor inner_decoder_convert_weight = const()[name = tensor("inner_decoder_convert_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(30711552)))]; + tensor inner_decoder_q_proj_0_bias = const()[name = tensor("inner_decoder_q_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(31235904)))]; + tensor inner_decoder_q_proj_0_weight = const()[name = tensor("inner_decoder_q_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(31236992)))]; + tensor inner_decoder_k_proj_0_bias = const()[name = tensor("inner_decoder_k_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(31499200)))]; + tensor inner_decoder_k_proj_0_weight = const()[name = tensor("inner_decoder_k_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(31500288)))]; + tensor inner_decoder_v_proj_0_bias = const()[name = tensor("inner_decoder_v_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(31762496)))]; + tensor inner_decoder_v_proj_0_weight = const()[name = tensor("inner_decoder_v_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(31763584)))]; + tensor inner_decoder_g_proj_0_bias = const()[name = tensor("inner_decoder_g_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32025792)))]; + tensor inner_decoder_g_proj_0_weight = const()[name = tensor("inner_decoder_g_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32026880)))]; + tensor inner_decoder_out_proj_0_bias = const()[name = tensor("inner_decoder_out_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32289088)))]; + tensor inner_decoder_out_proj_0_weight = const()[name = tensor("inner_decoder_out_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32290176)))]; + tensor inner_decoder_norm11_0_bias = const()[name = tensor("inner_decoder_norm11_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32552384)))]; + tensor inner_decoder_norm11_0_weight = const()[name = tensor("inner_decoder_norm11_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32553472)))]; + tensor inner_decoder_self_attn2_0_out_proj_bias = const()[name = tensor("inner_decoder_self_attn2_0_out_proj_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32554560)))]; + tensor inner_decoder_self_attn2_0_out_proj_weight = const()[name = tensor("inner_decoder_self_attn2_0_out_proj_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32555648)))]; + tensor inner_decoder_self_attn2_0_in_proj_bias = const()[name = tensor("inner_decoder_self_attn2_0_in_proj_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32817856)))]; + tensor inner_decoder_self_attn2_0_in_proj_weight = const()[name = tensor("inner_decoder_self_attn2_0_in_proj_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32820992)))]; + tensor inner_decoder_norm21_0_bias = const()[name = tensor("inner_decoder_norm21_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(33607488)))]; + tensor inner_decoder_norm21_0_weight = const()[name = tensor("inner_decoder_norm21_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(33608576)))]; + tensor inner_decoder_linear1_0_bias = const()[name = tensor("inner_decoder_linear1_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(33609664)))]; + tensor inner_decoder_linear1_0_weight = const()[name = tensor("inner_decoder_linear1_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(33617920)))]; + tensor inner_decoder_linear2_0_bias = const()[name = tensor("inner_decoder_linear2_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(35715136)))]; + tensor inner_decoder_linear2_0_weight = const()[name = tensor("inner_decoder_linear2_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(35716224)))]; + tensor inner_decoder_norm22_0_bias = const()[name = tensor("inner_decoder_norm22_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(37813440)))]; + tensor inner_decoder_norm22_0_weight = const()[name = tensor("inner_decoder_norm22_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(37814528)))]; + tensor inner_decoder_q_proj_1_bias = const()[name = tensor("inner_decoder_q_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(37815616)))]; + tensor inner_decoder_q_proj_1_weight = const()[name = tensor("inner_decoder_q_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(37816704)))]; + tensor inner_decoder_k_proj_1_bias = const()[name = tensor("inner_decoder_k_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38078912)))]; + tensor inner_decoder_k_proj_1_weight = const()[name = tensor("inner_decoder_k_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38080000)))]; + tensor inner_decoder_v_proj_1_bias = const()[name = tensor("inner_decoder_v_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38342208)))]; + tensor inner_decoder_v_proj_1_weight = const()[name = tensor("inner_decoder_v_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38343296)))]; + tensor inner_decoder_g_proj_1_bias = const()[name = tensor("inner_decoder_g_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38605504)))]; + tensor inner_decoder_g_proj_1_weight = const()[name = tensor("inner_decoder_g_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38606592)))]; + tensor inner_decoder_out_proj_1_bias = const()[name = tensor("inner_decoder_out_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38868800)))]; + tensor inner_decoder_out_proj_1_weight = const()[name = tensor("inner_decoder_out_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38869888)))]; + tensor inner_decoder_norm11_1_bias = const()[name = tensor("inner_decoder_norm11_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(39132096)))]; + tensor inner_decoder_norm11_1_weight = const()[name = tensor("inner_decoder_norm11_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(39133184)))]; + tensor inner_decoder_self_attn2_1_out_proj_bias = const()[name = tensor("inner_decoder_self_attn2_1_out_proj_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(39134272)))]; + tensor inner_decoder_self_attn2_1_out_proj_weight = const()[name = tensor("inner_decoder_self_attn2_1_out_proj_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(39135360)))]; + tensor inner_decoder_self_attn2_1_in_proj_bias = const()[name = tensor("inner_decoder_self_attn2_1_in_proj_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(39397568)))]; + tensor inner_decoder_self_attn2_1_in_proj_weight = const()[name = tensor("inner_decoder_self_attn2_1_in_proj_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(39400704)))]; + tensor inner_decoder_norm21_1_bias = const()[name = tensor("inner_decoder_norm21_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(40187200)))]; + tensor inner_decoder_norm21_1_weight = const()[name = tensor("inner_decoder_norm21_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(40188288)))]; + tensor inner_decoder_linear1_1_bias = const()[name = tensor("inner_decoder_linear1_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(40189376)))]; + tensor inner_decoder_linear1_1_weight = const()[name = tensor("inner_decoder_linear1_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(40197632)))]; + tensor inner_decoder_linear2_1_bias = const()[name = tensor("inner_decoder_linear2_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(42294848)))]; + tensor inner_decoder_linear2_1_weight = const()[name = tensor("inner_decoder_linear2_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(42295936)))]; + tensor inner_decoder_norm22_1_bias = const()[name = tensor("inner_decoder_norm22_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(44393152)))]; + tensor inner_decoder_norm22_1_weight = const()[name = tensor("inner_decoder_norm22_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(44394240)))]; + tensor stacked_axes_0 = const()[name = tensor("stacked_axes_0"), val = tensor([1])]; + tensor stacked = expand_dims(axes = stacked_axes_0, x = features)[name = tensor("stacked")]; + tensor var_26 = const()[name = tensor("op_26"), val = tensor([1, 1, 345])]; + tensor input_1 = reshape(shape = var_26, x = stacked)[name = tensor("input_1")]; + tensor var_29 = const()[name = tensor("op_29"), val = tensor(0x1p+0)]; + tensor var_35 = const()[name = tensor("op_35"), val = tensor(true)]; + tensor var_36 = const()[name = tensor("op_36"), val = tensor(0x1.4f8b58p-17)]; + tensor var_39 = const()[name = tensor("op_39"), val = tensor(0)]; + tensor var_41 = const()[name = tensor("op_41"), val = tensor(2)]; + tensor var_42 = const()[name = tensor("op_42"), val = tensor(-1)]; + tensor var_44 = const()[name = tensor("op_44"), val = tensor(0x1.0c6f7ap-20)]; + tensor var_49 = const()[name = tensor("op_49"), val = tensor(0x1.5798eep-27)]; + tensor var_52 = const()[name = tensor("op_52"), val = tensor(1)]; + tensor input_3 = linear(bias = inner_encoder_input_projection_linear_bias, weight = inner_encoder_input_projection_linear_weight, x = input_1)[name = tensor("linear_0")]; tensor input_5_axes_0 = const()[name = tensor("input_5_axes_0"), val = tensor([-1])]; - tensor input_5 = layer_norm(axes = input_5_axes_0, beta = encoder_ffn1_0_module_sequential_0_bias, epsilon = var_28, gamma = encoder_ffn1_0_module_sequential_0_weight, x = input_3)[name = tensor("input_5")]; - tensor inputs_1 = linear(bias = encoder_ffn1_0_module_sequential_1_linear_bias, weight = encoder_ffn1_0_module_sequential_1_linear_weight, x = input_5)[name = tensor("linear_1")]; - tensor input_7 = silu(x = inputs_1)[name = tensor("input_7")]; - tensor input_11 = linear(bias = encoder_ffn1_0_module_sequential_4_linear_bias, weight = encoder_ffn1_0_module_sequential_4_linear_weight, x = input_7)[name = tensor("linear_2")]; - tensor var_147 = const()[name = tensor("op_147"), val = tensor(0x1p-1)]; - tensor var_148 = mul(x = input_11, y = var_147)[name = tensor("op_148")]; - tensor input_13 = add(x = var_148, y = input_3)[name = tensor("input_13")]; + tensor input_5 = layer_norm(axes = input_5_axes_0, beta = inner_encoder_pre_layer_norm_bias, epsilon = var_36, gamma = inner_encoder_pre_layer_norm_weight, x = input_3)[name = tensor("input_5")]; + tensor input_7_axes_0 = const()[name = tensor("input_7_axes_0"), val = tensor([-1])]; + tensor input_7 = layer_norm(axes = input_7_axes_0, beta = inner_encoder_ffn1_0_module_sequential_0_bias, epsilon = var_36, gamma = inner_encoder_ffn1_0_module_sequential_0_weight, x = input_5)[name = tensor("input_7")]; + tensor inputs_1 = linear(bias = inner_encoder_ffn1_0_module_sequential_1_linear_bias, weight = inner_encoder_ffn1_0_module_sequential_1_linear_weight, x = input_7)[name = tensor("linear_1")]; + tensor input_9 = silu(x = inputs_1)[name = tensor("input_9")]; + tensor input_13 = linear(bias = inner_encoder_ffn1_0_module_sequential_4_linear_bias, weight = inner_encoder_ffn1_0_module_sequential_4_linear_weight, x = input_9)[name = tensor("linear_2")]; + tensor var_173 = const()[name = tensor("op_173"), val = tensor(0x1p-1)]; + tensor var_174 = mul(x = input_13, y = var_173)[name = tensor("op_174")]; + tensor input_15 = add(x = var_174, y = input_5)[name = tensor("input_15")]; tensor x_1_axes_0 = const()[name = tensor("x_1_axes_0"), val = tensor([-1])]; - tensor x_1 = layer_norm(axes = x_1_axes_0, beta = encoder_ret_lns_0_bias, epsilon = var_28, gamma = encoder_ret_lns_0_weight, x = input_13)[name = tensor("x_1")]; + tensor x_1 = layer_norm(axes = x_1_axes_0, beta = inner_encoder_ret_lns_0_bias, epsilon = var_36, gamma = inner_encoder_ret_lns_0_weight, x = input_15)[name = tensor("x_1")]; tensor prev_kv_1_begin_0 = const()[name = tensor("prev_kv_1_begin_0"), val = tensor([0, 0, 0, 0, 0])]; tensor prev_kv_1_end_0 = const()[name = tensor("prev_kv_1_end_0"), val = tensor([1, 1, 4, 64, 64])]; tensor prev_kv_1_end_mask_0 = const()[name = tensor("prev_kv_1_end_mask_0"), val = tensor([false, true, true, true, true])]; @@ -238,139 +244,139 @@ program(1.0) tensor prev_scale_1_end_mask_0 = const()[name = tensor("prev_scale_1_end_mask_0"), val = tensor([false, true])]; tensor prev_scale_1_squeeze_mask_0 = const()[name = tensor("prev_scale_1_squeeze_mask_0"), val = tensor([true, false])]; tensor prev_scale_1 = slice_by_index(begin = prev_scale_1_begin_0, end = prev_scale_1_end_0, end_mask = prev_scale_1_end_mask_0, squeeze_mask = prev_scale_1_squeeze_mask_0, x = enc_scale)[name = tensor("prev_scale_1")]; - tensor var_162 = linear(bias = encoder_q_proj_0_bias, weight = encoder_q_proj_0_weight, x = x_1)[name = tensor("linear_3")]; - tensor var_163 = const()[name = tensor("op_163"), val = tensor([1, 1, 4, 64])]; - tensor var_164 = reshape(shape = var_163, x = var_162)[name = tensor("op_164")]; + tensor var_188 = linear(bias = inner_encoder_q_proj_0_bias, weight = inner_encoder_q_proj_0_weight, x = x_1)[name = tensor("linear_3")]; + tensor var_189 = const()[name = tensor("op_189"), val = tensor([1, 1, 4, 64])]; + tensor var_190 = reshape(shape = var_189, x = var_188)[name = tensor("op_190")]; tensor q_1_perm_0 = const()[name = tensor("q_1_perm_0"), val = tensor([0, 2, 1, 3])]; - tensor var_168 = linear(bias = encoder_k_proj_0_bias, weight = encoder_k_proj_0_weight, x = x_1)[name = tensor("linear_4")]; - tensor var_169 = const()[name = tensor("op_169"), val = tensor(0x1p-3)]; - tensor var_170 = mul(x = var_168, y = var_169)[name = tensor("op_170")]; - tensor var_171 = const()[name = tensor("op_171"), val = tensor([1, 1, 4, 64])]; - tensor var_172 = reshape(shape = var_171, x = var_170)[name = tensor("op_172")]; + tensor var_194 = linear(bias = inner_encoder_k_proj_0_bias, weight = inner_encoder_k_proj_0_weight, x = x_1)[name = tensor("linear_4")]; + tensor var_195 = const()[name = tensor("op_195"), val = tensor(0x1p-3)]; + tensor var_196 = mul(x = var_194, y = var_195)[name = tensor("op_196")]; + tensor var_197 = const()[name = tensor("op_197"), val = tensor([1, 1, 4, 64])]; + tensor var_198 = reshape(shape = var_197, x = var_196)[name = tensor("op_198")]; tensor k_1_perm_0 = const()[name = tensor("k_1_perm_0"), val = tensor([0, 2, 1, 3])]; - tensor var_176 = linear(bias = encoder_v_proj_0_bias, weight = encoder_v_proj_0_weight, x = x_1)[name = tensor("linear_5")]; - tensor var_177 = const()[name = tensor("op_177"), val = tensor([1, 1, 4, 64])]; - tensor var_178 = reshape(shape = var_177, x = var_176)[name = tensor("op_178")]; + tensor var_202 = linear(bias = inner_encoder_v_proj_0_bias, weight = inner_encoder_v_proj_0_weight, x = x_1)[name = tensor("linear_5")]; + tensor var_203 = const()[name = tensor("op_203"), val = tensor([1, 1, 4, 64])]; + tensor var_204 = reshape(shape = var_203, x = var_202)[name = tensor("op_204")]; tensor v_1_perm_0 = const()[name = tensor("v_1_perm_0"), val = tensor([0, 2, 1, 3])]; - tensor input_17 = linear(bias = encoder_g_proj_0_bias, weight = encoder_g_proj_0_weight, x = x_1)[name = tensor("linear_6")]; + tensor input_19 = linear(bias = inner_encoder_g_proj_0_bias, weight = inner_encoder_g_proj_0_weight, x = x_1)[name = tensor("linear_6")]; tensor sqrt_s0_1 = sqrt(x = prev_scale_1)[name = tensor("sqrt_s0_1")]; - tensor s_t_1 = add(x = prev_scale_1, y = encoder__t_index)[name = tensor("s_t_1")]; + tensor s_t_1 = add(x = prev_scale_1, y = inner_encoder__t_index)[name = tensor("s_t_1")]; tensor sqrt_s_t_1 = sqrt(x = s_t_1)[name = tensor("sqrt_s_t_1")]; tensor qk_1_transpose_x_1 = const()[name = tensor("qk_1_transpose_x_1"), val = tensor(false)]; tensor qk_1_transpose_y_1 = const()[name = tensor("qk_1_transpose_y_1"), val = tensor(true)]; - tensor k_1 = transpose(perm = k_1_perm_0, x = var_172)[name = tensor("transpose_57")]; - tensor q_1 = transpose(perm = q_1_perm_0, x = var_164)[name = tensor("transpose_58")]; + tensor k_1 = transpose(perm = k_1_perm_0, x = var_198)[name = tensor("transpose_57")]; + tensor q_1 = transpose(perm = q_1_perm_0, x = var_190)[name = tensor("transpose_58")]; tensor qk_1 = matmul(transpose_x = qk_1_transpose_x_1, transpose_y = qk_1_transpose_y_1, x = q_1, y = k_1)[name = tensor("qk_1")]; - tensor var_188 = const()[name = tensor("op_188"), val = tensor([1, 1])]; - tensor var_189 = reshape(shape = var_188, x = sqrt_s_t_1)[name = tensor("op_189")]; - tensor M_1 = real_div(x = encoder__causal_mask, y = var_189)[name = tensor("M_1")]; - tensor var_191 = mul(x = qk_1, y = M_1)[name = tensor("op_191")]; + tensor var_214 = const()[name = tensor("op_214"), val = tensor([1, 1])]; + tensor var_215 = reshape(shape = var_214, x = sqrt_s_t_1)[name = tensor("op_215")]; + tensor M_1 = real_div(x = inner_encoder__causal_mask, y = var_215)[name = tensor("M_1")]; + tensor var_217 = mul(x = qk_1, y = M_1)[name = tensor("op_217")]; tensor inner_1_transpose_x_0 = const()[name = tensor("inner_1_transpose_x_0"), val = tensor(false)]; tensor inner_1_transpose_y_0 = const()[name = tensor("inner_1_transpose_y_0"), val = tensor(false)]; - tensor v_1 = transpose(perm = v_1_perm_0, x = var_178)[name = tensor("transpose_56")]; - tensor inner_1 = matmul(transpose_x = inner_1_transpose_x_0, transpose_y = inner_1_transpose_y_0, x = var_191, y = v_1)[name = tensor("inner_1")]; - tensor var_193_transpose_x_0 = const()[name = tensor("op_193_transpose_x_0"), val = tensor(false)]; - tensor var_193_transpose_y_0 = const()[name = tensor("op_193_transpose_y_0"), val = tensor(false)]; - tensor var_193 = matmul(transpose_x = var_193_transpose_x_0, transpose_y = var_193_transpose_y_0, x = q_1, y = prev_kv_1)[name = tensor("op_193")]; - tensor var_194 = real_div(x = sqrt_s0_1, y = sqrt_s_t_1)[name = tensor("op_194")]; - tensor var_195 = const()[name = tensor("op_195"), val = tensor([1, 1, 1, 1])]; - tensor var_196 = reshape(shape = var_195, x = var_194)[name = tensor("op_196")]; - tensor cross_1 = mul(x = var_193, y = var_196)[name = tensor("cross_1")]; + tensor v_1 = transpose(perm = v_1_perm_0, x = var_204)[name = tensor("transpose_56")]; + tensor inner_1 = matmul(transpose_x = inner_1_transpose_x_0, transpose_y = inner_1_transpose_y_0, x = var_217, y = v_1)[name = tensor("inner_1")]; + tensor var_219_transpose_x_0 = const()[name = tensor("op_219_transpose_x_0"), val = tensor(false)]; + tensor var_219_transpose_y_0 = const()[name = tensor("op_219_transpose_y_0"), val = tensor(false)]; + tensor var_219 = matmul(transpose_x = var_219_transpose_x_0, transpose_y = var_219_transpose_y_0, x = q_1, y = prev_kv_1)[name = tensor("op_219")]; + tensor var_220 = real_div(x = sqrt_s0_1, y = sqrt_s_t_1)[name = tensor("op_220")]; + tensor var_221 = const()[name = tensor("op_221"), val = tensor([1, 1, 1, 1])]; + tensor var_222 = reshape(shape = var_221, x = var_220)[name = tensor("op_222")]; + tensor cross_1 = mul(x = var_219, y = var_222)[name = tensor("cross_1")]; tensor out_1 = add(x = inner_1, y = cross_1)[name = tensor("out_1")]; - tensor var_199 = mul(x = prev_kv_1, y = sqrt_s0_1)[name = tensor("op_199")]; - tensor var_201_transpose_x_1 = const()[name = tensor("op_201_transpose_x_1"), val = tensor(true)]; - tensor var_201_transpose_y_1 = const()[name = tensor("op_201_transpose_y_1"), val = tensor(false)]; - tensor var_201 = matmul(transpose_x = var_201_transpose_x_1, transpose_y = var_201_transpose_y_1, x = k_1, y = v_1)[name = tensor("op_201")]; - tensor new_kv_unnorm_1 = add(x = var_199, y = var_201)[name = tensor("new_kv_unnorm_1")]; - tensor var_203 = const()[name = tensor("op_203"), val = tensor(0x1p+0)]; - tensor new_scale_1 = add(x = prev_scale_1, y = var_203)[name = tensor("new_scale_1")]; - tensor var_205 = sqrt(x = new_scale_1)[name = tensor("op_205")]; - tensor var_206 = real_div(x = new_kv_unnorm_1, y = var_205)[name = tensor("op_206")]; - tensor var_207_perm_0 = const()[name = tensor("op_207_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_225 = mul(x = prev_kv_1, y = sqrt_s0_1)[name = tensor("op_225")]; + tensor var_227_transpose_x_1 = const()[name = tensor("op_227_transpose_x_1"), val = tensor(true)]; + tensor var_227_transpose_y_1 = const()[name = tensor("op_227_transpose_y_1"), val = tensor(false)]; + tensor var_227 = matmul(transpose_x = var_227_transpose_x_1, transpose_y = var_227_transpose_y_1, x = k_1, y = v_1)[name = tensor("op_227")]; + tensor new_kv_unnorm_1 = add(x = var_225, y = var_227)[name = tensor("new_kv_unnorm_1")]; + tensor var_229 = const()[name = tensor("op_229"), val = tensor(0x1p+0)]; + tensor new_scale_1 = add(x = prev_scale_1, y = var_229)[name = tensor("new_scale_1")]; + tensor var_231 = sqrt(x = new_scale_1)[name = tensor("op_231")]; + tensor var_232 = real_div(x = new_kv_unnorm_1, y = var_231)[name = tensor("op_232")]; + tensor var_233_perm_0 = const()[name = tensor("op_233_perm_0"), val = tensor([0, 2, 1, 3])]; tensor out_3_axes_0 = const()[name = tensor("out_3_axes_0"), val = tensor([-1])]; - tensor var_207 = transpose(perm = var_207_perm_0, x = out_1)[name = tensor("transpose_55")]; - tensor out_3 = layer_norm(axes = out_3_axes_0, epsilon = var_17, x = var_207)[name = tensor("out_3")]; - tensor var_211 = const()[name = tensor("op_211"), val = tensor([1, 1, 256])]; - tensor out_5 = reshape(shape = var_211, x = out_3)[name = tensor("out_5")]; - tensor var_213 = silu(x = input_17)[name = tensor("op_213")]; - tensor input_19 = mul(x = var_213, y = out_5)[name = tensor("input_19")]; - tensor ret_out_1 = linear(bias = encoder_out_proj_0_bias, weight = encoder_out_proj_0_weight, x = input_19)[name = tensor("linear_7")]; - tensor x_3 = add(x = input_13, y = ret_out_1)[name = tensor("x_3")]; + tensor var_233 = transpose(perm = var_233_perm_0, x = out_1)[name = tensor("transpose_55")]; + tensor out_3 = layer_norm(axes = out_3_axes_0, epsilon = var_44, x = var_233)[name = tensor("out_3")]; + tensor var_237 = const()[name = tensor("op_237"), val = tensor([1, 1, 256])]; + tensor out_5 = reshape(shape = var_237, x = out_3)[name = tensor("out_5")]; + tensor var_239 = silu(x = input_19)[name = tensor("op_239")]; + tensor input_21 = mul(x = var_239, y = out_5)[name = tensor("input_21")]; + tensor ret_out_1 = linear(bias = inner_encoder_out_proj_0_bias, weight = inner_encoder_out_proj_0_weight, x = input_21)[name = tensor("linear_7")]; + tensor x_3 = add(x = input_15, y = ret_out_1)[name = tensor("x_3")]; tensor window_1_begin_0 = const()[name = tensor("window_1_begin_0"), val = tensor([0, 0, 0, 0])]; tensor window_1_end_0 = const()[name = tensor("window_1_end_0"), val = tensor([1, 1, 16, 256])]; tensor window_1_end_mask_0 = const()[name = tensor("window_1_end_mask_0"), val = tensor([false, true, true, true])]; tensor window_1_squeeze_mask_0 = const()[name = tensor("window_1_squeeze_mask_0"), val = tensor([true, false, false, false])]; tensor window_1 = slice_by_index(begin = window_1_begin_0, end = window_1_end_0, end_mask = window_1_end_mask_0, squeeze_mask = window_1_squeeze_mask_0, x = enc_conv_cache)[name = tensor("window_1")]; - tensor var_224_begin_0 = const()[name = tensor("op_224_begin_0"), val = tensor([0, 1, 0])]; - tensor var_224_end_0 = const()[name = tensor("op_224_end_0"), val = tensor([1, 16, 256])]; - tensor var_224_end_mask_0 = const()[name = tensor("op_224_end_mask_0"), val = tensor([true, true, true])]; - tensor var_224 = slice_by_index(begin = var_224_begin_0, end = var_224_end_0, end_mask = var_224_end_mask_0, x = window_1)[name = tensor("op_224")]; + tensor var_250_begin_0 = const()[name = tensor("op_250_begin_0"), val = tensor([0, 1, 0])]; + tensor var_250_end_0 = const()[name = tensor("op_250_end_0"), val = tensor([1, 16, 256])]; + tensor var_250_end_mask_0 = const()[name = tensor("op_250_end_mask_0"), val = tensor([true, true, true])]; + tensor var_250 = slice_by_index(begin = var_250_begin_0, end = var_250_end_0, end_mask = var_250_end_mask_0, x = window_1)[name = tensor("op_250")]; tensor window_3_interleave_0 = const()[name = tensor("window_3_interleave_0"), val = tensor(false)]; - tensor window_3 = concat(axis = var_26, interleave = window_3_interleave_0, values = (var_224, x_3))[name = tensor("window_3")]; - tensor input_21_interleave_0 = const()[name = tensor("input_21_interleave_0"), val = tensor(false)]; - tensor input_21 = concat(axis = var_23, interleave = input_21_interleave_0, values = window_3)[name = tensor("input_21")]; + tensor window_3 = concat(axis = var_52, interleave = window_3_interleave_0, values = (var_250, x_3))[name = tensor("window_3")]; + tensor input_23_interleave_0 = const()[name = tensor("input_23_interleave_0"), val = tensor(false)]; + tensor input_23 = concat(axis = var_39, interleave = input_23_interleave_0, values = window_3)[name = tensor("input_23")]; tensor x_5_axes_0 = const()[name = tensor("x_5_axes_0"), val = tensor([-1])]; - tensor x_5 = layer_norm(axes = x_5_axes_0, beta = encoder_conv_module_0_sequential_0_bias, epsilon = var_28, gamma = encoder_conv_module_0_sequential_0_weight, x = input_21)[name = tensor("x_5")]; - tensor input_23_perm_0 = const()[name = tensor("input_23_perm_0"), val = tensor([0, 2, 1])]; + tensor x_5 = layer_norm(axes = x_5_axes_0, beta = inner_encoder_conv_module_0_sequential_0_bias, epsilon = var_36, gamma = inner_encoder_conv_module_0_sequential_0_weight, x = input_23)[name = tensor("x_5")]; + tensor input_25_perm_0 = const()[name = tensor("input_25_perm_0"), val = tensor([0, 2, 1])]; tensor inputs_3_pad_type_0 = const()[name = tensor("inputs_3_pad_type_0"), val = tensor("valid")]; tensor inputs_3_strides_0 = const()[name = tensor("inputs_3_strides_0"), val = tensor([1])]; tensor inputs_3_pad_0 = const()[name = tensor("inputs_3_pad_0"), val = tensor([0, 0])]; tensor inputs_3_dilations_0 = const()[name = tensor("inputs_3_dilations_0"), val = tensor([1])]; tensor inputs_3_groups_0 = const()[name = tensor("inputs_3_groups_0"), val = tensor(1)]; - tensor input_23 = transpose(perm = input_23_perm_0, x = x_5)[name = tensor("transpose_54")]; - tensor inputs_3 = conv(bias = encoder_conv_module_0_sequential_2_conv_bias, dilations = inputs_3_dilations_0, groups = inputs_3_groups_0, pad = inputs_3_pad_0, pad_type = inputs_3_pad_type_0, strides = inputs_3_strides_0, weight = encoder_conv_module_0_sequential_2_conv_weight, x = input_23)[name = tensor("inputs_3")]; - tensor var_249_split_sizes_0 = const()[name = tensor("op_249_split_sizes_0"), val = tensor([256, 256])]; - tensor var_249_axis_0 = const()[name = tensor("op_249_axis_0"), val = tensor(1)]; - tensor var_249_0, tensor var_249_1 = split(axis = var_249_axis_0, split_sizes = var_249_split_sizes_0, x = inputs_3)[name = tensor("op_249")]; - tensor var_251 = sigmoid(x = var_249_1)[name = tensor("op_251")]; - tensor inputs_5 = mul(x = var_249_0, y = var_251)[name = tensor("inputs_5")]; + tensor input_25 = transpose(perm = input_25_perm_0, x = x_5)[name = tensor("transpose_54")]; + tensor inputs_3 = conv(bias = inner_encoder_conv_module_0_sequential_2_conv_bias, dilations = inputs_3_dilations_0, groups = inputs_3_groups_0, pad = inputs_3_pad_0, pad_type = inputs_3_pad_type_0, strides = inputs_3_strides_0, weight = inner_encoder_conv_module_0_sequential_2_conv_weight, x = input_25)[name = tensor("inputs_3")]; + tensor var_275_split_sizes_0 = const()[name = tensor("op_275_split_sizes_0"), val = tensor([256, 256])]; + tensor var_275_axis_0 = const()[name = tensor("op_275_axis_0"), val = tensor(1)]; + tensor var_275_0, tensor var_275_1 = split(axis = var_275_axis_0, split_sizes = var_275_split_sizes_0, x = inputs_3)[name = tensor("op_275")]; + tensor var_277 = sigmoid(x = var_275_1)[name = tensor("op_277")]; + tensor inputs_5 = mul(x = var_275_0, y = var_277)[name = tensor("inputs_5")]; tensor outputs_aug_1_pad_type_0 = const()[name = tensor("outputs_aug_1_pad_type_0"), val = tensor("custom")]; tensor outputs_aug_1_pad_0 = const()[name = tensor("outputs_aug_1_pad_0"), val = tensor([15, 15])]; tensor outputs_aug_1_groups_0 = const()[name = tensor("outputs_aug_1_groups_0"), val = tensor(256)]; tensor outputs_aug_1_strides_0 = const()[name = tensor("outputs_aug_1_strides_0"), val = tensor([1])]; tensor outputs_aug_1_dilations_0 = const()[name = tensor("outputs_aug_1_dilations_0"), val = tensor([1])]; - tensor outputs_aug_1 = conv(dilations = outputs_aug_1_dilations_0, groups = outputs_aug_1_groups_0, pad = outputs_aug_1_pad_0, pad_type = outputs_aug_1_pad_type_0, strides = outputs_aug_1_strides_0, weight = encoder_conv_module_0_sequential_4_conv_weight, x = inputs_5)[name = tensor("outputs_aug_1")]; - tensor input_25_begin_0 = const()[name = tensor("input_25_begin_0"), val = tensor([0, 0, 0])]; - tensor input_25_end_0 = const()[name = tensor("input_25_end_0"), val = tensor([1, 256, 16])]; - tensor input_25_end_mask_0 = const()[name = tensor("input_25_end_mask_0"), val = tensor([true, true, false])]; - tensor input_25 = slice_by_index(begin = input_25_begin_0, end = input_25_end_0, end_mask = input_25_end_mask_0, x = outputs_aug_1)[name = tensor("input_25")]; - tensor inputs_7 = batch_norm(beta = encoder_conv_module_0_sequential_5_bias, epsilon = var_28, gamma = encoder_conv_module_0_sequential_5_weight, mean = encoder_conv_module_0_sequential_5_running_mean, variance = encoder_conv_module_0_sequential_5_running_var, x = input_25)[name = tensor("inputs_7")]; - tensor input_27 = silu(x = inputs_7)[name = tensor("input_27")]; - tensor input_29_pad_type_0 = const()[name = tensor("input_29_pad_type_0"), val = tensor("valid")]; - tensor input_29_strides_0 = const()[name = tensor("input_29_strides_0"), val = tensor([1])]; - tensor input_29_pad_0 = const()[name = tensor("input_29_pad_0"), val = tensor([0, 0])]; - tensor input_29_dilations_0 = const()[name = tensor("input_29_dilations_0"), val = tensor([1])]; - tensor input_29_groups_0 = const()[name = tensor("input_29_groups_0"), val = tensor(1)]; - tensor input_29 = conv(bias = encoder_conv_module_0_sequential_7_conv_bias, dilations = input_29_dilations_0, groups = input_29_groups_0, pad = input_29_pad_0, pad_type = input_29_pad_type_0, strides = input_29_strides_0, weight = encoder_conv_module_0_sequential_7_conv_weight, x = input_27)[name = tensor("input_29")]; + tensor outputs_aug_1 = conv(dilations = outputs_aug_1_dilations_0, groups = outputs_aug_1_groups_0, pad = outputs_aug_1_pad_0, pad_type = outputs_aug_1_pad_type_0, strides = outputs_aug_1_strides_0, weight = inner_encoder_conv_module_0_sequential_4_conv_weight, x = inputs_5)[name = tensor("outputs_aug_1")]; + tensor input_27_begin_0 = const()[name = tensor("input_27_begin_0"), val = tensor([0, 0, 0])]; + tensor input_27_end_0 = const()[name = tensor("input_27_end_0"), val = tensor([1, 256, 16])]; + tensor input_27_end_mask_0 = const()[name = tensor("input_27_end_mask_0"), val = tensor([true, true, false])]; + tensor input_27 = slice_by_index(begin = input_27_begin_0, end = input_27_end_0, end_mask = input_27_end_mask_0, x = outputs_aug_1)[name = tensor("input_27")]; + tensor inputs_7 = batch_norm(beta = inner_encoder_conv_module_0_sequential_5_bias, epsilon = var_36, gamma = inner_encoder_conv_module_0_sequential_5_weight, mean = inner_encoder_conv_module_0_sequential_5_running_mean, variance = inner_encoder_conv_module_0_sequential_5_running_var, x = input_27)[name = tensor("inputs_7")]; + tensor input_29 = silu(x = inputs_7)[name = tensor("input_29")]; + tensor input_31_pad_type_0 = const()[name = tensor("input_31_pad_type_0"), val = tensor("valid")]; + tensor input_31_strides_0 = const()[name = tensor("input_31_strides_0"), val = tensor([1])]; + tensor input_31_pad_0 = const()[name = tensor("input_31_pad_0"), val = tensor([0, 0])]; + tensor input_31_dilations_0 = const()[name = tensor("input_31_dilations_0"), val = tensor([1])]; + tensor input_31_groups_0 = const()[name = tensor("input_31_groups_0"), val = tensor(1)]; + tensor input_31 = conv(bias = inner_encoder_conv_module_0_sequential_7_conv_bias, dilations = input_31_dilations_0, groups = input_31_groups_0, pad = input_31_pad_0, pad_type = input_31_pad_type_0, strides = input_31_strides_0, weight = inner_encoder_conv_module_0_sequential_7_conv_weight, x = input_29)[name = tensor("input_31")]; tensor conv_out_1_perm_0 = const()[name = tensor("conv_out_1_perm_0"), val = tensor([0, 2, 1])]; - tensor var_282_begin_0 = const()[name = tensor("op_282_begin_0"), val = tensor([0, -1, 0])]; - tensor var_282_end_0 = const()[name = tensor("op_282_end_0"), val = tensor([1, 16, 256])]; - tensor var_282_end_mask_0 = const()[name = tensor("op_282_end_mask_0"), val = tensor([true, true, true])]; - tensor conv_out_1 = transpose(perm = conv_out_1_perm_0, x = input_29)[name = tensor("transpose_53")]; - tensor var_282 = slice_by_index(begin = var_282_begin_0, end = var_282_end_0, end_mask = var_282_end_mask_0, x = conv_out_1)[name = tensor("op_282")]; - tensor var_284_perm_0 = const()[name = tensor("op_284_perm_0"), val = tensor([1, 0, 2])]; - tensor var_284 = transpose(perm = var_284_perm_0, x = var_282)[name = tensor("transpose_52")]; - tensor input_31 = add(x = x_3, y = var_284)[name = tensor("input_31")]; - tensor input_33_axes_0 = const()[name = tensor("input_33_axes_0"), val = tensor([-1])]; - tensor input_33 = layer_norm(axes = input_33_axes_0, beta = encoder_ffn2_0_module_sequential_0_bias, epsilon = var_28, gamma = encoder_ffn2_0_module_sequential_0_weight, x = input_31)[name = tensor("input_33")]; - tensor inputs_9 = linear(bias = encoder_ffn2_0_module_sequential_1_linear_bias, weight = encoder_ffn2_0_module_sequential_1_linear_weight, x = input_33)[name = tensor("linear_8")]; - tensor input_35 = silu(x = inputs_9)[name = tensor("input_35")]; - tensor input_39 = linear(bias = encoder_ffn2_0_module_sequential_4_linear_bias, weight = encoder_ffn2_0_module_sequential_4_linear_weight, x = input_35)[name = tensor("linear_9")]; - tensor var_307 = const()[name = tensor("op_307"), val = tensor(0x1p-1)]; - tensor var_308 = mul(x = input_39, y = var_307)[name = tensor("op_308")]; - tensor input_41 = add(x = var_308, y = input_31)[name = tensor("input_41")]; - tensor input_43_axes_0 = const()[name = tensor("input_43_axes_0"), val = tensor([-1])]; - tensor input_43 = layer_norm(axes = input_43_axes_0, beta = encoder_layer_norm_0_bias, epsilon = var_28, gamma = encoder_layer_norm_0_weight, x = input_41)[name = tensor("input_43")]; + tensor var_308_begin_0 = const()[name = tensor("op_308_begin_0"), val = tensor([0, -1, 0])]; + tensor var_308_end_0 = const()[name = tensor("op_308_end_0"), val = tensor([1, 16, 256])]; + tensor var_308_end_mask_0 = const()[name = tensor("op_308_end_mask_0"), val = tensor([true, true, true])]; + tensor conv_out_1 = transpose(perm = conv_out_1_perm_0, x = input_31)[name = tensor("transpose_53")]; + tensor var_308 = slice_by_index(begin = var_308_begin_0, end = var_308_end_0, end_mask = var_308_end_mask_0, x = conv_out_1)[name = tensor("op_308")]; + tensor var_310_perm_0 = const()[name = tensor("op_310_perm_0"), val = tensor([1, 0, 2])]; + tensor var_310 = transpose(perm = var_310_perm_0, x = var_308)[name = tensor("transpose_52")]; + tensor input_33 = add(x = x_3, y = var_310)[name = tensor("input_33")]; + tensor input_35_axes_0 = const()[name = tensor("input_35_axes_0"), val = tensor([-1])]; + tensor input_35 = layer_norm(axes = input_35_axes_0, beta = inner_encoder_ffn2_0_module_sequential_0_bias, epsilon = var_36, gamma = inner_encoder_ffn2_0_module_sequential_0_weight, x = input_33)[name = tensor("input_35")]; + tensor inputs_9 = linear(bias = inner_encoder_ffn2_0_module_sequential_1_linear_bias, weight = inner_encoder_ffn2_0_module_sequential_1_linear_weight, x = input_35)[name = tensor("linear_8")]; + tensor input_37 = silu(x = inputs_9)[name = tensor("input_37")]; + tensor input_41 = linear(bias = inner_encoder_ffn2_0_module_sequential_4_linear_bias, weight = inner_encoder_ffn2_0_module_sequential_4_linear_weight, x = input_37)[name = tensor("linear_9")]; + tensor var_333 = const()[name = tensor("op_333"), val = tensor(0x1p-1)]; + tensor var_334 = mul(x = input_41, y = var_333)[name = tensor("op_334")]; + tensor input_43 = add(x = var_334, y = input_33)[name = tensor("input_43")]; tensor input_45_axes_0 = const()[name = tensor("input_45_axes_0"), val = tensor([-1])]; - tensor input_45 = layer_norm(axes = input_45_axes_0, beta = encoder_ffn1_1_module_sequential_0_bias, epsilon = var_28, gamma = encoder_ffn1_1_module_sequential_0_weight, x = input_43)[name = tensor("input_45")]; - tensor inputs_11 = linear(bias = encoder_ffn1_1_module_sequential_1_linear_bias, weight = encoder_ffn1_1_module_sequential_1_linear_weight, x = input_45)[name = tensor("linear_10")]; - tensor input_47 = silu(x = inputs_11)[name = tensor("input_47")]; - tensor input_51 = linear(bias = encoder_ffn1_1_module_sequential_4_linear_bias, weight = encoder_ffn1_1_module_sequential_4_linear_weight, x = input_47)[name = tensor("linear_11")]; - tensor var_337 = const()[name = tensor("op_337"), val = tensor(0x1p-1)]; - tensor var_338 = mul(x = input_51, y = var_337)[name = tensor("op_338")]; - tensor input_53 = add(x = var_338, y = input_43)[name = tensor("input_53")]; + tensor input_45 = layer_norm(axes = input_45_axes_0, beta = inner_encoder_layer_norm_0_bias, epsilon = var_36, gamma = inner_encoder_layer_norm_0_weight, x = input_43)[name = tensor("input_45")]; + tensor input_47_axes_0 = const()[name = tensor("input_47_axes_0"), val = tensor([-1])]; + tensor input_47 = layer_norm(axes = input_47_axes_0, beta = inner_encoder_ffn1_1_module_sequential_0_bias, epsilon = var_36, gamma = inner_encoder_ffn1_1_module_sequential_0_weight, x = input_45)[name = tensor("input_47")]; + tensor inputs_11 = linear(bias = inner_encoder_ffn1_1_module_sequential_1_linear_bias, weight = inner_encoder_ffn1_1_module_sequential_1_linear_weight, x = input_47)[name = tensor("linear_10")]; + tensor input_49 = silu(x = inputs_11)[name = tensor("input_49")]; + tensor input_53 = linear(bias = inner_encoder_ffn1_1_module_sequential_4_linear_bias, weight = inner_encoder_ffn1_1_module_sequential_4_linear_weight, x = input_49)[name = tensor("linear_11")]; + tensor var_363 = const()[name = tensor("op_363"), val = tensor(0x1p-1)]; + tensor var_364 = mul(x = input_53, y = var_363)[name = tensor("op_364")]; + tensor input_55 = add(x = var_364, y = input_45)[name = tensor("input_55")]; tensor x_7_axes_0 = const()[name = tensor("x_7_axes_0"), val = tensor([-1])]; - tensor x_7 = layer_norm(axes = x_7_axes_0, beta = encoder_ret_lns_1_bias, epsilon = var_28, gamma = encoder_ret_lns_1_weight, x = input_53)[name = tensor("x_7")]; + tensor x_7 = layer_norm(axes = x_7_axes_0, beta = inner_encoder_ret_lns_1_bias, epsilon = var_36, gamma = inner_encoder_ret_lns_1_weight, x = input_55)[name = tensor("x_7")]; tensor prev_kv_3_begin_0 = const()[name = tensor("prev_kv_3_begin_0"), val = tensor([1, 0, 0, 0, 0])]; tensor prev_kv_3_end_0 = const()[name = tensor("prev_kv_3_end_0"), val = tensor([2, 1, 4, 64, 64])]; tensor prev_kv_3_end_mask_0 = const()[name = tensor("prev_kv_3_end_mask_0"), val = tensor([false, true, true, true, true])]; @@ -381,139 +387,139 @@ program(1.0) tensor prev_scale_3_end_mask_0 = const()[name = tensor("prev_scale_3_end_mask_0"), val = tensor([false, true])]; tensor prev_scale_3_squeeze_mask_0 = const()[name = tensor("prev_scale_3_squeeze_mask_0"), val = tensor([true, false])]; tensor prev_scale_3 = slice_by_index(begin = prev_scale_3_begin_0, end = prev_scale_3_end_0, end_mask = prev_scale_3_end_mask_0, squeeze_mask = prev_scale_3_squeeze_mask_0, x = enc_scale)[name = tensor("prev_scale_3")]; - tensor var_352 = linear(bias = encoder_q_proj_1_bias, weight = encoder_q_proj_1_weight, x = x_7)[name = tensor("linear_12")]; - tensor var_353 = const()[name = tensor("op_353"), val = tensor([1, 1, 4, 64])]; - tensor var_354 = reshape(shape = var_353, x = var_352)[name = tensor("op_354")]; + tensor var_378 = linear(bias = inner_encoder_q_proj_1_bias, weight = inner_encoder_q_proj_1_weight, x = x_7)[name = tensor("linear_12")]; + tensor var_379 = const()[name = tensor("op_379"), val = tensor([1, 1, 4, 64])]; + tensor var_380 = reshape(shape = var_379, x = var_378)[name = tensor("op_380")]; tensor q_3_perm_0 = const()[name = tensor("q_3_perm_0"), val = tensor([0, 2, 1, 3])]; - tensor var_358 = linear(bias = encoder_k_proj_1_bias, weight = encoder_k_proj_1_weight, x = x_7)[name = tensor("linear_13")]; - tensor var_359 = const()[name = tensor("op_359"), val = tensor(0x1p-3)]; - tensor var_360 = mul(x = var_358, y = var_359)[name = tensor("op_360")]; - tensor var_361 = const()[name = tensor("op_361"), val = tensor([1, 1, 4, 64])]; - tensor var_362 = reshape(shape = var_361, x = var_360)[name = tensor("op_362")]; + tensor var_384 = linear(bias = inner_encoder_k_proj_1_bias, weight = inner_encoder_k_proj_1_weight, x = x_7)[name = tensor("linear_13")]; + tensor var_385 = const()[name = tensor("op_385"), val = tensor(0x1p-3)]; + tensor var_386 = mul(x = var_384, y = var_385)[name = tensor("op_386")]; + tensor var_387 = const()[name = tensor("op_387"), val = tensor([1, 1, 4, 64])]; + tensor var_388 = reshape(shape = var_387, x = var_386)[name = tensor("op_388")]; tensor k_3_perm_0 = const()[name = tensor("k_3_perm_0"), val = tensor([0, 2, 1, 3])]; - tensor var_366 = linear(bias = encoder_v_proj_1_bias, weight = encoder_v_proj_1_weight, x = x_7)[name = tensor("linear_14")]; - tensor var_367 = const()[name = tensor("op_367"), val = tensor([1, 1, 4, 64])]; - tensor var_368 = reshape(shape = var_367, x = var_366)[name = tensor("op_368")]; + tensor var_392 = linear(bias = inner_encoder_v_proj_1_bias, weight = inner_encoder_v_proj_1_weight, x = x_7)[name = tensor("linear_14")]; + tensor var_393 = const()[name = tensor("op_393"), val = tensor([1, 1, 4, 64])]; + tensor var_394 = reshape(shape = var_393, x = var_392)[name = tensor("op_394")]; tensor v_3_perm_0 = const()[name = tensor("v_3_perm_0"), val = tensor([0, 2, 1, 3])]; - tensor input_57 = linear(bias = encoder_g_proj_1_bias, weight = encoder_g_proj_1_weight, x = x_7)[name = tensor("linear_15")]; + tensor input_59 = linear(bias = inner_encoder_g_proj_1_bias, weight = inner_encoder_g_proj_1_weight, x = x_7)[name = tensor("linear_15")]; tensor sqrt_s0_3 = sqrt(x = prev_scale_3)[name = tensor("sqrt_s0_3")]; - tensor s_t_3 = add(x = prev_scale_3, y = encoder__t_index)[name = tensor("s_t_3")]; + tensor s_t_3 = add(x = prev_scale_3, y = inner_encoder__t_index)[name = tensor("s_t_3")]; tensor sqrt_s_t_3 = sqrt(x = s_t_3)[name = tensor("sqrt_s_t_3")]; tensor qk_3_transpose_x_1 = const()[name = tensor("qk_3_transpose_x_1"), val = tensor(false)]; tensor qk_3_transpose_y_1 = const()[name = tensor("qk_3_transpose_y_1"), val = tensor(true)]; - tensor k_3 = transpose(perm = k_3_perm_0, x = var_362)[name = tensor("transpose_50")]; - tensor q_3 = transpose(perm = q_3_perm_0, x = var_354)[name = tensor("transpose_51")]; + tensor k_3 = transpose(perm = k_3_perm_0, x = var_388)[name = tensor("transpose_50")]; + tensor q_3 = transpose(perm = q_3_perm_0, x = var_380)[name = tensor("transpose_51")]; tensor qk_3 = matmul(transpose_x = qk_3_transpose_x_1, transpose_y = qk_3_transpose_y_1, x = q_3, y = k_3)[name = tensor("qk_3")]; - tensor var_378 = const()[name = tensor("op_378"), val = tensor([1, 1])]; - tensor var_379 = reshape(shape = var_378, x = sqrt_s_t_3)[name = tensor("op_379")]; - tensor M_3 = real_div(x = encoder__causal_mask, y = var_379)[name = tensor("M_3")]; - tensor var_381 = mul(x = qk_3, y = M_3)[name = tensor("op_381")]; + tensor var_404 = const()[name = tensor("op_404"), val = tensor([1, 1])]; + tensor var_405 = reshape(shape = var_404, x = sqrt_s_t_3)[name = tensor("op_405")]; + tensor M_3 = real_div(x = inner_encoder__causal_mask, y = var_405)[name = tensor("M_3")]; + tensor var_407 = mul(x = qk_3, y = M_3)[name = tensor("op_407")]; tensor inner_3_transpose_x_0 = const()[name = tensor("inner_3_transpose_x_0"), val = tensor(false)]; tensor inner_3_transpose_y_0 = const()[name = tensor("inner_3_transpose_y_0"), val = tensor(false)]; - tensor v_3 = transpose(perm = v_3_perm_0, x = var_368)[name = tensor("transpose_49")]; - tensor inner_3 = matmul(transpose_x = inner_3_transpose_x_0, transpose_y = inner_3_transpose_y_0, x = var_381, y = v_3)[name = tensor("inner_3")]; - tensor var_383_transpose_x_0 = const()[name = tensor("op_383_transpose_x_0"), val = tensor(false)]; - tensor var_383_transpose_y_0 = const()[name = tensor("op_383_transpose_y_0"), val = tensor(false)]; - tensor var_383 = matmul(transpose_x = var_383_transpose_x_0, transpose_y = var_383_transpose_y_0, x = q_3, y = prev_kv_3)[name = tensor("op_383")]; - tensor var_384 = real_div(x = sqrt_s0_3, y = sqrt_s_t_3)[name = tensor("op_384")]; - tensor var_385 = const()[name = tensor("op_385"), val = tensor([1, 1, 1, 1])]; - tensor var_386 = reshape(shape = var_385, x = var_384)[name = tensor("op_386")]; - tensor cross_3 = mul(x = var_383, y = var_386)[name = tensor("cross_3")]; + tensor v_3 = transpose(perm = v_3_perm_0, x = var_394)[name = tensor("transpose_49")]; + tensor inner_3 = matmul(transpose_x = inner_3_transpose_x_0, transpose_y = inner_3_transpose_y_0, x = var_407, y = v_3)[name = tensor("inner_3")]; + tensor var_409_transpose_x_0 = const()[name = tensor("op_409_transpose_x_0"), val = tensor(false)]; + tensor var_409_transpose_y_0 = const()[name = tensor("op_409_transpose_y_0"), val = tensor(false)]; + tensor var_409 = matmul(transpose_x = var_409_transpose_x_0, transpose_y = var_409_transpose_y_0, x = q_3, y = prev_kv_3)[name = tensor("op_409")]; + tensor var_410 = real_div(x = sqrt_s0_3, y = sqrt_s_t_3)[name = tensor("op_410")]; + tensor var_411 = const()[name = tensor("op_411"), val = tensor([1, 1, 1, 1])]; + tensor var_412 = reshape(shape = var_411, x = var_410)[name = tensor("op_412")]; + tensor cross_3 = mul(x = var_409, y = var_412)[name = tensor("cross_3")]; tensor out_7 = add(x = inner_3, y = cross_3)[name = tensor("out_7")]; - tensor var_389 = mul(x = prev_kv_3, y = sqrt_s0_3)[name = tensor("op_389")]; - tensor var_391_transpose_x_1 = const()[name = tensor("op_391_transpose_x_1"), val = tensor(true)]; - tensor var_391_transpose_y_1 = const()[name = tensor("op_391_transpose_y_1"), val = tensor(false)]; - tensor var_391 = matmul(transpose_x = var_391_transpose_x_1, transpose_y = var_391_transpose_y_1, x = k_3, y = v_3)[name = tensor("op_391")]; - tensor new_kv_unnorm_3 = add(x = var_389, y = var_391)[name = tensor("new_kv_unnorm_3")]; - tensor var_393 = const()[name = tensor("op_393"), val = tensor(0x1p+0)]; - tensor new_scale_3 = add(x = prev_scale_3, y = var_393)[name = tensor("new_scale_3")]; - tensor var_395 = sqrt(x = new_scale_3)[name = tensor("op_395")]; - tensor var_396 = real_div(x = new_kv_unnorm_3, y = var_395)[name = tensor("op_396")]; - tensor var_397_perm_0 = const()[name = tensor("op_397_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_415 = mul(x = prev_kv_3, y = sqrt_s0_3)[name = tensor("op_415")]; + tensor var_417_transpose_x_1 = const()[name = tensor("op_417_transpose_x_1"), val = tensor(true)]; + tensor var_417_transpose_y_1 = const()[name = tensor("op_417_transpose_y_1"), val = tensor(false)]; + tensor var_417 = matmul(transpose_x = var_417_transpose_x_1, transpose_y = var_417_transpose_y_1, x = k_3, y = v_3)[name = tensor("op_417")]; + tensor new_kv_unnorm_3 = add(x = var_415, y = var_417)[name = tensor("new_kv_unnorm_3")]; + tensor var_419 = const()[name = tensor("op_419"), val = tensor(0x1p+0)]; + tensor new_scale_3 = add(x = prev_scale_3, y = var_419)[name = tensor("new_scale_3")]; + tensor var_421 = sqrt(x = new_scale_3)[name = tensor("op_421")]; + tensor var_422 = real_div(x = new_kv_unnorm_3, y = var_421)[name = tensor("op_422")]; + tensor var_423_perm_0 = const()[name = tensor("op_423_perm_0"), val = tensor([0, 2, 1, 3])]; tensor out_9_axes_0 = const()[name = tensor("out_9_axes_0"), val = tensor([-1])]; - tensor var_397 = transpose(perm = var_397_perm_0, x = out_7)[name = tensor("transpose_48")]; - tensor out_9 = layer_norm(axes = out_9_axes_0, epsilon = var_17, x = var_397)[name = tensor("out_9")]; - tensor var_401 = const()[name = tensor("op_401"), val = tensor([1, 1, 256])]; - tensor out_11 = reshape(shape = var_401, x = out_9)[name = tensor("out_11")]; - tensor var_403 = silu(x = input_57)[name = tensor("op_403")]; - tensor input_59 = mul(x = var_403, y = out_11)[name = tensor("input_59")]; - tensor ret_out_3 = linear(bias = encoder_out_proj_1_bias, weight = encoder_out_proj_1_weight, x = input_59)[name = tensor("linear_16")]; - tensor x_9 = add(x = input_53, y = ret_out_3)[name = tensor("x_9")]; + tensor var_423 = transpose(perm = var_423_perm_0, x = out_7)[name = tensor("transpose_48")]; + tensor out_9 = layer_norm(axes = out_9_axes_0, epsilon = var_44, x = var_423)[name = tensor("out_9")]; + tensor var_427 = const()[name = tensor("op_427"), val = tensor([1, 1, 256])]; + tensor out_11 = reshape(shape = var_427, x = out_9)[name = tensor("out_11")]; + tensor var_429 = silu(x = input_59)[name = tensor("op_429")]; + tensor input_61 = mul(x = var_429, y = out_11)[name = tensor("input_61")]; + tensor ret_out_3 = linear(bias = inner_encoder_out_proj_1_bias, weight = inner_encoder_out_proj_1_weight, x = input_61)[name = tensor("linear_16")]; + tensor x_9 = add(x = input_55, y = ret_out_3)[name = tensor("x_9")]; tensor window_5_begin_0 = const()[name = tensor("window_5_begin_0"), val = tensor([1, 0, 0, 0])]; tensor window_5_end_0 = const()[name = tensor("window_5_end_0"), val = tensor([2, 1, 16, 256])]; tensor window_5_end_mask_0 = const()[name = tensor("window_5_end_mask_0"), val = tensor([false, true, true, true])]; tensor window_5_squeeze_mask_0 = const()[name = tensor("window_5_squeeze_mask_0"), val = tensor([true, false, false, false])]; tensor window_5 = slice_by_index(begin = window_5_begin_0, end = window_5_end_0, end_mask = window_5_end_mask_0, squeeze_mask = window_5_squeeze_mask_0, x = enc_conv_cache)[name = tensor("window_5")]; - tensor var_414_begin_0 = const()[name = tensor("op_414_begin_0"), val = tensor([0, 1, 0])]; - tensor var_414_end_0 = const()[name = tensor("op_414_end_0"), val = tensor([1, 16, 256])]; - tensor var_414_end_mask_0 = const()[name = tensor("op_414_end_mask_0"), val = tensor([true, true, true])]; - tensor var_414 = slice_by_index(begin = var_414_begin_0, end = var_414_end_0, end_mask = var_414_end_mask_0, x = window_5)[name = tensor("op_414")]; + tensor var_440_begin_0 = const()[name = tensor("op_440_begin_0"), val = tensor([0, 1, 0])]; + tensor var_440_end_0 = const()[name = tensor("op_440_end_0"), val = tensor([1, 16, 256])]; + tensor var_440_end_mask_0 = const()[name = tensor("op_440_end_mask_0"), val = tensor([true, true, true])]; + tensor var_440 = slice_by_index(begin = var_440_begin_0, end = var_440_end_0, end_mask = var_440_end_mask_0, x = window_5)[name = tensor("op_440")]; tensor window_7_interleave_0 = const()[name = tensor("window_7_interleave_0"), val = tensor(false)]; - tensor window_7 = concat(axis = var_26, interleave = window_7_interleave_0, values = (var_414, x_9))[name = tensor("window_7")]; - tensor input_61_interleave_0 = const()[name = tensor("input_61_interleave_0"), val = tensor(false)]; - tensor input_61 = concat(axis = var_23, interleave = input_61_interleave_0, values = window_7)[name = tensor("input_61")]; + tensor window_7 = concat(axis = var_52, interleave = window_7_interleave_0, values = (var_440, x_9))[name = tensor("window_7")]; + tensor input_63_interleave_0 = const()[name = tensor("input_63_interleave_0"), val = tensor(false)]; + tensor input_63 = concat(axis = var_39, interleave = input_63_interleave_0, values = window_7)[name = tensor("input_63")]; tensor x_11_axes_0 = const()[name = tensor("x_11_axes_0"), val = tensor([-1])]; - tensor x_11 = layer_norm(axes = x_11_axes_0, beta = encoder_conv_module_1_sequential_0_bias, epsilon = var_28, gamma = encoder_conv_module_1_sequential_0_weight, x = input_61)[name = tensor("x_11")]; - tensor input_63_perm_0 = const()[name = tensor("input_63_perm_0"), val = tensor([0, 2, 1])]; + tensor x_11 = layer_norm(axes = x_11_axes_0, beta = inner_encoder_conv_module_1_sequential_0_bias, epsilon = var_36, gamma = inner_encoder_conv_module_1_sequential_0_weight, x = input_63)[name = tensor("x_11")]; + tensor input_65_perm_0 = const()[name = tensor("input_65_perm_0"), val = tensor([0, 2, 1])]; tensor inputs_13_pad_type_0 = const()[name = tensor("inputs_13_pad_type_0"), val = tensor("valid")]; tensor inputs_13_strides_0 = const()[name = tensor("inputs_13_strides_0"), val = tensor([1])]; tensor inputs_13_pad_0 = const()[name = tensor("inputs_13_pad_0"), val = tensor([0, 0])]; tensor inputs_13_dilations_0 = const()[name = tensor("inputs_13_dilations_0"), val = tensor([1])]; tensor inputs_13_groups_0 = const()[name = tensor("inputs_13_groups_0"), val = tensor(1)]; - tensor input_63 = transpose(perm = input_63_perm_0, x = x_11)[name = tensor("transpose_47")]; - tensor inputs_13 = conv(bias = encoder_conv_module_1_sequential_2_conv_bias, dilations = inputs_13_dilations_0, groups = inputs_13_groups_0, pad = inputs_13_pad_0, pad_type = inputs_13_pad_type_0, strides = inputs_13_strides_0, weight = encoder_conv_module_1_sequential_2_conv_weight, x = input_63)[name = tensor("inputs_13")]; - tensor var_439_split_sizes_0 = const()[name = tensor("op_439_split_sizes_0"), val = tensor([256, 256])]; - tensor var_439_axis_0 = const()[name = tensor("op_439_axis_0"), val = tensor(1)]; - tensor var_439_0, tensor var_439_1 = split(axis = var_439_axis_0, split_sizes = var_439_split_sizes_0, x = inputs_13)[name = tensor("op_439")]; - tensor var_441 = sigmoid(x = var_439_1)[name = tensor("op_441")]; - tensor inputs_15 = mul(x = var_439_0, y = var_441)[name = tensor("inputs_15")]; + tensor input_65 = transpose(perm = input_65_perm_0, x = x_11)[name = tensor("transpose_47")]; + tensor inputs_13 = conv(bias = inner_encoder_conv_module_1_sequential_2_conv_bias, dilations = inputs_13_dilations_0, groups = inputs_13_groups_0, pad = inputs_13_pad_0, pad_type = inputs_13_pad_type_0, strides = inputs_13_strides_0, weight = inner_encoder_conv_module_1_sequential_2_conv_weight, x = input_65)[name = tensor("inputs_13")]; + tensor var_465_split_sizes_0 = const()[name = tensor("op_465_split_sizes_0"), val = tensor([256, 256])]; + tensor var_465_axis_0 = const()[name = tensor("op_465_axis_0"), val = tensor(1)]; + tensor var_465_0, tensor var_465_1 = split(axis = var_465_axis_0, split_sizes = var_465_split_sizes_0, x = inputs_13)[name = tensor("op_465")]; + tensor var_467 = sigmoid(x = var_465_1)[name = tensor("op_467")]; + tensor inputs_15 = mul(x = var_465_0, y = var_467)[name = tensor("inputs_15")]; tensor outputs_aug_3_pad_type_0 = const()[name = tensor("outputs_aug_3_pad_type_0"), val = tensor("custom")]; tensor outputs_aug_3_pad_0 = const()[name = tensor("outputs_aug_3_pad_0"), val = tensor([15, 15])]; tensor outputs_aug_3_groups_0 = const()[name = tensor("outputs_aug_3_groups_0"), val = tensor(256)]; tensor outputs_aug_3_strides_0 = const()[name = tensor("outputs_aug_3_strides_0"), val = tensor([1])]; tensor outputs_aug_3_dilations_0 = const()[name = tensor("outputs_aug_3_dilations_0"), val = tensor([1])]; - tensor outputs_aug_3 = conv(dilations = outputs_aug_3_dilations_0, groups = outputs_aug_3_groups_0, pad = outputs_aug_3_pad_0, pad_type = outputs_aug_3_pad_type_0, strides = outputs_aug_3_strides_0, weight = encoder_conv_module_1_sequential_4_conv_weight, x = inputs_15)[name = tensor("outputs_aug_3")]; - tensor input_65_begin_0 = const()[name = tensor("input_65_begin_0"), val = tensor([0, 0, 0])]; - tensor input_65_end_0 = const()[name = tensor("input_65_end_0"), val = tensor([1, 256, 16])]; - tensor input_65_end_mask_0 = const()[name = tensor("input_65_end_mask_0"), val = tensor([true, true, false])]; - tensor input_65 = slice_by_index(begin = input_65_begin_0, end = input_65_end_0, end_mask = input_65_end_mask_0, x = outputs_aug_3)[name = tensor("input_65")]; - tensor inputs_17 = batch_norm(beta = encoder_conv_module_1_sequential_5_bias, epsilon = var_28, gamma = encoder_conv_module_1_sequential_5_weight, mean = encoder_conv_module_1_sequential_5_running_mean, variance = encoder_conv_module_1_sequential_5_running_var, x = input_65)[name = tensor("inputs_17")]; - tensor input_67 = silu(x = inputs_17)[name = tensor("input_67")]; - tensor input_69_pad_type_0 = const()[name = tensor("input_69_pad_type_0"), val = tensor("valid")]; - tensor input_69_strides_0 = const()[name = tensor("input_69_strides_0"), val = tensor([1])]; - tensor input_69_pad_0 = const()[name = tensor("input_69_pad_0"), val = tensor([0, 0])]; - tensor input_69_dilations_0 = const()[name = tensor("input_69_dilations_0"), val = tensor([1])]; - tensor input_69_groups_0 = const()[name = tensor("input_69_groups_0"), val = tensor(1)]; - tensor input_69 = conv(bias = encoder_conv_module_1_sequential_7_conv_bias, dilations = input_69_dilations_0, groups = input_69_groups_0, pad = input_69_pad_0, pad_type = input_69_pad_type_0, strides = input_69_strides_0, weight = encoder_conv_module_1_sequential_7_conv_weight, x = input_67)[name = tensor("input_69")]; + tensor outputs_aug_3 = conv(dilations = outputs_aug_3_dilations_0, groups = outputs_aug_3_groups_0, pad = outputs_aug_3_pad_0, pad_type = outputs_aug_3_pad_type_0, strides = outputs_aug_3_strides_0, weight = inner_encoder_conv_module_1_sequential_4_conv_weight, x = inputs_15)[name = tensor("outputs_aug_3")]; + tensor input_67_begin_0 = const()[name = tensor("input_67_begin_0"), val = tensor([0, 0, 0])]; + tensor input_67_end_0 = const()[name = tensor("input_67_end_0"), val = tensor([1, 256, 16])]; + tensor input_67_end_mask_0 = const()[name = tensor("input_67_end_mask_0"), val = tensor([true, true, false])]; + tensor input_67 = slice_by_index(begin = input_67_begin_0, end = input_67_end_0, end_mask = input_67_end_mask_0, x = outputs_aug_3)[name = tensor("input_67")]; + tensor inputs_17 = batch_norm(beta = inner_encoder_conv_module_1_sequential_5_bias, epsilon = var_36, gamma = inner_encoder_conv_module_1_sequential_5_weight, mean = inner_encoder_conv_module_1_sequential_5_running_mean, variance = inner_encoder_conv_module_1_sequential_5_running_var, x = input_67)[name = tensor("inputs_17")]; + tensor input_69 = silu(x = inputs_17)[name = tensor("input_69")]; + tensor input_71_pad_type_0 = const()[name = tensor("input_71_pad_type_0"), val = tensor("valid")]; + tensor input_71_strides_0 = const()[name = tensor("input_71_strides_0"), val = tensor([1])]; + tensor input_71_pad_0 = const()[name = tensor("input_71_pad_0"), val = tensor([0, 0])]; + tensor input_71_dilations_0 = const()[name = tensor("input_71_dilations_0"), val = tensor([1])]; + tensor input_71_groups_0 = const()[name = tensor("input_71_groups_0"), val = tensor(1)]; + tensor input_71 = conv(bias = inner_encoder_conv_module_1_sequential_7_conv_bias, dilations = input_71_dilations_0, groups = input_71_groups_0, pad = input_71_pad_0, pad_type = input_71_pad_type_0, strides = input_71_strides_0, weight = inner_encoder_conv_module_1_sequential_7_conv_weight, x = input_69)[name = tensor("input_71")]; tensor conv_out_3_perm_0 = const()[name = tensor("conv_out_3_perm_0"), val = tensor([0, 2, 1])]; - tensor var_472_begin_0 = const()[name = tensor("op_472_begin_0"), val = tensor([0, -1, 0])]; - tensor var_472_end_0 = const()[name = tensor("op_472_end_0"), val = tensor([1, 16, 256])]; - tensor var_472_end_mask_0 = const()[name = tensor("op_472_end_mask_0"), val = tensor([true, true, true])]; - tensor conv_out_3 = transpose(perm = conv_out_3_perm_0, x = input_69)[name = tensor("transpose_46")]; - tensor var_472 = slice_by_index(begin = var_472_begin_0, end = var_472_end_0, end_mask = var_472_end_mask_0, x = conv_out_3)[name = tensor("op_472")]; - tensor var_474_perm_0 = const()[name = tensor("op_474_perm_0"), val = tensor([1, 0, 2])]; - tensor var_474 = transpose(perm = var_474_perm_0, x = var_472)[name = tensor("transpose_45")]; - tensor input_71 = add(x = x_9, y = var_474)[name = tensor("input_71")]; - tensor input_73_axes_0 = const()[name = tensor("input_73_axes_0"), val = tensor([-1])]; - tensor input_73 = layer_norm(axes = input_73_axes_0, beta = encoder_ffn2_1_module_sequential_0_bias, epsilon = var_28, gamma = encoder_ffn2_1_module_sequential_0_weight, x = input_71)[name = tensor("input_73")]; - tensor inputs_19 = linear(bias = encoder_ffn2_1_module_sequential_1_linear_bias, weight = encoder_ffn2_1_module_sequential_1_linear_weight, x = input_73)[name = tensor("linear_17")]; - tensor input_75 = silu(x = inputs_19)[name = tensor("input_75")]; - tensor input_79 = linear(bias = encoder_ffn2_1_module_sequential_4_linear_bias, weight = encoder_ffn2_1_module_sequential_4_linear_weight, x = input_75)[name = tensor("linear_18")]; - tensor var_497 = const()[name = tensor("op_497"), val = tensor(0x1p-1)]; - tensor var_498 = mul(x = input_79, y = var_497)[name = tensor("op_498")]; - tensor input_81 = add(x = var_498, y = input_71)[name = tensor("input_81")]; - tensor input_83_axes_0 = const()[name = tensor("input_83_axes_0"), val = tensor([-1])]; - tensor input_83 = layer_norm(axes = input_83_axes_0, beta = encoder_layer_norm_1_bias, epsilon = var_28, gamma = encoder_layer_norm_1_weight, x = input_81)[name = tensor("input_83")]; + tensor var_498_begin_0 = const()[name = tensor("op_498_begin_0"), val = tensor([0, -1, 0])]; + tensor var_498_end_0 = const()[name = tensor("op_498_end_0"), val = tensor([1, 16, 256])]; + tensor var_498_end_mask_0 = const()[name = tensor("op_498_end_mask_0"), val = tensor([true, true, true])]; + tensor conv_out_3 = transpose(perm = conv_out_3_perm_0, x = input_71)[name = tensor("transpose_46")]; + tensor var_498 = slice_by_index(begin = var_498_begin_0, end = var_498_end_0, end_mask = var_498_end_mask_0, x = conv_out_3)[name = tensor("op_498")]; + tensor var_500_perm_0 = const()[name = tensor("op_500_perm_0"), val = tensor([1, 0, 2])]; + tensor var_500 = transpose(perm = var_500_perm_0, x = var_498)[name = tensor("transpose_45")]; + tensor input_73 = add(x = x_9, y = var_500)[name = tensor("input_73")]; + tensor input_75_axes_0 = const()[name = tensor("input_75_axes_0"), val = tensor([-1])]; + tensor input_75 = layer_norm(axes = input_75_axes_0, beta = inner_encoder_ffn2_1_module_sequential_0_bias, epsilon = var_36, gamma = inner_encoder_ffn2_1_module_sequential_0_weight, x = input_73)[name = tensor("input_75")]; + tensor inputs_19 = linear(bias = inner_encoder_ffn2_1_module_sequential_1_linear_bias, weight = inner_encoder_ffn2_1_module_sequential_1_linear_weight, x = input_75)[name = tensor("linear_17")]; + tensor input_77 = silu(x = inputs_19)[name = tensor("input_77")]; + tensor input_81 = linear(bias = inner_encoder_ffn2_1_module_sequential_4_linear_bias, weight = inner_encoder_ffn2_1_module_sequential_4_linear_weight, x = input_77)[name = tensor("linear_18")]; + tensor var_523 = const()[name = tensor("op_523"), val = tensor(0x1p-1)]; + tensor var_524 = mul(x = input_81, y = var_523)[name = tensor("op_524")]; + tensor input_83 = add(x = var_524, y = input_73)[name = tensor("input_83")]; tensor input_85_axes_0 = const()[name = tensor("input_85_axes_0"), val = tensor([-1])]; - tensor input_85 = layer_norm(axes = input_85_axes_0, beta = encoder_ffn1_2_module_sequential_0_bias, epsilon = var_28, gamma = encoder_ffn1_2_module_sequential_0_weight, x = input_83)[name = tensor("input_85")]; - tensor inputs_21 = linear(bias = encoder_ffn1_2_module_sequential_1_linear_bias, weight = encoder_ffn1_2_module_sequential_1_linear_weight, x = input_85)[name = tensor("linear_19")]; - tensor input_87 = silu(x = inputs_21)[name = tensor("input_87")]; - tensor input_91 = linear(bias = encoder_ffn1_2_module_sequential_4_linear_bias, weight = encoder_ffn1_2_module_sequential_4_linear_weight, x = input_87)[name = tensor("linear_20")]; - tensor var_527 = const()[name = tensor("op_527"), val = tensor(0x1p-1)]; - tensor var_528 = mul(x = input_91, y = var_527)[name = tensor("op_528")]; - tensor input_93 = add(x = var_528, y = input_83)[name = tensor("input_93")]; + tensor input_85 = layer_norm(axes = input_85_axes_0, beta = inner_encoder_layer_norm_1_bias, epsilon = var_36, gamma = inner_encoder_layer_norm_1_weight, x = input_83)[name = tensor("input_85")]; + tensor input_87_axes_0 = const()[name = tensor("input_87_axes_0"), val = tensor([-1])]; + tensor input_87 = layer_norm(axes = input_87_axes_0, beta = inner_encoder_ffn1_2_module_sequential_0_bias, epsilon = var_36, gamma = inner_encoder_ffn1_2_module_sequential_0_weight, x = input_85)[name = tensor("input_87")]; + tensor inputs_21 = linear(bias = inner_encoder_ffn1_2_module_sequential_1_linear_bias, weight = inner_encoder_ffn1_2_module_sequential_1_linear_weight, x = input_87)[name = tensor("linear_19")]; + tensor input_89 = silu(x = inputs_21)[name = tensor("input_89")]; + tensor input_93 = linear(bias = inner_encoder_ffn1_2_module_sequential_4_linear_bias, weight = inner_encoder_ffn1_2_module_sequential_4_linear_weight, x = input_89)[name = tensor("linear_20")]; + tensor var_553 = const()[name = tensor("op_553"), val = tensor(0x1p-1)]; + tensor var_554 = mul(x = input_93, y = var_553)[name = tensor("op_554")]; + tensor input_95 = add(x = var_554, y = input_85)[name = tensor("input_95")]; tensor x_13_axes_0 = const()[name = tensor("x_13_axes_0"), val = tensor([-1])]; - tensor x_13 = layer_norm(axes = x_13_axes_0, beta = encoder_ret_lns_2_bias, epsilon = var_28, gamma = encoder_ret_lns_2_weight, x = input_93)[name = tensor("x_13")]; + tensor x_13 = layer_norm(axes = x_13_axes_0, beta = inner_encoder_ret_lns_2_bias, epsilon = var_36, gamma = inner_encoder_ret_lns_2_weight, x = input_95)[name = tensor("x_13")]; tensor prev_kv_5_begin_0 = const()[name = tensor("prev_kv_5_begin_0"), val = tensor([2, 0, 0, 0, 0])]; tensor prev_kv_5_end_0 = const()[name = tensor("prev_kv_5_end_0"), val = tensor([3, 1, 4, 64, 64])]; tensor prev_kv_5_end_mask_0 = const()[name = tensor("prev_kv_5_end_mask_0"), val = tensor([false, true, true, true, true])]; @@ -524,139 +530,139 @@ program(1.0) tensor prev_scale_5_end_mask_0 = const()[name = tensor("prev_scale_5_end_mask_0"), val = tensor([false, true])]; tensor prev_scale_5_squeeze_mask_0 = const()[name = tensor("prev_scale_5_squeeze_mask_0"), val = tensor([true, false])]; tensor prev_scale_5 = slice_by_index(begin = prev_scale_5_begin_0, end = prev_scale_5_end_0, end_mask = prev_scale_5_end_mask_0, squeeze_mask = prev_scale_5_squeeze_mask_0, x = enc_scale)[name = tensor("prev_scale_5")]; - tensor var_542 = linear(bias = encoder_q_proj_2_bias, weight = encoder_q_proj_2_weight, x = x_13)[name = tensor("linear_21")]; - tensor var_543 = const()[name = tensor("op_543"), val = tensor([1, 1, 4, 64])]; - tensor var_544 = reshape(shape = var_543, x = var_542)[name = tensor("op_544")]; + tensor var_568 = linear(bias = inner_encoder_q_proj_2_bias, weight = inner_encoder_q_proj_2_weight, x = x_13)[name = tensor("linear_21")]; + tensor var_569 = const()[name = tensor("op_569"), val = tensor([1, 1, 4, 64])]; + tensor var_570 = reshape(shape = var_569, x = var_568)[name = tensor("op_570")]; tensor q_5_perm_0 = const()[name = tensor("q_5_perm_0"), val = tensor([0, 2, 1, 3])]; - tensor var_548 = linear(bias = encoder_k_proj_2_bias, weight = encoder_k_proj_2_weight, x = x_13)[name = tensor("linear_22")]; - tensor var_549 = const()[name = tensor("op_549"), val = tensor(0x1p-3)]; - tensor var_550 = mul(x = var_548, y = var_549)[name = tensor("op_550")]; - tensor var_551 = const()[name = tensor("op_551"), val = tensor([1, 1, 4, 64])]; - tensor var_552 = reshape(shape = var_551, x = var_550)[name = tensor("op_552")]; + tensor var_574 = linear(bias = inner_encoder_k_proj_2_bias, weight = inner_encoder_k_proj_2_weight, x = x_13)[name = tensor("linear_22")]; + tensor var_575 = const()[name = tensor("op_575"), val = tensor(0x1p-3)]; + tensor var_576 = mul(x = var_574, y = var_575)[name = tensor("op_576")]; + tensor var_577 = const()[name = tensor("op_577"), val = tensor([1, 1, 4, 64])]; + tensor var_578 = reshape(shape = var_577, x = var_576)[name = tensor("op_578")]; tensor k_5_perm_0 = const()[name = tensor("k_5_perm_0"), val = tensor([0, 2, 1, 3])]; - tensor var_556 = linear(bias = encoder_v_proj_2_bias, weight = encoder_v_proj_2_weight, x = x_13)[name = tensor("linear_23")]; - tensor var_557 = const()[name = tensor("op_557"), val = tensor([1, 1, 4, 64])]; - tensor var_558 = reshape(shape = var_557, x = var_556)[name = tensor("op_558")]; + tensor var_582 = linear(bias = inner_encoder_v_proj_2_bias, weight = inner_encoder_v_proj_2_weight, x = x_13)[name = tensor("linear_23")]; + tensor var_583 = const()[name = tensor("op_583"), val = tensor([1, 1, 4, 64])]; + tensor var_584 = reshape(shape = var_583, x = var_582)[name = tensor("op_584")]; tensor v_5_perm_0 = const()[name = tensor("v_5_perm_0"), val = tensor([0, 2, 1, 3])]; - tensor input_97 = linear(bias = encoder_g_proj_2_bias, weight = encoder_g_proj_2_weight, x = x_13)[name = tensor("linear_24")]; + tensor input_99 = linear(bias = inner_encoder_g_proj_2_bias, weight = inner_encoder_g_proj_2_weight, x = x_13)[name = tensor("linear_24")]; tensor sqrt_s0_5 = sqrt(x = prev_scale_5)[name = tensor("sqrt_s0_5")]; - tensor s_t_5 = add(x = prev_scale_5, y = encoder__t_index)[name = tensor("s_t_5")]; + tensor s_t_5 = add(x = prev_scale_5, y = inner_encoder__t_index)[name = tensor("s_t_5")]; tensor sqrt_s_t_5 = sqrt(x = s_t_5)[name = tensor("sqrt_s_t_5")]; tensor qk_5_transpose_x_1 = const()[name = tensor("qk_5_transpose_x_1"), val = tensor(false)]; tensor qk_5_transpose_y_1 = const()[name = tensor("qk_5_transpose_y_1"), val = tensor(true)]; - tensor k_5 = transpose(perm = k_5_perm_0, x = var_552)[name = tensor("transpose_43")]; - tensor q_5 = transpose(perm = q_5_perm_0, x = var_544)[name = tensor("transpose_44")]; + tensor k_5 = transpose(perm = k_5_perm_0, x = var_578)[name = tensor("transpose_43")]; + tensor q_5 = transpose(perm = q_5_perm_0, x = var_570)[name = tensor("transpose_44")]; tensor qk_5 = matmul(transpose_x = qk_5_transpose_x_1, transpose_y = qk_5_transpose_y_1, x = q_5, y = k_5)[name = tensor("qk_5")]; - tensor var_568 = const()[name = tensor("op_568"), val = tensor([1, 1])]; - tensor var_569 = reshape(shape = var_568, x = sqrt_s_t_5)[name = tensor("op_569")]; - tensor M_5 = real_div(x = encoder__causal_mask, y = var_569)[name = tensor("M_5")]; - tensor var_571 = mul(x = qk_5, y = M_5)[name = tensor("op_571")]; + tensor var_594 = const()[name = tensor("op_594"), val = tensor([1, 1])]; + tensor var_595 = reshape(shape = var_594, x = sqrt_s_t_5)[name = tensor("op_595")]; + tensor M_5 = real_div(x = inner_encoder__causal_mask, y = var_595)[name = tensor("M_5")]; + tensor var_597 = mul(x = qk_5, y = M_5)[name = tensor("op_597")]; tensor inner_5_transpose_x_0 = const()[name = tensor("inner_5_transpose_x_0"), val = tensor(false)]; tensor inner_5_transpose_y_0 = const()[name = tensor("inner_5_transpose_y_0"), val = tensor(false)]; - tensor v_5 = transpose(perm = v_5_perm_0, x = var_558)[name = tensor("transpose_42")]; - tensor inner_5 = matmul(transpose_x = inner_5_transpose_x_0, transpose_y = inner_5_transpose_y_0, x = var_571, y = v_5)[name = tensor("inner_5")]; - tensor var_573_transpose_x_0 = const()[name = tensor("op_573_transpose_x_0"), val = tensor(false)]; - tensor var_573_transpose_y_0 = const()[name = tensor("op_573_transpose_y_0"), val = tensor(false)]; - tensor var_573 = matmul(transpose_x = var_573_transpose_x_0, transpose_y = var_573_transpose_y_0, x = q_5, y = prev_kv_5)[name = tensor("op_573")]; - tensor var_574 = real_div(x = sqrt_s0_5, y = sqrt_s_t_5)[name = tensor("op_574")]; - tensor var_575 = const()[name = tensor("op_575"), val = tensor([1, 1, 1, 1])]; - tensor var_576 = reshape(shape = var_575, x = var_574)[name = tensor("op_576")]; - tensor cross_5 = mul(x = var_573, y = var_576)[name = tensor("cross_5")]; + tensor v_5 = transpose(perm = v_5_perm_0, x = var_584)[name = tensor("transpose_42")]; + tensor inner_5 = matmul(transpose_x = inner_5_transpose_x_0, transpose_y = inner_5_transpose_y_0, x = var_597, y = v_5)[name = tensor("inner_5")]; + tensor var_599_transpose_x_0 = const()[name = tensor("op_599_transpose_x_0"), val = tensor(false)]; + tensor var_599_transpose_y_0 = const()[name = tensor("op_599_transpose_y_0"), val = tensor(false)]; + tensor var_599 = matmul(transpose_x = var_599_transpose_x_0, transpose_y = var_599_transpose_y_0, x = q_5, y = prev_kv_5)[name = tensor("op_599")]; + tensor var_600 = real_div(x = sqrt_s0_5, y = sqrt_s_t_5)[name = tensor("op_600")]; + tensor var_601 = const()[name = tensor("op_601"), val = tensor([1, 1, 1, 1])]; + tensor var_602 = reshape(shape = var_601, x = var_600)[name = tensor("op_602")]; + tensor cross_5 = mul(x = var_599, y = var_602)[name = tensor("cross_5")]; tensor out_13 = add(x = inner_5, y = cross_5)[name = tensor("out_13")]; - tensor var_579 = mul(x = prev_kv_5, y = sqrt_s0_5)[name = tensor("op_579")]; - tensor var_581_transpose_x_1 = const()[name = tensor("op_581_transpose_x_1"), val = tensor(true)]; - tensor var_581_transpose_y_1 = const()[name = tensor("op_581_transpose_y_1"), val = tensor(false)]; - tensor var_581 = matmul(transpose_x = var_581_transpose_x_1, transpose_y = var_581_transpose_y_1, x = k_5, y = v_5)[name = tensor("op_581")]; - tensor new_kv_unnorm_5 = add(x = var_579, y = var_581)[name = tensor("new_kv_unnorm_5")]; - tensor var_583 = const()[name = tensor("op_583"), val = tensor(0x1p+0)]; - tensor new_scale_5 = add(x = prev_scale_5, y = var_583)[name = tensor("new_scale_5")]; - tensor var_585 = sqrt(x = new_scale_5)[name = tensor("op_585")]; - tensor var_586 = real_div(x = new_kv_unnorm_5, y = var_585)[name = tensor("op_586")]; - tensor var_587_perm_0 = const()[name = tensor("op_587_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_605 = mul(x = prev_kv_5, y = sqrt_s0_5)[name = tensor("op_605")]; + tensor var_607_transpose_x_1 = const()[name = tensor("op_607_transpose_x_1"), val = tensor(true)]; + tensor var_607_transpose_y_1 = const()[name = tensor("op_607_transpose_y_1"), val = tensor(false)]; + tensor var_607 = matmul(transpose_x = var_607_transpose_x_1, transpose_y = var_607_transpose_y_1, x = k_5, y = v_5)[name = tensor("op_607")]; + tensor new_kv_unnorm_5 = add(x = var_605, y = var_607)[name = tensor("new_kv_unnorm_5")]; + tensor var_609 = const()[name = tensor("op_609"), val = tensor(0x1p+0)]; + tensor new_scale_5 = add(x = prev_scale_5, y = var_609)[name = tensor("new_scale_5")]; + tensor var_611 = sqrt(x = new_scale_5)[name = tensor("op_611")]; + tensor var_612 = real_div(x = new_kv_unnorm_5, y = var_611)[name = tensor("op_612")]; + tensor var_613_perm_0 = const()[name = tensor("op_613_perm_0"), val = tensor([0, 2, 1, 3])]; tensor out_15_axes_0 = const()[name = tensor("out_15_axes_0"), val = tensor([-1])]; - tensor var_587 = transpose(perm = var_587_perm_0, x = out_13)[name = tensor("transpose_41")]; - tensor out_15 = layer_norm(axes = out_15_axes_0, epsilon = var_17, x = var_587)[name = tensor("out_15")]; - tensor var_591 = const()[name = tensor("op_591"), val = tensor([1, 1, 256])]; - tensor out_17 = reshape(shape = var_591, x = out_15)[name = tensor("out_17")]; - tensor var_593 = silu(x = input_97)[name = tensor("op_593")]; - tensor input_99 = mul(x = var_593, y = out_17)[name = tensor("input_99")]; - tensor ret_out_5 = linear(bias = encoder_out_proj_2_bias, weight = encoder_out_proj_2_weight, x = input_99)[name = tensor("linear_25")]; - tensor x_15 = add(x = input_93, y = ret_out_5)[name = tensor("x_15")]; + tensor var_613 = transpose(perm = var_613_perm_0, x = out_13)[name = tensor("transpose_41")]; + tensor out_15 = layer_norm(axes = out_15_axes_0, epsilon = var_44, x = var_613)[name = tensor("out_15")]; + tensor var_617 = const()[name = tensor("op_617"), val = tensor([1, 1, 256])]; + tensor out_17 = reshape(shape = var_617, x = out_15)[name = tensor("out_17")]; + tensor var_619 = silu(x = input_99)[name = tensor("op_619")]; + tensor input_101 = mul(x = var_619, y = out_17)[name = tensor("input_101")]; + tensor ret_out_5 = linear(bias = inner_encoder_out_proj_2_bias, weight = inner_encoder_out_proj_2_weight, x = input_101)[name = tensor("linear_25")]; + tensor x_15 = add(x = input_95, y = ret_out_5)[name = tensor("x_15")]; tensor window_9_begin_0 = const()[name = tensor("window_9_begin_0"), val = tensor([2, 0, 0, 0])]; tensor window_9_end_0 = const()[name = tensor("window_9_end_0"), val = tensor([3, 1, 16, 256])]; tensor window_9_end_mask_0 = const()[name = tensor("window_9_end_mask_0"), val = tensor([false, true, true, true])]; tensor window_9_squeeze_mask_0 = const()[name = tensor("window_9_squeeze_mask_0"), val = tensor([true, false, false, false])]; tensor window_9 = slice_by_index(begin = window_9_begin_0, end = window_9_end_0, end_mask = window_9_end_mask_0, squeeze_mask = window_9_squeeze_mask_0, x = enc_conv_cache)[name = tensor("window_9")]; - tensor var_604_begin_0 = const()[name = tensor("op_604_begin_0"), val = tensor([0, 1, 0])]; - tensor var_604_end_0 = const()[name = tensor("op_604_end_0"), val = tensor([1, 16, 256])]; - tensor var_604_end_mask_0 = const()[name = tensor("op_604_end_mask_0"), val = tensor([true, true, true])]; - tensor var_604 = slice_by_index(begin = var_604_begin_0, end = var_604_end_0, end_mask = var_604_end_mask_0, x = window_9)[name = tensor("op_604")]; + tensor var_630_begin_0 = const()[name = tensor("op_630_begin_0"), val = tensor([0, 1, 0])]; + tensor var_630_end_0 = const()[name = tensor("op_630_end_0"), val = tensor([1, 16, 256])]; + tensor var_630_end_mask_0 = const()[name = tensor("op_630_end_mask_0"), val = tensor([true, true, true])]; + tensor var_630 = slice_by_index(begin = var_630_begin_0, end = var_630_end_0, end_mask = var_630_end_mask_0, x = window_9)[name = tensor("op_630")]; tensor window_11_interleave_0 = const()[name = tensor("window_11_interleave_0"), val = tensor(false)]; - tensor window_11 = concat(axis = var_26, interleave = window_11_interleave_0, values = (var_604, x_15))[name = tensor("window_11")]; - tensor input_101_interleave_0 = const()[name = tensor("input_101_interleave_0"), val = tensor(false)]; - tensor input_101 = concat(axis = var_23, interleave = input_101_interleave_0, values = window_11)[name = tensor("input_101")]; + tensor window_11 = concat(axis = var_52, interleave = window_11_interleave_0, values = (var_630, x_15))[name = tensor("window_11")]; + tensor input_103_interleave_0 = const()[name = tensor("input_103_interleave_0"), val = tensor(false)]; + tensor input_103 = concat(axis = var_39, interleave = input_103_interleave_0, values = window_11)[name = tensor("input_103")]; tensor x_17_axes_0 = const()[name = tensor("x_17_axes_0"), val = tensor([-1])]; - tensor x_17 = layer_norm(axes = x_17_axes_0, beta = encoder_conv_module_2_sequential_0_bias, epsilon = var_28, gamma = encoder_conv_module_2_sequential_0_weight, x = input_101)[name = tensor("x_17")]; - tensor input_103_perm_0 = const()[name = tensor("input_103_perm_0"), val = tensor([0, 2, 1])]; + tensor x_17 = layer_norm(axes = x_17_axes_0, beta = inner_encoder_conv_module_2_sequential_0_bias, epsilon = var_36, gamma = inner_encoder_conv_module_2_sequential_0_weight, x = input_103)[name = tensor("x_17")]; + tensor input_105_perm_0 = const()[name = tensor("input_105_perm_0"), val = tensor([0, 2, 1])]; tensor inputs_23_pad_type_0 = const()[name = tensor("inputs_23_pad_type_0"), val = tensor("valid")]; tensor inputs_23_strides_0 = const()[name = tensor("inputs_23_strides_0"), val = tensor([1])]; tensor inputs_23_pad_0 = const()[name = tensor("inputs_23_pad_0"), val = tensor([0, 0])]; tensor inputs_23_dilations_0 = const()[name = tensor("inputs_23_dilations_0"), val = tensor([1])]; tensor inputs_23_groups_0 = const()[name = tensor("inputs_23_groups_0"), val = tensor(1)]; - tensor input_103 = transpose(perm = input_103_perm_0, x = x_17)[name = tensor("transpose_40")]; - tensor inputs_23 = conv(bias = encoder_conv_module_2_sequential_2_conv_bias, dilations = inputs_23_dilations_0, groups = inputs_23_groups_0, pad = inputs_23_pad_0, pad_type = inputs_23_pad_type_0, strides = inputs_23_strides_0, weight = encoder_conv_module_2_sequential_2_conv_weight, x = input_103)[name = tensor("inputs_23")]; - tensor var_629_split_sizes_0 = const()[name = tensor("op_629_split_sizes_0"), val = tensor([256, 256])]; - tensor var_629_axis_0 = const()[name = tensor("op_629_axis_0"), val = tensor(1)]; - tensor var_629_0, tensor var_629_1 = split(axis = var_629_axis_0, split_sizes = var_629_split_sizes_0, x = inputs_23)[name = tensor("op_629")]; - tensor var_631 = sigmoid(x = var_629_1)[name = tensor("op_631")]; - tensor inputs_25 = mul(x = var_629_0, y = var_631)[name = tensor("inputs_25")]; + tensor input_105 = transpose(perm = input_105_perm_0, x = x_17)[name = tensor("transpose_40")]; + tensor inputs_23 = conv(bias = inner_encoder_conv_module_2_sequential_2_conv_bias, dilations = inputs_23_dilations_0, groups = inputs_23_groups_0, pad = inputs_23_pad_0, pad_type = inputs_23_pad_type_0, strides = inputs_23_strides_0, weight = inner_encoder_conv_module_2_sequential_2_conv_weight, x = input_105)[name = tensor("inputs_23")]; + tensor var_655_split_sizes_0 = const()[name = tensor("op_655_split_sizes_0"), val = tensor([256, 256])]; + tensor var_655_axis_0 = const()[name = tensor("op_655_axis_0"), val = tensor(1)]; + tensor var_655_0, tensor var_655_1 = split(axis = var_655_axis_0, split_sizes = var_655_split_sizes_0, x = inputs_23)[name = tensor("op_655")]; + tensor var_657 = sigmoid(x = var_655_1)[name = tensor("op_657")]; + tensor inputs_25 = mul(x = var_655_0, y = var_657)[name = tensor("inputs_25")]; tensor outputs_aug_5_pad_type_0 = const()[name = tensor("outputs_aug_5_pad_type_0"), val = tensor("custom")]; tensor outputs_aug_5_pad_0 = const()[name = tensor("outputs_aug_5_pad_0"), val = tensor([15, 15])]; tensor outputs_aug_5_groups_0 = const()[name = tensor("outputs_aug_5_groups_0"), val = tensor(256)]; tensor outputs_aug_5_strides_0 = const()[name = tensor("outputs_aug_5_strides_0"), val = tensor([1])]; tensor outputs_aug_5_dilations_0 = const()[name = tensor("outputs_aug_5_dilations_0"), val = tensor([1])]; - tensor outputs_aug_5 = conv(dilations = outputs_aug_5_dilations_0, groups = outputs_aug_5_groups_0, pad = outputs_aug_5_pad_0, pad_type = outputs_aug_5_pad_type_0, strides = outputs_aug_5_strides_0, weight = encoder_conv_module_2_sequential_4_conv_weight, x = inputs_25)[name = tensor("outputs_aug_5")]; - tensor input_105_begin_0 = const()[name = tensor("input_105_begin_0"), val = tensor([0, 0, 0])]; - tensor input_105_end_0 = const()[name = tensor("input_105_end_0"), val = tensor([1, 256, 16])]; - tensor input_105_end_mask_0 = const()[name = tensor("input_105_end_mask_0"), val = tensor([true, true, false])]; - tensor input_105 = slice_by_index(begin = input_105_begin_0, end = input_105_end_0, end_mask = input_105_end_mask_0, x = outputs_aug_5)[name = tensor("input_105")]; - tensor inputs_27 = batch_norm(beta = encoder_conv_module_2_sequential_5_bias, epsilon = var_28, gamma = encoder_conv_module_2_sequential_5_weight, mean = encoder_conv_module_2_sequential_5_running_mean, variance = encoder_conv_module_2_sequential_5_running_var, x = input_105)[name = tensor("inputs_27")]; - tensor input_107 = silu(x = inputs_27)[name = tensor("input_107")]; - tensor input_109_pad_type_0 = const()[name = tensor("input_109_pad_type_0"), val = tensor("valid")]; - tensor input_109_strides_0 = const()[name = tensor("input_109_strides_0"), val = tensor([1])]; - tensor input_109_pad_0 = const()[name = tensor("input_109_pad_0"), val = tensor([0, 0])]; - tensor input_109_dilations_0 = const()[name = tensor("input_109_dilations_0"), val = tensor([1])]; - tensor input_109_groups_0 = const()[name = tensor("input_109_groups_0"), val = tensor(1)]; - tensor input_109 = conv(bias = encoder_conv_module_2_sequential_7_conv_bias, dilations = input_109_dilations_0, groups = input_109_groups_0, pad = input_109_pad_0, pad_type = input_109_pad_type_0, strides = input_109_strides_0, weight = encoder_conv_module_2_sequential_7_conv_weight, x = input_107)[name = tensor("input_109")]; + tensor outputs_aug_5 = conv(dilations = outputs_aug_5_dilations_0, groups = outputs_aug_5_groups_0, pad = outputs_aug_5_pad_0, pad_type = outputs_aug_5_pad_type_0, strides = outputs_aug_5_strides_0, weight = inner_encoder_conv_module_2_sequential_4_conv_weight, x = inputs_25)[name = tensor("outputs_aug_5")]; + tensor input_107_begin_0 = const()[name = tensor("input_107_begin_0"), val = tensor([0, 0, 0])]; + tensor input_107_end_0 = const()[name = tensor("input_107_end_0"), val = tensor([1, 256, 16])]; + tensor input_107_end_mask_0 = const()[name = tensor("input_107_end_mask_0"), val = tensor([true, true, false])]; + tensor input_107 = slice_by_index(begin = input_107_begin_0, end = input_107_end_0, end_mask = input_107_end_mask_0, x = outputs_aug_5)[name = tensor("input_107")]; + tensor inputs_27 = batch_norm(beta = inner_encoder_conv_module_2_sequential_5_bias, epsilon = var_36, gamma = inner_encoder_conv_module_2_sequential_5_weight, mean = inner_encoder_conv_module_2_sequential_5_running_mean, variance = inner_encoder_conv_module_2_sequential_5_running_var, x = input_107)[name = tensor("inputs_27")]; + tensor input_109 = silu(x = inputs_27)[name = tensor("input_109")]; + tensor input_111_pad_type_0 = const()[name = tensor("input_111_pad_type_0"), val = tensor("valid")]; + tensor input_111_strides_0 = const()[name = tensor("input_111_strides_0"), val = tensor([1])]; + tensor input_111_pad_0 = const()[name = tensor("input_111_pad_0"), val = tensor([0, 0])]; + tensor input_111_dilations_0 = const()[name = tensor("input_111_dilations_0"), val = tensor([1])]; + tensor input_111_groups_0 = const()[name = tensor("input_111_groups_0"), val = tensor(1)]; + tensor input_111 = conv(bias = inner_encoder_conv_module_2_sequential_7_conv_bias, dilations = input_111_dilations_0, groups = input_111_groups_0, pad = input_111_pad_0, pad_type = input_111_pad_type_0, strides = input_111_strides_0, weight = inner_encoder_conv_module_2_sequential_7_conv_weight, x = input_109)[name = tensor("input_111")]; tensor conv_out_5_perm_0 = const()[name = tensor("conv_out_5_perm_0"), val = tensor([0, 2, 1])]; - tensor var_662_begin_0 = const()[name = tensor("op_662_begin_0"), val = tensor([0, -1, 0])]; - tensor var_662_end_0 = const()[name = tensor("op_662_end_0"), val = tensor([1, 16, 256])]; - tensor var_662_end_mask_0 = const()[name = tensor("op_662_end_mask_0"), val = tensor([true, true, true])]; - tensor conv_out_5 = transpose(perm = conv_out_5_perm_0, x = input_109)[name = tensor("transpose_39")]; - tensor var_662 = slice_by_index(begin = var_662_begin_0, end = var_662_end_0, end_mask = var_662_end_mask_0, x = conv_out_5)[name = tensor("op_662")]; - tensor var_664_perm_0 = const()[name = tensor("op_664_perm_0"), val = tensor([1, 0, 2])]; - tensor var_664 = transpose(perm = var_664_perm_0, x = var_662)[name = tensor("transpose_38")]; - tensor input_111 = add(x = x_15, y = var_664)[name = tensor("input_111")]; - tensor input_113_axes_0 = const()[name = tensor("input_113_axes_0"), val = tensor([-1])]; - tensor input_113 = layer_norm(axes = input_113_axes_0, beta = encoder_ffn2_2_module_sequential_0_bias, epsilon = var_28, gamma = encoder_ffn2_2_module_sequential_0_weight, x = input_111)[name = tensor("input_113")]; - tensor inputs_29 = linear(bias = encoder_ffn2_2_module_sequential_1_linear_bias, weight = encoder_ffn2_2_module_sequential_1_linear_weight, x = input_113)[name = tensor("linear_26")]; - tensor input_115 = silu(x = inputs_29)[name = tensor("input_115")]; - tensor input_119 = linear(bias = encoder_ffn2_2_module_sequential_4_linear_bias, weight = encoder_ffn2_2_module_sequential_4_linear_weight, x = input_115)[name = tensor("linear_27")]; - tensor var_687 = const()[name = tensor("op_687"), val = tensor(0x1p-1)]; - tensor var_688 = mul(x = input_119, y = var_687)[name = tensor("op_688")]; - tensor input_121 = add(x = var_688, y = input_111)[name = tensor("input_121")]; - tensor input_123_axes_0 = const()[name = tensor("input_123_axes_0"), val = tensor([-1])]; - tensor input_123 = layer_norm(axes = input_123_axes_0, beta = encoder_layer_norm_2_bias, epsilon = var_28, gamma = encoder_layer_norm_2_weight, x = input_121)[name = tensor("input_123")]; + tensor var_688_begin_0 = const()[name = tensor("op_688_begin_0"), val = tensor([0, -1, 0])]; + tensor var_688_end_0 = const()[name = tensor("op_688_end_0"), val = tensor([1, 16, 256])]; + tensor var_688_end_mask_0 = const()[name = tensor("op_688_end_mask_0"), val = tensor([true, true, true])]; + tensor conv_out_5 = transpose(perm = conv_out_5_perm_0, x = input_111)[name = tensor("transpose_39")]; + tensor var_688 = slice_by_index(begin = var_688_begin_0, end = var_688_end_0, end_mask = var_688_end_mask_0, x = conv_out_5)[name = tensor("op_688")]; + tensor var_690_perm_0 = const()[name = tensor("op_690_perm_0"), val = tensor([1, 0, 2])]; + tensor var_690 = transpose(perm = var_690_perm_0, x = var_688)[name = tensor("transpose_38")]; + tensor input_113 = add(x = x_15, y = var_690)[name = tensor("input_113")]; + tensor input_115_axes_0 = const()[name = tensor("input_115_axes_0"), val = tensor([-1])]; + tensor input_115 = layer_norm(axes = input_115_axes_0, beta = inner_encoder_ffn2_2_module_sequential_0_bias, epsilon = var_36, gamma = inner_encoder_ffn2_2_module_sequential_0_weight, x = input_113)[name = tensor("input_115")]; + tensor inputs_29 = linear(bias = inner_encoder_ffn2_2_module_sequential_1_linear_bias, weight = inner_encoder_ffn2_2_module_sequential_1_linear_weight, x = input_115)[name = tensor("linear_26")]; + tensor input_117 = silu(x = inputs_29)[name = tensor("input_117")]; + tensor input_121 = linear(bias = inner_encoder_ffn2_2_module_sequential_4_linear_bias, weight = inner_encoder_ffn2_2_module_sequential_4_linear_weight, x = input_117)[name = tensor("linear_27")]; + tensor var_713 = const()[name = tensor("op_713"), val = tensor(0x1p-1)]; + tensor var_714 = mul(x = input_121, y = var_713)[name = tensor("op_714")]; + tensor input_123 = add(x = var_714, y = input_113)[name = tensor("input_123")]; tensor input_125_axes_0 = const()[name = tensor("input_125_axes_0"), val = tensor([-1])]; - tensor input_125 = layer_norm(axes = input_125_axes_0, beta = encoder_ffn1_3_module_sequential_0_bias, epsilon = var_28, gamma = encoder_ffn1_3_module_sequential_0_weight, x = input_123)[name = tensor("input_125")]; - tensor inputs_31 = linear(bias = encoder_ffn1_3_module_sequential_1_linear_bias, weight = encoder_ffn1_3_module_sequential_1_linear_weight, x = input_125)[name = tensor("linear_28")]; - tensor input_127 = silu(x = inputs_31)[name = tensor("input_127")]; - tensor input_131 = linear(bias = encoder_ffn1_3_module_sequential_4_linear_bias, weight = encoder_ffn1_3_module_sequential_4_linear_weight, x = input_127)[name = tensor("linear_29")]; - tensor var_717 = const()[name = tensor("op_717"), val = tensor(0x1p-1)]; - tensor var_718 = mul(x = input_131, y = var_717)[name = tensor("op_718")]; - tensor input_133 = add(x = var_718, y = input_123)[name = tensor("input_133")]; + tensor input_125 = layer_norm(axes = input_125_axes_0, beta = inner_encoder_layer_norm_2_bias, epsilon = var_36, gamma = inner_encoder_layer_norm_2_weight, x = input_123)[name = tensor("input_125")]; + tensor input_127_axes_0 = const()[name = tensor("input_127_axes_0"), val = tensor([-1])]; + tensor input_127 = layer_norm(axes = input_127_axes_0, beta = inner_encoder_ffn1_3_module_sequential_0_bias, epsilon = var_36, gamma = inner_encoder_ffn1_3_module_sequential_0_weight, x = input_125)[name = tensor("input_127")]; + tensor inputs_31 = linear(bias = inner_encoder_ffn1_3_module_sequential_1_linear_bias, weight = inner_encoder_ffn1_3_module_sequential_1_linear_weight, x = input_127)[name = tensor("linear_28")]; + tensor input_129 = silu(x = inputs_31)[name = tensor("input_129")]; + tensor input_133 = linear(bias = inner_encoder_ffn1_3_module_sequential_4_linear_bias, weight = inner_encoder_ffn1_3_module_sequential_4_linear_weight, x = input_129)[name = tensor("linear_29")]; + tensor var_743 = const()[name = tensor("op_743"), val = tensor(0x1p-1)]; + tensor var_744 = mul(x = input_133, y = var_743)[name = tensor("op_744")]; + tensor input_135 = add(x = var_744, y = input_125)[name = tensor("input_135")]; tensor x_19_axes_0 = const()[name = tensor("x_19_axes_0"), val = tensor([-1])]; - tensor x_19 = layer_norm(axes = x_19_axes_0, beta = encoder_ret_lns_3_bias, epsilon = var_28, gamma = encoder_ret_lns_3_weight, x = input_133)[name = tensor("x_19")]; + tensor x_19 = layer_norm(axes = x_19_axes_0, beta = inner_encoder_ret_lns_3_bias, epsilon = var_36, gamma = inner_encoder_ret_lns_3_weight, x = input_135)[name = tensor("x_19")]; tensor prev_kv_7_begin_0 = const()[name = tensor("prev_kv_7_begin_0"), val = tensor([3, 0, 0, 0, 0])]; tensor prev_kv_7_end_0 = const()[name = tensor("prev_kv_7_end_0"), val = tensor([4, 1, 4, 64, 64])]; tensor prev_kv_7_end_mask_0 = const()[name = tensor("prev_kv_7_end_mask_0"), val = tensor([false, true, true, true, true])]; @@ -667,175 +673,168 @@ program(1.0) tensor prev_scale_7_end_mask_0 = const()[name = tensor("prev_scale_7_end_mask_0"), val = tensor([false, true])]; tensor prev_scale_7_squeeze_mask_0 = const()[name = tensor("prev_scale_7_squeeze_mask_0"), val = tensor([true, false])]; tensor prev_scale_7 = slice_by_index(begin = prev_scale_7_begin_0, end = prev_scale_7_end_0, end_mask = prev_scale_7_end_mask_0, squeeze_mask = prev_scale_7_squeeze_mask_0, x = enc_scale)[name = tensor("prev_scale_7")]; - tensor var_732 = linear(bias = encoder_q_proj_3_bias, weight = encoder_q_proj_3_weight, x = x_19)[name = tensor("linear_30")]; - tensor var_733 = const()[name = tensor("op_733"), val = tensor([1, 1, 4, 64])]; - tensor var_734 = reshape(shape = var_733, x = var_732)[name = tensor("op_734")]; + tensor var_758 = linear(bias = inner_encoder_q_proj_3_bias, weight = inner_encoder_q_proj_3_weight, x = x_19)[name = tensor("linear_30")]; + tensor var_759 = const()[name = tensor("op_759"), val = tensor([1, 1, 4, 64])]; + tensor var_760 = reshape(shape = var_759, x = var_758)[name = tensor("op_760")]; tensor q_7_perm_0 = const()[name = tensor("q_7_perm_0"), val = tensor([0, 2, 1, 3])]; - tensor var_738 = linear(bias = encoder_k_proj_3_bias, weight = encoder_k_proj_3_weight, x = x_19)[name = tensor("linear_31")]; - tensor var_739 = const()[name = tensor("op_739"), val = tensor(0x1p-3)]; - tensor var_740 = mul(x = var_738, y = var_739)[name = tensor("op_740")]; - tensor var_741 = const()[name = tensor("op_741"), val = tensor([1, 1, 4, 64])]; - tensor var_742 = reshape(shape = var_741, x = var_740)[name = tensor("op_742")]; + tensor var_764 = linear(bias = inner_encoder_k_proj_3_bias, weight = inner_encoder_k_proj_3_weight, x = x_19)[name = tensor("linear_31")]; + tensor var_765 = const()[name = tensor("op_765"), val = tensor(0x1p-3)]; + tensor var_766 = mul(x = var_764, y = var_765)[name = tensor("op_766")]; + tensor var_767 = const()[name = tensor("op_767"), val = tensor([1, 1, 4, 64])]; + tensor var_768 = reshape(shape = var_767, x = var_766)[name = tensor("op_768")]; tensor k_7_perm_0 = const()[name = tensor("k_7_perm_0"), val = tensor([0, 2, 1, 3])]; - tensor var_746 = linear(bias = encoder_v_proj_3_bias, weight = encoder_v_proj_3_weight, x = x_19)[name = tensor("linear_32")]; - tensor var_747 = const()[name = tensor("op_747"), val = tensor([1, 1, 4, 64])]; - tensor var_748 = reshape(shape = var_747, x = var_746)[name = tensor("op_748")]; + tensor var_772 = linear(bias = inner_encoder_v_proj_3_bias, weight = inner_encoder_v_proj_3_weight, x = x_19)[name = tensor("linear_32")]; + tensor var_773 = const()[name = tensor("op_773"), val = tensor([1, 1, 4, 64])]; + tensor var_774 = reshape(shape = var_773, x = var_772)[name = tensor("op_774")]; tensor v_7_perm_0 = const()[name = tensor("v_7_perm_0"), val = tensor([0, 2, 1, 3])]; - tensor input_137 = linear(bias = encoder_g_proj_3_bias, weight = encoder_g_proj_3_weight, x = x_19)[name = tensor("linear_33")]; + tensor input_139 = linear(bias = inner_encoder_g_proj_3_bias, weight = inner_encoder_g_proj_3_weight, x = x_19)[name = tensor("linear_33")]; tensor sqrt_s0_7 = sqrt(x = prev_scale_7)[name = tensor("sqrt_s0_7")]; - tensor s_t_7 = add(x = prev_scale_7, y = encoder__t_index)[name = tensor("s_t_7")]; + tensor s_t_7 = add(x = prev_scale_7, y = inner_encoder__t_index)[name = tensor("s_t_7")]; tensor sqrt_s_t_7 = sqrt(x = s_t_7)[name = tensor("sqrt_s_t_7")]; tensor qk_7_transpose_x_1 = const()[name = tensor("qk_7_transpose_x_1"), val = tensor(false)]; tensor qk_7_transpose_y_1 = const()[name = tensor("qk_7_transpose_y_1"), val = tensor(true)]; - tensor k_7 = transpose(perm = k_7_perm_0, x = var_742)[name = tensor("transpose_36")]; - tensor q_7 = transpose(perm = q_7_perm_0, x = var_734)[name = tensor("transpose_37")]; + tensor k_7 = transpose(perm = k_7_perm_0, x = var_768)[name = tensor("transpose_36")]; + tensor q_7 = transpose(perm = q_7_perm_0, x = var_760)[name = tensor("transpose_37")]; tensor qk_7 = matmul(transpose_x = qk_7_transpose_x_1, transpose_y = qk_7_transpose_y_1, x = q_7, y = k_7)[name = tensor("qk_7")]; - tensor var_758 = const()[name = tensor("op_758"), val = tensor([1, 1])]; - tensor var_759 = reshape(shape = var_758, x = sqrt_s_t_7)[name = tensor("op_759")]; - tensor M_7 = real_div(x = encoder__causal_mask, y = var_759)[name = tensor("M_7")]; - tensor var_761 = mul(x = qk_7, y = M_7)[name = tensor("op_761")]; + tensor var_784 = const()[name = tensor("op_784"), val = tensor([1, 1])]; + tensor var_785 = reshape(shape = var_784, x = sqrt_s_t_7)[name = tensor("op_785")]; + tensor M_7 = real_div(x = inner_encoder__causal_mask, y = var_785)[name = tensor("M_7")]; + tensor var_787 = mul(x = qk_7, y = M_7)[name = tensor("op_787")]; tensor inner_7_transpose_x_0 = const()[name = tensor("inner_7_transpose_x_0"), val = tensor(false)]; tensor inner_7_transpose_y_0 = const()[name = tensor("inner_7_transpose_y_0"), val = tensor(false)]; - tensor v_7 = transpose(perm = v_7_perm_0, x = var_748)[name = tensor("transpose_35")]; - tensor inner_7 = matmul(transpose_x = inner_7_transpose_x_0, transpose_y = inner_7_transpose_y_0, x = var_761, y = v_7)[name = tensor("inner_7")]; - tensor var_763_transpose_x_0 = const()[name = tensor("op_763_transpose_x_0"), val = tensor(false)]; - tensor var_763_transpose_y_0 = const()[name = tensor("op_763_transpose_y_0"), val = tensor(false)]; - tensor var_763 = matmul(transpose_x = var_763_transpose_x_0, transpose_y = var_763_transpose_y_0, x = q_7, y = prev_kv_7)[name = tensor("op_763")]; - tensor var_764 = real_div(x = sqrt_s0_7, y = sqrt_s_t_7)[name = tensor("op_764")]; - tensor var_765 = const()[name = tensor("op_765"), val = tensor([1, 1, 1, 1])]; - tensor var_766 = reshape(shape = var_765, x = var_764)[name = tensor("op_766")]; - tensor cross_7 = mul(x = var_763, y = var_766)[name = tensor("cross_7")]; + tensor v_7 = transpose(perm = v_7_perm_0, x = var_774)[name = tensor("transpose_35")]; + tensor inner_7 = matmul(transpose_x = inner_7_transpose_x_0, transpose_y = inner_7_transpose_y_0, x = var_787, y = v_7)[name = tensor("inner_7")]; + tensor var_789_transpose_x_0 = const()[name = tensor("op_789_transpose_x_0"), val = tensor(false)]; + tensor var_789_transpose_y_0 = const()[name = tensor("op_789_transpose_y_0"), val = tensor(false)]; + tensor var_789 = matmul(transpose_x = var_789_transpose_x_0, transpose_y = var_789_transpose_y_0, x = q_7, y = prev_kv_7)[name = tensor("op_789")]; + tensor var_790 = real_div(x = sqrt_s0_7, y = sqrt_s_t_7)[name = tensor("op_790")]; + tensor var_791 = const()[name = tensor("op_791"), val = tensor([1, 1, 1, 1])]; + tensor var_792 = reshape(shape = var_791, x = var_790)[name = tensor("op_792")]; + tensor cross_7 = mul(x = var_789, y = var_792)[name = tensor("cross_7")]; tensor out_19 = add(x = inner_7, y = cross_7)[name = tensor("out_19")]; - tensor var_769 = mul(x = prev_kv_7, y = sqrt_s0_7)[name = tensor("op_769")]; - tensor var_771_transpose_x_1 = const()[name = tensor("op_771_transpose_x_1"), val = tensor(true)]; - tensor var_771_transpose_y_1 = const()[name = tensor("op_771_transpose_y_1"), val = tensor(false)]; - tensor var_771 = matmul(transpose_x = var_771_transpose_x_1, transpose_y = var_771_transpose_y_1, x = k_7, y = v_7)[name = tensor("op_771")]; - tensor new_kv_unnorm_7 = add(x = var_769, y = var_771)[name = tensor("new_kv_unnorm_7")]; - tensor var_773 = const()[name = tensor("op_773"), val = tensor(0x1p+0)]; - tensor new_scale_7 = add(x = prev_scale_7, y = var_773)[name = tensor("new_scale_7")]; - tensor var_775 = sqrt(x = new_scale_7)[name = tensor("op_775")]; - tensor nkv_1 = real_div(x = new_kv_unnorm_7, y = var_775)[name = tensor("nkv_1")]; - tensor var_777_perm_0 = const()[name = tensor("op_777_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_795 = mul(x = prev_kv_7, y = sqrt_s0_7)[name = tensor("op_795")]; + tensor var_797_transpose_x_1 = const()[name = tensor("op_797_transpose_x_1"), val = tensor(true)]; + tensor var_797_transpose_y_1 = const()[name = tensor("op_797_transpose_y_1"), val = tensor(false)]; + tensor var_797 = matmul(transpose_x = var_797_transpose_x_1, transpose_y = var_797_transpose_y_1, x = k_7, y = v_7)[name = tensor("op_797")]; + tensor new_kv_unnorm_7 = add(x = var_795, y = var_797)[name = tensor("new_kv_unnorm_7")]; + tensor var_799 = const()[name = tensor("op_799"), val = tensor(0x1p+0)]; + tensor new_scale_7 = add(x = prev_scale_7, y = var_799)[name = tensor("new_scale_7")]; + tensor var_801 = sqrt(x = new_scale_7)[name = tensor("op_801")]; + tensor nkv_1 = real_div(x = new_kv_unnorm_7, y = var_801)[name = tensor("nkv_1")]; + tensor var_803_perm_0 = const()[name = tensor("op_803_perm_0"), val = tensor([0, 2, 1, 3])]; tensor out_21_axes_0 = const()[name = tensor("out_21_axes_0"), val = tensor([-1])]; - tensor var_777 = transpose(perm = var_777_perm_0, x = out_19)[name = tensor("transpose_34")]; - tensor out_21 = layer_norm(axes = out_21_axes_0, epsilon = var_17, x = var_777)[name = tensor("out_21")]; - tensor var_781 = const()[name = tensor("op_781"), val = tensor([1, 1, 256])]; - tensor out_23 = reshape(shape = var_781, x = out_21)[name = tensor("out_23")]; - tensor var_783 = silu(x = input_137)[name = tensor("op_783")]; - tensor input_139 = mul(x = var_783, y = out_23)[name = tensor("input_139")]; - tensor ret_out_7 = linear(bias = encoder_out_proj_3_bias, weight = encoder_out_proj_3_weight, x = input_139)[name = tensor("linear_34")]; - tensor x_21 = add(x = input_133, y = ret_out_7)[name = tensor("x_21")]; + tensor var_803 = transpose(perm = var_803_perm_0, x = out_19)[name = tensor("transpose_34")]; + tensor out_21 = layer_norm(axes = out_21_axes_0, epsilon = var_44, x = var_803)[name = tensor("out_21")]; + tensor var_807 = const()[name = tensor("op_807"), val = tensor([1, 1, 256])]; + tensor out_23 = reshape(shape = var_807, x = out_21)[name = tensor("out_23")]; + tensor var_809 = silu(x = input_139)[name = tensor("op_809")]; + tensor input_141 = mul(x = var_809, y = out_23)[name = tensor("input_141")]; + tensor ret_out_7 = linear(bias = inner_encoder_out_proj_3_bias, weight = inner_encoder_out_proj_3_weight, x = input_141)[name = tensor("linear_34")]; + tensor x_21 = add(x = input_135, y = ret_out_7)[name = tensor("x_21")]; tensor window_13_begin_0 = const()[name = tensor("window_13_begin_0"), val = tensor([3, 0, 0, 0])]; tensor window_13_end_0 = const()[name = tensor("window_13_end_0"), val = tensor([4, 1, 16, 256])]; tensor window_13_end_mask_0 = const()[name = tensor("window_13_end_mask_0"), val = tensor([false, true, true, true])]; tensor window_13_squeeze_mask_0 = const()[name = tensor("window_13_squeeze_mask_0"), val = tensor([true, false, false, false])]; tensor window_13 = slice_by_index(begin = window_13_begin_0, end = window_13_end_0, end_mask = window_13_end_mask_0, squeeze_mask = window_13_squeeze_mask_0, x = enc_conv_cache)[name = tensor("window_13")]; - tensor var_794_begin_0 = const()[name = tensor("op_794_begin_0"), val = tensor([0, 1, 0])]; - tensor var_794_end_0 = const()[name = tensor("op_794_end_0"), val = tensor([1, 16, 256])]; - tensor var_794_end_mask_0 = const()[name = tensor("op_794_end_mask_0"), val = tensor([true, true, true])]; - tensor var_794 = slice_by_index(begin = var_794_begin_0, end = var_794_end_0, end_mask = var_794_end_mask_0, x = window_13)[name = tensor("op_794")]; + tensor var_820_begin_0 = const()[name = tensor("op_820_begin_0"), val = tensor([0, 1, 0])]; + tensor var_820_end_0 = const()[name = tensor("op_820_end_0"), val = tensor([1, 16, 256])]; + tensor var_820_end_mask_0 = const()[name = tensor("op_820_end_mask_0"), val = tensor([true, true, true])]; + tensor var_820 = slice_by_index(begin = var_820_begin_0, end = var_820_end_0, end_mask = var_820_end_mask_0, x = window_13)[name = tensor("op_820")]; tensor window_interleave_0 = const()[name = tensor("window_interleave_0"), val = tensor(false)]; - tensor window = concat(axis = var_26, interleave = window_interleave_0, values = (var_794, x_21))[name = tensor("window")]; - tensor input_141_interleave_0 = const()[name = tensor("input_141_interleave_0"), val = tensor(false)]; - tensor input_141 = concat(axis = var_23, interleave = input_141_interleave_0, values = window)[name = tensor("input_141")]; + tensor window = concat(axis = var_52, interleave = window_interleave_0, values = (var_820, x_21))[name = tensor("window")]; + tensor input_143_interleave_0 = const()[name = tensor("input_143_interleave_0"), val = tensor(false)]; + tensor input_143 = concat(axis = var_39, interleave = input_143_interleave_0, values = window)[name = tensor("input_143")]; tensor x_23_axes_0 = const()[name = tensor("x_23_axes_0"), val = tensor([-1])]; - tensor x_23 = layer_norm(axes = x_23_axes_0, beta = encoder_conv_module_3_sequential_0_bias, epsilon = var_28, gamma = encoder_conv_module_3_sequential_0_weight, x = input_141)[name = tensor("x_23")]; - tensor input_143_perm_0 = const()[name = tensor("input_143_perm_0"), val = tensor([0, 2, 1])]; + tensor x_23 = layer_norm(axes = x_23_axes_0, beta = inner_encoder_conv_module_3_sequential_0_bias, epsilon = var_36, gamma = inner_encoder_conv_module_3_sequential_0_weight, x = input_143)[name = tensor("x_23")]; + tensor input_145_perm_0 = const()[name = tensor("input_145_perm_0"), val = tensor([0, 2, 1])]; tensor inputs_33_pad_type_0 = const()[name = tensor("inputs_33_pad_type_0"), val = tensor("valid")]; tensor inputs_33_strides_0 = const()[name = tensor("inputs_33_strides_0"), val = tensor([1])]; tensor inputs_33_pad_0 = const()[name = tensor("inputs_33_pad_0"), val = tensor([0, 0])]; tensor inputs_33_dilations_0 = const()[name = tensor("inputs_33_dilations_0"), val = tensor([1])]; tensor inputs_33_groups_0 = const()[name = tensor("inputs_33_groups_0"), val = tensor(1)]; - tensor input_143 = transpose(perm = input_143_perm_0, x = x_23)[name = tensor("transpose_33")]; - tensor inputs_33 = conv(bias = encoder_conv_module_3_sequential_2_conv_bias, dilations = inputs_33_dilations_0, groups = inputs_33_groups_0, pad = inputs_33_pad_0, pad_type = inputs_33_pad_type_0, strides = inputs_33_strides_0, weight = encoder_conv_module_3_sequential_2_conv_weight, x = input_143)[name = tensor("inputs_33")]; - tensor var_819_split_sizes_0 = const()[name = tensor("op_819_split_sizes_0"), val = tensor([256, 256])]; - tensor var_819_axis_0 = const()[name = tensor("op_819_axis_0"), val = tensor(1)]; - tensor var_819_0, tensor var_819_1 = split(axis = var_819_axis_0, split_sizes = var_819_split_sizes_0, x = inputs_33)[name = tensor("op_819")]; - tensor var_821 = sigmoid(x = var_819_1)[name = tensor("op_821")]; - tensor inputs_35 = mul(x = var_819_0, y = var_821)[name = tensor("inputs_35")]; + tensor input_145 = transpose(perm = input_145_perm_0, x = x_23)[name = tensor("transpose_33")]; + tensor inputs_33 = conv(bias = inner_encoder_conv_module_3_sequential_2_conv_bias, dilations = inputs_33_dilations_0, groups = inputs_33_groups_0, pad = inputs_33_pad_0, pad_type = inputs_33_pad_type_0, strides = inputs_33_strides_0, weight = inner_encoder_conv_module_3_sequential_2_conv_weight, x = input_145)[name = tensor("inputs_33")]; + tensor var_845_split_sizes_0 = const()[name = tensor("op_845_split_sizes_0"), val = tensor([256, 256])]; + tensor var_845_axis_0 = const()[name = tensor("op_845_axis_0"), val = tensor(1)]; + tensor var_845_0, tensor var_845_1 = split(axis = var_845_axis_0, split_sizes = var_845_split_sizes_0, x = inputs_33)[name = tensor("op_845")]; + tensor var_847 = sigmoid(x = var_845_1)[name = tensor("op_847")]; + tensor inputs_35 = mul(x = var_845_0, y = var_847)[name = tensor("inputs_35")]; tensor outputs_aug_pad_type_0 = const()[name = tensor("outputs_aug_pad_type_0"), val = tensor("custom")]; tensor outputs_aug_pad_0 = const()[name = tensor("outputs_aug_pad_0"), val = tensor([15, 15])]; tensor outputs_aug_groups_0 = const()[name = tensor("outputs_aug_groups_0"), val = tensor(256)]; tensor outputs_aug_strides_0 = const()[name = tensor("outputs_aug_strides_0"), val = tensor([1])]; tensor outputs_aug_dilations_0 = const()[name = tensor("outputs_aug_dilations_0"), val = tensor([1])]; - tensor outputs_aug = conv(dilations = outputs_aug_dilations_0, groups = outputs_aug_groups_0, pad = outputs_aug_pad_0, pad_type = outputs_aug_pad_type_0, strides = outputs_aug_strides_0, weight = encoder_conv_module_3_sequential_4_conv_weight, x = inputs_35)[name = tensor("outputs_aug")]; - tensor input_145_begin_0 = const()[name = tensor("input_145_begin_0"), val = tensor([0, 0, 0])]; - tensor input_145_end_0 = const()[name = tensor("input_145_end_0"), val = tensor([1, 256, 16])]; - tensor input_145_end_mask_0 = const()[name = tensor("input_145_end_mask_0"), val = tensor([true, true, false])]; - tensor input_145 = slice_by_index(begin = input_145_begin_0, end = input_145_end_0, end_mask = input_145_end_mask_0, x = outputs_aug)[name = tensor("input_145")]; - tensor inputs_37 = batch_norm(beta = encoder_conv_module_3_sequential_5_bias, epsilon = var_28, gamma = encoder_conv_module_3_sequential_5_weight, mean = encoder_conv_module_3_sequential_5_running_mean, variance = encoder_conv_module_3_sequential_5_running_var, x = input_145)[name = tensor("inputs_37")]; - tensor input_147 = silu(x = inputs_37)[name = tensor("input_147")]; - tensor input_149_pad_type_0 = const()[name = tensor("input_149_pad_type_0"), val = tensor("valid")]; - tensor input_149_strides_0 = const()[name = tensor("input_149_strides_0"), val = tensor([1])]; - tensor input_149_pad_0 = const()[name = tensor("input_149_pad_0"), val = tensor([0, 0])]; - tensor input_149_dilations_0 = const()[name = tensor("input_149_dilations_0"), val = tensor([1])]; - tensor input_149_groups_0 = const()[name = tensor("input_149_groups_0"), val = tensor(1)]; - tensor input_149 = conv(bias = encoder_conv_module_3_sequential_7_conv_bias, dilations = input_149_dilations_0, groups = input_149_groups_0, pad = input_149_pad_0, pad_type = input_149_pad_type_0, strides = input_149_strides_0, weight = encoder_conv_module_3_sequential_7_conv_weight, x = input_147)[name = tensor("input_149")]; + tensor outputs_aug = conv(dilations = outputs_aug_dilations_0, groups = outputs_aug_groups_0, pad = outputs_aug_pad_0, pad_type = outputs_aug_pad_type_0, strides = outputs_aug_strides_0, weight = inner_encoder_conv_module_3_sequential_4_conv_weight, x = inputs_35)[name = tensor("outputs_aug")]; + tensor input_147_begin_0 = const()[name = tensor("input_147_begin_0"), val = tensor([0, 0, 0])]; + tensor input_147_end_0 = const()[name = tensor("input_147_end_0"), val = tensor([1, 256, 16])]; + tensor input_147_end_mask_0 = const()[name = tensor("input_147_end_mask_0"), val = tensor([true, true, false])]; + tensor input_147 = slice_by_index(begin = input_147_begin_0, end = input_147_end_0, end_mask = input_147_end_mask_0, x = outputs_aug)[name = tensor("input_147")]; + tensor inputs_37 = batch_norm(beta = inner_encoder_conv_module_3_sequential_5_bias, epsilon = var_36, gamma = inner_encoder_conv_module_3_sequential_5_weight, mean = inner_encoder_conv_module_3_sequential_5_running_mean, variance = inner_encoder_conv_module_3_sequential_5_running_var, x = input_147)[name = tensor("inputs_37")]; + tensor input_149 = silu(x = inputs_37)[name = tensor("input_149")]; + tensor input_151_pad_type_0 = const()[name = tensor("input_151_pad_type_0"), val = tensor("valid")]; + tensor input_151_strides_0 = const()[name = tensor("input_151_strides_0"), val = tensor([1])]; + tensor input_151_pad_0 = const()[name = tensor("input_151_pad_0"), val = tensor([0, 0])]; + tensor input_151_dilations_0 = const()[name = tensor("input_151_dilations_0"), val = tensor([1])]; + tensor input_151_groups_0 = const()[name = tensor("input_151_groups_0"), val = tensor(1)]; + tensor input_151 = conv(bias = inner_encoder_conv_module_3_sequential_7_conv_bias, dilations = input_151_dilations_0, groups = input_151_groups_0, pad = input_151_pad_0, pad_type = input_151_pad_type_0, strides = input_151_strides_0, weight = inner_encoder_conv_module_3_sequential_7_conv_weight, x = input_149)[name = tensor("input_151")]; tensor conv_out_7_perm_0 = const()[name = tensor("conv_out_7_perm_0"), val = tensor([0, 2, 1])]; - tensor var_852_begin_0 = const()[name = tensor("op_852_begin_0"), val = tensor([0, -1, 0])]; - tensor var_852_end_0 = const()[name = tensor("op_852_end_0"), val = tensor([1, 16, 256])]; - tensor var_852_end_mask_0 = const()[name = tensor("op_852_end_mask_0"), val = tensor([true, true, true])]; - tensor conv_out_7 = transpose(perm = conv_out_7_perm_0, x = input_149)[name = tensor("transpose_32")]; - tensor var_852 = slice_by_index(begin = var_852_begin_0, end = var_852_end_0, end_mask = var_852_end_mask_0, x = conv_out_7)[name = tensor("op_852")]; - tensor var_854_perm_0 = const()[name = tensor("op_854_perm_0"), val = tensor([1, 0, 2])]; - tensor var_854 = transpose(perm = var_854_perm_0, x = var_852)[name = tensor("transpose_31")]; - tensor input_151 = add(x = x_21, y = var_854)[name = tensor("input_151")]; - tensor input_153_axes_0 = const()[name = tensor("input_153_axes_0"), val = tensor([-1])]; - tensor input_153 = layer_norm(axes = input_153_axes_0, beta = encoder_ffn2_3_module_sequential_0_bias, epsilon = var_28, gamma = encoder_ffn2_3_module_sequential_0_weight, x = input_151)[name = tensor("input_153")]; - tensor inputs = linear(bias = encoder_ffn2_3_module_sequential_1_linear_bias, weight = encoder_ffn2_3_module_sequential_1_linear_weight, x = input_153)[name = tensor("linear_35")]; - tensor input_155 = silu(x = inputs)[name = tensor("input_155")]; - tensor input_159 = linear(bias = encoder_ffn2_3_module_sequential_4_linear_bias, weight = encoder_ffn2_3_module_sequential_4_linear_weight, x = input_155)[name = tensor("linear_36")]; - tensor var_877 = const()[name = tensor("op_877"), val = tensor(0x1p-1)]; - tensor var_878 = mul(x = input_159, y = var_877)[name = tensor("op_878")]; - tensor input_161 = add(x = var_878, y = input_151)[name = tensor("input_161")]; + tensor var_878_begin_0 = const()[name = tensor("op_878_begin_0"), val = tensor([0, -1, 0])]; + tensor var_878_end_0 = const()[name = tensor("op_878_end_0"), val = tensor([1, 16, 256])]; + tensor var_878_end_mask_0 = const()[name = tensor("op_878_end_mask_0"), val = tensor([true, true, true])]; + tensor conv_out_7 = transpose(perm = conv_out_7_perm_0, x = input_151)[name = tensor("transpose_32")]; + tensor var_878 = slice_by_index(begin = var_878_begin_0, end = var_878_end_0, end_mask = var_878_end_mask_0, x = conv_out_7)[name = tensor("op_878")]; + tensor var_880_perm_0 = const()[name = tensor("op_880_perm_0"), val = tensor([1, 0, 2])]; + tensor var_880 = transpose(perm = var_880_perm_0, x = var_878)[name = tensor("transpose_31")]; + tensor input_153 = add(x = x_21, y = var_880)[name = tensor("input_153")]; + tensor input_155_axes_0 = const()[name = tensor("input_155_axes_0"), val = tensor([-1])]; + tensor input_155 = layer_norm(axes = input_155_axes_0, beta = inner_encoder_ffn2_3_module_sequential_0_bias, epsilon = var_36, gamma = inner_encoder_ffn2_3_module_sequential_0_weight, x = input_153)[name = tensor("input_155")]; + tensor inputs = linear(bias = inner_encoder_ffn2_3_module_sequential_1_linear_bias, weight = inner_encoder_ffn2_3_module_sequential_1_linear_weight, x = input_155)[name = tensor("linear_35")]; + tensor input_157 = silu(x = inputs)[name = tensor("input_157")]; + tensor input_161 = linear(bias = inner_encoder_ffn2_3_module_sequential_4_linear_bias, weight = inner_encoder_ffn2_3_module_sequential_4_linear_weight, x = input_157)[name = tensor("linear_36")]; + tensor var_903 = const()[name = tensor("op_903"), val = tensor(0x1p-1)]; + tensor var_904 = mul(x = input_161, y = var_903)[name = tensor("op_904")]; + tensor input_163 = add(x = var_904, y = input_153)[name = tensor("input_163")]; tensor x_25_axes_0 = const()[name = tensor("x_25_axes_0"), val = tensor([-1])]; - tensor x_25 = layer_norm(axes = x_25_axes_0, beta = encoder_layer_norm_3_bias, epsilon = var_28, gamma = encoder_layer_norm_3_weight, x = input_161)[name = tensor("x_25")]; + tensor x_25 = layer_norm(axes = x_25_axes_0, beta = inner_encoder_layer_norm_3_bias, epsilon = var_36, gamma = inner_encoder_layer_norm_3_weight, x = input_163)[name = tensor("x_25")]; tensor x_ct_perm_0 = const()[name = tensor("x_ct_perm_0"), val = tensor([0, 2, 1])]; tensor cat_interleave_0 = const()[name = tensor("cat_interleave_0"), val = tensor(false)]; tensor x_ct = transpose(perm = x_ct_perm_0, x = x_25)[name = tensor("transpose_30")]; - tensor cat = concat(axis = var_20, interleave = cat_interleave_0, values = (cnn_window, x_ct))[name = tensor("cat")]; + tensor cat = concat(axis = var_41, interleave = cat_interleave_0, values = (cnn_window, x_ct))[name = tensor("cat")]; tensor conv_out_pad_type_0 = const()[name = tensor("conv_out_pad_type_0"), val = tensor("valid")]; tensor conv_out_strides_0 = const()[name = tensor("conv_out_strides_0"), val = tensor([1])]; tensor conv_out_pad_0 = const()[name = tensor("conv_out_pad_0"), val = tensor([0, 0])]; tensor conv_out_dilations_0 = const()[name = tensor("conv_out_dilations_0"), val = tensor([1])]; tensor conv_out_groups_0 = const()[name = tensor("conv_out_groups_0"), val = tensor(1)]; - tensor conv_out = conv(bias = encoder_cnn_bias, dilations = conv_out_dilations_0, groups = conv_out_groups_0, pad = conv_out_pad_0, pad_type = conv_out_pad_type_0, strides = conv_out_strides_0, weight = encoder_cnn_weight, x = cat)[name = tensor("conv_out")]; - tensor var_896_begin_0 = const()[name = tensor("op_896_begin_0"), val = tensor([0, 0, 1])]; - tensor var_896_end_0 = const()[name = tensor("op_896_end_0"), val = tensor([1, 256, 19])]; - tensor var_896_end_mask_0 = const()[name = tensor("op_896_end_mask_0"), val = tensor([true, true, true])]; - tensor cnn_window_new = slice_by_index(begin = var_896_begin_0, end = var_896_end_0, end_mask = var_896_end_mask_0, x = cat)[name = tensor("op_896")]; - tensor input_163_perm_0 = const()[name = tensor("input_163_perm_0"), val = tensor([0, 2, 1])]; - tensor var_898 = const()[name = tensor("op_898"), val = tensor([-1])]; - tensor input_163 = transpose(perm = input_163_perm_0, x = conv_out)[name = tensor("transpose_29")]; - tensor var_899 = reduce_l2_norm(axes = var_898, keep_dims = var_29, x = input_163)[name = tensor("op_899")]; + tensor conv_out = conv(bias = inner_encoder_cnn_bias, dilations = conv_out_dilations_0, groups = conv_out_groups_0, pad = conv_out_pad_0, pad_type = conv_out_pad_type_0, strides = conv_out_strides_0, weight = inner_encoder_cnn_weight, x = cat)[name = tensor("conv_out")]; + tensor var_922_begin_0 = const()[name = tensor("op_922_begin_0"), val = tensor([0, 0, 1])]; + tensor var_922_end_0 = const()[name = tensor("op_922_end_0"), val = tensor([1, 256, 19])]; + tensor var_922_end_mask_0 = const()[name = tensor("op_922_end_mask_0"), val = tensor([true, true, true])]; + tensor cnn_window_new = slice_by_index(begin = var_922_begin_0, end = var_922_end_0, end_mask = var_922_end_mask_0, x = cat)[name = tensor("op_922")]; + tensor input_165_perm_0 = const()[name = tensor("input_165_perm_0"), val = tensor([0, 2, 1])]; + tensor var_924 = const()[name = tensor("op_924"), val = tensor([-1])]; + tensor input_165 = transpose(perm = input_165_perm_0, x = conv_out)[name = tensor("transpose_29")]; + tensor var_925 = reduce_l2_norm(axes = var_924, keep_dims = var_35, x = input_165)[name = tensor("op_925")]; tensor const_12 = const()[name = tensor("const_12"), val = tensor(0x1.fffffep+127)]; - tensor clip_0 = clip(alpha = var_11, beta = const_12, x = var_899)[name = tensor("clip_0")]; - tensor emb = real_div(x = input_163, y = clip_0)[name = tensor("emb")]; - tensor var_903_axis_0 = const()[name = tensor("op_903_axis_0"), val = tensor(0)]; - tensor enc_kv_new = stack(axis = var_903_axis_0, values = (var_206, var_396, var_586, nkv_1))[name = tensor("op_903")]; - tensor var_905_axis_0 = const()[name = tensor("op_905_axis_0"), val = tensor(0)]; - tensor enc_scale_new = stack(axis = var_905_axis_0, values = (new_scale_1, new_scale_3, new_scale_5, new_scale_7))[name = tensor("op_905")]; - tensor var_907_axis_0 = const()[name = tensor("op_907_axis_0"), val = tensor(0)]; - tensor enc_conv_cache_new = stack(axis = var_907_axis_0, values = (window_3, window_7, window_11, window))[name = tensor("op_907")]; - tensor var_916 = const()[name = tensor("op_916"), val = tensor(0x1.5798eep-27)]; - tensor var_921 = const()[name = tensor("op_921"), val = tensor(0x1.4f8b58p-17)]; - tensor var_923 = const()[name = tensor("op_923"), val = tensor(0x1.0c6f7ap-20)]; - tensor var_924 = const()[name = tensor("op_924"), val = tensor(true)]; - tensor var_926 = const()[name = tensor("op_926"), val = tensor(0x1p+0)]; - tensor var_930 = const()[name = tensor("op_930"), val = tensor(-1)]; - tensor var_936 = const()[name = tensor("op_936"), val = tensor(0)]; - tensor var_993 = const()[name = tensor("op_993"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(44395328)))]; - tensor var_998_axes_0 = const()[name = tensor("op_998_axes_0"), val = tensor([2])]; - tensor var_998 = expand_dims(axes = var_998_axes_0, x = emb)[name = tensor("op_998")]; + tensor clip_0 = clip(alpha = var_49, beta = const_12, x = var_925)[name = tensor("clip_0")]; + tensor emb = real_div(x = input_165, y = clip_0)[name = tensor("emb")]; + tensor var_929_axis_0 = const()[name = tensor("op_929_axis_0"), val = tensor(0)]; + tensor enc_kv_new = stack(axis = var_929_axis_0, values = (var_232, var_422, var_612, nkv_1))[name = tensor("op_929")]; + tensor var_931_axis_0 = const()[name = tensor("op_931_axis_0"), val = tensor(0)]; + tensor enc_scale_new = stack(axis = var_931_axis_0, values = (new_scale_1, new_scale_3, new_scale_5, new_scale_7))[name = tensor("op_931")]; + tensor var_933_axis_0 = const()[name = tensor("op_933_axis_0"), val = tensor(0)]; + tensor enc_conv_cache_new = stack(axis = var_933_axis_0, values = (window_3, window_7, window_11, window))[name = tensor("op_933")]; + tensor var_996 = const()[name = tensor("op_996"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(44395328)))]; + tensor var_1001_axes_0 = const()[name = tensor("op_1001_axes_0"), val = tensor([2])]; + tensor var_1001 = expand_dims(axes = var_1001_axes_0, x = emb)[name = tensor("op_1001")]; tensor emb_exp_reps_0 = const()[name = tensor("emb_exp_reps_0"), val = tensor([1, 1, 12, 1])]; - tensor emb_exp = tile(reps = emb_exp_reps_0, x = var_998)[name = tensor("emb_exp")]; - tensor input_165_interleave_0 = const()[name = tensor("input_165_interleave_0"), val = tensor(false)]; - tensor input_165 = concat(axis = var_930, interleave = input_165_interleave_0, values = (emb_exp, var_993))[name = tensor("input_165")]; - tensor x_27 = linear(bias = decoder_convert_bias, weight = decoder_convert_weight, x = input_165)[name = tensor("linear_37")]; - tensor var_1006_perm_0 = const()[name = tensor("op_1006_perm_0"), val = tensor([0, 2, 1, 3])]; - tensor var_1010 = const()[name = tensor("op_1010"), val = tensor([12, 1, 256])]; - tensor var_1006 = transpose(perm = var_1006_perm_0, x = x_27)[name = tensor("transpose_28")]; - tensor x_29 = reshape(shape = var_1010, x = var_1006)[name = tensor("x_29")]; + tensor emb_exp = tile(reps = emb_exp_reps_0, x = var_1001)[name = tensor("emb_exp")]; + tensor input_167_interleave_0 = const()[name = tensor("input_167_interleave_0"), val = tensor(false)]; + tensor input_167 = concat(axis = var_42, interleave = input_167_interleave_0, values = (emb_exp, var_996))[name = tensor("input_167")]; + tensor x_27 = linear(bias = inner_decoder_convert_bias, weight = inner_decoder_convert_weight, x = input_167)[name = tensor("linear_37")]; + tensor var_1009_perm_0 = const()[name = tensor("op_1009_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_1013 = const()[name = tensor("op_1013"), val = tensor([12, 1, 256])]; + tensor var_1009 = transpose(perm = var_1009_perm_0, x = x_27)[name = tensor("transpose_28")]; + tensor x_29 = reshape(shape = var_1013, x = var_1009)[name = tensor("x_29")]; tensor prev_kv_9_begin_0 = const()[name = tensor("prev_kv_9_begin_0"), val = tensor([0, 0, 0, 0, 0])]; tensor prev_kv_9_end_0 = const()[name = tensor("prev_kv_9_end_0"), val = tensor([1, 12, 4, 64, 64])]; tensor prev_kv_9_end_mask_0 = const()[name = tensor("prev_kv_9_end_mask_0"), val = tensor([false, true, true, true, true])]; @@ -846,131 +845,131 @@ program(1.0) tensor prev_scale_9_end_mask_0 = const()[name = tensor("prev_scale_9_end_mask_0"), val = tensor([false, true])]; tensor prev_scale_9_squeeze_mask_0 = const()[name = tensor("prev_scale_9_squeeze_mask_0"), val = tensor([true, false])]; tensor prev_scale_9 = slice_by_index(begin = prev_scale_9_begin_0, end = prev_scale_9_end_0, end_mask = prev_scale_9_end_mask_0, squeeze_mask = prev_scale_9_squeeze_mask_0, x = dec_scale)[name = tensor("prev_scale_9")]; - tensor var_1018 = linear(bias = decoder_q_proj_0_bias, weight = decoder_q_proj_0_weight, x = x_29)[name = tensor("linear_38")]; - tensor var_1019 = const()[name = tensor("op_1019"), val = tensor([12, 1, 4, 64])]; - tensor var_1020 = reshape(shape = var_1019, x = var_1018)[name = tensor("op_1020")]; + tensor var_1021 = linear(bias = inner_decoder_q_proj_0_bias, weight = inner_decoder_q_proj_0_weight, x = x_29)[name = tensor("linear_38")]; + tensor var_1022 = const()[name = tensor("op_1022"), val = tensor([12, 1, 4, 64])]; + tensor var_1023 = reshape(shape = var_1022, x = var_1021)[name = tensor("op_1023")]; tensor q_9_perm_0 = const()[name = tensor("q_9_perm_0"), val = tensor([0, 2, 1, 3])]; - tensor var_1024 = linear(bias = decoder_k_proj_0_bias, weight = decoder_k_proj_0_weight, x = x_29)[name = tensor("linear_39")]; - tensor var_1025 = const()[name = tensor("op_1025"), val = tensor(0x1p-3)]; - tensor var_1026 = mul(x = var_1024, y = var_1025)[name = tensor("op_1026")]; - tensor var_1027 = const()[name = tensor("op_1027"), val = tensor([12, 1, 4, 64])]; - tensor var_1028 = reshape(shape = var_1027, x = var_1026)[name = tensor("op_1028")]; + tensor var_1027 = linear(bias = inner_decoder_k_proj_0_bias, weight = inner_decoder_k_proj_0_weight, x = x_29)[name = tensor("linear_39")]; + tensor var_1028 = const()[name = tensor("op_1028"), val = tensor(0x1p-3)]; + tensor var_1029 = mul(x = var_1027, y = var_1028)[name = tensor("op_1029")]; + tensor var_1030 = const()[name = tensor("op_1030"), val = tensor([12, 1, 4, 64])]; + tensor var_1031 = reshape(shape = var_1030, x = var_1029)[name = tensor("op_1031")]; tensor k_9_perm_0 = const()[name = tensor("k_9_perm_0"), val = tensor([0, 2, 1, 3])]; - tensor var_1032 = linear(bias = decoder_v_proj_0_bias, weight = decoder_v_proj_0_weight, x = x_29)[name = tensor("linear_40")]; - tensor var_1033 = const()[name = tensor("op_1033"), val = tensor([12, 1, 4, 64])]; - tensor var_1034 = reshape(shape = var_1033, x = var_1032)[name = tensor("op_1034")]; + tensor var_1035 = linear(bias = inner_decoder_v_proj_0_bias, weight = inner_decoder_v_proj_0_weight, x = x_29)[name = tensor("linear_40")]; + tensor var_1036 = const()[name = tensor("op_1036"), val = tensor([12, 1, 4, 64])]; + tensor var_1037 = reshape(shape = var_1036, x = var_1035)[name = tensor("op_1037")]; tensor v_9_perm_0 = const()[name = tensor("v_9_perm_0"), val = tensor([0, 2, 1, 3])]; - tensor input_169 = linear(bias = decoder_g_proj_0_bias, weight = decoder_g_proj_0_weight, x = x_29)[name = tensor("linear_41")]; + tensor input_171 = linear(bias = inner_decoder_g_proj_0_bias, weight = inner_decoder_g_proj_0_weight, x = x_29)[name = tensor("linear_41")]; tensor cumsum_mask_1_exclusive_0 = const()[name = tensor("cumsum_mask_1_exclusive_0"), val = tensor(false)]; tensor cumsum_mask_1_reverse_0 = const()[name = tensor("cumsum_mask_1_reverse_0"), val = tensor(false)]; - tensor cumsum_mask_1 = cumsum(axis = var_936, exclusive = cumsum_mask_1_exclusive_0, reverse = cumsum_mask_1_reverse_0, x = valid_mask)[name = tensor("cumsum_mask_1")]; + tensor cumsum_mask_1 = cumsum(axis = var_39, exclusive = cumsum_mask_1_exclusive_0, reverse = cumsum_mask_1_reverse_0, x = valid_mask)[name = tensor("cumsum_mask_1")]; tensor sqrt_s0_9 = sqrt(x = prev_scale_9)[name = tensor("sqrt_s0_9")]; tensor s_t_9 = add(x = prev_scale_9, y = cumsum_mask_1)[name = tensor("s_t_9")]; tensor const_20 = const()[name = tensor("const_20"), val = tensor(0x1.fffffep+127)]; - tensor clip_1 = clip(alpha = var_926, beta = const_20, x = s_t_9)[name = tensor("clip_1")]; + tensor clip_1 = clip(alpha = var_29, beta = const_20, x = s_t_9)[name = tensor("clip_1")]; tensor sqrt_s_t_9 = sqrt(x = clip_1)[name = tensor("sqrt_s_t_9")]; tensor qk_9_transpose_x_1 = const()[name = tensor("qk_9_transpose_x_1"), val = tensor(false)]; tensor qk_9_transpose_y_1 = const()[name = tensor("qk_9_transpose_y_1"), val = tensor(true)]; - tensor k_9 = transpose(perm = k_9_perm_0, x = var_1028)[name = tensor("transpose_26")]; - tensor q_9 = transpose(perm = q_9_perm_0, x = var_1020)[name = tensor("transpose_27")]; + tensor k_9 = transpose(perm = k_9_perm_0, x = var_1031)[name = tensor("transpose_26")]; + tensor q_9 = transpose(perm = q_9_perm_0, x = var_1023)[name = tensor("transpose_27")]; tensor qk_9 = matmul(transpose_x = qk_9_transpose_x_1, transpose_y = qk_9_transpose_y_1, x = q_9, y = k_9)[name = tensor("qk_9")]; - tensor var_1046 = const()[name = tensor("op_1046"), val = tensor([1, 1])]; - tensor var_1047 = reshape(shape = var_1046, x = valid_mask)[name = tensor("op_1047")]; tensor var_1049 = const()[name = tensor("op_1049"), val = tensor([1, 1])]; - tensor var_1050 = reshape(shape = var_1049, x = sqrt_s_t_9)[name = tensor("op_1050")]; - tensor M_9 = real_div(x = var_1047, y = var_1050)[name = tensor("M_9")]; - tensor var_1052 = mul(x = qk_9, y = M_9)[name = tensor("op_1052")]; + tensor var_1050 = reshape(shape = var_1049, x = valid_mask)[name = tensor("op_1050")]; + tensor var_1052 = const()[name = tensor("op_1052"), val = tensor([1, 1])]; + tensor var_1053 = reshape(shape = var_1052, x = sqrt_s_t_9)[name = tensor("op_1053")]; + tensor M_9 = real_div(x = var_1050, y = var_1053)[name = tensor("M_9")]; + tensor var_1055 = mul(x = qk_9, y = M_9)[name = tensor("op_1055")]; tensor inner_9_transpose_x_0 = const()[name = tensor("inner_9_transpose_x_0"), val = tensor(false)]; tensor inner_9_transpose_y_0 = const()[name = tensor("inner_9_transpose_y_0"), val = tensor(false)]; - tensor v_9 = transpose(perm = v_9_perm_0, x = var_1034)[name = tensor("transpose_25")]; - tensor inner_9 = matmul(transpose_x = inner_9_transpose_x_0, transpose_y = inner_9_transpose_y_0, x = var_1052, y = v_9)[name = tensor("inner_9")]; - tensor var_1054_transpose_x_0 = const()[name = tensor("op_1054_transpose_x_0"), val = tensor(false)]; - tensor var_1054_transpose_y_0 = const()[name = tensor("op_1054_transpose_y_0"), val = tensor(false)]; - tensor var_1054 = matmul(transpose_x = var_1054_transpose_x_0, transpose_y = var_1054_transpose_y_0, x = q_9, y = prev_kv_9)[name = tensor("op_1054")]; - tensor var_1055 = real_div(x = sqrt_s0_9, y = sqrt_s_t_9)[name = tensor("op_1055")]; - tensor var_1056 = const()[name = tensor("op_1056"), val = tensor([1, 1, 1, 1])]; - tensor var_1057 = reshape(shape = var_1056, x = var_1055)[name = tensor("op_1057")]; - tensor cross_9 = mul(x = var_1054, y = var_1057)[name = tensor("cross_9")]; + tensor v_9 = transpose(perm = v_9_perm_0, x = var_1037)[name = tensor("transpose_25")]; + tensor inner_9 = matmul(transpose_x = inner_9_transpose_x_0, transpose_y = inner_9_transpose_y_0, x = var_1055, y = v_9)[name = tensor("inner_9")]; + tensor var_1057_transpose_x_0 = const()[name = tensor("op_1057_transpose_x_0"), val = tensor(false)]; + tensor var_1057_transpose_y_0 = const()[name = tensor("op_1057_transpose_y_0"), val = tensor(false)]; + tensor var_1057 = matmul(transpose_x = var_1057_transpose_x_0, transpose_y = var_1057_transpose_y_0, x = q_9, y = prev_kv_9)[name = tensor("op_1057")]; + tensor var_1058 = real_div(x = sqrt_s0_9, y = sqrt_s_t_9)[name = tensor("op_1058")]; + tensor var_1059 = const()[name = tensor("op_1059"), val = tensor([1, 1, 1, 1])]; + tensor var_1060 = reshape(shape = var_1059, x = var_1058)[name = tensor("op_1060")]; + tensor cross_9 = mul(x = var_1057, y = var_1060)[name = tensor("cross_9")]; tensor out_25 = add(x = inner_9, y = cross_9)[name = tensor("out_25")]; - tensor var_1060 = const()[name = tensor("op_1060"), val = tensor([1, 1, 1, 1])]; - tensor var_1061 = reshape(shape = var_1060, x = valid_mask)[name = tensor("op_1061")]; - tensor v_masked_1 = mul(x = v_9, y = var_1061)[name = tensor("v_masked_1")]; - tensor var_1063 = mul(x = prev_kv_9, y = sqrt_s0_9)[name = tensor("op_1063")]; - tensor var_1065_transpose_x_1 = const()[name = tensor("op_1065_transpose_x_1"), val = tensor(true)]; - tensor var_1065_transpose_y_1 = const()[name = tensor("op_1065_transpose_y_1"), val = tensor(false)]; - tensor var_1065 = matmul(transpose_x = var_1065_transpose_x_1, transpose_y = var_1065_transpose_y_1, x = k_9, y = v_masked_1)[name = tensor("op_1065")]; - tensor new_kv_unnorm_9 = add(x = var_1063, y = var_1065)[name = tensor("new_kv_unnorm_9")]; - tensor var_1067_keep_dims_0 = const()[name = tensor("op_1067_keep_dims_0"), val = tensor(false)]; - tensor var_1067 = reduce_sum(keep_dims = var_1067_keep_dims_0, x = valid_mask)[name = tensor("op_1067")]; - tensor var_1068 = const()[name = tensor("op_1068"), val = tensor([1])]; - tensor var_1069 = reshape(shape = var_1068, x = var_1067)[name = tensor("op_1069")]; - tensor new_scale_9 = add(x = prev_scale_9, y = var_1069)[name = tensor("new_scale_9")]; + tensor var_1063 = const()[name = tensor("op_1063"), val = tensor([1, 1, 1, 1])]; + tensor var_1064 = reshape(shape = var_1063, x = valid_mask)[name = tensor("op_1064")]; + tensor v_masked_1 = mul(x = v_9, y = var_1064)[name = tensor("v_masked_1")]; + tensor var_1066 = mul(x = prev_kv_9, y = sqrt_s0_9)[name = tensor("op_1066")]; + tensor var_1068_transpose_x_1 = const()[name = tensor("op_1068_transpose_x_1"), val = tensor(true)]; + tensor var_1068_transpose_y_1 = const()[name = tensor("op_1068_transpose_y_1"), val = tensor(false)]; + tensor var_1068 = matmul(transpose_x = var_1068_transpose_x_1, transpose_y = var_1068_transpose_y_1, x = k_9, y = v_masked_1)[name = tensor("op_1068")]; + tensor new_kv_unnorm_9 = add(x = var_1066, y = var_1068)[name = tensor("new_kv_unnorm_9")]; + tensor var_1070_keep_dims_0 = const()[name = tensor("op_1070_keep_dims_0"), val = tensor(false)]; + tensor var_1070 = reduce_sum(keep_dims = var_1070_keep_dims_0, x = valid_mask)[name = tensor("op_1070")]; + tensor var_1071 = const()[name = tensor("op_1071"), val = tensor([1])]; + tensor var_1072 = reshape(shape = var_1071, x = var_1070)[name = tensor("op_1072")]; + tensor new_scale_9 = add(x = prev_scale_9, y = var_1072)[name = tensor("new_scale_9")]; tensor const_21 = const()[name = tensor("const_21"), val = tensor(0x1.fffffep+127)]; - tensor clip_2 = clip(alpha = var_926, beta = const_21, x = new_scale_9)[name = tensor("clip_2")]; + tensor clip_2 = clip(alpha = var_29, beta = const_21, x = new_scale_9)[name = tensor("clip_2")]; tensor sqrt_new_scale_1 = sqrt(x = clip_2)[name = tensor("sqrt_new_scale_1")]; - tensor var_1073 = real_div(x = new_kv_unnorm_9, y = sqrt_new_scale_1)[name = tensor("op_1073")]; - tensor var_1074_perm_0 = const()[name = tensor("op_1074_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_1076 = real_div(x = new_kv_unnorm_9, y = sqrt_new_scale_1)[name = tensor("op_1076")]; + tensor var_1077_perm_0 = const()[name = tensor("op_1077_perm_0"), val = tensor([0, 2, 1, 3])]; tensor out_27_axes_0 = const()[name = tensor("out_27_axes_0"), val = tensor([-1])]; - tensor var_1074 = transpose(perm = var_1074_perm_0, x = out_25)[name = tensor("transpose_24")]; - tensor out_27 = layer_norm(axes = out_27_axes_0, epsilon = var_923, x = var_1074)[name = tensor("out_27")]; - tensor var_1078 = const()[name = tensor("op_1078"), val = tensor([12, 1, 256])]; - tensor out_29 = reshape(shape = var_1078, x = out_27)[name = tensor("out_29")]; - tensor var_1080 = silu(x = input_169)[name = tensor("op_1080")]; - tensor input_171 = mul(x = var_1080, y = out_29)[name = tensor("input_171")]; - tensor ret_out_9 = linear(bias = decoder_out_proj_0_bias, weight = decoder_out_proj_0_weight, x = input_171)[name = tensor("linear_42")]; - tensor input_173 = add(x = x_29, y = ret_out_9)[name = tensor("input_173")]; + tensor var_1077 = transpose(perm = var_1077_perm_0, x = out_25)[name = tensor("transpose_24")]; + tensor out_27 = layer_norm(axes = out_27_axes_0, epsilon = var_44, x = var_1077)[name = tensor("out_27")]; + tensor var_1081 = const()[name = tensor("op_1081"), val = tensor([12, 1, 256])]; + tensor out_29 = reshape(shape = var_1081, x = out_27)[name = tensor("out_29")]; + tensor var_1083 = silu(x = input_171)[name = tensor("op_1083")]; + tensor input_173 = mul(x = var_1083, y = out_29)[name = tensor("input_173")]; + tensor ret_out_9 = linear(bias = inner_decoder_out_proj_0_bias, weight = inner_decoder_out_proj_0_weight, x = input_173)[name = tensor("linear_42")]; + tensor input_175 = add(x = x_29, y = ret_out_9)[name = tensor("input_175")]; tensor xt_1_axes_0 = const()[name = tensor("xt_1_axes_0"), val = tensor([-1])]; - tensor xt_1 = layer_norm(axes = xt_1_axes_0, beta = decoder_norm11_0_bias, epsilon = var_921, gamma = decoder_norm11_0_weight, x = input_173)[name = tensor("xt_1")]; - tensor var_1090 = const()[name = tensor("op_1090"), val = tensor([1, 12, 1, 256])]; - tensor var_1091 = reshape(shape = var_1090, x = xt_1)[name = tensor("op_1091")]; - tensor var_1092_perm_0 = const()[name = tensor("op_1092_perm_0"), val = tensor([0, 2, 1, 3])]; - tensor var_1095 = const()[name = tensor("op_1095"), val = tensor([1, 12, 256])]; - tensor var_1092 = transpose(perm = var_1092_perm_0, x = var_1091)[name = tensor("transpose_23")]; - tensor query_1 = reshape(shape = var_1095, x = var_1092)[name = tensor("query_1")]; + tensor xt_1 = layer_norm(axes = xt_1_axes_0, beta = inner_decoder_norm11_0_bias, epsilon = var_36, gamma = inner_decoder_norm11_0_weight, x = input_175)[name = tensor("xt_1")]; + tensor var_1093 = const()[name = tensor("op_1093"), val = tensor([1, 12, 1, 256])]; + tensor var_1094 = reshape(shape = var_1093, x = xt_1)[name = tensor("op_1094")]; + tensor var_1095_perm_0 = const()[name = tensor("op_1095_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_1098 = const()[name = tensor("op_1098"), val = tensor([1, 12, 256])]; + tensor var_1095 = transpose(perm = var_1095_perm_0, x = var_1094)[name = tensor("transpose_23")]; + tensor query_1 = reshape(shape = var_1098, x = var_1095)[name = tensor("query_1")]; tensor query_3_perm_0 = const()[name = tensor("query_3_perm_0"), val = tensor([1, 0, 2])]; tensor query_3 = transpose(perm = query_3_perm_0, x = query_1)[name = tensor("transpose_22")]; - tensor var_1118 = linear(bias = decoder_self_attn2_0_in_proj_bias, weight = decoder_self_attn2_0_in_proj_weight, x = query_3)[name = tensor("linear_43")]; + tensor var_1121 = linear(bias = inner_decoder_self_attn2_0_in_proj_bias, weight = inner_decoder_self_attn2_0_in_proj_weight, x = query_3)[name = tensor("linear_43")]; tensor concat_1 = const()[name = tensor("concat_1"), val = tensor([12, 1, 3, 256])]; - tensor var_1120 = reshape(shape = concat_1, x = var_1118)[name = tensor("op_1120")]; - tensor var_1121_axes_0 = const()[name = tensor("op_1121_axes_0"), val = tensor([0])]; - tensor var_1121 = expand_dims(axes = var_1121_axes_0, x = var_1120)[name = tensor("op_1121")]; - tensor var_1122_perm_0 = const()[name = tensor("op_1122_perm_0"), val = tensor([-2, 1, 2, 0, 4])]; - tensor var_1123_axes_0 = const()[name = tensor("op_1123_axes_0"), val = tensor([-2])]; - tensor var_1122 = transpose(perm = var_1122_perm_0, x = var_1121)[name = tensor("transpose_21")]; - tensor var_1123 = squeeze(axes = var_1123_axes_0, x = var_1122)[name = tensor("op_1123")]; + tensor var_1123 = reshape(shape = concat_1, x = var_1121)[name = tensor("op_1123")]; + tensor var_1124_axes_0 = const()[name = tensor("op_1124_axes_0"), val = tensor([0])]; + tensor var_1124 = expand_dims(axes = var_1124_axes_0, x = var_1123)[name = tensor("op_1124")]; + tensor var_1125_perm_0 = const()[name = tensor("op_1125_perm_0"), val = tensor([-2, 1, 2, 0, 4])]; + tensor var_1126_axes_0 = const()[name = tensor("op_1126_axes_0"), val = tensor([-2])]; + tensor var_1125 = transpose(perm = var_1125_perm_0, x = var_1124)[name = tensor("transpose_21")]; + tensor var_1126 = squeeze(axes = var_1126_axes_0, x = var_1125)[name = tensor("op_1126")]; tensor q_11_begin_0 = const()[name = tensor("q_11_begin_0"), val = tensor([0, 0, 0, 0])]; tensor q_11_end_0 = const()[name = tensor("q_11_end_0"), val = tensor([1, 12, 1, 256])]; tensor q_11_end_mask_0 = const()[name = tensor("q_11_end_mask_0"), val = tensor([false, true, true, true])]; tensor q_11_squeeze_mask_0 = const()[name = tensor("q_11_squeeze_mask_0"), val = tensor([true, false, false, false])]; - tensor q_11 = slice_by_index(begin = q_11_begin_0, end = q_11_end_0, end_mask = q_11_end_mask_0, squeeze_mask = q_11_squeeze_mask_0, x = var_1123)[name = tensor("q_11")]; + tensor q_11 = slice_by_index(begin = q_11_begin_0, end = q_11_end_0, end_mask = q_11_end_mask_0, squeeze_mask = q_11_squeeze_mask_0, x = var_1126)[name = tensor("q_11")]; tensor k_11_begin_0 = const()[name = tensor("k_11_begin_0"), val = tensor([1, 0, 0, 0])]; tensor k_11_end_0 = const()[name = tensor("k_11_end_0"), val = tensor([2, 12, 1, 256])]; tensor k_11_end_mask_0 = const()[name = tensor("k_11_end_mask_0"), val = tensor([false, true, true, true])]; tensor k_11_squeeze_mask_0 = const()[name = tensor("k_11_squeeze_mask_0"), val = tensor([true, false, false, false])]; - tensor k_11 = slice_by_index(begin = k_11_begin_0, end = k_11_end_0, end_mask = k_11_end_mask_0, squeeze_mask = k_11_squeeze_mask_0, x = var_1123)[name = tensor("k_11")]; + tensor k_11 = slice_by_index(begin = k_11_begin_0, end = k_11_end_0, end_mask = k_11_end_mask_0, squeeze_mask = k_11_squeeze_mask_0, x = var_1126)[name = tensor("k_11")]; tensor v_11_begin_0 = const()[name = tensor("v_11_begin_0"), val = tensor([2, 0, 0, 0])]; tensor v_11_end_0 = const()[name = tensor("v_11_end_0"), val = tensor([3, 12, 1, 256])]; tensor v_11_end_mask_0 = const()[name = tensor("v_11_end_mask_0"), val = tensor([false, true, true, true])]; tensor v_11_squeeze_mask_0 = const()[name = tensor("v_11_squeeze_mask_0"), val = tensor([true, false, false, false])]; - tensor v_11 = slice_by_index(begin = v_11_begin_0, end = v_11_end_0, end_mask = v_11_end_mask_0, squeeze_mask = v_11_squeeze_mask_0, x = var_1123)[name = tensor("v_11")]; - tensor var_1131 = const()[name = tensor("op_1131"), val = tensor([12, 4, 64])]; - tensor var_1132 = reshape(shape = var_1131, x = q_11)[name = tensor("op_1132")]; + tensor v_11 = slice_by_index(begin = v_11_begin_0, end = v_11_end_0, end_mask = v_11_end_mask_0, squeeze_mask = v_11_squeeze_mask_0, x = var_1126)[name = tensor("v_11")]; + tensor var_1134 = const()[name = tensor("op_1134"), val = tensor([12, 4, 64])]; + tensor var_1135 = reshape(shape = var_1134, x = q_11)[name = tensor("op_1135")]; tensor q_13_perm_0 = const()[name = tensor("q_13_perm_0"), val = tensor([1, 0, 2])]; - tensor var_1138 = const()[name = tensor("op_1138"), val = tensor([12, 4, 64])]; - tensor var_1139 = reshape(shape = var_1138, x = k_11)[name = tensor("op_1139")]; + tensor var_1141 = const()[name = tensor("op_1141"), val = tensor([12, 4, 64])]; + tensor var_1142 = reshape(shape = var_1141, x = k_11)[name = tensor("op_1142")]; tensor k_13_perm_0 = const()[name = tensor("k_13_perm_0"), val = tensor([1, 0, 2])]; - tensor var_1145 = const()[name = tensor("op_1145"), val = tensor([12, 4, 64])]; - tensor var_1146 = reshape(shape = var_1145, x = v_11)[name = tensor("op_1146")]; + tensor var_1148 = const()[name = tensor("op_1148"), val = tensor([12, 4, 64])]; + tensor var_1149 = reshape(shape = var_1148, x = v_11)[name = tensor("op_1149")]; tensor v_13_perm_0 = const()[name = tensor("v_13_perm_0"), val = tensor([1, 0, 2])]; - tensor var_1149 = const()[name = tensor("op_1149"), val = tensor([1, 4, 12, 64])]; - tensor q_13 = transpose(perm = q_13_perm_0, x = var_1132)[name = tensor("transpose_20")]; - tensor q_15 = reshape(shape = var_1149, x = q_13)[name = tensor("q_15")]; - tensor var_1151 = const()[name = tensor("op_1151"), val = tensor([1, 4, 12, 64])]; - tensor k_13 = transpose(perm = k_13_perm_0, x = var_1139)[name = tensor("transpose_19")]; - tensor k_15 = reshape(shape = var_1151, x = k_13)[name = tensor("k_15")]; - tensor var_1153 = const()[name = tensor("op_1153"), val = tensor([1, 4, 12, 64])]; - tensor v_13 = transpose(perm = v_13_perm_0, x = var_1146)[name = tensor("transpose_18")]; - tensor v_15 = reshape(shape = var_1153, x = v_13)[name = tensor("v_15")]; + tensor var_1152 = const()[name = tensor("op_1152"), val = tensor([1, 4, 12, 64])]; + tensor q_13 = transpose(perm = q_13_perm_0, x = var_1135)[name = tensor("transpose_20")]; + tensor q_15 = reshape(shape = var_1152, x = q_13)[name = tensor("q_15")]; + tensor var_1154 = const()[name = tensor("op_1154"), val = tensor([1, 4, 12, 64])]; + tensor k_13 = transpose(perm = k_13_perm_0, x = var_1142)[name = tensor("transpose_19")]; + tensor k_15 = reshape(shape = var_1154, x = k_13)[name = tensor("k_15")]; + tensor var_1156 = const()[name = tensor("op_1156"), val = tensor([1, 4, 12, 64])]; + tensor v_13 = transpose(perm = v_13_perm_0, x = var_1149)[name = tensor("transpose_18")]; + tensor v_15 = reshape(shape = var_1156, x = v_13)[name = tensor("v_15")]; tensor mul_1_y_0 = const()[name = tensor("mul_1_y_0"), val = tensor(0x1p-3)]; tensor mul_1 = mul(x = q_15, y = mul_1_y_0)[name = tensor("mul_1")]; tensor matmul_0_transpose_y_0 = const()[name = tensor("matmul_0_transpose_y_0"), val = tensor(true)]; @@ -981,30 +980,30 @@ program(1.0) tensor attn_output_1_transpose_x_0 = const()[name = tensor("attn_output_1_transpose_x_0"), val = tensor(false)]; tensor attn_output_1_transpose_y_0 = const()[name = tensor("attn_output_1_transpose_y_0"), val = tensor(false)]; tensor attn_output_1 = matmul(transpose_x = attn_output_1_transpose_x_0, transpose_y = attn_output_1_transpose_y_0, x = softmax_0, y = v_15)[name = tensor("attn_output_1")]; - tensor var_1156 = const()[name = tensor("op_1156"), val = tensor([2, 0, 1, 3])]; - tensor var_1161 = const()[name = tensor("op_1161"), val = tensor([12, 256])]; - tensor var_1157 = transpose(perm = var_1156, x = attn_output_1)[name = tensor("transpose_17")]; - tensor attn_output_3 = reshape(shape = var_1161, x = var_1157)[name = tensor("attn_output_3")]; - tensor attn_output_5 = linear(bias = decoder_self_attn2_0_out_proj_bias, weight = decoder_self_attn2_0_out_proj_weight, x = attn_output_3)[name = tensor("linear_44")]; - tensor var_1165 = const()[name = tensor("op_1165"), val = tensor([12, 1, 256])]; - tensor attn_output_7 = reshape(shape = var_1165, x = attn_output_5)[name = tensor("attn_output_7")]; + tensor var_1159 = const()[name = tensor("op_1159"), val = tensor([2, 0, 1, 3])]; + tensor var_1164 = const()[name = tensor("op_1164"), val = tensor([12, 256])]; + tensor var_1160 = transpose(perm = var_1159, x = attn_output_1)[name = tensor("transpose_17")]; + tensor attn_output_3 = reshape(shape = var_1164, x = var_1160)[name = tensor("attn_output_3")]; + tensor attn_output_5 = linear(bias = inner_decoder_self_attn2_0_out_proj_bias, weight = inner_decoder_self_attn2_0_out_proj_weight, x = attn_output_3)[name = tensor("linear_44")]; + tensor var_1168 = const()[name = tensor("op_1168"), val = tensor([12, 1, 256])]; + tensor attn_output_7 = reshape(shape = var_1168, x = attn_output_5)[name = tensor("attn_output_7")]; tensor sa2_out_1_perm_0 = const()[name = tensor("sa2_out_1_perm_0"), val = tensor([1, 0, 2])]; tensor sa2_out_1 = transpose(perm = sa2_out_1_perm_0, x = attn_output_7)[name = tensor("transpose_16")]; - tensor input_175 = add(x = query_1, y = sa2_out_1)[name = tensor("input_175")]; - tensor input_177_axes_0 = const()[name = tensor("input_177_axes_0"), val = tensor([-1])]; - tensor input_177 = layer_norm(axes = input_177_axes_0, beta = decoder_norm21_0_bias, epsilon = var_921, gamma = decoder_norm21_0_weight, x = input_175)[name = tensor("input_177")]; - tensor input_179 = linear(bias = decoder_linear1_0_bias, weight = decoder_linear1_0_weight, x = input_177)[name = tensor("linear_45")]; - tensor input_181 = relu(x = input_179)[name = tensor("input_181")]; - tensor ffn_out_1 = linear(bias = decoder_linear2_0_bias, weight = decoder_linear2_0_weight, x = input_181)[name = tensor("linear_46")]; - tensor input_183 = add(x = input_177, y = ffn_out_1)[name = tensor("input_183")]; + tensor input_177 = add(x = query_1, y = sa2_out_1)[name = tensor("input_177")]; + tensor input_179_axes_0 = const()[name = tensor("input_179_axes_0"), val = tensor([-1])]; + tensor input_179 = layer_norm(axes = input_179_axes_0, beta = inner_decoder_norm21_0_bias, epsilon = var_36, gamma = inner_decoder_norm21_0_weight, x = input_177)[name = tensor("input_179")]; + tensor input_181 = linear(bias = inner_decoder_linear1_0_bias, weight = inner_decoder_linear1_0_weight, x = input_179)[name = tensor("linear_45")]; + tensor input_183 = relu(x = input_181)[name = tensor("input_183")]; + tensor ffn_out_1 = linear(bias = inner_decoder_linear2_0_bias, weight = inner_decoder_linear2_0_weight, x = input_183)[name = tensor("linear_46")]; + tensor input_185 = add(x = input_179, y = ffn_out_1)[name = tensor("input_185")]; tensor xt_3_axes_0 = const()[name = tensor("xt_3_axes_0"), val = tensor([-1])]; - tensor xt_3 = layer_norm(axes = xt_3_axes_0, beta = decoder_norm22_0_bias, epsilon = var_921, gamma = decoder_norm22_0_weight, x = input_183)[name = tensor("xt_3")]; - tensor var_1185 = const()[name = tensor("op_1185"), val = tensor([1, 1, 12, 256])]; - tensor x_31 = reshape(shape = var_1185, x = xt_3)[name = tensor("x_31")]; - tensor var_1187_perm_0 = const()[name = tensor("op_1187_perm_0"), val = tensor([0, 2, 1, 3])]; - tensor var_1191 = const()[name = tensor("op_1191"), val = tensor([12, 1, 256])]; - tensor var_1187 = transpose(perm = var_1187_perm_0, x = x_31)[name = tensor("transpose_15")]; - tensor x = reshape(shape = var_1191, x = var_1187)[name = tensor("x")]; + tensor xt_3 = layer_norm(axes = xt_3_axes_0, beta = inner_decoder_norm22_0_bias, epsilon = var_36, gamma = inner_decoder_norm22_0_weight, x = input_185)[name = tensor("xt_3")]; + tensor var_1188 = const()[name = tensor("op_1188"), val = tensor([1, 1, 12, 256])]; + tensor x_31 = reshape(shape = var_1188, x = xt_3)[name = tensor("x_31")]; + tensor var_1190_perm_0 = const()[name = tensor("op_1190_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_1194 = const()[name = tensor("op_1194"), val = tensor([12, 1, 256])]; + tensor var_1190 = transpose(perm = var_1190_perm_0, x = x_31)[name = tensor("transpose_15")]; + tensor x = reshape(shape = var_1194, x = var_1190)[name = tensor("x")]; tensor prev_kv_begin_0 = const()[name = tensor("prev_kv_begin_0"), val = tensor([1, 0, 0, 0, 0])]; tensor prev_kv_end_0 = const()[name = tensor("prev_kv_end_0"), val = tensor([2, 12, 4, 64, 64])]; tensor prev_kv_end_mask_0 = const()[name = tensor("prev_kv_end_mask_0"), val = tensor([false, true, true, true, true])]; @@ -1015,120 +1014,120 @@ program(1.0) tensor prev_scale_end_mask_0 = const()[name = tensor("prev_scale_end_mask_0"), val = tensor([false, true])]; tensor prev_scale_squeeze_mask_0 = const()[name = tensor("prev_scale_squeeze_mask_0"), val = tensor([true, false])]; tensor prev_scale = slice_by_index(begin = prev_scale_begin_0, end = prev_scale_end_0, end_mask = prev_scale_end_mask_0, squeeze_mask = prev_scale_squeeze_mask_0, x = dec_scale)[name = tensor("prev_scale")]; - tensor var_1199 = linear(bias = decoder_q_proj_1_bias, weight = decoder_q_proj_1_weight, x = x)[name = tensor("linear_47")]; - tensor var_1200 = const()[name = tensor("op_1200"), val = tensor([12, 1, 4, 64])]; - tensor var_1201 = reshape(shape = var_1200, x = var_1199)[name = tensor("op_1201")]; + tensor var_1202 = linear(bias = inner_decoder_q_proj_1_bias, weight = inner_decoder_q_proj_1_weight, x = x)[name = tensor("linear_47")]; + tensor var_1203 = const()[name = tensor("op_1203"), val = tensor([12, 1, 4, 64])]; + tensor var_1204 = reshape(shape = var_1203, x = var_1202)[name = tensor("op_1204")]; tensor q_17_perm_0 = const()[name = tensor("q_17_perm_0"), val = tensor([0, 2, 1, 3])]; - tensor var_1205 = linear(bias = decoder_k_proj_1_bias, weight = decoder_k_proj_1_weight, x = x)[name = tensor("linear_48")]; - tensor var_1206 = const()[name = tensor("op_1206"), val = tensor(0x1p-3)]; - tensor var_1207 = mul(x = var_1205, y = var_1206)[name = tensor("op_1207")]; - tensor var_1208 = const()[name = tensor("op_1208"), val = tensor([12, 1, 4, 64])]; - tensor var_1209 = reshape(shape = var_1208, x = var_1207)[name = tensor("op_1209")]; + tensor var_1208 = linear(bias = inner_decoder_k_proj_1_bias, weight = inner_decoder_k_proj_1_weight, x = x)[name = tensor("linear_48")]; + tensor var_1209 = const()[name = tensor("op_1209"), val = tensor(0x1p-3)]; + tensor var_1210 = mul(x = var_1208, y = var_1209)[name = tensor("op_1210")]; + tensor var_1211 = const()[name = tensor("op_1211"), val = tensor([12, 1, 4, 64])]; + tensor var_1212 = reshape(shape = var_1211, x = var_1210)[name = tensor("op_1212")]; tensor k_17_perm_0 = const()[name = tensor("k_17_perm_0"), val = tensor([0, 2, 1, 3])]; - tensor var_1213 = linear(bias = decoder_v_proj_1_bias, weight = decoder_v_proj_1_weight, x = x)[name = tensor("linear_49")]; - tensor var_1214 = const()[name = tensor("op_1214"), val = tensor([12, 1, 4, 64])]; - tensor var_1215 = reshape(shape = var_1214, x = var_1213)[name = tensor("op_1215")]; + tensor var_1216 = linear(bias = inner_decoder_v_proj_1_bias, weight = inner_decoder_v_proj_1_weight, x = x)[name = tensor("linear_49")]; + tensor var_1217 = const()[name = tensor("op_1217"), val = tensor([12, 1, 4, 64])]; + tensor var_1218 = reshape(shape = var_1217, x = var_1216)[name = tensor("op_1218")]; tensor v_17_perm_0 = const()[name = tensor("v_17_perm_0"), val = tensor([0, 2, 1, 3])]; - tensor input_187 = linear(bias = decoder_g_proj_1_bias, weight = decoder_g_proj_1_weight, x = x)[name = tensor("linear_50")]; + tensor input_189 = linear(bias = inner_decoder_g_proj_1_bias, weight = inner_decoder_g_proj_1_weight, x = x)[name = tensor("linear_50")]; tensor sqrt_s0 = sqrt(x = prev_scale)[name = tensor("sqrt_s0")]; tensor s_t = add(x = prev_scale, y = cumsum_mask_1)[name = tensor("s_t")]; tensor const_32 = const()[name = tensor("const_32"), val = tensor(0x1.fffffep+127)]; - tensor clip_3 = clip(alpha = var_926, beta = const_32, x = s_t)[name = tensor("clip_3")]; + tensor clip_3 = clip(alpha = var_29, beta = const_32, x = s_t)[name = tensor("clip_3")]; tensor sqrt_s_t = sqrt(x = clip_3)[name = tensor("sqrt_s_t")]; tensor qk_transpose_x_1 = const()[name = tensor("qk_transpose_x_1"), val = tensor(false)]; tensor qk_transpose_y_1 = const()[name = tensor("qk_transpose_y_1"), val = tensor(true)]; - tensor k_17 = transpose(perm = k_17_perm_0, x = var_1209)[name = tensor("transpose_13")]; - tensor q_17 = transpose(perm = q_17_perm_0, x = var_1201)[name = tensor("transpose_14")]; + tensor k_17 = transpose(perm = k_17_perm_0, x = var_1212)[name = tensor("transpose_13")]; + tensor q_17 = transpose(perm = q_17_perm_0, x = var_1204)[name = tensor("transpose_14")]; tensor qk = matmul(transpose_x = qk_transpose_x_1, transpose_y = qk_transpose_y_1, x = q_17, y = k_17)[name = tensor("qk")]; - tensor var_1230 = const()[name = tensor("op_1230"), val = tensor([1, 1])]; - tensor var_1231 = reshape(shape = var_1230, x = sqrt_s_t)[name = tensor("op_1231")]; - tensor M = real_div(x = var_1047, y = var_1231)[name = tensor("M")]; - tensor var_1233 = mul(x = qk, y = M)[name = tensor("op_1233")]; - tensor inner_transpose_x_0 = const()[name = tensor("inner_transpose_x_0"), val = tensor(false)]; - tensor inner_transpose_y_0 = const()[name = tensor("inner_transpose_y_0"), val = tensor(false)]; - tensor v_17 = transpose(perm = v_17_perm_0, x = var_1215)[name = tensor("transpose_12")]; - tensor inner = matmul(transpose_x = inner_transpose_x_0, transpose_y = inner_transpose_y_0, x = var_1233, y = v_17)[name = tensor("inner")]; - tensor var_1235_transpose_x_0 = const()[name = tensor("op_1235_transpose_x_0"), val = tensor(false)]; - tensor var_1235_transpose_y_0 = const()[name = tensor("op_1235_transpose_y_0"), val = tensor(false)]; - tensor var_1235 = matmul(transpose_x = var_1235_transpose_x_0, transpose_y = var_1235_transpose_y_0, x = q_17, y = prev_kv)[name = tensor("op_1235")]; - tensor var_1236 = real_div(x = sqrt_s0, y = sqrt_s_t)[name = tensor("op_1236")]; - tensor var_1237 = const()[name = tensor("op_1237"), val = tensor([1, 1, 1, 1])]; - tensor var_1238 = reshape(shape = var_1237, x = var_1236)[name = tensor("op_1238")]; - tensor cross = mul(x = var_1235, y = var_1238)[name = tensor("cross")]; - tensor out_31 = add(x = inner, y = cross)[name = tensor("out_31")]; - tensor v_masked = mul(x = v_17, y = var_1061)[name = tensor("v_masked")]; - tensor var_1244 = mul(x = prev_kv, y = sqrt_s0)[name = tensor("op_1244")]; - tensor var_1246_transpose_x_1 = const()[name = tensor("op_1246_transpose_x_1"), val = tensor(true)]; - tensor var_1246_transpose_y_1 = const()[name = tensor("op_1246_transpose_y_1"), val = tensor(false)]; - tensor var_1246 = matmul(transpose_x = var_1246_transpose_x_1, transpose_y = var_1246_transpose_y_1, x = k_17, y = v_masked)[name = tensor("op_1246")]; - tensor new_kv_unnorm = add(x = var_1244, y = var_1246)[name = tensor("new_kv_unnorm")]; - tensor new_scale = add(x = prev_scale, y = var_1069)[name = tensor("new_scale")]; + tensor var_1233 = const()[name = tensor("op_1233"), val = tensor([1, 1])]; + tensor var_1234 = reshape(shape = var_1233, x = sqrt_s_t)[name = tensor("op_1234")]; + tensor M = real_div(x = var_1050, y = var_1234)[name = tensor("M")]; + tensor var_1236 = mul(x = qk, y = M)[name = tensor("op_1236")]; + tensor inner_11_transpose_x_0 = const()[name = tensor("inner_11_transpose_x_0"), val = tensor(false)]; + tensor inner_11_transpose_y_0 = const()[name = tensor("inner_11_transpose_y_0"), val = tensor(false)]; + tensor v_17 = transpose(perm = v_17_perm_0, x = var_1218)[name = tensor("transpose_12")]; + tensor inner_11 = matmul(transpose_x = inner_11_transpose_x_0, transpose_y = inner_11_transpose_y_0, x = var_1236, y = v_17)[name = tensor("inner_11")]; + tensor var_1238_transpose_x_0 = const()[name = tensor("op_1238_transpose_x_0"), val = tensor(false)]; + tensor var_1238_transpose_y_0 = const()[name = tensor("op_1238_transpose_y_0"), val = tensor(false)]; + tensor var_1238 = matmul(transpose_x = var_1238_transpose_x_0, transpose_y = var_1238_transpose_y_0, x = q_17, y = prev_kv)[name = tensor("op_1238")]; + tensor var_1239 = real_div(x = sqrt_s0, y = sqrt_s_t)[name = tensor("op_1239")]; + tensor var_1240 = const()[name = tensor("op_1240"), val = tensor([1, 1, 1, 1])]; + tensor var_1241 = reshape(shape = var_1240, x = var_1239)[name = tensor("op_1241")]; + tensor cross = mul(x = var_1238, y = var_1241)[name = tensor("cross")]; + tensor out_31 = add(x = inner_11, y = cross)[name = tensor("out_31")]; + tensor v_masked = mul(x = v_17, y = var_1064)[name = tensor("v_masked")]; + tensor var_1247 = mul(x = prev_kv, y = sqrt_s0)[name = tensor("op_1247")]; + tensor var_1249_transpose_x_1 = const()[name = tensor("op_1249_transpose_x_1"), val = tensor(true)]; + tensor var_1249_transpose_y_1 = const()[name = tensor("op_1249_transpose_y_1"), val = tensor(false)]; + tensor var_1249 = matmul(transpose_x = var_1249_transpose_x_1, transpose_y = var_1249_transpose_y_1, x = k_17, y = v_masked)[name = tensor("op_1249")]; + tensor new_kv_unnorm = add(x = var_1247, y = var_1249)[name = tensor("new_kv_unnorm")]; + tensor new_scale = add(x = prev_scale, y = var_1072)[name = tensor("new_scale")]; tensor const_33 = const()[name = tensor("const_33"), val = tensor(0x1.fffffep+127)]; - tensor clip_4 = clip(alpha = var_926, beta = const_33, x = new_scale)[name = tensor("clip_4")]; + tensor clip_4 = clip(alpha = var_29, beta = const_33, x = new_scale)[name = tensor("clip_4")]; tensor sqrt_new_scale = sqrt(x = clip_4)[name = tensor("sqrt_new_scale")]; tensor nkv = real_div(x = new_kv_unnorm, y = sqrt_new_scale)[name = tensor("nkv")]; - tensor var_1255_perm_0 = const()[name = tensor("op_1255_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_1258_perm_0 = const()[name = tensor("op_1258_perm_0"), val = tensor([0, 2, 1, 3])]; tensor out_33_axes_0 = const()[name = tensor("out_33_axes_0"), val = tensor([-1])]; - tensor var_1255 = transpose(perm = var_1255_perm_0, x = out_31)[name = tensor("transpose_11")]; - tensor out_33 = layer_norm(axes = out_33_axes_0, epsilon = var_923, x = var_1255)[name = tensor("out_33")]; - tensor var_1259 = const()[name = tensor("op_1259"), val = tensor([12, 1, 256])]; - tensor out = reshape(shape = var_1259, x = out_33)[name = tensor("out")]; - tensor var_1261 = silu(x = input_187)[name = tensor("op_1261")]; - tensor input_189 = mul(x = var_1261, y = out)[name = tensor("input_189")]; - tensor ret_out = linear(bias = decoder_out_proj_1_bias, weight = decoder_out_proj_1_weight, x = input_189)[name = tensor("linear_51")]; - tensor input_191 = add(x = x, y = ret_out)[name = tensor("input_191")]; + tensor var_1258 = transpose(perm = var_1258_perm_0, x = out_31)[name = tensor("transpose_11")]; + tensor out_33 = layer_norm(axes = out_33_axes_0, epsilon = var_44, x = var_1258)[name = tensor("out_33")]; + tensor var_1262 = const()[name = tensor("op_1262"), val = tensor([12, 1, 256])]; + tensor out = reshape(shape = var_1262, x = out_33)[name = tensor("out")]; + tensor var_1264 = silu(x = input_189)[name = tensor("op_1264")]; + tensor input_191 = mul(x = var_1264, y = out)[name = tensor("input_191")]; + tensor ret_out = linear(bias = inner_decoder_out_proj_1_bias, weight = inner_decoder_out_proj_1_weight, x = input_191)[name = tensor("linear_51")]; + tensor input_193 = add(x = x, y = ret_out)[name = tensor("input_193")]; tensor xt_5_axes_0 = const()[name = tensor("xt_5_axes_0"), val = tensor([-1])]; - tensor xt_5 = layer_norm(axes = xt_5_axes_0, beta = decoder_norm11_1_bias, epsilon = var_921, gamma = decoder_norm11_1_weight, x = input_191)[name = tensor("xt_5")]; - tensor var_1271 = const()[name = tensor("op_1271"), val = tensor([1, 12, 1, 256])]; - tensor var_1272 = reshape(shape = var_1271, x = xt_5)[name = tensor("op_1272")]; - tensor var_1273_perm_0 = const()[name = tensor("op_1273_perm_0"), val = tensor([0, 2, 1, 3])]; - tensor var_1276 = const()[name = tensor("op_1276"), val = tensor([1, 12, 256])]; - tensor var_1273 = transpose(perm = var_1273_perm_0, x = var_1272)[name = tensor("transpose_10")]; - tensor query_5 = reshape(shape = var_1276, x = var_1273)[name = tensor("query_5")]; + tensor xt_5 = layer_norm(axes = xt_5_axes_0, beta = inner_decoder_norm11_1_bias, epsilon = var_36, gamma = inner_decoder_norm11_1_weight, x = input_193)[name = tensor("xt_5")]; + tensor var_1274 = const()[name = tensor("op_1274"), val = tensor([1, 12, 1, 256])]; + tensor var_1275 = reshape(shape = var_1274, x = xt_5)[name = tensor("op_1275")]; + tensor var_1276_perm_0 = const()[name = tensor("op_1276_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_1279 = const()[name = tensor("op_1279"), val = tensor([1, 12, 256])]; + tensor var_1276 = transpose(perm = var_1276_perm_0, x = var_1275)[name = tensor("transpose_10")]; + tensor query_5 = reshape(shape = var_1279, x = var_1276)[name = tensor("query_5")]; tensor query_perm_0 = const()[name = tensor("query_perm_0"), val = tensor([1, 0, 2])]; tensor query = transpose(perm = query_perm_0, x = query_5)[name = tensor("transpose_9")]; - tensor var_1299 = linear(bias = decoder_self_attn2_1_in_proj_bias, weight = decoder_self_attn2_1_in_proj_weight, x = query)[name = tensor("linear_52")]; + tensor var_1302 = linear(bias = inner_decoder_self_attn2_1_in_proj_bias, weight = inner_decoder_self_attn2_1_in_proj_weight, x = query)[name = tensor("linear_52")]; tensor concat_2 = const()[name = tensor("concat_2"), val = tensor([12, 1, 3, 256])]; - tensor var_1301 = reshape(shape = concat_2, x = var_1299)[name = tensor("op_1301")]; - tensor var_1302_axes_0 = const()[name = tensor("op_1302_axes_0"), val = tensor([0])]; - tensor var_1302 = expand_dims(axes = var_1302_axes_0, x = var_1301)[name = tensor("op_1302")]; - tensor var_1303_perm_0 = const()[name = tensor("op_1303_perm_0"), val = tensor([-2, 1, 2, 0, 4])]; - tensor var_1304_axes_0 = const()[name = tensor("op_1304_axes_0"), val = tensor([-2])]; - tensor var_1303 = transpose(perm = var_1303_perm_0, x = var_1302)[name = tensor("transpose_8")]; - tensor var_1304 = squeeze(axes = var_1304_axes_0, x = var_1303)[name = tensor("op_1304")]; + tensor var_1304 = reshape(shape = concat_2, x = var_1302)[name = tensor("op_1304")]; + tensor var_1305_axes_0 = const()[name = tensor("op_1305_axes_0"), val = tensor([0])]; + tensor var_1305 = expand_dims(axes = var_1305_axes_0, x = var_1304)[name = tensor("op_1305")]; + tensor var_1306_perm_0 = const()[name = tensor("op_1306_perm_0"), val = tensor([-2, 1, 2, 0, 4])]; + tensor var_1307_axes_0 = const()[name = tensor("op_1307_axes_0"), val = tensor([-2])]; + tensor var_1306 = transpose(perm = var_1306_perm_0, x = var_1305)[name = tensor("transpose_8")]; + tensor var_1307 = squeeze(axes = var_1307_axes_0, x = var_1306)[name = tensor("op_1307")]; tensor q_19_begin_0 = const()[name = tensor("q_19_begin_0"), val = tensor([0, 0, 0, 0])]; tensor q_19_end_0 = const()[name = tensor("q_19_end_0"), val = tensor([1, 12, 1, 256])]; tensor q_19_end_mask_0 = const()[name = tensor("q_19_end_mask_0"), val = tensor([false, true, true, true])]; tensor q_19_squeeze_mask_0 = const()[name = tensor("q_19_squeeze_mask_0"), val = tensor([true, false, false, false])]; - tensor q_19 = slice_by_index(begin = q_19_begin_0, end = q_19_end_0, end_mask = q_19_end_mask_0, squeeze_mask = q_19_squeeze_mask_0, x = var_1304)[name = tensor("q_19")]; + tensor q_19 = slice_by_index(begin = q_19_begin_0, end = q_19_end_0, end_mask = q_19_end_mask_0, squeeze_mask = q_19_squeeze_mask_0, x = var_1307)[name = tensor("q_19")]; tensor k_19_begin_0 = const()[name = tensor("k_19_begin_0"), val = tensor([1, 0, 0, 0])]; tensor k_19_end_0 = const()[name = tensor("k_19_end_0"), val = tensor([2, 12, 1, 256])]; tensor k_19_end_mask_0 = const()[name = tensor("k_19_end_mask_0"), val = tensor([false, true, true, true])]; tensor k_19_squeeze_mask_0 = const()[name = tensor("k_19_squeeze_mask_0"), val = tensor([true, false, false, false])]; - tensor k_19 = slice_by_index(begin = k_19_begin_0, end = k_19_end_0, end_mask = k_19_end_mask_0, squeeze_mask = k_19_squeeze_mask_0, x = var_1304)[name = tensor("k_19")]; + tensor k_19 = slice_by_index(begin = k_19_begin_0, end = k_19_end_0, end_mask = k_19_end_mask_0, squeeze_mask = k_19_squeeze_mask_0, x = var_1307)[name = tensor("k_19")]; tensor v_19_begin_0 = const()[name = tensor("v_19_begin_0"), val = tensor([2, 0, 0, 0])]; tensor v_19_end_0 = const()[name = tensor("v_19_end_0"), val = tensor([3, 12, 1, 256])]; tensor v_19_end_mask_0 = const()[name = tensor("v_19_end_mask_0"), val = tensor([false, true, true, true])]; tensor v_19_squeeze_mask_0 = const()[name = tensor("v_19_squeeze_mask_0"), val = tensor([true, false, false, false])]; - tensor v_19 = slice_by_index(begin = v_19_begin_0, end = v_19_end_0, end_mask = v_19_end_mask_0, squeeze_mask = v_19_squeeze_mask_0, x = var_1304)[name = tensor("v_19")]; - tensor var_1312 = const()[name = tensor("op_1312"), val = tensor([12, 4, 64])]; - tensor var_1313 = reshape(shape = var_1312, x = q_19)[name = tensor("op_1313")]; + tensor v_19 = slice_by_index(begin = v_19_begin_0, end = v_19_end_0, end_mask = v_19_end_mask_0, squeeze_mask = v_19_squeeze_mask_0, x = var_1307)[name = tensor("v_19")]; + tensor var_1315 = const()[name = tensor("op_1315"), val = tensor([12, 4, 64])]; + tensor var_1316 = reshape(shape = var_1315, x = q_19)[name = tensor("op_1316")]; tensor q_21_perm_0 = const()[name = tensor("q_21_perm_0"), val = tensor([1, 0, 2])]; - tensor var_1319 = const()[name = tensor("op_1319"), val = tensor([12, 4, 64])]; - tensor var_1320 = reshape(shape = var_1319, x = k_19)[name = tensor("op_1320")]; + tensor var_1322 = const()[name = tensor("op_1322"), val = tensor([12, 4, 64])]; + tensor var_1323 = reshape(shape = var_1322, x = k_19)[name = tensor("op_1323")]; tensor k_21_perm_0 = const()[name = tensor("k_21_perm_0"), val = tensor([1, 0, 2])]; - tensor var_1326 = const()[name = tensor("op_1326"), val = tensor([12, 4, 64])]; - tensor var_1327 = reshape(shape = var_1326, x = v_19)[name = tensor("op_1327")]; + tensor var_1329 = const()[name = tensor("op_1329"), val = tensor([12, 4, 64])]; + tensor var_1330 = reshape(shape = var_1329, x = v_19)[name = tensor("op_1330")]; tensor v_21_perm_0 = const()[name = tensor("v_21_perm_0"), val = tensor([1, 0, 2])]; - tensor var_1330 = const()[name = tensor("op_1330"), val = tensor([1, 4, 12, 64])]; - tensor q_21 = transpose(perm = q_21_perm_0, x = var_1313)[name = tensor("transpose_7")]; - tensor q = reshape(shape = var_1330, x = q_21)[name = tensor("q")]; - tensor var_1332 = const()[name = tensor("op_1332"), val = tensor([1, 4, 12, 64])]; - tensor k_21 = transpose(perm = k_21_perm_0, x = var_1320)[name = tensor("transpose_6")]; - tensor k = reshape(shape = var_1332, x = k_21)[name = tensor("k")]; - tensor var_1334 = const()[name = tensor("op_1334"), val = tensor([1, 4, 12, 64])]; - tensor v_21 = transpose(perm = v_21_perm_0, x = var_1327)[name = tensor("transpose_5")]; - tensor v = reshape(shape = var_1334, x = v_21)[name = tensor("v")]; + tensor var_1333 = const()[name = tensor("op_1333"), val = tensor([1, 4, 12, 64])]; + tensor q_21 = transpose(perm = q_21_perm_0, x = var_1316)[name = tensor("transpose_7")]; + tensor q = reshape(shape = var_1333, x = q_21)[name = tensor("q")]; + tensor var_1335 = const()[name = tensor("op_1335"), val = tensor([1, 4, 12, 64])]; + tensor k_21 = transpose(perm = k_21_perm_0, x = var_1323)[name = tensor("transpose_6")]; + tensor k = reshape(shape = var_1335, x = k_21)[name = tensor("k")]; + tensor var_1337 = const()[name = tensor("op_1337"), val = tensor([1, 4, 12, 64])]; + tensor v_21 = transpose(perm = v_21_perm_0, x = var_1330)[name = tensor("transpose_5")]; + tensor v = reshape(shape = var_1337, x = v_21)[name = tensor("v")]; tensor mul_3_y_0 = const()[name = tensor("mul_3_y_0"), val = tensor(0x1p-3)]; tensor mul_3 = mul(x = q, y = mul_3_y_0)[name = tensor("mul_3")]; tensor matmul_1_transpose_y_0 = const()[name = tensor("matmul_1_transpose_y_0"), val = tensor(true)]; @@ -1139,34 +1138,34 @@ program(1.0) tensor attn_output_9_transpose_x_0 = const()[name = tensor("attn_output_9_transpose_x_0"), val = tensor(false)]; tensor attn_output_9_transpose_y_0 = const()[name = tensor("attn_output_9_transpose_y_0"), val = tensor(false)]; tensor attn_output_9 = matmul(transpose_x = attn_output_9_transpose_x_0, transpose_y = attn_output_9_transpose_y_0, x = softmax_1, y = v)[name = tensor("attn_output_9")]; - tensor var_1337 = const()[name = tensor("op_1337"), val = tensor([2, 0, 1, 3])]; - tensor var_1342 = const()[name = tensor("op_1342"), val = tensor([12, 256])]; - tensor var_1338 = transpose(perm = var_1337, x = attn_output_9)[name = tensor("transpose_4")]; - tensor attn_output_11 = reshape(shape = var_1342, x = var_1338)[name = tensor("attn_output_11")]; - tensor attn_output_13 = linear(bias = decoder_self_attn2_1_out_proj_bias, weight = decoder_self_attn2_1_out_proj_weight, x = attn_output_11)[name = tensor("linear_53")]; - tensor var_1346 = const()[name = tensor("op_1346"), val = tensor([12, 1, 256])]; - tensor attn_output = reshape(shape = var_1346, x = attn_output_13)[name = tensor("attn_output")]; + tensor var_1340 = const()[name = tensor("op_1340"), val = tensor([2, 0, 1, 3])]; + tensor var_1345 = const()[name = tensor("op_1345"), val = tensor([12, 256])]; + tensor var_1341 = transpose(perm = var_1340, x = attn_output_9)[name = tensor("transpose_4")]; + tensor attn_output_11 = reshape(shape = var_1345, x = var_1341)[name = tensor("attn_output_11")]; + tensor attn_output_13 = linear(bias = inner_decoder_self_attn2_1_out_proj_bias, weight = inner_decoder_self_attn2_1_out_proj_weight, x = attn_output_11)[name = tensor("linear_53")]; + tensor var_1349 = const()[name = tensor("op_1349"), val = tensor([12, 1, 256])]; + tensor attn_output = reshape(shape = var_1349, x = attn_output_13)[name = tensor("attn_output")]; tensor sa2_out_perm_0 = const()[name = tensor("sa2_out_perm_0"), val = tensor([1, 0, 2])]; tensor sa2_out = transpose(perm = sa2_out_perm_0, x = attn_output)[name = tensor("transpose_3")]; - tensor input_193 = add(x = query_5, y = sa2_out)[name = tensor("input_193")]; - tensor input_195_axes_0 = const()[name = tensor("input_195_axes_0"), val = tensor([-1])]; - tensor input_195 = layer_norm(axes = input_195_axes_0, beta = decoder_norm21_1_bias, epsilon = var_921, gamma = decoder_norm21_1_weight, x = input_193)[name = tensor("input_195")]; - tensor input_197 = linear(bias = decoder_linear1_1_bias, weight = decoder_linear1_1_weight, x = input_195)[name = tensor("linear_54")]; - tensor input_199 = relu(x = input_197)[name = tensor("input_199")]; - tensor ffn_out = linear(bias = decoder_linear2_1_bias, weight = decoder_linear2_1_weight, x = input_199)[name = tensor("linear_55")]; - tensor input_201 = add(x = input_195, y = ffn_out)[name = tensor("input_201")]; + tensor input_195 = add(x = query_5, y = sa2_out)[name = tensor("input_195")]; + tensor input_197_axes_0 = const()[name = tensor("input_197_axes_0"), val = tensor([-1])]; + tensor input_197 = layer_norm(axes = input_197_axes_0, beta = inner_decoder_norm21_1_bias, epsilon = var_36, gamma = inner_decoder_norm21_1_weight, x = input_195)[name = tensor("input_197")]; + tensor input_199 = linear(bias = inner_decoder_linear1_1_bias, weight = inner_decoder_linear1_1_weight, x = input_197)[name = tensor("linear_54")]; + tensor input_201 = relu(x = input_199)[name = tensor("input_201")]; + tensor ffn_out = linear(bias = inner_decoder_linear2_1_bias, weight = inner_decoder_linear2_1_weight, x = input_201)[name = tensor("linear_55")]; + tensor input_203 = add(x = input_197, y = ffn_out)[name = tensor("input_203")]; tensor xt_axes_0 = const()[name = tensor("xt_axes_0"), val = tensor([-1])]; - tensor xt = layer_norm(axes = xt_axes_0, beta = decoder_norm22_1_bias, epsilon = var_921, gamma = decoder_norm22_1_weight, x = input_201)[name = tensor("xt")]; - tensor var_1366 = const()[name = tensor("op_1366"), val = tensor([1, 1, 12, 256])]; - tensor input = reshape(shape = var_1366, x = xt)[name = tensor("input")]; - tensor var_1368 = const()[name = tensor("op_1368"), val = tensor([-1])]; - tensor var_1369 = reduce_l2_norm(axes = var_1368, keep_dims = var_924, x = input)[name = tensor("op_1369")]; + tensor xt = layer_norm(axes = xt_axes_0, beta = inner_decoder_norm22_1_bias, epsilon = var_36, gamma = inner_decoder_norm22_1_weight, x = input_203)[name = tensor("xt")]; + tensor var_1369 = const()[name = tensor("op_1369"), val = tensor([1, 1, 12, 256])]; + tensor input = reshape(shape = var_1369, x = xt)[name = tensor("input")]; + tensor var_1371 = const()[name = tensor("op_1371"), val = tensor([-1])]; + tensor var_1372 = reduce_l2_norm(axes = var_1371, keep_dims = var_35, x = input)[name = tensor("op_1372")]; tensor const_42 = const()[name = tensor("const_42"), val = tensor(0x1.fffffep+127)]; - tensor clip_5 = clip(alpha = var_916, beta = const_42, x = var_1369)[name = tensor("clip_5")]; - tensor var_1371 = real_div(x = input, y = clip_5)[name = tensor("op_1371")]; + tensor clip_5 = clip(alpha = var_49, beta = const_42, x = var_1372)[name = tensor("clip_5")]; + tensor var_1374 = real_div(x = input, y = clip_5)[name = tensor("op_1374")]; tensor transpose_1_perm_0 = const()[name = tensor("transpose_1_perm_0"), val = tensor([0, 1, 3, 2])]; tensor concat_7 = const()[name = tensor("concat_7"), val = tensor([1, 256, 12])]; - tensor transpose_1 = transpose(perm = transpose_1_perm_0, x = var_1371)[name = tensor("transpose_2")]; + tensor transpose_1 = transpose(perm = transpose_1_perm_0, x = var_1374)[name = tensor("transpose_2")]; tensor reshape_1 = reshape(shape = concat_7, x = transpose_1)[name = tensor("reshape_1")]; tensor matmul_2_transpose_x_0 = const()[name = tensor("matmul_2_transpose_x_0"), val = tensor(false)]; tensor matmul_2_transpose_y_0 = const()[name = tensor("matmul_2_transpose_y_0"), val = tensor(false)]; @@ -1175,10 +1174,10 @@ program(1.0) tensor output_end_0 = const()[name = tensor("output_end_0"), val = tensor([1, 1, 11])]; tensor output_end_mask_0 = const()[name = tensor("output_end_mask_0"), val = tensor([true, true, false])]; tensor output = slice_by_index(begin = output_begin_0, end = output_end_0, end_mask = output_end_mask_0, x = matmul_2)[name = tensor("output")]; - tensor probs = sigmoid(x = output)[name = tensor("op_1375")]; - tensor var_1377_axis_0 = const()[name = tensor("op_1377_axis_0"), val = tensor(0)]; - tensor dec_kv_new = stack(axis = var_1377_axis_0, values = (var_1073, nkv))[name = tensor("op_1377")]; - tensor var_1379_axis_0 = const()[name = tensor("op_1379_axis_0"), val = tensor(0)]; - tensor dec_scale_new = stack(axis = var_1379_axis_0, values = (new_scale_9, new_scale))[name = tensor("op_1379")]; + tensor probs = sigmoid(x = output)[name = tensor("op_1378")]; + tensor var_1380_axis_0 = const()[name = tensor("op_1380_axis_0"), val = tensor(0)]; + tensor dec_kv_new = stack(axis = var_1380_axis_0, values = (var_1076, nkv))[name = tensor("op_1380")]; + tensor var_1382_axis_0 = const()[name = tensor("op_1382_axis_0"), val = tensor(0)]; + tensor dec_scale_new = stack(axis = var_1382_axis_0, values = (new_scale_9, new_scale))[name = tensor("op_1382")]; } -> (probs, enc_kv_new, enc_scale_new, enc_conv_cache_new, cnn_window_new, dec_kv_new, dec_scale_new); } \ No newline at end of file diff --git a/optimized/dih3/100ms/ls_eend_dih3_100ms.mlpackage/Data/com.apple.CoreML/model.mlmodel b/optimized/dih3/100ms/ls_eend_dih3_100ms.mlpackage/Data/com.apple.CoreML/model.mlmodel index 172bd36e29739117a908278b3525a67e992d8b17..4d0f7dab493892e96c7f5365a99d744d609ae14c 100644 --- a/optimized/dih3/100ms/ls_eend_dih3_100ms.mlpackage/Data/com.apple.CoreML/model.mlmodel +++ b/optimized/dih3/100ms/ls_eend_dih3_100ms.mlpackage/Data/com.apple.CoreML/model.mlmodel @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:866e7d0226752a637b1394b3ae9ecf1dc1de0ca1073f1b164eb38c8dfe62ea4a -size 171366 +oid sha256:cedf01fcea7289c508f0174575bc7f7cb6a932676c48bf6d17e6e8761e0e7ce1 +size 175284 diff --git a/optimized/dih3/100ms/ls_eend_dih3_100ms.mlpackage/Manifest.json b/optimized/dih3/100ms/ls_eend_dih3_100ms.mlpackage/Manifest.json index 5e96fabcd00f0d81e3d7405b594d3c65e54a58a8..597f60c85753c109df7183ebf5cc7eba6fc00946 100644 --- a/optimized/dih3/100ms/ls_eend_dih3_100ms.mlpackage/Manifest.json +++ b/optimized/dih3/100ms/ls_eend_dih3_100ms.mlpackage/Manifest.json @@ -1,18 +1,18 @@ { "fileFormatVersion": "1.0.0", "itemInfoEntries": { - "80A6D93A-77B4-40EA-A15C-5545A78FEEBD": { + "887AF76B-5BF7-465B-AF63-A5E7FF000DC6": { "author": "com.apple.CoreML", "description": "CoreML Model Specification", "name": "model.mlmodel", "path": "com.apple.CoreML/model.mlmodel" }, - "9DDE2145-36B0-46F7-8DFC-385831DC4F71": { + "8D81431E-8F42-4587-93CD-FB6D145BD632": { "author": "com.apple.CoreML", "description": "CoreML Model Weights", "name": "weights", "path": "com.apple.CoreML/weights" } }, - "rootModelIdentifier": "80A6D93A-77B4-40EA-A15C-5545A78FEEBD" + "rootModelIdentifier": "887AF76B-5BF7-465B-AF63-A5E7FF000DC6" } diff --git a/optimized/dih3/200ms/ls_eend_dih3_200ms.mlmodelc/analytics/coremldata.bin b/optimized/dih3/200ms/ls_eend_dih3_200ms.mlmodelc/analytics/coremldata.bin index 5409bd7627c8339408f2ba44c89764b14943a6db..5275287c92c06d59c0f454d9396d645febb5490e 100644 --- a/optimized/dih3/200ms/ls_eend_dih3_200ms.mlmodelc/analytics/coremldata.bin +++ b/optimized/dih3/200ms/ls_eend_dih3_200ms.mlmodelc/analytics/coremldata.bin @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:a8893bce28227e8d51b3751b9be3a44e3ec395d16b2515ff8db197af2749801d +oid sha256:5b551f8181ccd8ed3acd3ea62db76df66be4bc1acee718b152fedb8e2052cf1f size 243 diff --git a/optimized/dih3/200ms/ls_eend_dih3_200ms.mlmodelc/coremldata.bin b/optimized/dih3/200ms/ls_eend_dih3_200ms.mlmodelc/coremldata.bin index 1ab9910e40d04ee5bb87f2f64a0f79e96a63d69c..b604e797d0d7645d87b600080421e751416b3e47 100644 --- a/optimized/dih3/200ms/ls_eend_dih3_200ms.mlmodelc/coremldata.bin +++ b/optimized/dih3/200ms/ls_eend_dih3_200ms.mlmodelc/coremldata.bin @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:714cac00c71f584a18c944c841c3236d25c7bded9052f77aa3c110e1c0373194 -size 1310 +oid sha256:a9a6da6dd9e7274ccc549595c38658e3f8a05098b3ff212278de6e0cac6f157e +size 1413 diff --git a/optimized/dih3/200ms/ls_eend_dih3_200ms.mlmodelc/metadata.json b/optimized/dih3/200ms/ls_eend_dih3_200ms.mlmodelc/metadata.json index aaab8ad1bf6a40dc63cb5e669c82cca3eee7f301..0a646b8aab9afdfcede56f504c4d672650956b97 100644 --- a/optimized/dih3/200ms/ls_eend_dih3_200ms.mlmodelc/metadata.json +++ b/optimized/dih3/200ms/ls_eend_dih3_200ms.mlmodelc/metadata.json @@ -1,7 +1,7 @@ [ { "metadataOutputVersion" : "3.0", - "shortDescription" : "LS-EEND DIHARD III streaming diarizer (pipeline, T=2, max_speakers=10)", + "shortDescription" : "LS-EEND DIHARD III streaming diarizer (pipeline, T=2, max_speakers=10, layout=raw_mel, cu=all)", "outputSchema" : [ { "hasShapeFlexibility" : "0", @@ -81,7 +81,7 @@ "specificationVersion" : 8, "mlProgramOperationTypeHistogram" : { "Ios16.reduceL2Norm" : 2, - "Ios17.reshape" : 68, + "Ios17.reshape" : 69, "Ios16.softmax" : 2, "Ios17.matmul" : 29, "Ios17.transpose" : 57, @@ -89,7 +89,7 @@ "Ios17.expandDims" : 3, "Ios17.add" : 46, "Ios16.sigmoid" : 5, - "Ios17.sliceByIndex" : 48, + "Ios17.sliceByIndex" : 50, "Tile" : 1, "Ios16.reduceSum" : 1, "Ios17.squeeze" : 2, @@ -101,7 +101,7 @@ "Ios16.silu" : 18, "Ios17.realDiv" : 20, "Ios17.linear" : 56, - "Stack" : 5, + "Stack" : 6, "Ios17.concat" : 14, "Ios16.relu" : 2, "Ios16.cumsum" : 1, @@ -128,9 +128,9 @@ "hasShapeFlexibility" : "0", "isOptional" : "0", "dataType" : "Float32", - "formattedType" : "MultiArray (Float32 1 × 2 × 345)", + "formattedType" : "MultiArray (Float32 1 × 25 × 23)", "shortDescription" : "", - "shape" : "[1, 2, 345]", + "shape" : "[1, 25, 23]", "name" : "features", "type" : "MultiArray" }, @@ -206,8 +206,8 @@ } ], "userDefinedMetadata" : { - "com.github.apple.coremltools.conversion_date" : "2026-04-16", - "config" : "{\"model_name\": \"dih3\", \"model_label\": \"DIHARD III\", \"variant\": \"pipeline\", \"chunk_size\": 2, \"step_duration_ms\": 200, \"frame_duration_ms\": 100, \"frame_duration_seconds\": 0.1, \"max_speakers\": 10, \"max_nspks\": 12, \"n_units\": 256, \"n_heads\": 4, \"enc_n_layers\": 4, \"dec_n_layers\": 2, \"conv_kernel_size\": 16, \"conv_delay\": 9, \"sample_rate\": 8000, \"win_length\": 200, \"hop_length\": 80, \"n_mels\": 23, \"context_size\": 7, \"subsampling\": 10, \"feat_type\": \"logmel23_cummn\", \"pure_roll\": true}", + "com.github.apple.coremltools.conversion_date" : "2026-04-18", + "config" : "{\"model_name\": \"dih3\", \"model_label\": \"DIHARD III\", \"variant\": \"pipeline\", \"chunk_size\": 2, \"step_duration_ms\": 200, \"frame_duration_ms\": 100, \"frame_duration_seconds\": 0.1, \"max_speakers\": 10, \"max_nspks\": 12, \"n_units\": 256, \"n_heads\": 4, \"enc_n_layers\": 4, \"dec_n_layers\": 2, \"conv_kernel_size\": 16, \"conv_delay\": 9, \"sample_rate\": 8000, \"win_length\": 200, \"hop_length\": 80, \"n_mels\": 23, \"context_size\": 7, \"subsampling\": 10, \"feat_type\": \"logmel23_cummn\", \"pure_roll\": true, \"input_layout\": \"raw_mel\", \"compute_units_export\": \"all\", \"raw_mel_length\": 25}", "com.github.apple.coremltools.source" : "torch==2.6.0", "com.github.apple.coremltools.version" : "9.0", "com.github.apple.coremltools.source_dialect" : "TorchScript" diff --git a/optimized/dih3/200ms/ls_eend_dih3_200ms.mlmodelc/model.mil b/optimized/dih3/200ms/ls_eend_dih3_200ms.mlmodelc/model.mil index 5651eaf71d9adaeffd8f5bfe34de4f9950314977..2ccb23e6a99a5f049025ab2b67fb75bd7b7827a2 100644 --- a/optimized/dih3/200ms/ls_eend_dih3_200ms.mlmodelc/model.mil +++ b/optimized/dih3/200ms/ls_eend_dih3_200ms.mlmodelc/model.mil @@ -1,234 +1,248 @@ program(1.0) [buildInfo = dict, tensor>({{"coremlc-component-MIL", "3510.2.1"}, {"coremlc-version", "3500.32.1"}, {"coremltools-component-torch", "2.6.0"}, {"coremltools-source-dialect", "TorchScript"}, {"coremltools-version", "9.0"}})] { - func main(tensor cnn_window, tensor dec_kv, tensor dec_scale, tensor enc_conv_cache, tensor enc_kv, tensor enc_scale, tensor features, tensor valid_mask) { - tensor encoder_cnn_bias = const()[name = tensor("encoder_cnn_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(64)))]; - tensor encoder_cnn_weight = const()[name = tensor("encoder_cnn_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1152)))]; - tensor encoder__causal_mask = const()[name = tensor("encoder__causal_mask"), val = tensor([[0x1p+0, 0x0p+0], [0x1p+0, 0x1p+0]])]; - tensor encoder__t_index = const()[name = tensor("encoder__t_index"), val = tensor([0x1p+0, 0x1p+1])]; - tensor encoder_input_projection_linear_bias = const()[name = tensor("encoder_input_projection_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4981952)))]; - tensor encoder_input_projection_linear_weight = const()[name = tensor("encoder_input_projection_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4983040)))]; - tensor encoder_pre_layer_norm_bias = const()[name = tensor("encoder_pre_layer_norm_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(5336384)))]; - tensor encoder_pre_layer_norm_weight = const()[name = tensor("encoder_pre_layer_norm_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(5337472)))]; - tensor encoder_ffn1_0_module_sequential_0_bias = const()[name = tensor("encoder_ffn1_0_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(5338560)))]; - tensor encoder_ffn1_0_module_sequential_0_weight = const()[name = tensor("encoder_ffn1_0_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(5339648)))]; - tensor encoder_ffn1_0_module_sequential_1_linear_bias = const()[name = tensor("encoder_ffn1_0_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(5340736)))]; - tensor encoder_ffn1_0_module_sequential_1_linear_weight = const()[name = tensor("encoder_ffn1_0_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(5344896)))]; - tensor encoder_ffn1_0_module_sequential_4_linear_bias = const()[name = tensor("encoder_ffn1_0_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(6393536)))]; - tensor encoder_ffn1_0_module_sequential_4_linear_weight = const()[name = tensor("encoder_ffn1_0_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(6394624)))]; - tensor encoder_ret_lns_0_bias = const()[name = tensor("encoder_ret_lns_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7443264)))]; - tensor encoder_ret_lns_0_weight = const()[name = tensor("encoder_ret_lns_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7444352)))]; - tensor encoder_q_proj_0_bias = const()[name = tensor("encoder_q_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7445440)))]; - tensor encoder_q_proj_0_weight = const()[name = tensor("encoder_q_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7446528)))]; - tensor encoder_k_proj_0_bias = const()[name = tensor("encoder_k_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7708736)))]; - tensor encoder_k_proj_0_weight = const()[name = tensor("encoder_k_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7709824)))]; - tensor encoder_v_proj_0_bias = const()[name = tensor("encoder_v_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7972032)))]; - tensor encoder_v_proj_0_weight = const()[name = tensor("encoder_v_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7973120)))]; - tensor encoder_g_proj_0_bias = const()[name = tensor("encoder_g_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8235328)))]; - tensor encoder_g_proj_0_weight = const()[name = tensor("encoder_g_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8236416)))]; - tensor encoder_out_proj_0_bias = const()[name = tensor("encoder_out_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8498624)))]; - tensor encoder_out_proj_0_weight = const()[name = tensor("encoder_out_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8499712)))]; - tensor encoder_conv_module_0_sequential_0_bias = const()[name = tensor("encoder_conv_module_0_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8761920)))]; - tensor encoder_conv_module_0_sequential_0_weight = const()[name = tensor("encoder_conv_module_0_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8763008)))]; - tensor encoder_conv_module_0_sequential_2_conv_bias = const()[name = tensor("encoder_conv_module_0_sequential_2_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8764096)))]; - tensor encoder_conv_module_0_sequential_2_conv_weight = const()[name = tensor("encoder_conv_module_0_sequential_2_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8766208)))]; - tensor encoder_conv_module_0_sequential_4_conv_weight = const()[name = tensor("encoder_conv_module_0_sequential_4_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9290560)))]; - tensor encoder_conv_module_0_sequential_5_running_var = const()[name = tensor("encoder_conv_module_0_sequential_5_running_var"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9307008)))]; - tensor encoder_conv_module_0_sequential_5_running_mean = const()[name = tensor("encoder_conv_module_0_sequential_5_running_mean"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9308096)))]; - tensor encoder_conv_module_0_sequential_5_bias = const()[name = tensor("encoder_conv_module_0_sequential_5_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9309184)))]; - tensor encoder_conv_module_0_sequential_5_weight = const()[name = tensor("encoder_conv_module_0_sequential_5_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9310272)))]; - tensor encoder_conv_module_0_sequential_7_conv_bias = const()[name = tensor("encoder_conv_module_0_sequential_7_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9311360)))]; - tensor encoder_conv_module_0_sequential_7_conv_weight = const()[name = tensor("encoder_conv_module_0_sequential_7_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9312448)))]; - tensor encoder_ffn2_0_module_sequential_0_bias = const()[name = tensor("encoder_ffn2_0_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9574656)))]; - tensor encoder_ffn2_0_module_sequential_0_weight = const()[name = tensor("encoder_ffn2_0_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9575744)))]; - tensor encoder_ffn2_0_module_sequential_1_linear_bias = const()[name = tensor("encoder_ffn2_0_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9576832)))]; - tensor encoder_ffn2_0_module_sequential_1_linear_weight = const()[name = tensor("encoder_ffn2_0_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9580992)))]; - tensor encoder_ffn2_0_module_sequential_4_linear_bias = const()[name = tensor("encoder_ffn2_0_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(10629632)))]; - tensor encoder_ffn2_0_module_sequential_4_linear_weight = const()[name = tensor("encoder_ffn2_0_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(10630720)))]; - tensor encoder_layer_norm_0_bias = const()[name = tensor("encoder_layer_norm_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11679360)))]; - tensor encoder_layer_norm_0_weight = const()[name = tensor("encoder_layer_norm_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11680448)))]; - tensor encoder_ffn1_1_module_sequential_0_bias = const()[name = tensor("encoder_ffn1_1_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11681536)))]; - tensor encoder_ffn1_1_module_sequential_0_weight = const()[name = tensor("encoder_ffn1_1_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11682624)))]; - tensor encoder_ffn1_1_module_sequential_1_linear_bias = const()[name = tensor("encoder_ffn1_1_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11683712)))]; - tensor encoder_ffn1_1_module_sequential_1_linear_weight = const()[name = tensor("encoder_ffn1_1_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11687872)))]; - tensor encoder_ffn1_1_module_sequential_4_linear_bias = const()[name = tensor("encoder_ffn1_1_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(12736512)))]; - tensor encoder_ffn1_1_module_sequential_4_linear_weight = const()[name = tensor("encoder_ffn1_1_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(12737600)))]; - tensor encoder_ret_lns_1_bias = const()[name = tensor("encoder_ret_lns_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(13786240)))]; - tensor encoder_ret_lns_1_weight = const()[name = tensor("encoder_ret_lns_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(13787328)))]; - tensor encoder_q_proj_1_bias = const()[name = tensor("encoder_q_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(13788416)))]; - tensor encoder_q_proj_1_weight = const()[name = tensor("encoder_q_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(13789504)))]; - tensor encoder_k_proj_1_bias = const()[name = tensor("encoder_k_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14051712)))]; - tensor encoder_k_proj_1_weight = const()[name = tensor("encoder_k_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14052800)))]; - tensor encoder_v_proj_1_bias = const()[name = tensor("encoder_v_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14315008)))]; - tensor encoder_v_proj_1_weight = const()[name = tensor("encoder_v_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14316096)))]; - tensor encoder_g_proj_1_bias = const()[name = tensor("encoder_g_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14578304)))]; - tensor encoder_g_proj_1_weight = const()[name = tensor("encoder_g_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14579392)))]; - tensor encoder_out_proj_1_bias = const()[name = tensor("encoder_out_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14841600)))]; - tensor encoder_out_proj_1_weight = const()[name = tensor("encoder_out_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14842688)))]; - tensor encoder_conv_module_1_sequential_0_bias = const()[name = tensor("encoder_conv_module_1_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15104896)))]; - tensor encoder_conv_module_1_sequential_0_weight = const()[name = tensor("encoder_conv_module_1_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15105984)))]; - tensor encoder_conv_module_1_sequential_2_conv_bias = const()[name = tensor("encoder_conv_module_1_sequential_2_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15107072)))]; - tensor encoder_conv_module_1_sequential_2_conv_weight = const()[name = tensor("encoder_conv_module_1_sequential_2_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15109184)))]; - tensor encoder_conv_module_1_sequential_4_conv_weight = const()[name = tensor("encoder_conv_module_1_sequential_4_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15633536)))]; - tensor encoder_conv_module_1_sequential_5_running_var = const()[name = tensor("encoder_conv_module_1_sequential_5_running_var"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15649984)))]; - tensor encoder_conv_module_1_sequential_5_running_mean = const()[name = tensor("encoder_conv_module_1_sequential_5_running_mean"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15651072)))]; - tensor encoder_conv_module_1_sequential_5_bias = const()[name = tensor("encoder_conv_module_1_sequential_5_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15652160)))]; - tensor encoder_conv_module_1_sequential_5_weight = const()[name = tensor("encoder_conv_module_1_sequential_5_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15653248)))]; - tensor encoder_conv_module_1_sequential_7_conv_bias = const()[name = tensor("encoder_conv_module_1_sequential_7_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15654336)))]; - tensor encoder_conv_module_1_sequential_7_conv_weight = const()[name = tensor("encoder_conv_module_1_sequential_7_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15655424)))]; - tensor encoder_ffn2_1_module_sequential_0_bias = const()[name = tensor("encoder_ffn2_1_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15917632)))]; - tensor encoder_ffn2_1_module_sequential_0_weight = const()[name = tensor("encoder_ffn2_1_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15918720)))]; - tensor encoder_ffn2_1_module_sequential_1_linear_bias = const()[name = tensor("encoder_ffn2_1_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15919808)))]; - tensor encoder_ffn2_1_module_sequential_1_linear_weight = const()[name = tensor("encoder_ffn2_1_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15923968)))]; - tensor encoder_ffn2_1_module_sequential_4_linear_bias = const()[name = tensor("encoder_ffn2_1_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(16972608)))]; - tensor encoder_ffn2_1_module_sequential_4_linear_weight = const()[name = tensor("encoder_ffn2_1_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(16973696)))]; - tensor encoder_layer_norm_1_bias = const()[name = tensor("encoder_layer_norm_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18022336)))]; - tensor encoder_layer_norm_1_weight = const()[name = tensor("encoder_layer_norm_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18023424)))]; - tensor encoder_ffn1_2_module_sequential_0_bias = const()[name = tensor("encoder_ffn1_2_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18024512)))]; - tensor encoder_ffn1_2_module_sequential_0_weight = const()[name = tensor("encoder_ffn1_2_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18025600)))]; - tensor encoder_ffn1_2_module_sequential_1_linear_bias = const()[name = tensor("encoder_ffn1_2_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18026688)))]; - tensor encoder_ffn1_2_module_sequential_1_linear_weight = const()[name = tensor("encoder_ffn1_2_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18030848)))]; - tensor encoder_ffn1_2_module_sequential_4_linear_bias = const()[name = tensor("encoder_ffn1_2_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(19079488)))]; - tensor encoder_ffn1_2_module_sequential_4_linear_weight = const()[name = tensor("encoder_ffn1_2_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(19080576)))]; - tensor encoder_ret_lns_2_bias = const()[name = tensor("encoder_ret_lns_2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20129216)))]; - tensor encoder_ret_lns_2_weight = const()[name = tensor("encoder_ret_lns_2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20130304)))]; - tensor encoder_q_proj_2_bias = const()[name = tensor("encoder_q_proj_2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20131392)))]; - tensor encoder_q_proj_2_weight = const()[name = tensor("encoder_q_proj_2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20132480)))]; - tensor encoder_k_proj_2_bias = const()[name = tensor("encoder_k_proj_2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20394688)))]; - tensor encoder_k_proj_2_weight = const()[name = tensor("encoder_k_proj_2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20395776)))]; - tensor encoder_v_proj_2_bias = const()[name = tensor("encoder_v_proj_2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20657984)))]; - tensor encoder_v_proj_2_weight = const()[name = tensor("encoder_v_proj_2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20659072)))]; - tensor encoder_g_proj_2_bias = const()[name = tensor("encoder_g_proj_2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20921280)))]; - tensor encoder_g_proj_2_weight = const()[name = tensor("encoder_g_proj_2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20922368)))]; - tensor encoder_out_proj_2_bias = const()[name = tensor("encoder_out_proj_2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21184576)))]; - tensor encoder_out_proj_2_weight = const()[name = tensor("encoder_out_proj_2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21185664)))]; - tensor encoder_conv_module_2_sequential_0_bias = const()[name = tensor("encoder_conv_module_2_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21447872)))]; - tensor encoder_conv_module_2_sequential_0_weight = const()[name = tensor("encoder_conv_module_2_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21448960)))]; - tensor encoder_conv_module_2_sequential_2_conv_bias = const()[name = tensor("encoder_conv_module_2_sequential_2_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21450048)))]; - tensor encoder_conv_module_2_sequential_2_conv_weight = const()[name = tensor("encoder_conv_module_2_sequential_2_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21452160)))]; - tensor encoder_conv_module_2_sequential_4_conv_weight = const()[name = tensor("encoder_conv_module_2_sequential_4_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21976512)))]; - tensor encoder_conv_module_2_sequential_5_running_var = const()[name = tensor("encoder_conv_module_2_sequential_5_running_var"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21992960)))]; - tensor encoder_conv_module_2_sequential_5_running_mean = const()[name = tensor("encoder_conv_module_2_sequential_5_running_mean"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21994048)))]; - tensor encoder_conv_module_2_sequential_5_bias = const()[name = tensor("encoder_conv_module_2_sequential_5_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21995136)))]; - tensor encoder_conv_module_2_sequential_5_weight = const()[name = tensor("encoder_conv_module_2_sequential_5_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21996224)))]; - tensor encoder_conv_module_2_sequential_7_conv_bias = const()[name = tensor("encoder_conv_module_2_sequential_7_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21997312)))]; - tensor encoder_conv_module_2_sequential_7_conv_weight = const()[name = tensor("encoder_conv_module_2_sequential_7_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21998400)))]; - tensor encoder_ffn2_2_module_sequential_0_bias = const()[name = tensor("encoder_ffn2_2_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(22260608)))]; - tensor encoder_ffn2_2_module_sequential_0_weight = const()[name = tensor("encoder_ffn2_2_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(22261696)))]; - tensor encoder_ffn2_2_module_sequential_1_linear_bias = const()[name = tensor("encoder_ffn2_2_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(22262784)))]; - tensor encoder_ffn2_2_module_sequential_1_linear_weight = const()[name = tensor("encoder_ffn2_2_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(22266944)))]; - tensor encoder_ffn2_2_module_sequential_4_linear_bias = const()[name = tensor("encoder_ffn2_2_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(23315584)))]; - tensor encoder_ffn2_2_module_sequential_4_linear_weight = const()[name = tensor("encoder_ffn2_2_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(23316672)))]; - tensor encoder_layer_norm_2_bias = const()[name = tensor("encoder_layer_norm_2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(24365312)))]; - tensor encoder_layer_norm_2_weight = const()[name = tensor("encoder_layer_norm_2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(24366400)))]; - tensor encoder_ffn1_3_module_sequential_0_bias = const()[name = tensor("encoder_ffn1_3_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(24367488)))]; - tensor encoder_ffn1_3_module_sequential_0_weight = const()[name = tensor("encoder_ffn1_3_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(24368576)))]; - tensor encoder_ffn1_3_module_sequential_1_linear_bias = const()[name = tensor("encoder_ffn1_3_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(24369664)))]; - tensor encoder_ffn1_3_module_sequential_1_linear_weight = const()[name = tensor("encoder_ffn1_3_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(24373824)))]; - tensor encoder_ffn1_3_module_sequential_4_linear_bias = const()[name = tensor("encoder_ffn1_3_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(25422464)))]; - tensor encoder_ffn1_3_module_sequential_4_linear_weight = const()[name = tensor("encoder_ffn1_3_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(25423552)))]; - tensor encoder_ret_lns_3_bias = const()[name = tensor("encoder_ret_lns_3_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(26472192)))]; - tensor encoder_ret_lns_3_weight = const()[name = tensor("encoder_ret_lns_3_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(26473280)))]; - tensor encoder_q_proj_3_bias = const()[name = tensor("encoder_q_proj_3_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(26474368)))]; - tensor encoder_q_proj_3_weight = const()[name = tensor("encoder_q_proj_3_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(26475456)))]; - tensor encoder_k_proj_3_bias = const()[name = tensor("encoder_k_proj_3_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(26737664)))]; - tensor encoder_k_proj_3_weight = const()[name = tensor("encoder_k_proj_3_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(26738752)))]; - tensor encoder_v_proj_3_bias = const()[name = tensor("encoder_v_proj_3_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27000960)))]; - tensor encoder_v_proj_3_weight = const()[name = tensor("encoder_v_proj_3_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27002048)))]; - tensor encoder_g_proj_3_bias = const()[name = tensor("encoder_g_proj_3_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27264256)))]; - tensor encoder_g_proj_3_weight = const()[name = tensor("encoder_g_proj_3_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27265344)))]; - tensor encoder_out_proj_3_bias = const()[name = tensor("encoder_out_proj_3_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27527552)))]; - tensor encoder_out_proj_3_weight = const()[name = tensor("encoder_out_proj_3_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27528640)))]; - tensor encoder_conv_module_3_sequential_0_bias = const()[name = tensor("encoder_conv_module_3_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27790848)))]; - tensor encoder_conv_module_3_sequential_0_weight = const()[name = tensor("encoder_conv_module_3_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27791936)))]; - tensor encoder_conv_module_3_sequential_2_conv_bias = const()[name = tensor("encoder_conv_module_3_sequential_2_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27793024)))]; - tensor encoder_conv_module_3_sequential_2_conv_weight = const()[name = tensor("encoder_conv_module_3_sequential_2_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27795136)))]; - tensor encoder_conv_module_3_sequential_4_conv_weight = const()[name = tensor("encoder_conv_module_3_sequential_4_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28319488)))]; - tensor encoder_conv_module_3_sequential_5_running_var = const()[name = tensor("encoder_conv_module_3_sequential_5_running_var"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28335936)))]; - tensor encoder_conv_module_3_sequential_5_running_mean = const()[name = tensor("encoder_conv_module_3_sequential_5_running_mean"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28337024)))]; - tensor encoder_conv_module_3_sequential_5_bias = const()[name = tensor("encoder_conv_module_3_sequential_5_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28338112)))]; - tensor encoder_conv_module_3_sequential_5_weight = const()[name = tensor("encoder_conv_module_3_sequential_5_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28339200)))]; - tensor encoder_conv_module_3_sequential_7_conv_bias = const()[name = tensor("encoder_conv_module_3_sequential_7_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28340288)))]; - tensor encoder_conv_module_3_sequential_7_conv_weight = const()[name = tensor("encoder_conv_module_3_sequential_7_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28341376)))]; - tensor encoder_ffn2_3_module_sequential_0_bias = const()[name = tensor("encoder_ffn2_3_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28603584)))]; - tensor encoder_ffn2_3_module_sequential_0_weight = const()[name = tensor("encoder_ffn2_3_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28604672)))]; - tensor encoder_ffn2_3_module_sequential_1_linear_bias = const()[name = tensor("encoder_ffn2_3_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28605760)))]; - tensor encoder_ffn2_3_module_sequential_1_linear_weight = const()[name = tensor("encoder_ffn2_3_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28609920)))]; - tensor encoder_ffn2_3_module_sequential_4_linear_bias = const()[name = tensor("encoder_ffn2_3_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(29658560)))]; - tensor encoder_ffn2_3_module_sequential_4_linear_weight = const()[name = tensor("encoder_ffn2_3_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(29659648)))]; - tensor encoder_layer_norm_3_bias = const()[name = tensor("encoder_layer_norm_3_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(30708288)))]; - tensor encoder_layer_norm_3_weight = const()[name = tensor("encoder_layer_norm_3_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(30709376)))]; - tensor decoder__causal_mask = const()[name = tensor("decoder__causal_mask"), val = tensor([[0x1p+0, 0x0p+0], [0x1p+0, 0x1p+0]])]; - tensor decoder_convert_bias = const()[name = tensor("decoder_convert_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(30710464)))]; - tensor decoder_convert_weight = const()[name = tensor("decoder_convert_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(30711552)))]; - tensor decoder_q_proj_0_bias = const()[name = tensor("decoder_q_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(31235904)))]; - tensor decoder_q_proj_0_weight = const()[name = tensor("decoder_q_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(31236992)))]; - tensor decoder_k_proj_0_bias = const()[name = tensor("decoder_k_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(31499200)))]; - tensor decoder_k_proj_0_weight = const()[name = tensor("decoder_k_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(31500288)))]; - tensor decoder_v_proj_0_bias = const()[name = tensor("decoder_v_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(31762496)))]; - tensor decoder_v_proj_0_weight = const()[name = tensor("decoder_v_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(31763584)))]; - tensor decoder_g_proj_0_bias = const()[name = tensor("decoder_g_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32025792)))]; - tensor decoder_g_proj_0_weight = const()[name = tensor("decoder_g_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32026880)))]; - tensor decoder_out_proj_0_bias = const()[name = tensor("decoder_out_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32289088)))]; - tensor decoder_out_proj_0_weight = const()[name = tensor("decoder_out_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32290176)))]; - tensor decoder_norm11_0_bias = const()[name = tensor("decoder_norm11_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32552384)))]; - tensor decoder_norm11_0_weight = const()[name = tensor("decoder_norm11_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32553472)))]; - tensor decoder_self_attn2_0_out_proj_bias = const()[name = tensor("decoder_self_attn2_0_out_proj_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32554560)))]; - tensor decoder_self_attn2_0_out_proj_weight = const()[name = tensor("decoder_self_attn2_0_out_proj_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32555648)))]; - tensor decoder_self_attn2_0_in_proj_bias = const()[name = tensor("decoder_self_attn2_0_in_proj_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32817856)))]; - tensor decoder_self_attn2_0_in_proj_weight = const()[name = tensor("decoder_self_attn2_0_in_proj_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32820992)))]; - tensor decoder_norm21_0_bias = const()[name = tensor("decoder_norm21_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(33607488)))]; - tensor decoder_norm21_0_weight = const()[name = tensor("decoder_norm21_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(33608576)))]; - tensor decoder_linear1_0_bias = const()[name = tensor("decoder_linear1_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(33609664)))]; - tensor decoder_linear1_0_weight = const()[name = tensor("decoder_linear1_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(33617920)))]; - tensor decoder_linear2_0_bias = const()[name = tensor("decoder_linear2_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(35715136)))]; - tensor decoder_linear2_0_weight = const()[name = tensor("decoder_linear2_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(35716224)))]; - tensor decoder_norm22_0_bias = const()[name = tensor("decoder_norm22_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(37813440)))]; - tensor decoder_norm22_0_weight = const()[name = tensor("decoder_norm22_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(37814528)))]; - tensor decoder_q_proj_1_bias = const()[name = tensor("decoder_q_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(37815616)))]; - tensor decoder_q_proj_1_weight = const()[name = tensor("decoder_q_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(37816704)))]; - tensor decoder_k_proj_1_bias = const()[name = tensor("decoder_k_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38078912)))]; - tensor decoder_k_proj_1_weight = const()[name = tensor("decoder_k_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38080000)))]; - tensor decoder_v_proj_1_bias = const()[name = tensor("decoder_v_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38342208)))]; - tensor decoder_v_proj_1_weight = const()[name = tensor("decoder_v_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38343296)))]; - tensor decoder_g_proj_1_bias = const()[name = tensor("decoder_g_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38605504)))]; - tensor decoder_g_proj_1_weight = const()[name = tensor("decoder_g_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38606592)))]; - tensor decoder_out_proj_1_bias = const()[name = tensor("decoder_out_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38868800)))]; - tensor decoder_out_proj_1_weight = const()[name = tensor("decoder_out_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38869888)))]; - tensor decoder_norm11_1_bias = const()[name = tensor("decoder_norm11_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(39132096)))]; - tensor decoder_norm11_1_weight = const()[name = tensor("decoder_norm11_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(39133184)))]; - tensor decoder_self_attn2_1_out_proj_bias = const()[name = tensor("decoder_self_attn2_1_out_proj_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(39134272)))]; - tensor decoder_self_attn2_1_out_proj_weight = const()[name = tensor("decoder_self_attn2_1_out_proj_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(39135360)))]; - tensor decoder_self_attn2_1_in_proj_bias = const()[name = tensor("decoder_self_attn2_1_in_proj_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(39397568)))]; - tensor decoder_self_attn2_1_in_proj_weight = const()[name = tensor("decoder_self_attn2_1_in_proj_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(39400704)))]; - tensor decoder_norm21_1_bias = const()[name = tensor("decoder_norm21_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(40187200)))]; - tensor decoder_norm21_1_weight = const()[name = tensor("decoder_norm21_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(40188288)))]; - tensor decoder_linear1_1_bias = const()[name = tensor("decoder_linear1_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(40189376)))]; - tensor decoder_linear1_1_weight = const()[name = tensor("decoder_linear1_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(40197632)))]; - tensor decoder_linear2_1_bias = const()[name = tensor("decoder_linear2_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(42294848)))]; - tensor decoder_linear2_1_weight = const()[name = tensor("decoder_linear2_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(42295936)))]; - tensor decoder_norm22_1_bias = const()[name = tensor("decoder_norm22_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(44393152)))]; - tensor decoder_norm22_1_weight = const()[name = tensor("decoder_norm22_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(44394240)))]; - tensor var_11 = const()[name = tensor("op_11"), val = tensor(0x1.5798eep-27)]; - tensor var_17 = const()[name = tensor("op_17"), val = tensor(0x1.0c6f7ap-20)]; - tensor var_20 = const()[name = tensor("op_20"), val = tensor(2)]; - tensor var_23 = const()[name = tensor("op_23"), val = tensor(0)]; - tensor var_26 = const()[name = tensor("op_26"), val = tensor(1)]; - tensor var_28 = const()[name = tensor("op_28"), val = tensor(0x1.4f8b58p-17)]; - tensor var_29 = const()[name = tensor("op_29"), val = tensor(true)]; - tensor input_1 = linear(bias = encoder_input_projection_linear_bias, weight = encoder_input_projection_linear_weight, x = features)[name = tensor("linear_0")]; - tensor input_3_axes_0 = const()[name = tensor("input_3_axes_0"), val = tensor([-1])]; - tensor input_3 = layer_norm(axes = input_3_axes_0, beta = encoder_pre_layer_norm_bias, epsilon = var_28, gamma = encoder_pre_layer_norm_weight, x = input_1)[name = tensor("input_3")]; + func main(tensor cnn_window, tensor dec_kv, tensor dec_scale, tensor enc_conv_cache, tensor enc_kv, tensor enc_scale, tensor features, tensor valid_mask) { + tensor inner_encoder_cnn_bias = const()[name = tensor("inner_encoder_cnn_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(64)))]; + tensor inner_encoder_cnn_weight = const()[name = tensor("inner_encoder_cnn_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1152)))]; + tensor inner_encoder__causal_mask = const()[name = tensor("inner_encoder__causal_mask"), val = tensor([[0x1p+0, 0x0p+0], [0x1p+0, 0x1p+0]])]; + tensor inner_encoder__t_index = const()[name = tensor("inner_encoder__t_index"), val = tensor([0x1p+0, 0x1p+1])]; + tensor inner_encoder_input_projection_linear_bias = const()[name = tensor("inner_encoder_input_projection_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4981952)))]; + tensor inner_encoder_input_projection_linear_weight = const()[name = tensor("inner_encoder_input_projection_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4983040)))]; + tensor inner_encoder_pre_layer_norm_bias = const()[name = tensor("inner_encoder_pre_layer_norm_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(5336384)))]; + tensor inner_encoder_pre_layer_norm_weight = const()[name = tensor("inner_encoder_pre_layer_norm_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(5337472)))]; + tensor inner_encoder_ffn1_0_module_sequential_0_bias = const()[name = tensor("inner_encoder_ffn1_0_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(5338560)))]; + tensor inner_encoder_ffn1_0_module_sequential_0_weight = const()[name = tensor("inner_encoder_ffn1_0_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(5339648)))]; + tensor inner_encoder_ffn1_0_module_sequential_1_linear_bias = const()[name = tensor("inner_encoder_ffn1_0_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(5340736)))]; + tensor inner_encoder_ffn1_0_module_sequential_1_linear_weight = const()[name = tensor("inner_encoder_ffn1_0_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(5344896)))]; + tensor inner_encoder_ffn1_0_module_sequential_4_linear_bias = const()[name = tensor("inner_encoder_ffn1_0_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(6393536)))]; + tensor inner_encoder_ffn1_0_module_sequential_4_linear_weight = const()[name = tensor("inner_encoder_ffn1_0_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(6394624)))]; + tensor inner_encoder_ret_lns_0_bias = const()[name = tensor("inner_encoder_ret_lns_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7443264)))]; + tensor inner_encoder_ret_lns_0_weight = const()[name = tensor("inner_encoder_ret_lns_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7444352)))]; + tensor inner_encoder_q_proj_0_bias = const()[name = tensor("inner_encoder_q_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7445440)))]; + tensor inner_encoder_q_proj_0_weight = const()[name = tensor("inner_encoder_q_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7446528)))]; + tensor inner_encoder_k_proj_0_bias = const()[name = tensor("inner_encoder_k_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7708736)))]; + tensor inner_encoder_k_proj_0_weight = const()[name = tensor("inner_encoder_k_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7709824)))]; + tensor inner_encoder_v_proj_0_bias = const()[name = tensor("inner_encoder_v_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7972032)))]; + tensor inner_encoder_v_proj_0_weight = const()[name = tensor("inner_encoder_v_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7973120)))]; + tensor inner_encoder_g_proj_0_bias = const()[name = tensor("inner_encoder_g_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8235328)))]; + tensor inner_encoder_g_proj_0_weight = const()[name = tensor("inner_encoder_g_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8236416)))]; + tensor inner_encoder_out_proj_0_bias = const()[name = tensor("inner_encoder_out_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8498624)))]; + tensor inner_encoder_out_proj_0_weight = const()[name = tensor("inner_encoder_out_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8499712)))]; + tensor inner_encoder_conv_module_0_sequential_0_bias = const()[name = tensor("inner_encoder_conv_module_0_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8761920)))]; + tensor inner_encoder_conv_module_0_sequential_0_weight = const()[name = tensor("inner_encoder_conv_module_0_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8763008)))]; + tensor inner_encoder_conv_module_0_sequential_2_conv_bias = const()[name = tensor("inner_encoder_conv_module_0_sequential_2_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8764096)))]; + tensor inner_encoder_conv_module_0_sequential_2_conv_weight = const()[name = tensor("inner_encoder_conv_module_0_sequential_2_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8766208)))]; + tensor inner_encoder_conv_module_0_sequential_4_conv_weight = const()[name = tensor("inner_encoder_conv_module_0_sequential_4_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9290560)))]; + tensor inner_encoder_conv_module_0_sequential_5_running_var = const()[name = tensor("inner_encoder_conv_module_0_sequential_5_running_var"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9307008)))]; + tensor inner_encoder_conv_module_0_sequential_5_running_mean = const()[name = tensor("inner_encoder_conv_module_0_sequential_5_running_mean"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9308096)))]; + tensor inner_encoder_conv_module_0_sequential_5_bias = const()[name = tensor("inner_encoder_conv_module_0_sequential_5_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9309184)))]; + tensor inner_encoder_conv_module_0_sequential_5_weight = const()[name = tensor("inner_encoder_conv_module_0_sequential_5_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9310272)))]; + tensor inner_encoder_conv_module_0_sequential_7_conv_bias = const()[name = tensor("inner_encoder_conv_module_0_sequential_7_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9311360)))]; + tensor inner_encoder_conv_module_0_sequential_7_conv_weight = const()[name = tensor("inner_encoder_conv_module_0_sequential_7_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9312448)))]; + tensor inner_encoder_ffn2_0_module_sequential_0_bias = const()[name = tensor("inner_encoder_ffn2_0_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9574656)))]; + tensor inner_encoder_ffn2_0_module_sequential_0_weight = const()[name = tensor("inner_encoder_ffn2_0_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9575744)))]; + tensor inner_encoder_ffn2_0_module_sequential_1_linear_bias = const()[name = tensor("inner_encoder_ffn2_0_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9576832)))]; + tensor inner_encoder_ffn2_0_module_sequential_1_linear_weight = const()[name = tensor("inner_encoder_ffn2_0_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9580992)))]; + tensor inner_encoder_ffn2_0_module_sequential_4_linear_bias = const()[name = tensor("inner_encoder_ffn2_0_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(10629632)))]; + tensor inner_encoder_ffn2_0_module_sequential_4_linear_weight = const()[name = tensor("inner_encoder_ffn2_0_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(10630720)))]; + tensor inner_encoder_layer_norm_0_bias = const()[name = tensor("inner_encoder_layer_norm_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11679360)))]; + tensor inner_encoder_layer_norm_0_weight = const()[name = tensor("inner_encoder_layer_norm_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11680448)))]; + tensor inner_encoder_ffn1_1_module_sequential_0_bias = const()[name = tensor("inner_encoder_ffn1_1_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11681536)))]; + tensor inner_encoder_ffn1_1_module_sequential_0_weight = const()[name = tensor("inner_encoder_ffn1_1_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11682624)))]; + tensor inner_encoder_ffn1_1_module_sequential_1_linear_bias = const()[name = tensor("inner_encoder_ffn1_1_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11683712)))]; + tensor inner_encoder_ffn1_1_module_sequential_1_linear_weight = const()[name = tensor("inner_encoder_ffn1_1_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11687872)))]; + tensor inner_encoder_ffn1_1_module_sequential_4_linear_bias = const()[name = tensor("inner_encoder_ffn1_1_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(12736512)))]; + tensor inner_encoder_ffn1_1_module_sequential_4_linear_weight = const()[name = tensor("inner_encoder_ffn1_1_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(12737600)))]; + tensor inner_encoder_ret_lns_1_bias = const()[name = tensor("inner_encoder_ret_lns_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(13786240)))]; + tensor inner_encoder_ret_lns_1_weight = const()[name = tensor("inner_encoder_ret_lns_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(13787328)))]; + tensor inner_encoder_q_proj_1_bias = const()[name = tensor("inner_encoder_q_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(13788416)))]; + tensor inner_encoder_q_proj_1_weight = const()[name = tensor("inner_encoder_q_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(13789504)))]; + tensor inner_encoder_k_proj_1_bias = const()[name = tensor("inner_encoder_k_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14051712)))]; + tensor inner_encoder_k_proj_1_weight = const()[name = tensor("inner_encoder_k_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14052800)))]; + tensor inner_encoder_v_proj_1_bias = const()[name = tensor("inner_encoder_v_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14315008)))]; + tensor inner_encoder_v_proj_1_weight = const()[name = tensor("inner_encoder_v_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14316096)))]; + tensor inner_encoder_g_proj_1_bias = const()[name = tensor("inner_encoder_g_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14578304)))]; + tensor inner_encoder_g_proj_1_weight = const()[name = tensor("inner_encoder_g_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14579392)))]; + tensor inner_encoder_out_proj_1_bias = const()[name = tensor("inner_encoder_out_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14841600)))]; + tensor inner_encoder_out_proj_1_weight = const()[name = tensor("inner_encoder_out_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14842688)))]; + tensor inner_encoder_conv_module_1_sequential_0_bias = const()[name = tensor("inner_encoder_conv_module_1_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15104896)))]; + tensor inner_encoder_conv_module_1_sequential_0_weight = const()[name = tensor("inner_encoder_conv_module_1_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15105984)))]; + tensor inner_encoder_conv_module_1_sequential_2_conv_bias = const()[name = tensor("inner_encoder_conv_module_1_sequential_2_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15107072)))]; + tensor inner_encoder_conv_module_1_sequential_2_conv_weight = const()[name = tensor("inner_encoder_conv_module_1_sequential_2_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15109184)))]; + tensor inner_encoder_conv_module_1_sequential_4_conv_weight = const()[name = tensor("inner_encoder_conv_module_1_sequential_4_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15633536)))]; + tensor inner_encoder_conv_module_1_sequential_5_running_var = const()[name = tensor("inner_encoder_conv_module_1_sequential_5_running_var"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15649984)))]; + tensor inner_encoder_conv_module_1_sequential_5_running_mean = const()[name = tensor("inner_encoder_conv_module_1_sequential_5_running_mean"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15651072)))]; + tensor inner_encoder_conv_module_1_sequential_5_bias = const()[name = tensor("inner_encoder_conv_module_1_sequential_5_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15652160)))]; + tensor inner_encoder_conv_module_1_sequential_5_weight = const()[name = tensor("inner_encoder_conv_module_1_sequential_5_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15653248)))]; + tensor inner_encoder_conv_module_1_sequential_7_conv_bias = const()[name = tensor("inner_encoder_conv_module_1_sequential_7_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15654336)))]; + tensor inner_encoder_conv_module_1_sequential_7_conv_weight = const()[name = tensor("inner_encoder_conv_module_1_sequential_7_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15655424)))]; + tensor inner_encoder_ffn2_1_module_sequential_0_bias = const()[name = tensor("inner_encoder_ffn2_1_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15917632)))]; + tensor inner_encoder_ffn2_1_module_sequential_0_weight = const()[name = tensor("inner_encoder_ffn2_1_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15918720)))]; + tensor inner_encoder_ffn2_1_module_sequential_1_linear_bias = const()[name = tensor("inner_encoder_ffn2_1_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15919808)))]; + tensor inner_encoder_ffn2_1_module_sequential_1_linear_weight = const()[name = tensor("inner_encoder_ffn2_1_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15923968)))]; + tensor inner_encoder_ffn2_1_module_sequential_4_linear_bias = const()[name = tensor("inner_encoder_ffn2_1_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(16972608)))]; + tensor inner_encoder_ffn2_1_module_sequential_4_linear_weight = const()[name = tensor("inner_encoder_ffn2_1_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(16973696)))]; + tensor inner_encoder_layer_norm_1_bias = const()[name = tensor("inner_encoder_layer_norm_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18022336)))]; + tensor inner_encoder_layer_norm_1_weight = const()[name = tensor("inner_encoder_layer_norm_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18023424)))]; + tensor inner_encoder_ffn1_2_module_sequential_0_bias = const()[name = tensor("inner_encoder_ffn1_2_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18024512)))]; + tensor inner_encoder_ffn1_2_module_sequential_0_weight = const()[name = tensor("inner_encoder_ffn1_2_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18025600)))]; + tensor inner_encoder_ffn1_2_module_sequential_1_linear_bias = const()[name = tensor("inner_encoder_ffn1_2_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18026688)))]; + tensor inner_encoder_ffn1_2_module_sequential_1_linear_weight = const()[name = tensor("inner_encoder_ffn1_2_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18030848)))]; + tensor inner_encoder_ffn1_2_module_sequential_4_linear_bias = const()[name = tensor("inner_encoder_ffn1_2_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(19079488)))]; + tensor inner_encoder_ffn1_2_module_sequential_4_linear_weight = const()[name = tensor("inner_encoder_ffn1_2_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(19080576)))]; + tensor inner_encoder_ret_lns_2_bias = const()[name = tensor("inner_encoder_ret_lns_2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20129216)))]; + tensor inner_encoder_ret_lns_2_weight = const()[name = tensor("inner_encoder_ret_lns_2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20130304)))]; + tensor inner_encoder_q_proj_2_bias = const()[name = tensor("inner_encoder_q_proj_2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20131392)))]; + tensor inner_encoder_q_proj_2_weight = const()[name = tensor("inner_encoder_q_proj_2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20132480)))]; + tensor inner_encoder_k_proj_2_bias = const()[name = tensor("inner_encoder_k_proj_2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20394688)))]; + tensor inner_encoder_k_proj_2_weight = const()[name = tensor("inner_encoder_k_proj_2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20395776)))]; + tensor inner_encoder_v_proj_2_bias = const()[name = tensor("inner_encoder_v_proj_2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20657984)))]; + tensor inner_encoder_v_proj_2_weight = const()[name = tensor("inner_encoder_v_proj_2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20659072)))]; + tensor inner_encoder_g_proj_2_bias = const()[name = tensor("inner_encoder_g_proj_2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20921280)))]; + tensor inner_encoder_g_proj_2_weight = const()[name = tensor("inner_encoder_g_proj_2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20922368)))]; + tensor inner_encoder_out_proj_2_bias = const()[name = tensor("inner_encoder_out_proj_2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21184576)))]; + tensor inner_encoder_out_proj_2_weight = const()[name = tensor("inner_encoder_out_proj_2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21185664)))]; + tensor inner_encoder_conv_module_2_sequential_0_bias = const()[name = tensor("inner_encoder_conv_module_2_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21447872)))]; + tensor inner_encoder_conv_module_2_sequential_0_weight = const()[name = tensor("inner_encoder_conv_module_2_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21448960)))]; + tensor inner_encoder_conv_module_2_sequential_2_conv_bias = const()[name = tensor("inner_encoder_conv_module_2_sequential_2_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21450048)))]; + tensor inner_encoder_conv_module_2_sequential_2_conv_weight = const()[name = tensor("inner_encoder_conv_module_2_sequential_2_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21452160)))]; + tensor inner_encoder_conv_module_2_sequential_4_conv_weight = const()[name = tensor("inner_encoder_conv_module_2_sequential_4_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21976512)))]; + tensor inner_encoder_conv_module_2_sequential_5_running_var = const()[name = tensor("inner_encoder_conv_module_2_sequential_5_running_var"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21992960)))]; + tensor inner_encoder_conv_module_2_sequential_5_running_mean = const()[name = tensor("inner_encoder_conv_module_2_sequential_5_running_mean"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21994048)))]; + tensor inner_encoder_conv_module_2_sequential_5_bias = const()[name = tensor("inner_encoder_conv_module_2_sequential_5_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21995136)))]; + tensor inner_encoder_conv_module_2_sequential_5_weight = const()[name = tensor("inner_encoder_conv_module_2_sequential_5_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21996224)))]; + tensor inner_encoder_conv_module_2_sequential_7_conv_bias = const()[name = tensor("inner_encoder_conv_module_2_sequential_7_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21997312)))]; + tensor inner_encoder_conv_module_2_sequential_7_conv_weight = const()[name = tensor("inner_encoder_conv_module_2_sequential_7_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21998400)))]; + tensor inner_encoder_ffn2_2_module_sequential_0_bias = const()[name = tensor("inner_encoder_ffn2_2_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(22260608)))]; + tensor inner_encoder_ffn2_2_module_sequential_0_weight = const()[name = tensor("inner_encoder_ffn2_2_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(22261696)))]; + tensor inner_encoder_ffn2_2_module_sequential_1_linear_bias = const()[name = tensor("inner_encoder_ffn2_2_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(22262784)))]; + tensor inner_encoder_ffn2_2_module_sequential_1_linear_weight = const()[name = tensor("inner_encoder_ffn2_2_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(22266944)))]; + tensor inner_encoder_ffn2_2_module_sequential_4_linear_bias = const()[name = tensor("inner_encoder_ffn2_2_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(23315584)))]; + tensor inner_encoder_ffn2_2_module_sequential_4_linear_weight = const()[name = tensor("inner_encoder_ffn2_2_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(23316672)))]; + tensor inner_encoder_layer_norm_2_bias = const()[name = tensor("inner_encoder_layer_norm_2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(24365312)))]; + tensor inner_encoder_layer_norm_2_weight = const()[name = tensor("inner_encoder_layer_norm_2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(24366400)))]; + tensor inner_encoder_ffn1_3_module_sequential_0_bias = const()[name = tensor("inner_encoder_ffn1_3_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(24367488)))]; + tensor inner_encoder_ffn1_3_module_sequential_0_weight = const()[name = tensor("inner_encoder_ffn1_3_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(24368576)))]; + tensor inner_encoder_ffn1_3_module_sequential_1_linear_bias = const()[name = tensor("inner_encoder_ffn1_3_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(24369664)))]; + tensor inner_encoder_ffn1_3_module_sequential_1_linear_weight = const()[name = tensor("inner_encoder_ffn1_3_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(24373824)))]; + tensor inner_encoder_ffn1_3_module_sequential_4_linear_bias = const()[name = tensor("inner_encoder_ffn1_3_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(25422464)))]; + tensor inner_encoder_ffn1_3_module_sequential_4_linear_weight = const()[name = tensor("inner_encoder_ffn1_3_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(25423552)))]; + tensor inner_encoder_ret_lns_3_bias = const()[name = tensor("inner_encoder_ret_lns_3_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(26472192)))]; + tensor inner_encoder_ret_lns_3_weight = const()[name = tensor("inner_encoder_ret_lns_3_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(26473280)))]; + tensor inner_encoder_q_proj_3_bias = const()[name = tensor("inner_encoder_q_proj_3_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(26474368)))]; + tensor inner_encoder_q_proj_3_weight = const()[name = tensor("inner_encoder_q_proj_3_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(26475456)))]; + tensor inner_encoder_k_proj_3_bias = const()[name = tensor("inner_encoder_k_proj_3_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(26737664)))]; + tensor inner_encoder_k_proj_3_weight = const()[name = tensor("inner_encoder_k_proj_3_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(26738752)))]; + tensor inner_encoder_v_proj_3_bias = const()[name = tensor("inner_encoder_v_proj_3_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27000960)))]; + tensor inner_encoder_v_proj_3_weight = const()[name = tensor("inner_encoder_v_proj_3_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27002048)))]; + tensor inner_encoder_g_proj_3_bias = const()[name = tensor("inner_encoder_g_proj_3_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27264256)))]; + tensor inner_encoder_g_proj_3_weight = const()[name = tensor("inner_encoder_g_proj_3_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27265344)))]; + tensor inner_encoder_out_proj_3_bias = const()[name = tensor("inner_encoder_out_proj_3_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27527552)))]; + tensor inner_encoder_out_proj_3_weight = const()[name = tensor("inner_encoder_out_proj_3_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27528640)))]; + tensor inner_encoder_conv_module_3_sequential_0_bias = const()[name = tensor("inner_encoder_conv_module_3_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27790848)))]; + tensor inner_encoder_conv_module_3_sequential_0_weight = const()[name = tensor("inner_encoder_conv_module_3_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27791936)))]; + tensor inner_encoder_conv_module_3_sequential_2_conv_bias = const()[name = tensor("inner_encoder_conv_module_3_sequential_2_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27793024)))]; + tensor inner_encoder_conv_module_3_sequential_2_conv_weight = const()[name = tensor("inner_encoder_conv_module_3_sequential_2_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27795136)))]; + tensor inner_encoder_conv_module_3_sequential_4_conv_weight = const()[name = tensor("inner_encoder_conv_module_3_sequential_4_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28319488)))]; + tensor inner_encoder_conv_module_3_sequential_5_running_var = const()[name = tensor("inner_encoder_conv_module_3_sequential_5_running_var"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28335936)))]; + tensor inner_encoder_conv_module_3_sequential_5_running_mean = const()[name = tensor("inner_encoder_conv_module_3_sequential_5_running_mean"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28337024)))]; + tensor inner_encoder_conv_module_3_sequential_5_bias = const()[name = tensor("inner_encoder_conv_module_3_sequential_5_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28338112)))]; + tensor inner_encoder_conv_module_3_sequential_5_weight = const()[name = tensor("inner_encoder_conv_module_3_sequential_5_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28339200)))]; + tensor inner_encoder_conv_module_3_sequential_7_conv_bias = const()[name = tensor("inner_encoder_conv_module_3_sequential_7_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28340288)))]; + tensor inner_encoder_conv_module_3_sequential_7_conv_weight = const()[name = tensor("inner_encoder_conv_module_3_sequential_7_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28341376)))]; + tensor inner_encoder_ffn2_3_module_sequential_0_bias = const()[name = tensor("inner_encoder_ffn2_3_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28603584)))]; + tensor inner_encoder_ffn2_3_module_sequential_0_weight = const()[name = tensor("inner_encoder_ffn2_3_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28604672)))]; + tensor inner_encoder_ffn2_3_module_sequential_1_linear_bias = const()[name = tensor("inner_encoder_ffn2_3_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28605760)))]; + tensor inner_encoder_ffn2_3_module_sequential_1_linear_weight = const()[name = tensor("inner_encoder_ffn2_3_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28609920)))]; + tensor inner_encoder_ffn2_3_module_sequential_4_linear_bias = const()[name = tensor("inner_encoder_ffn2_3_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(29658560)))]; + tensor inner_encoder_ffn2_3_module_sequential_4_linear_weight = const()[name = tensor("inner_encoder_ffn2_3_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(29659648)))]; + tensor inner_encoder_layer_norm_3_bias = const()[name = tensor("inner_encoder_layer_norm_3_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(30708288)))]; + tensor inner_encoder_layer_norm_3_weight = const()[name = tensor("inner_encoder_layer_norm_3_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(30709376)))]; + tensor inner_decoder__causal_mask = const()[name = tensor("inner_decoder__causal_mask"), val = tensor([[0x1p+0, 0x0p+0], [0x1p+0, 0x1p+0]])]; + tensor inner_decoder_convert_bias = const()[name = tensor("inner_decoder_convert_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(30710464)))]; + tensor inner_decoder_convert_weight = const()[name = tensor("inner_decoder_convert_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(30711552)))]; + tensor inner_decoder_q_proj_0_bias = const()[name = tensor("inner_decoder_q_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(31235904)))]; + tensor inner_decoder_q_proj_0_weight = const()[name = tensor("inner_decoder_q_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(31236992)))]; + tensor inner_decoder_k_proj_0_bias = const()[name = tensor("inner_decoder_k_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(31499200)))]; + tensor inner_decoder_k_proj_0_weight = const()[name = tensor("inner_decoder_k_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(31500288)))]; + tensor inner_decoder_v_proj_0_bias = const()[name = tensor("inner_decoder_v_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(31762496)))]; + tensor inner_decoder_v_proj_0_weight = const()[name = tensor("inner_decoder_v_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(31763584)))]; + tensor inner_decoder_g_proj_0_bias = const()[name = tensor("inner_decoder_g_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32025792)))]; + tensor inner_decoder_g_proj_0_weight = const()[name = tensor("inner_decoder_g_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32026880)))]; + tensor inner_decoder_out_proj_0_bias = const()[name = tensor("inner_decoder_out_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32289088)))]; + tensor inner_decoder_out_proj_0_weight = const()[name = tensor("inner_decoder_out_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32290176)))]; + tensor inner_decoder_norm11_0_bias = const()[name = tensor("inner_decoder_norm11_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32552384)))]; + tensor inner_decoder_norm11_0_weight = const()[name = tensor("inner_decoder_norm11_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32553472)))]; + tensor inner_decoder_self_attn2_0_out_proj_bias = const()[name = tensor("inner_decoder_self_attn2_0_out_proj_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32554560)))]; + tensor inner_decoder_self_attn2_0_out_proj_weight = const()[name = tensor("inner_decoder_self_attn2_0_out_proj_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32555648)))]; + tensor inner_decoder_self_attn2_0_in_proj_bias = const()[name = tensor("inner_decoder_self_attn2_0_in_proj_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32817856)))]; + tensor inner_decoder_self_attn2_0_in_proj_weight = const()[name = tensor("inner_decoder_self_attn2_0_in_proj_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32820992)))]; + tensor inner_decoder_norm21_0_bias = const()[name = tensor("inner_decoder_norm21_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(33607488)))]; + tensor inner_decoder_norm21_0_weight = const()[name = tensor("inner_decoder_norm21_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(33608576)))]; + tensor inner_decoder_linear1_0_bias = const()[name = tensor("inner_decoder_linear1_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(33609664)))]; + tensor inner_decoder_linear1_0_weight = const()[name = tensor("inner_decoder_linear1_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(33617920)))]; + tensor inner_decoder_linear2_0_bias = const()[name = tensor("inner_decoder_linear2_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(35715136)))]; + tensor inner_decoder_linear2_0_weight = const()[name = tensor("inner_decoder_linear2_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(35716224)))]; + tensor inner_decoder_norm22_0_bias = const()[name = tensor("inner_decoder_norm22_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(37813440)))]; + tensor inner_decoder_norm22_0_weight = const()[name = tensor("inner_decoder_norm22_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(37814528)))]; + tensor inner_decoder_q_proj_1_bias = const()[name = tensor("inner_decoder_q_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(37815616)))]; + tensor inner_decoder_q_proj_1_weight = const()[name = tensor("inner_decoder_q_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(37816704)))]; + tensor inner_decoder_k_proj_1_bias = const()[name = tensor("inner_decoder_k_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38078912)))]; + tensor inner_decoder_k_proj_1_weight = const()[name = tensor("inner_decoder_k_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38080000)))]; + tensor inner_decoder_v_proj_1_bias = const()[name = tensor("inner_decoder_v_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38342208)))]; + tensor inner_decoder_v_proj_1_weight = const()[name = tensor("inner_decoder_v_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38343296)))]; + tensor inner_decoder_g_proj_1_bias = const()[name = tensor("inner_decoder_g_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38605504)))]; + tensor inner_decoder_g_proj_1_weight = const()[name = tensor("inner_decoder_g_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38606592)))]; + tensor inner_decoder_out_proj_1_bias = const()[name = tensor("inner_decoder_out_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38868800)))]; + tensor inner_decoder_out_proj_1_weight = const()[name = tensor("inner_decoder_out_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38869888)))]; + tensor inner_decoder_norm11_1_bias = const()[name = tensor("inner_decoder_norm11_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(39132096)))]; + tensor inner_decoder_norm11_1_weight = const()[name = tensor("inner_decoder_norm11_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(39133184)))]; + tensor inner_decoder_self_attn2_1_out_proj_bias = const()[name = tensor("inner_decoder_self_attn2_1_out_proj_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(39134272)))]; + tensor inner_decoder_self_attn2_1_out_proj_weight = const()[name = tensor("inner_decoder_self_attn2_1_out_proj_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(39135360)))]; + tensor inner_decoder_self_attn2_1_in_proj_bias = const()[name = tensor("inner_decoder_self_attn2_1_in_proj_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(39397568)))]; + tensor inner_decoder_self_attn2_1_in_proj_weight = const()[name = tensor("inner_decoder_self_attn2_1_in_proj_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(39400704)))]; + tensor inner_decoder_norm21_1_bias = const()[name = tensor("inner_decoder_norm21_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(40187200)))]; + tensor inner_decoder_norm21_1_weight = const()[name = tensor("inner_decoder_norm21_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(40188288)))]; + tensor inner_decoder_linear1_1_bias = const()[name = tensor("inner_decoder_linear1_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(40189376)))]; + tensor inner_decoder_linear1_1_weight = const()[name = tensor("inner_decoder_linear1_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(40197632)))]; + tensor inner_decoder_linear2_1_bias = const()[name = tensor("inner_decoder_linear2_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(42294848)))]; + tensor inner_decoder_linear2_1_weight = const()[name = tensor("inner_decoder_linear2_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(42295936)))]; + tensor inner_decoder_norm22_1_bias = const()[name = tensor("inner_decoder_norm22_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(44393152)))]; + tensor inner_decoder_norm22_1_weight = const()[name = tensor("inner_decoder_norm22_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(44394240)))]; + tensor var_19_begin_0 = const()[name = tensor("op_19_begin_0"), val = tensor([0, 0, 0])]; + tensor var_19_end_0 = const()[name = tensor("op_19_end_0"), val = tensor([1, 15, 23])]; + tensor var_19_end_mask_0 = const()[name = tensor("op_19_end_mask_0"), val = tensor([true, false, true])]; + tensor var_19 = slice_by_index(begin = var_19_begin_0, end = var_19_end_0, end_mask = var_19_end_mask_0, x = features)[name = tensor("op_19")]; + tensor var_29_begin_0 = const()[name = tensor("op_29_begin_0"), val = tensor([0, 10, 0])]; + tensor var_29_end_0 = const()[name = tensor("op_29_end_0"), val = tensor([1, 1, 23])]; + tensor var_29_end_mask_0 = const()[name = tensor("op_29_end_mask_0"), val = tensor([true, true, true])]; + tensor var_29 = slice_by_index(begin = var_29_begin_0, end = var_29_end_0, end_mask = var_29_end_mask_0, x = features)[name = tensor("op_29")]; + tensor stacked_axis_0 = const()[name = tensor("stacked_axis_0"), val = tensor(1)]; + tensor stacked = stack(axis = stacked_axis_0, values = (var_19, var_29))[name = tensor("stacked")]; + tensor var_36 = const()[name = tensor("op_36"), val = tensor([1, 2, 345])]; + tensor input_1 = reshape(shape = var_36, x = stacked)[name = tensor("input_1")]; + tensor var_39 = const()[name = tensor("op_39"), val = tensor(0x1p+0)]; + tensor var_45 = const()[name = tensor("op_45"), val = tensor(true)]; + tensor var_46 = const()[name = tensor("op_46"), val = tensor(0x1.4f8b58p-17)]; + tensor var_49 = const()[name = tensor("op_49"), val = tensor(0)]; + tensor var_51 = const()[name = tensor("op_51"), val = tensor(2)]; + tensor var_52 = const()[name = tensor("op_52"), val = tensor(-1)]; + tensor var_54 = const()[name = tensor("op_54"), val = tensor(0x1.0c6f7ap-20)]; + tensor var_59 = const()[name = tensor("op_59"), val = tensor(0x1.5798eep-27)]; + tensor var_62 = const()[name = tensor("op_62"), val = tensor(1)]; + tensor input_3 = linear(bias = inner_encoder_input_projection_linear_bias, weight = inner_encoder_input_projection_linear_weight, x = input_1)[name = tensor("linear_0")]; tensor input_5_axes_0 = const()[name = tensor("input_5_axes_0"), val = tensor([-1])]; - tensor input_5 = layer_norm(axes = input_5_axes_0, beta = encoder_ffn1_0_module_sequential_0_bias, epsilon = var_28, gamma = encoder_ffn1_0_module_sequential_0_weight, x = input_3)[name = tensor("input_5")]; - tensor inputs_1 = linear(bias = encoder_ffn1_0_module_sequential_1_linear_bias, weight = encoder_ffn1_0_module_sequential_1_linear_weight, x = input_5)[name = tensor("linear_1")]; - tensor input_7 = silu(x = inputs_1)[name = tensor("input_7")]; - tensor input_11 = linear(bias = encoder_ffn1_0_module_sequential_4_linear_bias, weight = encoder_ffn1_0_module_sequential_4_linear_weight, x = input_7)[name = tensor("linear_2")]; - tensor var_147 = const()[name = tensor("op_147"), val = tensor(0x1p-1)]; - tensor var_148 = mul(x = input_11, y = var_147)[name = tensor("op_148")]; - tensor input_13 = add(x = var_148, y = input_3)[name = tensor("input_13")]; + tensor input_5 = layer_norm(axes = input_5_axes_0, beta = inner_encoder_pre_layer_norm_bias, epsilon = var_46, gamma = inner_encoder_pre_layer_norm_weight, x = input_3)[name = tensor("input_5")]; + tensor input_7_axes_0 = const()[name = tensor("input_7_axes_0"), val = tensor([-1])]; + tensor input_7 = layer_norm(axes = input_7_axes_0, beta = inner_encoder_ffn1_0_module_sequential_0_bias, epsilon = var_46, gamma = inner_encoder_ffn1_0_module_sequential_0_weight, x = input_5)[name = tensor("input_7")]; + tensor inputs_1 = linear(bias = inner_encoder_ffn1_0_module_sequential_1_linear_bias, weight = inner_encoder_ffn1_0_module_sequential_1_linear_weight, x = input_7)[name = tensor("linear_1")]; + tensor input_9 = silu(x = inputs_1)[name = tensor("input_9")]; + tensor input_13 = linear(bias = inner_encoder_ffn1_0_module_sequential_4_linear_bias, weight = inner_encoder_ffn1_0_module_sequential_4_linear_weight, x = input_9)[name = tensor("linear_2")]; + tensor var_183 = const()[name = tensor("op_183"), val = tensor(0x1p-1)]; + tensor var_184 = mul(x = input_13, y = var_183)[name = tensor("op_184")]; + tensor input_15 = add(x = var_184, y = input_5)[name = tensor("input_15")]; tensor x_1_axes_0 = const()[name = tensor("x_1_axes_0"), val = tensor([-1])]; - tensor x_1 = layer_norm(axes = x_1_axes_0, beta = encoder_ret_lns_0_bias, epsilon = var_28, gamma = encoder_ret_lns_0_weight, x = input_13)[name = tensor("x_1")]; + tensor x_1 = layer_norm(axes = x_1_axes_0, beta = inner_encoder_ret_lns_0_bias, epsilon = var_46, gamma = inner_encoder_ret_lns_0_weight, x = input_15)[name = tensor("x_1")]; tensor prev_kv_1_begin_0 = const()[name = tensor("prev_kv_1_begin_0"), val = tensor([0, 0, 0, 0, 0])]; tensor prev_kv_1_end_0 = const()[name = tensor("prev_kv_1_end_0"), val = tensor([1, 1, 4, 64, 64])]; tensor prev_kv_1_end_mask_0 = const()[name = tensor("prev_kv_1_end_mask_0"), val = tensor([false, true, true, true, true])]; @@ -239,153 +253,153 @@ program(1.0) tensor prev_scale_1_end_mask_0 = const()[name = tensor("prev_scale_1_end_mask_0"), val = tensor([false, true])]; tensor prev_scale_1_squeeze_mask_0 = const()[name = tensor("prev_scale_1_squeeze_mask_0"), val = tensor([true, false])]; tensor prev_scale_1 = slice_by_index(begin = prev_scale_1_begin_0, end = prev_scale_1_end_0, end_mask = prev_scale_1_end_mask_0, squeeze_mask = prev_scale_1_squeeze_mask_0, x = enc_scale)[name = tensor("prev_scale_1")]; - tensor var_162 = linear(bias = encoder_q_proj_0_bias, weight = encoder_q_proj_0_weight, x = x_1)[name = tensor("linear_3")]; - tensor var_163 = const()[name = tensor("op_163"), val = tensor([1, 2, 4, 64])]; - tensor var_164 = reshape(shape = var_163, x = var_162)[name = tensor("op_164")]; + tensor var_198 = linear(bias = inner_encoder_q_proj_0_bias, weight = inner_encoder_q_proj_0_weight, x = x_1)[name = tensor("linear_3")]; + tensor var_199 = const()[name = tensor("op_199"), val = tensor([1, 2, 4, 64])]; + tensor var_200 = reshape(shape = var_199, x = var_198)[name = tensor("op_200")]; tensor q_1_perm_0 = const()[name = tensor("q_1_perm_0"), val = tensor([0, 2, 1, 3])]; - tensor var_168 = linear(bias = encoder_k_proj_0_bias, weight = encoder_k_proj_0_weight, x = x_1)[name = tensor("linear_4")]; - tensor var_169 = const()[name = tensor("op_169"), val = tensor(0x1p-3)]; - tensor var_170 = mul(x = var_168, y = var_169)[name = tensor("op_170")]; - tensor var_171 = const()[name = tensor("op_171"), val = tensor([1, 2, 4, 64])]; - tensor var_172 = reshape(shape = var_171, x = var_170)[name = tensor("op_172")]; + tensor var_204 = linear(bias = inner_encoder_k_proj_0_bias, weight = inner_encoder_k_proj_0_weight, x = x_1)[name = tensor("linear_4")]; + tensor var_205 = const()[name = tensor("op_205"), val = tensor(0x1p-3)]; + tensor var_206 = mul(x = var_204, y = var_205)[name = tensor("op_206")]; + tensor var_207 = const()[name = tensor("op_207"), val = tensor([1, 2, 4, 64])]; + tensor var_208 = reshape(shape = var_207, x = var_206)[name = tensor("op_208")]; tensor k_1_perm_0 = const()[name = tensor("k_1_perm_0"), val = tensor([0, 2, 1, 3])]; - tensor var_176 = linear(bias = encoder_v_proj_0_bias, weight = encoder_v_proj_0_weight, x = x_1)[name = tensor("linear_5")]; - tensor var_177 = const()[name = tensor("op_177"), val = tensor([1, 2, 4, 64])]; - tensor var_178 = reshape(shape = var_177, x = var_176)[name = tensor("op_178")]; + tensor var_212 = linear(bias = inner_encoder_v_proj_0_bias, weight = inner_encoder_v_proj_0_weight, x = x_1)[name = tensor("linear_5")]; + tensor var_213 = const()[name = tensor("op_213"), val = tensor([1, 2, 4, 64])]; + tensor var_214 = reshape(shape = var_213, x = var_212)[name = tensor("op_214")]; tensor v_1_perm_0 = const()[name = tensor("v_1_perm_0"), val = tensor([0, 2, 1, 3])]; - tensor input_17 = linear(bias = encoder_g_proj_0_bias, weight = encoder_g_proj_0_weight, x = x_1)[name = tensor("linear_6")]; + tensor input_19 = linear(bias = inner_encoder_g_proj_0_bias, weight = inner_encoder_g_proj_0_weight, x = x_1)[name = tensor("linear_6")]; tensor sqrt_s0_1 = sqrt(x = prev_scale_1)[name = tensor("sqrt_s0_1")]; - tensor s_t_1 = add(x = prev_scale_1, y = encoder__t_index)[name = tensor("s_t_1")]; + tensor s_t_1 = add(x = prev_scale_1, y = inner_encoder__t_index)[name = tensor("s_t_1")]; tensor sqrt_s_t_1 = sqrt(x = s_t_1)[name = tensor("sqrt_s_t_1")]; tensor qk_1_transpose_x_1 = const()[name = tensor("qk_1_transpose_x_1"), val = tensor(false)]; tensor qk_1_transpose_y_1 = const()[name = tensor("qk_1_transpose_y_1"), val = tensor(true)]; - tensor k_1 = transpose(perm = k_1_perm_0, x = var_172)[name = tensor("transpose_57")]; - tensor q_1 = transpose(perm = q_1_perm_0, x = var_164)[name = tensor("transpose_58")]; + tensor k_1 = transpose(perm = k_1_perm_0, x = var_208)[name = tensor("transpose_57")]; + tensor q_1 = transpose(perm = q_1_perm_0, x = var_200)[name = tensor("transpose_58")]; tensor qk_1 = matmul(transpose_x = qk_1_transpose_x_1, transpose_y = qk_1_transpose_y_1, x = q_1, y = k_1)[name = tensor("qk_1")]; - tensor var_188 = const()[name = tensor("op_188"), val = tensor([2, 1])]; - tensor var_189 = reshape(shape = var_188, x = sqrt_s_t_1)[name = tensor("op_189")]; - tensor M_1 = real_div(x = encoder__causal_mask, y = var_189)[name = tensor("M_1")]; - tensor var_191 = mul(x = qk_1, y = M_1)[name = tensor("op_191")]; + tensor var_224 = const()[name = tensor("op_224"), val = tensor([2, 1])]; + tensor var_225 = reshape(shape = var_224, x = sqrt_s_t_1)[name = tensor("op_225")]; + tensor M_1 = real_div(x = inner_encoder__causal_mask, y = var_225)[name = tensor("M_1")]; + tensor var_227 = mul(x = qk_1, y = M_1)[name = tensor("op_227")]; tensor inner_1_transpose_x_0 = const()[name = tensor("inner_1_transpose_x_0"), val = tensor(false)]; tensor inner_1_transpose_y_0 = const()[name = tensor("inner_1_transpose_y_0"), val = tensor(false)]; - tensor v_1 = transpose(perm = v_1_perm_0, x = var_178)[name = tensor("transpose_56")]; - tensor inner_1 = matmul(transpose_x = inner_1_transpose_x_0, transpose_y = inner_1_transpose_y_0, x = var_191, y = v_1)[name = tensor("inner_1")]; - tensor var_193_transpose_x_0 = const()[name = tensor("op_193_transpose_x_0"), val = tensor(false)]; - tensor var_193_transpose_y_0 = const()[name = tensor("op_193_transpose_y_0"), val = tensor(false)]; - tensor var_193 = matmul(transpose_x = var_193_transpose_x_0, transpose_y = var_193_transpose_y_0, x = q_1, y = prev_kv_1)[name = tensor("op_193")]; - tensor var_194 = real_div(x = sqrt_s0_1, y = sqrt_s_t_1)[name = tensor("op_194")]; - tensor var_195 = const()[name = tensor("op_195"), val = tensor([1, 1, 2, 1])]; - tensor var_196 = reshape(shape = var_195, x = var_194)[name = tensor("op_196")]; - tensor cross_1 = mul(x = var_193, y = var_196)[name = tensor("cross_1")]; + tensor v_1 = transpose(perm = v_1_perm_0, x = var_214)[name = tensor("transpose_56")]; + tensor inner_1 = matmul(transpose_x = inner_1_transpose_x_0, transpose_y = inner_1_transpose_y_0, x = var_227, y = v_1)[name = tensor("inner_1")]; + tensor var_229_transpose_x_0 = const()[name = tensor("op_229_transpose_x_0"), val = tensor(false)]; + tensor var_229_transpose_y_0 = const()[name = tensor("op_229_transpose_y_0"), val = tensor(false)]; + tensor var_229 = matmul(transpose_x = var_229_transpose_x_0, transpose_y = var_229_transpose_y_0, x = q_1, y = prev_kv_1)[name = tensor("op_229")]; + tensor var_230 = real_div(x = sqrt_s0_1, y = sqrt_s_t_1)[name = tensor("op_230")]; + tensor var_231 = const()[name = tensor("op_231"), val = tensor([1, 1, 2, 1])]; + tensor var_232 = reshape(shape = var_231, x = var_230)[name = tensor("op_232")]; + tensor cross_1 = mul(x = var_229, y = var_232)[name = tensor("cross_1")]; tensor out_1 = add(x = inner_1, y = cross_1)[name = tensor("out_1")]; - tensor var_199 = mul(x = prev_kv_1, y = sqrt_s0_1)[name = tensor("op_199")]; - tensor var_201_transpose_x_1 = const()[name = tensor("op_201_transpose_x_1"), val = tensor(true)]; - tensor var_201_transpose_y_1 = const()[name = tensor("op_201_transpose_y_1"), val = tensor(false)]; - tensor var_201 = matmul(transpose_x = var_201_transpose_x_1, transpose_y = var_201_transpose_y_1, x = k_1, y = v_1)[name = tensor("op_201")]; - tensor new_kv_unnorm_1 = add(x = var_199, y = var_201)[name = tensor("new_kv_unnorm_1")]; - tensor var_203 = const()[name = tensor("op_203"), val = tensor(0x1p+1)]; - tensor new_scale_1 = add(x = prev_scale_1, y = var_203)[name = tensor("new_scale_1")]; - tensor var_205 = sqrt(x = new_scale_1)[name = tensor("op_205")]; - tensor var_206 = real_div(x = new_kv_unnorm_1, y = var_205)[name = tensor("op_206")]; - tensor var_207_perm_0 = const()[name = tensor("op_207_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_235 = mul(x = prev_kv_1, y = sqrt_s0_1)[name = tensor("op_235")]; + tensor var_237_transpose_x_1 = const()[name = tensor("op_237_transpose_x_1"), val = tensor(true)]; + tensor var_237_transpose_y_1 = const()[name = tensor("op_237_transpose_y_1"), val = tensor(false)]; + tensor var_237 = matmul(transpose_x = var_237_transpose_x_1, transpose_y = var_237_transpose_y_1, x = k_1, y = v_1)[name = tensor("op_237")]; + tensor new_kv_unnorm_1 = add(x = var_235, y = var_237)[name = tensor("new_kv_unnorm_1")]; + tensor var_239 = const()[name = tensor("op_239"), val = tensor(0x1p+1)]; + tensor new_scale_1 = add(x = prev_scale_1, y = var_239)[name = tensor("new_scale_1")]; + tensor var_241 = sqrt(x = new_scale_1)[name = tensor("op_241")]; + tensor var_242 = real_div(x = new_kv_unnorm_1, y = var_241)[name = tensor("op_242")]; + tensor var_243_perm_0 = const()[name = tensor("op_243_perm_0"), val = tensor([0, 2, 1, 3])]; tensor out_3_axes_0 = const()[name = tensor("out_3_axes_0"), val = tensor([-1])]; - tensor var_207 = transpose(perm = var_207_perm_0, x = out_1)[name = tensor("transpose_55")]; - tensor out_3 = layer_norm(axes = out_3_axes_0, epsilon = var_17, x = var_207)[name = tensor("out_3")]; - tensor var_211 = const()[name = tensor("op_211"), val = tensor([1, 2, 256])]; - tensor out_5 = reshape(shape = var_211, x = out_3)[name = tensor("out_5")]; - tensor var_213 = silu(x = input_17)[name = tensor("op_213")]; - tensor input_19 = mul(x = var_213, y = out_5)[name = tensor("input_19")]; - tensor ret_out_1 = linear(bias = encoder_out_proj_0_bias, weight = encoder_out_proj_0_weight, x = input_19)[name = tensor("linear_7")]; - tensor x_3 = add(x = input_13, y = ret_out_1)[name = tensor("x_3")]; + tensor var_243 = transpose(perm = var_243_perm_0, x = out_1)[name = tensor("transpose_55")]; + tensor out_3 = layer_norm(axes = out_3_axes_0, epsilon = var_54, x = var_243)[name = tensor("out_3")]; + tensor var_247 = const()[name = tensor("op_247"), val = tensor([1, 2, 256])]; + tensor out_5 = reshape(shape = var_247, x = out_3)[name = tensor("out_5")]; + tensor var_249 = silu(x = input_19)[name = tensor("op_249")]; + tensor input_21 = mul(x = var_249, y = out_5)[name = tensor("input_21")]; + tensor ret_out_1 = linear(bias = inner_encoder_out_proj_0_bias, weight = inner_encoder_out_proj_0_weight, x = input_21)[name = tensor("linear_7")]; + tensor x_3 = add(x = input_15, y = ret_out_1)[name = tensor("x_3")]; tensor window_1_begin_0 = const()[name = tensor("window_1_begin_0"), val = tensor([0, 0, 0, 0])]; tensor window_1_end_0 = const()[name = tensor("window_1_end_0"), val = tensor([1, 1, 16, 256])]; tensor window_1_end_mask_0 = const()[name = tensor("window_1_end_mask_0"), val = tensor([false, true, true, true])]; tensor window_1_squeeze_mask_0 = const()[name = tensor("window_1_squeeze_mask_0"), val = tensor([true, false, false, false])]; tensor window_1 = slice_by_index(begin = window_1_begin_0, end = window_1_end_0, end_mask = window_1_end_mask_0, squeeze_mask = window_1_squeeze_mask_0, x = enc_conv_cache)[name = tensor("window_1")]; - tensor var_221_begin_0 = const()[name = tensor("op_221_begin_0"), val = tensor([0, 0, 0])]; - tensor var_221_end_0 = const()[name = tensor("op_221_end_0"), val = tensor([1, 1, 256])]; - tensor var_221_end_mask_0 = const()[name = tensor("op_221_end_mask_0"), val = tensor([true, false, true])]; - tensor var_221 = slice_by_index(begin = var_221_begin_0, end = var_221_end_0, end_mask = var_221_end_mask_0, x = x_3)[name = tensor("op_221")]; - tensor var_224_begin_0 = const()[name = tensor("op_224_begin_0"), val = tensor([0, 1, 0])]; - tensor var_224_end_0 = const()[name = tensor("op_224_end_0"), val = tensor([1, 16, 256])]; - tensor var_224_end_mask_0 = const()[name = tensor("op_224_end_mask_0"), val = tensor([true, true, true])]; - tensor var_224 = slice_by_index(begin = var_224_begin_0, end = var_224_end_0, end_mask = var_224_end_mask_0, x = window_1)[name = tensor("op_224")]; + tensor var_257_begin_0 = const()[name = tensor("op_257_begin_0"), val = tensor([0, 0, 0])]; + tensor var_257_end_0 = const()[name = tensor("op_257_end_0"), val = tensor([1, 1, 256])]; + tensor var_257_end_mask_0 = const()[name = tensor("op_257_end_mask_0"), val = tensor([true, false, true])]; + tensor var_257 = slice_by_index(begin = var_257_begin_0, end = var_257_end_0, end_mask = var_257_end_mask_0, x = x_3)[name = tensor("op_257")]; + tensor var_260_begin_0 = const()[name = tensor("op_260_begin_0"), val = tensor([0, 1, 0])]; + tensor var_260_end_0 = const()[name = tensor("op_260_end_0"), val = tensor([1, 16, 256])]; + tensor var_260_end_mask_0 = const()[name = tensor("op_260_end_mask_0"), val = tensor([true, true, true])]; + tensor var_260 = slice_by_index(begin = var_260_begin_0, end = var_260_end_0, end_mask = var_260_end_mask_0, x = window_1)[name = tensor("op_260")]; tensor window_3_interleave_0 = const()[name = tensor("window_3_interleave_0"), val = tensor(false)]; - tensor window_3 = concat(axis = var_26, interleave = window_3_interleave_0, values = (var_224, var_221))[name = tensor("window_3")]; - tensor var_229_begin_0 = const()[name = tensor("op_229_begin_0"), val = tensor([0, 1, 0])]; - tensor var_229_end_0 = const()[name = tensor("op_229_end_0"), val = tensor([1, 1, 256])]; - tensor var_229_end_mask_0 = const()[name = tensor("op_229_end_mask_0"), val = tensor([true, true, true])]; - tensor var_229 = slice_by_index(begin = var_229_begin_0, end = var_229_end_0, end_mask = var_229_end_mask_0, x = x_3)[name = tensor("op_229")]; - tensor var_232_begin_0 = const()[name = tensor("op_232_begin_0"), val = tensor([0, 1, 0])]; - tensor var_232_end_0 = const()[name = tensor("op_232_end_0"), val = tensor([1, 16, 256])]; - tensor var_232_end_mask_0 = const()[name = tensor("op_232_end_mask_0"), val = tensor([true, true, true])]; - tensor var_232 = slice_by_index(begin = var_232_begin_0, end = var_232_end_0, end_mask = var_232_end_mask_0, x = window_3)[name = tensor("op_232")]; + tensor window_3 = concat(axis = var_62, interleave = window_3_interleave_0, values = (var_260, var_257))[name = tensor("window_3")]; + tensor var_265_begin_0 = const()[name = tensor("op_265_begin_0"), val = tensor([0, 1, 0])]; + tensor var_265_end_0 = const()[name = tensor("op_265_end_0"), val = tensor([1, 1, 256])]; + tensor var_265_end_mask_0 = const()[name = tensor("op_265_end_mask_0"), val = tensor([true, true, true])]; + tensor var_265 = slice_by_index(begin = var_265_begin_0, end = var_265_end_0, end_mask = var_265_end_mask_0, x = x_3)[name = tensor("op_265")]; + tensor var_268_begin_0 = const()[name = tensor("op_268_begin_0"), val = tensor([0, 1, 0])]; + tensor var_268_end_0 = const()[name = tensor("op_268_end_0"), val = tensor([1, 16, 256])]; + tensor var_268_end_mask_0 = const()[name = tensor("op_268_end_mask_0"), val = tensor([true, true, true])]; + tensor var_268 = slice_by_index(begin = var_268_begin_0, end = var_268_end_0, end_mask = var_268_end_mask_0, x = window_3)[name = tensor("op_268")]; tensor window_5_interleave_0 = const()[name = tensor("window_5_interleave_0"), val = tensor(false)]; - tensor window_5 = concat(axis = var_26, interleave = window_5_interleave_0, values = (var_232, var_229))[name = tensor("window_5")]; - tensor input_21_interleave_0 = const()[name = tensor("input_21_interleave_0"), val = tensor(false)]; - tensor input_21 = concat(axis = var_23, interleave = input_21_interleave_0, values = (window_3, window_5))[name = tensor("input_21")]; + tensor window_5 = concat(axis = var_62, interleave = window_5_interleave_0, values = (var_268, var_265))[name = tensor("window_5")]; + tensor input_23_interleave_0 = const()[name = tensor("input_23_interleave_0"), val = tensor(false)]; + tensor input_23 = concat(axis = var_49, interleave = input_23_interleave_0, values = (window_3, window_5))[name = tensor("input_23")]; tensor x_5_axes_0 = const()[name = tensor("x_5_axes_0"), val = tensor([-1])]; - tensor x_5 = layer_norm(axes = x_5_axes_0, beta = encoder_conv_module_0_sequential_0_bias, epsilon = var_28, gamma = encoder_conv_module_0_sequential_0_weight, x = input_21)[name = tensor("x_5")]; - tensor input_23_perm_0 = const()[name = tensor("input_23_perm_0"), val = tensor([0, 2, 1])]; + tensor x_5 = layer_norm(axes = x_5_axes_0, beta = inner_encoder_conv_module_0_sequential_0_bias, epsilon = var_46, gamma = inner_encoder_conv_module_0_sequential_0_weight, x = input_23)[name = tensor("x_5")]; + tensor input_25_perm_0 = const()[name = tensor("input_25_perm_0"), val = tensor([0, 2, 1])]; tensor inputs_3_pad_type_0 = const()[name = tensor("inputs_3_pad_type_0"), val = tensor("valid")]; tensor inputs_3_strides_0 = const()[name = tensor("inputs_3_strides_0"), val = tensor([1])]; tensor inputs_3_pad_0 = const()[name = tensor("inputs_3_pad_0"), val = tensor([0, 0])]; tensor inputs_3_dilations_0 = const()[name = tensor("inputs_3_dilations_0"), val = tensor([1])]; tensor inputs_3_groups_0 = const()[name = tensor("inputs_3_groups_0"), val = tensor(1)]; - tensor input_23 = transpose(perm = input_23_perm_0, x = x_5)[name = tensor("transpose_54")]; - tensor inputs_3 = conv(bias = encoder_conv_module_0_sequential_2_conv_bias, dilations = inputs_3_dilations_0, groups = inputs_3_groups_0, pad = inputs_3_pad_0, pad_type = inputs_3_pad_type_0, strides = inputs_3_strides_0, weight = encoder_conv_module_0_sequential_2_conv_weight, x = input_23)[name = tensor("inputs_3")]; - tensor var_257_split_sizes_0 = const()[name = tensor("op_257_split_sizes_0"), val = tensor([256, 256])]; - tensor var_257_axis_0 = const()[name = tensor("op_257_axis_0"), val = tensor(1)]; - tensor var_257_0, tensor var_257_1 = split(axis = var_257_axis_0, split_sizes = var_257_split_sizes_0, x = inputs_3)[name = tensor("op_257")]; - tensor var_259 = sigmoid(x = var_257_1)[name = tensor("op_259")]; - tensor inputs_5 = mul(x = var_257_0, y = var_259)[name = tensor("inputs_5")]; + tensor input_25 = transpose(perm = input_25_perm_0, x = x_5)[name = tensor("transpose_54")]; + tensor inputs_3 = conv(bias = inner_encoder_conv_module_0_sequential_2_conv_bias, dilations = inputs_3_dilations_0, groups = inputs_3_groups_0, pad = inputs_3_pad_0, pad_type = inputs_3_pad_type_0, strides = inputs_3_strides_0, weight = inner_encoder_conv_module_0_sequential_2_conv_weight, x = input_25)[name = tensor("inputs_3")]; + tensor var_293_split_sizes_0 = const()[name = tensor("op_293_split_sizes_0"), val = tensor([256, 256])]; + tensor var_293_axis_0 = const()[name = tensor("op_293_axis_0"), val = tensor(1)]; + tensor var_293_0, tensor var_293_1 = split(axis = var_293_axis_0, split_sizes = var_293_split_sizes_0, x = inputs_3)[name = tensor("op_293")]; + tensor var_295 = sigmoid(x = var_293_1)[name = tensor("op_295")]; + tensor inputs_5 = mul(x = var_293_0, y = var_295)[name = tensor("inputs_5")]; tensor outputs_aug_1_pad_type_0 = const()[name = tensor("outputs_aug_1_pad_type_0"), val = tensor("custom")]; tensor outputs_aug_1_pad_0 = const()[name = tensor("outputs_aug_1_pad_0"), val = tensor([15, 15])]; tensor outputs_aug_1_groups_0 = const()[name = tensor("outputs_aug_1_groups_0"), val = tensor(256)]; tensor outputs_aug_1_strides_0 = const()[name = tensor("outputs_aug_1_strides_0"), val = tensor([1])]; tensor outputs_aug_1_dilations_0 = const()[name = tensor("outputs_aug_1_dilations_0"), val = tensor([1])]; - tensor outputs_aug_1 = conv(dilations = outputs_aug_1_dilations_0, groups = outputs_aug_1_groups_0, pad = outputs_aug_1_pad_0, pad_type = outputs_aug_1_pad_type_0, strides = outputs_aug_1_strides_0, weight = encoder_conv_module_0_sequential_4_conv_weight, x = inputs_5)[name = tensor("outputs_aug_1")]; - tensor input_25_begin_0 = const()[name = tensor("input_25_begin_0"), val = tensor([0, 0, 0])]; - tensor input_25_end_0 = const()[name = tensor("input_25_end_0"), val = tensor([2, 256, 16])]; - tensor input_25_end_mask_0 = const()[name = tensor("input_25_end_mask_0"), val = tensor([true, true, false])]; - tensor input_25 = slice_by_index(begin = input_25_begin_0, end = input_25_end_0, end_mask = input_25_end_mask_0, x = outputs_aug_1)[name = tensor("input_25")]; - tensor inputs_7 = batch_norm(beta = encoder_conv_module_0_sequential_5_bias, epsilon = var_28, gamma = encoder_conv_module_0_sequential_5_weight, mean = encoder_conv_module_0_sequential_5_running_mean, variance = encoder_conv_module_0_sequential_5_running_var, x = input_25)[name = tensor("inputs_7")]; - tensor input_27 = silu(x = inputs_7)[name = tensor("input_27")]; - tensor input_29_pad_type_0 = const()[name = tensor("input_29_pad_type_0"), val = tensor("valid")]; - tensor input_29_strides_0 = const()[name = tensor("input_29_strides_0"), val = tensor([1])]; - tensor input_29_pad_0 = const()[name = tensor("input_29_pad_0"), val = tensor([0, 0])]; - tensor input_29_dilations_0 = const()[name = tensor("input_29_dilations_0"), val = tensor([1])]; - tensor input_29_groups_0 = const()[name = tensor("input_29_groups_0"), val = tensor(1)]; - tensor input_29 = conv(bias = encoder_conv_module_0_sequential_7_conv_bias, dilations = input_29_dilations_0, groups = input_29_groups_0, pad = input_29_pad_0, pad_type = input_29_pad_type_0, strides = input_29_strides_0, weight = encoder_conv_module_0_sequential_7_conv_weight, x = input_27)[name = tensor("input_29")]; + tensor outputs_aug_1 = conv(dilations = outputs_aug_1_dilations_0, groups = outputs_aug_1_groups_0, pad = outputs_aug_1_pad_0, pad_type = outputs_aug_1_pad_type_0, strides = outputs_aug_1_strides_0, weight = inner_encoder_conv_module_0_sequential_4_conv_weight, x = inputs_5)[name = tensor("outputs_aug_1")]; + tensor input_27_begin_0 = const()[name = tensor("input_27_begin_0"), val = tensor([0, 0, 0])]; + tensor input_27_end_0 = const()[name = tensor("input_27_end_0"), val = tensor([2, 256, 16])]; + tensor input_27_end_mask_0 = const()[name = tensor("input_27_end_mask_0"), val = tensor([true, true, false])]; + tensor input_27 = slice_by_index(begin = input_27_begin_0, end = input_27_end_0, end_mask = input_27_end_mask_0, x = outputs_aug_1)[name = tensor("input_27")]; + tensor inputs_7 = batch_norm(beta = inner_encoder_conv_module_0_sequential_5_bias, epsilon = var_46, gamma = inner_encoder_conv_module_0_sequential_5_weight, mean = inner_encoder_conv_module_0_sequential_5_running_mean, variance = inner_encoder_conv_module_0_sequential_5_running_var, x = input_27)[name = tensor("inputs_7")]; + tensor input_29 = silu(x = inputs_7)[name = tensor("input_29")]; + tensor input_31_pad_type_0 = const()[name = tensor("input_31_pad_type_0"), val = tensor("valid")]; + tensor input_31_strides_0 = const()[name = tensor("input_31_strides_0"), val = tensor([1])]; + tensor input_31_pad_0 = const()[name = tensor("input_31_pad_0"), val = tensor([0, 0])]; + tensor input_31_dilations_0 = const()[name = tensor("input_31_dilations_0"), val = tensor([1])]; + tensor input_31_groups_0 = const()[name = tensor("input_31_groups_0"), val = tensor(1)]; + tensor input_31 = conv(bias = inner_encoder_conv_module_0_sequential_7_conv_bias, dilations = input_31_dilations_0, groups = input_31_groups_0, pad = input_31_pad_0, pad_type = input_31_pad_type_0, strides = input_31_strides_0, weight = inner_encoder_conv_module_0_sequential_7_conv_weight, x = input_29)[name = tensor("input_31")]; tensor conv_out_1_perm_0 = const()[name = tensor("conv_out_1_perm_0"), val = tensor([0, 2, 1])]; - tensor var_290_begin_0 = const()[name = tensor("op_290_begin_0"), val = tensor([0, -1, 0])]; - tensor var_290_end_0 = const()[name = tensor("op_290_end_0"), val = tensor([2, 16, 256])]; - tensor var_290_end_mask_0 = const()[name = tensor("op_290_end_mask_0"), val = tensor([true, true, true])]; - tensor conv_out_1 = transpose(perm = conv_out_1_perm_0, x = input_29)[name = tensor("transpose_53")]; - tensor var_290 = slice_by_index(begin = var_290_begin_0, end = var_290_end_0, end_mask = var_290_end_mask_0, x = conv_out_1)[name = tensor("op_290")]; - tensor var_292_perm_0 = const()[name = tensor("op_292_perm_0"), val = tensor([1, 0, 2])]; - tensor var_292 = transpose(perm = var_292_perm_0, x = var_290)[name = tensor("transpose_52")]; - tensor input_31 = add(x = x_3, y = var_292)[name = tensor("input_31")]; - tensor input_33_axes_0 = const()[name = tensor("input_33_axes_0"), val = tensor([-1])]; - tensor input_33 = layer_norm(axes = input_33_axes_0, beta = encoder_ffn2_0_module_sequential_0_bias, epsilon = var_28, gamma = encoder_ffn2_0_module_sequential_0_weight, x = input_31)[name = tensor("input_33")]; - tensor inputs_9 = linear(bias = encoder_ffn2_0_module_sequential_1_linear_bias, weight = encoder_ffn2_0_module_sequential_1_linear_weight, x = input_33)[name = tensor("linear_8")]; - tensor input_35 = silu(x = inputs_9)[name = tensor("input_35")]; - tensor input_39 = linear(bias = encoder_ffn2_0_module_sequential_4_linear_bias, weight = encoder_ffn2_0_module_sequential_4_linear_weight, x = input_35)[name = tensor("linear_9")]; - tensor var_315 = const()[name = tensor("op_315"), val = tensor(0x1p-1)]; - tensor var_316 = mul(x = input_39, y = var_315)[name = tensor("op_316")]; - tensor input_41 = add(x = var_316, y = input_31)[name = tensor("input_41")]; - tensor input_43_axes_0 = const()[name = tensor("input_43_axes_0"), val = tensor([-1])]; - tensor input_43 = layer_norm(axes = input_43_axes_0, beta = encoder_layer_norm_0_bias, epsilon = var_28, gamma = encoder_layer_norm_0_weight, x = input_41)[name = tensor("input_43")]; + tensor var_326_begin_0 = const()[name = tensor("op_326_begin_0"), val = tensor([0, -1, 0])]; + tensor var_326_end_0 = const()[name = tensor("op_326_end_0"), val = tensor([2, 16, 256])]; + tensor var_326_end_mask_0 = const()[name = tensor("op_326_end_mask_0"), val = tensor([true, true, true])]; + tensor conv_out_1 = transpose(perm = conv_out_1_perm_0, x = input_31)[name = tensor("transpose_53")]; + tensor var_326 = slice_by_index(begin = var_326_begin_0, end = var_326_end_0, end_mask = var_326_end_mask_0, x = conv_out_1)[name = tensor("op_326")]; + tensor var_328_perm_0 = const()[name = tensor("op_328_perm_0"), val = tensor([1, 0, 2])]; + tensor var_328 = transpose(perm = var_328_perm_0, x = var_326)[name = tensor("transpose_52")]; + tensor input_33 = add(x = x_3, y = var_328)[name = tensor("input_33")]; + tensor input_35_axes_0 = const()[name = tensor("input_35_axes_0"), val = tensor([-1])]; + tensor input_35 = layer_norm(axes = input_35_axes_0, beta = inner_encoder_ffn2_0_module_sequential_0_bias, epsilon = var_46, gamma = inner_encoder_ffn2_0_module_sequential_0_weight, x = input_33)[name = tensor("input_35")]; + tensor inputs_9 = linear(bias = inner_encoder_ffn2_0_module_sequential_1_linear_bias, weight = inner_encoder_ffn2_0_module_sequential_1_linear_weight, x = input_35)[name = tensor("linear_8")]; + tensor input_37 = silu(x = inputs_9)[name = tensor("input_37")]; + tensor input_41 = linear(bias = inner_encoder_ffn2_0_module_sequential_4_linear_bias, weight = inner_encoder_ffn2_0_module_sequential_4_linear_weight, x = input_37)[name = tensor("linear_9")]; + tensor var_351 = const()[name = tensor("op_351"), val = tensor(0x1p-1)]; + tensor var_352 = mul(x = input_41, y = var_351)[name = tensor("op_352")]; + tensor input_43 = add(x = var_352, y = input_33)[name = tensor("input_43")]; tensor input_45_axes_0 = const()[name = tensor("input_45_axes_0"), val = tensor([-1])]; - tensor input_45 = layer_norm(axes = input_45_axes_0, beta = encoder_ffn1_1_module_sequential_0_bias, epsilon = var_28, gamma = encoder_ffn1_1_module_sequential_0_weight, x = input_43)[name = tensor("input_45")]; - tensor inputs_11 = linear(bias = encoder_ffn1_1_module_sequential_1_linear_bias, weight = encoder_ffn1_1_module_sequential_1_linear_weight, x = input_45)[name = tensor("linear_10")]; - tensor input_47 = silu(x = inputs_11)[name = tensor("input_47")]; - tensor input_51 = linear(bias = encoder_ffn1_1_module_sequential_4_linear_bias, weight = encoder_ffn1_1_module_sequential_4_linear_weight, x = input_47)[name = tensor("linear_11")]; - tensor var_345 = const()[name = tensor("op_345"), val = tensor(0x1p-1)]; - tensor var_346 = mul(x = input_51, y = var_345)[name = tensor("op_346")]; - tensor input_53 = add(x = var_346, y = input_43)[name = tensor("input_53")]; + tensor input_45 = layer_norm(axes = input_45_axes_0, beta = inner_encoder_layer_norm_0_bias, epsilon = var_46, gamma = inner_encoder_layer_norm_0_weight, x = input_43)[name = tensor("input_45")]; + tensor input_47_axes_0 = const()[name = tensor("input_47_axes_0"), val = tensor([-1])]; + tensor input_47 = layer_norm(axes = input_47_axes_0, beta = inner_encoder_ffn1_1_module_sequential_0_bias, epsilon = var_46, gamma = inner_encoder_ffn1_1_module_sequential_0_weight, x = input_45)[name = tensor("input_47")]; + tensor inputs_11 = linear(bias = inner_encoder_ffn1_1_module_sequential_1_linear_bias, weight = inner_encoder_ffn1_1_module_sequential_1_linear_weight, x = input_47)[name = tensor("linear_10")]; + tensor input_49 = silu(x = inputs_11)[name = tensor("input_49")]; + tensor input_53 = linear(bias = inner_encoder_ffn1_1_module_sequential_4_linear_bias, weight = inner_encoder_ffn1_1_module_sequential_4_linear_weight, x = input_49)[name = tensor("linear_11")]; + tensor var_381 = const()[name = tensor("op_381"), val = tensor(0x1p-1)]; + tensor var_382 = mul(x = input_53, y = var_381)[name = tensor("op_382")]; + tensor input_55 = add(x = var_382, y = input_45)[name = tensor("input_55")]; tensor x_7_axes_0 = const()[name = tensor("x_7_axes_0"), val = tensor([-1])]; - tensor x_7 = layer_norm(axes = x_7_axes_0, beta = encoder_ret_lns_1_bias, epsilon = var_28, gamma = encoder_ret_lns_1_weight, x = input_53)[name = tensor("x_7")]; + tensor x_7 = layer_norm(axes = x_7_axes_0, beta = inner_encoder_ret_lns_1_bias, epsilon = var_46, gamma = inner_encoder_ret_lns_1_weight, x = input_55)[name = tensor("x_7")]; tensor prev_kv_3_begin_0 = const()[name = tensor("prev_kv_3_begin_0"), val = tensor([1, 0, 0, 0, 0])]; tensor prev_kv_3_end_0 = const()[name = tensor("prev_kv_3_end_0"), val = tensor([2, 1, 4, 64, 64])]; tensor prev_kv_3_end_mask_0 = const()[name = tensor("prev_kv_3_end_mask_0"), val = tensor([false, true, true, true, true])]; @@ -396,153 +410,153 @@ program(1.0) tensor prev_scale_3_end_mask_0 = const()[name = tensor("prev_scale_3_end_mask_0"), val = tensor([false, true])]; tensor prev_scale_3_squeeze_mask_0 = const()[name = tensor("prev_scale_3_squeeze_mask_0"), val = tensor([true, false])]; tensor prev_scale_3 = slice_by_index(begin = prev_scale_3_begin_0, end = prev_scale_3_end_0, end_mask = prev_scale_3_end_mask_0, squeeze_mask = prev_scale_3_squeeze_mask_0, x = enc_scale)[name = tensor("prev_scale_3")]; - tensor var_360 = linear(bias = encoder_q_proj_1_bias, weight = encoder_q_proj_1_weight, x = x_7)[name = tensor("linear_12")]; - tensor var_361 = const()[name = tensor("op_361"), val = tensor([1, 2, 4, 64])]; - tensor var_362 = reshape(shape = var_361, x = var_360)[name = tensor("op_362")]; + tensor var_396 = linear(bias = inner_encoder_q_proj_1_bias, weight = inner_encoder_q_proj_1_weight, x = x_7)[name = tensor("linear_12")]; + tensor var_397 = const()[name = tensor("op_397"), val = tensor([1, 2, 4, 64])]; + tensor var_398 = reshape(shape = var_397, x = var_396)[name = tensor("op_398")]; tensor q_3_perm_0 = const()[name = tensor("q_3_perm_0"), val = tensor([0, 2, 1, 3])]; - tensor var_366 = linear(bias = encoder_k_proj_1_bias, weight = encoder_k_proj_1_weight, x = x_7)[name = tensor("linear_13")]; - tensor var_367 = const()[name = tensor("op_367"), val = tensor(0x1p-3)]; - tensor var_368 = mul(x = var_366, y = var_367)[name = tensor("op_368")]; - tensor var_369 = const()[name = tensor("op_369"), val = tensor([1, 2, 4, 64])]; - tensor var_370 = reshape(shape = var_369, x = var_368)[name = tensor("op_370")]; + tensor var_402 = linear(bias = inner_encoder_k_proj_1_bias, weight = inner_encoder_k_proj_1_weight, x = x_7)[name = tensor("linear_13")]; + tensor var_403 = const()[name = tensor("op_403"), val = tensor(0x1p-3)]; + tensor var_404 = mul(x = var_402, y = var_403)[name = tensor("op_404")]; + tensor var_405 = const()[name = tensor("op_405"), val = tensor([1, 2, 4, 64])]; + tensor var_406 = reshape(shape = var_405, x = var_404)[name = tensor("op_406")]; tensor k_3_perm_0 = const()[name = tensor("k_3_perm_0"), val = tensor([0, 2, 1, 3])]; - tensor var_374 = linear(bias = encoder_v_proj_1_bias, weight = encoder_v_proj_1_weight, x = x_7)[name = tensor("linear_14")]; - tensor var_375 = const()[name = tensor("op_375"), val = tensor([1, 2, 4, 64])]; - tensor var_376 = reshape(shape = var_375, x = var_374)[name = tensor("op_376")]; + tensor var_410 = linear(bias = inner_encoder_v_proj_1_bias, weight = inner_encoder_v_proj_1_weight, x = x_7)[name = tensor("linear_14")]; + tensor var_411 = const()[name = tensor("op_411"), val = tensor([1, 2, 4, 64])]; + tensor var_412 = reshape(shape = var_411, x = var_410)[name = tensor("op_412")]; tensor v_3_perm_0 = const()[name = tensor("v_3_perm_0"), val = tensor([0, 2, 1, 3])]; - tensor input_57 = linear(bias = encoder_g_proj_1_bias, weight = encoder_g_proj_1_weight, x = x_7)[name = tensor("linear_15")]; + tensor input_59 = linear(bias = inner_encoder_g_proj_1_bias, weight = inner_encoder_g_proj_1_weight, x = x_7)[name = tensor("linear_15")]; tensor sqrt_s0_3 = sqrt(x = prev_scale_3)[name = tensor("sqrt_s0_3")]; - tensor s_t_3 = add(x = prev_scale_3, y = encoder__t_index)[name = tensor("s_t_3")]; + tensor s_t_3 = add(x = prev_scale_3, y = inner_encoder__t_index)[name = tensor("s_t_3")]; tensor sqrt_s_t_3 = sqrt(x = s_t_3)[name = tensor("sqrt_s_t_3")]; tensor qk_3_transpose_x_1 = const()[name = tensor("qk_3_transpose_x_1"), val = tensor(false)]; tensor qk_3_transpose_y_1 = const()[name = tensor("qk_3_transpose_y_1"), val = tensor(true)]; - tensor k_3 = transpose(perm = k_3_perm_0, x = var_370)[name = tensor("transpose_50")]; - tensor q_3 = transpose(perm = q_3_perm_0, x = var_362)[name = tensor("transpose_51")]; + tensor k_3 = transpose(perm = k_3_perm_0, x = var_406)[name = tensor("transpose_50")]; + tensor q_3 = transpose(perm = q_3_perm_0, x = var_398)[name = tensor("transpose_51")]; tensor qk_3 = matmul(transpose_x = qk_3_transpose_x_1, transpose_y = qk_3_transpose_y_1, x = q_3, y = k_3)[name = tensor("qk_3")]; - tensor var_386 = const()[name = tensor("op_386"), val = tensor([2, 1])]; - tensor var_387 = reshape(shape = var_386, x = sqrt_s_t_3)[name = tensor("op_387")]; - tensor M_3 = real_div(x = encoder__causal_mask, y = var_387)[name = tensor("M_3")]; - tensor var_389 = mul(x = qk_3, y = M_3)[name = tensor("op_389")]; + tensor var_422 = const()[name = tensor("op_422"), val = tensor([2, 1])]; + tensor var_423 = reshape(shape = var_422, x = sqrt_s_t_3)[name = tensor("op_423")]; + tensor M_3 = real_div(x = inner_encoder__causal_mask, y = var_423)[name = tensor("M_3")]; + tensor var_425 = mul(x = qk_3, y = M_3)[name = tensor("op_425")]; tensor inner_3_transpose_x_0 = const()[name = tensor("inner_3_transpose_x_0"), val = tensor(false)]; tensor inner_3_transpose_y_0 = const()[name = tensor("inner_3_transpose_y_0"), val = tensor(false)]; - tensor v_3 = transpose(perm = v_3_perm_0, x = var_376)[name = tensor("transpose_49")]; - tensor inner_3 = matmul(transpose_x = inner_3_transpose_x_0, transpose_y = inner_3_transpose_y_0, x = var_389, y = v_3)[name = tensor("inner_3")]; - tensor var_391_transpose_x_0 = const()[name = tensor("op_391_transpose_x_0"), val = tensor(false)]; - tensor var_391_transpose_y_0 = const()[name = tensor("op_391_transpose_y_0"), val = tensor(false)]; - tensor var_391 = matmul(transpose_x = var_391_transpose_x_0, transpose_y = var_391_transpose_y_0, x = q_3, y = prev_kv_3)[name = tensor("op_391")]; - tensor var_392 = real_div(x = sqrt_s0_3, y = sqrt_s_t_3)[name = tensor("op_392")]; - tensor var_393 = const()[name = tensor("op_393"), val = tensor([1, 1, 2, 1])]; - tensor var_394 = reshape(shape = var_393, x = var_392)[name = tensor("op_394")]; - tensor cross_3 = mul(x = var_391, y = var_394)[name = tensor("cross_3")]; + tensor v_3 = transpose(perm = v_3_perm_0, x = var_412)[name = tensor("transpose_49")]; + tensor inner_3 = matmul(transpose_x = inner_3_transpose_x_0, transpose_y = inner_3_transpose_y_0, x = var_425, y = v_3)[name = tensor("inner_3")]; + tensor var_427_transpose_x_0 = const()[name = tensor("op_427_transpose_x_0"), val = tensor(false)]; + tensor var_427_transpose_y_0 = const()[name = tensor("op_427_transpose_y_0"), val = tensor(false)]; + tensor var_427 = matmul(transpose_x = var_427_transpose_x_0, transpose_y = var_427_transpose_y_0, x = q_3, y = prev_kv_3)[name = tensor("op_427")]; + tensor var_428 = real_div(x = sqrt_s0_3, y = sqrt_s_t_3)[name = tensor("op_428")]; + tensor var_429 = const()[name = tensor("op_429"), val = tensor([1, 1, 2, 1])]; + tensor var_430 = reshape(shape = var_429, x = var_428)[name = tensor("op_430")]; + tensor cross_3 = mul(x = var_427, y = var_430)[name = tensor("cross_3")]; tensor out_7 = add(x = inner_3, y = cross_3)[name = tensor("out_7")]; - tensor var_397 = mul(x = prev_kv_3, y = sqrt_s0_3)[name = tensor("op_397")]; - tensor var_399_transpose_x_1 = const()[name = tensor("op_399_transpose_x_1"), val = tensor(true)]; - tensor var_399_transpose_y_1 = const()[name = tensor("op_399_transpose_y_1"), val = tensor(false)]; - tensor var_399 = matmul(transpose_x = var_399_transpose_x_1, transpose_y = var_399_transpose_y_1, x = k_3, y = v_3)[name = tensor("op_399")]; - tensor new_kv_unnorm_3 = add(x = var_397, y = var_399)[name = tensor("new_kv_unnorm_3")]; - tensor var_401 = const()[name = tensor("op_401"), val = tensor(0x1p+1)]; - tensor new_scale_3 = add(x = prev_scale_3, y = var_401)[name = tensor("new_scale_3")]; - tensor var_403 = sqrt(x = new_scale_3)[name = tensor("op_403")]; - tensor var_404 = real_div(x = new_kv_unnorm_3, y = var_403)[name = tensor("op_404")]; - tensor var_405_perm_0 = const()[name = tensor("op_405_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_433 = mul(x = prev_kv_3, y = sqrt_s0_3)[name = tensor("op_433")]; + tensor var_435_transpose_x_1 = const()[name = tensor("op_435_transpose_x_1"), val = tensor(true)]; + tensor var_435_transpose_y_1 = const()[name = tensor("op_435_transpose_y_1"), val = tensor(false)]; + tensor var_435 = matmul(transpose_x = var_435_transpose_x_1, transpose_y = var_435_transpose_y_1, x = k_3, y = v_3)[name = tensor("op_435")]; + tensor new_kv_unnorm_3 = add(x = var_433, y = var_435)[name = tensor("new_kv_unnorm_3")]; + tensor var_437 = const()[name = tensor("op_437"), val = tensor(0x1p+1)]; + tensor new_scale_3 = add(x = prev_scale_3, y = var_437)[name = tensor("new_scale_3")]; + tensor var_439 = sqrt(x = new_scale_3)[name = tensor("op_439")]; + tensor var_440 = real_div(x = new_kv_unnorm_3, y = var_439)[name = tensor("op_440")]; + tensor var_441_perm_0 = const()[name = tensor("op_441_perm_0"), val = tensor([0, 2, 1, 3])]; tensor out_9_axes_0 = const()[name = tensor("out_9_axes_0"), val = tensor([-1])]; - tensor var_405 = transpose(perm = var_405_perm_0, x = out_7)[name = tensor("transpose_48")]; - tensor out_9 = layer_norm(axes = out_9_axes_0, epsilon = var_17, x = var_405)[name = tensor("out_9")]; - tensor var_409 = const()[name = tensor("op_409"), val = tensor([1, 2, 256])]; - tensor out_11 = reshape(shape = var_409, x = out_9)[name = tensor("out_11")]; - tensor var_411 = silu(x = input_57)[name = tensor("op_411")]; - tensor input_59 = mul(x = var_411, y = out_11)[name = tensor("input_59")]; - tensor ret_out_3 = linear(bias = encoder_out_proj_1_bias, weight = encoder_out_proj_1_weight, x = input_59)[name = tensor("linear_16")]; - tensor x_9 = add(x = input_53, y = ret_out_3)[name = tensor("x_9")]; + tensor var_441 = transpose(perm = var_441_perm_0, x = out_7)[name = tensor("transpose_48")]; + tensor out_9 = layer_norm(axes = out_9_axes_0, epsilon = var_54, x = var_441)[name = tensor("out_9")]; + tensor var_445 = const()[name = tensor("op_445"), val = tensor([1, 2, 256])]; + tensor out_11 = reshape(shape = var_445, x = out_9)[name = tensor("out_11")]; + tensor var_447 = silu(x = input_59)[name = tensor("op_447")]; + tensor input_61 = mul(x = var_447, y = out_11)[name = tensor("input_61")]; + tensor ret_out_3 = linear(bias = inner_encoder_out_proj_1_bias, weight = inner_encoder_out_proj_1_weight, x = input_61)[name = tensor("linear_16")]; + tensor x_9 = add(x = input_55, y = ret_out_3)[name = tensor("x_9")]; tensor window_7_begin_0 = const()[name = tensor("window_7_begin_0"), val = tensor([1, 0, 0, 0])]; tensor window_7_end_0 = const()[name = tensor("window_7_end_0"), val = tensor([2, 1, 16, 256])]; tensor window_7_end_mask_0 = const()[name = tensor("window_7_end_mask_0"), val = tensor([false, true, true, true])]; tensor window_7_squeeze_mask_0 = const()[name = tensor("window_7_squeeze_mask_0"), val = tensor([true, false, false, false])]; tensor window_7 = slice_by_index(begin = window_7_begin_0, end = window_7_end_0, end_mask = window_7_end_mask_0, squeeze_mask = window_7_squeeze_mask_0, x = enc_conv_cache)[name = tensor("window_7")]; - tensor var_419_begin_0 = const()[name = tensor("op_419_begin_0"), val = tensor([0, 0, 0])]; - tensor var_419_end_0 = const()[name = tensor("op_419_end_0"), val = tensor([1, 1, 256])]; - tensor var_419_end_mask_0 = const()[name = tensor("op_419_end_mask_0"), val = tensor([true, false, true])]; - tensor var_419 = slice_by_index(begin = var_419_begin_0, end = var_419_end_0, end_mask = var_419_end_mask_0, x = x_9)[name = tensor("op_419")]; - tensor var_422_begin_0 = const()[name = tensor("op_422_begin_0"), val = tensor([0, 1, 0])]; - tensor var_422_end_0 = const()[name = tensor("op_422_end_0"), val = tensor([1, 16, 256])]; - tensor var_422_end_mask_0 = const()[name = tensor("op_422_end_mask_0"), val = tensor([true, true, true])]; - tensor var_422 = slice_by_index(begin = var_422_begin_0, end = var_422_end_0, end_mask = var_422_end_mask_0, x = window_7)[name = tensor("op_422")]; + tensor var_455_begin_0 = const()[name = tensor("op_455_begin_0"), val = tensor([0, 0, 0])]; + tensor var_455_end_0 = const()[name = tensor("op_455_end_0"), val = tensor([1, 1, 256])]; + tensor var_455_end_mask_0 = const()[name = tensor("op_455_end_mask_0"), val = tensor([true, false, true])]; + tensor var_455 = slice_by_index(begin = var_455_begin_0, end = var_455_end_0, end_mask = var_455_end_mask_0, x = x_9)[name = tensor("op_455")]; + tensor var_458_begin_0 = const()[name = tensor("op_458_begin_0"), val = tensor([0, 1, 0])]; + tensor var_458_end_0 = const()[name = tensor("op_458_end_0"), val = tensor([1, 16, 256])]; + tensor var_458_end_mask_0 = const()[name = tensor("op_458_end_mask_0"), val = tensor([true, true, true])]; + tensor var_458 = slice_by_index(begin = var_458_begin_0, end = var_458_end_0, end_mask = var_458_end_mask_0, x = window_7)[name = tensor("op_458")]; tensor window_9_interleave_0 = const()[name = tensor("window_9_interleave_0"), val = tensor(false)]; - tensor window_9 = concat(axis = var_26, interleave = window_9_interleave_0, values = (var_422, var_419))[name = tensor("window_9")]; - tensor var_427_begin_0 = const()[name = tensor("op_427_begin_0"), val = tensor([0, 1, 0])]; - tensor var_427_end_0 = const()[name = tensor("op_427_end_0"), val = tensor([1, 1, 256])]; - tensor var_427_end_mask_0 = const()[name = tensor("op_427_end_mask_0"), val = tensor([true, true, true])]; - tensor var_427 = slice_by_index(begin = var_427_begin_0, end = var_427_end_0, end_mask = var_427_end_mask_0, x = x_9)[name = tensor("op_427")]; - tensor var_430_begin_0 = const()[name = tensor("op_430_begin_0"), val = tensor([0, 1, 0])]; - tensor var_430_end_0 = const()[name = tensor("op_430_end_0"), val = tensor([1, 16, 256])]; - tensor var_430_end_mask_0 = const()[name = tensor("op_430_end_mask_0"), val = tensor([true, true, true])]; - tensor var_430 = slice_by_index(begin = var_430_begin_0, end = var_430_end_0, end_mask = var_430_end_mask_0, x = window_9)[name = tensor("op_430")]; + tensor window_9 = concat(axis = var_62, interleave = window_9_interleave_0, values = (var_458, var_455))[name = tensor("window_9")]; + tensor var_463_begin_0 = const()[name = tensor("op_463_begin_0"), val = tensor([0, 1, 0])]; + tensor var_463_end_0 = const()[name = tensor("op_463_end_0"), val = tensor([1, 1, 256])]; + tensor var_463_end_mask_0 = const()[name = tensor("op_463_end_mask_0"), val = tensor([true, true, true])]; + tensor var_463 = slice_by_index(begin = var_463_begin_0, end = var_463_end_0, end_mask = var_463_end_mask_0, x = x_9)[name = tensor("op_463")]; + tensor var_466_begin_0 = const()[name = tensor("op_466_begin_0"), val = tensor([0, 1, 0])]; + tensor var_466_end_0 = const()[name = tensor("op_466_end_0"), val = tensor([1, 16, 256])]; + tensor var_466_end_mask_0 = const()[name = tensor("op_466_end_mask_0"), val = tensor([true, true, true])]; + tensor var_466 = slice_by_index(begin = var_466_begin_0, end = var_466_end_0, end_mask = var_466_end_mask_0, x = window_9)[name = tensor("op_466")]; tensor window_11_interleave_0 = const()[name = tensor("window_11_interleave_0"), val = tensor(false)]; - tensor window_11 = concat(axis = var_26, interleave = window_11_interleave_0, values = (var_430, var_427))[name = tensor("window_11")]; - tensor input_61_interleave_0 = const()[name = tensor("input_61_interleave_0"), val = tensor(false)]; - tensor input_61 = concat(axis = var_23, interleave = input_61_interleave_0, values = (window_9, window_11))[name = tensor("input_61")]; + tensor window_11 = concat(axis = var_62, interleave = window_11_interleave_0, values = (var_466, var_463))[name = tensor("window_11")]; + tensor input_63_interleave_0 = const()[name = tensor("input_63_interleave_0"), val = tensor(false)]; + tensor input_63 = concat(axis = var_49, interleave = input_63_interleave_0, values = (window_9, window_11))[name = tensor("input_63")]; tensor x_11_axes_0 = const()[name = tensor("x_11_axes_0"), val = tensor([-1])]; - tensor x_11 = layer_norm(axes = x_11_axes_0, beta = encoder_conv_module_1_sequential_0_bias, epsilon = var_28, gamma = encoder_conv_module_1_sequential_0_weight, x = input_61)[name = tensor("x_11")]; - tensor input_63_perm_0 = const()[name = tensor("input_63_perm_0"), val = tensor([0, 2, 1])]; + tensor x_11 = layer_norm(axes = x_11_axes_0, beta = inner_encoder_conv_module_1_sequential_0_bias, epsilon = var_46, gamma = inner_encoder_conv_module_1_sequential_0_weight, x = input_63)[name = tensor("x_11")]; + tensor input_65_perm_0 = const()[name = tensor("input_65_perm_0"), val = tensor([0, 2, 1])]; tensor inputs_13_pad_type_0 = const()[name = tensor("inputs_13_pad_type_0"), val = tensor("valid")]; tensor inputs_13_strides_0 = const()[name = tensor("inputs_13_strides_0"), val = tensor([1])]; tensor inputs_13_pad_0 = const()[name = tensor("inputs_13_pad_0"), val = tensor([0, 0])]; tensor inputs_13_dilations_0 = const()[name = tensor("inputs_13_dilations_0"), val = tensor([1])]; tensor inputs_13_groups_0 = const()[name = tensor("inputs_13_groups_0"), val = tensor(1)]; - tensor input_63 = transpose(perm = input_63_perm_0, x = x_11)[name = tensor("transpose_47")]; - tensor inputs_13 = conv(bias = encoder_conv_module_1_sequential_2_conv_bias, dilations = inputs_13_dilations_0, groups = inputs_13_groups_0, pad = inputs_13_pad_0, pad_type = inputs_13_pad_type_0, strides = inputs_13_strides_0, weight = encoder_conv_module_1_sequential_2_conv_weight, x = input_63)[name = tensor("inputs_13")]; - tensor var_455_split_sizes_0 = const()[name = tensor("op_455_split_sizes_0"), val = tensor([256, 256])]; - tensor var_455_axis_0 = const()[name = tensor("op_455_axis_0"), val = tensor(1)]; - tensor var_455_0, tensor var_455_1 = split(axis = var_455_axis_0, split_sizes = var_455_split_sizes_0, x = inputs_13)[name = tensor("op_455")]; - tensor var_457 = sigmoid(x = var_455_1)[name = tensor("op_457")]; - tensor inputs_15 = mul(x = var_455_0, y = var_457)[name = tensor("inputs_15")]; + tensor input_65 = transpose(perm = input_65_perm_0, x = x_11)[name = tensor("transpose_47")]; + tensor inputs_13 = conv(bias = inner_encoder_conv_module_1_sequential_2_conv_bias, dilations = inputs_13_dilations_0, groups = inputs_13_groups_0, pad = inputs_13_pad_0, pad_type = inputs_13_pad_type_0, strides = inputs_13_strides_0, weight = inner_encoder_conv_module_1_sequential_2_conv_weight, x = input_65)[name = tensor("inputs_13")]; + tensor var_491_split_sizes_0 = const()[name = tensor("op_491_split_sizes_0"), val = tensor([256, 256])]; + tensor var_491_axis_0 = const()[name = tensor("op_491_axis_0"), val = tensor(1)]; + tensor var_491_0, tensor var_491_1 = split(axis = var_491_axis_0, split_sizes = var_491_split_sizes_0, x = inputs_13)[name = tensor("op_491")]; + tensor var_493 = sigmoid(x = var_491_1)[name = tensor("op_493")]; + tensor inputs_15 = mul(x = var_491_0, y = var_493)[name = tensor("inputs_15")]; tensor outputs_aug_3_pad_type_0 = const()[name = tensor("outputs_aug_3_pad_type_0"), val = tensor("custom")]; tensor outputs_aug_3_pad_0 = const()[name = tensor("outputs_aug_3_pad_0"), val = tensor([15, 15])]; tensor outputs_aug_3_groups_0 = const()[name = tensor("outputs_aug_3_groups_0"), val = tensor(256)]; tensor outputs_aug_3_strides_0 = const()[name = tensor("outputs_aug_3_strides_0"), val = tensor([1])]; tensor outputs_aug_3_dilations_0 = const()[name = tensor("outputs_aug_3_dilations_0"), val = tensor([1])]; - tensor outputs_aug_3 = conv(dilations = outputs_aug_3_dilations_0, groups = outputs_aug_3_groups_0, pad = outputs_aug_3_pad_0, pad_type = outputs_aug_3_pad_type_0, strides = outputs_aug_3_strides_0, weight = encoder_conv_module_1_sequential_4_conv_weight, x = inputs_15)[name = tensor("outputs_aug_3")]; - tensor input_65_begin_0 = const()[name = tensor("input_65_begin_0"), val = tensor([0, 0, 0])]; - tensor input_65_end_0 = const()[name = tensor("input_65_end_0"), val = tensor([2, 256, 16])]; - tensor input_65_end_mask_0 = const()[name = tensor("input_65_end_mask_0"), val = tensor([true, true, false])]; - tensor input_65 = slice_by_index(begin = input_65_begin_0, end = input_65_end_0, end_mask = input_65_end_mask_0, x = outputs_aug_3)[name = tensor("input_65")]; - tensor inputs_17 = batch_norm(beta = encoder_conv_module_1_sequential_5_bias, epsilon = var_28, gamma = encoder_conv_module_1_sequential_5_weight, mean = encoder_conv_module_1_sequential_5_running_mean, variance = encoder_conv_module_1_sequential_5_running_var, x = input_65)[name = tensor("inputs_17")]; - tensor input_67 = silu(x = inputs_17)[name = tensor("input_67")]; - tensor input_69_pad_type_0 = const()[name = tensor("input_69_pad_type_0"), val = tensor("valid")]; - tensor input_69_strides_0 = const()[name = tensor("input_69_strides_0"), val = tensor([1])]; - tensor input_69_pad_0 = const()[name = tensor("input_69_pad_0"), val = tensor([0, 0])]; - tensor input_69_dilations_0 = const()[name = tensor("input_69_dilations_0"), val = tensor([1])]; - tensor input_69_groups_0 = const()[name = tensor("input_69_groups_0"), val = tensor(1)]; - tensor input_69 = conv(bias = encoder_conv_module_1_sequential_7_conv_bias, dilations = input_69_dilations_0, groups = input_69_groups_0, pad = input_69_pad_0, pad_type = input_69_pad_type_0, strides = input_69_strides_0, weight = encoder_conv_module_1_sequential_7_conv_weight, x = input_67)[name = tensor("input_69")]; + tensor outputs_aug_3 = conv(dilations = outputs_aug_3_dilations_0, groups = outputs_aug_3_groups_0, pad = outputs_aug_3_pad_0, pad_type = outputs_aug_3_pad_type_0, strides = outputs_aug_3_strides_0, weight = inner_encoder_conv_module_1_sequential_4_conv_weight, x = inputs_15)[name = tensor("outputs_aug_3")]; + tensor input_67_begin_0 = const()[name = tensor("input_67_begin_0"), val = tensor([0, 0, 0])]; + tensor input_67_end_0 = const()[name = tensor("input_67_end_0"), val = tensor([2, 256, 16])]; + tensor input_67_end_mask_0 = const()[name = tensor("input_67_end_mask_0"), val = tensor([true, true, false])]; + tensor input_67 = slice_by_index(begin = input_67_begin_0, end = input_67_end_0, end_mask = input_67_end_mask_0, x = outputs_aug_3)[name = tensor("input_67")]; + tensor inputs_17 = batch_norm(beta = inner_encoder_conv_module_1_sequential_5_bias, epsilon = var_46, gamma = inner_encoder_conv_module_1_sequential_5_weight, mean = inner_encoder_conv_module_1_sequential_5_running_mean, variance = inner_encoder_conv_module_1_sequential_5_running_var, x = input_67)[name = tensor("inputs_17")]; + tensor input_69 = silu(x = inputs_17)[name = tensor("input_69")]; + tensor input_71_pad_type_0 = const()[name = tensor("input_71_pad_type_0"), val = tensor("valid")]; + tensor input_71_strides_0 = const()[name = tensor("input_71_strides_0"), val = tensor([1])]; + tensor input_71_pad_0 = const()[name = tensor("input_71_pad_0"), val = tensor([0, 0])]; + tensor input_71_dilations_0 = const()[name = tensor("input_71_dilations_0"), val = tensor([1])]; + tensor input_71_groups_0 = const()[name = tensor("input_71_groups_0"), val = tensor(1)]; + tensor input_71 = conv(bias = inner_encoder_conv_module_1_sequential_7_conv_bias, dilations = input_71_dilations_0, groups = input_71_groups_0, pad = input_71_pad_0, pad_type = input_71_pad_type_0, strides = input_71_strides_0, weight = inner_encoder_conv_module_1_sequential_7_conv_weight, x = input_69)[name = tensor("input_71")]; tensor conv_out_3_perm_0 = const()[name = tensor("conv_out_3_perm_0"), val = tensor([0, 2, 1])]; - tensor var_488_begin_0 = const()[name = tensor("op_488_begin_0"), val = tensor([0, -1, 0])]; - tensor var_488_end_0 = const()[name = tensor("op_488_end_0"), val = tensor([2, 16, 256])]; - tensor var_488_end_mask_0 = const()[name = tensor("op_488_end_mask_0"), val = tensor([true, true, true])]; - tensor conv_out_3 = transpose(perm = conv_out_3_perm_0, x = input_69)[name = tensor("transpose_46")]; - tensor var_488 = slice_by_index(begin = var_488_begin_0, end = var_488_end_0, end_mask = var_488_end_mask_0, x = conv_out_3)[name = tensor("op_488")]; - tensor var_490_perm_0 = const()[name = tensor("op_490_perm_0"), val = tensor([1, 0, 2])]; - tensor var_490 = transpose(perm = var_490_perm_0, x = var_488)[name = tensor("transpose_45")]; - tensor input_71 = add(x = x_9, y = var_490)[name = tensor("input_71")]; - tensor input_73_axes_0 = const()[name = tensor("input_73_axes_0"), val = tensor([-1])]; - tensor input_73 = layer_norm(axes = input_73_axes_0, beta = encoder_ffn2_1_module_sequential_0_bias, epsilon = var_28, gamma = encoder_ffn2_1_module_sequential_0_weight, x = input_71)[name = tensor("input_73")]; - tensor inputs_19 = linear(bias = encoder_ffn2_1_module_sequential_1_linear_bias, weight = encoder_ffn2_1_module_sequential_1_linear_weight, x = input_73)[name = tensor("linear_17")]; - tensor input_75 = silu(x = inputs_19)[name = tensor("input_75")]; - tensor input_79 = linear(bias = encoder_ffn2_1_module_sequential_4_linear_bias, weight = encoder_ffn2_1_module_sequential_4_linear_weight, x = input_75)[name = tensor("linear_18")]; - tensor var_513 = const()[name = tensor("op_513"), val = tensor(0x1p-1)]; - tensor var_514 = mul(x = input_79, y = var_513)[name = tensor("op_514")]; - tensor input_81 = add(x = var_514, y = input_71)[name = tensor("input_81")]; - tensor input_83_axes_0 = const()[name = tensor("input_83_axes_0"), val = tensor([-1])]; - tensor input_83 = layer_norm(axes = input_83_axes_0, beta = encoder_layer_norm_1_bias, epsilon = var_28, gamma = encoder_layer_norm_1_weight, x = input_81)[name = tensor("input_83")]; + tensor var_524_begin_0 = const()[name = tensor("op_524_begin_0"), val = tensor([0, -1, 0])]; + tensor var_524_end_0 = const()[name = tensor("op_524_end_0"), val = tensor([2, 16, 256])]; + tensor var_524_end_mask_0 = const()[name = tensor("op_524_end_mask_0"), val = tensor([true, true, true])]; + tensor conv_out_3 = transpose(perm = conv_out_3_perm_0, x = input_71)[name = tensor("transpose_46")]; + tensor var_524 = slice_by_index(begin = var_524_begin_0, end = var_524_end_0, end_mask = var_524_end_mask_0, x = conv_out_3)[name = tensor("op_524")]; + tensor var_526_perm_0 = const()[name = tensor("op_526_perm_0"), val = tensor([1, 0, 2])]; + tensor var_526 = transpose(perm = var_526_perm_0, x = var_524)[name = tensor("transpose_45")]; + tensor input_73 = add(x = x_9, y = var_526)[name = tensor("input_73")]; + tensor input_75_axes_0 = const()[name = tensor("input_75_axes_0"), val = tensor([-1])]; + tensor input_75 = layer_norm(axes = input_75_axes_0, beta = inner_encoder_ffn2_1_module_sequential_0_bias, epsilon = var_46, gamma = inner_encoder_ffn2_1_module_sequential_0_weight, x = input_73)[name = tensor("input_75")]; + tensor inputs_19 = linear(bias = inner_encoder_ffn2_1_module_sequential_1_linear_bias, weight = inner_encoder_ffn2_1_module_sequential_1_linear_weight, x = input_75)[name = tensor("linear_17")]; + tensor input_77 = silu(x = inputs_19)[name = tensor("input_77")]; + tensor input_81 = linear(bias = inner_encoder_ffn2_1_module_sequential_4_linear_bias, weight = inner_encoder_ffn2_1_module_sequential_4_linear_weight, x = input_77)[name = tensor("linear_18")]; + tensor var_549 = const()[name = tensor("op_549"), val = tensor(0x1p-1)]; + tensor var_550 = mul(x = input_81, y = var_549)[name = tensor("op_550")]; + tensor input_83 = add(x = var_550, y = input_73)[name = tensor("input_83")]; tensor input_85_axes_0 = const()[name = tensor("input_85_axes_0"), val = tensor([-1])]; - tensor input_85 = layer_norm(axes = input_85_axes_0, beta = encoder_ffn1_2_module_sequential_0_bias, epsilon = var_28, gamma = encoder_ffn1_2_module_sequential_0_weight, x = input_83)[name = tensor("input_85")]; - tensor inputs_21 = linear(bias = encoder_ffn1_2_module_sequential_1_linear_bias, weight = encoder_ffn1_2_module_sequential_1_linear_weight, x = input_85)[name = tensor("linear_19")]; - tensor input_87 = silu(x = inputs_21)[name = tensor("input_87")]; - tensor input_91 = linear(bias = encoder_ffn1_2_module_sequential_4_linear_bias, weight = encoder_ffn1_2_module_sequential_4_linear_weight, x = input_87)[name = tensor("linear_20")]; - tensor var_543 = const()[name = tensor("op_543"), val = tensor(0x1p-1)]; - tensor var_544 = mul(x = input_91, y = var_543)[name = tensor("op_544")]; - tensor input_93 = add(x = var_544, y = input_83)[name = tensor("input_93")]; + tensor input_85 = layer_norm(axes = input_85_axes_0, beta = inner_encoder_layer_norm_1_bias, epsilon = var_46, gamma = inner_encoder_layer_norm_1_weight, x = input_83)[name = tensor("input_85")]; + tensor input_87_axes_0 = const()[name = tensor("input_87_axes_0"), val = tensor([-1])]; + tensor input_87 = layer_norm(axes = input_87_axes_0, beta = inner_encoder_ffn1_2_module_sequential_0_bias, epsilon = var_46, gamma = inner_encoder_ffn1_2_module_sequential_0_weight, x = input_85)[name = tensor("input_87")]; + tensor inputs_21 = linear(bias = inner_encoder_ffn1_2_module_sequential_1_linear_bias, weight = inner_encoder_ffn1_2_module_sequential_1_linear_weight, x = input_87)[name = tensor("linear_19")]; + tensor input_89 = silu(x = inputs_21)[name = tensor("input_89")]; + tensor input_93 = linear(bias = inner_encoder_ffn1_2_module_sequential_4_linear_bias, weight = inner_encoder_ffn1_2_module_sequential_4_linear_weight, x = input_89)[name = tensor("linear_20")]; + tensor var_579 = const()[name = tensor("op_579"), val = tensor(0x1p-1)]; + tensor var_580 = mul(x = input_93, y = var_579)[name = tensor("op_580")]; + tensor input_95 = add(x = var_580, y = input_85)[name = tensor("input_95")]; tensor x_13_axes_0 = const()[name = tensor("x_13_axes_0"), val = tensor([-1])]; - tensor x_13 = layer_norm(axes = x_13_axes_0, beta = encoder_ret_lns_2_bias, epsilon = var_28, gamma = encoder_ret_lns_2_weight, x = input_93)[name = tensor("x_13")]; + tensor x_13 = layer_norm(axes = x_13_axes_0, beta = inner_encoder_ret_lns_2_bias, epsilon = var_46, gamma = inner_encoder_ret_lns_2_weight, x = input_95)[name = tensor("x_13")]; tensor prev_kv_5_begin_0 = const()[name = tensor("prev_kv_5_begin_0"), val = tensor([2, 0, 0, 0, 0])]; tensor prev_kv_5_end_0 = const()[name = tensor("prev_kv_5_end_0"), val = tensor([3, 1, 4, 64, 64])]; tensor prev_kv_5_end_mask_0 = const()[name = tensor("prev_kv_5_end_mask_0"), val = tensor([false, true, true, true, true])]; @@ -553,153 +567,153 @@ program(1.0) tensor prev_scale_5_end_mask_0 = const()[name = tensor("prev_scale_5_end_mask_0"), val = tensor([false, true])]; tensor prev_scale_5_squeeze_mask_0 = const()[name = tensor("prev_scale_5_squeeze_mask_0"), val = tensor([true, false])]; tensor prev_scale_5 = slice_by_index(begin = prev_scale_5_begin_0, end = prev_scale_5_end_0, end_mask = prev_scale_5_end_mask_0, squeeze_mask = prev_scale_5_squeeze_mask_0, x = enc_scale)[name = tensor("prev_scale_5")]; - tensor var_558 = linear(bias = encoder_q_proj_2_bias, weight = encoder_q_proj_2_weight, x = x_13)[name = tensor("linear_21")]; - tensor var_559 = const()[name = tensor("op_559"), val = tensor([1, 2, 4, 64])]; - tensor var_560 = reshape(shape = var_559, x = var_558)[name = tensor("op_560")]; + tensor var_594 = linear(bias = inner_encoder_q_proj_2_bias, weight = inner_encoder_q_proj_2_weight, x = x_13)[name = tensor("linear_21")]; + tensor var_595 = const()[name = tensor("op_595"), val = tensor([1, 2, 4, 64])]; + tensor var_596 = reshape(shape = var_595, x = var_594)[name = tensor("op_596")]; tensor q_5_perm_0 = const()[name = tensor("q_5_perm_0"), val = tensor([0, 2, 1, 3])]; - tensor var_564 = linear(bias = encoder_k_proj_2_bias, weight = encoder_k_proj_2_weight, x = x_13)[name = tensor("linear_22")]; - tensor var_565 = const()[name = tensor("op_565"), val = tensor(0x1p-3)]; - tensor var_566 = mul(x = var_564, y = var_565)[name = tensor("op_566")]; - tensor var_567 = const()[name = tensor("op_567"), val = tensor([1, 2, 4, 64])]; - tensor var_568 = reshape(shape = var_567, x = var_566)[name = tensor("op_568")]; + tensor var_600 = linear(bias = inner_encoder_k_proj_2_bias, weight = inner_encoder_k_proj_2_weight, x = x_13)[name = tensor("linear_22")]; + tensor var_601 = const()[name = tensor("op_601"), val = tensor(0x1p-3)]; + tensor var_602 = mul(x = var_600, y = var_601)[name = tensor("op_602")]; + tensor var_603 = const()[name = tensor("op_603"), val = tensor([1, 2, 4, 64])]; + tensor var_604 = reshape(shape = var_603, x = var_602)[name = tensor("op_604")]; tensor k_5_perm_0 = const()[name = tensor("k_5_perm_0"), val = tensor([0, 2, 1, 3])]; - tensor var_572 = linear(bias = encoder_v_proj_2_bias, weight = encoder_v_proj_2_weight, x = x_13)[name = tensor("linear_23")]; - tensor var_573 = const()[name = tensor("op_573"), val = tensor([1, 2, 4, 64])]; - tensor var_574 = reshape(shape = var_573, x = var_572)[name = tensor("op_574")]; + tensor var_608 = linear(bias = inner_encoder_v_proj_2_bias, weight = inner_encoder_v_proj_2_weight, x = x_13)[name = tensor("linear_23")]; + tensor var_609 = const()[name = tensor("op_609"), val = tensor([1, 2, 4, 64])]; + tensor var_610 = reshape(shape = var_609, x = var_608)[name = tensor("op_610")]; tensor v_5_perm_0 = const()[name = tensor("v_5_perm_0"), val = tensor([0, 2, 1, 3])]; - tensor input_97 = linear(bias = encoder_g_proj_2_bias, weight = encoder_g_proj_2_weight, x = x_13)[name = tensor("linear_24")]; + tensor input_99 = linear(bias = inner_encoder_g_proj_2_bias, weight = inner_encoder_g_proj_2_weight, x = x_13)[name = tensor("linear_24")]; tensor sqrt_s0_5 = sqrt(x = prev_scale_5)[name = tensor("sqrt_s0_5")]; - tensor s_t_5 = add(x = prev_scale_5, y = encoder__t_index)[name = tensor("s_t_5")]; + tensor s_t_5 = add(x = prev_scale_5, y = inner_encoder__t_index)[name = tensor("s_t_5")]; tensor sqrt_s_t_5 = sqrt(x = s_t_5)[name = tensor("sqrt_s_t_5")]; tensor qk_5_transpose_x_1 = const()[name = tensor("qk_5_transpose_x_1"), val = tensor(false)]; tensor qk_5_transpose_y_1 = const()[name = tensor("qk_5_transpose_y_1"), val = tensor(true)]; - tensor k_5 = transpose(perm = k_5_perm_0, x = var_568)[name = tensor("transpose_43")]; - tensor q_5 = transpose(perm = q_5_perm_0, x = var_560)[name = tensor("transpose_44")]; + tensor k_5 = transpose(perm = k_5_perm_0, x = var_604)[name = tensor("transpose_43")]; + tensor q_5 = transpose(perm = q_5_perm_0, x = var_596)[name = tensor("transpose_44")]; tensor qk_5 = matmul(transpose_x = qk_5_transpose_x_1, transpose_y = qk_5_transpose_y_1, x = q_5, y = k_5)[name = tensor("qk_5")]; - tensor var_584 = const()[name = tensor("op_584"), val = tensor([2, 1])]; - tensor var_585 = reshape(shape = var_584, x = sqrt_s_t_5)[name = tensor("op_585")]; - tensor M_5 = real_div(x = encoder__causal_mask, y = var_585)[name = tensor("M_5")]; - tensor var_587 = mul(x = qk_5, y = M_5)[name = tensor("op_587")]; + tensor var_620 = const()[name = tensor("op_620"), val = tensor([2, 1])]; + tensor var_621 = reshape(shape = var_620, x = sqrt_s_t_5)[name = tensor("op_621")]; + tensor M_5 = real_div(x = inner_encoder__causal_mask, y = var_621)[name = tensor("M_5")]; + tensor var_623 = mul(x = qk_5, y = M_5)[name = tensor("op_623")]; tensor inner_5_transpose_x_0 = const()[name = tensor("inner_5_transpose_x_0"), val = tensor(false)]; tensor inner_5_transpose_y_0 = const()[name = tensor("inner_5_transpose_y_0"), val = tensor(false)]; - tensor v_5 = transpose(perm = v_5_perm_0, x = var_574)[name = tensor("transpose_42")]; - tensor inner_5 = matmul(transpose_x = inner_5_transpose_x_0, transpose_y = inner_5_transpose_y_0, x = var_587, y = v_5)[name = tensor("inner_5")]; - tensor var_589_transpose_x_0 = const()[name = tensor("op_589_transpose_x_0"), val = tensor(false)]; - tensor var_589_transpose_y_0 = const()[name = tensor("op_589_transpose_y_0"), val = tensor(false)]; - tensor var_589 = matmul(transpose_x = var_589_transpose_x_0, transpose_y = var_589_transpose_y_0, x = q_5, y = prev_kv_5)[name = tensor("op_589")]; - tensor var_590 = real_div(x = sqrt_s0_5, y = sqrt_s_t_5)[name = tensor("op_590")]; - tensor var_591 = const()[name = tensor("op_591"), val = tensor([1, 1, 2, 1])]; - tensor var_592 = reshape(shape = var_591, x = var_590)[name = tensor("op_592")]; - tensor cross_5 = mul(x = var_589, y = var_592)[name = tensor("cross_5")]; + tensor v_5 = transpose(perm = v_5_perm_0, x = var_610)[name = tensor("transpose_42")]; + tensor inner_5 = matmul(transpose_x = inner_5_transpose_x_0, transpose_y = inner_5_transpose_y_0, x = var_623, y = v_5)[name = tensor("inner_5")]; + tensor var_625_transpose_x_0 = const()[name = tensor("op_625_transpose_x_0"), val = tensor(false)]; + tensor var_625_transpose_y_0 = const()[name = tensor("op_625_transpose_y_0"), val = tensor(false)]; + tensor var_625 = matmul(transpose_x = var_625_transpose_x_0, transpose_y = var_625_transpose_y_0, x = q_5, y = prev_kv_5)[name = tensor("op_625")]; + tensor var_626 = real_div(x = sqrt_s0_5, y = sqrt_s_t_5)[name = tensor("op_626")]; + tensor var_627 = const()[name = tensor("op_627"), val = tensor([1, 1, 2, 1])]; + tensor var_628 = reshape(shape = var_627, x = var_626)[name = tensor("op_628")]; + tensor cross_5 = mul(x = var_625, y = var_628)[name = tensor("cross_5")]; tensor out_13 = add(x = inner_5, y = cross_5)[name = tensor("out_13")]; - tensor var_595 = mul(x = prev_kv_5, y = sqrt_s0_5)[name = tensor("op_595")]; - tensor var_597_transpose_x_1 = const()[name = tensor("op_597_transpose_x_1"), val = tensor(true)]; - tensor var_597_transpose_y_1 = const()[name = tensor("op_597_transpose_y_1"), val = tensor(false)]; - tensor var_597 = matmul(transpose_x = var_597_transpose_x_1, transpose_y = var_597_transpose_y_1, x = k_5, y = v_5)[name = tensor("op_597")]; - tensor new_kv_unnorm_5 = add(x = var_595, y = var_597)[name = tensor("new_kv_unnorm_5")]; - tensor var_599 = const()[name = tensor("op_599"), val = tensor(0x1p+1)]; - tensor new_scale_5 = add(x = prev_scale_5, y = var_599)[name = tensor("new_scale_5")]; - tensor var_601 = sqrt(x = new_scale_5)[name = tensor("op_601")]; - tensor var_602 = real_div(x = new_kv_unnorm_5, y = var_601)[name = tensor("op_602")]; - tensor var_603_perm_0 = const()[name = tensor("op_603_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_631 = mul(x = prev_kv_5, y = sqrt_s0_5)[name = tensor("op_631")]; + tensor var_633_transpose_x_1 = const()[name = tensor("op_633_transpose_x_1"), val = tensor(true)]; + tensor var_633_transpose_y_1 = const()[name = tensor("op_633_transpose_y_1"), val = tensor(false)]; + tensor var_633 = matmul(transpose_x = var_633_transpose_x_1, transpose_y = var_633_transpose_y_1, x = k_5, y = v_5)[name = tensor("op_633")]; + tensor new_kv_unnorm_5 = add(x = var_631, y = var_633)[name = tensor("new_kv_unnorm_5")]; + tensor var_635 = const()[name = tensor("op_635"), val = tensor(0x1p+1)]; + tensor new_scale_5 = add(x = prev_scale_5, y = var_635)[name = tensor("new_scale_5")]; + tensor var_637 = sqrt(x = new_scale_5)[name = tensor("op_637")]; + tensor var_638 = real_div(x = new_kv_unnorm_5, y = var_637)[name = tensor("op_638")]; + tensor var_639_perm_0 = const()[name = tensor("op_639_perm_0"), val = tensor([0, 2, 1, 3])]; tensor out_15_axes_0 = const()[name = tensor("out_15_axes_0"), val = tensor([-1])]; - tensor var_603 = transpose(perm = var_603_perm_0, x = out_13)[name = tensor("transpose_41")]; - tensor out_15 = layer_norm(axes = out_15_axes_0, epsilon = var_17, x = var_603)[name = tensor("out_15")]; - tensor var_607 = const()[name = tensor("op_607"), val = tensor([1, 2, 256])]; - tensor out_17 = reshape(shape = var_607, x = out_15)[name = tensor("out_17")]; - tensor var_609 = silu(x = input_97)[name = tensor("op_609")]; - tensor input_99 = mul(x = var_609, y = out_17)[name = tensor("input_99")]; - tensor ret_out_5 = linear(bias = encoder_out_proj_2_bias, weight = encoder_out_proj_2_weight, x = input_99)[name = tensor("linear_25")]; - tensor x_15 = add(x = input_93, y = ret_out_5)[name = tensor("x_15")]; + tensor var_639 = transpose(perm = var_639_perm_0, x = out_13)[name = tensor("transpose_41")]; + tensor out_15 = layer_norm(axes = out_15_axes_0, epsilon = var_54, x = var_639)[name = tensor("out_15")]; + tensor var_643 = const()[name = tensor("op_643"), val = tensor([1, 2, 256])]; + tensor out_17 = reshape(shape = var_643, x = out_15)[name = tensor("out_17")]; + tensor var_645 = silu(x = input_99)[name = tensor("op_645")]; + tensor input_101 = mul(x = var_645, y = out_17)[name = tensor("input_101")]; + tensor ret_out_5 = linear(bias = inner_encoder_out_proj_2_bias, weight = inner_encoder_out_proj_2_weight, x = input_101)[name = tensor("linear_25")]; + tensor x_15 = add(x = input_95, y = ret_out_5)[name = tensor("x_15")]; tensor window_13_begin_0 = const()[name = tensor("window_13_begin_0"), val = tensor([2, 0, 0, 0])]; tensor window_13_end_0 = const()[name = tensor("window_13_end_0"), val = tensor([3, 1, 16, 256])]; tensor window_13_end_mask_0 = const()[name = tensor("window_13_end_mask_0"), val = tensor([false, true, true, true])]; tensor window_13_squeeze_mask_0 = const()[name = tensor("window_13_squeeze_mask_0"), val = tensor([true, false, false, false])]; tensor window_13 = slice_by_index(begin = window_13_begin_0, end = window_13_end_0, end_mask = window_13_end_mask_0, squeeze_mask = window_13_squeeze_mask_0, x = enc_conv_cache)[name = tensor("window_13")]; - tensor var_617_begin_0 = const()[name = tensor("op_617_begin_0"), val = tensor([0, 0, 0])]; - tensor var_617_end_0 = const()[name = tensor("op_617_end_0"), val = tensor([1, 1, 256])]; - tensor var_617_end_mask_0 = const()[name = tensor("op_617_end_mask_0"), val = tensor([true, false, true])]; - tensor var_617 = slice_by_index(begin = var_617_begin_0, end = var_617_end_0, end_mask = var_617_end_mask_0, x = x_15)[name = tensor("op_617")]; - tensor var_620_begin_0 = const()[name = tensor("op_620_begin_0"), val = tensor([0, 1, 0])]; - tensor var_620_end_0 = const()[name = tensor("op_620_end_0"), val = tensor([1, 16, 256])]; - tensor var_620_end_mask_0 = const()[name = tensor("op_620_end_mask_0"), val = tensor([true, true, true])]; - tensor var_620 = slice_by_index(begin = var_620_begin_0, end = var_620_end_0, end_mask = var_620_end_mask_0, x = window_13)[name = tensor("op_620")]; + tensor var_653_begin_0 = const()[name = tensor("op_653_begin_0"), val = tensor([0, 0, 0])]; + tensor var_653_end_0 = const()[name = tensor("op_653_end_0"), val = tensor([1, 1, 256])]; + tensor var_653_end_mask_0 = const()[name = tensor("op_653_end_mask_0"), val = tensor([true, false, true])]; + tensor var_653 = slice_by_index(begin = var_653_begin_0, end = var_653_end_0, end_mask = var_653_end_mask_0, x = x_15)[name = tensor("op_653")]; + tensor var_656_begin_0 = const()[name = tensor("op_656_begin_0"), val = tensor([0, 1, 0])]; + tensor var_656_end_0 = const()[name = tensor("op_656_end_0"), val = tensor([1, 16, 256])]; + tensor var_656_end_mask_0 = const()[name = tensor("op_656_end_mask_0"), val = tensor([true, true, true])]; + tensor var_656 = slice_by_index(begin = var_656_begin_0, end = var_656_end_0, end_mask = var_656_end_mask_0, x = window_13)[name = tensor("op_656")]; tensor window_15_interleave_0 = const()[name = tensor("window_15_interleave_0"), val = tensor(false)]; - tensor window_15 = concat(axis = var_26, interleave = window_15_interleave_0, values = (var_620, var_617))[name = tensor("window_15")]; - tensor var_625_begin_0 = const()[name = tensor("op_625_begin_0"), val = tensor([0, 1, 0])]; - tensor var_625_end_0 = const()[name = tensor("op_625_end_0"), val = tensor([1, 1, 256])]; - tensor var_625_end_mask_0 = const()[name = tensor("op_625_end_mask_0"), val = tensor([true, true, true])]; - tensor var_625 = slice_by_index(begin = var_625_begin_0, end = var_625_end_0, end_mask = var_625_end_mask_0, x = x_15)[name = tensor("op_625")]; - tensor var_628_begin_0 = const()[name = tensor("op_628_begin_0"), val = tensor([0, 1, 0])]; - tensor var_628_end_0 = const()[name = tensor("op_628_end_0"), val = tensor([1, 16, 256])]; - tensor var_628_end_mask_0 = const()[name = tensor("op_628_end_mask_0"), val = tensor([true, true, true])]; - tensor var_628 = slice_by_index(begin = var_628_begin_0, end = var_628_end_0, end_mask = var_628_end_mask_0, x = window_15)[name = tensor("op_628")]; + tensor window_15 = concat(axis = var_62, interleave = window_15_interleave_0, values = (var_656, var_653))[name = tensor("window_15")]; + tensor var_661_begin_0 = const()[name = tensor("op_661_begin_0"), val = tensor([0, 1, 0])]; + tensor var_661_end_0 = const()[name = tensor("op_661_end_0"), val = tensor([1, 1, 256])]; + tensor var_661_end_mask_0 = const()[name = tensor("op_661_end_mask_0"), val = tensor([true, true, true])]; + tensor var_661 = slice_by_index(begin = var_661_begin_0, end = var_661_end_0, end_mask = var_661_end_mask_0, x = x_15)[name = tensor("op_661")]; + tensor var_664_begin_0 = const()[name = tensor("op_664_begin_0"), val = tensor([0, 1, 0])]; + tensor var_664_end_0 = const()[name = tensor("op_664_end_0"), val = tensor([1, 16, 256])]; + tensor var_664_end_mask_0 = const()[name = tensor("op_664_end_mask_0"), val = tensor([true, true, true])]; + tensor var_664 = slice_by_index(begin = var_664_begin_0, end = var_664_end_0, end_mask = var_664_end_mask_0, x = window_15)[name = tensor("op_664")]; tensor window_17_interleave_0 = const()[name = tensor("window_17_interleave_0"), val = tensor(false)]; - tensor window_17 = concat(axis = var_26, interleave = window_17_interleave_0, values = (var_628, var_625))[name = tensor("window_17")]; - tensor input_101_interleave_0 = const()[name = tensor("input_101_interleave_0"), val = tensor(false)]; - tensor input_101 = concat(axis = var_23, interleave = input_101_interleave_0, values = (window_15, window_17))[name = tensor("input_101")]; + tensor window_17 = concat(axis = var_62, interleave = window_17_interleave_0, values = (var_664, var_661))[name = tensor("window_17")]; + tensor input_103_interleave_0 = const()[name = tensor("input_103_interleave_0"), val = tensor(false)]; + tensor input_103 = concat(axis = var_49, interleave = input_103_interleave_0, values = (window_15, window_17))[name = tensor("input_103")]; tensor x_17_axes_0 = const()[name = tensor("x_17_axes_0"), val = tensor([-1])]; - tensor x_17 = layer_norm(axes = x_17_axes_0, beta = encoder_conv_module_2_sequential_0_bias, epsilon = var_28, gamma = encoder_conv_module_2_sequential_0_weight, x = input_101)[name = tensor("x_17")]; - tensor input_103_perm_0 = const()[name = tensor("input_103_perm_0"), val = tensor([0, 2, 1])]; + tensor x_17 = layer_norm(axes = x_17_axes_0, beta = inner_encoder_conv_module_2_sequential_0_bias, epsilon = var_46, gamma = inner_encoder_conv_module_2_sequential_0_weight, x = input_103)[name = tensor("x_17")]; + tensor input_105_perm_0 = const()[name = tensor("input_105_perm_0"), val = tensor([0, 2, 1])]; tensor inputs_23_pad_type_0 = const()[name = tensor("inputs_23_pad_type_0"), val = tensor("valid")]; tensor inputs_23_strides_0 = const()[name = tensor("inputs_23_strides_0"), val = tensor([1])]; tensor inputs_23_pad_0 = const()[name = tensor("inputs_23_pad_0"), val = tensor([0, 0])]; tensor inputs_23_dilations_0 = const()[name = tensor("inputs_23_dilations_0"), val = tensor([1])]; tensor inputs_23_groups_0 = const()[name = tensor("inputs_23_groups_0"), val = tensor(1)]; - tensor input_103 = transpose(perm = input_103_perm_0, x = x_17)[name = tensor("transpose_40")]; - tensor inputs_23 = conv(bias = encoder_conv_module_2_sequential_2_conv_bias, dilations = inputs_23_dilations_0, groups = inputs_23_groups_0, pad = inputs_23_pad_0, pad_type = inputs_23_pad_type_0, strides = inputs_23_strides_0, weight = encoder_conv_module_2_sequential_2_conv_weight, x = input_103)[name = tensor("inputs_23")]; - tensor var_653_split_sizes_0 = const()[name = tensor("op_653_split_sizes_0"), val = tensor([256, 256])]; - tensor var_653_axis_0 = const()[name = tensor("op_653_axis_0"), val = tensor(1)]; - tensor var_653_0, tensor var_653_1 = split(axis = var_653_axis_0, split_sizes = var_653_split_sizes_0, x = inputs_23)[name = tensor("op_653")]; - tensor var_655 = sigmoid(x = var_653_1)[name = tensor("op_655")]; - tensor inputs_25 = mul(x = var_653_0, y = var_655)[name = tensor("inputs_25")]; + tensor input_105 = transpose(perm = input_105_perm_0, x = x_17)[name = tensor("transpose_40")]; + tensor inputs_23 = conv(bias = inner_encoder_conv_module_2_sequential_2_conv_bias, dilations = inputs_23_dilations_0, groups = inputs_23_groups_0, pad = inputs_23_pad_0, pad_type = inputs_23_pad_type_0, strides = inputs_23_strides_0, weight = inner_encoder_conv_module_2_sequential_2_conv_weight, x = input_105)[name = tensor("inputs_23")]; + tensor var_689_split_sizes_0 = const()[name = tensor("op_689_split_sizes_0"), val = tensor([256, 256])]; + tensor var_689_axis_0 = const()[name = tensor("op_689_axis_0"), val = tensor(1)]; + tensor var_689_0, tensor var_689_1 = split(axis = var_689_axis_0, split_sizes = var_689_split_sizes_0, x = inputs_23)[name = tensor("op_689")]; + tensor var_691 = sigmoid(x = var_689_1)[name = tensor("op_691")]; + tensor inputs_25 = mul(x = var_689_0, y = var_691)[name = tensor("inputs_25")]; tensor outputs_aug_5_pad_type_0 = const()[name = tensor("outputs_aug_5_pad_type_0"), val = tensor("custom")]; tensor outputs_aug_5_pad_0 = const()[name = tensor("outputs_aug_5_pad_0"), val = tensor([15, 15])]; tensor outputs_aug_5_groups_0 = const()[name = tensor("outputs_aug_5_groups_0"), val = tensor(256)]; tensor outputs_aug_5_strides_0 = const()[name = tensor("outputs_aug_5_strides_0"), val = tensor([1])]; tensor outputs_aug_5_dilations_0 = const()[name = tensor("outputs_aug_5_dilations_0"), val = tensor([1])]; - tensor outputs_aug_5 = conv(dilations = outputs_aug_5_dilations_0, groups = outputs_aug_5_groups_0, pad = outputs_aug_5_pad_0, pad_type = outputs_aug_5_pad_type_0, strides = outputs_aug_5_strides_0, weight = encoder_conv_module_2_sequential_4_conv_weight, x = inputs_25)[name = tensor("outputs_aug_5")]; - tensor input_105_begin_0 = const()[name = tensor("input_105_begin_0"), val = tensor([0, 0, 0])]; - tensor input_105_end_0 = const()[name = tensor("input_105_end_0"), val = tensor([2, 256, 16])]; - tensor input_105_end_mask_0 = const()[name = tensor("input_105_end_mask_0"), val = tensor([true, true, false])]; - tensor input_105 = slice_by_index(begin = input_105_begin_0, end = input_105_end_0, end_mask = input_105_end_mask_0, x = outputs_aug_5)[name = tensor("input_105")]; - tensor inputs_27 = batch_norm(beta = encoder_conv_module_2_sequential_5_bias, epsilon = var_28, gamma = encoder_conv_module_2_sequential_5_weight, mean = encoder_conv_module_2_sequential_5_running_mean, variance = encoder_conv_module_2_sequential_5_running_var, x = input_105)[name = tensor("inputs_27")]; - tensor input_107 = silu(x = inputs_27)[name = tensor("input_107")]; - tensor input_109_pad_type_0 = const()[name = tensor("input_109_pad_type_0"), val = tensor("valid")]; - tensor input_109_strides_0 = const()[name = tensor("input_109_strides_0"), val = tensor([1])]; - tensor input_109_pad_0 = const()[name = tensor("input_109_pad_0"), val = tensor([0, 0])]; - tensor input_109_dilations_0 = const()[name = tensor("input_109_dilations_0"), val = tensor([1])]; - tensor input_109_groups_0 = const()[name = tensor("input_109_groups_0"), val = tensor(1)]; - tensor input_109 = conv(bias = encoder_conv_module_2_sequential_7_conv_bias, dilations = input_109_dilations_0, groups = input_109_groups_0, pad = input_109_pad_0, pad_type = input_109_pad_type_0, strides = input_109_strides_0, weight = encoder_conv_module_2_sequential_7_conv_weight, x = input_107)[name = tensor("input_109")]; + tensor outputs_aug_5 = conv(dilations = outputs_aug_5_dilations_0, groups = outputs_aug_5_groups_0, pad = outputs_aug_5_pad_0, pad_type = outputs_aug_5_pad_type_0, strides = outputs_aug_5_strides_0, weight = inner_encoder_conv_module_2_sequential_4_conv_weight, x = inputs_25)[name = tensor("outputs_aug_5")]; + tensor input_107_begin_0 = const()[name = tensor("input_107_begin_0"), val = tensor([0, 0, 0])]; + tensor input_107_end_0 = const()[name = tensor("input_107_end_0"), val = tensor([2, 256, 16])]; + tensor input_107_end_mask_0 = const()[name = tensor("input_107_end_mask_0"), val = tensor([true, true, false])]; + tensor input_107 = slice_by_index(begin = input_107_begin_0, end = input_107_end_0, end_mask = input_107_end_mask_0, x = outputs_aug_5)[name = tensor("input_107")]; + tensor inputs_27 = batch_norm(beta = inner_encoder_conv_module_2_sequential_5_bias, epsilon = var_46, gamma = inner_encoder_conv_module_2_sequential_5_weight, mean = inner_encoder_conv_module_2_sequential_5_running_mean, variance = inner_encoder_conv_module_2_sequential_5_running_var, x = input_107)[name = tensor("inputs_27")]; + tensor input_109 = silu(x = inputs_27)[name = tensor("input_109")]; + tensor input_111_pad_type_0 = const()[name = tensor("input_111_pad_type_0"), val = tensor("valid")]; + tensor input_111_strides_0 = const()[name = tensor("input_111_strides_0"), val = tensor([1])]; + tensor input_111_pad_0 = const()[name = tensor("input_111_pad_0"), val = tensor([0, 0])]; + tensor input_111_dilations_0 = const()[name = tensor("input_111_dilations_0"), val = tensor([1])]; + tensor input_111_groups_0 = const()[name = tensor("input_111_groups_0"), val = tensor(1)]; + tensor input_111 = conv(bias = inner_encoder_conv_module_2_sequential_7_conv_bias, dilations = input_111_dilations_0, groups = input_111_groups_0, pad = input_111_pad_0, pad_type = input_111_pad_type_0, strides = input_111_strides_0, weight = inner_encoder_conv_module_2_sequential_7_conv_weight, x = input_109)[name = tensor("input_111")]; tensor conv_out_5_perm_0 = const()[name = tensor("conv_out_5_perm_0"), val = tensor([0, 2, 1])]; - tensor var_686_begin_0 = const()[name = tensor("op_686_begin_0"), val = tensor([0, -1, 0])]; - tensor var_686_end_0 = const()[name = tensor("op_686_end_0"), val = tensor([2, 16, 256])]; - tensor var_686_end_mask_0 = const()[name = tensor("op_686_end_mask_0"), val = tensor([true, true, true])]; - tensor conv_out_5 = transpose(perm = conv_out_5_perm_0, x = input_109)[name = tensor("transpose_39")]; - tensor var_686 = slice_by_index(begin = var_686_begin_0, end = var_686_end_0, end_mask = var_686_end_mask_0, x = conv_out_5)[name = tensor("op_686")]; - tensor var_688_perm_0 = const()[name = tensor("op_688_perm_0"), val = tensor([1, 0, 2])]; - tensor var_688 = transpose(perm = var_688_perm_0, x = var_686)[name = tensor("transpose_38")]; - tensor input_111 = add(x = x_15, y = var_688)[name = tensor("input_111")]; - tensor input_113_axes_0 = const()[name = tensor("input_113_axes_0"), val = tensor([-1])]; - tensor input_113 = layer_norm(axes = input_113_axes_0, beta = encoder_ffn2_2_module_sequential_0_bias, epsilon = var_28, gamma = encoder_ffn2_2_module_sequential_0_weight, x = input_111)[name = tensor("input_113")]; - tensor inputs_29 = linear(bias = encoder_ffn2_2_module_sequential_1_linear_bias, weight = encoder_ffn2_2_module_sequential_1_linear_weight, x = input_113)[name = tensor("linear_26")]; - tensor input_115 = silu(x = inputs_29)[name = tensor("input_115")]; - tensor input_119 = linear(bias = encoder_ffn2_2_module_sequential_4_linear_bias, weight = encoder_ffn2_2_module_sequential_4_linear_weight, x = input_115)[name = tensor("linear_27")]; - tensor var_711 = const()[name = tensor("op_711"), val = tensor(0x1p-1)]; - tensor var_712 = mul(x = input_119, y = var_711)[name = tensor("op_712")]; - tensor input_121 = add(x = var_712, y = input_111)[name = tensor("input_121")]; - tensor input_123_axes_0 = const()[name = tensor("input_123_axes_0"), val = tensor([-1])]; - tensor input_123 = layer_norm(axes = input_123_axes_0, beta = encoder_layer_norm_2_bias, epsilon = var_28, gamma = encoder_layer_norm_2_weight, x = input_121)[name = tensor("input_123")]; + tensor var_722_begin_0 = const()[name = tensor("op_722_begin_0"), val = tensor([0, -1, 0])]; + tensor var_722_end_0 = const()[name = tensor("op_722_end_0"), val = tensor([2, 16, 256])]; + tensor var_722_end_mask_0 = const()[name = tensor("op_722_end_mask_0"), val = tensor([true, true, true])]; + tensor conv_out_5 = transpose(perm = conv_out_5_perm_0, x = input_111)[name = tensor("transpose_39")]; + tensor var_722 = slice_by_index(begin = var_722_begin_0, end = var_722_end_0, end_mask = var_722_end_mask_0, x = conv_out_5)[name = tensor("op_722")]; + tensor var_724_perm_0 = const()[name = tensor("op_724_perm_0"), val = tensor([1, 0, 2])]; + tensor var_724 = transpose(perm = var_724_perm_0, x = var_722)[name = tensor("transpose_38")]; + tensor input_113 = add(x = x_15, y = var_724)[name = tensor("input_113")]; + tensor input_115_axes_0 = const()[name = tensor("input_115_axes_0"), val = tensor([-1])]; + tensor input_115 = layer_norm(axes = input_115_axes_0, beta = inner_encoder_ffn2_2_module_sequential_0_bias, epsilon = var_46, gamma = inner_encoder_ffn2_2_module_sequential_0_weight, x = input_113)[name = tensor("input_115")]; + tensor inputs_29 = linear(bias = inner_encoder_ffn2_2_module_sequential_1_linear_bias, weight = inner_encoder_ffn2_2_module_sequential_1_linear_weight, x = input_115)[name = tensor("linear_26")]; + tensor input_117 = silu(x = inputs_29)[name = tensor("input_117")]; + tensor input_121 = linear(bias = inner_encoder_ffn2_2_module_sequential_4_linear_bias, weight = inner_encoder_ffn2_2_module_sequential_4_linear_weight, x = input_117)[name = tensor("linear_27")]; + tensor var_747 = const()[name = tensor("op_747"), val = tensor(0x1p-1)]; + tensor var_748 = mul(x = input_121, y = var_747)[name = tensor("op_748")]; + tensor input_123 = add(x = var_748, y = input_113)[name = tensor("input_123")]; tensor input_125_axes_0 = const()[name = tensor("input_125_axes_0"), val = tensor([-1])]; - tensor input_125 = layer_norm(axes = input_125_axes_0, beta = encoder_ffn1_3_module_sequential_0_bias, epsilon = var_28, gamma = encoder_ffn1_3_module_sequential_0_weight, x = input_123)[name = tensor("input_125")]; - tensor inputs_31 = linear(bias = encoder_ffn1_3_module_sequential_1_linear_bias, weight = encoder_ffn1_3_module_sequential_1_linear_weight, x = input_125)[name = tensor("linear_28")]; - tensor input_127 = silu(x = inputs_31)[name = tensor("input_127")]; - tensor input_131 = linear(bias = encoder_ffn1_3_module_sequential_4_linear_bias, weight = encoder_ffn1_3_module_sequential_4_linear_weight, x = input_127)[name = tensor("linear_29")]; - tensor var_741 = const()[name = tensor("op_741"), val = tensor(0x1p-1)]; - tensor var_742 = mul(x = input_131, y = var_741)[name = tensor("op_742")]; - tensor input_133 = add(x = var_742, y = input_123)[name = tensor("input_133")]; + tensor input_125 = layer_norm(axes = input_125_axes_0, beta = inner_encoder_layer_norm_2_bias, epsilon = var_46, gamma = inner_encoder_layer_norm_2_weight, x = input_123)[name = tensor("input_125")]; + tensor input_127_axes_0 = const()[name = tensor("input_127_axes_0"), val = tensor([-1])]; + tensor input_127 = layer_norm(axes = input_127_axes_0, beta = inner_encoder_ffn1_3_module_sequential_0_bias, epsilon = var_46, gamma = inner_encoder_ffn1_3_module_sequential_0_weight, x = input_125)[name = tensor("input_127")]; + tensor inputs_31 = linear(bias = inner_encoder_ffn1_3_module_sequential_1_linear_bias, weight = inner_encoder_ffn1_3_module_sequential_1_linear_weight, x = input_127)[name = tensor("linear_28")]; + tensor input_129 = silu(x = inputs_31)[name = tensor("input_129")]; + tensor input_133 = linear(bias = inner_encoder_ffn1_3_module_sequential_4_linear_bias, weight = inner_encoder_ffn1_3_module_sequential_4_linear_weight, x = input_129)[name = tensor("linear_29")]; + tensor var_777 = const()[name = tensor("op_777"), val = tensor(0x1p-1)]; + tensor var_778 = mul(x = input_133, y = var_777)[name = tensor("op_778")]; + tensor input_135 = add(x = var_778, y = input_125)[name = tensor("input_135")]; tensor x_19_axes_0 = const()[name = tensor("x_19_axes_0"), val = tensor([-1])]; - tensor x_19 = layer_norm(axes = x_19_axes_0, beta = encoder_ret_lns_3_bias, epsilon = var_28, gamma = encoder_ret_lns_3_weight, x = input_133)[name = tensor("x_19")]; + tensor x_19 = layer_norm(axes = x_19_axes_0, beta = inner_encoder_ret_lns_3_bias, epsilon = var_46, gamma = inner_encoder_ret_lns_3_weight, x = input_135)[name = tensor("x_19")]; tensor prev_kv_7_begin_0 = const()[name = tensor("prev_kv_7_begin_0"), val = tensor([3, 0, 0, 0, 0])]; tensor prev_kv_7_end_0 = const()[name = tensor("prev_kv_7_end_0"), val = tensor([4, 1, 4, 64, 64])]; tensor prev_kv_7_end_mask_0 = const()[name = tensor("prev_kv_7_end_mask_0"), val = tensor([false, true, true, true, true])]; @@ -710,189 +724,182 @@ program(1.0) tensor prev_scale_7_end_mask_0 = const()[name = tensor("prev_scale_7_end_mask_0"), val = tensor([false, true])]; tensor prev_scale_7_squeeze_mask_0 = const()[name = tensor("prev_scale_7_squeeze_mask_0"), val = tensor([true, false])]; tensor prev_scale_7 = slice_by_index(begin = prev_scale_7_begin_0, end = prev_scale_7_end_0, end_mask = prev_scale_7_end_mask_0, squeeze_mask = prev_scale_7_squeeze_mask_0, x = enc_scale)[name = tensor("prev_scale_7")]; - tensor var_756 = linear(bias = encoder_q_proj_3_bias, weight = encoder_q_proj_3_weight, x = x_19)[name = tensor("linear_30")]; - tensor var_757 = const()[name = tensor("op_757"), val = tensor([1, 2, 4, 64])]; - tensor var_758 = reshape(shape = var_757, x = var_756)[name = tensor("op_758")]; + tensor var_792 = linear(bias = inner_encoder_q_proj_3_bias, weight = inner_encoder_q_proj_3_weight, x = x_19)[name = tensor("linear_30")]; + tensor var_793 = const()[name = tensor("op_793"), val = tensor([1, 2, 4, 64])]; + tensor var_794 = reshape(shape = var_793, x = var_792)[name = tensor("op_794")]; tensor q_7_perm_0 = const()[name = tensor("q_7_perm_0"), val = tensor([0, 2, 1, 3])]; - tensor var_762 = linear(bias = encoder_k_proj_3_bias, weight = encoder_k_proj_3_weight, x = x_19)[name = tensor("linear_31")]; - tensor var_763 = const()[name = tensor("op_763"), val = tensor(0x1p-3)]; - tensor var_764 = mul(x = var_762, y = var_763)[name = tensor("op_764")]; - tensor var_765 = const()[name = tensor("op_765"), val = tensor([1, 2, 4, 64])]; - tensor var_766 = reshape(shape = var_765, x = var_764)[name = tensor("op_766")]; + tensor var_798 = linear(bias = inner_encoder_k_proj_3_bias, weight = inner_encoder_k_proj_3_weight, x = x_19)[name = tensor("linear_31")]; + tensor var_799 = const()[name = tensor("op_799"), val = tensor(0x1p-3)]; + tensor var_800 = mul(x = var_798, y = var_799)[name = tensor("op_800")]; + tensor var_801 = const()[name = tensor("op_801"), val = tensor([1, 2, 4, 64])]; + tensor var_802 = reshape(shape = var_801, x = var_800)[name = tensor("op_802")]; tensor k_7_perm_0 = const()[name = tensor("k_7_perm_0"), val = tensor([0, 2, 1, 3])]; - tensor var_770 = linear(bias = encoder_v_proj_3_bias, weight = encoder_v_proj_3_weight, x = x_19)[name = tensor("linear_32")]; - tensor var_771 = const()[name = tensor("op_771"), val = tensor([1, 2, 4, 64])]; - tensor var_772 = reshape(shape = var_771, x = var_770)[name = tensor("op_772")]; + tensor var_806 = linear(bias = inner_encoder_v_proj_3_bias, weight = inner_encoder_v_proj_3_weight, x = x_19)[name = tensor("linear_32")]; + tensor var_807 = const()[name = tensor("op_807"), val = tensor([1, 2, 4, 64])]; + tensor var_808 = reshape(shape = var_807, x = var_806)[name = tensor("op_808")]; tensor v_7_perm_0 = const()[name = tensor("v_7_perm_0"), val = tensor([0, 2, 1, 3])]; - tensor input_137 = linear(bias = encoder_g_proj_3_bias, weight = encoder_g_proj_3_weight, x = x_19)[name = tensor("linear_33")]; + tensor input_139 = linear(bias = inner_encoder_g_proj_3_bias, weight = inner_encoder_g_proj_3_weight, x = x_19)[name = tensor("linear_33")]; tensor sqrt_s0_7 = sqrt(x = prev_scale_7)[name = tensor("sqrt_s0_7")]; - tensor s_t_7 = add(x = prev_scale_7, y = encoder__t_index)[name = tensor("s_t_7")]; + tensor s_t_7 = add(x = prev_scale_7, y = inner_encoder__t_index)[name = tensor("s_t_7")]; tensor sqrt_s_t_7 = sqrt(x = s_t_7)[name = tensor("sqrt_s_t_7")]; tensor qk_7_transpose_x_1 = const()[name = tensor("qk_7_transpose_x_1"), val = tensor(false)]; tensor qk_7_transpose_y_1 = const()[name = tensor("qk_7_transpose_y_1"), val = tensor(true)]; - tensor k_7 = transpose(perm = k_7_perm_0, x = var_766)[name = tensor("transpose_36")]; - tensor q_7 = transpose(perm = q_7_perm_0, x = var_758)[name = tensor("transpose_37")]; + tensor k_7 = transpose(perm = k_7_perm_0, x = var_802)[name = tensor("transpose_36")]; + tensor q_7 = transpose(perm = q_7_perm_0, x = var_794)[name = tensor("transpose_37")]; tensor qk_7 = matmul(transpose_x = qk_7_transpose_x_1, transpose_y = qk_7_transpose_y_1, x = q_7, y = k_7)[name = tensor("qk_7")]; - tensor var_782 = const()[name = tensor("op_782"), val = tensor([2, 1])]; - tensor var_783 = reshape(shape = var_782, x = sqrt_s_t_7)[name = tensor("op_783")]; - tensor M_7 = real_div(x = encoder__causal_mask, y = var_783)[name = tensor("M_7")]; - tensor var_785 = mul(x = qk_7, y = M_7)[name = tensor("op_785")]; + tensor var_818 = const()[name = tensor("op_818"), val = tensor([2, 1])]; + tensor var_819 = reshape(shape = var_818, x = sqrt_s_t_7)[name = tensor("op_819")]; + tensor M_7 = real_div(x = inner_encoder__causal_mask, y = var_819)[name = tensor("M_7")]; + tensor var_821 = mul(x = qk_7, y = M_7)[name = tensor("op_821")]; tensor inner_7_transpose_x_0 = const()[name = tensor("inner_7_transpose_x_0"), val = tensor(false)]; tensor inner_7_transpose_y_0 = const()[name = tensor("inner_7_transpose_y_0"), val = tensor(false)]; - tensor v_7 = transpose(perm = v_7_perm_0, x = var_772)[name = tensor("transpose_35")]; - tensor inner_7 = matmul(transpose_x = inner_7_transpose_x_0, transpose_y = inner_7_transpose_y_0, x = var_785, y = v_7)[name = tensor("inner_7")]; - tensor var_787_transpose_x_0 = const()[name = tensor("op_787_transpose_x_0"), val = tensor(false)]; - tensor var_787_transpose_y_0 = const()[name = tensor("op_787_transpose_y_0"), val = tensor(false)]; - tensor var_787 = matmul(transpose_x = var_787_transpose_x_0, transpose_y = var_787_transpose_y_0, x = q_7, y = prev_kv_7)[name = tensor("op_787")]; - tensor var_788 = real_div(x = sqrt_s0_7, y = sqrt_s_t_7)[name = tensor("op_788")]; - tensor var_789 = const()[name = tensor("op_789"), val = tensor([1, 1, 2, 1])]; - tensor var_790 = reshape(shape = var_789, x = var_788)[name = tensor("op_790")]; - tensor cross_7 = mul(x = var_787, y = var_790)[name = tensor("cross_7")]; + tensor v_7 = transpose(perm = v_7_perm_0, x = var_808)[name = tensor("transpose_35")]; + tensor inner_7 = matmul(transpose_x = inner_7_transpose_x_0, transpose_y = inner_7_transpose_y_0, x = var_821, y = v_7)[name = tensor("inner_7")]; + tensor var_823_transpose_x_0 = const()[name = tensor("op_823_transpose_x_0"), val = tensor(false)]; + tensor var_823_transpose_y_0 = const()[name = tensor("op_823_transpose_y_0"), val = tensor(false)]; + tensor var_823 = matmul(transpose_x = var_823_transpose_x_0, transpose_y = var_823_transpose_y_0, x = q_7, y = prev_kv_7)[name = tensor("op_823")]; + tensor var_824 = real_div(x = sqrt_s0_7, y = sqrt_s_t_7)[name = tensor("op_824")]; + tensor var_825 = const()[name = tensor("op_825"), val = tensor([1, 1, 2, 1])]; + tensor var_826 = reshape(shape = var_825, x = var_824)[name = tensor("op_826")]; + tensor cross_7 = mul(x = var_823, y = var_826)[name = tensor("cross_7")]; tensor out_19 = add(x = inner_7, y = cross_7)[name = tensor("out_19")]; - tensor var_793 = mul(x = prev_kv_7, y = sqrt_s0_7)[name = tensor("op_793")]; - tensor var_795_transpose_x_1 = const()[name = tensor("op_795_transpose_x_1"), val = tensor(true)]; - tensor var_795_transpose_y_1 = const()[name = tensor("op_795_transpose_y_1"), val = tensor(false)]; - tensor var_795 = matmul(transpose_x = var_795_transpose_x_1, transpose_y = var_795_transpose_y_1, x = k_7, y = v_7)[name = tensor("op_795")]; - tensor new_kv_unnorm_7 = add(x = var_793, y = var_795)[name = tensor("new_kv_unnorm_7")]; - tensor var_797 = const()[name = tensor("op_797"), val = tensor(0x1p+1)]; - tensor new_scale_7 = add(x = prev_scale_7, y = var_797)[name = tensor("new_scale_7")]; - tensor var_799 = sqrt(x = new_scale_7)[name = tensor("op_799")]; - tensor nkv_1 = real_div(x = new_kv_unnorm_7, y = var_799)[name = tensor("nkv_1")]; - tensor var_801_perm_0 = const()[name = tensor("op_801_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_829 = mul(x = prev_kv_7, y = sqrt_s0_7)[name = tensor("op_829")]; + tensor var_831_transpose_x_1 = const()[name = tensor("op_831_transpose_x_1"), val = tensor(true)]; + tensor var_831_transpose_y_1 = const()[name = tensor("op_831_transpose_y_1"), val = tensor(false)]; + tensor var_831 = matmul(transpose_x = var_831_transpose_x_1, transpose_y = var_831_transpose_y_1, x = k_7, y = v_7)[name = tensor("op_831")]; + tensor new_kv_unnorm_7 = add(x = var_829, y = var_831)[name = tensor("new_kv_unnorm_7")]; + tensor var_833 = const()[name = tensor("op_833"), val = tensor(0x1p+1)]; + tensor new_scale_7 = add(x = prev_scale_7, y = var_833)[name = tensor("new_scale_7")]; + tensor var_835 = sqrt(x = new_scale_7)[name = tensor("op_835")]; + tensor nkv_1 = real_div(x = new_kv_unnorm_7, y = var_835)[name = tensor("nkv_1")]; + tensor var_837_perm_0 = const()[name = tensor("op_837_perm_0"), val = tensor([0, 2, 1, 3])]; tensor out_21_axes_0 = const()[name = tensor("out_21_axes_0"), val = tensor([-1])]; - tensor var_801 = transpose(perm = var_801_perm_0, x = out_19)[name = tensor("transpose_34")]; - tensor out_21 = layer_norm(axes = out_21_axes_0, epsilon = var_17, x = var_801)[name = tensor("out_21")]; - tensor var_805 = const()[name = tensor("op_805"), val = tensor([1, 2, 256])]; - tensor out_23 = reshape(shape = var_805, x = out_21)[name = tensor("out_23")]; - tensor var_807 = silu(x = input_137)[name = tensor("op_807")]; - tensor input_139 = mul(x = var_807, y = out_23)[name = tensor("input_139")]; - tensor ret_out_7 = linear(bias = encoder_out_proj_3_bias, weight = encoder_out_proj_3_weight, x = input_139)[name = tensor("linear_34")]; - tensor x_21 = add(x = input_133, y = ret_out_7)[name = tensor("x_21")]; + tensor var_837 = transpose(perm = var_837_perm_0, x = out_19)[name = tensor("transpose_34")]; + tensor out_21 = layer_norm(axes = out_21_axes_0, epsilon = var_54, x = var_837)[name = tensor("out_21")]; + tensor var_841 = const()[name = tensor("op_841"), val = tensor([1, 2, 256])]; + tensor out_23 = reshape(shape = var_841, x = out_21)[name = tensor("out_23")]; + tensor var_843 = silu(x = input_139)[name = tensor("op_843")]; + tensor input_141 = mul(x = var_843, y = out_23)[name = tensor("input_141")]; + tensor ret_out_7 = linear(bias = inner_encoder_out_proj_3_bias, weight = inner_encoder_out_proj_3_weight, x = input_141)[name = tensor("linear_34")]; + tensor x_21 = add(x = input_135, y = ret_out_7)[name = tensor("x_21")]; tensor window_19_begin_0 = const()[name = tensor("window_19_begin_0"), val = tensor([3, 0, 0, 0])]; tensor window_19_end_0 = const()[name = tensor("window_19_end_0"), val = tensor([4, 1, 16, 256])]; tensor window_19_end_mask_0 = const()[name = tensor("window_19_end_mask_0"), val = tensor([false, true, true, true])]; tensor window_19_squeeze_mask_0 = const()[name = tensor("window_19_squeeze_mask_0"), val = tensor([true, false, false, false])]; tensor window_19 = slice_by_index(begin = window_19_begin_0, end = window_19_end_0, end_mask = window_19_end_mask_0, squeeze_mask = window_19_squeeze_mask_0, x = enc_conv_cache)[name = tensor("window_19")]; - tensor var_815_begin_0 = const()[name = tensor("op_815_begin_0"), val = tensor([0, 0, 0])]; - tensor var_815_end_0 = const()[name = tensor("op_815_end_0"), val = tensor([1, 1, 256])]; - tensor var_815_end_mask_0 = const()[name = tensor("op_815_end_mask_0"), val = tensor([true, false, true])]; - tensor var_815 = slice_by_index(begin = var_815_begin_0, end = var_815_end_0, end_mask = var_815_end_mask_0, x = x_21)[name = tensor("op_815")]; - tensor var_818_begin_0 = const()[name = tensor("op_818_begin_0"), val = tensor([0, 1, 0])]; - tensor var_818_end_0 = const()[name = tensor("op_818_end_0"), val = tensor([1, 16, 256])]; - tensor var_818_end_mask_0 = const()[name = tensor("op_818_end_mask_0"), val = tensor([true, true, true])]; - tensor var_818 = slice_by_index(begin = var_818_begin_0, end = var_818_end_0, end_mask = var_818_end_mask_0, x = window_19)[name = tensor("op_818")]; + tensor var_851_begin_0 = const()[name = tensor("op_851_begin_0"), val = tensor([0, 0, 0])]; + tensor var_851_end_0 = const()[name = tensor("op_851_end_0"), val = tensor([1, 1, 256])]; + tensor var_851_end_mask_0 = const()[name = tensor("op_851_end_mask_0"), val = tensor([true, false, true])]; + tensor var_851 = slice_by_index(begin = var_851_begin_0, end = var_851_end_0, end_mask = var_851_end_mask_0, x = x_21)[name = tensor("op_851")]; + tensor var_854_begin_0 = const()[name = tensor("op_854_begin_0"), val = tensor([0, 1, 0])]; + tensor var_854_end_0 = const()[name = tensor("op_854_end_0"), val = tensor([1, 16, 256])]; + tensor var_854_end_mask_0 = const()[name = tensor("op_854_end_mask_0"), val = tensor([true, true, true])]; + tensor var_854 = slice_by_index(begin = var_854_begin_0, end = var_854_end_0, end_mask = var_854_end_mask_0, x = window_19)[name = tensor("op_854")]; tensor window_21_interleave_0 = const()[name = tensor("window_21_interleave_0"), val = tensor(false)]; - tensor window_21 = concat(axis = var_26, interleave = window_21_interleave_0, values = (var_818, var_815))[name = tensor("window_21")]; - tensor var_823_begin_0 = const()[name = tensor("op_823_begin_0"), val = tensor([0, 1, 0])]; - tensor var_823_end_0 = const()[name = tensor("op_823_end_0"), val = tensor([1, 1, 256])]; - tensor var_823_end_mask_0 = const()[name = tensor("op_823_end_mask_0"), val = tensor([true, true, true])]; - tensor var_823 = slice_by_index(begin = var_823_begin_0, end = var_823_end_0, end_mask = var_823_end_mask_0, x = x_21)[name = tensor("op_823")]; - tensor var_826_begin_0 = const()[name = tensor("op_826_begin_0"), val = tensor([0, 1, 0])]; - tensor var_826_end_0 = const()[name = tensor("op_826_end_0"), val = tensor([1, 16, 256])]; - tensor var_826_end_mask_0 = const()[name = tensor("op_826_end_mask_0"), val = tensor([true, true, true])]; - tensor var_826 = slice_by_index(begin = var_826_begin_0, end = var_826_end_0, end_mask = var_826_end_mask_0, x = window_21)[name = tensor("op_826")]; + tensor window_21 = concat(axis = var_62, interleave = window_21_interleave_0, values = (var_854, var_851))[name = tensor("window_21")]; + tensor var_859_begin_0 = const()[name = tensor("op_859_begin_0"), val = tensor([0, 1, 0])]; + tensor var_859_end_0 = const()[name = tensor("op_859_end_0"), val = tensor([1, 1, 256])]; + tensor var_859_end_mask_0 = const()[name = tensor("op_859_end_mask_0"), val = tensor([true, true, true])]; + tensor var_859 = slice_by_index(begin = var_859_begin_0, end = var_859_end_0, end_mask = var_859_end_mask_0, x = x_21)[name = tensor("op_859")]; + tensor var_862_begin_0 = const()[name = tensor("op_862_begin_0"), val = tensor([0, 1, 0])]; + tensor var_862_end_0 = const()[name = tensor("op_862_end_0"), val = tensor([1, 16, 256])]; + tensor var_862_end_mask_0 = const()[name = tensor("op_862_end_mask_0"), val = tensor([true, true, true])]; + tensor var_862 = slice_by_index(begin = var_862_begin_0, end = var_862_end_0, end_mask = var_862_end_mask_0, x = window_21)[name = tensor("op_862")]; tensor window_interleave_0 = const()[name = tensor("window_interleave_0"), val = tensor(false)]; - tensor window = concat(axis = var_26, interleave = window_interleave_0, values = (var_826, var_823))[name = tensor("window")]; - tensor input_141_interleave_0 = const()[name = tensor("input_141_interleave_0"), val = tensor(false)]; - tensor input_141 = concat(axis = var_23, interleave = input_141_interleave_0, values = (window_21, window))[name = tensor("input_141")]; + tensor window = concat(axis = var_62, interleave = window_interleave_0, values = (var_862, var_859))[name = tensor("window")]; + tensor input_143_interleave_0 = const()[name = tensor("input_143_interleave_0"), val = tensor(false)]; + tensor input_143 = concat(axis = var_49, interleave = input_143_interleave_0, values = (window_21, window))[name = tensor("input_143")]; tensor x_23_axes_0 = const()[name = tensor("x_23_axes_0"), val = tensor([-1])]; - tensor x_23 = layer_norm(axes = x_23_axes_0, beta = encoder_conv_module_3_sequential_0_bias, epsilon = var_28, gamma = encoder_conv_module_3_sequential_0_weight, x = input_141)[name = tensor("x_23")]; - tensor input_143_perm_0 = const()[name = tensor("input_143_perm_0"), val = tensor([0, 2, 1])]; + tensor x_23 = layer_norm(axes = x_23_axes_0, beta = inner_encoder_conv_module_3_sequential_0_bias, epsilon = var_46, gamma = inner_encoder_conv_module_3_sequential_0_weight, x = input_143)[name = tensor("x_23")]; + tensor input_145_perm_0 = const()[name = tensor("input_145_perm_0"), val = tensor([0, 2, 1])]; tensor inputs_33_pad_type_0 = const()[name = tensor("inputs_33_pad_type_0"), val = tensor("valid")]; tensor inputs_33_strides_0 = const()[name = tensor("inputs_33_strides_0"), val = tensor([1])]; tensor inputs_33_pad_0 = const()[name = tensor("inputs_33_pad_0"), val = tensor([0, 0])]; tensor inputs_33_dilations_0 = const()[name = tensor("inputs_33_dilations_0"), val = tensor([1])]; tensor inputs_33_groups_0 = const()[name = tensor("inputs_33_groups_0"), val = tensor(1)]; - tensor input_143 = transpose(perm = input_143_perm_0, x = x_23)[name = tensor("transpose_33")]; - tensor inputs_33 = conv(bias = encoder_conv_module_3_sequential_2_conv_bias, dilations = inputs_33_dilations_0, groups = inputs_33_groups_0, pad = inputs_33_pad_0, pad_type = inputs_33_pad_type_0, strides = inputs_33_strides_0, weight = encoder_conv_module_3_sequential_2_conv_weight, x = input_143)[name = tensor("inputs_33")]; - tensor var_851_split_sizes_0 = const()[name = tensor("op_851_split_sizes_0"), val = tensor([256, 256])]; - tensor var_851_axis_0 = const()[name = tensor("op_851_axis_0"), val = tensor(1)]; - tensor var_851_0, tensor var_851_1 = split(axis = var_851_axis_0, split_sizes = var_851_split_sizes_0, x = inputs_33)[name = tensor("op_851")]; - tensor var_853 = sigmoid(x = var_851_1)[name = tensor("op_853")]; - tensor inputs_35 = mul(x = var_851_0, y = var_853)[name = tensor("inputs_35")]; + tensor input_145 = transpose(perm = input_145_perm_0, x = x_23)[name = tensor("transpose_33")]; + tensor inputs_33 = conv(bias = inner_encoder_conv_module_3_sequential_2_conv_bias, dilations = inputs_33_dilations_0, groups = inputs_33_groups_0, pad = inputs_33_pad_0, pad_type = inputs_33_pad_type_0, strides = inputs_33_strides_0, weight = inner_encoder_conv_module_3_sequential_2_conv_weight, x = input_145)[name = tensor("inputs_33")]; + tensor var_887_split_sizes_0 = const()[name = tensor("op_887_split_sizes_0"), val = tensor([256, 256])]; + tensor var_887_axis_0 = const()[name = tensor("op_887_axis_0"), val = tensor(1)]; + tensor var_887_0, tensor var_887_1 = split(axis = var_887_axis_0, split_sizes = var_887_split_sizes_0, x = inputs_33)[name = tensor("op_887")]; + tensor var_889 = sigmoid(x = var_887_1)[name = tensor("op_889")]; + tensor inputs_35 = mul(x = var_887_0, y = var_889)[name = tensor("inputs_35")]; tensor outputs_aug_pad_type_0 = const()[name = tensor("outputs_aug_pad_type_0"), val = tensor("custom")]; tensor outputs_aug_pad_0 = const()[name = tensor("outputs_aug_pad_0"), val = tensor([15, 15])]; tensor outputs_aug_groups_0 = const()[name = tensor("outputs_aug_groups_0"), val = tensor(256)]; tensor outputs_aug_strides_0 = const()[name = tensor("outputs_aug_strides_0"), val = tensor([1])]; tensor outputs_aug_dilations_0 = const()[name = tensor("outputs_aug_dilations_0"), val = tensor([1])]; - tensor outputs_aug = conv(dilations = outputs_aug_dilations_0, groups = outputs_aug_groups_0, pad = outputs_aug_pad_0, pad_type = outputs_aug_pad_type_0, strides = outputs_aug_strides_0, weight = encoder_conv_module_3_sequential_4_conv_weight, x = inputs_35)[name = tensor("outputs_aug")]; - tensor input_145_begin_0 = const()[name = tensor("input_145_begin_0"), val = tensor([0, 0, 0])]; - tensor input_145_end_0 = const()[name = tensor("input_145_end_0"), val = tensor([2, 256, 16])]; - tensor input_145_end_mask_0 = const()[name = tensor("input_145_end_mask_0"), val = tensor([true, true, false])]; - tensor input_145 = slice_by_index(begin = input_145_begin_0, end = input_145_end_0, end_mask = input_145_end_mask_0, x = outputs_aug)[name = tensor("input_145")]; - tensor inputs_37 = batch_norm(beta = encoder_conv_module_3_sequential_5_bias, epsilon = var_28, gamma = encoder_conv_module_3_sequential_5_weight, mean = encoder_conv_module_3_sequential_5_running_mean, variance = encoder_conv_module_3_sequential_5_running_var, x = input_145)[name = tensor("inputs_37")]; - tensor input_147 = silu(x = inputs_37)[name = tensor("input_147")]; - tensor input_149_pad_type_0 = const()[name = tensor("input_149_pad_type_0"), val = tensor("valid")]; - tensor input_149_strides_0 = const()[name = tensor("input_149_strides_0"), val = tensor([1])]; - tensor input_149_pad_0 = const()[name = tensor("input_149_pad_0"), val = tensor([0, 0])]; - tensor input_149_dilations_0 = const()[name = tensor("input_149_dilations_0"), val = tensor([1])]; - tensor input_149_groups_0 = const()[name = tensor("input_149_groups_0"), val = tensor(1)]; - tensor input_149 = conv(bias = encoder_conv_module_3_sequential_7_conv_bias, dilations = input_149_dilations_0, groups = input_149_groups_0, pad = input_149_pad_0, pad_type = input_149_pad_type_0, strides = input_149_strides_0, weight = encoder_conv_module_3_sequential_7_conv_weight, x = input_147)[name = tensor("input_149")]; + tensor outputs_aug = conv(dilations = outputs_aug_dilations_0, groups = outputs_aug_groups_0, pad = outputs_aug_pad_0, pad_type = outputs_aug_pad_type_0, strides = outputs_aug_strides_0, weight = inner_encoder_conv_module_3_sequential_4_conv_weight, x = inputs_35)[name = tensor("outputs_aug")]; + tensor input_147_begin_0 = const()[name = tensor("input_147_begin_0"), val = tensor([0, 0, 0])]; + tensor input_147_end_0 = const()[name = tensor("input_147_end_0"), val = tensor([2, 256, 16])]; + tensor input_147_end_mask_0 = const()[name = tensor("input_147_end_mask_0"), val = tensor([true, true, false])]; + tensor input_147 = slice_by_index(begin = input_147_begin_0, end = input_147_end_0, end_mask = input_147_end_mask_0, x = outputs_aug)[name = tensor("input_147")]; + tensor inputs_37 = batch_norm(beta = inner_encoder_conv_module_3_sequential_5_bias, epsilon = var_46, gamma = inner_encoder_conv_module_3_sequential_5_weight, mean = inner_encoder_conv_module_3_sequential_5_running_mean, variance = inner_encoder_conv_module_3_sequential_5_running_var, x = input_147)[name = tensor("inputs_37")]; + tensor input_149 = silu(x = inputs_37)[name = tensor("input_149")]; + tensor input_151_pad_type_0 = const()[name = tensor("input_151_pad_type_0"), val = tensor("valid")]; + tensor input_151_strides_0 = const()[name = tensor("input_151_strides_0"), val = tensor([1])]; + tensor input_151_pad_0 = const()[name = tensor("input_151_pad_0"), val = tensor([0, 0])]; + tensor input_151_dilations_0 = const()[name = tensor("input_151_dilations_0"), val = tensor([1])]; + tensor input_151_groups_0 = const()[name = tensor("input_151_groups_0"), val = tensor(1)]; + tensor input_151 = conv(bias = inner_encoder_conv_module_3_sequential_7_conv_bias, dilations = input_151_dilations_0, groups = input_151_groups_0, pad = input_151_pad_0, pad_type = input_151_pad_type_0, strides = input_151_strides_0, weight = inner_encoder_conv_module_3_sequential_7_conv_weight, x = input_149)[name = tensor("input_151")]; tensor conv_out_7_perm_0 = const()[name = tensor("conv_out_7_perm_0"), val = tensor([0, 2, 1])]; - tensor var_884_begin_0 = const()[name = tensor("op_884_begin_0"), val = tensor([0, -1, 0])]; - tensor var_884_end_0 = const()[name = tensor("op_884_end_0"), val = tensor([2, 16, 256])]; - tensor var_884_end_mask_0 = const()[name = tensor("op_884_end_mask_0"), val = tensor([true, true, true])]; - tensor conv_out_7 = transpose(perm = conv_out_7_perm_0, x = input_149)[name = tensor("transpose_32")]; - tensor var_884 = slice_by_index(begin = var_884_begin_0, end = var_884_end_0, end_mask = var_884_end_mask_0, x = conv_out_7)[name = tensor("op_884")]; - tensor var_886_perm_0 = const()[name = tensor("op_886_perm_0"), val = tensor([1, 0, 2])]; - tensor var_886 = transpose(perm = var_886_perm_0, x = var_884)[name = tensor("transpose_31")]; - tensor input_151 = add(x = x_21, y = var_886)[name = tensor("input_151")]; - tensor input_153_axes_0 = const()[name = tensor("input_153_axes_0"), val = tensor([-1])]; - tensor input_153 = layer_norm(axes = input_153_axes_0, beta = encoder_ffn2_3_module_sequential_0_bias, epsilon = var_28, gamma = encoder_ffn2_3_module_sequential_0_weight, x = input_151)[name = tensor("input_153")]; - tensor inputs = linear(bias = encoder_ffn2_3_module_sequential_1_linear_bias, weight = encoder_ffn2_3_module_sequential_1_linear_weight, x = input_153)[name = tensor("linear_35")]; - tensor input_155 = silu(x = inputs)[name = tensor("input_155")]; - tensor input_159 = linear(bias = encoder_ffn2_3_module_sequential_4_linear_bias, weight = encoder_ffn2_3_module_sequential_4_linear_weight, x = input_155)[name = tensor("linear_36")]; - tensor var_909 = const()[name = tensor("op_909"), val = tensor(0x1p-1)]; - tensor var_910 = mul(x = input_159, y = var_909)[name = tensor("op_910")]; - tensor input_161 = add(x = var_910, y = input_151)[name = tensor("input_161")]; + tensor var_920_begin_0 = const()[name = tensor("op_920_begin_0"), val = tensor([0, -1, 0])]; + tensor var_920_end_0 = const()[name = tensor("op_920_end_0"), val = tensor([2, 16, 256])]; + tensor var_920_end_mask_0 = const()[name = tensor("op_920_end_mask_0"), val = tensor([true, true, true])]; + tensor conv_out_7 = transpose(perm = conv_out_7_perm_0, x = input_151)[name = tensor("transpose_32")]; + tensor var_920 = slice_by_index(begin = var_920_begin_0, end = var_920_end_0, end_mask = var_920_end_mask_0, x = conv_out_7)[name = tensor("op_920")]; + tensor var_922_perm_0 = const()[name = tensor("op_922_perm_0"), val = tensor([1, 0, 2])]; + tensor var_922 = transpose(perm = var_922_perm_0, x = var_920)[name = tensor("transpose_31")]; + tensor input_153 = add(x = x_21, y = var_922)[name = tensor("input_153")]; + tensor input_155_axes_0 = const()[name = tensor("input_155_axes_0"), val = tensor([-1])]; + tensor input_155 = layer_norm(axes = input_155_axes_0, beta = inner_encoder_ffn2_3_module_sequential_0_bias, epsilon = var_46, gamma = inner_encoder_ffn2_3_module_sequential_0_weight, x = input_153)[name = tensor("input_155")]; + tensor inputs = linear(bias = inner_encoder_ffn2_3_module_sequential_1_linear_bias, weight = inner_encoder_ffn2_3_module_sequential_1_linear_weight, x = input_155)[name = tensor("linear_35")]; + tensor input_157 = silu(x = inputs)[name = tensor("input_157")]; + tensor input_161 = linear(bias = inner_encoder_ffn2_3_module_sequential_4_linear_bias, weight = inner_encoder_ffn2_3_module_sequential_4_linear_weight, x = input_157)[name = tensor("linear_36")]; + tensor var_945 = const()[name = tensor("op_945"), val = tensor(0x1p-1)]; + tensor var_946 = mul(x = input_161, y = var_945)[name = tensor("op_946")]; + tensor input_163 = add(x = var_946, y = input_153)[name = tensor("input_163")]; tensor x_25_axes_0 = const()[name = tensor("x_25_axes_0"), val = tensor([-1])]; - tensor x_25 = layer_norm(axes = x_25_axes_0, beta = encoder_layer_norm_3_bias, epsilon = var_28, gamma = encoder_layer_norm_3_weight, x = input_161)[name = tensor("x_25")]; + tensor x_25 = layer_norm(axes = x_25_axes_0, beta = inner_encoder_layer_norm_3_bias, epsilon = var_46, gamma = inner_encoder_layer_norm_3_weight, x = input_163)[name = tensor("x_25")]; tensor x_ct_perm_0 = const()[name = tensor("x_ct_perm_0"), val = tensor([0, 2, 1])]; tensor cat_interleave_0 = const()[name = tensor("cat_interleave_0"), val = tensor(false)]; tensor x_ct = transpose(perm = x_ct_perm_0, x = x_25)[name = tensor("transpose_30")]; - tensor cat = concat(axis = var_20, interleave = cat_interleave_0, values = (cnn_window, x_ct))[name = tensor("cat")]; + tensor cat = concat(axis = var_51, interleave = cat_interleave_0, values = (cnn_window, x_ct))[name = tensor("cat")]; tensor conv_out_pad_type_0 = const()[name = tensor("conv_out_pad_type_0"), val = tensor("valid")]; tensor conv_out_strides_0 = const()[name = tensor("conv_out_strides_0"), val = tensor([1])]; tensor conv_out_pad_0 = const()[name = tensor("conv_out_pad_0"), val = tensor([0, 0])]; tensor conv_out_dilations_0 = const()[name = tensor("conv_out_dilations_0"), val = tensor([1])]; tensor conv_out_groups_0 = const()[name = tensor("conv_out_groups_0"), val = tensor(1)]; - tensor conv_out = conv(bias = encoder_cnn_bias, dilations = conv_out_dilations_0, groups = conv_out_groups_0, pad = conv_out_pad_0, pad_type = conv_out_pad_type_0, strides = conv_out_strides_0, weight = encoder_cnn_weight, x = cat)[name = tensor("conv_out")]; - tensor var_928_begin_0 = const()[name = tensor("op_928_begin_0"), val = tensor([0, 0, 2])]; - tensor var_928_end_0 = const()[name = tensor("op_928_end_0"), val = tensor([1, 256, 20])]; - tensor var_928_end_mask_0 = const()[name = tensor("op_928_end_mask_0"), val = tensor([true, true, true])]; - tensor cnn_window_new = slice_by_index(begin = var_928_begin_0, end = var_928_end_0, end_mask = var_928_end_mask_0, x = cat)[name = tensor("op_928")]; - tensor input_163_perm_0 = const()[name = tensor("input_163_perm_0"), val = tensor([0, 2, 1])]; - tensor var_930 = const()[name = tensor("op_930"), val = tensor([-1])]; - tensor input_163 = transpose(perm = input_163_perm_0, x = conv_out)[name = tensor("transpose_29")]; - tensor var_931 = reduce_l2_norm(axes = var_930, keep_dims = var_29, x = input_163)[name = tensor("op_931")]; + tensor conv_out = conv(bias = inner_encoder_cnn_bias, dilations = conv_out_dilations_0, groups = conv_out_groups_0, pad = conv_out_pad_0, pad_type = conv_out_pad_type_0, strides = conv_out_strides_0, weight = inner_encoder_cnn_weight, x = cat)[name = tensor("conv_out")]; + tensor var_964_begin_0 = const()[name = tensor("op_964_begin_0"), val = tensor([0, 0, 2])]; + tensor var_964_end_0 = const()[name = tensor("op_964_end_0"), val = tensor([1, 256, 20])]; + tensor var_964_end_mask_0 = const()[name = tensor("op_964_end_mask_0"), val = tensor([true, true, true])]; + tensor cnn_window_new = slice_by_index(begin = var_964_begin_0, end = var_964_end_0, end_mask = var_964_end_mask_0, x = cat)[name = tensor("op_964")]; + tensor input_165_perm_0 = const()[name = tensor("input_165_perm_0"), val = tensor([0, 2, 1])]; + tensor var_966 = const()[name = tensor("op_966"), val = tensor([-1])]; + tensor input_165 = transpose(perm = input_165_perm_0, x = conv_out)[name = tensor("transpose_29")]; + tensor var_967 = reduce_l2_norm(axes = var_966, keep_dims = var_45, x = input_165)[name = tensor("op_967")]; tensor const_12 = const()[name = tensor("const_12"), val = tensor(0x1.fffffep+127)]; - tensor clip_0 = clip(alpha = var_11, beta = const_12, x = var_931)[name = tensor("clip_0")]; - tensor emb = real_div(x = input_163, y = clip_0)[name = tensor("emb")]; - tensor var_935_axis_0 = const()[name = tensor("op_935_axis_0"), val = tensor(0)]; - tensor enc_kv_new = stack(axis = var_935_axis_0, values = (var_206, var_404, var_602, nkv_1))[name = tensor("op_935")]; - tensor var_937_axis_0 = const()[name = tensor("op_937_axis_0"), val = tensor(0)]; - tensor enc_scale_new = stack(axis = var_937_axis_0, values = (new_scale_1, new_scale_3, new_scale_5, new_scale_7))[name = tensor("op_937")]; - tensor var_939_axis_0 = const()[name = tensor("op_939_axis_0"), val = tensor(0)]; - tensor enc_conv_cache_new = stack(axis = var_939_axis_0, values = (window_5, window_11, window_17, window))[name = tensor("op_939")]; - tensor var_948 = const()[name = tensor("op_948"), val = tensor(0x1.5798eep-27)]; - tensor var_953 = const()[name = tensor("op_953"), val = tensor(0x1.4f8b58p-17)]; - tensor var_955 = const()[name = tensor("op_955"), val = tensor(0x1.0c6f7ap-20)]; - tensor var_956 = const()[name = tensor("op_956"), val = tensor(true)]; - tensor var_958 = const()[name = tensor("op_958"), val = tensor(0x1p+0)]; - tensor var_962 = const()[name = tensor("op_962"), val = tensor(-1)]; - tensor var_968 = const()[name = tensor("op_968"), val = tensor(0)]; + tensor clip_0 = clip(alpha = var_59, beta = const_12, x = var_967)[name = tensor("clip_0")]; + tensor emb = real_div(x = input_165, y = clip_0)[name = tensor("emb")]; + tensor var_971_axis_0 = const()[name = tensor("op_971_axis_0"), val = tensor(0)]; + tensor enc_kv_new = stack(axis = var_971_axis_0, values = (var_242, var_440, var_638, nkv_1))[name = tensor("op_971")]; + tensor var_973_axis_0 = const()[name = tensor("op_973_axis_0"), val = tensor(0)]; + tensor enc_scale_new = stack(axis = var_973_axis_0, values = (new_scale_1, new_scale_3, new_scale_5, new_scale_7))[name = tensor("op_973")]; + tensor var_975_axis_0 = const()[name = tensor("op_975_axis_0"), val = tensor(0)]; + tensor enc_conv_cache_new = stack(axis = var_975_axis_0, values = (window_5, window_11, window_17, window))[name = tensor("op_975")]; tensor pos = const()[name = tensor("pos"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(44395328)))]; - tensor var_1030_axes_0 = const()[name = tensor("op_1030_axes_0"), val = tensor([2])]; - tensor var_1030 = expand_dims(axes = var_1030_axes_0, x = emb)[name = tensor("op_1030")]; + tensor var_1043_axes_0 = const()[name = tensor("op_1043_axes_0"), val = tensor([2])]; + tensor var_1043 = expand_dims(axes = var_1043_axes_0, x = emb)[name = tensor("op_1043")]; tensor emb_exp_reps_0 = const()[name = tensor("emb_exp_reps_0"), val = tensor([1, 1, 12, 1])]; - tensor emb_exp = tile(reps = emb_exp_reps_0, x = var_1030)[name = tensor("emb_exp")]; - tensor input_165_interleave_0 = const()[name = tensor("input_165_interleave_0"), val = tensor(false)]; - tensor input_165 = concat(axis = var_962, interleave = input_165_interleave_0, values = (emb_exp, pos))[name = tensor("input_165")]; - tensor x_27 = linear(bias = decoder_convert_bias, weight = decoder_convert_weight, x = input_165)[name = tensor("linear_37")]; - tensor var_1038_perm_0 = const()[name = tensor("op_1038_perm_0"), val = tensor([0, 2, 1, 3])]; - tensor var_1042 = const()[name = tensor("op_1042"), val = tensor([12, 2, 256])]; - tensor var_1038 = transpose(perm = var_1038_perm_0, x = x_27)[name = tensor("transpose_28")]; - tensor x_29 = reshape(shape = var_1042, x = var_1038)[name = tensor("x_29")]; + tensor emb_exp = tile(reps = emb_exp_reps_0, x = var_1043)[name = tensor("emb_exp")]; + tensor input_167_interleave_0 = const()[name = tensor("input_167_interleave_0"), val = tensor(false)]; + tensor input_167 = concat(axis = var_52, interleave = input_167_interleave_0, values = (emb_exp, pos))[name = tensor("input_167")]; + tensor x_27 = linear(bias = inner_decoder_convert_bias, weight = inner_decoder_convert_weight, x = input_167)[name = tensor("linear_37")]; + tensor var_1051_perm_0 = const()[name = tensor("op_1051_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_1055 = const()[name = tensor("op_1055"), val = tensor([12, 2, 256])]; + tensor var_1051 = transpose(perm = var_1051_perm_0, x = x_27)[name = tensor("transpose_28")]; + tensor x_29 = reshape(shape = var_1055, x = var_1051)[name = tensor("x_29")]; tensor prev_kv_9_begin_0 = const()[name = tensor("prev_kv_9_begin_0"), val = tensor([0, 0, 0, 0, 0])]; tensor prev_kv_9_end_0 = const()[name = tensor("prev_kv_9_end_0"), val = tensor([1, 12, 4, 64, 64])]; tensor prev_kv_9_end_mask_0 = const()[name = tensor("prev_kv_9_end_mask_0"), val = tensor([false, true, true, true, true])]; @@ -903,132 +910,132 @@ program(1.0) tensor prev_scale_9_end_mask_0 = const()[name = tensor("prev_scale_9_end_mask_0"), val = tensor([false, true])]; tensor prev_scale_9_squeeze_mask_0 = const()[name = tensor("prev_scale_9_squeeze_mask_0"), val = tensor([true, false])]; tensor prev_scale_9 = slice_by_index(begin = prev_scale_9_begin_0, end = prev_scale_9_end_0, end_mask = prev_scale_9_end_mask_0, squeeze_mask = prev_scale_9_squeeze_mask_0, x = dec_scale)[name = tensor("prev_scale_9")]; - tensor var_1050 = linear(bias = decoder_q_proj_0_bias, weight = decoder_q_proj_0_weight, x = x_29)[name = tensor("linear_38")]; - tensor var_1051 = const()[name = tensor("op_1051"), val = tensor([12, 2, 4, 64])]; - tensor var_1052 = reshape(shape = var_1051, x = var_1050)[name = tensor("op_1052")]; + tensor var_1063 = linear(bias = inner_decoder_q_proj_0_bias, weight = inner_decoder_q_proj_0_weight, x = x_29)[name = tensor("linear_38")]; + tensor var_1064 = const()[name = tensor("op_1064"), val = tensor([12, 2, 4, 64])]; + tensor var_1065 = reshape(shape = var_1064, x = var_1063)[name = tensor("op_1065")]; tensor q_9_perm_0 = const()[name = tensor("q_9_perm_0"), val = tensor([0, 2, 1, 3])]; - tensor var_1056 = linear(bias = decoder_k_proj_0_bias, weight = decoder_k_proj_0_weight, x = x_29)[name = tensor("linear_39")]; - tensor var_1057 = const()[name = tensor("op_1057"), val = tensor(0x1p-3)]; - tensor var_1058 = mul(x = var_1056, y = var_1057)[name = tensor("op_1058")]; - tensor var_1059 = const()[name = tensor("op_1059"), val = tensor([12, 2, 4, 64])]; - tensor var_1060 = reshape(shape = var_1059, x = var_1058)[name = tensor("op_1060")]; + tensor var_1069 = linear(bias = inner_decoder_k_proj_0_bias, weight = inner_decoder_k_proj_0_weight, x = x_29)[name = tensor("linear_39")]; + tensor var_1070 = const()[name = tensor("op_1070"), val = tensor(0x1p-3)]; + tensor var_1071 = mul(x = var_1069, y = var_1070)[name = tensor("op_1071")]; + tensor var_1072 = const()[name = tensor("op_1072"), val = tensor([12, 2, 4, 64])]; + tensor var_1073 = reshape(shape = var_1072, x = var_1071)[name = tensor("op_1073")]; tensor k_9_perm_0 = const()[name = tensor("k_9_perm_0"), val = tensor([0, 2, 1, 3])]; - tensor var_1064 = linear(bias = decoder_v_proj_0_bias, weight = decoder_v_proj_0_weight, x = x_29)[name = tensor("linear_40")]; - tensor var_1065 = const()[name = tensor("op_1065"), val = tensor([12, 2, 4, 64])]; - tensor var_1066 = reshape(shape = var_1065, x = var_1064)[name = tensor("op_1066")]; + tensor var_1077 = linear(bias = inner_decoder_v_proj_0_bias, weight = inner_decoder_v_proj_0_weight, x = x_29)[name = tensor("linear_40")]; + tensor var_1078 = const()[name = tensor("op_1078"), val = tensor([12, 2, 4, 64])]; + tensor var_1079 = reshape(shape = var_1078, x = var_1077)[name = tensor("op_1079")]; tensor v_9_perm_0 = const()[name = tensor("v_9_perm_0"), val = tensor([0, 2, 1, 3])]; - tensor input_169 = linear(bias = decoder_g_proj_0_bias, weight = decoder_g_proj_0_weight, x = x_29)[name = tensor("linear_41")]; + tensor input_171 = linear(bias = inner_decoder_g_proj_0_bias, weight = inner_decoder_g_proj_0_weight, x = x_29)[name = tensor("linear_41")]; tensor cumsum_mask_1_exclusive_0 = const()[name = tensor("cumsum_mask_1_exclusive_0"), val = tensor(false)]; tensor cumsum_mask_1_reverse_0 = const()[name = tensor("cumsum_mask_1_reverse_0"), val = tensor(false)]; - tensor cumsum_mask_1 = cumsum(axis = var_968, exclusive = cumsum_mask_1_exclusive_0, reverse = cumsum_mask_1_reverse_0, x = valid_mask)[name = tensor("cumsum_mask_1")]; + tensor cumsum_mask_1 = cumsum(axis = var_49, exclusive = cumsum_mask_1_exclusive_0, reverse = cumsum_mask_1_reverse_0, x = valid_mask)[name = tensor("cumsum_mask_1")]; tensor sqrt_s0_9 = sqrt(x = prev_scale_9)[name = tensor("sqrt_s0_9")]; tensor s_t_9 = add(x = prev_scale_9, y = cumsum_mask_1)[name = tensor("s_t_9")]; tensor const_20 = const()[name = tensor("const_20"), val = tensor(0x1.fffffep+127)]; - tensor clip_1 = clip(alpha = var_958, beta = const_20, x = s_t_9)[name = tensor("clip_1")]; + tensor clip_1 = clip(alpha = var_39, beta = const_20, x = s_t_9)[name = tensor("clip_1")]; tensor sqrt_s_t_9 = sqrt(x = clip_1)[name = tensor("sqrt_s_t_9")]; tensor qk_9_transpose_x_1 = const()[name = tensor("qk_9_transpose_x_1"), val = tensor(false)]; tensor qk_9_transpose_y_1 = const()[name = tensor("qk_9_transpose_y_1"), val = tensor(true)]; - tensor k_9 = transpose(perm = k_9_perm_0, x = var_1060)[name = tensor("transpose_26")]; - tensor q_9 = transpose(perm = q_9_perm_0, x = var_1052)[name = tensor("transpose_27")]; + tensor k_9 = transpose(perm = k_9_perm_0, x = var_1073)[name = tensor("transpose_26")]; + tensor q_9 = transpose(perm = q_9_perm_0, x = var_1065)[name = tensor("transpose_27")]; tensor qk_9 = matmul(transpose_x = qk_9_transpose_x_1, transpose_y = qk_9_transpose_y_1, x = q_9, y = k_9)[name = tensor("qk_9")]; - tensor var_1078 = const()[name = tensor("op_1078"), val = tensor([1, 2])]; - tensor var_1079 = reshape(shape = var_1078, x = valid_mask)[name = tensor("op_1079")]; - tensor causal_with_valid_1 = mul(x = decoder__causal_mask, y = var_1079)[name = tensor("causal_with_valid_1")]; - tensor var_1081 = const()[name = tensor("op_1081"), val = tensor([2, 1])]; - tensor var_1082 = reshape(shape = var_1081, x = sqrt_s_t_9)[name = tensor("op_1082")]; - tensor M_9 = real_div(x = causal_with_valid_1, y = var_1082)[name = tensor("M_9")]; - tensor var_1084 = mul(x = qk_9, y = M_9)[name = tensor("op_1084")]; + tensor var_1091 = const()[name = tensor("op_1091"), val = tensor([1, 2])]; + tensor var_1092 = reshape(shape = var_1091, x = valid_mask)[name = tensor("op_1092")]; + tensor causal_with_valid_1 = mul(x = inner_decoder__causal_mask, y = var_1092)[name = tensor("causal_with_valid_1")]; + tensor var_1094 = const()[name = tensor("op_1094"), val = tensor([2, 1])]; + tensor var_1095 = reshape(shape = var_1094, x = sqrt_s_t_9)[name = tensor("op_1095")]; + tensor M_9 = real_div(x = causal_with_valid_1, y = var_1095)[name = tensor("M_9")]; + tensor var_1097 = mul(x = qk_9, y = M_9)[name = tensor("op_1097")]; tensor inner_9_transpose_x_0 = const()[name = tensor("inner_9_transpose_x_0"), val = tensor(false)]; tensor inner_9_transpose_y_0 = const()[name = tensor("inner_9_transpose_y_0"), val = tensor(false)]; - tensor v_9 = transpose(perm = v_9_perm_0, x = var_1066)[name = tensor("transpose_25")]; - tensor inner_9 = matmul(transpose_x = inner_9_transpose_x_0, transpose_y = inner_9_transpose_y_0, x = var_1084, y = v_9)[name = tensor("inner_9")]; - tensor var_1086_transpose_x_0 = const()[name = tensor("op_1086_transpose_x_0"), val = tensor(false)]; - tensor var_1086_transpose_y_0 = const()[name = tensor("op_1086_transpose_y_0"), val = tensor(false)]; - tensor var_1086 = matmul(transpose_x = var_1086_transpose_x_0, transpose_y = var_1086_transpose_y_0, x = q_9, y = prev_kv_9)[name = tensor("op_1086")]; - tensor var_1087 = real_div(x = sqrt_s0_9, y = sqrt_s_t_9)[name = tensor("op_1087")]; - tensor var_1088 = const()[name = tensor("op_1088"), val = tensor([1, 1, 2, 1])]; - tensor var_1089 = reshape(shape = var_1088, x = var_1087)[name = tensor("op_1089")]; - tensor cross_9 = mul(x = var_1086, y = var_1089)[name = tensor("cross_9")]; + tensor v_9 = transpose(perm = v_9_perm_0, x = var_1079)[name = tensor("transpose_25")]; + tensor inner_9 = matmul(transpose_x = inner_9_transpose_x_0, transpose_y = inner_9_transpose_y_0, x = var_1097, y = v_9)[name = tensor("inner_9")]; + tensor var_1099_transpose_x_0 = const()[name = tensor("op_1099_transpose_x_0"), val = tensor(false)]; + tensor var_1099_transpose_y_0 = const()[name = tensor("op_1099_transpose_y_0"), val = tensor(false)]; + tensor var_1099 = matmul(transpose_x = var_1099_transpose_x_0, transpose_y = var_1099_transpose_y_0, x = q_9, y = prev_kv_9)[name = tensor("op_1099")]; + tensor var_1100 = real_div(x = sqrt_s0_9, y = sqrt_s_t_9)[name = tensor("op_1100")]; + tensor var_1101 = const()[name = tensor("op_1101"), val = tensor([1, 1, 2, 1])]; + tensor var_1102 = reshape(shape = var_1101, x = var_1100)[name = tensor("op_1102")]; + tensor cross_9 = mul(x = var_1099, y = var_1102)[name = tensor("cross_9")]; tensor out_25 = add(x = inner_9, y = cross_9)[name = tensor("out_25")]; - tensor var_1092 = const()[name = tensor("op_1092"), val = tensor([1, 1, 2, 1])]; - tensor var_1093 = reshape(shape = var_1092, x = valid_mask)[name = tensor("op_1093")]; - tensor v_masked_1 = mul(x = v_9, y = var_1093)[name = tensor("v_masked_1")]; - tensor var_1095 = mul(x = prev_kv_9, y = sqrt_s0_9)[name = tensor("op_1095")]; - tensor var_1097_transpose_x_1 = const()[name = tensor("op_1097_transpose_x_1"), val = tensor(true)]; - tensor var_1097_transpose_y_1 = const()[name = tensor("op_1097_transpose_y_1"), val = tensor(false)]; - tensor var_1097 = matmul(transpose_x = var_1097_transpose_x_1, transpose_y = var_1097_transpose_y_1, x = k_9, y = v_masked_1)[name = tensor("op_1097")]; - tensor new_kv_unnorm_9 = add(x = var_1095, y = var_1097)[name = tensor("new_kv_unnorm_9")]; - tensor var_1099_keep_dims_0 = const()[name = tensor("op_1099_keep_dims_0"), val = tensor(false)]; - tensor var_1099 = reduce_sum(keep_dims = var_1099_keep_dims_0, x = valid_mask)[name = tensor("op_1099")]; - tensor var_1100 = const()[name = tensor("op_1100"), val = tensor([1])]; - tensor var_1101 = reshape(shape = var_1100, x = var_1099)[name = tensor("op_1101")]; - tensor new_scale_9 = add(x = prev_scale_9, y = var_1101)[name = tensor("new_scale_9")]; + tensor var_1105 = const()[name = tensor("op_1105"), val = tensor([1, 1, 2, 1])]; + tensor var_1106 = reshape(shape = var_1105, x = valid_mask)[name = tensor("op_1106")]; + tensor v_masked_1 = mul(x = v_9, y = var_1106)[name = tensor("v_masked_1")]; + tensor var_1108 = mul(x = prev_kv_9, y = sqrt_s0_9)[name = tensor("op_1108")]; + tensor var_1110_transpose_x_1 = const()[name = tensor("op_1110_transpose_x_1"), val = tensor(true)]; + tensor var_1110_transpose_y_1 = const()[name = tensor("op_1110_transpose_y_1"), val = tensor(false)]; + tensor var_1110 = matmul(transpose_x = var_1110_transpose_x_1, transpose_y = var_1110_transpose_y_1, x = k_9, y = v_masked_1)[name = tensor("op_1110")]; + tensor new_kv_unnorm_9 = add(x = var_1108, y = var_1110)[name = tensor("new_kv_unnorm_9")]; + tensor var_1112_keep_dims_0 = const()[name = tensor("op_1112_keep_dims_0"), val = tensor(false)]; + tensor var_1112 = reduce_sum(keep_dims = var_1112_keep_dims_0, x = valid_mask)[name = tensor("op_1112")]; + tensor var_1113 = const()[name = tensor("op_1113"), val = tensor([1])]; + tensor var_1114 = reshape(shape = var_1113, x = var_1112)[name = tensor("op_1114")]; + tensor new_scale_9 = add(x = prev_scale_9, y = var_1114)[name = tensor("new_scale_9")]; tensor const_21 = const()[name = tensor("const_21"), val = tensor(0x1.fffffep+127)]; - tensor clip_2 = clip(alpha = var_958, beta = const_21, x = new_scale_9)[name = tensor("clip_2")]; + tensor clip_2 = clip(alpha = var_39, beta = const_21, x = new_scale_9)[name = tensor("clip_2")]; tensor sqrt_new_scale_1 = sqrt(x = clip_2)[name = tensor("sqrt_new_scale_1")]; - tensor var_1105 = real_div(x = new_kv_unnorm_9, y = sqrt_new_scale_1)[name = tensor("op_1105")]; - tensor var_1106_perm_0 = const()[name = tensor("op_1106_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_1118 = real_div(x = new_kv_unnorm_9, y = sqrt_new_scale_1)[name = tensor("op_1118")]; + tensor var_1119_perm_0 = const()[name = tensor("op_1119_perm_0"), val = tensor([0, 2, 1, 3])]; tensor out_27_axes_0 = const()[name = tensor("out_27_axes_0"), val = tensor([-1])]; - tensor var_1106 = transpose(perm = var_1106_perm_0, x = out_25)[name = tensor("transpose_24")]; - tensor out_27 = layer_norm(axes = out_27_axes_0, epsilon = var_955, x = var_1106)[name = tensor("out_27")]; - tensor var_1110 = const()[name = tensor("op_1110"), val = tensor([12, 2, 256])]; - tensor out_29 = reshape(shape = var_1110, x = out_27)[name = tensor("out_29")]; - tensor var_1112 = silu(x = input_169)[name = tensor("op_1112")]; - tensor input_171 = mul(x = var_1112, y = out_29)[name = tensor("input_171")]; - tensor ret_out_9 = linear(bias = decoder_out_proj_0_bias, weight = decoder_out_proj_0_weight, x = input_171)[name = tensor("linear_42")]; - tensor input_173 = add(x = x_29, y = ret_out_9)[name = tensor("input_173")]; + tensor var_1119 = transpose(perm = var_1119_perm_0, x = out_25)[name = tensor("transpose_24")]; + tensor out_27 = layer_norm(axes = out_27_axes_0, epsilon = var_54, x = var_1119)[name = tensor("out_27")]; + tensor var_1123 = const()[name = tensor("op_1123"), val = tensor([12, 2, 256])]; + tensor out_29 = reshape(shape = var_1123, x = out_27)[name = tensor("out_29")]; + tensor var_1125 = silu(x = input_171)[name = tensor("op_1125")]; + tensor input_173 = mul(x = var_1125, y = out_29)[name = tensor("input_173")]; + tensor ret_out_9 = linear(bias = inner_decoder_out_proj_0_bias, weight = inner_decoder_out_proj_0_weight, x = input_173)[name = tensor("linear_42")]; + tensor input_175 = add(x = x_29, y = ret_out_9)[name = tensor("input_175")]; tensor xt_1_axes_0 = const()[name = tensor("xt_1_axes_0"), val = tensor([-1])]; - tensor xt_1 = layer_norm(axes = xt_1_axes_0, beta = decoder_norm11_0_bias, epsilon = var_953, gamma = decoder_norm11_0_weight, x = input_173)[name = tensor("xt_1")]; - tensor var_1122 = const()[name = tensor("op_1122"), val = tensor([1, 12, 2, 256])]; - tensor var_1123 = reshape(shape = var_1122, x = xt_1)[name = tensor("op_1123")]; - tensor var_1124_perm_0 = const()[name = tensor("op_1124_perm_0"), val = tensor([0, 2, 1, 3])]; - tensor var_1127 = const()[name = tensor("op_1127"), val = tensor([2, 12, 256])]; - tensor var_1124 = transpose(perm = var_1124_perm_0, x = var_1123)[name = tensor("transpose_23")]; - tensor query_1 = reshape(shape = var_1127, x = var_1124)[name = tensor("query_1")]; + tensor xt_1 = layer_norm(axes = xt_1_axes_0, beta = inner_decoder_norm11_0_bias, epsilon = var_46, gamma = inner_decoder_norm11_0_weight, x = input_175)[name = tensor("xt_1")]; + tensor var_1135 = const()[name = tensor("op_1135"), val = tensor([1, 12, 2, 256])]; + tensor var_1136 = reshape(shape = var_1135, x = xt_1)[name = tensor("op_1136")]; + tensor var_1137_perm_0 = const()[name = tensor("op_1137_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_1140 = const()[name = tensor("op_1140"), val = tensor([2, 12, 256])]; + tensor var_1137 = transpose(perm = var_1137_perm_0, x = var_1136)[name = tensor("transpose_23")]; + tensor query_1 = reshape(shape = var_1140, x = var_1137)[name = tensor("query_1")]; tensor query_3_perm_0 = const()[name = tensor("query_3_perm_0"), val = tensor([1, 0, 2])]; tensor query_3 = transpose(perm = query_3_perm_0, x = query_1)[name = tensor("transpose_22")]; - tensor var_1150 = linear(bias = decoder_self_attn2_0_in_proj_bias, weight = decoder_self_attn2_0_in_proj_weight, x = query_3)[name = tensor("linear_43")]; + tensor var_1163 = linear(bias = inner_decoder_self_attn2_0_in_proj_bias, weight = inner_decoder_self_attn2_0_in_proj_weight, x = query_3)[name = tensor("linear_43")]; tensor concat_1 = const()[name = tensor("concat_1"), val = tensor([12, 2, 3, 256])]; - tensor var_1152 = reshape(shape = concat_1, x = var_1150)[name = tensor("op_1152")]; - tensor var_1153_axes_0 = const()[name = tensor("op_1153_axes_0"), val = tensor([0])]; - tensor var_1153 = expand_dims(axes = var_1153_axes_0, x = var_1152)[name = tensor("op_1153")]; - tensor var_1154_perm_0 = const()[name = tensor("op_1154_perm_0"), val = tensor([-2, 1, 2, 0, 4])]; - tensor var_1155_axes_0 = const()[name = tensor("op_1155_axes_0"), val = tensor([-2])]; - tensor var_1154 = transpose(perm = var_1154_perm_0, x = var_1153)[name = tensor("transpose_21")]; - tensor var_1155 = squeeze(axes = var_1155_axes_0, x = var_1154)[name = tensor("op_1155")]; + tensor var_1165 = reshape(shape = concat_1, x = var_1163)[name = tensor("op_1165")]; + tensor var_1166_axes_0 = const()[name = tensor("op_1166_axes_0"), val = tensor([0])]; + tensor var_1166 = expand_dims(axes = var_1166_axes_0, x = var_1165)[name = tensor("op_1166")]; + tensor var_1167_perm_0 = const()[name = tensor("op_1167_perm_0"), val = tensor([-2, 1, 2, 0, 4])]; + tensor var_1168_axes_0 = const()[name = tensor("op_1168_axes_0"), val = tensor([-2])]; + tensor var_1167 = transpose(perm = var_1167_perm_0, x = var_1166)[name = tensor("transpose_21")]; + tensor var_1168 = squeeze(axes = var_1168_axes_0, x = var_1167)[name = tensor("op_1168")]; tensor q_11_begin_0 = const()[name = tensor("q_11_begin_0"), val = tensor([0, 0, 0, 0])]; tensor q_11_end_0 = const()[name = tensor("q_11_end_0"), val = tensor([1, 12, 2, 256])]; tensor q_11_end_mask_0 = const()[name = tensor("q_11_end_mask_0"), val = tensor([false, true, true, true])]; tensor q_11_squeeze_mask_0 = const()[name = tensor("q_11_squeeze_mask_0"), val = tensor([true, false, false, false])]; - tensor q_11 = slice_by_index(begin = q_11_begin_0, end = q_11_end_0, end_mask = q_11_end_mask_0, squeeze_mask = q_11_squeeze_mask_0, x = var_1155)[name = tensor("q_11")]; + tensor q_11 = slice_by_index(begin = q_11_begin_0, end = q_11_end_0, end_mask = q_11_end_mask_0, squeeze_mask = q_11_squeeze_mask_0, x = var_1168)[name = tensor("q_11")]; tensor k_11_begin_0 = const()[name = tensor("k_11_begin_0"), val = tensor([1, 0, 0, 0])]; tensor k_11_end_0 = const()[name = tensor("k_11_end_0"), val = tensor([2, 12, 2, 256])]; tensor k_11_end_mask_0 = const()[name = tensor("k_11_end_mask_0"), val = tensor([false, true, true, true])]; tensor k_11_squeeze_mask_0 = const()[name = tensor("k_11_squeeze_mask_0"), val = tensor([true, false, false, false])]; - tensor k_11 = slice_by_index(begin = k_11_begin_0, end = k_11_end_0, end_mask = k_11_end_mask_0, squeeze_mask = k_11_squeeze_mask_0, x = var_1155)[name = tensor("k_11")]; + tensor k_11 = slice_by_index(begin = k_11_begin_0, end = k_11_end_0, end_mask = k_11_end_mask_0, squeeze_mask = k_11_squeeze_mask_0, x = var_1168)[name = tensor("k_11")]; tensor v_11_begin_0 = const()[name = tensor("v_11_begin_0"), val = tensor([2, 0, 0, 0])]; tensor v_11_end_0 = const()[name = tensor("v_11_end_0"), val = tensor([3, 12, 2, 256])]; tensor v_11_end_mask_0 = const()[name = tensor("v_11_end_mask_0"), val = tensor([false, true, true, true])]; tensor v_11_squeeze_mask_0 = const()[name = tensor("v_11_squeeze_mask_0"), val = tensor([true, false, false, false])]; - tensor v_11 = slice_by_index(begin = v_11_begin_0, end = v_11_end_0, end_mask = v_11_end_mask_0, squeeze_mask = v_11_squeeze_mask_0, x = var_1155)[name = tensor("v_11")]; - tensor var_1163 = const()[name = tensor("op_1163"), val = tensor([12, 8, 64])]; - tensor var_1164 = reshape(shape = var_1163, x = q_11)[name = tensor("op_1164")]; + tensor v_11 = slice_by_index(begin = v_11_begin_0, end = v_11_end_0, end_mask = v_11_end_mask_0, squeeze_mask = v_11_squeeze_mask_0, x = var_1168)[name = tensor("v_11")]; + tensor var_1176 = const()[name = tensor("op_1176"), val = tensor([12, 8, 64])]; + tensor var_1177 = reshape(shape = var_1176, x = q_11)[name = tensor("op_1177")]; tensor q_13_perm_0 = const()[name = tensor("q_13_perm_0"), val = tensor([1, 0, 2])]; - tensor var_1170 = const()[name = tensor("op_1170"), val = tensor([12, 8, 64])]; - tensor var_1171 = reshape(shape = var_1170, x = k_11)[name = tensor("op_1171")]; + tensor var_1183 = const()[name = tensor("op_1183"), val = tensor([12, 8, 64])]; + tensor var_1184 = reshape(shape = var_1183, x = k_11)[name = tensor("op_1184")]; tensor k_13_perm_0 = const()[name = tensor("k_13_perm_0"), val = tensor([1, 0, 2])]; - tensor var_1177 = const()[name = tensor("op_1177"), val = tensor([12, 8, 64])]; - tensor var_1178 = reshape(shape = var_1177, x = v_11)[name = tensor("op_1178")]; + tensor var_1190 = const()[name = tensor("op_1190"), val = tensor([12, 8, 64])]; + tensor var_1191 = reshape(shape = var_1190, x = v_11)[name = tensor("op_1191")]; tensor v_13_perm_0 = const()[name = tensor("v_13_perm_0"), val = tensor([1, 0, 2])]; - tensor var_1181 = const()[name = tensor("op_1181"), val = tensor([2, 4, 12, 64])]; - tensor q_13 = transpose(perm = q_13_perm_0, x = var_1164)[name = tensor("transpose_20")]; - tensor q_15 = reshape(shape = var_1181, x = q_13)[name = tensor("q_15")]; - tensor var_1183 = const()[name = tensor("op_1183"), val = tensor([2, 4, 12, 64])]; - tensor k_13 = transpose(perm = k_13_perm_0, x = var_1171)[name = tensor("transpose_19")]; - tensor k_15 = reshape(shape = var_1183, x = k_13)[name = tensor("k_15")]; - tensor var_1185 = const()[name = tensor("op_1185"), val = tensor([2, 4, 12, 64])]; - tensor v_13 = transpose(perm = v_13_perm_0, x = var_1178)[name = tensor("transpose_18")]; - tensor v_15 = reshape(shape = var_1185, x = v_13)[name = tensor("v_15")]; + tensor var_1194 = const()[name = tensor("op_1194"), val = tensor([2, 4, 12, 64])]; + tensor q_13 = transpose(perm = q_13_perm_0, x = var_1177)[name = tensor("transpose_20")]; + tensor q_15 = reshape(shape = var_1194, x = q_13)[name = tensor("q_15")]; + tensor var_1196 = const()[name = tensor("op_1196"), val = tensor([2, 4, 12, 64])]; + tensor k_13 = transpose(perm = k_13_perm_0, x = var_1184)[name = tensor("transpose_19")]; + tensor k_15 = reshape(shape = var_1196, x = k_13)[name = tensor("k_15")]; + tensor var_1198 = const()[name = tensor("op_1198"), val = tensor([2, 4, 12, 64])]; + tensor v_13 = transpose(perm = v_13_perm_0, x = var_1191)[name = tensor("transpose_18")]; + tensor v_15 = reshape(shape = var_1198, x = v_13)[name = tensor("v_15")]; tensor mul_1_y_0 = const()[name = tensor("mul_1_y_0"), val = tensor(0x1p-3)]; tensor mul_1 = mul(x = q_15, y = mul_1_y_0)[name = tensor("mul_1")]; tensor matmul_0_transpose_y_0 = const()[name = tensor("matmul_0_transpose_y_0"), val = tensor(true)]; @@ -1039,30 +1046,30 @@ program(1.0) tensor attn_output_1_transpose_x_0 = const()[name = tensor("attn_output_1_transpose_x_0"), val = tensor(false)]; tensor attn_output_1_transpose_y_0 = const()[name = tensor("attn_output_1_transpose_y_0"), val = tensor(false)]; tensor attn_output_1 = matmul(transpose_x = attn_output_1_transpose_x_0, transpose_y = attn_output_1_transpose_y_0, x = softmax_0, y = v_15)[name = tensor("attn_output_1")]; - tensor var_1188 = const()[name = tensor("op_1188"), val = tensor([2, 0, 1, 3])]; - tensor var_1193 = const()[name = tensor("op_1193"), val = tensor([24, 256])]; - tensor var_1189 = transpose(perm = var_1188, x = attn_output_1)[name = tensor("transpose_17")]; - tensor attn_output_3 = reshape(shape = var_1193, x = var_1189)[name = tensor("attn_output_3")]; - tensor attn_output_5 = linear(bias = decoder_self_attn2_0_out_proj_bias, weight = decoder_self_attn2_0_out_proj_weight, x = attn_output_3)[name = tensor("linear_44")]; - tensor var_1197 = const()[name = tensor("op_1197"), val = tensor([12, 2, 256])]; - tensor attn_output_7 = reshape(shape = var_1197, x = attn_output_5)[name = tensor("attn_output_7")]; + tensor var_1201 = const()[name = tensor("op_1201"), val = tensor([2, 0, 1, 3])]; + tensor var_1206 = const()[name = tensor("op_1206"), val = tensor([24, 256])]; + tensor var_1202 = transpose(perm = var_1201, x = attn_output_1)[name = tensor("transpose_17")]; + tensor attn_output_3 = reshape(shape = var_1206, x = var_1202)[name = tensor("attn_output_3")]; + tensor attn_output_5 = linear(bias = inner_decoder_self_attn2_0_out_proj_bias, weight = inner_decoder_self_attn2_0_out_proj_weight, x = attn_output_3)[name = tensor("linear_44")]; + tensor var_1210 = const()[name = tensor("op_1210"), val = tensor([12, 2, 256])]; + tensor attn_output_7 = reshape(shape = var_1210, x = attn_output_5)[name = tensor("attn_output_7")]; tensor sa2_out_1_perm_0 = const()[name = tensor("sa2_out_1_perm_0"), val = tensor([1, 0, 2])]; tensor sa2_out_1 = transpose(perm = sa2_out_1_perm_0, x = attn_output_7)[name = tensor("transpose_16")]; - tensor input_175 = add(x = query_1, y = sa2_out_1)[name = tensor("input_175")]; - tensor input_177_axes_0 = const()[name = tensor("input_177_axes_0"), val = tensor([-1])]; - tensor input_177 = layer_norm(axes = input_177_axes_0, beta = decoder_norm21_0_bias, epsilon = var_953, gamma = decoder_norm21_0_weight, x = input_175)[name = tensor("input_177")]; - tensor input_179 = linear(bias = decoder_linear1_0_bias, weight = decoder_linear1_0_weight, x = input_177)[name = tensor("linear_45")]; - tensor input_181 = relu(x = input_179)[name = tensor("input_181")]; - tensor ffn_out_1 = linear(bias = decoder_linear2_0_bias, weight = decoder_linear2_0_weight, x = input_181)[name = tensor("linear_46")]; - tensor input_183 = add(x = input_177, y = ffn_out_1)[name = tensor("input_183")]; + tensor input_177 = add(x = query_1, y = sa2_out_1)[name = tensor("input_177")]; + tensor input_179_axes_0 = const()[name = tensor("input_179_axes_0"), val = tensor([-1])]; + tensor input_179 = layer_norm(axes = input_179_axes_0, beta = inner_decoder_norm21_0_bias, epsilon = var_46, gamma = inner_decoder_norm21_0_weight, x = input_177)[name = tensor("input_179")]; + tensor input_181 = linear(bias = inner_decoder_linear1_0_bias, weight = inner_decoder_linear1_0_weight, x = input_179)[name = tensor("linear_45")]; + tensor input_183 = relu(x = input_181)[name = tensor("input_183")]; + tensor ffn_out_1 = linear(bias = inner_decoder_linear2_0_bias, weight = inner_decoder_linear2_0_weight, x = input_183)[name = tensor("linear_46")]; + tensor input_185 = add(x = input_179, y = ffn_out_1)[name = tensor("input_185")]; tensor xt_3_axes_0 = const()[name = tensor("xt_3_axes_0"), val = tensor([-1])]; - tensor xt_3 = layer_norm(axes = xt_3_axes_0, beta = decoder_norm22_0_bias, epsilon = var_953, gamma = decoder_norm22_0_weight, x = input_183)[name = tensor("xt_3")]; - tensor var_1217 = const()[name = tensor("op_1217"), val = tensor([1, 2, 12, 256])]; - tensor x_31 = reshape(shape = var_1217, x = xt_3)[name = tensor("x_31")]; - tensor var_1219_perm_0 = const()[name = tensor("op_1219_perm_0"), val = tensor([0, 2, 1, 3])]; - tensor var_1223 = const()[name = tensor("op_1223"), val = tensor([12, 2, 256])]; - tensor var_1219 = transpose(perm = var_1219_perm_0, x = x_31)[name = tensor("transpose_15")]; - tensor x = reshape(shape = var_1223, x = var_1219)[name = tensor("x")]; + tensor xt_3 = layer_norm(axes = xt_3_axes_0, beta = inner_decoder_norm22_0_bias, epsilon = var_46, gamma = inner_decoder_norm22_0_weight, x = input_185)[name = tensor("xt_3")]; + tensor var_1230 = const()[name = tensor("op_1230"), val = tensor([1, 2, 12, 256])]; + tensor x_31 = reshape(shape = var_1230, x = xt_3)[name = tensor("x_31")]; + tensor var_1232_perm_0 = const()[name = tensor("op_1232_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_1236 = const()[name = tensor("op_1236"), val = tensor([12, 2, 256])]; + tensor var_1232 = transpose(perm = var_1232_perm_0, x = x_31)[name = tensor("transpose_15")]; + tensor x = reshape(shape = var_1236, x = var_1232)[name = tensor("x")]; tensor prev_kv_begin_0 = const()[name = tensor("prev_kv_begin_0"), val = tensor([1, 0, 0, 0, 0])]; tensor prev_kv_end_0 = const()[name = tensor("prev_kv_end_0"), val = tensor([2, 12, 4, 64, 64])]; tensor prev_kv_end_mask_0 = const()[name = tensor("prev_kv_end_mask_0"), val = tensor([false, true, true, true, true])]; @@ -1073,120 +1080,120 @@ program(1.0) tensor prev_scale_end_mask_0 = const()[name = tensor("prev_scale_end_mask_0"), val = tensor([false, true])]; tensor prev_scale_squeeze_mask_0 = const()[name = tensor("prev_scale_squeeze_mask_0"), val = tensor([true, false])]; tensor prev_scale = slice_by_index(begin = prev_scale_begin_0, end = prev_scale_end_0, end_mask = prev_scale_end_mask_0, squeeze_mask = prev_scale_squeeze_mask_0, x = dec_scale)[name = tensor("prev_scale")]; - tensor var_1231 = linear(bias = decoder_q_proj_1_bias, weight = decoder_q_proj_1_weight, x = x)[name = tensor("linear_47")]; - tensor var_1232 = const()[name = tensor("op_1232"), val = tensor([12, 2, 4, 64])]; - tensor var_1233 = reshape(shape = var_1232, x = var_1231)[name = tensor("op_1233")]; + tensor var_1244 = linear(bias = inner_decoder_q_proj_1_bias, weight = inner_decoder_q_proj_1_weight, x = x)[name = tensor("linear_47")]; + tensor var_1245 = const()[name = tensor("op_1245"), val = tensor([12, 2, 4, 64])]; + tensor var_1246 = reshape(shape = var_1245, x = var_1244)[name = tensor("op_1246")]; tensor q_17_perm_0 = const()[name = tensor("q_17_perm_0"), val = tensor([0, 2, 1, 3])]; - tensor var_1237 = linear(bias = decoder_k_proj_1_bias, weight = decoder_k_proj_1_weight, x = x)[name = tensor("linear_48")]; - tensor var_1238 = const()[name = tensor("op_1238"), val = tensor(0x1p-3)]; - tensor var_1239 = mul(x = var_1237, y = var_1238)[name = tensor("op_1239")]; - tensor var_1240 = const()[name = tensor("op_1240"), val = tensor([12, 2, 4, 64])]; - tensor var_1241 = reshape(shape = var_1240, x = var_1239)[name = tensor("op_1241")]; + tensor var_1250 = linear(bias = inner_decoder_k_proj_1_bias, weight = inner_decoder_k_proj_1_weight, x = x)[name = tensor("linear_48")]; + tensor var_1251 = const()[name = tensor("op_1251"), val = tensor(0x1p-3)]; + tensor var_1252 = mul(x = var_1250, y = var_1251)[name = tensor("op_1252")]; + tensor var_1253 = const()[name = tensor("op_1253"), val = tensor([12, 2, 4, 64])]; + tensor var_1254 = reshape(shape = var_1253, x = var_1252)[name = tensor("op_1254")]; tensor k_17_perm_0 = const()[name = tensor("k_17_perm_0"), val = tensor([0, 2, 1, 3])]; - tensor var_1245 = linear(bias = decoder_v_proj_1_bias, weight = decoder_v_proj_1_weight, x = x)[name = tensor("linear_49")]; - tensor var_1246 = const()[name = tensor("op_1246"), val = tensor([12, 2, 4, 64])]; - tensor var_1247 = reshape(shape = var_1246, x = var_1245)[name = tensor("op_1247")]; + tensor var_1258 = linear(bias = inner_decoder_v_proj_1_bias, weight = inner_decoder_v_proj_1_weight, x = x)[name = tensor("linear_49")]; + tensor var_1259 = const()[name = tensor("op_1259"), val = tensor([12, 2, 4, 64])]; + tensor var_1260 = reshape(shape = var_1259, x = var_1258)[name = tensor("op_1260")]; tensor v_17_perm_0 = const()[name = tensor("v_17_perm_0"), val = tensor([0, 2, 1, 3])]; - tensor input_187 = linear(bias = decoder_g_proj_1_bias, weight = decoder_g_proj_1_weight, x = x)[name = tensor("linear_50")]; + tensor input_189 = linear(bias = inner_decoder_g_proj_1_bias, weight = inner_decoder_g_proj_1_weight, x = x)[name = tensor("linear_50")]; tensor sqrt_s0 = sqrt(x = prev_scale)[name = tensor("sqrt_s0")]; tensor s_t = add(x = prev_scale, y = cumsum_mask_1)[name = tensor("s_t")]; tensor const_32 = const()[name = tensor("const_32"), val = tensor(0x1.fffffep+127)]; - tensor clip_3 = clip(alpha = var_958, beta = const_32, x = s_t)[name = tensor("clip_3")]; + tensor clip_3 = clip(alpha = var_39, beta = const_32, x = s_t)[name = tensor("clip_3")]; tensor sqrt_s_t = sqrt(x = clip_3)[name = tensor("sqrt_s_t")]; tensor qk_transpose_x_1 = const()[name = tensor("qk_transpose_x_1"), val = tensor(false)]; tensor qk_transpose_y_1 = const()[name = tensor("qk_transpose_y_1"), val = tensor(true)]; - tensor k_17 = transpose(perm = k_17_perm_0, x = var_1241)[name = tensor("transpose_13")]; - tensor q_17 = transpose(perm = q_17_perm_0, x = var_1233)[name = tensor("transpose_14")]; + tensor k_17 = transpose(perm = k_17_perm_0, x = var_1254)[name = tensor("transpose_13")]; + tensor q_17 = transpose(perm = q_17_perm_0, x = var_1246)[name = tensor("transpose_14")]; tensor qk = matmul(transpose_x = qk_transpose_x_1, transpose_y = qk_transpose_y_1, x = q_17, y = k_17)[name = tensor("qk")]; - tensor var_1262 = const()[name = tensor("op_1262"), val = tensor([2, 1])]; - tensor var_1263 = reshape(shape = var_1262, x = sqrt_s_t)[name = tensor("op_1263")]; - tensor M = real_div(x = causal_with_valid_1, y = var_1263)[name = tensor("M")]; - tensor var_1265 = mul(x = qk, y = M)[name = tensor("op_1265")]; - tensor inner_transpose_x_0 = const()[name = tensor("inner_transpose_x_0"), val = tensor(false)]; - tensor inner_transpose_y_0 = const()[name = tensor("inner_transpose_y_0"), val = tensor(false)]; - tensor v_17 = transpose(perm = v_17_perm_0, x = var_1247)[name = tensor("transpose_12")]; - tensor inner = matmul(transpose_x = inner_transpose_x_0, transpose_y = inner_transpose_y_0, x = var_1265, y = v_17)[name = tensor("inner")]; - tensor var_1267_transpose_x_0 = const()[name = tensor("op_1267_transpose_x_0"), val = tensor(false)]; - tensor var_1267_transpose_y_0 = const()[name = tensor("op_1267_transpose_y_0"), val = tensor(false)]; - tensor var_1267 = matmul(transpose_x = var_1267_transpose_x_0, transpose_y = var_1267_transpose_y_0, x = q_17, y = prev_kv)[name = tensor("op_1267")]; - tensor var_1268 = real_div(x = sqrt_s0, y = sqrt_s_t)[name = tensor("op_1268")]; - tensor var_1269 = const()[name = tensor("op_1269"), val = tensor([1, 1, 2, 1])]; - tensor var_1270 = reshape(shape = var_1269, x = var_1268)[name = tensor("op_1270")]; - tensor cross = mul(x = var_1267, y = var_1270)[name = tensor("cross")]; - tensor out_31 = add(x = inner, y = cross)[name = tensor("out_31")]; - tensor v_masked = mul(x = v_17, y = var_1093)[name = tensor("v_masked")]; - tensor var_1276 = mul(x = prev_kv, y = sqrt_s0)[name = tensor("op_1276")]; - tensor var_1278_transpose_x_1 = const()[name = tensor("op_1278_transpose_x_1"), val = tensor(true)]; - tensor var_1278_transpose_y_1 = const()[name = tensor("op_1278_transpose_y_1"), val = tensor(false)]; - tensor var_1278 = matmul(transpose_x = var_1278_transpose_x_1, transpose_y = var_1278_transpose_y_1, x = k_17, y = v_masked)[name = tensor("op_1278")]; - tensor new_kv_unnorm = add(x = var_1276, y = var_1278)[name = tensor("new_kv_unnorm")]; - tensor new_scale = add(x = prev_scale, y = var_1101)[name = tensor("new_scale")]; + tensor var_1275 = const()[name = tensor("op_1275"), val = tensor([2, 1])]; + tensor var_1276 = reshape(shape = var_1275, x = sqrt_s_t)[name = tensor("op_1276")]; + tensor M = real_div(x = causal_with_valid_1, y = var_1276)[name = tensor("M")]; + tensor var_1278 = mul(x = qk, y = M)[name = tensor("op_1278")]; + tensor inner_11_transpose_x_0 = const()[name = tensor("inner_11_transpose_x_0"), val = tensor(false)]; + tensor inner_11_transpose_y_0 = const()[name = tensor("inner_11_transpose_y_0"), val = tensor(false)]; + tensor v_17 = transpose(perm = v_17_perm_0, x = var_1260)[name = tensor("transpose_12")]; + tensor inner_11 = matmul(transpose_x = inner_11_transpose_x_0, transpose_y = inner_11_transpose_y_0, x = var_1278, y = v_17)[name = tensor("inner_11")]; + tensor var_1280_transpose_x_0 = const()[name = tensor("op_1280_transpose_x_0"), val = tensor(false)]; + tensor var_1280_transpose_y_0 = const()[name = tensor("op_1280_transpose_y_0"), val = tensor(false)]; + tensor var_1280 = matmul(transpose_x = var_1280_transpose_x_0, transpose_y = var_1280_transpose_y_0, x = q_17, y = prev_kv)[name = tensor("op_1280")]; + tensor var_1281 = real_div(x = sqrt_s0, y = sqrt_s_t)[name = tensor("op_1281")]; + tensor var_1282 = const()[name = tensor("op_1282"), val = tensor([1, 1, 2, 1])]; + tensor var_1283 = reshape(shape = var_1282, x = var_1281)[name = tensor("op_1283")]; + tensor cross = mul(x = var_1280, y = var_1283)[name = tensor("cross")]; + tensor out_31 = add(x = inner_11, y = cross)[name = tensor("out_31")]; + tensor v_masked = mul(x = v_17, y = var_1106)[name = tensor("v_masked")]; + tensor var_1289 = mul(x = prev_kv, y = sqrt_s0)[name = tensor("op_1289")]; + tensor var_1291_transpose_x_1 = const()[name = tensor("op_1291_transpose_x_1"), val = tensor(true)]; + tensor var_1291_transpose_y_1 = const()[name = tensor("op_1291_transpose_y_1"), val = tensor(false)]; + tensor var_1291 = matmul(transpose_x = var_1291_transpose_x_1, transpose_y = var_1291_transpose_y_1, x = k_17, y = v_masked)[name = tensor("op_1291")]; + tensor new_kv_unnorm = add(x = var_1289, y = var_1291)[name = tensor("new_kv_unnorm")]; + tensor new_scale = add(x = prev_scale, y = var_1114)[name = tensor("new_scale")]; tensor const_33 = const()[name = tensor("const_33"), val = tensor(0x1.fffffep+127)]; - tensor clip_4 = clip(alpha = var_958, beta = const_33, x = new_scale)[name = tensor("clip_4")]; + tensor clip_4 = clip(alpha = var_39, beta = const_33, x = new_scale)[name = tensor("clip_4")]; tensor sqrt_new_scale = sqrt(x = clip_4)[name = tensor("sqrt_new_scale")]; tensor nkv = real_div(x = new_kv_unnorm, y = sqrt_new_scale)[name = tensor("nkv")]; - tensor var_1287_perm_0 = const()[name = tensor("op_1287_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_1300_perm_0 = const()[name = tensor("op_1300_perm_0"), val = tensor([0, 2, 1, 3])]; tensor out_33_axes_0 = const()[name = tensor("out_33_axes_0"), val = tensor([-1])]; - tensor var_1287 = transpose(perm = var_1287_perm_0, x = out_31)[name = tensor("transpose_11")]; - tensor out_33 = layer_norm(axes = out_33_axes_0, epsilon = var_955, x = var_1287)[name = tensor("out_33")]; - tensor var_1291 = const()[name = tensor("op_1291"), val = tensor([12, 2, 256])]; - tensor out = reshape(shape = var_1291, x = out_33)[name = tensor("out")]; - tensor var_1293 = silu(x = input_187)[name = tensor("op_1293")]; - tensor input_189 = mul(x = var_1293, y = out)[name = tensor("input_189")]; - tensor ret_out = linear(bias = decoder_out_proj_1_bias, weight = decoder_out_proj_1_weight, x = input_189)[name = tensor("linear_51")]; - tensor input_191 = add(x = x, y = ret_out)[name = tensor("input_191")]; + tensor var_1300 = transpose(perm = var_1300_perm_0, x = out_31)[name = tensor("transpose_11")]; + tensor out_33 = layer_norm(axes = out_33_axes_0, epsilon = var_54, x = var_1300)[name = tensor("out_33")]; + tensor var_1304 = const()[name = tensor("op_1304"), val = tensor([12, 2, 256])]; + tensor out = reshape(shape = var_1304, x = out_33)[name = tensor("out")]; + tensor var_1306 = silu(x = input_189)[name = tensor("op_1306")]; + tensor input_191 = mul(x = var_1306, y = out)[name = tensor("input_191")]; + tensor ret_out = linear(bias = inner_decoder_out_proj_1_bias, weight = inner_decoder_out_proj_1_weight, x = input_191)[name = tensor("linear_51")]; + tensor input_193 = add(x = x, y = ret_out)[name = tensor("input_193")]; tensor xt_5_axes_0 = const()[name = tensor("xt_5_axes_0"), val = tensor([-1])]; - tensor xt_5 = layer_norm(axes = xt_5_axes_0, beta = decoder_norm11_1_bias, epsilon = var_953, gamma = decoder_norm11_1_weight, x = input_191)[name = tensor("xt_5")]; - tensor var_1303 = const()[name = tensor("op_1303"), val = tensor([1, 12, 2, 256])]; - tensor var_1304 = reshape(shape = var_1303, x = xt_5)[name = tensor("op_1304")]; - tensor var_1305_perm_0 = const()[name = tensor("op_1305_perm_0"), val = tensor([0, 2, 1, 3])]; - tensor var_1308 = const()[name = tensor("op_1308"), val = tensor([2, 12, 256])]; - tensor var_1305 = transpose(perm = var_1305_perm_0, x = var_1304)[name = tensor("transpose_10")]; - tensor query_5 = reshape(shape = var_1308, x = var_1305)[name = tensor("query_5")]; + tensor xt_5 = layer_norm(axes = xt_5_axes_0, beta = inner_decoder_norm11_1_bias, epsilon = var_46, gamma = inner_decoder_norm11_1_weight, x = input_193)[name = tensor("xt_5")]; + tensor var_1316 = const()[name = tensor("op_1316"), val = tensor([1, 12, 2, 256])]; + tensor var_1317 = reshape(shape = var_1316, x = xt_5)[name = tensor("op_1317")]; + tensor var_1318_perm_0 = const()[name = tensor("op_1318_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_1321 = const()[name = tensor("op_1321"), val = tensor([2, 12, 256])]; + tensor var_1318 = transpose(perm = var_1318_perm_0, x = var_1317)[name = tensor("transpose_10")]; + tensor query_5 = reshape(shape = var_1321, x = var_1318)[name = tensor("query_5")]; tensor query_perm_0 = const()[name = tensor("query_perm_0"), val = tensor([1, 0, 2])]; tensor query = transpose(perm = query_perm_0, x = query_5)[name = tensor("transpose_9")]; - tensor var_1331 = linear(bias = decoder_self_attn2_1_in_proj_bias, weight = decoder_self_attn2_1_in_proj_weight, x = query)[name = tensor("linear_52")]; + tensor var_1344 = linear(bias = inner_decoder_self_attn2_1_in_proj_bias, weight = inner_decoder_self_attn2_1_in_proj_weight, x = query)[name = tensor("linear_52")]; tensor concat_2 = const()[name = tensor("concat_2"), val = tensor([12, 2, 3, 256])]; - tensor var_1333 = reshape(shape = concat_2, x = var_1331)[name = tensor("op_1333")]; - tensor var_1334_axes_0 = const()[name = tensor("op_1334_axes_0"), val = tensor([0])]; - tensor var_1334 = expand_dims(axes = var_1334_axes_0, x = var_1333)[name = tensor("op_1334")]; - tensor var_1335_perm_0 = const()[name = tensor("op_1335_perm_0"), val = tensor([-2, 1, 2, 0, 4])]; - tensor var_1336_axes_0 = const()[name = tensor("op_1336_axes_0"), val = tensor([-2])]; - tensor var_1335 = transpose(perm = var_1335_perm_0, x = var_1334)[name = tensor("transpose_8")]; - tensor var_1336 = squeeze(axes = var_1336_axes_0, x = var_1335)[name = tensor("op_1336")]; + tensor var_1346 = reshape(shape = concat_2, x = var_1344)[name = tensor("op_1346")]; + tensor var_1347_axes_0 = const()[name = tensor("op_1347_axes_0"), val = tensor([0])]; + tensor var_1347 = expand_dims(axes = var_1347_axes_0, x = var_1346)[name = tensor("op_1347")]; + tensor var_1348_perm_0 = const()[name = tensor("op_1348_perm_0"), val = tensor([-2, 1, 2, 0, 4])]; + tensor var_1349_axes_0 = const()[name = tensor("op_1349_axes_0"), val = tensor([-2])]; + tensor var_1348 = transpose(perm = var_1348_perm_0, x = var_1347)[name = tensor("transpose_8")]; + tensor var_1349 = squeeze(axes = var_1349_axes_0, x = var_1348)[name = tensor("op_1349")]; tensor q_19_begin_0 = const()[name = tensor("q_19_begin_0"), val = tensor([0, 0, 0, 0])]; tensor q_19_end_0 = const()[name = tensor("q_19_end_0"), val = tensor([1, 12, 2, 256])]; tensor q_19_end_mask_0 = const()[name = tensor("q_19_end_mask_0"), val = tensor([false, true, true, true])]; tensor q_19_squeeze_mask_0 = const()[name = tensor("q_19_squeeze_mask_0"), val = tensor([true, false, false, false])]; - tensor q_19 = slice_by_index(begin = q_19_begin_0, end = q_19_end_0, end_mask = q_19_end_mask_0, squeeze_mask = q_19_squeeze_mask_0, x = var_1336)[name = tensor("q_19")]; + tensor q_19 = slice_by_index(begin = q_19_begin_0, end = q_19_end_0, end_mask = q_19_end_mask_0, squeeze_mask = q_19_squeeze_mask_0, x = var_1349)[name = tensor("q_19")]; tensor k_19_begin_0 = const()[name = tensor("k_19_begin_0"), val = tensor([1, 0, 0, 0])]; tensor k_19_end_0 = const()[name = tensor("k_19_end_0"), val = tensor([2, 12, 2, 256])]; tensor k_19_end_mask_0 = const()[name = tensor("k_19_end_mask_0"), val = tensor([false, true, true, true])]; tensor k_19_squeeze_mask_0 = const()[name = tensor("k_19_squeeze_mask_0"), val = tensor([true, false, false, false])]; - tensor k_19 = slice_by_index(begin = k_19_begin_0, end = k_19_end_0, end_mask = k_19_end_mask_0, squeeze_mask = k_19_squeeze_mask_0, x = var_1336)[name = tensor("k_19")]; + tensor k_19 = slice_by_index(begin = k_19_begin_0, end = k_19_end_0, end_mask = k_19_end_mask_0, squeeze_mask = k_19_squeeze_mask_0, x = var_1349)[name = tensor("k_19")]; tensor v_19_begin_0 = const()[name = tensor("v_19_begin_0"), val = tensor([2, 0, 0, 0])]; tensor v_19_end_0 = const()[name = tensor("v_19_end_0"), val = tensor([3, 12, 2, 256])]; tensor v_19_end_mask_0 = const()[name = tensor("v_19_end_mask_0"), val = tensor([false, true, true, true])]; tensor v_19_squeeze_mask_0 = const()[name = tensor("v_19_squeeze_mask_0"), val = tensor([true, false, false, false])]; - tensor v_19 = slice_by_index(begin = v_19_begin_0, end = v_19_end_0, end_mask = v_19_end_mask_0, squeeze_mask = v_19_squeeze_mask_0, x = var_1336)[name = tensor("v_19")]; - tensor var_1344 = const()[name = tensor("op_1344"), val = tensor([12, 8, 64])]; - tensor var_1345 = reshape(shape = var_1344, x = q_19)[name = tensor("op_1345")]; + tensor v_19 = slice_by_index(begin = v_19_begin_0, end = v_19_end_0, end_mask = v_19_end_mask_0, squeeze_mask = v_19_squeeze_mask_0, x = var_1349)[name = tensor("v_19")]; + tensor var_1357 = const()[name = tensor("op_1357"), val = tensor([12, 8, 64])]; + tensor var_1358 = reshape(shape = var_1357, x = q_19)[name = tensor("op_1358")]; tensor q_21_perm_0 = const()[name = tensor("q_21_perm_0"), val = tensor([1, 0, 2])]; - tensor var_1351 = const()[name = tensor("op_1351"), val = tensor([12, 8, 64])]; - tensor var_1352 = reshape(shape = var_1351, x = k_19)[name = tensor("op_1352")]; + tensor var_1364 = const()[name = tensor("op_1364"), val = tensor([12, 8, 64])]; + tensor var_1365 = reshape(shape = var_1364, x = k_19)[name = tensor("op_1365")]; tensor k_21_perm_0 = const()[name = tensor("k_21_perm_0"), val = tensor([1, 0, 2])]; - tensor var_1358 = const()[name = tensor("op_1358"), val = tensor([12, 8, 64])]; - tensor var_1359 = reshape(shape = var_1358, x = v_19)[name = tensor("op_1359")]; + tensor var_1371 = const()[name = tensor("op_1371"), val = tensor([12, 8, 64])]; + tensor var_1372 = reshape(shape = var_1371, x = v_19)[name = tensor("op_1372")]; tensor v_21_perm_0 = const()[name = tensor("v_21_perm_0"), val = tensor([1, 0, 2])]; - tensor var_1362 = const()[name = tensor("op_1362"), val = tensor([2, 4, 12, 64])]; - tensor q_21 = transpose(perm = q_21_perm_0, x = var_1345)[name = tensor("transpose_7")]; - tensor q = reshape(shape = var_1362, x = q_21)[name = tensor("q")]; - tensor var_1364 = const()[name = tensor("op_1364"), val = tensor([2, 4, 12, 64])]; - tensor k_21 = transpose(perm = k_21_perm_0, x = var_1352)[name = tensor("transpose_6")]; - tensor k = reshape(shape = var_1364, x = k_21)[name = tensor("k")]; - tensor var_1366 = const()[name = tensor("op_1366"), val = tensor([2, 4, 12, 64])]; - tensor v_21 = transpose(perm = v_21_perm_0, x = var_1359)[name = tensor("transpose_5")]; - tensor v = reshape(shape = var_1366, x = v_21)[name = tensor("v")]; + tensor var_1375 = const()[name = tensor("op_1375"), val = tensor([2, 4, 12, 64])]; + tensor q_21 = transpose(perm = q_21_perm_0, x = var_1358)[name = tensor("transpose_7")]; + tensor q = reshape(shape = var_1375, x = q_21)[name = tensor("q")]; + tensor var_1377 = const()[name = tensor("op_1377"), val = tensor([2, 4, 12, 64])]; + tensor k_21 = transpose(perm = k_21_perm_0, x = var_1365)[name = tensor("transpose_6")]; + tensor k = reshape(shape = var_1377, x = k_21)[name = tensor("k")]; + tensor var_1379 = const()[name = tensor("op_1379"), val = tensor([2, 4, 12, 64])]; + tensor v_21 = transpose(perm = v_21_perm_0, x = var_1372)[name = tensor("transpose_5")]; + tensor v = reshape(shape = var_1379, x = v_21)[name = tensor("v")]; tensor mul_3_y_0 = const()[name = tensor("mul_3_y_0"), val = tensor(0x1p-3)]; tensor mul_3 = mul(x = q, y = mul_3_y_0)[name = tensor("mul_3")]; tensor matmul_1_transpose_y_0 = const()[name = tensor("matmul_1_transpose_y_0"), val = tensor(true)]; @@ -1197,36 +1204,36 @@ program(1.0) tensor attn_output_9_transpose_x_0 = const()[name = tensor("attn_output_9_transpose_x_0"), val = tensor(false)]; tensor attn_output_9_transpose_y_0 = const()[name = tensor("attn_output_9_transpose_y_0"), val = tensor(false)]; tensor attn_output_9 = matmul(transpose_x = attn_output_9_transpose_x_0, transpose_y = attn_output_9_transpose_y_0, x = softmax_1, y = v)[name = tensor("attn_output_9")]; - tensor var_1369 = const()[name = tensor("op_1369"), val = tensor([2, 0, 1, 3])]; - tensor var_1374 = const()[name = tensor("op_1374"), val = tensor([24, 256])]; - tensor var_1370 = transpose(perm = var_1369, x = attn_output_9)[name = tensor("transpose_4")]; - tensor attn_output_11 = reshape(shape = var_1374, x = var_1370)[name = tensor("attn_output_11")]; - tensor attn_output_13 = linear(bias = decoder_self_attn2_1_out_proj_bias, weight = decoder_self_attn2_1_out_proj_weight, x = attn_output_11)[name = tensor("linear_53")]; - tensor var_1378 = const()[name = tensor("op_1378"), val = tensor([12, 2, 256])]; - tensor attn_output = reshape(shape = var_1378, x = attn_output_13)[name = tensor("attn_output")]; + tensor var_1382 = const()[name = tensor("op_1382"), val = tensor([2, 0, 1, 3])]; + tensor var_1387 = const()[name = tensor("op_1387"), val = tensor([24, 256])]; + tensor var_1383 = transpose(perm = var_1382, x = attn_output_9)[name = tensor("transpose_4")]; + tensor attn_output_11 = reshape(shape = var_1387, x = var_1383)[name = tensor("attn_output_11")]; + tensor attn_output_13 = linear(bias = inner_decoder_self_attn2_1_out_proj_bias, weight = inner_decoder_self_attn2_1_out_proj_weight, x = attn_output_11)[name = tensor("linear_53")]; + tensor var_1391 = const()[name = tensor("op_1391"), val = tensor([12, 2, 256])]; + tensor attn_output = reshape(shape = var_1391, x = attn_output_13)[name = tensor("attn_output")]; tensor sa2_out_perm_0 = const()[name = tensor("sa2_out_perm_0"), val = tensor([1, 0, 2])]; tensor sa2_out = transpose(perm = sa2_out_perm_0, x = attn_output)[name = tensor("transpose_3")]; - tensor input_193 = add(x = query_5, y = sa2_out)[name = tensor("input_193")]; - tensor input_195_axes_0 = const()[name = tensor("input_195_axes_0"), val = tensor([-1])]; - tensor input_195 = layer_norm(axes = input_195_axes_0, beta = decoder_norm21_1_bias, epsilon = var_953, gamma = decoder_norm21_1_weight, x = input_193)[name = tensor("input_195")]; - tensor input_197 = linear(bias = decoder_linear1_1_bias, weight = decoder_linear1_1_weight, x = input_195)[name = tensor("linear_54")]; - tensor input_199 = relu(x = input_197)[name = tensor("input_199")]; - tensor ffn_out = linear(bias = decoder_linear2_1_bias, weight = decoder_linear2_1_weight, x = input_199)[name = tensor("linear_55")]; - tensor input_201 = add(x = input_195, y = ffn_out)[name = tensor("input_201")]; + tensor input_195 = add(x = query_5, y = sa2_out)[name = tensor("input_195")]; + tensor input_197_axes_0 = const()[name = tensor("input_197_axes_0"), val = tensor([-1])]; + tensor input_197 = layer_norm(axes = input_197_axes_0, beta = inner_decoder_norm21_1_bias, epsilon = var_46, gamma = inner_decoder_norm21_1_weight, x = input_195)[name = tensor("input_197")]; + tensor input_199 = linear(bias = inner_decoder_linear1_1_bias, weight = inner_decoder_linear1_1_weight, x = input_197)[name = tensor("linear_54")]; + tensor input_201 = relu(x = input_199)[name = tensor("input_201")]; + tensor ffn_out = linear(bias = inner_decoder_linear2_1_bias, weight = inner_decoder_linear2_1_weight, x = input_201)[name = tensor("linear_55")]; + tensor input_203 = add(x = input_197, y = ffn_out)[name = tensor("input_203")]; tensor xt_axes_0 = const()[name = tensor("xt_axes_0"), val = tensor([-1])]; - tensor xt = layer_norm(axes = xt_axes_0, beta = decoder_norm22_1_bias, epsilon = var_953, gamma = decoder_norm22_1_weight, x = input_201)[name = tensor("xt")]; - tensor var_1398 = const()[name = tensor("op_1398"), val = tensor([1, 2, 12, 256])]; - tensor input = reshape(shape = var_1398, x = xt)[name = tensor("input")]; - tensor var_1400 = const()[name = tensor("op_1400"), val = tensor([-1])]; - tensor var_1401 = reduce_l2_norm(axes = var_1400, keep_dims = var_956, x = input)[name = tensor("op_1401")]; + tensor xt = layer_norm(axes = xt_axes_0, beta = inner_decoder_norm22_1_bias, epsilon = var_46, gamma = inner_decoder_norm22_1_weight, x = input_203)[name = tensor("xt")]; + tensor var_1411 = const()[name = tensor("op_1411"), val = tensor([1, 2, 12, 256])]; + tensor input = reshape(shape = var_1411, x = xt)[name = tensor("input")]; + tensor var_1413 = const()[name = tensor("op_1413"), val = tensor([-1])]; + tensor var_1414 = reduce_l2_norm(axes = var_1413, keep_dims = var_45, x = input)[name = tensor("op_1414")]; tensor const_42 = const()[name = tensor("const_42"), val = tensor(0x1.fffffep+127)]; - tensor clip_5 = clip(alpha = var_948, beta = const_42, x = var_1401)[name = tensor("clip_5")]; - tensor var_1403 = real_div(x = input, y = clip_5)[name = tensor("op_1403")]; + tensor clip_5 = clip(alpha = var_59, beta = const_42, x = var_1414)[name = tensor("clip_5")]; + tensor var_1416 = real_div(x = input, y = clip_5)[name = tensor("op_1416")]; tensor concat_6 = const()[name = tensor("concat_6"), val = tensor([2, 1, 256])]; tensor reshape_0 = reshape(shape = concat_6, x = emb)[name = tensor("reshape_0")]; tensor transpose_1_perm_0 = const()[name = tensor("transpose_1_perm_0"), val = tensor([0, 1, 3, 2])]; tensor concat_7 = const()[name = tensor("concat_7"), val = tensor([2, 256, 12])]; - tensor transpose_1 = transpose(perm = transpose_1_perm_0, x = var_1403)[name = tensor("transpose_2")]; + tensor transpose_1 = transpose(perm = transpose_1_perm_0, x = var_1416)[name = tensor("transpose_2")]; tensor reshape_1 = reshape(shape = concat_7, x = transpose_1)[name = tensor("reshape_1")]; tensor matmul_2_transpose_x_0 = const()[name = tensor("matmul_2_transpose_x_0"), val = tensor(false)]; tensor matmul_2_transpose_y_0 = const()[name = tensor("matmul_2_transpose_y_0"), val = tensor(false)]; @@ -1237,10 +1244,10 @@ program(1.0) tensor output_end_0 = const()[name = tensor("output_end_0"), val = tensor([1, 2, 11])]; tensor output_end_mask_0 = const()[name = tensor("output_end_mask_0"), val = tensor([true, true, false])]; tensor output = slice_by_index(begin = output_begin_0, end = output_end_0, end_mask = output_end_mask_0, x = reshape_2)[name = tensor("output")]; - tensor probs = sigmoid(x = output)[name = tensor("op_1407")]; - tensor var_1409_axis_0 = const()[name = tensor("op_1409_axis_0"), val = tensor(0)]; - tensor dec_kv_new = stack(axis = var_1409_axis_0, values = (var_1105, nkv))[name = tensor("op_1409")]; - tensor var_1411_axis_0 = const()[name = tensor("op_1411_axis_0"), val = tensor(0)]; - tensor dec_scale_new = stack(axis = var_1411_axis_0, values = (new_scale_9, new_scale))[name = tensor("op_1411")]; + tensor probs = sigmoid(x = output)[name = tensor("op_1420")]; + tensor var_1422_axis_0 = const()[name = tensor("op_1422_axis_0"), val = tensor(0)]; + tensor dec_kv_new = stack(axis = var_1422_axis_0, values = (var_1118, nkv))[name = tensor("op_1422")]; + tensor var_1424_axis_0 = const()[name = tensor("op_1424_axis_0"), val = tensor(0)]; + tensor dec_scale_new = stack(axis = var_1424_axis_0, values = (new_scale_9, new_scale))[name = tensor("op_1424")]; } -> (probs, enc_kv_new, enc_scale_new, enc_conv_cache_new, cnn_window_new, dec_kv_new, dec_scale_new); } \ No newline at end of file diff --git a/optimized/dih3/200ms/ls_eend_dih3_200ms.mlpackage/Data/com.apple.CoreML/model.mlmodel b/optimized/dih3/200ms/ls_eend_dih3_200ms.mlpackage/Data/com.apple.CoreML/model.mlmodel index c03c36525cece803ff8208a5ea26cac5ac40a745..65c40d2e1c058ddda3921d57f1cd4e276a5a9a92 100644 --- a/optimized/dih3/200ms/ls_eend_dih3_200ms.mlpackage/Data/com.apple.CoreML/model.mlmodel +++ b/optimized/dih3/200ms/ls_eend_dih3_200ms.mlpackage/Data/com.apple.CoreML/model.mlmodel @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:4835a14c03b80d2ffb221a44b04b7b0b1c4cdb4e6a71682753bcfe838c3773b1 -size 179876 +oid sha256:47aad9d08e699005add53327400f5fd34df38bbf24c5cf7510373733c1983619 +size 184856 diff --git a/optimized/dih3/200ms/ls_eend_dih3_200ms.mlpackage/Manifest.json b/optimized/dih3/200ms/ls_eend_dih3_200ms.mlpackage/Manifest.json index 0fd137e70ac537ef68adfd0386c412681e663286..c6b16cff84064a7ae23053b1d8da69e7034ab07d 100644 --- a/optimized/dih3/200ms/ls_eend_dih3_200ms.mlpackage/Manifest.json +++ b/optimized/dih3/200ms/ls_eend_dih3_200ms.mlpackage/Manifest.json @@ -1,18 +1,18 @@ { "fileFormatVersion": "1.0.0", "itemInfoEntries": { - "5BAC2999-7E80-40F0-8220-7BBE6801AB8D": { + "39AD8B7F-BA07-4BF0-9035-9C5650454BFF": { "author": "com.apple.CoreML", "description": "CoreML Model Weights", "name": "weights", "path": "com.apple.CoreML/weights" }, - "ED375499-50AC-4B4A-A792-3AB2C9B91F44": { + "5B2A221F-BCEE-4C0E-8D64-CFA3E747939D": { "author": "com.apple.CoreML", "description": "CoreML Model Specification", "name": "model.mlmodel", "path": "com.apple.CoreML/model.mlmodel" } }, - "rootModelIdentifier": "ED375499-50AC-4B4A-A792-3AB2C9B91F44" + "rootModelIdentifier": "5B2A221F-BCEE-4C0E-8D64-CFA3E747939D" } diff --git a/optimized/dih3/300ms/ls_eend_dih3_300ms.mlmodelc/analytics/coremldata.bin b/optimized/dih3/300ms/ls_eend_dih3_300ms.mlmodelc/analytics/coremldata.bin index 65f7eaf6f1cff00f16e9edf9cb4cdfde14a47a56..c2bd905a2fc18436badf62025ecdac9f5ba4282a 100644 --- a/optimized/dih3/300ms/ls_eend_dih3_300ms.mlmodelc/analytics/coremldata.bin +++ b/optimized/dih3/300ms/ls_eend_dih3_300ms.mlmodelc/analytics/coremldata.bin @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:0f1ec2df8f646b9374cef6046fd9256230748f14059ab5f8e8258279b0565bda +oid sha256:09fea4de1562533e09d1e835506a267b44e274c7ab11fa1f8faf37c1580952c9 size 243 diff --git a/optimized/dih3/300ms/ls_eend_dih3_300ms.mlmodelc/coremldata.bin b/optimized/dih3/300ms/ls_eend_dih3_300ms.mlmodelc/coremldata.bin index 10dce50396088c6728d963dc8842a625a7b9facc..c6561f077fce2c8bade14393c9ef3927d4594c84 100644 --- a/optimized/dih3/300ms/ls_eend_dih3_300ms.mlmodelc/coremldata.bin +++ b/optimized/dih3/300ms/ls_eend_dih3_300ms.mlmodelc/coremldata.bin @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:b37ebc9cf057e7be9a35dd0948213af49004827a34416b9b3634e4609254a179 -size 1310 +oid sha256:d88e937e4d0ab1139e715f6bc98ec09ce1af57dfc6bcb6639bdbf18359f72e1d +size 1413 diff --git a/optimized/dih3/300ms/ls_eend_dih3_300ms.mlmodelc/metadata.json b/optimized/dih3/300ms/ls_eend_dih3_300ms.mlmodelc/metadata.json index b80d71250f01b02f0cb7ffcce159276184b82521..2783dc934f83ee0345b1f04642de92ed9d264557 100644 --- a/optimized/dih3/300ms/ls_eend_dih3_300ms.mlmodelc/metadata.json +++ b/optimized/dih3/300ms/ls_eend_dih3_300ms.mlmodelc/metadata.json @@ -1,7 +1,7 @@ [ { "metadataOutputVersion" : "3.0", - "shortDescription" : "LS-EEND DIHARD III streaming diarizer (pipeline, T=3, max_speakers=10)", + "shortDescription" : "LS-EEND DIHARD III streaming diarizer (pipeline, T=3, max_speakers=10, layout=raw_mel, cu=all)", "outputSchema" : [ { "hasShapeFlexibility" : "0", @@ -81,7 +81,7 @@ "specificationVersion" : 8, "mlProgramOperationTypeHistogram" : { "Ios16.reduceL2Norm" : 2, - "Ios17.reshape" : 68, + "Ios17.reshape" : 69, "Ios16.softmax" : 2, "Ios17.matmul" : 29, "Ios17.transpose" : 57, @@ -89,7 +89,7 @@ "Ios17.expandDims" : 3, "Ios17.add" : 46, "Ios16.sigmoid" : 5, - "Ios17.sliceByIndex" : 56, + "Ios17.sliceByIndex" : 59, "Tile" : 1, "Ios16.reduceSum" : 1, "Ios17.squeeze" : 2, @@ -101,7 +101,7 @@ "Ios16.silu" : 18, "Ios17.realDiv" : 20, "Ios17.linear" : 56, - "Stack" : 5, + "Stack" : 6, "Ios17.concat" : 18, "Ios16.relu" : 2, "Ios16.cumsum" : 1, @@ -128,9 +128,9 @@ "hasShapeFlexibility" : "0", "isOptional" : "0", "dataType" : "Float32", - "formattedType" : "MultiArray (Float32 1 × 3 × 345)", + "formattedType" : "MultiArray (Float32 1 × 35 × 23)", "shortDescription" : "", - "shape" : "[1, 3, 345]", + "shape" : "[1, 35, 23]", "name" : "features", "type" : "MultiArray" }, @@ -206,8 +206,8 @@ } ], "userDefinedMetadata" : { - "com.github.apple.coremltools.conversion_date" : "2026-04-16", - "config" : "{\"model_name\": \"dih3\", \"model_label\": \"DIHARD III\", \"variant\": \"pipeline\", \"chunk_size\": 3, \"step_duration_ms\": 300, \"frame_duration_ms\": 100, \"frame_duration_seconds\": 0.1, \"max_speakers\": 10, \"max_nspks\": 12, \"n_units\": 256, \"n_heads\": 4, \"enc_n_layers\": 4, \"dec_n_layers\": 2, \"conv_kernel_size\": 16, \"conv_delay\": 9, \"sample_rate\": 8000, \"win_length\": 200, \"hop_length\": 80, \"n_mels\": 23, \"context_size\": 7, \"subsampling\": 10, \"feat_type\": \"logmel23_cummn\", \"pure_roll\": true}", + "com.github.apple.coremltools.conversion_date" : "2026-04-18", + "config" : "{\"model_name\": \"dih3\", \"model_label\": \"DIHARD III\", \"variant\": \"pipeline\", \"chunk_size\": 3, \"step_duration_ms\": 300, \"frame_duration_ms\": 100, \"frame_duration_seconds\": 0.1, \"max_speakers\": 10, \"max_nspks\": 12, \"n_units\": 256, \"n_heads\": 4, \"enc_n_layers\": 4, \"dec_n_layers\": 2, \"conv_kernel_size\": 16, \"conv_delay\": 9, \"sample_rate\": 8000, \"win_length\": 200, \"hop_length\": 80, \"n_mels\": 23, \"context_size\": 7, \"subsampling\": 10, \"feat_type\": \"logmel23_cummn\", \"pure_roll\": true, \"input_layout\": \"raw_mel\", \"compute_units_export\": \"all\", \"raw_mel_length\": 35}", "com.github.apple.coremltools.source" : "torch==2.6.0", "com.github.apple.coremltools.version" : "9.0", "com.github.apple.coremltools.source_dialect" : "TorchScript" diff --git a/optimized/dih3/300ms/ls_eend_dih3_300ms.mlmodelc/model.mil b/optimized/dih3/300ms/ls_eend_dih3_300ms.mlmodelc/model.mil index 48f1c415ceeabfb14a7edfb748becb4366d6de9f..d54a8f41d62f449a9ccc5f0085a42da328ef56f7 100644 --- a/optimized/dih3/300ms/ls_eend_dih3_300ms.mlmodelc/model.mil +++ b/optimized/dih3/300ms/ls_eend_dih3_300ms.mlmodelc/model.mil @@ -1,234 +1,252 @@ program(1.0) [buildInfo = dict, tensor>({{"coremlc-component-MIL", "3510.2.1"}, {"coremlc-version", "3500.32.1"}, {"coremltools-component-torch", "2.6.0"}, {"coremltools-source-dialect", "TorchScript"}, {"coremltools-version", "9.0"}})] { - func main(tensor cnn_window, tensor dec_kv, tensor dec_scale, tensor enc_conv_cache, tensor enc_kv, tensor enc_scale, tensor features, tensor valid_mask) { - tensor encoder_cnn_bias = const()[name = tensor("encoder_cnn_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(64)))]; - tensor encoder_cnn_weight = const()[name = tensor("encoder_cnn_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1152)))]; - tensor encoder__causal_mask = const()[name = tensor("encoder__causal_mask"), val = tensor([[0x1p+0, 0x0p+0, 0x0p+0], [0x1p+0, 0x1p+0, 0x0p+0], [0x1p+0, 0x1p+0, 0x1p+0]])]; - tensor encoder__t_index = const()[name = tensor("encoder__t_index"), val = tensor([0x1p+0, 0x1p+1, 0x1.8p+1])]; - tensor encoder_input_projection_linear_bias = const()[name = tensor("encoder_input_projection_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4981952)))]; - tensor encoder_input_projection_linear_weight = const()[name = tensor("encoder_input_projection_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4983040)))]; - tensor encoder_pre_layer_norm_bias = const()[name = tensor("encoder_pre_layer_norm_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(5336384)))]; - tensor encoder_pre_layer_norm_weight = const()[name = tensor("encoder_pre_layer_norm_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(5337472)))]; - tensor encoder_ffn1_0_module_sequential_0_bias = const()[name = tensor("encoder_ffn1_0_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(5338560)))]; - tensor encoder_ffn1_0_module_sequential_0_weight = const()[name = tensor("encoder_ffn1_0_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(5339648)))]; - tensor encoder_ffn1_0_module_sequential_1_linear_bias = const()[name = tensor("encoder_ffn1_0_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(5340736)))]; - tensor encoder_ffn1_0_module_sequential_1_linear_weight = const()[name = tensor("encoder_ffn1_0_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(5344896)))]; - tensor encoder_ffn1_0_module_sequential_4_linear_bias = const()[name = tensor("encoder_ffn1_0_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(6393536)))]; - tensor encoder_ffn1_0_module_sequential_4_linear_weight = const()[name = tensor("encoder_ffn1_0_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(6394624)))]; - tensor encoder_ret_lns_0_bias = const()[name = tensor("encoder_ret_lns_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7443264)))]; - tensor encoder_ret_lns_0_weight = const()[name = tensor("encoder_ret_lns_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7444352)))]; - tensor encoder_q_proj_0_bias = const()[name = tensor("encoder_q_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7445440)))]; - tensor encoder_q_proj_0_weight = const()[name = tensor("encoder_q_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7446528)))]; - tensor encoder_k_proj_0_bias = const()[name = tensor("encoder_k_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7708736)))]; - tensor encoder_k_proj_0_weight = const()[name = tensor("encoder_k_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7709824)))]; - tensor encoder_v_proj_0_bias = const()[name = tensor("encoder_v_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7972032)))]; - tensor encoder_v_proj_0_weight = const()[name = tensor("encoder_v_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7973120)))]; - tensor encoder_g_proj_0_bias = const()[name = tensor("encoder_g_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8235328)))]; - tensor encoder_g_proj_0_weight = const()[name = tensor("encoder_g_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8236416)))]; - tensor encoder_out_proj_0_bias = const()[name = tensor("encoder_out_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8498624)))]; - tensor encoder_out_proj_0_weight = const()[name = tensor("encoder_out_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8499712)))]; - tensor encoder_conv_module_0_sequential_0_bias = const()[name = tensor("encoder_conv_module_0_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8761920)))]; - tensor encoder_conv_module_0_sequential_0_weight = const()[name = tensor("encoder_conv_module_0_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8763008)))]; - tensor encoder_conv_module_0_sequential_2_conv_bias = const()[name = tensor("encoder_conv_module_0_sequential_2_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8764096)))]; - tensor encoder_conv_module_0_sequential_2_conv_weight = const()[name = tensor("encoder_conv_module_0_sequential_2_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8766208)))]; - tensor encoder_conv_module_0_sequential_4_conv_weight = const()[name = tensor("encoder_conv_module_0_sequential_4_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9290560)))]; - tensor encoder_conv_module_0_sequential_5_running_var = const()[name = tensor("encoder_conv_module_0_sequential_5_running_var"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9307008)))]; - tensor encoder_conv_module_0_sequential_5_running_mean = const()[name = tensor("encoder_conv_module_0_sequential_5_running_mean"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9308096)))]; - tensor encoder_conv_module_0_sequential_5_bias = const()[name = tensor("encoder_conv_module_0_sequential_5_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9309184)))]; - tensor encoder_conv_module_0_sequential_5_weight = const()[name = tensor("encoder_conv_module_0_sequential_5_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9310272)))]; - tensor encoder_conv_module_0_sequential_7_conv_bias = const()[name = tensor("encoder_conv_module_0_sequential_7_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9311360)))]; - tensor encoder_conv_module_0_sequential_7_conv_weight = const()[name = tensor("encoder_conv_module_0_sequential_7_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9312448)))]; - tensor encoder_ffn2_0_module_sequential_0_bias = const()[name = tensor("encoder_ffn2_0_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9574656)))]; - tensor encoder_ffn2_0_module_sequential_0_weight = const()[name = tensor("encoder_ffn2_0_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9575744)))]; - tensor encoder_ffn2_0_module_sequential_1_linear_bias = const()[name = tensor("encoder_ffn2_0_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9576832)))]; - tensor encoder_ffn2_0_module_sequential_1_linear_weight = const()[name = tensor("encoder_ffn2_0_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9580992)))]; - tensor encoder_ffn2_0_module_sequential_4_linear_bias = const()[name = tensor("encoder_ffn2_0_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(10629632)))]; - tensor encoder_ffn2_0_module_sequential_4_linear_weight = const()[name = tensor("encoder_ffn2_0_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(10630720)))]; - tensor encoder_layer_norm_0_bias = const()[name = tensor("encoder_layer_norm_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11679360)))]; - tensor encoder_layer_norm_0_weight = const()[name = tensor("encoder_layer_norm_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11680448)))]; - tensor encoder_ffn1_1_module_sequential_0_bias = const()[name = tensor("encoder_ffn1_1_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11681536)))]; - tensor encoder_ffn1_1_module_sequential_0_weight = const()[name = tensor("encoder_ffn1_1_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11682624)))]; - tensor encoder_ffn1_1_module_sequential_1_linear_bias = const()[name = tensor("encoder_ffn1_1_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11683712)))]; - tensor encoder_ffn1_1_module_sequential_1_linear_weight = const()[name = tensor("encoder_ffn1_1_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11687872)))]; - tensor encoder_ffn1_1_module_sequential_4_linear_bias = const()[name = tensor("encoder_ffn1_1_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(12736512)))]; - tensor encoder_ffn1_1_module_sequential_4_linear_weight = const()[name = tensor("encoder_ffn1_1_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(12737600)))]; - tensor encoder_ret_lns_1_bias = const()[name = tensor("encoder_ret_lns_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(13786240)))]; - tensor encoder_ret_lns_1_weight = const()[name = tensor("encoder_ret_lns_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(13787328)))]; - tensor encoder_q_proj_1_bias = const()[name = tensor("encoder_q_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(13788416)))]; - tensor encoder_q_proj_1_weight = const()[name = tensor("encoder_q_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(13789504)))]; - tensor encoder_k_proj_1_bias = const()[name = tensor("encoder_k_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14051712)))]; - tensor encoder_k_proj_1_weight = const()[name = tensor("encoder_k_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14052800)))]; - tensor encoder_v_proj_1_bias = const()[name = tensor("encoder_v_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14315008)))]; - tensor encoder_v_proj_1_weight = const()[name = tensor("encoder_v_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14316096)))]; - tensor encoder_g_proj_1_bias = const()[name = tensor("encoder_g_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14578304)))]; - tensor encoder_g_proj_1_weight = const()[name = tensor("encoder_g_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14579392)))]; - tensor encoder_out_proj_1_bias = const()[name = tensor("encoder_out_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14841600)))]; - tensor encoder_out_proj_1_weight = const()[name = tensor("encoder_out_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14842688)))]; - tensor encoder_conv_module_1_sequential_0_bias = const()[name = tensor("encoder_conv_module_1_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15104896)))]; - tensor encoder_conv_module_1_sequential_0_weight = const()[name = tensor("encoder_conv_module_1_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15105984)))]; - tensor encoder_conv_module_1_sequential_2_conv_bias = const()[name = tensor("encoder_conv_module_1_sequential_2_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15107072)))]; - tensor encoder_conv_module_1_sequential_2_conv_weight = const()[name = tensor("encoder_conv_module_1_sequential_2_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15109184)))]; - tensor encoder_conv_module_1_sequential_4_conv_weight = const()[name = tensor("encoder_conv_module_1_sequential_4_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15633536)))]; - tensor encoder_conv_module_1_sequential_5_running_var = const()[name = tensor("encoder_conv_module_1_sequential_5_running_var"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15649984)))]; - tensor encoder_conv_module_1_sequential_5_running_mean = const()[name = tensor("encoder_conv_module_1_sequential_5_running_mean"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15651072)))]; - tensor encoder_conv_module_1_sequential_5_bias = const()[name = tensor("encoder_conv_module_1_sequential_5_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15652160)))]; - tensor encoder_conv_module_1_sequential_5_weight = const()[name = tensor("encoder_conv_module_1_sequential_5_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15653248)))]; - tensor encoder_conv_module_1_sequential_7_conv_bias = const()[name = tensor("encoder_conv_module_1_sequential_7_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15654336)))]; - tensor encoder_conv_module_1_sequential_7_conv_weight = const()[name = tensor("encoder_conv_module_1_sequential_7_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15655424)))]; - tensor encoder_ffn2_1_module_sequential_0_bias = const()[name = tensor("encoder_ffn2_1_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15917632)))]; - tensor encoder_ffn2_1_module_sequential_0_weight = const()[name = tensor("encoder_ffn2_1_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15918720)))]; - tensor encoder_ffn2_1_module_sequential_1_linear_bias = const()[name = tensor("encoder_ffn2_1_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15919808)))]; - tensor encoder_ffn2_1_module_sequential_1_linear_weight = const()[name = tensor("encoder_ffn2_1_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15923968)))]; - tensor encoder_ffn2_1_module_sequential_4_linear_bias = const()[name = tensor("encoder_ffn2_1_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(16972608)))]; - tensor encoder_ffn2_1_module_sequential_4_linear_weight = const()[name = tensor("encoder_ffn2_1_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(16973696)))]; - tensor encoder_layer_norm_1_bias = const()[name = tensor("encoder_layer_norm_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18022336)))]; - tensor encoder_layer_norm_1_weight = const()[name = tensor("encoder_layer_norm_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18023424)))]; - tensor encoder_ffn1_2_module_sequential_0_bias = const()[name = tensor("encoder_ffn1_2_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18024512)))]; - tensor encoder_ffn1_2_module_sequential_0_weight = const()[name = tensor("encoder_ffn1_2_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18025600)))]; - tensor encoder_ffn1_2_module_sequential_1_linear_bias = const()[name = tensor("encoder_ffn1_2_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18026688)))]; - tensor encoder_ffn1_2_module_sequential_1_linear_weight = const()[name = tensor("encoder_ffn1_2_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18030848)))]; - tensor encoder_ffn1_2_module_sequential_4_linear_bias = const()[name = tensor("encoder_ffn1_2_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(19079488)))]; - tensor encoder_ffn1_2_module_sequential_4_linear_weight = const()[name = tensor("encoder_ffn1_2_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(19080576)))]; - tensor encoder_ret_lns_2_bias = const()[name = tensor("encoder_ret_lns_2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20129216)))]; - tensor encoder_ret_lns_2_weight = const()[name = tensor("encoder_ret_lns_2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20130304)))]; - tensor encoder_q_proj_2_bias = const()[name = tensor("encoder_q_proj_2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20131392)))]; - tensor encoder_q_proj_2_weight = const()[name = tensor("encoder_q_proj_2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20132480)))]; - tensor encoder_k_proj_2_bias = const()[name = tensor("encoder_k_proj_2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20394688)))]; - tensor encoder_k_proj_2_weight = const()[name = tensor("encoder_k_proj_2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20395776)))]; - tensor encoder_v_proj_2_bias = const()[name = tensor("encoder_v_proj_2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20657984)))]; - tensor encoder_v_proj_2_weight = const()[name = tensor("encoder_v_proj_2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20659072)))]; - tensor encoder_g_proj_2_bias = const()[name = tensor("encoder_g_proj_2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20921280)))]; - tensor encoder_g_proj_2_weight = const()[name = tensor("encoder_g_proj_2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20922368)))]; - tensor encoder_out_proj_2_bias = const()[name = tensor("encoder_out_proj_2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21184576)))]; - tensor encoder_out_proj_2_weight = const()[name = tensor("encoder_out_proj_2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21185664)))]; - tensor encoder_conv_module_2_sequential_0_bias = const()[name = tensor("encoder_conv_module_2_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21447872)))]; - tensor encoder_conv_module_2_sequential_0_weight = const()[name = tensor("encoder_conv_module_2_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21448960)))]; - tensor encoder_conv_module_2_sequential_2_conv_bias = const()[name = tensor("encoder_conv_module_2_sequential_2_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21450048)))]; - tensor encoder_conv_module_2_sequential_2_conv_weight = const()[name = tensor("encoder_conv_module_2_sequential_2_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21452160)))]; - tensor encoder_conv_module_2_sequential_4_conv_weight = const()[name = tensor("encoder_conv_module_2_sequential_4_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21976512)))]; - tensor encoder_conv_module_2_sequential_5_running_var = const()[name = tensor("encoder_conv_module_2_sequential_5_running_var"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21992960)))]; - tensor encoder_conv_module_2_sequential_5_running_mean = const()[name = tensor("encoder_conv_module_2_sequential_5_running_mean"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21994048)))]; - tensor encoder_conv_module_2_sequential_5_bias = const()[name = tensor("encoder_conv_module_2_sequential_5_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21995136)))]; - tensor encoder_conv_module_2_sequential_5_weight = const()[name = tensor("encoder_conv_module_2_sequential_5_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21996224)))]; - tensor encoder_conv_module_2_sequential_7_conv_bias = const()[name = tensor("encoder_conv_module_2_sequential_7_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21997312)))]; - tensor encoder_conv_module_2_sequential_7_conv_weight = const()[name = tensor("encoder_conv_module_2_sequential_7_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21998400)))]; - tensor encoder_ffn2_2_module_sequential_0_bias = const()[name = tensor("encoder_ffn2_2_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(22260608)))]; - tensor encoder_ffn2_2_module_sequential_0_weight = const()[name = tensor("encoder_ffn2_2_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(22261696)))]; - tensor encoder_ffn2_2_module_sequential_1_linear_bias = const()[name = tensor("encoder_ffn2_2_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(22262784)))]; - tensor encoder_ffn2_2_module_sequential_1_linear_weight = const()[name = tensor("encoder_ffn2_2_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(22266944)))]; - tensor encoder_ffn2_2_module_sequential_4_linear_bias = const()[name = tensor("encoder_ffn2_2_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(23315584)))]; - tensor encoder_ffn2_2_module_sequential_4_linear_weight = const()[name = tensor("encoder_ffn2_2_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(23316672)))]; - tensor encoder_layer_norm_2_bias = const()[name = tensor("encoder_layer_norm_2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(24365312)))]; - tensor encoder_layer_norm_2_weight = const()[name = tensor("encoder_layer_norm_2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(24366400)))]; - tensor encoder_ffn1_3_module_sequential_0_bias = const()[name = tensor("encoder_ffn1_3_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(24367488)))]; - tensor encoder_ffn1_3_module_sequential_0_weight = const()[name = tensor("encoder_ffn1_3_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(24368576)))]; - tensor encoder_ffn1_3_module_sequential_1_linear_bias = const()[name = tensor("encoder_ffn1_3_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(24369664)))]; - tensor encoder_ffn1_3_module_sequential_1_linear_weight = const()[name = tensor("encoder_ffn1_3_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(24373824)))]; - tensor encoder_ffn1_3_module_sequential_4_linear_bias = const()[name = tensor("encoder_ffn1_3_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(25422464)))]; - tensor encoder_ffn1_3_module_sequential_4_linear_weight = const()[name = tensor("encoder_ffn1_3_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(25423552)))]; - tensor encoder_ret_lns_3_bias = const()[name = tensor("encoder_ret_lns_3_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(26472192)))]; - tensor encoder_ret_lns_3_weight = const()[name = tensor("encoder_ret_lns_3_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(26473280)))]; - tensor encoder_q_proj_3_bias = const()[name = tensor("encoder_q_proj_3_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(26474368)))]; - tensor encoder_q_proj_3_weight = const()[name = tensor("encoder_q_proj_3_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(26475456)))]; - tensor encoder_k_proj_3_bias = const()[name = tensor("encoder_k_proj_3_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(26737664)))]; - tensor encoder_k_proj_3_weight = const()[name = tensor("encoder_k_proj_3_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(26738752)))]; - tensor encoder_v_proj_3_bias = const()[name = tensor("encoder_v_proj_3_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27000960)))]; - tensor encoder_v_proj_3_weight = const()[name = tensor("encoder_v_proj_3_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27002048)))]; - tensor encoder_g_proj_3_bias = const()[name = tensor("encoder_g_proj_3_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27264256)))]; - tensor encoder_g_proj_3_weight = const()[name = tensor("encoder_g_proj_3_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27265344)))]; - tensor encoder_out_proj_3_bias = const()[name = tensor("encoder_out_proj_3_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27527552)))]; - tensor encoder_out_proj_3_weight = const()[name = tensor("encoder_out_proj_3_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27528640)))]; - tensor encoder_conv_module_3_sequential_0_bias = const()[name = tensor("encoder_conv_module_3_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27790848)))]; - tensor encoder_conv_module_3_sequential_0_weight = const()[name = tensor("encoder_conv_module_3_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27791936)))]; - tensor encoder_conv_module_3_sequential_2_conv_bias = const()[name = tensor("encoder_conv_module_3_sequential_2_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27793024)))]; - tensor encoder_conv_module_3_sequential_2_conv_weight = const()[name = tensor("encoder_conv_module_3_sequential_2_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27795136)))]; - tensor encoder_conv_module_3_sequential_4_conv_weight = const()[name = tensor("encoder_conv_module_3_sequential_4_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28319488)))]; - tensor encoder_conv_module_3_sequential_5_running_var = const()[name = tensor("encoder_conv_module_3_sequential_5_running_var"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28335936)))]; - tensor encoder_conv_module_3_sequential_5_running_mean = const()[name = tensor("encoder_conv_module_3_sequential_5_running_mean"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28337024)))]; - tensor encoder_conv_module_3_sequential_5_bias = const()[name = tensor("encoder_conv_module_3_sequential_5_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28338112)))]; - tensor encoder_conv_module_3_sequential_5_weight = const()[name = tensor("encoder_conv_module_3_sequential_5_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28339200)))]; - tensor encoder_conv_module_3_sequential_7_conv_bias = const()[name = tensor("encoder_conv_module_3_sequential_7_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28340288)))]; - tensor encoder_conv_module_3_sequential_7_conv_weight = const()[name = tensor("encoder_conv_module_3_sequential_7_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28341376)))]; - tensor encoder_ffn2_3_module_sequential_0_bias = const()[name = tensor("encoder_ffn2_3_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28603584)))]; - tensor encoder_ffn2_3_module_sequential_0_weight = const()[name = tensor("encoder_ffn2_3_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28604672)))]; - tensor encoder_ffn2_3_module_sequential_1_linear_bias = const()[name = tensor("encoder_ffn2_3_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28605760)))]; - tensor encoder_ffn2_3_module_sequential_1_linear_weight = const()[name = tensor("encoder_ffn2_3_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28609920)))]; - tensor encoder_ffn2_3_module_sequential_4_linear_bias = const()[name = tensor("encoder_ffn2_3_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(29658560)))]; - tensor encoder_ffn2_3_module_sequential_4_linear_weight = const()[name = tensor("encoder_ffn2_3_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(29659648)))]; - tensor encoder_layer_norm_3_bias = const()[name = tensor("encoder_layer_norm_3_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(30708288)))]; - tensor encoder_layer_norm_3_weight = const()[name = tensor("encoder_layer_norm_3_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(30709376)))]; - tensor decoder__causal_mask = const()[name = tensor("decoder__causal_mask"), val = tensor([[0x1p+0, 0x0p+0, 0x0p+0], [0x1p+0, 0x1p+0, 0x0p+0], [0x1p+0, 0x1p+0, 0x1p+0]])]; - tensor decoder_convert_bias = const()[name = tensor("decoder_convert_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(30710464)))]; - tensor decoder_convert_weight = const()[name = tensor("decoder_convert_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(30711552)))]; - tensor decoder_q_proj_0_bias = const()[name = tensor("decoder_q_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(31235904)))]; - tensor decoder_q_proj_0_weight = const()[name = tensor("decoder_q_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(31236992)))]; - tensor decoder_k_proj_0_bias = const()[name = tensor("decoder_k_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(31499200)))]; - tensor decoder_k_proj_0_weight = const()[name = tensor("decoder_k_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(31500288)))]; - tensor decoder_v_proj_0_bias = const()[name = tensor("decoder_v_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(31762496)))]; - tensor decoder_v_proj_0_weight = const()[name = tensor("decoder_v_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(31763584)))]; - tensor decoder_g_proj_0_bias = const()[name = tensor("decoder_g_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32025792)))]; - tensor decoder_g_proj_0_weight = const()[name = tensor("decoder_g_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32026880)))]; - tensor decoder_out_proj_0_bias = const()[name = tensor("decoder_out_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32289088)))]; - tensor decoder_out_proj_0_weight = const()[name = tensor("decoder_out_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32290176)))]; - tensor decoder_norm11_0_bias = const()[name = tensor("decoder_norm11_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32552384)))]; - tensor decoder_norm11_0_weight = const()[name = tensor("decoder_norm11_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32553472)))]; - tensor decoder_self_attn2_0_out_proj_bias = const()[name = tensor("decoder_self_attn2_0_out_proj_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32554560)))]; - tensor decoder_self_attn2_0_out_proj_weight = const()[name = tensor("decoder_self_attn2_0_out_proj_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32555648)))]; - tensor decoder_self_attn2_0_in_proj_bias = const()[name = tensor("decoder_self_attn2_0_in_proj_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32817856)))]; - tensor decoder_self_attn2_0_in_proj_weight = const()[name = tensor("decoder_self_attn2_0_in_proj_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32820992)))]; - tensor decoder_norm21_0_bias = const()[name = tensor("decoder_norm21_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(33607488)))]; - tensor decoder_norm21_0_weight = const()[name = tensor("decoder_norm21_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(33608576)))]; - tensor decoder_linear1_0_bias = const()[name = tensor("decoder_linear1_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(33609664)))]; - tensor decoder_linear1_0_weight = const()[name = tensor("decoder_linear1_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(33617920)))]; - tensor decoder_linear2_0_bias = const()[name = tensor("decoder_linear2_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(35715136)))]; - tensor decoder_linear2_0_weight = const()[name = tensor("decoder_linear2_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(35716224)))]; - tensor decoder_norm22_0_bias = const()[name = tensor("decoder_norm22_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(37813440)))]; - tensor decoder_norm22_0_weight = const()[name = tensor("decoder_norm22_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(37814528)))]; - tensor decoder_q_proj_1_bias = const()[name = tensor("decoder_q_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(37815616)))]; - tensor decoder_q_proj_1_weight = const()[name = tensor("decoder_q_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(37816704)))]; - tensor decoder_k_proj_1_bias = const()[name = tensor("decoder_k_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38078912)))]; - tensor decoder_k_proj_1_weight = const()[name = tensor("decoder_k_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38080000)))]; - tensor decoder_v_proj_1_bias = const()[name = tensor("decoder_v_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38342208)))]; - tensor decoder_v_proj_1_weight = const()[name = tensor("decoder_v_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38343296)))]; - tensor decoder_g_proj_1_bias = const()[name = tensor("decoder_g_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38605504)))]; - tensor decoder_g_proj_1_weight = const()[name = tensor("decoder_g_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38606592)))]; - tensor decoder_out_proj_1_bias = const()[name = tensor("decoder_out_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38868800)))]; - tensor decoder_out_proj_1_weight = const()[name = tensor("decoder_out_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38869888)))]; - tensor decoder_norm11_1_bias = const()[name = tensor("decoder_norm11_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(39132096)))]; - tensor decoder_norm11_1_weight = const()[name = tensor("decoder_norm11_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(39133184)))]; - tensor decoder_self_attn2_1_out_proj_bias = const()[name = tensor("decoder_self_attn2_1_out_proj_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(39134272)))]; - tensor decoder_self_attn2_1_out_proj_weight = const()[name = tensor("decoder_self_attn2_1_out_proj_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(39135360)))]; - tensor decoder_self_attn2_1_in_proj_bias = const()[name = tensor("decoder_self_attn2_1_in_proj_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(39397568)))]; - tensor decoder_self_attn2_1_in_proj_weight = const()[name = tensor("decoder_self_attn2_1_in_proj_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(39400704)))]; - tensor decoder_norm21_1_bias = const()[name = tensor("decoder_norm21_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(40187200)))]; - tensor decoder_norm21_1_weight = const()[name = tensor("decoder_norm21_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(40188288)))]; - tensor decoder_linear1_1_bias = const()[name = tensor("decoder_linear1_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(40189376)))]; - tensor decoder_linear1_1_weight = const()[name = tensor("decoder_linear1_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(40197632)))]; - tensor decoder_linear2_1_bias = const()[name = tensor("decoder_linear2_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(42294848)))]; - tensor decoder_linear2_1_weight = const()[name = tensor("decoder_linear2_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(42295936)))]; - tensor decoder_norm22_1_bias = const()[name = tensor("decoder_norm22_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(44393152)))]; - tensor decoder_norm22_1_weight = const()[name = tensor("decoder_norm22_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(44394240)))]; - tensor var_11 = const()[name = tensor("op_11"), val = tensor(0x1.5798eep-27)]; - tensor var_17 = const()[name = tensor("op_17"), val = tensor(0x1.0c6f7ap-20)]; - tensor var_20 = const()[name = tensor("op_20"), val = tensor(2)]; - tensor var_23 = const()[name = tensor("op_23"), val = tensor(0)]; - tensor var_26 = const()[name = tensor("op_26"), val = tensor(1)]; - tensor var_28 = const()[name = tensor("op_28"), val = tensor(0x1.4f8b58p-17)]; - tensor var_29 = const()[name = tensor("op_29"), val = tensor(true)]; - tensor input_1 = linear(bias = encoder_input_projection_linear_bias, weight = encoder_input_projection_linear_weight, x = features)[name = tensor("linear_0")]; - tensor input_3_axes_0 = const()[name = tensor("input_3_axes_0"), val = tensor([-1])]; - tensor input_3 = layer_norm(axes = input_3_axes_0, beta = encoder_pre_layer_norm_bias, epsilon = var_28, gamma = encoder_pre_layer_norm_weight, x = input_1)[name = tensor("input_3")]; + func main(tensor cnn_window, tensor dec_kv, tensor dec_scale, tensor enc_conv_cache, tensor enc_kv, tensor enc_scale, tensor features, tensor valid_mask) { + tensor inner_encoder_cnn_bias = const()[name = tensor("inner_encoder_cnn_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(64)))]; + tensor inner_encoder_cnn_weight = const()[name = tensor("inner_encoder_cnn_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1152)))]; + tensor inner_encoder__causal_mask = const()[name = tensor("inner_encoder__causal_mask"), val = tensor([[0x1p+0, 0x0p+0, 0x0p+0], [0x1p+0, 0x1p+0, 0x0p+0], [0x1p+0, 0x1p+0, 0x1p+0]])]; + tensor inner_encoder__t_index = const()[name = tensor("inner_encoder__t_index"), val = tensor([0x1p+0, 0x1p+1, 0x1.8p+1])]; + tensor inner_encoder_input_projection_linear_bias = const()[name = tensor("inner_encoder_input_projection_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4981952)))]; + tensor inner_encoder_input_projection_linear_weight = const()[name = tensor("inner_encoder_input_projection_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4983040)))]; + tensor inner_encoder_pre_layer_norm_bias = const()[name = tensor("inner_encoder_pre_layer_norm_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(5336384)))]; + tensor inner_encoder_pre_layer_norm_weight = const()[name = tensor("inner_encoder_pre_layer_norm_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(5337472)))]; + tensor inner_encoder_ffn1_0_module_sequential_0_bias = const()[name = tensor("inner_encoder_ffn1_0_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(5338560)))]; + tensor inner_encoder_ffn1_0_module_sequential_0_weight = const()[name = tensor("inner_encoder_ffn1_0_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(5339648)))]; + tensor inner_encoder_ffn1_0_module_sequential_1_linear_bias = const()[name = tensor("inner_encoder_ffn1_0_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(5340736)))]; + tensor inner_encoder_ffn1_0_module_sequential_1_linear_weight = const()[name = tensor("inner_encoder_ffn1_0_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(5344896)))]; + tensor inner_encoder_ffn1_0_module_sequential_4_linear_bias = const()[name = tensor("inner_encoder_ffn1_0_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(6393536)))]; + tensor inner_encoder_ffn1_0_module_sequential_4_linear_weight = const()[name = tensor("inner_encoder_ffn1_0_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(6394624)))]; + tensor inner_encoder_ret_lns_0_bias = const()[name = tensor("inner_encoder_ret_lns_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7443264)))]; + tensor inner_encoder_ret_lns_0_weight = const()[name = tensor("inner_encoder_ret_lns_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7444352)))]; + tensor inner_encoder_q_proj_0_bias = const()[name = tensor("inner_encoder_q_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7445440)))]; + tensor inner_encoder_q_proj_0_weight = const()[name = tensor("inner_encoder_q_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7446528)))]; + tensor inner_encoder_k_proj_0_bias = const()[name = tensor("inner_encoder_k_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7708736)))]; + tensor inner_encoder_k_proj_0_weight = const()[name = tensor("inner_encoder_k_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7709824)))]; + tensor inner_encoder_v_proj_0_bias = const()[name = tensor("inner_encoder_v_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7972032)))]; + tensor inner_encoder_v_proj_0_weight = const()[name = tensor("inner_encoder_v_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7973120)))]; + tensor inner_encoder_g_proj_0_bias = const()[name = tensor("inner_encoder_g_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8235328)))]; + tensor inner_encoder_g_proj_0_weight = const()[name = tensor("inner_encoder_g_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8236416)))]; + tensor inner_encoder_out_proj_0_bias = const()[name = tensor("inner_encoder_out_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8498624)))]; + tensor inner_encoder_out_proj_0_weight = const()[name = tensor("inner_encoder_out_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8499712)))]; + tensor inner_encoder_conv_module_0_sequential_0_bias = const()[name = tensor("inner_encoder_conv_module_0_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8761920)))]; + tensor inner_encoder_conv_module_0_sequential_0_weight = const()[name = tensor("inner_encoder_conv_module_0_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8763008)))]; + tensor inner_encoder_conv_module_0_sequential_2_conv_bias = const()[name = tensor("inner_encoder_conv_module_0_sequential_2_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8764096)))]; + tensor inner_encoder_conv_module_0_sequential_2_conv_weight = const()[name = tensor("inner_encoder_conv_module_0_sequential_2_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8766208)))]; + tensor inner_encoder_conv_module_0_sequential_4_conv_weight = const()[name = tensor("inner_encoder_conv_module_0_sequential_4_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9290560)))]; + tensor inner_encoder_conv_module_0_sequential_5_running_var = const()[name = tensor("inner_encoder_conv_module_0_sequential_5_running_var"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9307008)))]; + tensor inner_encoder_conv_module_0_sequential_5_running_mean = const()[name = tensor("inner_encoder_conv_module_0_sequential_5_running_mean"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9308096)))]; + tensor inner_encoder_conv_module_0_sequential_5_bias = const()[name = tensor("inner_encoder_conv_module_0_sequential_5_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9309184)))]; + tensor inner_encoder_conv_module_0_sequential_5_weight = const()[name = tensor("inner_encoder_conv_module_0_sequential_5_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9310272)))]; + tensor inner_encoder_conv_module_0_sequential_7_conv_bias = const()[name = tensor("inner_encoder_conv_module_0_sequential_7_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9311360)))]; + tensor inner_encoder_conv_module_0_sequential_7_conv_weight = const()[name = tensor("inner_encoder_conv_module_0_sequential_7_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9312448)))]; + tensor inner_encoder_ffn2_0_module_sequential_0_bias = const()[name = tensor("inner_encoder_ffn2_0_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9574656)))]; + tensor inner_encoder_ffn2_0_module_sequential_0_weight = const()[name = tensor("inner_encoder_ffn2_0_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9575744)))]; + tensor inner_encoder_ffn2_0_module_sequential_1_linear_bias = const()[name = tensor("inner_encoder_ffn2_0_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9576832)))]; + tensor inner_encoder_ffn2_0_module_sequential_1_linear_weight = const()[name = tensor("inner_encoder_ffn2_0_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9580992)))]; + tensor inner_encoder_ffn2_0_module_sequential_4_linear_bias = const()[name = tensor("inner_encoder_ffn2_0_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(10629632)))]; + tensor inner_encoder_ffn2_0_module_sequential_4_linear_weight = const()[name = tensor("inner_encoder_ffn2_0_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(10630720)))]; + tensor inner_encoder_layer_norm_0_bias = const()[name = tensor("inner_encoder_layer_norm_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11679360)))]; + tensor inner_encoder_layer_norm_0_weight = const()[name = tensor("inner_encoder_layer_norm_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11680448)))]; + tensor inner_encoder_ffn1_1_module_sequential_0_bias = const()[name = tensor("inner_encoder_ffn1_1_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11681536)))]; + tensor inner_encoder_ffn1_1_module_sequential_0_weight = const()[name = tensor("inner_encoder_ffn1_1_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11682624)))]; + tensor inner_encoder_ffn1_1_module_sequential_1_linear_bias = const()[name = tensor("inner_encoder_ffn1_1_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11683712)))]; + tensor inner_encoder_ffn1_1_module_sequential_1_linear_weight = const()[name = tensor("inner_encoder_ffn1_1_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11687872)))]; + tensor inner_encoder_ffn1_1_module_sequential_4_linear_bias = const()[name = tensor("inner_encoder_ffn1_1_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(12736512)))]; + tensor inner_encoder_ffn1_1_module_sequential_4_linear_weight = const()[name = tensor("inner_encoder_ffn1_1_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(12737600)))]; + tensor inner_encoder_ret_lns_1_bias = const()[name = tensor("inner_encoder_ret_lns_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(13786240)))]; + tensor inner_encoder_ret_lns_1_weight = const()[name = tensor("inner_encoder_ret_lns_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(13787328)))]; + tensor inner_encoder_q_proj_1_bias = const()[name = tensor("inner_encoder_q_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(13788416)))]; + tensor inner_encoder_q_proj_1_weight = const()[name = tensor("inner_encoder_q_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(13789504)))]; + tensor inner_encoder_k_proj_1_bias = const()[name = tensor("inner_encoder_k_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14051712)))]; + tensor inner_encoder_k_proj_1_weight = const()[name = tensor("inner_encoder_k_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14052800)))]; + tensor inner_encoder_v_proj_1_bias = const()[name = tensor("inner_encoder_v_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14315008)))]; + tensor inner_encoder_v_proj_1_weight = const()[name = tensor("inner_encoder_v_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14316096)))]; + tensor inner_encoder_g_proj_1_bias = const()[name = tensor("inner_encoder_g_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14578304)))]; + tensor inner_encoder_g_proj_1_weight = const()[name = tensor("inner_encoder_g_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14579392)))]; + tensor inner_encoder_out_proj_1_bias = const()[name = tensor("inner_encoder_out_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14841600)))]; + tensor inner_encoder_out_proj_1_weight = const()[name = tensor("inner_encoder_out_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14842688)))]; + tensor inner_encoder_conv_module_1_sequential_0_bias = const()[name = tensor("inner_encoder_conv_module_1_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15104896)))]; + tensor inner_encoder_conv_module_1_sequential_0_weight = const()[name = tensor("inner_encoder_conv_module_1_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15105984)))]; + tensor inner_encoder_conv_module_1_sequential_2_conv_bias = const()[name = tensor("inner_encoder_conv_module_1_sequential_2_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15107072)))]; + tensor inner_encoder_conv_module_1_sequential_2_conv_weight = const()[name = tensor("inner_encoder_conv_module_1_sequential_2_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15109184)))]; + tensor inner_encoder_conv_module_1_sequential_4_conv_weight = const()[name = tensor("inner_encoder_conv_module_1_sequential_4_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15633536)))]; + tensor inner_encoder_conv_module_1_sequential_5_running_var = const()[name = tensor("inner_encoder_conv_module_1_sequential_5_running_var"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15649984)))]; + tensor inner_encoder_conv_module_1_sequential_5_running_mean = const()[name = tensor("inner_encoder_conv_module_1_sequential_5_running_mean"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15651072)))]; + tensor inner_encoder_conv_module_1_sequential_5_bias = const()[name = tensor("inner_encoder_conv_module_1_sequential_5_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15652160)))]; + tensor inner_encoder_conv_module_1_sequential_5_weight = const()[name = tensor("inner_encoder_conv_module_1_sequential_5_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15653248)))]; + tensor inner_encoder_conv_module_1_sequential_7_conv_bias = const()[name = tensor("inner_encoder_conv_module_1_sequential_7_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15654336)))]; + tensor inner_encoder_conv_module_1_sequential_7_conv_weight = const()[name = tensor("inner_encoder_conv_module_1_sequential_7_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15655424)))]; + tensor inner_encoder_ffn2_1_module_sequential_0_bias = const()[name = tensor("inner_encoder_ffn2_1_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15917632)))]; + tensor inner_encoder_ffn2_1_module_sequential_0_weight = const()[name = tensor("inner_encoder_ffn2_1_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15918720)))]; + tensor inner_encoder_ffn2_1_module_sequential_1_linear_bias = const()[name = tensor("inner_encoder_ffn2_1_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15919808)))]; + tensor inner_encoder_ffn2_1_module_sequential_1_linear_weight = const()[name = tensor("inner_encoder_ffn2_1_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15923968)))]; + tensor inner_encoder_ffn2_1_module_sequential_4_linear_bias = const()[name = tensor("inner_encoder_ffn2_1_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(16972608)))]; + tensor inner_encoder_ffn2_1_module_sequential_4_linear_weight = const()[name = tensor("inner_encoder_ffn2_1_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(16973696)))]; + tensor inner_encoder_layer_norm_1_bias = const()[name = tensor("inner_encoder_layer_norm_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18022336)))]; + tensor inner_encoder_layer_norm_1_weight = const()[name = tensor("inner_encoder_layer_norm_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18023424)))]; + tensor inner_encoder_ffn1_2_module_sequential_0_bias = const()[name = tensor("inner_encoder_ffn1_2_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18024512)))]; + tensor inner_encoder_ffn1_2_module_sequential_0_weight = const()[name = tensor("inner_encoder_ffn1_2_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18025600)))]; + tensor inner_encoder_ffn1_2_module_sequential_1_linear_bias = const()[name = tensor("inner_encoder_ffn1_2_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18026688)))]; + tensor inner_encoder_ffn1_2_module_sequential_1_linear_weight = const()[name = tensor("inner_encoder_ffn1_2_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18030848)))]; + tensor inner_encoder_ffn1_2_module_sequential_4_linear_bias = const()[name = tensor("inner_encoder_ffn1_2_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(19079488)))]; + tensor inner_encoder_ffn1_2_module_sequential_4_linear_weight = const()[name = tensor("inner_encoder_ffn1_2_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(19080576)))]; + tensor inner_encoder_ret_lns_2_bias = const()[name = tensor("inner_encoder_ret_lns_2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20129216)))]; + tensor inner_encoder_ret_lns_2_weight = const()[name = tensor("inner_encoder_ret_lns_2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20130304)))]; + tensor inner_encoder_q_proj_2_bias = const()[name = tensor("inner_encoder_q_proj_2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20131392)))]; + tensor inner_encoder_q_proj_2_weight = const()[name = tensor("inner_encoder_q_proj_2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20132480)))]; + tensor inner_encoder_k_proj_2_bias = const()[name = tensor("inner_encoder_k_proj_2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20394688)))]; + tensor inner_encoder_k_proj_2_weight = const()[name = tensor("inner_encoder_k_proj_2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20395776)))]; + tensor inner_encoder_v_proj_2_bias = const()[name = tensor("inner_encoder_v_proj_2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20657984)))]; + tensor inner_encoder_v_proj_2_weight = const()[name = tensor("inner_encoder_v_proj_2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20659072)))]; + tensor inner_encoder_g_proj_2_bias = const()[name = tensor("inner_encoder_g_proj_2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20921280)))]; + tensor inner_encoder_g_proj_2_weight = const()[name = tensor("inner_encoder_g_proj_2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20922368)))]; + tensor inner_encoder_out_proj_2_bias = const()[name = tensor("inner_encoder_out_proj_2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21184576)))]; + tensor inner_encoder_out_proj_2_weight = const()[name = tensor("inner_encoder_out_proj_2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21185664)))]; + tensor inner_encoder_conv_module_2_sequential_0_bias = const()[name = tensor("inner_encoder_conv_module_2_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21447872)))]; + tensor inner_encoder_conv_module_2_sequential_0_weight = const()[name = tensor("inner_encoder_conv_module_2_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21448960)))]; + tensor inner_encoder_conv_module_2_sequential_2_conv_bias = const()[name = tensor("inner_encoder_conv_module_2_sequential_2_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21450048)))]; + tensor inner_encoder_conv_module_2_sequential_2_conv_weight = const()[name = tensor("inner_encoder_conv_module_2_sequential_2_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21452160)))]; + tensor inner_encoder_conv_module_2_sequential_4_conv_weight = const()[name = tensor("inner_encoder_conv_module_2_sequential_4_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21976512)))]; + tensor inner_encoder_conv_module_2_sequential_5_running_var = const()[name = tensor("inner_encoder_conv_module_2_sequential_5_running_var"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21992960)))]; + tensor inner_encoder_conv_module_2_sequential_5_running_mean = const()[name = tensor("inner_encoder_conv_module_2_sequential_5_running_mean"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21994048)))]; + tensor inner_encoder_conv_module_2_sequential_5_bias = const()[name = tensor("inner_encoder_conv_module_2_sequential_5_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21995136)))]; + tensor inner_encoder_conv_module_2_sequential_5_weight = const()[name = tensor("inner_encoder_conv_module_2_sequential_5_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21996224)))]; + tensor inner_encoder_conv_module_2_sequential_7_conv_bias = const()[name = tensor("inner_encoder_conv_module_2_sequential_7_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21997312)))]; + tensor inner_encoder_conv_module_2_sequential_7_conv_weight = const()[name = tensor("inner_encoder_conv_module_2_sequential_7_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21998400)))]; + tensor inner_encoder_ffn2_2_module_sequential_0_bias = const()[name = tensor("inner_encoder_ffn2_2_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(22260608)))]; + tensor inner_encoder_ffn2_2_module_sequential_0_weight = const()[name = tensor("inner_encoder_ffn2_2_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(22261696)))]; + tensor inner_encoder_ffn2_2_module_sequential_1_linear_bias = const()[name = tensor("inner_encoder_ffn2_2_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(22262784)))]; + tensor inner_encoder_ffn2_2_module_sequential_1_linear_weight = const()[name = tensor("inner_encoder_ffn2_2_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(22266944)))]; + tensor inner_encoder_ffn2_2_module_sequential_4_linear_bias = const()[name = tensor("inner_encoder_ffn2_2_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(23315584)))]; + tensor inner_encoder_ffn2_2_module_sequential_4_linear_weight = const()[name = tensor("inner_encoder_ffn2_2_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(23316672)))]; + tensor inner_encoder_layer_norm_2_bias = const()[name = tensor("inner_encoder_layer_norm_2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(24365312)))]; + tensor inner_encoder_layer_norm_2_weight = const()[name = tensor("inner_encoder_layer_norm_2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(24366400)))]; + tensor inner_encoder_ffn1_3_module_sequential_0_bias = const()[name = tensor("inner_encoder_ffn1_3_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(24367488)))]; + tensor inner_encoder_ffn1_3_module_sequential_0_weight = const()[name = tensor("inner_encoder_ffn1_3_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(24368576)))]; + tensor inner_encoder_ffn1_3_module_sequential_1_linear_bias = const()[name = tensor("inner_encoder_ffn1_3_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(24369664)))]; + tensor inner_encoder_ffn1_3_module_sequential_1_linear_weight = const()[name = tensor("inner_encoder_ffn1_3_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(24373824)))]; + tensor inner_encoder_ffn1_3_module_sequential_4_linear_bias = const()[name = tensor("inner_encoder_ffn1_3_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(25422464)))]; + tensor inner_encoder_ffn1_3_module_sequential_4_linear_weight = const()[name = tensor("inner_encoder_ffn1_3_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(25423552)))]; + tensor inner_encoder_ret_lns_3_bias = const()[name = tensor("inner_encoder_ret_lns_3_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(26472192)))]; + tensor inner_encoder_ret_lns_3_weight = const()[name = tensor("inner_encoder_ret_lns_3_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(26473280)))]; + tensor inner_encoder_q_proj_3_bias = const()[name = tensor("inner_encoder_q_proj_3_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(26474368)))]; + tensor inner_encoder_q_proj_3_weight = const()[name = tensor("inner_encoder_q_proj_3_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(26475456)))]; + tensor inner_encoder_k_proj_3_bias = const()[name = tensor("inner_encoder_k_proj_3_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(26737664)))]; + tensor inner_encoder_k_proj_3_weight = const()[name = tensor("inner_encoder_k_proj_3_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(26738752)))]; + tensor inner_encoder_v_proj_3_bias = const()[name = tensor("inner_encoder_v_proj_3_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27000960)))]; + tensor inner_encoder_v_proj_3_weight = const()[name = tensor("inner_encoder_v_proj_3_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27002048)))]; + tensor inner_encoder_g_proj_3_bias = const()[name = tensor("inner_encoder_g_proj_3_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27264256)))]; + tensor inner_encoder_g_proj_3_weight = const()[name = tensor("inner_encoder_g_proj_3_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27265344)))]; + tensor inner_encoder_out_proj_3_bias = const()[name = tensor("inner_encoder_out_proj_3_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27527552)))]; + tensor inner_encoder_out_proj_3_weight = const()[name = tensor("inner_encoder_out_proj_3_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27528640)))]; + tensor inner_encoder_conv_module_3_sequential_0_bias = const()[name = tensor("inner_encoder_conv_module_3_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27790848)))]; + tensor inner_encoder_conv_module_3_sequential_0_weight = const()[name = tensor("inner_encoder_conv_module_3_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27791936)))]; + tensor inner_encoder_conv_module_3_sequential_2_conv_bias = const()[name = tensor("inner_encoder_conv_module_3_sequential_2_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27793024)))]; + tensor inner_encoder_conv_module_3_sequential_2_conv_weight = const()[name = tensor("inner_encoder_conv_module_3_sequential_2_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27795136)))]; + tensor inner_encoder_conv_module_3_sequential_4_conv_weight = const()[name = tensor("inner_encoder_conv_module_3_sequential_4_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28319488)))]; + tensor inner_encoder_conv_module_3_sequential_5_running_var = const()[name = tensor("inner_encoder_conv_module_3_sequential_5_running_var"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28335936)))]; + tensor inner_encoder_conv_module_3_sequential_5_running_mean = const()[name = tensor("inner_encoder_conv_module_3_sequential_5_running_mean"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28337024)))]; + tensor inner_encoder_conv_module_3_sequential_5_bias = const()[name = tensor("inner_encoder_conv_module_3_sequential_5_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28338112)))]; + tensor inner_encoder_conv_module_3_sequential_5_weight = const()[name = tensor("inner_encoder_conv_module_3_sequential_5_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28339200)))]; + tensor inner_encoder_conv_module_3_sequential_7_conv_bias = const()[name = tensor("inner_encoder_conv_module_3_sequential_7_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28340288)))]; + tensor inner_encoder_conv_module_3_sequential_7_conv_weight = const()[name = tensor("inner_encoder_conv_module_3_sequential_7_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28341376)))]; + tensor inner_encoder_ffn2_3_module_sequential_0_bias = const()[name = tensor("inner_encoder_ffn2_3_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28603584)))]; + tensor inner_encoder_ffn2_3_module_sequential_0_weight = const()[name = tensor("inner_encoder_ffn2_3_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28604672)))]; + tensor inner_encoder_ffn2_3_module_sequential_1_linear_bias = const()[name = tensor("inner_encoder_ffn2_3_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28605760)))]; + tensor inner_encoder_ffn2_3_module_sequential_1_linear_weight = const()[name = tensor("inner_encoder_ffn2_3_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28609920)))]; + tensor inner_encoder_ffn2_3_module_sequential_4_linear_bias = const()[name = tensor("inner_encoder_ffn2_3_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(29658560)))]; + tensor inner_encoder_ffn2_3_module_sequential_4_linear_weight = const()[name = tensor("inner_encoder_ffn2_3_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(29659648)))]; + tensor inner_encoder_layer_norm_3_bias = const()[name = tensor("inner_encoder_layer_norm_3_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(30708288)))]; + tensor inner_encoder_layer_norm_3_weight = const()[name = tensor("inner_encoder_layer_norm_3_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(30709376)))]; + tensor inner_decoder__causal_mask = const()[name = tensor("inner_decoder__causal_mask"), val = tensor([[0x1p+0, 0x0p+0, 0x0p+0], [0x1p+0, 0x1p+0, 0x0p+0], [0x1p+0, 0x1p+0, 0x1p+0]])]; + tensor inner_decoder_convert_bias = const()[name = tensor("inner_decoder_convert_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(30710464)))]; + tensor inner_decoder_convert_weight = const()[name = tensor("inner_decoder_convert_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(30711552)))]; + tensor inner_decoder_q_proj_0_bias = const()[name = tensor("inner_decoder_q_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(31235904)))]; + tensor inner_decoder_q_proj_0_weight = const()[name = tensor("inner_decoder_q_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(31236992)))]; + tensor inner_decoder_k_proj_0_bias = const()[name = tensor("inner_decoder_k_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(31499200)))]; + tensor inner_decoder_k_proj_0_weight = const()[name = tensor("inner_decoder_k_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(31500288)))]; + tensor inner_decoder_v_proj_0_bias = const()[name = tensor("inner_decoder_v_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(31762496)))]; + tensor inner_decoder_v_proj_0_weight = const()[name = tensor("inner_decoder_v_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(31763584)))]; + tensor inner_decoder_g_proj_0_bias = const()[name = tensor("inner_decoder_g_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32025792)))]; + tensor inner_decoder_g_proj_0_weight = const()[name = tensor("inner_decoder_g_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32026880)))]; + tensor inner_decoder_out_proj_0_bias = const()[name = tensor("inner_decoder_out_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32289088)))]; + tensor inner_decoder_out_proj_0_weight = const()[name = tensor("inner_decoder_out_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32290176)))]; + tensor inner_decoder_norm11_0_bias = const()[name = tensor("inner_decoder_norm11_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32552384)))]; + tensor inner_decoder_norm11_0_weight = const()[name = tensor("inner_decoder_norm11_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32553472)))]; + tensor inner_decoder_self_attn2_0_out_proj_bias = const()[name = tensor("inner_decoder_self_attn2_0_out_proj_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32554560)))]; + tensor inner_decoder_self_attn2_0_out_proj_weight = const()[name = tensor("inner_decoder_self_attn2_0_out_proj_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32555648)))]; + tensor inner_decoder_self_attn2_0_in_proj_bias = const()[name = tensor("inner_decoder_self_attn2_0_in_proj_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32817856)))]; + tensor inner_decoder_self_attn2_0_in_proj_weight = const()[name = tensor("inner_decoder_self_attn2_0_in_proj_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32820992)))]; + tensor inner_decoder_norm21_0_bias = const()[name = tensor("inner_decoder_norm21_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(33607488)))]; + tensor inner_decoder_norm21_0_weight = const()[name = tensor("inner_decoder_norm21_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(33608576)))]; + tensor inner_decoder_linear1_0_bias = const()[name = tensor("inner_decoder_linear1_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(33609664)))]; + tensor inner_decoder_linear1_0_weight = const()[name = tensor("inner_decoder_linear1_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(33617920)))]; + tensor inner_decoder_linear2_0_bias = const()[name = tensor("inner_decoder_linear2_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(35715136)))]; + tensor inner_decoder_linear2_0_weight = const()[name = tensor("inner_decoder_linear2_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(35716224)))]; + tensor inner_decoder_norm22_0_bias = const()[name = tensor("inner_decoder_norm22_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(37813440)))]; + tensor inner_decoder_norm22_0_weight = const()[name = tensor("inner_decoder_norm22_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(37814528)))]; + tensor inner_decoder_q_proj_1_bias = const()[name = tensor("inner_decoder_q_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(37815616)))]; + tensor inner_decoder_q_proj_1_weight = const()[name = tensor("inner_decoder_q_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(37816704)))]; + tensor inner_decoder_k_proj_1_bias = const()[name = tensor("inner_decoder_k_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38078912)))]; + tensor inner_decoder_k_proj_1_weight = const()[name = tensor("inner_decoder_k_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38080000)))]; + tensor inner_decoder_v_proj_1_bias = const()[name = tensor("inner_decoder_v_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38342208)))]; + tensor inner_decoder_v_proj_1_weight = const()[name = tensor("inner_decoder_v_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38343296)))]; + tensor inner_decoder_g_proj_1_bias = const()[name = tensor("inner_decoder_g_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38605504)))]; + tensor inner_decoder_g_proj_1_weight = const()[name = tensor("inner_decoder_g_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38606592)))]; + tensor inner_decoder_out_proj_1_bias = const()[name = tensor("inner_decoder_out_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38868800)))]; + tensor inner_decoder_out_proj_1_weight = const()[name = tensor("inner_decoder_out_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38869888)))]; + tensor inner_decoder_norm11_1_bias = const()[name = tensor("inner_decoder_norm11_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(39132096)))]; + tensor inner_decoder_norm11_1_weight = const()[name = tensor("inner_decoder_norm11_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(39133184)))]; + tensor inner_decoder_self_attn2_1_out_proj_bias = const()[name = tensor("inner_decoder_self_attn2_1_out_proj_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(39134272)))]; + tensor inner_decoder_self_attn2_1_out_proj_weight = const()[name = tensor("inner_decoder_self_attn2_1_out_proj_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(39135360)))]; + tensor inner_decoder_self_attn2_1_in_proj_bias = const()[name = tensor("inner_decoder_self_attn2_1_in_proj_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(39397568)))]; + tensor inner_decoder_self_attn2_1_in_proj_weight = const()[name = tensor("inner_decoder_self_attn2_1_in_proj_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(39400704)))]; + tensor inner_decoder_norm21_1_bias = const()[name = tensor("inner_decoder_norm21_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(40187200)))]; + tensor inner_decoder_norm21_1_weight = const()[name = tensor("inner_decoder_norm21_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(40188288)))]; + tensor inner_decoder_linear1_1_bias = const()[name = tensor("inner_decoder_linear1_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(40189376)))]; + tensor inner_decoder_linear1_1_weight = const()[name = tensor("inner_decoder_linear1_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(40197632)))]; + tensor inner_decoder_linear2_1_bias = const()[name = tensor("inner_decoder_linear2_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(42294848)))]; + tensor inner_decoder_linear2_1_weight = const()[name = tensor("inner_decoder_linear2_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(42295936)))]; + tensor inner_decoder_norm22_1_bias = const()[name = tensor("inner_decoder_norm22_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(44393152)))]; + tensor inner_decoder_norm22_1_weight = const()[name = tensor("inner_decoder_norm22_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(44394240)))]; + tensor var_19_begin_0 = const()[name = tensor("op_19_begin_0"), val = tensor([0, 0, 0])]; + tensor var_19_end_0 = const()[name = tensor("op_19_end_0"), val = tensor([1, 15, 23])]; + tensor var_19_end_mask_0 = const()[name = tensor("op_19_end_mask_0"), val = tensor([true, false, true])]; + tensor var_19 = slice_by_index(begin = var_19_begin_0, end = var_19_end_0, end_mask = var_19_end_mask_0, x = features)[name = tensor("op_19")]; + tensor var_29_begin_0 = const()[name = tensor("op_29_begin_0"), val = tensor([0, 10, 0])]; + tensor var_29_end_0 = const()[name = tensor("op_29_end_0"), val = tensor([1, 25, 23])]; + tensor var_29_end_mask_0 = const()[name = tensor("op_29_end_mask_0"), val = tensor([true, false, true])]; + tensor var_29 = slice_by_index(begin = var_29_begin_0, end = var_29_end_0, end_mask = var_29_end_mask_0, x = features)[name = tensor("op_29")]; + tensor var_39_begin_0 = const()[name = tensor("op_39_begin_0"), val = tensor([0, 20, 0])]; + tensor var_39_end_0 = const()[name = tensor("op_39_end_0"), val = tensor([1, 1, 23])]; + tensor var_39_end_mask_0 = const()[name = tensor("op_39_end_mask_0"), val = tensor([true, true, true])]; + tensor var_39 = slice_by_index(begin = var_39_begin_0, end = var_39_end_0, end_mask = var_39_end_mask_0, x = features)[name = tensor("op_39")]; + tensor stacked_axis_0 = const()[name = tensor("stacked_axis_0"), val = tensor(1)]; + tensor stacked = stack(axis = stacked_axis_0, values = (var_19, var_29, var_39))[name = tensor("stacked")]; + tensor var_46 = const()[name = tensor("op_46"), val = tensor([1, 3, 345])]; + tensor input_1 = reshape(shape = var_46, x = stacked)[name = tensor("input_1")]; + tensor var_49 = const()[name = tensor("op_49"), val = tensor(0x1p+0)]; + tensor var_55 = const()[name = tensor("op_55"), val = tensor(true)]; + tensor var_56 = const()[name = tensor("op_56"), val = tensor(0x1.4f8b58p-17)]; + tensor var_59 = const()[name = tensor("op_59"), val = tensor(0)]; + tensor var_61 = const()[name = tensor("op_61"), val = tensor(2)]; + tensor var_62 = const()[name = tensor("op_62"), val = tensor(-1)]; + tensor var_64 = const()[name = tensor("op_64"), val = tensor(0x1.0c6f7ap-20)]; + tensor var_69 = const()[name = tensor("op_69"), val = tensor(0x1.5798eep-27)]; + tensor var_72 = const()[name = tensor("op_72"), val = tensor(1)]; + tensor input_3 = linear(bias = inner_encoder_input_projection_linear_bias, weight = inner_encoder_input_projection_linear_weight, x = input_1)[name = tensor("linear_0")]; tensor input_5_axes_0 = const()[name = tensor("input_5_axes_0"), val = tensor([-1])]; - tensor input_5 = layer_norm(axes = input_5_axes_0, beta = encoder_ffn1_0_module_sequential_0_bias, epsilon = var_28, gamma = encoder_ffn1_0_module_sequential_0_weight, x = input_3)[name = tensor("input_5")]; - tensor inputs_1 = linear(bias = encoder_ffn1_0_module_sequential_1_linear_bias, weight = encoder_ffn1_0_module_sequential_1_linear_weight, x = input_5)[name = tensor("linear_1")]; - tensor input_7 = silu(x = inputs_1)[name = tensor("input_7")]; - tensor input_11 = linear(bias = encoder_ffn1_0_module_sequential_4_linear_bias, weight = encoder_ffn1_0_module_sequential_4_linear_weight, x = input_7)[name = tensor("linear_2")]; - tensor var_147 = const()[name = tensor("op_147"), val = tensor(0x1p-1)]; - tensor var_148 = mul(x = input_11, y = var_147)[name = tensor("op_148")]; - tensor input_13 = add(x = var_148, y = input_3)[name = tensor("input_13")]; + tensor input_5 = layer_norm(axes = input_5_axes_0, beta = inner_encoder_pre_layer_norm_bias, epsilon = var_56, gamma = inner_encoder_pre_layer_norm_weight, x = input_3)[name = tensor("input_5")]; + tensor input_7_axes_0 = const()[name = tensor("input_7_axes_0"), val = tensor([-1])]; + tensor input_7 = layer_norm(axes = input_7_axes_0, beta = inner_encoder_ffn1_0_module_sequential_0_bias, epsilon = var_56, gamma = inner_encoder_ffn1_0_module_sequential_0_weight, x = input_5)[name = tensor("input_7")]; + tensor inputs_1 = linear(bias = inner_encoder_ffn1_0_module_sequential_1_linear_bias, weight = inner_encoder_ffn1_0_module_sequential_1_linear_weight, x = input_7)[name = tensor("linear_1")]; + tensor input_9 = silu(x = inputs_1)[name = tensor("input_9")]; + tensor input_13 = linear(bias = inner_encoder_ffn1_0_module_sequential_4_linear_bias, weight = inner_encoder_ffn1_0_module_sequential_4_linear_weight, x = input_9)[name = tensor("linear_2")]; + tensor var_193 = const()[name = tensor("op_193"), val = tensor(0x1p-1)]; + tensor var_194 = mul(x = input_13, y = var_193)[name = tensor("op_194")]; + tensor input_15 = add(x = var_194, y = input_5)[name = tensor("input_15")]; tensor x_1_axes_0 = const()[name = tensor("x_1_axes_0"), val = tensor([-1])]; - tensor x_1 = layer_norm(axes = x_1_axes_0, beta = encoder_ret_lns_0_bias, epsilon = var_28, gamma = encoder_ret_lns_0_weight, x = input_13)[name = tensor("x_1")]; + tensor x_1 = layer_norm(axes = x_1_axes_0, beta = inner_encoder_ret_lns_0_bias, epsilon = var_56, gamma = inner_encoder_ret_lns_0_weight, x = input_15)[name = tensor("x_1")]; tensor prev_kv_1_begin_0 = const()[name = tensor("prev_kv_1_begin_0"), val = tensor([0, 0, 0, 0, 0])]; tensor prev_kv_1_end_0 = const()[name = tensor("prev_kv_1_end_0"), val = tensor([1, 1, 4, 64, 64])]; tensor prev_kv_1_end_mask_0 = const()[name = tensor("prev_kv_1_end_mask_0"), val = tensor([false, true, true, true, true])]; @@ -239,163 +257,163 @@ program(1.0) tensor prev_scale_1_end_mask_0 = const()[name = tensor("prev_scale_1_end_mask_0"), val = tensor([false, true])]; tensor prev_scale_1_squeeze_mask_0 = const()[name = tensor("prev_scale_1_squeeze_mask_0"), val = tensor([true, false])]; tensor prev_scale_1 = slice_by_index(begin = prev_scale_1_begin_0, end = prev_scale_1_end_0, end_mask = prev_scale_1_end_mask_0, squeeze_mask = prev_scale_1_squeeze_mask_0, x = enc_scale)[name = tensor("prev_scale_1")]; - tensor var_162 = linear(bias = encoder_q_proj_0_bias, weight = encoder_q_proj_0_weight, x = x_1)[name = tensor("linear_3")]; - tensor var_163 = const()[name = tensor("op_163"), val = tensor([1, 3, 4, 64])]; - tensor var_164 = reshape(shape = var_163, x = var_162)[name = tensor("op_164")]; + tensor var_208 = linear(bias = inner_encoder_q_proj_0_bias, weight = inner_encoder_q_proj_0_weight, x = x_1)[name = tensor("linear_3")]; + tensor var_209 = const()[name = tensor("op_209"), val = tensor([1, 3, 4, 64])]; + tensor var_210 = reshape(shape = var_209, x = var_208)[name = tensor("op_210")]; tensor q_1_perm_0 = const()[name = tensor("q_1_perm_0"), val = tensor([0, 2, 1, 3])]; - tensor var_168 = linear(bias = encoder_k_proj_0_bias, weight = encoder_k_proj_0_weight, x = x_1)[name = tensor("linear_4")]; - tensor var_169 = const()[name = tensor("op_169"), val = tensor(0x1p-3)]; - tensor var_170 = mul(x = var_168, y = var_169)[name = tensor("op_170")]; - tensor var_171 = const()[name = tensor("op_171"), val = tensor([1, 3, 4, 64])]; - tensor var_172 = reshape(shape = var_171, x = var_170)[name = tensor("op_172")]; + tensor var_214 = linear(bias = inner_encoder_k_proj_0_bias, weight = inner_encoder_k_proj_0_weight, x = x_1)[name = tensor("linear_4")]; + tensor var_215 = const()[name = tensor("op_215"), val = tensor(0x1p-3)]; + tensor var_216 = mul(x = var_214, y = var_215)[name = tensor("op_216")]; + tensor var_217 = const()[name = tensor("op_217"), val = tensor([1, 3, 4, 64])]; + tensor var_218 = reshape(shape = var_217, x = var_216)[name = tensor("op_218")]; tensor k_1_perm_0 = const()[name = tensor("k_1_perm_0"), val = tensor([0, 2, 1, 3])]; - tensor var_176 = linear(bias = encoder_v_proj_0_bias, weight = encoder_v_proj_0_weight, x = x_1)[name = tensor("linear_5")]; - tensor var_177 = const()[name = tensor("op_177"), val = tensor([1, 3, 4, 64])]; - tensor var_178 = reshape(shape = var_177, x = var_176)[name = tensor("op_178")]; + tensor var_222 = linear(bias = inner_encoder_v_proj_0_bias, weight = inner_encoder_v_proj_0_weight, x = x_1)[name = tensor("linear_5")]; + tensor var_223 = const()[name = tensor("op_223"), val = tensor([1, 3, 4, 64])]; + tensor var_224 = reshape(shape = var_223, x = var_222)[name = tensor("op_224")]; tensor v_1_perm_0 = const()[name = tensor("v_1_perm_0"), val = tensor([0, 2, 1, 3])]; - tensor input_17 = linear(bias = encoder_g_proj_0_bias, weight = encoder_g_proj_0_weight, x = x_1)[name = tensor("linear_6")]; + tensor input_19 = linear(bias = inner_encoder_g_proj_0_bias, weight = inner_encoder_g_proj_0_weight, x = x_1)[name = tensor("linear_6")]; tensor sqrt_s0_1 = sqrt(x = prev_scale_1)[name = tensor("sqrt_s0_1")]; - tensor s_t_1 = add(x = prev_scale_1, y = encoder__t_index)[name = tensor("s_t_1")]; + tensor s_t_1 = add(x = prev_scale_1, y = inner_encoder__t_index)[name = tensor("s_t_1")]; tensor sqrt_s_t_1 = sqrt(x = s_t_1)[name = tensor("sqrt_s_t_1")]; tensor qk_1_transpose_x_1 = const()[name = tensor("qk_1_transpose_x_1"), val = tensor(false)]; tensor qk_1_transpose_y_1 = const()[name = tensor("qk_1_transpose_y_1"), val = tensor(true)]; - tensor k_1 = transpose(perm = k_1_perm_0, x = var_172)[name = tensor("transpose_57")]; - tensor q_1 = transpose(perm = q_1_perm_0, x = var_164)[name = tensor("transpose_58")]; + tensor k_1 = transpose(perm = k_1_perm_0, x = var_218)[name = tensor("transpose_57")]; + tensor q_1 = transpose(perm = q_1_perm_0, x = var_210)[name = tensor("transpose_58")]; tensor qk_1 = matmul(transpose_x = qk_1_transpose_x_1, transpose_y = qk_1_transpose_y_1, x = q_1, y = k_1)[name = tensor("qk_1")]; - tensor var_188 = const()[name = tensor("op_188"), val = tensor([3, 1])]; - tensor var_189 = reshape(shape = var_188, x = sqrt_s_t_1)[name = tensor("op_189")]; - tensor M_1 = real_div(x = encoder__causal_mask, y = var_189)[name = tensor("M_1")]; - tensor var_191 = mul(x = qk_1, y = M_1)[name = tensor("op_191")]; + tensor var_234 = const()[name = tensor("op_234"), val = tensor([3, 1])]; + tensor var_235 = reshape(shape = var_234, x = sqrt_s_t_1)[name = tensor("op_235")]; + tensor M_1 = real_div(x = inner_encoder__causal_mask, y = var_235)[name = tensor("M_1")]; + tensor var_237 = mul(x = qk_1, y = M_1)[name = tensor("op_237")]; tensor inner_1_transpose_x_0 = const()[name = tensor("inner_1_transpose_x_0"), val = tensor(false)]; tensor inner_1_transpose_y_0 = const()[name = tensor("inner_1_transpose_y_0"), val = tensor(false)]; - tensor v_1 = transpose(perm = v_1_perm_0, x = var_178)[name = tensor("transpose_56")]; - tensor inner_1 = matmul(transpose_x = inner_1_transpose_x_0, transpose_y = inner_1_transpose_y_0, x = var_191, y = v_1)[name = tensor("inner_1")]; - tensor var_193_transpose_x_0 = const()[name = tensor("op_193_transpose_x_0"), val = tensor(false)]; - tensor var_193_transpose_y_0 = const()[name = tensor("op_193_transpose_y_0"), val = tensor(false)]; - tensor var_193 = matmul(transpose_x = var_193_transpose_x_0, transpose_y = var_193_transpose_y_0, x = q_1, y = prev_kv_1)[name = tensor("op_193")]; - tensor var_194 = real_div(x = sqrt_s0_1, y = sqrt_s_t_1)[name = tensor("op_194")]; - tensor var_195 = const()[name = tensor("op_195"), val = tensor([1, 1, 3, 1])]; - tensor var_196 = reshape(shape = var_195, x = var_194)[name = tensor("op_196")]; - tensor cross_1 = mul(x = var_193, y = var_196)[name = tensor("cross_1")]; + tensor v_1 = transpose(perm = v_1_perm_0, x = var_224)[name = tensor("transpose_56")]; + tensor inner_1 = matmul(transpose_x = inner_1_transpose_x_0, transpose_y = inner_1_transpose_y_0, x = var_237, y = v_1)[name = tensor("inner_1")]; + tensor var_239_transpose_x_0 = const()[name = tensor("op_239_transpose_x_0"), val = tensor(false)]; + tensor var_239_transpose_y_0 = const()[name = tensor("op_239_transpose_y_0"), val = tensor(false)]; + tensor var_239 = matmul(transpose_x = var_239_transpose_x_0, transpose_y = var_239_transpose_y_0, x = q_1, y = prev_kv_1)[name = tensor("op_239")]; + tensor var_240 = real_div(x = sqrt_s0_1, y = sqrt_s_t_1)[name = tensor("op_240")]; + tensor var_241 = const()[name = tensor("op_241"), val = tensor([1, 1, 3, 1])]; + tensor var_242 = reshape(shape = var_241, x = var_240)[name = tensor("op_242")]; + tensor cross_1 = mul(x = var_239, y = var_242)[name = tensor("cross_1")]; tensor out_1 = add(x = inner_1, y = cross_1)[name = tensor("out_1")]; - tensor var_199 = mul(x = prev_kv_1, y = sqrt_s0_1)[name = tensor("op_199")]; - tensor var_201_transpose_x_1 = const()[name = tensor("op_201_transpose_x_1"), val = tensor(true)]; - tensor var_201_transpose_y_1 = const()[name = tensor("op_201_transpose_y_1"), val = tensor(false)]; - tensor var_201 = matmul(transpose_x = var_201_transpose_x_1, transpose_y = var_201_transpose_y_1, x = k_1, y = v_1)[name = tensor("op_201")]; - tensor new_kv_unnorm_1 = add(x = var_199, y = var_201)[name = tensor("new_kv_unnorm_1")]; - tensor var_203 = const()[name = tensor("op_203"), val = tensor(0x1.8p+1)]; - tensor new_scale_1 = add(x = prev_scale_1, y = var_203)[name = tensor("new_scale_1")]; - tensor var_205 = sqrt(x = new_scale_1)[name = tensor("op_205")]; - tensor var_206 = real_div(x = new_kv_unnorm_1, y = var_205)[name = tensor("op_206")]; - tensor var_207_perm_0 = const()[name = tensor("op_207_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_245 = mul(x = prev_kv_1, y = sqrt_s0_1)[name = tensor("op_245")]; + tensor var_247_transpose_x_1 = const()[name = tensor("op_247_transpose_x_1"), val = tensor(true)]; + tensor var_247_transpose_y_1 = const()[name = tensor("op_247_transpose_y_1"), val = tensor(false)]; + tensor var_247 = matmul(transpose_x = var_247_transpose_x_1, transpose_y = var_247_transpose_y_1, x = k_1, y = v_1)[name = tensor("op_247")]; + tensor new_kv_unnorm_1 = add(x = var_245, y = var_247)[name = tensor("new_kv_unnorm_1")]; + tensor var_249 = const()[name = tensor("op_249"), val = tensor(0x1.8p+1)]; + tensor new_scale_1 = add(x = prev_scale_1, y = var_249)[name = tensor("new_scale_1")]; + tensor var_251 = sqrt(x = new_scale_1)[name = tensor("op_251")]; + tensor var_252 = real_div(x = new_kv_unnorm_1, y = var_251)[name = tensor("op_252")]; + tensor var_253_perm_0 = const()[name = tensor("op_253_perm_0"), val = tensor([0, 2, 1, 3])]; tensor out_3_axes_0 = const()[name = tensor("out_3_axes_0"), val = tensor([-1])]; - tensor var_207 = transpose(perm = var_207_perm_0, x = out_1)[name = tensor("transpose_55")]; - tensor out_3 = layer_norm(axes = out_3_axes_0, epsilon = var_17, x = var_207)[name = tensor("out_3")]; - tensor var_211 = const()[name = tensor("op_211"), val = tensor([1, 3, 256])]; - tensor out_5 = reshape(shape = var_211, x = out_3)[name = tensor("out_5")]; - tensor var_213 = silu(x = input_17)[name = tensor("op_213")]; - tensor input_19 = mul(x = var_213, y = out_5)[name = tensor("input_19")]; - tensor ret_out_1 = linear(bias = encoder_out_proj_0_bias, weight = encoder_out_proj_0_weight, x = input_19)[name = tensor("linear_7")]; - tensor x_3 = add(x = input_13, y = ret_out_1)[name = tensor("x_3")]; + tensor var_253 = transpose(perm = var_253_perm_0, x = out_1)[name = tensor("transpose_55")]; + tensor out_3 = layer_norm(axes = out_3_axes_0, epsilon = var_64, x = var_253)[name = tensor("out_3")]; + tensor var_257 = const()[name = tensor("op_257"), val = tensor([1, 3, 256])]; + tensor out_5 = reshape(shape = var_257, x = out_3)[name = tensor("out_5")]; + tensor var_259 = silu(x = input_19)[name = tensor("op_259")]; + tensor input_21 = mul(x = var_259, y = out_5)[name = tensor("input_21")]; + tensor ret_out_1 = linear(bias = inner_encoder_out_proj_0_bias, weight = inner_encoder_out_proj_0_weight, x = input_21)[name = tensor("linear_7")]; + tensor x_3 = add(x = input_15, y = ret_out_1)[name = tensor("x_3")]; tensor window_1_begin_0 = const()[name = tensor("window_1_begin_0"), val = tensor([0, 0, 0, 0])]; tensor window_1_end_0 = const()[name = tensor("window_1_end_0"), val = tensor([1, 1, 16, 256])]; tensor window_1_end_mask_0 = const()[name = tensor("window_1_end_mask_0"), val = tensor([false, true, true, true])]; tensor window_1_squeeze_mask_0 = const()[name = tensor("window_1_squeeze_mask_0"), val = tensor([true, false, false, false])]; tensor window_1 = slice_by_index(begin = window_1_begin_0, end = window_1_end_0, end_mask = window_1_end_mask_0, squeeze_mask = window_1_squeeze_mask_0, x = enc_conv_cache)[name = tensor("window_1")]; - tensor var_221_begin_0 = const()[name = tensor("op_221_begin_0"), val = tensor([0, 0, 0])]; - tensor var_221_end_0 = const()[name = tensor("op_221_end_0"), val = tensor([1, 1, 256])]; - tensor var_221_end_mask_0 = const()[name = tensor("op_221_end_mask_0"), val = tensor([true, false, true])]; - tensor var_221 = slice_by_index(begin = var_221_begin_0, end = var_221_end_0, end_mask = var_221_end_mask_0, x = x_3)[name = tensor("op_221")]; - tensor var_224_begin_0 = const()[name = tensor("op_224_begin_0"), val = tensor([0, 1, 0])]; - tensor var_224_end_0 = const()[name = tensor("op_224_end_0"), val = tensor([1, 16, 256])]; - tensor var_224_end_mask_0 = const()[name = tensor("op_224_end_mask_0"), val = tensor([true, true, true])]; - tensor var_224 = slice_by_index(begin = var_224_begin_0, end = var_224_end_0, end_mask = var_224_end_mask_0, x = window_1)[name = tensor("op_224")]; + tensor var_267_begin_0 = const()[name = tensor("op_267_begin_0"), val = tensor([0, 0, 0])]; + tensor var_267_end_0 = const()[name = tensor("op_267_end_0"), val = tensor([1, 1, 256])]; + tensor var_267_end_mask_0 = const()[name = tensor("op_267_end_mask_0"), val = tensor([true, false, true])]; + tensor var_267 = slice_by_index(begin = var_267_begin_0, end = var_267_end_0, end_mask = var_267_end_mask_0, x = x_3)[name = tensor("op_267")]; + tensor var_270_begin_0 = const()[name = tensor("op_270_begin_0"), val = tensor([0, 1, 0])]; + tensor var_270_end_0 = const()[name = tensor("op_270_end_0"), val = tensor([1, 16, 256])]; + tensor var_270_end_mask_0 = const()[name = tensor("op_270_end_mask_0"), val = tensor([true, true, true])]; + tensor var_270 = slice_by_index(begin = var_270_begin_0, end = var_270_end_0, end_mask = var_270_end_mask_0, x = window_1)[name = tensor("op_270")]; tensor window_3_interleave_0 = const()[name = tensor("window_3_interleave_0"), val = tensor(false)]; - tensor window_3 = concat(axis = var_26, interleave = window_3_interleave_0, values = (var_224, var_221))[name = tensor("window_3")]; - tensor var_229_begin_0 = const()[name = tensor("op_229_begin_0"), val = tensor([0, 1, 0])]; - tensor var_229_end_0 = const()[name = tensor("op_229_end_0"), val = tensor([1, 2, 256])]; - tensor var_229_end_mask_0 = const()[name = tensor("op_229_end_mask_0"), val = tensor([true, false, true])]; - tensor var_229 = slice_by_index(begin = var_229_begin_0, end = var_229_end_0, end_mask = var_229_end_mask_0, x = x_3)[name = tensor("op_229")]; - tensor var_232_begin_0 = const()[name = tensor("op_232_begin_0"), val = tensor([0, 1, 0])]; - tensor var_232_end_0 = const()[name = tensor("op_232_end_0"), val = tensor([1, 16, 256])]; - tensor var_232_end_mask_0 = const()[name = tensor("op_232_end_mask_0"), val = tensor([true, true, true])]; - tensor var_232 = slice_by_index(begin = var_232_begin_0, end = var_232_end_0, end_mask = var_232_end_mask_0, x = window_3)[name = tensor("op_232")]; + tensor window_3 = concat(axis = var_72, interleave = window_3_interleave_0, values = (var_270, var_267))[name = tensor("window_3")]; + tensor var_275_begin_0 = const()[name = tensor("op_275_begin_0"), val = tensor([0, 1, 0])]; + tensor var_275_end_0 = const()[name = tensor("op_275_end_0"), val = tensor([1, 2, 256])]; + tensor var_275_end_mask_0 = const()[name = tensor("op_275_end_mask_0"), val = tensor([true, false, true])]; + tensor var_275 = slice_by_index(begin = var_275_begin_0, end = var_275_end_0, end_mask = var_275_end_mask_0, x = x_3)[name = tensor("op_275")]; + tensor var_278_begin_0 = const()[name = tensor("op_278_begin_0"), val = tensor([0, 1, 0])]; + tensor var_278_end_0 = const()[name = tensor("op_278_end_0"), val = tensor([1, 16, 256])]; + tensor var_278_end_mask_0 = const()[name = tensor("op_278_end_mask_0"), val = tensor([true, true, true])]; + tensor var_278 = slice_by_index(begin = var_278_begin_0, end = var_278_end_0, end_mask = var_278_end_mask_0, x = window_3)[name = tensor("op_278")]; tensor window_5_interleave_0 = const()[name = tensor("window_5_interleave_0"), val = tensor(false)]; - tensor window_5 = concat(axis = var_26, interleave = window_5_interleave_0, values = (var_232, var_229))[name = tensor("window_5")]; - tensor var_237_begin_0 = const()[name = tensor("op_237_begin_0"), val = tensor([0, 2, 0])]; - tensor var_237_end_0 = const()[name = tensor("op_237_end_0"), val = tensor([1, 1, 256])]; - tensor var_237_end_mask_0 = const()[name = tensor("op_237_end_mask_0"), val = tensor([true, true, true])]; - tensor var_237 = slice_by_index(begin = var_237_begin_0, end = var_237_end_0, end_mask = var_237_end_mask_0, x = x_3)[name = tensor("op_237")]; - tensor var_240_begin_0 = const()[name = tensor("op_240_begin_0"), val = tensor([0, 1, 0])]; - tensor var_240_end_0 = const()[name = tensor("op_240_end_0"), val = tensor([1, 16, 256])]; - tensor var_240_end_mask_0 = const()[name = tensor("op_240_end_mask_0"), val = tensor([true, true, true])]; - tensor var_240 = slice_by_index(begin = var_240_begin_0, end = var_240_end_0, end_mask = var_240_end_mask_0, x = window_5)[name = tensor("op_240")]; + tensor window_5 = concat(axis = var_72, interleave = window_5_interleave_0, values = (var_278, var_275))[name = tensor("window_5")]; + tensor var_283_begin_0 = const()[name = tensor("op_283_begin_0"), val = tensor([0, 2, 0])]; + tensor var_283_end_0 = const()[name = tensor("op_283_end_0"), val = tensor([1, 1, 256])]; + tensor var_283_end_mask_0 = const()[name = tensor("op_283_end_mask_0"), val = tensor([true, true, true])]; + tensor var_283 = slice_by_index(begin = var_283_begin_0, end = var_283_end_0, end_mask = var_283_end_mask_0, x = x_3)[name = tensor("op_283")]; + tensor var_286_begin_0 = const()[name = tensor("op_286_begin_0"), val = tensor([0, 1, 0])]; + tensor var_286_end_0 = const()[name = tensor("op_286_end_0"), val = tensor([1, 16, 256])]; + tensor var_286_end_mask_0 = const()[name = tensor("op_286_end_mask_0"), val = tensor([true, true, true])]; + tensor var_286 = slice_by_index(begin = var_286_begin_0, end = var_286_end_0, end_mask = var_286_end_mask_0, x = window_5)[name = tensor("op_286")]; tensor window_7_interleave_0 = const()[name = tensor("window_7_interleave_0"), val = tensor(false)]; - tensor window_7 = concat(axis = var_26, interleave = window_7_interleave_0, values = (var_240, var_237))[name = tensor("window_7")]; - tensor input_21_interleave_0 = const()[name = tensor("input_21_interleave_0"), val = tensor(false)]; - tensor input_21 = concat(axis = var_23, interleave = input_21_interleave_0, values = (window_3, window_5, window_7))[name = tensor("input_21")]; + tensor window_7 = concat(axis = var_72, interleave = window_7_interleave_0, values = (var_286, var_283))[name = tensor("window_7")]; + tensor input_23_interleave_0 = const()[name = tensor("input_23_interleave_0"), val = tensor(false)]; + tensor input_23 = concat(axis = var_59, interleave = input_23_interleave_0, values = (window_3, window_5, window_7))[name = tensor("input_23")]; tensor x_5_axes_0 = const()[name = tensor("x_5_axes_0"), val = tensor([-1])]; - tensor x_5 = layer_norm(axes = x_5_axes_0, beta = encoder_conv_module_0_sequential_0_bias, epsilon = var_28, gamma = encoder_conv_module_0_sequential_0_weight, x = input_21)[name = tensor("x_5")]; - tensor input_23_perm_0 = const()[name = tensor("input_23_perm_0"), val = tensor([0, 2, 1])]; + tensor x_5 = layer_norm(axes = x_5_axes_0, beta = inner_encoder_conv_module_0_sequential_0_bias, epsilon = var_56, gamma = inner_encoder_conv_module_0_sequential_0_weight, x = input_23)[name = tensor("x_5")]; + tensor input_25_perm_0 = const()[name = tensor("input_25_perm_0"), val = tensor([0, 2, 1])]; tensor inputs_3_pad_type_0 = const()[name = tensor("inputs_3_pad_type_0"), val = tensor("valid")]; tensor inputs_3_strides_0 = const()[name = tensor("inputs_3_strides_0"), val = tensor([1])]; tensor inputs_3_pad_0 = const()[name = tensor("inputs_3_pad_0"), val = tensor([0, 0])]; tensor inputs_3_dilations_0 = const()[name = tensor("inputs_3_dilations_0"), val = tensor([1])]; tensor inputs_3_groups_0 = const()[name = tensor("inputs_3_groups_0"), val = tensor(1)]; - tensor input_23 = transpose(perm = input_23_perm_0, x = x_5)[name = tensor("transpose_54")]; - tensor inputs_3 = conv(bias = encoder_conv_module_0_sequential_2_conv_bias, dilations = inputs_3_dilations_0, groups = inputs_3_groups_0, pad = inputs_3_pad_0, pad_type = inputs_3_pad_type_0, strides = inputs_3_strides_0, weight = encoder_conv_module_0_sequential_2_conv_weight, x = input_23)[name = tensor("inputs_3")]; - tensor var_265_split_sizes_0 = const()[name = tensor("op_265_split_sizes_0"), val = tensor([256, 256])]; - tensor var_265_axis_0 = const()[name = tensor("op_265_axis_0"), val = tensor(1)]; - tensor var_265_0, tensor var_265_1 = split(axis = var_265_axis_0, split_sizes = var_265_split_sizes_0, x = inputs_3)[name = tensor("op_265")]; - tensor var_267 = sigmoid(x = var_265_1)[name = tensor("op_267")]; - tensor inputs_5 = mul(x = var_265_0, y = var_267)[name = tensor("inputs_5")]; + tensor input_25 = transpose(perm = input_25_perm_0, x = x_5)[name = tensor("transpose_54")]; + tensor inputs_3 = conv(bias = inner_encoder_conv_module_0_sequential_2_conv_bias, dilations = inputs_3_dilations_0, groups = inputs_3_groups_0, pad = inputs_3_pad_0, pad_type = inputs_3_pad_type_0, strides = inputs_3_strides_0, weight = inner_encoder_conv_module_0_sequential_2_conv_weight, x = input_25)[name = tensor("inputs_3")]; + tensor var_311_split_sizes_0 = const()[name = tensor("op_311_split_sizes_0"), val = tensor([256, 256])]; + tensor var_311_axis_0 = const()[name = tensor("op_311_axis_0"), val = tensor(1)]; + tensor var_311_0, tensor var_311_1 = split(axis = var_311_axis_0, split_sizes = var_311_split_sizes_0, x = inputs_3)[name = tensor("op_311")]; + tensor var_313 = sigmoid(x = var_311_1)[name = tensor("op_313")]; + tensor inputs_5 = mul(x = var_311_0, y = var_313)[name = tensor("inputs_5")]; tensor outputs_aug_1_pad_type_0 = const()[name = tensor("outputs_aug_1_pad_type_0"), val = tensor("custom")]; tensor outputs_aug_1_pad_0 = const()[name = tensor("outputs_aug_1_pad_0"), val = tensor([15, 15])]; tensor outputs_aug_1_groups_0 = const()[name = tensor("outputs_aug_1_groups_0"), val = tensor(256)]; tensor outputs_aug_1_strides_0 = const()[name = tensor("outputs_aug_1_strides_0"), val = tensor([1])]; tensor outputs_aug_1_dilations_0 = const()[name = tensor("outputs_aug_1_dilations_0"), val = tensor([1])]; - tensor outputs_aug_1 = conv(dilations = outputs_aug_1_dilations_0, groups = outputs_aug_1_groups_0, pad = outputs_aug_1_pad_0, pad_type = outputs_aug_1_pad_type_0, strides = outputs_aug_1_strides_0, weight = encoder_conv_module_0_sequential_4_conv_weight, x = inputs_5)[name = tensor("outputs_aug_1")]; - tensor input_25_begin_0 = const()[name = tensor("input_25_begin_0"), val = tensor([0, 0, 0])]; - tensor input_25_end_0 = const()[name = tensor("input_25_end_0"), val = tensor([3, 256, 16])]; - tensor input_25_end_mask_0 = const()[name = tensor("input_25_end_mask_0"), val = tensor([true, true, false])]; - tensor input_25 = slice_by_index(begin = input_25_begin_0, end = input_25_end_0, end_mask = input_25_end_mask_0, x = outputs_aug_1)[name = tensor("input_25")]; - tensor inputs_7 = batch_norm(beta = encoder_conv_module_0_sequential_5_bias, epsilon = var_28, gamma = encoder_conv_module_0_sequential_5_weight, mean = encoder_conv_module_0_sequential_5_running_mean, variance = encoder_conv_module_0_sequential_5_running_var, x = input_25)[name = tensor("inputs_7")]; - tensor input_27 = silu(x = inputs_7)[name = tensor("input_27")]; - tensor input_29_pad_type_0 = const()[name = tensor("input_29_pad_type_0"), val = tensor("valid")]; - tensor input_29_strides_0 = const()[name = tensor("input_29_strides_0"), val = tensor([1])]; - tensor input_29_pad_0 = const()[name = tensor("input_29_pad_0"), val = tensor([0, 0])]; - tensor input_29_dilations_0 = const()[name = tensor("input_29_dilations_0"), val = tensor([1])]; - tensor input_29_groups_0 = const()[name = tensor("input_29_groups_0"), val = tensor(1)]; - tensor input_29 = conv(bias = encoder_conv_module_0_sequential_7_conv_bias, dilations = input_29_dilations_0, groups = input_29_groups_0, pad = input_29_pad_0, pad_type = input_29_pad_type_0, strides = input_29_strides_0, weight = encoder_conv_module_0_sequential_7_conv_weight, x = input_27)[name = tensor("input_29")]; + tensor outputs_aug_1 = conv(dilations = outputs_aug_1_dilations_0, groups = outputs_aug_1_groups_0, pad = outputs_aug_1_pad_0, pad_type = outputs_aug_1_pad_type_0, strides = outputs_aug_1_strides_0, weight = inner_encoder_conv_module_0_sequential_4_conv_weight, x = inputs_5)[name = tensor("outputs_aug_1")]; + tensor input_27_begin_0 = const()[name = tensor("input_27_begin_0"), val = tensor([0, 0, 0])]; + tensor input_27_end_0 = const()[name = tensor("input_27_end_0"), val = tensor([3, 256, 16])]; + tensor input_27_end_mask_0 = const()[name = tensor("input_27_end_mask_0"), val = tensor([true, true, false])]; + tensor input_27 = slice_by_index(begin = input_27_begin_0, end = input_27_end_0, end_mask = input_27_end_mask_0, x = outputs_aug_1)[name = tensor("input_27")]; + tensor inputs_7 = batch_norm(beta = inner_encoder_conv_module_0_sequential_5_bias, epsilon = var_56, gamma = inner_encoder_conv_module_0_sequential_5_weight, mean = inner_encoder_conv_module_0_sequential_5_running_mean, variance = inner_encoder_conv_module_0_sequential_5_running_var, x = input_27)[name = tensor("inputs_7")]; + tensor input_29 = silu(x = inputs_7)[name = tensor("input_29")]; + tensor input_31_pad_type_0 = const()[name = tensor("input_31_pad_type_0"), val = tensor("valid")]; + tensor input_31_strides_0 = const()[name = tensor("input_31_strides_0"), val = tensor([1])]; + tensor input_31_pad_0 = const()[name = tensor("input_31_pad_0"), val = tensor([0, 0])]; + tensor input_31_dilations_0 = const()[name = tensor("input_31_dilations_0"), val = tensor([1])]; + tensor input_31_groups_0 = const()[name = tensor("input_31_groups_0"), val = tensor(1)]; + tensor input_31 = conv(bias = inner_encoder_conv_module_0_sequential_7_conv_bias, dilations = input_31_dilations_0, groups = input_31_groups_0, pad = input_31_pad_0, pad_type = input_31_pad_type_0, strides = input_31_strides_0, weight = inner_encoder_conv_module_0_sequential_7_conv_weight, x = input_29)[name = tensor("input_31")]; tensor conv_out_1_perm_0 = const()[name = tensor("conv_out_1_perm_0"), val = tensor([0, 2, 1])]; - tensor var_298_begin_0 = const()[name = tensor("op_298_begin_0"), val = tensor([0, -1, 0])]; - tensor var_298_end_0 = const()[name = tensor("op_298_end_0"), val = tensor([3, 16, 256])]; - tensor var_298_end_mask_0 = const()[name = tensor("op_298_end_mask_0"), val = tensor([true, true, true])]; - tensor conv_out_1 = transpose(perm = conv_out_1_perm_0, x = input_29)[name = tensor("transpose_53")]; - tensor var_298 = slice_by_index(begin = var_298_begin_0, end = var_298_end_0, end_mask = var_298_end_mask_0, x = conv_out_1)[name = tensor("op_298")]; - tensor var_300_perm_0 = const()[name = tensor("op_300_perm_0"), val = tensor([1, 0, 2])]; - tensor var_300 = transpose(perm = var_300_perm_0, x = var_298)[name = tensor("transpose_52")]; - tensor input_31 = add(x = x_3, y = var_300)[name = tensor("input_31")]; - tensor input_33_axes_0 = const()[name = tensor("input_33_axes_0"), val = tensor([-1])]; - tensor input_33 = layer_norm(axes = input_33_axes_0, beta = encoder_ffn2_0_module_sequential_0_bias, epsilon = var_28, gamma = encoder_ffn2_0_module_sequential_0_weight, x = input_31)[name = tensor("input_33")]; - tensor inputs_9 = linear(bias = encoder_ffn2_0_module_sequential_1_linear_bias, weight = encoder_ffn2_0_module_sequential_1_linear_weight, x = input_33)[name = tensor("linear_8")]; - tensor input_35 = silu(x = inputs_9)[name = tensor("input_35")]; - tensor input_39 = linear(bias = encoder_ffn2_0_module_sequential_4_linear_bias, weight = encoder_ffn2_0_module_sequential_4_linear_weight, x = input_35)[name = tensor("linear_9")]; - tensor var_323 = const()[name = tensor("op_323"), val = tensor(0x1p-1)]; - tensor var_324 = mul(x = input_39, y = var_323)[name = tensor("op_324")]; - tensor input_41 = add(x = var_324, y = input_31)[name = tensor("input_41")]; - tensor input_43_axes_0 = const()[name = tensor("input_43_axes_0"), val = tensor([-1])]; - tensor input_43 = layer_norm(axes = input_43_axes_0, beta = encoder_layer_norm_0_bias, epsilon = var_28, gamma = encoder_layer_norm_0_weight, x = input_41)[name = tensor("input_43")]; + tensor var_344_begin_0 = const()[name = tensor("op_344_begin_0"), val = tensor([0, -1, 0])]; + tensor var_344_end_0 = const()[name = tensor("op_344_end_0"), val = tensor([3, 16, 256])]; + tensor var_344_end_mask_0 = const()[name = tensor("op_344_end_mask_0"), val = tensor([true, true, true])]; + tensor conv_out_1 = transpose(perm = conv_out_1_perm_0, x = input_31)[name = tensor("transpose_53")]; + tensor var_344 = slice_by_index(begin = var_344_begin_0, end = var_344_end_0, end_mask = var_344_end_mask_0, x = conv_out_1)[name = tensor("op_344")]; + tensor var_346_perm_0 = const()[name = tensor("op_346_perm_0"), val = tensor([1, 0, 2])]; + tensor var_346 = transpose(perm = var_346_perm_0, x = var_344)[name = tensor("transpose_52")]; + tensor input_33 = add(x = x_3, y = var_346)[name = tensor("input_33")]; + tensor input_35_axes_0 = const()[name = tensor("input_35_axes_0"), val = tensor([-1])]; + tensor input_35 = layer_norm(axes = input_35_axes_0, beta = inner_encoder_ffn2_0_module_sequential_0_bias, epsilon = var_56, gamma = inner_encoder_ffn2_0_module_sequential_0_weight, x = input_33)[name = tensor("input_35")]; + tensor inputs_9 = linear(bias = inner_encoder_ffn2_0_module_sequential_1_linear_bias, weight = inner_encoder_ffn2_0_module_sequential_1_linear_weight, x = input_35)[name = tensor("linear_8")]; + tensor input_37 = silu(x = inputs_9)[name = tensor("input_37")]; + tensor input_41 = linear(bias = inner_encoder_ffn2_0_module_sequential_4_linear_bias, weight = inner_encoder_ffn2_0_module_sequential_4_linear_weight, x = input_37)[name = tensor("linear_9")]; + tensor var_369 = const()[name = tensor("op_369"), val = tensor(0x1p-1)]; + tensor var_370 = mul(x = input_41, y = var_369)[name = tensor("op_370")]; + tensor input_43 = add(x = var_370, y = input_33)[name = tensor("input_43")]; tensor input_45_axes_0 = const()[name = tensor("input_45_axes_0"), val = tensor([-1])]; - tensor input_45 = layer_norm(axes = input_45_axes_0, beta = encoder_ffn1_1_module_sequential_0_bias, epsilon = var_28, gamma = encoder_ffn1_1_module_sequential_0_weight, x = input_43)[name = tensor("input_45")]; - tensor inputs_11 = linear(bias = encoder_ffn1_1_module_sequential_1_linear_bias, weight = encoder_ffn1_1_module_sequential_1_linear_weight, x = input_45)[name = tensor("linear_10")]; - tensor input_47 = silu(x = inputs_11)[name = tensor("input_47")]; - tensor input_51 = linear(bias = encoder_ffn1_1_module_sequential_4_linear_bias, weight = encoder_ffn1_1_module_sequential_4_linear_weight, x = input_47)[name = tensor("linear_11")]; - tensor var_353 = const()[name = tensor("op_353"), val = tensor(0x1p-1)]; - tensor var_354 = mul(x = input_51, y = var_353)[name = tensor("op_354")]; - tensor input_53 = add(x = var_354, y = input_43)[name = tensor("input_53")]; + tensor input_45 = layer_norm(axes = input_45_axes_0, beta = inner_encoder_layer_norm_0_bias, epsilon = var_56, gamma = inner_encoder_layer_norm_0_weight, x = input_43)[name = tensor("input_45")]; + tensor input_47_axes_0 = const()[name = tensor("input_47_axes_0"), val = tensor([-1])]; + tensor input_47 = layer_norm(axes = input_47_axes_0, beta = inner_encoder_ffn1_1_module_sequential_0_bias, epsilon = var_56, gamma = inner_encoder_ffn1_1_module_sequential_0_weight, x = input_45)[name = tensor("input_47")]; + tensor inputs_11 = linear(bias = inner_encoder_ffn1_1_module_sequential_1_linear_bias, weight = inner_encoder_ffn1_1_module_sequential_1_linear_weight, x = input_47)[name = tensor("linear_10")]; + tensor input_49 = silu(x = inputs_11)[name = tensor("input_49")]; + tensor input_53 = linear(bias = inner_encoder_ffn1_1_module_sequential_4_linear_bias, weight = inner_encoder_ffn1_1_module_sequential_4_linear_weight, x = input_49)[name = tensor("linear_11")]; + tensor var_399 = const()[name = tensor("op_399"), val = tensor(0x1p-1)]; + tensor var_400 = mul(x = input_53, y = var_399)[name = tensor("op_400")]; + tensor input_55 = add(x = var_400, y = input_45)[name = tensor("input_55")]; tensor x_7_axes_0 = const()[name = tensor("x_7_axes_0"), val = tensor([-1])]; - tensor x_7 = layer_norm(axes = x_7_axes_0, beta = encoder_ret_lns_1_bias, epsilon = var_28, gamma = encoder_ret_lns_1_weight, x = input_53)[name = tensor("x_7")]; + tensor x_7 = layer_norm(axes = x_7_axes_0, beta = inner_encoder_ret_lns_1_bias, epsilon = var_56, gamma = inner_encoder_ret_lns_1_weight, x = input_55)[name = tensor("x_7")]; tensor prev_kv_3_begin_0 = const()[name = tensor("prev_kv_3_begin_0"), val = tensor([1, 0, 0, 0, 0])]; tensor prev_kv_3_end_0 = const()[name = tensor("prev_kv_3_end_0"), val = tensor([2, 1, 4, 64, 64])]; tensor prev_kv_3_end_mask_0 = const()[name = tensor("prev_kv_3_end_mask_0"), val = tensor([false, true, true, true, true])]; @@ -406,163 +424,163 @@ program(1.0) tensor prev_scale_3_end_mask_0 = const()[name = tensor("prev_scale_3_end_mask_0"), val = tensor([false, true])]; tensor prev_scale_3_squeeze_mask_0 = const()[name = tensor("prev_scale_3_squeeze_mask_0"), val = tensor([true, false])]; tensor prev_scale_3 = slice_by_index(begin = prev_scale_3_begin_0, end = prev_scale_3_end_0, end_mask = prev_scale_3_end_mask_0, squeeze_mask = prev_scale_3_squeeze_mask_0, x = enc_scale)[name = tensor("prev_scale_3")]; - tensor var_368 = linear(bias = encoder_q_proj_1_bias, weight = encoder_q_proj_1_weight, x = x_7)[name = tensor("linear_12")]; - tensor var_369 = const()[name = tensor("op_369"), val = tensor([1, 3, 4, 64])]; - tensor var_370 = reshape(shape = var_369, x = var_368)[name = tensor("op_370")]; + tensor var_414 = linear(bias = inner_encoder_q_proj_1_bias, weight = inner_encoder_q_proj_1_weight, x = x_7)[name = tensor("linear_12")]; + tensor var_415 = const()[name = tensor("op_415"), val = tensor([1, 3, 4, 64])]; + tensor var_416 = reshape(shape = var_415, x = var_414)[name = tensor("op_416")]; tensor q_3_perm_0 = const()[name = tensor("q_3_perm_0"), val = tensor([0, 2, 1, 3])]; - tensor var_374 = linear(bias = encoder_k_proj_1_bias, weight = encoder_k_proj_1_weight, x = x_7)[name = tensor("linear_13")]; - tensor var_375 = const()[name = tensor("op_375"), val = tensor(0x1p-3)]; - tensor var_376 = mul(x = var_374, y = var_375)[name = tensor("op_376")]; - tensor var_377 = const()[name = tensor("op_377"), val = tensor([1, 3, 4, 64])]; - tensor var_378 = reshape(shape = var_377, x = var_376)[name = tensor("op_378")]; + tensor var_420 = linear(bias = inner_encoder_k_proj_1_bias, weight = inner_encoder_k_proj_1_weight, x = x_7)[name = tensor("linear_13")]; + tensor var_421 = const()[name = tensor("op_421"), val = tensor(0x1p-3)]; + tensor var_422 = mul(x = var_420, y = var_421)[name = tensor("op_422")]; + tensor var_423 = const()[name = tensor("op_423"), val = tensor([1, 3, 4, 64])]; + tensor var_424 = reshape(shape = var_423, x = var_422)[name = tensor("op_424")]; tensor k_3_perm_0 = const()[name = tensor("k_3_perm_0"), val = tensor([0, 2, 1, 3])]; - tensor var_382 = linear(bias = encoder_v_proj_1_bias, weight = encoder_v_proj_1_weight, x = x_7)[name = tensor("linear_14")]; - tensor var_383 = const()[name = tensor("op_383"), val = tensor([1, 3, 4, 64])]; - tensor var_384 = reshape(shape = var_383, x = var_382)[name = tensor("op_384")]; + tensor var_428 = linear(bias = inner_encoder_v_proj_1_bias, weight = inner_encoder_v_proj_1_weight, x = x_7)[name = tensor("linear_14")]; + tensor var_429 = const()[name = tensor("op_429"), val = tensor([1, 3, 4, 64])]; + tensor var_430 = reshape(shape = var_429, x = var_428)[name = tensor("op_430")]; tensor v_3_perm_0 = const()[name = tensor("v_3_perm_0"), val = tensor([0, 2, 1, 3])]; - tensor input_57 = linear(bias = encoder_g_proj_1_bias, weight = encoder_g_proj_1_weight, x = x_7)[name = tensor("linear_15")]; + tensor input_59 = linear(bias = inner_encoder_g_proj_1_bias, weight = inner_encoder_g_proj_1_weight, x = x_7)[name = tensor("linear_15")]; tensor sqrt_s0_3 = sqrt(x = prev_scale_3)[name = tensor("sqrt_s0_3")]; - tensor s_t_3 = add(x = prev_scale_3, y = encoder__t_index)[name = tensor("s_t_3")]; + tensor s_t_3 = add(x = prev_scale_3, y = inner_encoder__t_index)[name = tensor("s_t_3")]; tensor sqrt_s_t_3 = sqrt(x = s_t_3)[name = tensor("sqrt_s_t_3")]; tensor qk_3_transpose_x_1 = const()[name = tensor("qk_3_transpose_x_1"), val = tensor(false)]; tensor qk_3_transpose_y_1 = const()[name = tensor("qk_3_transpose_y_1"), val = tensor(true)]; - tensor k_3 = transpose(perm = k_3_perm_0, x = var_378)[name = tensor("transpose_50")]; - tensor q_3 = transpose(perm = q_3_perm_0, x = var_370)[name = tensor("transpose_51")]; + tensor k_3 = transpose(perm = k_3_perm_0, x = var_424)[name = tensor("transpose_50")]; + tensor q_3 = transpose(perm = q_3_perm_0, x = var_416)[name = tensor("transpose_51")]; tensor qk_3 = matmul(transpose_x = qk_3_transpose_x_1, transpose_y = qk_3_transpose_y_1, x = q_3, y = k_3)[name = tensor("qk_3")]; - tensor var_394 = const()[name = tensor("op_394"), val = tensor([3, 1])]; - tensor var_395 = reshape(shape = var_394, x = sqrt_s_t_3)[name = tensor("op_395")]; - tensor M_3 = real_div(x = encoder__causal_mask, y = var_395)[name = tensor("M_3")]; - tensor var_397 = mul(x = qk_3, y = M_3)[name = tensor("op_397")]; + tensor var_440 = const()[name = tensor("op_440"), val = tensor([3, 1])]; + tensor var_441 = reshape(shape = var_440, x = sqrt_s_t_3)[name = tensor("op_441")]; + tensor M_3 = real_div(x = inner_encoder__causal_mask, y = var_441)[name = tensor("M_3")]; + tensor var_443 = mul(x = qk_3, y = M_3)[name = tensor("op_443")]; tensor inner_3_transpose_x_0 = const()[name = tensor("inner_3_transpose_x_0"), val = tensor(false)]; tensor inner_3_transpose_y_0 = const()[name = tensor("inner_3_transpose_y_0"), val = tensor(false)]; - tensor v_3 = transpose(perm = v_3_perm_0, x = var_384)[name = tensor("transpose_49")]; - tensor inner_3 = matmul(transpose_x = inner_3_transpose_x_0, transpose_y = inner_3_transpose_y_0, x = var_397, y = v_3)[name = tensor("inner_3")]; - tensor var_399_transpose_x_0 = const()[name = tensor("op_399_transpose_x_0"), val = tensor(false)]; - tensor var_399_transpose_y_0 = const()[name = tensor("op_399_transpose_y_0"), val = tensor(false)]; - tensor var_399 = matmul(transpose_x = var_399_transpose_x_0, transpose_y = var_399_transpose_y_0, x = q_3, y = prev_kv_3)[name = tensor("op_399")]; - tensor var_400 = real_div(x = sqrt_s0_3, y = sqrt_s_t_3)[name = tensor("op_400")]; - tensor var_401 = const()[name = tensor("op_401"), val = tensor([1, 1, 3, 1])]; - tensor var_402 = reshape(shape = var_401, x = var_400)[name = tensor("op_402")]; - tensor cross_3 = mul(x = var_399, y = var_402)[name = tensor("cross_3")]; + tensor v_3 = transpose(perm = v_3_perm_0, x = var_430)[name = tensor("transpose_49")]; + tensor inner_3 = matmul(transpose_x = inner_3_transpose_x_0, transpose_y = inner_3_transpose_y_0, x = var_443, y = v_3)[name = tensor("inner_3")]; + tensor var_445_transpose_x_0 = const()[name = tensor("op_445_transpose_x_0"), val = tensor(false)]; + tensor var_445_transpose_y_0 = const()[name = tensor("op_445_transpose_y_0"), val = tensor(false)]; + tensor var_445 = matmul(transpose_x = var_445_transpose_x_0, transpose_y = var_445_transpose_y_0, x = q_3, y = prev_kv_3)[name = tensor("op_445")]; + tensor var_446 = real_div(x = sqrt_s0_3, y = sqrt_s_t_3)[name = tensor("op_446")]; + tensor var_447 = const()[name = tensor("op_447"), val = tensor([1, 1, 3, 1])]; + tensor var_448 = reshape(shape = var_447, x = var_446)[name = tensor("op_448")]; + tensor cross_3 = mul(x = var_445, y = var_448)[name = tensor("cross_3")]; tensor out_7 = add(x = inner_3, y = cross_3)[name = tensor("out_7")]; - tensor var_405 = mul(x = prev_kv_3, y = sqrt_s0_3)[name = tensor("op_405")]; - tensor var_407_transpose_x_1 = const()[name = tensor("op_407_transpose_x_1"), val = tensor(true)]; - tensor var_407_transpose_y_1 = const()[name = tensor("op_407_transpose_y_1"), val = tensor(false)]; - tensor var_407 = matmul(transpose_x = var_407_transpose_x_1, transpose_y = var_407_transpose_y_1, x = k_3, y = v_3)[name = tensor("op_407")]; - tensor new_kv_unnorm_3 = add(x = var_405, y = var_407)[name = tensor("new_kv_unnorm_3")]; - tensor var_409 = const()[name = tensor("op_409"), val = tensor(0x1.8p+1)]; - tensor new_scale_3 = add(x = prev_scale_3, y = var_409)[name = tensor("new_scale_3")]; - tensor var_411 = sqrt(x = new_scale_3)[name = tensor("op_411")]; - tensor var_412 = real_div(x = new_kv_unnorm_3, y = var_411)[name = tensor("op_412")]; - tensor var_413_perm_0 = const()[name = tensor("op_413_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_451 = mul(x = prev_kv_3, y = sqrt_s0_3)[name = tensor("op_451")]; + tensor var_453_transpose_x_1 = const()[name = tensor("op_453_transpose_x_1"), val = tensor(true)]; + tensor var_453_transpose_y_1 = const()[name = tensor("op_453_transpose_y_1"), val = tensor(false)]; + tensor var_453 = matmul(transpose_x = var_453_transpose_x_1, transpose_y = var_453_transpose_y_1, x = k_3, y = v_3)[name = tensor("op_453")]; + tensor new_kv_unnorm_3 = add(x = var_451, y = var_453)[name = tensor("new_kv_unnorm_3")]; + tensor var_455 = const()[name = tensor("op_455"), val = tensor(0x1.8p+1)]; + tensor new_scale_3 = add(x = prev_scale_3, y = var_455)[name = tensor("new_scale_3")]; + tensor var_457 = sqrt(x = new_scale_3)[name = tensor("op_457")]; + tensor var_458 = real_div(x = new_kv_unnorm_3, y = var_457)[name = tensor("op_458")]; + tensor var_459_perm_0 = const()[name = tensor("op_459_perm_0"), val = tensor([0, 2, 1, 3])]; tensor out_9_axes_0 = const()[name = tensor("out_9_axes_0"), val = tensor([-1])]; - tensor var_413 = transpose(perm = var_413_perm_0, x = out_7)[name = tensor("transpose_48")]; - tensor out_9 = layer_norm(axes = out_9_axes_0, epsilon = var_17, x = var_413)[name = tensor("out_9")]; - tensor var_417 = const()[name = tensor("op_417"), val = tensor([1, 3, 256])]; - tensor out_11 = reshape(shape = var_417, x = out_9)[name = tensor("out_11")]; - tensor var_419 = silu(x = input_57)[name = tensor("op_419")]; - tensor input_59 = mul(x = var_419, y = out_11)[name = tensor("input_59")]; - tensor ret_out_3 = linear(bias = encoder_out_proj_1_bias, weight = encoder_out_proj_1_weight, x = input_59)[name = tensor("linear_16")]; - tensor x_9 = add(x = input_53, y = ret_out_3)[name = tensor("x_9")]; + tensor var_459 = transpose(perm = var_459_perm_0, x = out_7)[name = tensor("transpose_48")]; + tensor out_9 = layer_norm(axes = out_9_axes_0, epsilon = var_64, x = var_459)[name = tensor("out_9")]; + tensor var_463 = const()[name = tensor("op_463"), val = tensor([1, 3, 256])]; + tensor out_11 = reshape(shape = var_463, x = out_9)[name = tensor("out_11")]; + tensor var_465 = silu(x = input_59)[name = tensor("op_465")]; + tensor input_61 = mul(x = var_465, y = out_11)[name = tensor("input_61")]; + tensor ret_out_3 = linear(bias = inner_encoder_out_proj_1_bias, weight = inner_encoder_out_proj_1_weight, x = input_61)[name = tensor("linear_16")]; + tensor x_9 = add(x = input_55, y = ret_out_3)[name = tensor("x_9")]; tensor window_9_begin_0 = const()[name = tensor("window_9_begin_0"), val = tensor([1, 0, 0, 0])]; tensor window_9_end_0 = const()[name = tensor("window_9_end_0"), val = tensor([2, 1, 16, 256])]; tensor window_9_end_mask_0 = const()[name = tensor("window_9_end_mask_0"), val = tensor([false, true, true, true])]; tensor window_9_squeeze_mask_0 = const()[name = tensor("window_9_squeeze_mask_0"), val = tensor([true, false, false, false])]; tensor window_9 = slice_by_index(begin = window_9_begin_0, end = window_9_end_0, end_mask = window_9_end_mask_0, squeeze_mask = window_9_squeeze_mask_0, x = enc_conv_cache)[name = tensor("window_9")]; - tensor var_427_begin_0 = const()[name = tensor("op_427_begin_0"), val = tensor([0, 0, 0])]; - tensor var_427_end_0 = const()[name = tensor("op_427_end_0"), val = tensor([1, 1, 256])]; - tensor var_427_end_mask_0 = const()[name = tensor("op_427_end_mask_0"), val = tensor([true, false, true])]; - tensor var_427 = slice_by_index(begin = var_427_begin_0, end = var_427_end_0, end_mask = var_427_end_mask_0, x = x_9)[name = tensor("op_427")]; - tensor var_430_begin_0 = const()[name = tensor("op_430_begin_0"), val = tensor([0, 1, 0])]; - tensor var_430_end_0 = const()[name = tensor("op_430_end_0"), val = tensor([1, 16, 256])]; - tensor var_430_end_mask_0 = const()[name = tensor("op_430_end_mask_0"), val = tensor([true, true, true])]; - tensor var_430 = slice_by_index(begin = var_430_begin_0, end = var_430_end_0, end_mask = var_430_end_mask_0, x = window_9)[name = tensor("op_430")]; + tensor var_473_begin_0 = const()[name = tensor("op_473_begin_0"), val = tensor([0, 0, 0])]; + tensor var_473_end_0 = const()[name = tensor("op_473_end_0"), val = tensor([1, 1, 256])]; + tensor var_473_end_mask_0 = const()[name = tensor("op_473_end_mask_0"), val = tensor([true, false, true])]; + tensor var_473 = slice_by_index(begin = var_473_begin_0, end = var_473_end_0, end_mask = var_473_end_mask_0, x = x_9)[name = tensor("op_473")]; + tensor var_476_begin_0 = const()[name = tensor("op_476_begin_0"), val = tensor([0, 1, 0])]; + tensor var_476_end_0 = const()[name = tensor("op_476_end_0"), val = tensor([1, 16, 256])]; + tensor var_476_end_mask_0 = const()[name = tensor("op_476_end_mask_0"), val = tensor([true, true, true])]; + tensor var_476 = slice_by_index(begin = var_476_begin_0, end = var_476_end_0, end_mask = var_476_end_mask_0, x = window_9)[name = tensor("op_476")]; tensor window_11_interleave_0 = const()[name = tensor("window_11_interleave_0"), val = tensor(false)]; - tensor window_11 = concat(axis = var_26, interleave = window_11_interleave_0, values = (var_430, var_427))[name = tensor("window_11")]; - tensor var_435_begin_0 = const()[name = tensor("op_435_begin_0"), val = tensor([0, 1, 0])]; - tensor var_435_end_0 = const()[name = tensor("op_435_end_0"), val = tensor([1, 2, 256])]; - tensor var_435_end_mask_0 = const()[name = tensor("op_435_end_mask_0"), val = tensor([true, false, true])]; - tensor var_435 = slice_by_index(begin = var_435_begin_0, end = var_435_end_0, end_mask = var_435_end_mask_0, x = x_9)[name = tensor("op_435")]; - tensor var_438_begin_0 = const()[name = tensor("op_438_begin_0"), val = tensor([0, 1, 0])]; - tensor var_438_end_0 = const()[name = tensor("op_438_end_0"), val = tensor([1, 16, 256])]; - tensor var_438_end_mask_0 = const()[name = tensor("op_438_end_mask_0"), val = tensor([true, true, true])]; - tensor var_438 = slice_by_index(begin = var_438_begin_0, end = var_438_end_0, end_mask = var_438_end_mask_0, x = window_11)[name = tensor("op_438")]; + tensor window_11 = concat(axis = var_72, interleave = window_11_interleave_0, values = (var_476, var_473))[name = tensor("window_11")]; + tensor var_481_begin_0 = const()[name = tensor("op_481_begin_0"), val = tensor([0, 1, 0])]; + tensor var_481_end_0 = const()[name = tensor("op_481_end_0"), val = tensor([1, 2, 256])]; + tensor var_481_end_mask_0 = const()[name = tensor("op_481_end_mask_0"), val = tensor([true, false, true])]; + tensor var_481 = slice_by_index(begin = var_481_begin_0, end = var_481_end_0, end_mask = var_481_end_mask_0, x = x_9)[name = tensor("op_481")]; + tensor var_484_begin_0 = const()[name = tensor("op_484_begin_0"), val = tensor([0, 1, 0])]; + tensor var_484_end_0 = const()[name = tensor("op_484_end_0"), val = tensor([1, 16, 256])]; + tensor var_484_end_mask_0 = const()[name = tensor("op_484_end_mask_0"), val = tensor([true, true, true])]; + tensor var_484 = slice_by_index(begin = var_484_begin_0, end = var_484_end_0, end_mask = var_484_end_mask_0, x = window_11)[name = tensor("op_484")]; tensor window_13_interleave_0 = const()[name = tensor("window_13_interleave_0"), val = tensor(false)]; - tensor window_13 = concat(axis = var_26, interleave = window_13_interleave_0, values = (var_438, var_435))[name = tensor("window_13")]; - tensor var_443_begin_0 = const()[name = tensor("op_443_begin_0"), val = tensor([0, 2, 0])]; - tensor var_443_end_0 = const()[name = tensor("op_443_end_0"), val = tensor([1, 1, 256])]; - tensor var_443_end_mask_0 = const()[name = tensor("op_443_end_mask_0"), val = tensor([true, true, true])]; - tensor var_443 = slice_by_index(begin = var_443_begin_0, end = var_443_end_0, end_mask = var_443_end_mask_0, x = x_9)[name = tensor("op_443")]; - tensor var_446_begin_0 = const()[name = tensor("op_446_begin_0"), val = tensor([0, 1, 0])]; - tensor var_446_end_0 = const()[name = tensor("op_446_end_0"), val = tensor([1, 16, 256])]; - tensor var_446_end_mask_0 = const()[name = tensor("op_446_end_mask_0"), val = tensor([true, true, true])]; - tensor var_446 = slice_by_index(begin = var_446_begin_0, end = var_446_end_0, end_mask = var_446_end_mask_0, x = window_13)[name = tensor("op_446")]; + tensor window_13 = concat(axis = var_72, interleave = window_13_interleave_0, values = (var_484, var_481))[name = tensor("window_13")]; + tensor var_489_begin_0 = const()[name = tensor("op_489_begin_0"), val = tensor([0, 2, 0])]; + tensor var_489_end_0 = const()[name = tensor("op_489_end_0"), val = tensor([1, 1, 256])]; + tensor var_489_end_mask_0 = const()[name = tensor("op_489_end_mask_0"), val = tensor([true, true, true])]; + tensor var_489 = slice_by_index(begin = var_489_begin_0, end = var_489_end_0, end_mask = var_489_end_mask_0, x = x_9)[name = tensor("op_489")]; + tensor var_492_begin_0 = const()[name = tensor("op_492_begin_0"), val = tensor([0, 1, 0])]; + tensor var_492_end_0 = const()[name = tensor("op_492_end_0"), val = tensor([1, 16, 256])]; + tensor var_492_end_mask_0 = const()[name = tensor("op_492_end_mask_0"), val = tensor([true, true, true])]; + tensor var_492 = slice_by_index(begin = var_492_begin_0, end = var_492_end_0, end_mask = var_492_end_mask_0, x = window_13)[name = tensor("op_492")]; tensor window_15_interleave_0 = const()[name = tensor("window_15_interleave_0"), val = tensor(false)]; - tensor window_15 = concat(axis = var_26, interleave = window_15_interleave_0, values = (var_446, var_443))[name = tensor("window_15")]; - tensor input_61_interleave_0 = const()[name = tensor("input_61_interleave_0"), val = tensor(false)]; - tensor input_61 = concat(axis = var_23, interleave = input_61_interleave_0, values = (window_11, window_13, window_15))[name = tensor("input_61")]; + tensor window_15 = concat(axis = var_72, interleave = window_15_interleave_0, values = (var_492, var_489))[name = tensor("window_15")]; + tensor input_63_interleave_0 = const()[name = tensor("input_63_interleave_0"), val = tensor(false)]; + tensor input_63 = concat(axis = var_59, interleave = input_63_interleave_0, values = (window_11, window_13, window_15))[name = tensor("input_63")]; tensor x_11_axes_0 = const()[name = tensor("x_11_axes_0"), val = tensor([-1])]; - tensor x_11 = layer_norm(axes = x_11_axes_0, beta = encoder_conv_module_1_sequential_0_bias, epsilon = var_28, gamma = encoder_conv_module_1_sequential_0_weight, x = input_61)[name = tensor("x_11")]; - tensor input_63_perm_0 = const()[name = tensor("input_63_perm_0"), val = tensor([0, 2, 1])]; + tensor x_11 = layer_norm(axes = x_11_axes_0, beta = inner_encoder_conv_module_1_sequential_0_bias, epsilon = var_56, gamma = inner_encoder_conv_module_1_sequential_0_weight, x = input_63)[name = tensor("x_11")]; + tensor input_65_perm_0 = const()[name = tensor("input_65_perm_0"), val = tensor([0, 2, 1])]; tensor inputs_13_pad_type_0 = const()[name = tensor("inputs_13_pad_type_0"), val = tensor("valid")]; tensor inputs_13_strides_0 = const()[name = tensor("inputs_13_strides_0"), val = tensor([1])]; tensor inputs_13_pad_0 = const()[name = tensor("inputs_13_pad_0"), val = tensor([0, 0])]; tensor inputs_13_dilations_0 = const()[name = tensor("inputs_13_dilations_0"), val = tensor([1])]; tensor inputs_13_groups_0 = const()[name = tensor("inputs_13_groups_0"), val = tensor(1)]; - tensor input_63 = transpose(perm = input_63_perm_0, x = x_11)[name = tensor("transpose_47")]; - tensor inputs_13 = conv(bias = encoder_conv_module_1_sequential_2_conv_bias, dilations = inputs_13_dilations_0, groups = inputs_13_groups_0, pad = inputs_13_pad_0, pad_type = inputs_13_pad_type_0, strides = inputs_13_strides_0, weight = encoder_conv_module_1_sequential_2_conv_weight, x = input_63)[name = tensor("inputs_13")]; - tensor var_471_split_sizes_0 = const()[name = tensor("op_471_split_sizes_0"), val = tensor([256, 256])]; - tensor var_471_axis_0 = const()[name = tensor("op_471_axis_0"), val = tensor(1)]; - tensor var_471_0, tensor var_471_1 = split(axis = var_471_axis_0, split_sizes = var_471_split_sizes_0, x = inputs_13)[name = tensor("op_471")]; - tensor var_473 = sigmoid(x = var_471_1)[name = tensor("op_473")]; - tensor inputs_15 = mul(x = var_471_0, y = var_473)[name = tensor("inputs_15")]; + tensor input_65 = transpose(perm = input_65_perm_0, x = x_11)[name = tensor("transpose_47")]; + tensor inputs_13 = conv(bias = inner_encoder_conv_module_1_sequential_2_conv_bias, dilations = inputs_13_dilations_0, groups = inputs_13_groups_0, pad = inputs_13_pad_0, pad_type = inputs_13_pad_type_0, strides = inputs_13_strides_0, weight = inner_encoder_conv_module_1_sequential_2_conv_weight, x = input_65)[name = tensor("inputs_13")]; + tensor var_517_split_sizes_0 = const()[name = tensor("op_517_split_sizes_0"), val = tensor([256, 256])]; + tensor var_517_axis_0 = const()[name = tensor("op_517_axis_0"), val = tensor(1)]; + tensor var_517_0, tensor var_517_1 = split(axis = var_517_axis_0, split_sizes = var_517_split_sizes_0, x = inputs_13)[name = tensor("op_517")]; + tensor var_519 = sigmoid(x = var_517_1)[name = tensor("op_519")]; + tensor inputs_15 = mul(x = var_517_0, y = var_519)[name = tensor("inputs_15")]; tensor outputs_aug_3_pad_type_0 = const()[name = tensor("outputs_aug_3_pad_type_0"), val = tensor("custom")]; tensor outputs_aug_3_pad_0 = const()[name = tensor("outputs_aug_3_pad_0"), val = tensor([15, 15])]; tensor outputs_aug_3_groups_0 = const()[name = tensor("outputs_aug_3_groups_0"), val = tensor(256)]; tensor outputs_aug_3_strides_0 = const()[name = tensor("outputs_aug_3_strides_0"), val = tensor([1])]; tensor outputs_aug_3_dilations_0 = const()[name = tensor("outputs_aug_3_dilations_0"), val = tensor([1])]; - tensor outputs_aug_3 = conv(dilations = outputs_aug_3_dilations_0, groups = outputs_aug_3_groups_0, pad = outputs_aug_3_pad_0, pad_type = outputs_aug_3_pad_type_0, strides = outputs_aug_3_strides_0, weight = encoder_conv_module_1_sequential_4_conv_weight, x = inputs_15)[name = tensor("outputs_aug_3")]; - tensor input_65_begin_0 = const()[name = tensor("input_65_begin_0"), val = tensor([0, 0, 0])]; - tensor input_65_end_0 = const()[name = tensor("input_65_end_0"), val = tensor([3, 256, 16])]; - tensor input_65_end_mask_0 = const()[name = tensor("input_65_end_mask_0"), val = tensor([true, true, false])]; - tensor input_65 = slice_by_index(begin = input_65_begin_0, end = input_65_end_0, end_mask = input_65_end_mask_0, x = outputs_aug_3)[name = tensor("input_65")]; - tensor inputs_17 = batch_norm(beta = encoder_conv_module_1_sequential_5_bias, epsilon = var_28, gamma = encoder_conv_module_1_sequential_5_weight, mean = encoder_conv_module_1_sequential_5_running_mean, variance = encoder_conv_module_1_sequential_5_running_var, x = input_65)[name = tensor("inputs_17")]; - tensor input_67 = silu(x = inputs_17)[name = tensor("input_67")]; - tensor input_69_pad_type_0 = const()[name = tensor("input_69_pad_type_0"), val = tensor("valid")]; - tensor input_69_strides_0 = const()[name = tensor("input_69_strides_0"), val = tensor([1])]; - tensor input_69_pad_0 = const()[name = tensor("input_69_pad_0"), val = tensor([0, 0])]; - tensor input_69_dilations_0 = const()[name = tensor("input_69_dilations_0"), val = tensor([1])]; - tensor input_69_groups_0 = const()[name = tensor("input_69_groups_0"), val = tensor(1)]; - tensor input_69 = conv(bias = encoder_conv_module_1_sequential_7_conv_bias, dilations = input_69_dilations_0, groups = input_69_groups_0, pad = input_69_pad_0, pad_type = input_69_pad_type_0, strides = input_69_strides_0, weight = encoder_conv_module_1_sequential_7_conv_weight, x = input_67)[name = tensor("input_69")]; + tensor outputs_aug_3 = conv(dilations = outputs_aug_3_dilations_0, groups = outputs_aug_3_groups_0, pad = outputs_aug_3_pad_0, pad_type = outputs_aug_3_pad_type_0, strides = outputs_aug_3_strides_0, weight = inner_encoder_conv_module_1_sequential_4_conv_weight, x = inputs_15)[name = tensor("outputs_aug_3")]; + tensor input_67_begin_0 = const()[name = tensor("input_67_begin_0"), val = tensor([0, 0, 0])]; + tensor input_67_end_0 = const()[name = tensor("input_67_end_0"), val = tensor([3, 256, 16])]; + tensor input_67_end_mask_0 = const()[name = tensor("input_67_end_mask_0"), val = tensor([true, true, false])]; + tensor input_67 = slice_by_index(begin = input_67_begin_0, end = input_67_end_0, end_mask = input_67_end_mask_0, x = outputs_aug_3)[name = tensor("input_67")]; + tensor inputs_17 = batch_norm(beta = inner_encoder_conv_module_1_sequential_5_bias, epsilon = var_56, gamma = inner_encoder_conv_module_1_sequential_5_weight, mean = inner_encoder_conv_module_1_sequential_5_running_mean, variance = inner_encoder_conv_module_1_sequential_5_running_var, x = input_67)[name = tensor("inputs_17")]; + tensor input_69 = silu(x = inputs_17)[name = tensor("input_69")]; + tensor input_71_pad_type_0 = const()[name = tensor("input_71_pad_type_0"), val = tensor("valid")]; + tensor input_71_strides_0 = const()[name = tensor("input_71_strides_0"), val = tensor([1])]; + tensor input_71_pad_0 = const()[name = tensor("input_71_pad_0"), val = tensor([0, 0])]; + tensor input_71_dilations_0 = const()[name = tensor("input_71_dilations_0"), val = tensor([1])]; + tensor input_71_groups_0 = const()[name = tensor("input_71_groups_0"), val = tensor(1)]; + tensor input_71 = conv(bias = inner_encoder_conv_module_1_sequential_7_conv_bias, dilations = input_71_dilations_0, groups = input_71_groups_0, pad = input_71_pad_0, pad_type = input_71_pad_type_0, strides = input_71_strides_0, weight = inner_encoder_conv_module_1_sequential_7_conv_weight, x = input_69)[name = tensor("input_71")]; tensor conv_out_3_perm_0 = const()[name = tensor("conv_out_3_perm_0"), val = tensor([0, 2, 1])]; - tensor var_504_begin_0 = const()[name = tensor("op_504_begin_0"), val = tensor([0, -1, 0])]; - tensor var_504_end_0 = const()[name = tensor("op_504_end_0"), val = tensor([3, 16, 256])]; - tensor var_504_end_mask_0 = const()[name = tensor("op_504_end_mask_0"), val = tensor([true, true, true])]; - tensor conv_out_3 = transpose(perm = conv_out_3_perm_0, x = input_69)[name = tensor("transpose_46")]; - tensor var_504 = slice_by_index(begin = var_504_begin_0, end = var_504_end_0, end_mask = var_504_end_mask_0, x = conv_out_3)[name = tensor("op_504")]; - tensor var_506_perm_0 = const()[name = tensor("op_506_perm_0"), val = tensor([1, 0, 2])]; - tensor var_506 = transpose(perm = var_506_perm_0, x = var_504)[name = tensor("transpose_45")]; - tensor input_71 = add(x = x_9, y = var_506)[name = tensor("input_71")]; - tensor input_73_axes_0 = const()[name = tensor("input_73_axes_0"), val = tensor([-1])]; - tensor input_73 = layer_norm(axes = input_73_axes_0, beta = encoder_ffn2_1_module_sequential_0_bias, epsilon = var_28, gamma = encoder_ffn2_1_module_sequential_0_weight, x = input_71)[name = tensor("input_73")]; - tensor inputs_19 = linear(bias = encoder_ffn2_1_module_sequential_1_linear_bias, weight = encoder_ffn2_1_module_sequential_1_linear_weight, x = input_73)[name = tensor("linear_17")]; - tensor input_75 = silu(x = inputs_19)[name = tensor("input_75")]; - tensor input_79 = linear(bias = encoder_ffn2_1_module_sequential_4_linear_bias, weight = encoder_ffn2_1_module_sequential_4_linear_weight, x = input_75)[name = tensor("linear_18")]; - tensor var_529 = const()[name = tensor("op_529"), val = tensor(0x1p-1)]; - tensor var_530 = mul(x = input_79, y = var_529)[name = tensor("op_530")]; - tensor input_81 = add(x = var_530, y = input_71)[name = tensor("input_81")]; - tensor input_83_axes_0 = const()[name = tensor("input_83_axes_0"), val = tensor([-1])]; - tensor input_83 = layer_norm(axes = input_83_axes_0, beta = encoder_layer_norm_1_bias, epsilon = var_28, gamma = encoder_layer_norm_1_weight, x = input_81)[name = tensor("input_83")]; + tensor var_550_begin_0 = const()[name = tensor("op_550_begin_0"), val = tensor([0, -1, 0])]; + tensor var_550_end_0 = const()[name = tensor("op_550_end_0"), val = tensor([3, 16, 256])]; + tensor var_550_end_mask_0 = const()[name = tensor("op_550_end_mask_0"), val = tensor([true, true, true])]; + tensor conv_out_3 = transpose(perm = conv_out_3_perm_0, x = input_71)[name = tensor("transpose_46")]; + tensor var_550 = slice_by_index(begin = var_550_begin_0, end = var_550_end_0, end_mask = var_550_end_mask_0, x = conv_out_3)[name = tensor("op_550")]; + tensor var_552_perm_0 = const()[name = tensor("op_552_perm_0"), val = tensor([1, 0, 2])]; + tensor var_552 = transpose(perm = var_552_perm_0, x = var_550)[name = tensor("transpose_45")]; + tensor input_73 = add(x = x_9, y = var_552)[name = tensor("input_73")]; + tensor input_75_axes_0 = const()[name = tensor("input_75_axes_0"), val = tensor([-1])]; + tensor input_75 = layer_norm(axes = input_75_axes_0, beta = inner_encoder_ffn2_1_module_sequential_0_bias, epsilon = var_56, gamma = inner_encoder_ffn2_1_module_sequential_0_weight, x = input_73)[name = tensor("input_75")]; + tensor inputs_19 = linear(bias = inner_encoder_ffn2_1_module_sequential_1_linear_bias, weight = inner_encoder_ffn2_1_module_sequential_1_linear_weight, x = input_75)[name = tensor("linear_17")]; + tensor input_77 = silu(x = inputs_19)[name = tensor("input_77")]; + tensor input_81 = linear(bias = inner_encoder_ffn2_1_module_sequential_4_linear_bias, weight = inner_encoder_ffn2_1_module_sequential_4_linear_weight, x = input_77)[name = tensor("linear_18")]; + tensor var_575 = const()[name = tensor("op_575"), val = tensor(0x1p-1)]; + tensor var_576 = mul(x = input_81, y = var_575)[name = tensor("op_576")]; + tensor input_83 = add(x = var_576, y = input_73)[name = tensor("input_83")]; tensor input_85_axes_0 = const()[name = tensor("input_85_axes_0"), val = tensor([-1])]; - tensor input_85 = layer_norm(axes = input_85_axes_0, beta = encoder_ffn1_2_module_sequential_0_bias, epsilon = var_28, gamma = encoder_ffn1_2_module_sequential_0_weight, x = input_83)[name = tensor("input_85")]; - tensor inputs_21 = linear(bias = encoder_ffn1_2_module_sequential_1_linear_bias, weight = encoder_ffn1_2_module_sequential_1_linear_weight, x = input_85)[name = tensor("linear_19")]; - tensor input_87 = silu(x = inputs_21)[name = tensor("input_87")]; - tensor input_91 = linear(bias = encoder_ffn1_2_module_sequential_4_linear_bias, weight = encoder_ffn1_2_module_sequential_4_linear_weight, x = input_87)[name = tensor("linear_20")]; - tensor var_559 = const()[name = tensor("op_559"), val = tensor(0x1p-1)]; - tensor var_560 = mul(x = input_91, y = var_559)[name = tensor("op_560")]; - tensor input_93 = add(x = var_560, y = input_83)[name = tensor("input_93")]; + tensor input_85 = layer_norm(axes = input_85_axes_0, beta = inner_encoder_layer_norm_1_bias, epsilon = var_56, gamma = inner_encoder_layer_norm_1_weight, x = input_83)[name = tensor("input_85")]; + tensor input_87_axes_0 = const()[name = tensor("input_87_axes_0"), val = tensor([-1])]; + tensor input_87 = layer_norm(axes = input_87_axes_0, beta = inner_encoder_ffn1_2_module_sequential_0_bias, epsilon = var_56, gamma = inner_encoder_ffn1_2_module_sequential_0_weight, x = input_85)[name = tensor("input_87")]; + tensor inputs_21 = linear(bias = inner_encoder_ffn1_2_module_sequential_1_linear_bias, weight = inner_encoder_ffn1_2_module_sequential_1_linear_weight, x = input_87)[name = tensor("linear_19")]; + tensor input_89 = silu(x = inputs_21)[name = tensor("input_89")]; + tensor input_93 = linear(bias = inner_encoder_ffn1_2_module_sequential_4_linear_bias, weight = inner_encoder_ffn1_2_module_sequential_4_linear_weight, x = input_89)[name = tensor("linear_20")]; + tensor var_605 = const()[name = tensor("op_605"), val = tensor(0x1p-1)]; + tensor var_606 = mul(x = input_93, y = var_605)[name = tensor("op_606")]; + tensor input_95 = add(x = var_606, y = input_85)[name = tensor("input_95")]; tensor x_13_axes_0 = const()[name = tensor("x_13_axes_0"), val = tensor([-1])]; - tensor x_13 = layer_norm(axes = x_13_axes_0, beta = encoder_ret_lns_2_bias, epsilon = var_28, gamma = encoder_ret_lns_2_weight, x = input_93)[name = tensor("x_13")]; + tensor x_13 = layer_norm(axes = x_13_axes_0, beta = inner_encoder_ret_lns_2_bias, epsilon = var_56, gamma = inner_encoder_ret_lns_2_weight, x = input_95)[name = tensor("x_13")]; tensor prev_kv_5_begin_0 = const()[name = tensor("prev_kv_5_begin_0"), val = tensor([2, 0, 0, 0, 0])]; tensor prev_kv_5_end_0 = const()[name = tensor("prev_kv_5_end_0"), val = tensor([3, 1, 4, 64, 64])]; tensor prev_kv_5_end_mask_0 = const()[name = tensor("prev_kv_5_end_mask_0"), val = tensor([false, true, true, true, true])]; @@ -573,163 +591,163 @@ program(1.0) tensor prev_scale_5_end_mask_0 = const()[name = tensor("prev_scale_5_end_mask_0"), val = tensor([false, true])]; tensor prev_scale_5_squeeze_mask_0 = const()[name = tensor("prev_scale_5_squeeze_mask_0"), val = tensor([true, false])]; tensor prev_scale_5 = slice_by_index(begin = prev_scale_5_begin_0, end = prev_scale_5_end_0, end_mask = prev_scale_5_end_mask_0, squeeze_mask = prev_scale_5_squeeze_mask_0, x = enc_scale)[name = tensor("prev_scale_5")]; - tensor var_574 = linear(bias = encoder_q_proj_2_bias, weight = encoder_q_proj_2_weight, x = x_13)[name = tensor("linear_21")]; - tensor var_575 = const()[name = tensor("op_575"), val = tensor([1, 3, 4, 64])]; - tensor var_576 = reshape(shape = var_575, x = var_574)[name = tensor("op_576")]; + tensor var_620 = linear(bias = inner_encoder_q_proj_2_bias, weight = inner_encoder_q_proj_2_weight, x = x_13)[name = tensor("linear_21")]; + tensor var_621 = const()[name = tensor("op_621"), val = tensor([1, 3, 4, 64])]; + tensor var_622 = reshape(shape = var_621, x = var_620)[name = tensor("op_622")]; tensor q_5_perm_0 = const()[name = tensor("q_5_perm_0"), val = tensor([0, 2, 1, 3])]; - tensor var_580 = linear(bias = encoder_k_proj_2_bias, weight = encoder_k_proj_2_weight, x = x_13)[name = tensor("linear_22")]; - tensor var_581 = const()[name = tensor("op_581"), val = tensor(0x1p-3)]; - tensor var_582 = mul(x = var_580, y = var_581)[name = tensor("op_582")]; - tensor var_583 = const()[name = tensor("op_583"), val = tensor([1, 3, 4, 64])]; - tensor var_584 = reshape(shape = var_583, x = var_582)[name = tensor("op_584")]; + tensor var_626 = linear(bias = inner_encoder_k_proj_2_bias, weight = inner_encoder_k_proj_2_weight, x = x_13)[name = tensor("linear_22")]; + tensor var_627 = const()[name = tensor("op_627"), val = tensor(0x1p-3)]; + tensor var_628 = mul(x = var_626, y = var_627)[name = tensor("op_628")]; + tensor var_629 = const()[name = tensor("op_629"), val = tensor([1, 3, 4, 64])]; + tensor var_630 = reshape(shape = var_629, x = var_628)[name = tensor("op_630")]; tensor k_5_perm_0 = const()[name = tensor("k_5_perm_0"), val = tensor([0, 2, 1, 3])]; - tensor var_588 = linear(bias = encoder_v_proj_2_bias, weight = encoder_v_proj_2_weight, x = x_13)[name = tensor("linear_23")]; - tensor var_589 = const()[name = tensor("op_589"), val = tensor([1, 3, 4, 64])]; - tensor var_590 = reshape(shape = var_589, x = var_588)[name = tensor("op_590")]; + tensor var_634 = linear(bias = inner_encoder_v_proj_2_bias, weight = inner_encoder_v_proj_2_weight, x = x_13)[name = tensor("linear_23")]; + tensor var_635 = const()[name = tensor("op_635"), val = tensor([1, 3, 4, 64])]; + tensor var_636 = reshape(shape = var_635, x = var_634)[name = tensor("op_636")]; tensor v_5_perm_0 = const()[name = tensor("v_5_perm_0"), val = tensor([0, 2, 1, 3])]; - tensor input_97 = linear(bias = encoder_g_proj_2_bias, weight = encoder_g_proj_2_weight, x = x_13)[name = tensor("linear_24")]; + tensor input_99 = linear(bias = inner_encoder_g_proj_2_bias, weight = inner_encoder_g_proj_2_weight, x = x_13)[name = tensor("linear_24")]; tensor sqrt_s0_5 = sqrt(x = prev_scale_5)[name = tensor("sqrt_s0_5")]; - tensor s_t_5 = add(x = prev_scale_5, y = encoder__t_index)[name = tensor("s_t_5")]; + tensor s_t_5 = add(x = prev_scale_5, y = inner_encoder__t_index)[name = tensor("s_t_5")]; tensor sqrt_s_t_5 = sqrt(x = s_t_5)[name = tensor("sqrt_s_t_5")]; tensor qk_5_transpose_x_1 = const()[name = tensor("qk_5_transpose_x_1"), val = tensor(false)]; tensor qk_5_transpose_y_1 = const()[name = tensor("qk_5_transpose_y_1"), val = tensor(true)]; - tensor k_5 = transpose(perm = k_5_perm_0, x = var_584)[name = tensor("transpose_43")]; - tensor q_5 = transpose(perm = q_5_perm_0, x = var_576)[name = tensor("transpose_44")]; + tensor k_5 = transpose(perm = k_5_perm_0, x = var_630)[name = tensor("transpose_43")]; + tensor q_5 = transpose(perm = q_5_perm_0, x = var_622)[name = tensor("transpose_44")]; tensor qk_5 = matmul(transpose_x = qk_5_transpose_x_1, transpose_y = qk_5_transpose_y_1, x = q_5, y = k_5)[name = tensor("qk_5")]; - tensor var_600 = const()[name = tensor("op_600"), val = tensor([3, 1])]; - tensor var_601 = reshape(shape = var_600, x = sqrt_s_t_5)[name = tensor("op_601")]; - tensor M_5 = real_div(x = encoder__causal_mask, y = var_601)[name = tensor("M_5")]; - tensor var_603 = mul(x = qk_5, y = M_5)[name = tensor("op_603")]; + tensor var_646 = const()[name = tensor("op_646"), val = tensor([3, 1])]; + tensor var_647 = reshape(shape = var_646, x = sqrt_s_t_5)[name = tensor("op_647")]; + tensor M_5 = real_div(x = inner_encoder__causal_mask, y = var_647)[name = tensor("M_5")]; + tensor var_649 = mul(x = qk_5, y = M_5)[name = tensor("op_649")]; tensor inner_5_transpose_x_0 = const()[name = tensor("inner_5_transpose_x_0"), val = tensor(false)]; tensor inner_5_transpose_y_0 = const()[name = tensor("inner_5_transpose_y_0"), val = tensor(false)]; - tensor v_5 = transpose(perm = v_5_perm_0, x = var_590)[name = tensor("transpose_42")]; - tensor inner_5 = matmul(transpose_x = inner_5_transpose_x_0, transpose_y = inner_5_transpose_y_0, x = var_603, y = v_5)[name = tensor("inner_5")]; - tensor var_605_transpose_x_0 = const()[name = tensor("op_605_transpose_x_0"), val = tensor(false)]; - tensor var_605_transpose_y_0 = const()[name = tensor("op_605_transpose_y_0"), val = tensor(false)]; - tensor var_605 = matmul(transpose_x = var_605_transpose_x_0, transpose_y = var_605_transpose_y_0, x = q_5, y = prev_kv_5)[name = tensor("op_605")]; - tensor var_606 = real_div(x = sqrt_s0_5, y = sqrt_s_t_5)[name = tensor("op_606")]; - tensor var_607 = const()[name = tensor("op_607"), val = tensor([1, 1, 3, 1])]; - tensor var_608 = reshape(shape = var_607, x = var_606)[name = tensor("op_608")]; - tensor cross_5 = mul(x = var_605, y = var_608)[name = tensor("cross_5")]; + tensor v_5 = transpose(perm = v_5_perm_0, x = var_636)[name = tensor("transpose_42")]; + tensor inner_5 = matmul(transpose_x = inner_5_transpose_x_0, transpose_y = inner_5_transpose_y_0, x = var_649, y = v_5)[name = tensor("inner_5")]; + tensor var_651_transpose_x_0 = const()[name = tensor("op_651_transpose_x_0"), val = tensor(false)]; + tensor var_651_transpose_y_0 = const()[name = tensor("op_651_transpose_y_0"), val = tensor(false)]; + tensor var_651 = matmul(transpose_x = var_651_transpose_x_0, transpose_y = var_651_transpose_y_0, x = q_5, y = prev_kv_5)[name = tensor("op_651")]; + tensor var_652 = real_div(x = sqrt_s0_5, y = sqrt_s_t_5)[name = tensor("op_652")]; + tensor var_653 = const()[name = tensor("op_653"), val = tensor([1, 1, 3, 1])]; + tensor var_654 = reshape(shape = var_653, x = var_652)[name = tensor("op_654")]; + tensor cross_5 = mul(x = var_651, y = var_654)[name = tensor("cross_5")]; tensor out_13 = add(x = inner_5, y = cross_5)[name = tensor("out_13")]; - tensor var_611 = mul(x = prev_kv_5, y = sqrt_s0_5)[name = tensor("op_611")]; - tensor var_613_transpose_x_1 = const()[name = tensor("op_613_transpose_x_1"), val = tensor(true)]; - tensor var_613_transpose_y_1 = const()[name = tensor("op_613_transpose_y_1"), val = tensor(false)]; - tensor var_613 = matmul(transpose_x = var_613_transpose_x_1, transpose_y = var_613_transpose_y_1, x = k_5, y = v_5)[name = tensor("op_613")]; - tensor new_kv_unnorm_5 = add(x = var_611, y = var_613)[name = tensor("new_kv_unnorm_5")]; - tensor var_615 = const()[name = tensor("op_615"), val = tensor(0x1.8p+1)]; - tensor new_scale_5 = add(x = prev_scale_5, y = var_615)[name = tensor("new_scale_5")]; - tensor var_617 = sqrt(x = new_scale_5)[name = tensor("op_617")]; - tensor var_618 = real_div(x = new_kv_unnorm_5, y = var_617)[name = tensor("op_618")]; - tensor var_619_perm_0 = const()[name = tensor("op_619_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_657 = mul(x = prev_kv_5, y = sqrt_s0_5)[name = tensor("op_657")]; + tensor var_659_transpose_x_1 = const()[name = tensor("op_659_transpose_x_1"), val = tensor(true)]; + tensor var_659_transpose_y_1 = const()[name = tensor("op_659_transpose_y_1"), val = tensor(false)]; + tensor var_659 = matmul(transpose_x = var_659_transpose_x_1, transpose_y = var_659_transpose_y_1, x = k_5, y = v_5)[name = tensor("op_659")]; + tensor new_kv_unnorm_5 = add(x = var_657, y = var_659)[name = tensor("new_kv_unnorm_5")]; + tensor var_661 = const()[name = tensor("op_661"), val = tensor(0x1.8p+1)]; + tensor new_scale_5 = add(x = prev_scale_5, y = var_661)[name = tensor("new_scale_5")]; + tensor var_663 = sqrt(x = new_scale_5)[name = tensor("op_663")]; + tensor var_664 = real_div(x = new_kv_unnorm_5, y = var_663)[name = tensor("op_664")]; + tensor var_665_perm_0 = const()[name = tensor("op_665_perm_0"), val = tensor([0, 2, 1, 3])]; tensor out_15_axes_0 = const()[name = tensor("out_15_axes_0"), val = tensor([-1])]; - tensor var_619 = transpose(perm = var_619_perm_0, x = out_13)[name = tensor("transpose_41")]; - tensor out_15 = layer_norm(axes = out_15_axes_0, epsilon = var_17, x = var_619)[name = tensor("out_15")]; - tensor var_623 = const()[name = tensor("op_623"), val = tensor([1, 3, 256])]; - tensor out_17 = reshape(shape = var_623, x = out_15)[name = tensor("out_17")]; - tensor var_625 = silu(x = input_97)[name = tensor("op_625")]; - tensor input_99 = mul(x = var_625, y = out_17)[name = tensor("input_99")]; - tensor ret_out_5 = linear(bias = encoder_out_proj_2_bias, weight = encoder_out_proj_2_weight, x = input_99)[name = tensor("linear_25")]; - tensor x_15 = add(x = input_93, y = ret_out_5)[name = tensor("x_15")]; + tensor var_665 = transpose(perm = var_665_perm_0, x = out_13)[name = tensor("transpose_41")]; + tensor out_15 = layer_norm(axes = out_15_axes_0, epsilon = var_64, x = var_665)[name = tensor("out_15")]; + tensor var_669 = const()[name = tensor("op_669"), val = tensor([1, 3, 256])]; + tensor out_17 = reshape(shape = var_669, x = out_15)[name = tensor("out_17")]; + tensor var_671 = silu(x = input_99)[name = tensor("op_671")]; + tensor input_101 = mul(x = var_671, y = out_17)[name = tensor("input_101")]; + tensor ret_out_5 = linear(bias = inner_encoder_out_proj_2_bias, weight = inner_encoder_out_proj_2_weight, x = input_101)[name = tensor("linear_25")]; + tensor x_15 = add(x = input_95, y = ret_out_5)[name = tensor("x_15")]; tensor window_17_begin_0 = const()[name = tensor("window_17_begin_0"), val = tensor([2, 0, 0, 0])]; tensor window_17_end_0 = const()[name = tensor("window_17_end_0"), val = tensor([3, 1, 16, 256])]; tensor window_17_end_mask_0 = const()[name = tensor("window_17_end_mask_0"), val = tensor([false, true, true, true])]; tensor window_17_squeeze_mask_0 = const()[name = tensor("window_17_squeeze_mask_0"), val = tensor([true, false, false, false])]; tensor window_17 = slice_by_index(begin = window_17_begin_0, end = window_17_end_0, end_mask = window_17_end_mask_0, squeeze_mask = window_17_squeeze_mask_0, x = enc_conv_cache)[name = tensor("window_17")]; - tensor var_633_begin_0 = const()[name = tensor("op_633_begin_0"), val = tensor([0, 0, 0])]; - tensor var_633_end_0 = const()[name = tensor("op_633_end_0"), val = tensor([1, 1, 256])]; - tensor var_633_end_mask_0 = const()[name = tensor("op_633_end_mask_0"), val = tensor([true, false, true])]; - tensor var_633 = slice_by_index(begin = var_633_begin_0, end = var_633_end_0, end_mask = var_633_end_mask_0, x = x_15)[name = tensor("op_633")]; - tensor var_636_begin_0 = const()[name = tensor("op_636_begin_0"), val = tensor([0, 1, 0])]; - tensor var_636_end_0 = const()[name = tensor("op_636_end_0"), val = tensor([1, 16, 256])]; - tensor var_636_end_mask_0 = const()[name = tensor("op_636_end_mask_0"), val = tensor([true, true, true])]; - tensor var_636 = slice_by_index(begin = var_636_begin_0, end = var_636_end_0, end_mask = var_636_end_mask_0, x = window_17)[name = tensor("op_636")]; + tensor var_679_begin_0 = const()[name = tensor("op_679_begin_0"), val = tensor([0, 0, 0])]; + tensor var_679_end_0 = const()[name = tensor("op_679_end_0"), val = tensor([1, 1, 256])]; + tensor var_679_end_mask_0 = const()[name = tensor("op_679_end_mask_0"), val = tensor([true, false, true])]; + tensor var_679 = slice_by_index(begin = var_679_begin_0, end = var_679_end_0, end_mask = var_679_end_mask_0, x = x_15)[name = tensor("op_679")]; + tensor var_682_begin_0 = const()[name = tensor("op_682_begin_0"), val = tensor([0, 1, 0])]; + tensor var_682_end_0 = const()[name = tensor("op_682_end_0"), val = tensor([1, 16, 256])]; + tensor var_682_end_mask_0 = const()[name = tensor("op_682_end_mask_0"), val = tensor([true, true, true])]; + tensor var_682 = slice_by_index(begin = var_682_begin_0, end = var_682_end_0, end_mask = var_682_end_mask_0, x = window_17)[name = tensor("op_682")]; tensor window_19_interleave_0 = const()[name = tensor("window_19_interleave_0"), val = tensor(false)]; - tensor window_19 = concat(axis = var_26, interleave = window_19_interleave_0, values = (var_636, var_633))[name = tensor("window_19")]; - tensor var_641_begin_0 = const()[name = tensor("op_641_begin_0"), val = tensor([0, 1, 0])]; - tensor var_641_end_0 = const()[name = tensor("op_641_end_0"), val = tensor([1, 2, 256])]; - tensor var_641_end_mask_0 = const()[name = tensor("op_641_end_mask_0"), val = tensor([true, false, true])]; - tensor var_641 = slice_by_index(begin = var_641_begin_0, end = var_641_end_0, end_mask = var_641_end_mask_0, x = x_15)[name = tensor("op_641")]; - tensor var_644_begin_0 = const()[name = tensor("op_644_begin_0"), val = tensor([0, 1, 0])]; - tensor var_644_end_0 = const()[name = tensor("op_644_end_0"), val = tensor([1, 16, 256])]; - tensor var_644_end_mask_0 = const()[name = tensor("op_644_end_mask_0"), val = tensor([true, true, true])]; - tensor var_644 = slice_by_index(begin = var_644_begin_0, end = var_644_end_0, end_mask = var_644_end_mask_0, x = window_19)[name = tensor("op_644")]; + tensor window_19 = concat(axis = var_72, interleave = window_19_interleave_0, values = (var_682, var_679))[name = tensor("window_19")]; + tensor var_687_begin_0 = const()[name = tensor("op_687_begin_0"), val = tensor([0, 1, 0])]; + tensor var_687_end_0 = const()[name = tensor("op_687_end_0"), val = tensor([1, 2, 256])]; + tensor var_687_end_mask_0 = const()[name = tensor("op_687_end_mask_0"), val = tensor([true, false, true])]; + tensor var_687 = slice_by_index(begin = var_687_begin_0, end = var_687_end_0, end_mask = var_687_end_mask_0, x = x_15)[name = tensor("op_687")]; + tensor var_690_begin_0 = const()[name = tensor("op_690_begin_0"), val = tensor([0, 1, 0])]; + tensor var_690_end_0 = const()[name = tensor("op_690_end_0"), val = tensor([1, 16, 256])]; + tensor var_690_end_mask_0 = const()[name = tensor("op_690_end_mask_0"), val = tensor([true, true, true])]; + tensor var_690 = slice_by_index(begin = var_690_begin_0, end = var_690_end_0, end_mask = var_690_end_mask_0, x = window_19)[name = tensor("op_690")]; tensor window_21_interleave_0 = const()[name = tensor("window_21_interleave_0"), val = tensor(false)]; - tensor window_21 = concat(axis = var_26, interleave = window_21_interleave_0, values = (var_644, var_641))[name = tensor("window_21")]; - tensor var_649_begin_0 = const()[name = tensor("op_649_begin_0"), val = tensor([0, 2, 0])]; - tensor var_649_end_0 = const()[name = tensor("op_649_end_0"), val = tensor([1, 1, 256])]; - tensor var_649_end_mask_0 = const()[name = tensor("op_649_end_mask_0"), val = tensor([true, true, true])]; - tensor var_649 = slice_by_index(begin = var_649_begin_0, end = var_649_end_0, end_mask = var_649_end_mask_0, x = x_15)[name = tensor("op_649")]; - tensor var_652_begin_0 = const()[name = tensor("op_652_begin_0"), val = tensor([0, 1, 0])]; - tensor var_652_end_0 = const()[name = tensor("op_652_end_0"), val = tensor([1, 16, 256])]; - tensor var_652_end_mask_0 = const()[name = tensor("op_652_end_mask_0"), val = tensor([true, true, true])]; - tensor var_652 = slice_by_index(begin = var_652_begin_0, end = var_652_end_0, end_mask = var_652_end_mask_0, x = window_21)[name = tensor("op_652")]; + tensor window_21 = concat(axis = var_72, interleave = window_21_interleave_0, values = (var_690, var_687))[name = tensor("window_21")]; + tensor var_695_begin_0 = const()[name = tensor("op_695_begin_0"), val = tensor([0, 2, 0])]; + tensor var_695_end_0 = const()[name = tensor("op_695_end_0"), val = tensor([1, 1, 256])]; + tensor var_695_end_mask_0 = const()[name = tensor("op_695_end_mask_0"), val = tensor([true, true, true])]; + tensor var_695 = slice_by_index(begin = var_695_begin_0, end = var_695_end_0, end_mask = var_695_end_mask_0, x = x_15)[name = tensor("op_695")]; + tensor var_698_begin_0 = const()[name = tensor("op_698_begin_0"), val = tensor([0, 1, 0])]; + tensor var_698_end_0 = const()[name = tensor("op_698_end_0"), val = tensor([1, 16, 256])]; + tensor var_698_end_mask_0 = const()[name = tensor("op_698_end_mask_0"), val = tensor([true, true, true])]; + tensor var_698 = slice_by_index(begin = var_698_begin_0, end = var_698_end_0, end_mask = var_698_end_mask_0, x = window_21)[name = tensor("op_698")]; tensor window_23_interleave_0 = const()[name = tensor("window_23_interleave_0"), val = tensor(false)]; - tensor window_23 = concat(axis = var_26, interleave = window_23_interleave_0, values = (var_652, var_649))[name = tensor("window_23")]; - tensor input_101_interleave_0 = const()[name = tensor("input_101_interleave_0"), val = tensor(false)]; - tensor input_101 = concat(axis = var_23, interleave = input_101_interleave_0, values = (window_19, window_21, window_23))[name = tensor("input_101")]; + tensor window_23 = concat(axis = var_72, interleave = window_23_interleave_0, values = (var_698, var_695))[name = tensor("window_23")]; + tensor input_103_interleave_0 = const()[name = tensor("input_103_interleave_0"), val = tensor(false)]; + tensor input_103 = concat(axis = var_59, interleave = input_103_interleave_0, values = (window_19, window_21, window_23))[name = tensor("input_103")]; tensor x_17_axes_0 = const()[name = tensor("x_17_axes_0"), val = tensor([-1])]; - tensor x_17 = layer_norm(axes = x_17_axes_0, beta = encoder_conv_module_2_sequential_0_bias, epsilon = var_28, gamma = encoder_conv_module_2_sequential_0_weight, x = input_101)[name = tensor("x_17")]; - tensor input_103_perm_0 = const()[name = tensor("input_103_perm_0"), val = tensor([0, 2, 1])]; + tensor x_17 = layer_norm(axes = x_17_axes_0, beta = inner_encoder_conv_module_2_sequential_0_bias, epsilon = var_56, gamma = inner_encoder_conv_module_2_sequential_0_weight, x = input_103)[name = tensor("x_17")]; + tensor input_105_perm_0 = const()[name = tensor("input_105_perm_0"), val = tensor([0, 2, 1])]; tensor inputs_23_pad_type_0 = const()[name = tensor("inputs_23_pad_type_0"), val = tensor("valid")]; tensor inputs_23_strides_0 = const()[name = tensor("inputs_23_strides_0"), val = tensor([1])]; tensor inputs_23_pad_0 = const()[name = tensor("inputs_23_pad_0"), val = tensor([0, 0])]; tensor inputs_23_dilations_0 = const()[name = tensor("inputs_23_dilations_0"), val = tensor([1])]; tensor inputs_23_groups_0 = const()[name = tensor("inputs_23_groups_0"), val = tensor(1)]; - tensor input_103 = transpose(perm = input_103_perm_0, x = x_17)[name = tensor("transpose_40")]; - tensor inputs_23 = conv(bias = encoder_conv_module_2_sequential_2_conv_bias, dilations = inputs_23_dilations_0, groups = inputs_23_groups_0, pad = inputs_23_pad_0, pad_type = inputs_23_pad_type_0, strides = inputs_23_strides_0, weight = encoder_conv_module_2_sequential_2_conv_weight, x = input_103)[name = tensor("inputs_23")]; - tensor var_677_split_sizes_0 = const()[name = tensor("op_677_split_sizes_0"), val = tensor([256, 256])]; - tensor var_677_axis_0 = const()[name = tensor("op_677_axis_0"), val = tensor(1)]; - tensor var_677_0, tensor var_677_1 = split(axis = var_677_axis_0, split_sizes = var_677_split_sizes_0, x = inputs_23)[name = tensor("op_677")]; - tensor var_679 = sigmoid(x = var_677_1)[name = tensor("op_679")]; - tensor inputs_25 = mul(x = var_677_0, y = var_679)[name = tensor("inputs_25")]; + tensor input_105 = transpose(perm = input_105_perm_0, x = x_17)[name = tensor("transpose_40")]; + tensor inputs_23 = conv(bias = inner_encoder_conv_module_2_sequential_2_conv_bias, dilations = inputs_23_dilations_0, groups = inputs_23_groups_0, pad = inputs_23_pad_0, pad_type = inputs_23_pad_type_0, strides = inputs_23_strides_0, weight = inner_encoder_conv_module_2_sequential_2_conv_weight, x = input_105)[name = tensor("inputs_23")]; + tensor var_723_split_sizes_0 = const()[name = tensor("op_723_split_sizes_0"), val = tensor([256, 256])]; + tensor var_723_axis_0 = const()[name = tensor("op_723_axis_0"), val = tensor(1)]; + tensor var_723_0, tensor var_723_1 = split(axis = var_723_axis_0, split_sizes = var_723_split_sizes_0, x = inputs_23)[name = tensor("op_723")]; + tensor var_725 = sigmoid(x = var_723_1)[name = tensor("op_725")]; + tensor inputs_25 = mul(x = var_723_0, y = var_725)[name = tensor("inputs_25")]; tensor outputs_aug_5_pad_type_0 = const()[name = tensor("outputs_aug_5_pad_type_0"), val = tensor("custom")]; tensor outputs_aug_5_pad_0 = const()[name = tensor("outputs_aug_5_pad_0"), val = tensor([15, 15])]; tensor outputs_aug_5_groups_0 = const()[name = tensor("outputs_aug_5_groups_0"), val = tensor(256)]; tensor outputs_aug_5_strides_0 = const()[name = tensor("outputs_aug_5_strides_0"), val = tensor([1])]; tensor outputs_aug_5_dilations_0 = const()[name = tensor("outputs_aug_5_dilations_0"), val = tensor([1])]; - tensor outputs_aug_5 = conv(dilations = outputs_aug_5_dilations_0, groups = outputs_aug_5_groups_0, pad = outputs_aug_5_pad_0, pad_type = outputs_aug_5_pad_type_0, strides = outputs_aug_5_strides_0, weight = encoder_conv_module_2_sequential_4_conv_weight, x = inputs_25)[name = tensor("outputs_aug_5")]; - tensor input_105_begin_0 = const()[name = tensor("input_105_begin_0"), val = tensor([0, 0, 0])]; - tensor input_105_end_0 = const()[name = tensor("input_105_end_0"), val = tensor([3, 256, 16])]; - tensor input_105_end_mask_0 = const()[name = tensor("input_105_end_mask_0"), val = tensor([true, true, false])]; - tensor input_105 = slice_by_index(begin = input_105_begin_0, end = input_105_end_0, end_mask = input_105_end_mask_0, x = outputs_aug_5)[name = tensor("input_105")]; - tensor inputs_27 = batch_norm(beta = encoder_conv_module_2_sequential_5_bias, epsilon = var_28, gamma = encoder_conv_module_2_sequential_5_weight, mean = encoder_conv_module_2_sequential_5_running_mean, variance = encoder_conv_module_2_sequential_5_running_var, x = input_105)[name = tensor("inputs_27")]; - tensor input_107 = silu(x = inputs_27)[name = tensor("input_107")]; - tensor input_109_pad_type_0 = const()[name = tensor("input_109_pad_type_0"), val = tensor("valid")]; - tensor input_109_strides_0 = const()[name = tensor("input_109_strides_0"), val = tensor([1])]; - tensor input_109_pad_0 = const()[name = tensor("input_109_pad_0"), val = tensor([0, 0])]; - tensor input_109_dilations_0 = const()[name = tensor("input_109_dilations_0"), val = tensor([1])]; - tensor input_109_groups_0 = const()[name = tensor("input_109_groups_0"), val = tensor(1)]; - tensor input_109 = conv(bias = encoder_conv_module_2_sequential_7_conv_bias, dilations = input_109_dilations_0, groups = input_109_groups_0, pad = input_109_pad_0, pad_type = input_109_pad_type_0, strides = input_109_strides_0, weight = encoder_conv_module_2_sequential_7_conv_weight, x = input_107)[name = tensor("input_109")]; + tensor outputs_aug_5 = conv(dilations = outputs_aug_5_dilations_0, groups = outputs_aug_5_groups_0, pad = outputs_aug_5_pad_0, pad_type = outputs_aug_5_pad_type_0, strides = outputs_aug_5_strides_0, weight = inner_encoder_conv_module_2_sequential_4_conv_weight, x = inputs_25)[name = tensor("outputs_aug_5")]; + tensor input_107_begin_0 = const()[name = tensor("input_107_begin_0"), val = tensor([0, 0, 0])]; + tensor input_107_end_0 = const()[name = tensor("input_107_end_0"), val = tensor([3, 256, 16])]; + tensor input_107_end_mask_0 = const()[name = tensor("input_107_end_mask_0"), val = tensor([true, true, false])]; + tensor input_107 = slice_by_index(begin = input_107_begin_0, end = input_107_end_0, end_mask = input_107_end_mask_0, x = outputs_aug_5)[name = tensor("input_107")]; + tensor inputs_27 = batch_norm(beta = inner_encoder_conv_module_2_sequential_5_bias, epsilon = var_56, gamma = inner_encoder_conv_module_2_sequential_5_weight, mean = inner_encoder_conv_module_2_sequential_5_running_mean, variance = inner_encoder_conv_module_2_sequential_5_running_var, x = input_107)[name = tensor("inputs_27")]; + tensor input_109 = silu(x = inputs_27)[name = tensor("input_109")]; + tensor input_111_pad_type_0 = const()[name = tensor("input_111_pad_type_0"), val = tensor("valid")]; + tensor input_111_strides_0 = const()[name = tensor("input_111_strides_0"), val = tensor([1])]; + tensor input_111_pad_0 = const()[name = tensor("input_111_pad_0"), val = tensor([0, 0])]; + tensor input_111_dilations_0 = const()[name = tensor("input_111_dilations_0"), val = tensor([1])]; + tensor input_111_groups_0 = const()[name = tensor("input_111_groups_0"), val = tensor(1)]; + tensor input_111 = conv(bias = inner_encoder_conv_module_2_sequential_7_conv_bias, dilations = input_111_dilations_0, groups = input_111_groups_0, pad = input_111_pad_0, pad_type = input_111_pad_type_0, strides = input_111_strides_0, weight = inner_encoder_conv_module_2_sequential_7_conv_weight, x = input_109)[name = tensor("input_111")]; tensor conv_out_5_perm_0 = const()[name = tensor("conv_out_5_perm_0"), val = tensor([0, 2, 1])]; - tensor var_710_begin_0 = const()[name = tensor("op_710_begin_0"), val = tensor([0, -1, 0])]; - tensor var_710_end_0 = const()[name = tensor("op_710_end_0"), val = tensor([3, 16, 256])]; - tensor var_710_end_mask_0 = const()[name = tensor("op_710_end_mask_0"), val = tensor([true, true, true])]; - tensor conv_out_5 = transpose(perm = conv_out_5_perm_0, x = input_109)[name = tensor("transpose_39")]; - tensor var_710 = slice_by_index(begin = var_710_begin_0, end = var_710_end_0, end_mask = var_710_end_mask_0, x = conv_out_5)[name = tensor("op_710")]; - tensor var_712_perm_0 = const()[name = tensor("op_712_perm_0"), val = tensor([1, 0, 2])]; - tensor var_712 = transpose(perm = var_712_perm_0, x = var_710)[name = tensor("transpose_38")]; - tensor input_111 = add(x = x_15, y = var_712)[name = tensor("input_111")]; - tensor input_113_axes_0 = const()[name = tensor("input_113_axes_0"), val = tensor([-1])]; - tensor input_113 = layer_norm(axes = input_113_axes_0, beta = encoder_ffn2_2_module_sequential_0_bias, epsilon = var_28, gamma = encoder_ffn2_2_module_sequential_0_weight, x = input_111)[name = tensor("input_113")]; - tensor inputs_29 = linear(bias = encoder_ffn2_2_module_sequential_1_linear_bias, weight = encoder_ffn2_2_module_sequential_1_linear_weight, x = input_113)[name = tensor("linear_26")]; - tensor input_115 = silu(x = inputs_29)[name = tensor("input_115")]; - tensor input_119 = linear(bias = encoder_ffn2_2_module_sequential_4_linear_bias, weight = encoder_ffn2_2_module_sequential_4_linear_weight, x = input_115)[name = tensor("linear_27")]; - tensor var_735 = const()[name = tensor("op_735"), val = tensor(0x1p-1)]; - tensor var_736 = mul(x = input_119, y = var_735)[name = tensor("op_736")]; - tensor input_121 = add(x = var_736, y = input_111)[name = tensor("input_121")]; - tensor input_123_axes_0 = const()[name = tensor("input_123_axes_0"), val = tensor([-1])]; - tensor input_123 = layer_norm(axes = input_123_axes_0, beta = encoder_layer_norm_2_bias, epsilon = var_28, gamma = encoder_layer_norm_2_weight, x = input_121)[name = tensor("input_123")]; + tensor var_756_begin_0 = const()[name = tensor("op_756_begin_0"), val = tensor([0, -1, 0])]; + tensor var_756_end_0 = const()[name = tensor("op_756_end_0"), val = tensor([3, 16, 256])]; + tensor var_756_end_mask_0 = const()[name = tensor("op_756_end_mask_0"), val = tensor([true, true, true])]; + tensor conv_out_5 = transpose(perm = conv_out_5_perm_0, x = input_111)[name = tensor("transpose_39")]; + tensor var_756 = slice_by_index(begin = var_756_begin_0, end = var_756_end_0, end_mask = var_756_end_mask_0, x = conv_out_5)[name = tensor("op_756")]; + tensor var_758_perm_0 = const()[name = tensor("op_758_perm_0"), val = tensor([1, 0, 2])]; + tensor var_758 = transpose(perm = var_758_perm_0, x = var_756)[name = tensor("transpose_38")]; + tensor input_113 = add(x = x_15, y = var_758)[name = tensor("input_113")]; + tensor input_115_axes_0 = const()[name = tensor("input_115_axes_0"), val = tensor([-1])]; + tensor input_115 = layer_norm(axes = input_115_axes_0, beta = inner_encoder_ffn2_2_module_sequential_0_bias, epsilon = var_56, gamma = inner_encoder_ffn2_2_module_sequential_0_weight, x = input_113)[name = tensor("input_115")]; + tensor inputs_29 = linear(bias = inner_encoder_ffn2_2_module_sequential_1_linear_bias, weight = inner_encoder_ffn2_2_module_sequential_1_linear_weight, x = input_115)[name = tensor("linear_26")]; + tensor input_117 = silu(x = inputs_29)[name = tensor("input_117")]; + tensor input_121 = linear(bias = inner_encoder_ffn2_2_module_sequential_4_linear_bias, weight = inner_encoder_ffn2_2_module_sequential_4_linear_weight, x = input_117)[name = tensor("linear_27")]; + tensor var_781 = const()[name = tensor("op_781"), val = tensor(0x1p-1)]; + tensor var_782 = mul(x = input_121, y = var_781)[name = tensor("op_782")]; + tensor input_123 = add(x = var_782, y = input_113)[name = tensor("input_123")]; tensor input_125_axes_0 = const()[name = tensor("input_125_axes_0"), val = tensor([-1])]; - tensor input_125 = layer_norm(axes = input_125_axes_0, beta = encoder_ffn1_3_module_sequential_0_bias, epsilon = var_28, gamma = encoder_ffn1_3_module_sequential_0_weight, x = input_123)[name = tensor("input_125")]; - tensor inputs_31 = linear(bias = encoder_ffn1_3_module_sequential_1_linear_bias, weight = encoder_ffn1_3_module_sequential_1_linear_weight, x = input_125)[name = tensor("linear_28")]; - tensor input_127 = silu(x = inputs_31)[name = tensor("input_127")]; - tensor input_131 = linear(bias = encoder_ffn1_3_module_sequential_4_linear_bias, weight = encoder_ffn1_3_module_sequential_4_linear_weight, x = input_127)[name = tensor("linear_29")]; - tensor var_765 = const()[name = tensor("op_765"), val = tensor(0x1p-1)]; - tensor var_766 = mul(x = input_131, y = var_765)[name = tensor("op_766")]; - tensor input_133 = add(x = var_766, y = input_123)[name = tensor("input_133")]; + tensor input_125 = layer_norm(axes = input_125_axes_0, beta = inner_encoder_layer_norm_2_bias, epsilon = var_56, gamma = inner_encoder_layer_norm_2_weight, x = input_123)[name = tensor("input_125")]; + tensor input_127_axes_0 = const()[name = tensor("input_127_axes_0"), val = tensor([-1])]; + tensor input_127 = layer_norm(axes = input_127_axes_0, beta = inner_encoder_ffn1_3_module_sequential_0_bias, epsilon = var_56, gamma = inner_encoder_ffn1_3_module_sequential_0_weight, x = input_125)[name = tensor("input_127")]; + tensor inputs_31 = linear(bias = inner_encoder_ffn1_3_module_sequential_1_linear_bias, weight = inner_encoder_ffn1_3_module_sequential_1_linear_weight, x = input_127)[name = tensor("linear_28")]; + tensor input_129 = silu(x = inputs_31)[name = tensor("input_129")]; + tensor input_133 = linear(bias = inner_encoder_ffn1_3_module_sequential_4_linear_bias, weight = inner_encoder_ffn1_3_module_sequential_4_linear_weight, x = input_129)[name = tensor("linear_29")]; + tensor var_811 = const()[name = tensor("op_811"), val = tensor(0x1p-1)]; + tensor var_812 = mul(x = input_133, y = var_811)[name = tensor("op_812")]; + tensor input_135 = add(x = var_812, y = input_125)[name = tensor("input_135")]; tensor x_19_axes_0 = const()[name = tensor("x_19_axes_0"), val = tensor([-1])]; - tensor x_19 = layer_norm(axes = x_19_axes_0, beta = encoder_ret_lns_3_bias, epsilon = var_28, gamma = encoder_ret_lns_3_weight, x = input_133)[name = tensor("x_19")]; + tensor x_19 = layer_norm(axes = x_19_axes_0, beta = inner_encoder_ret_lns_3_bias, epsilon = var_56, gamma = inner_encoder_ret_lns_3_weight, x = input_135)[name = tensor("x_19")]; tensor prev_kv_7_begin_0 = const()[name = tensor("prev_kv_7_begin_0"), val = tensor([3, 0, 0, 0, 0])]; tensor prev_kv_7_end_0 = const()[name = tensor("prev_kv_7_end_0"), val = tensor([4, 1, 4, 64, 64])]; tensor prev_kv_7_end_mask_0 = const()[name = tensor("prev_kv_7_end_mask_0"), val = tensor([false, true, true, true, true])]; @@ -740,199 +758,192 @@ program(1.0) tensor prev_scale_7_end_mask_0 = const()[name = tensor("prev_scale_7_end_mask_0"), val = tensor([false, true])]; tensor prev_scale_7_squeeze_mask_0 = const()[name = tensor("prev_scale_7_squeeze_mask_0"), val = tensor([true, false])]; tensor prev_scale_7 = slice_by_index(begin = prev_scale_7_begin_0, end = prev_scale_7_end_0, end_mask = prev_scale_7_end_mask_0, squeeze_mask = prev_scale_7_squeeze_mask_0, x = enc_scale)[name = tensor("prev_scale_7")]; - tensor var_780 = linear(bias = encoder_q_proj_3_bias, weight = encoder_q_proj_3_weight, x = x_19)[name = tensor("linear_30")]; - tensor var_781 = const()[name = tensor("op_781"), val = tensor([1, 3, 4, 64])]; - tensor var_782 = reshape(shape = var_781, x = var_780)[name = tensor("op_782")]; + tensor var_826 = linear(bias = inner_encoder_q_proj_3_bias, weight = inner_encoder_q_proj_3_weight, x = x_19)[name = tensor("linear_30")]; + tensor var_827 = const()[name = tensor("op_827"), val = tensor([1, 3, 4, 64])]; + tensor var_828 = reshape(shape = var_827, x = var_826)[name = tensor("op_828")]; tensor q_7_perm_0 = const()[name = tensor("q_7_perm_0"), val = tensor([0, 2, 1, 3])]; - tensor var_786 = linear(bias = encoder_k_proj_3_bias, weight = encoder_k_proj_3_weight, x = x_19)[name = tensor("linear_31")]; - tensor var_787 = const()[name = tensor("op_787"), val = tensor(0x1p-3)]; - tensor var_788 = mul(x = var_786, y = var_787)[name = tensor("op_788")]; - tensor var_789 = const()[name = tensor("op_789"), val = tensor([1, 3, 4, 64])]; - tensor var_790 = reshape(shape = var_789, x = var_788)[name = tensor("op_790")]; + tensor var_832 = linear(bias = inner_encoder_k_proj_3_bias, weight = inner_encoder_k_proj_3_weight, x = x_19)[name = tensor("linear_31")]; + tensor var_833 = const()[name = tensor("op_833"), val = tensor(0x1p-3)]; + tensor var_834 = mul(x = var_832, y = var_833)[name = tensor("op_834")]; + tensor var_835 = const()[name = tensor("op_835"), val = tensor([1, 3, 4, 64])]; + tensor var_836 = reshape(shape = var_835, x = var_834)[name = tensor("op_836")]; tensor k_7_perm_0 = const()[name = tensor("k_7_perm_0"), val = tensor([0, 2, 1, 3])]; - tensor var_794 = linear(bias = encoder_v_proj_3_bias, weight = encoder_v_proj_3_weight, x = x_19)[name = tensor("linear_32")]; - tensor var_795 = const()[name = tensor("op_795"), val = tensor([1, 3, 4, 64])]; - tensor var_796 = reshape(shape = var_795, x = var_794)[name = tensor("op_796")]; + tensor var_840 = linear(bias = inner_encoder_v_proj_3_bias, weight = inner_encoder_v_proj_3_weight, x = x_19)[name = tensor("linear_32")]; + tensor var_841 = const()[name = tensor("op_841"), val = tensor([1, 3, 4, 64])]; + tensor var_842 = reshape(shape = var_841, x = var_840)[name = tensor("op_842")]; tensor v_7_perm_0 = const()[name = tensor("v_7_perm_0"), val = tensor([0, 2, 1, 3])]; - tensor input_137 = linear(bias = encoder_g_proj_3_bias, weight = encoder_g_proj_3_weight, x = x_19)[name = tensor("linear_33")]; + tensor input_139 = linear(bias = inner_encoder_g_proj_3_bias, weight = inner_encoder_g_proj_3_weight, x = x_19)[name = tensor("linear_33")]; tensor sqrt_s0_7 = sqrt(x = prev_scale_7)[name = tensor("sqrt_s0_7")]; - tensor s_t_7 = add(x = prev_scale_7, y = encoder__t_index)[name = tensor("s_t_7")]; + tensor s_t_7 = add(x = prev_scale_7, y = inner_encoder__t_index)[name = tensor("s_t_7")]; tensor sqrt_s_t_7 = sqrt(x = s_t_7)[name = tensor("sqrt_s_t_7")]; tensor qk_7_transpose_x_1 = const()[name = tensor("qk_7_transpose_x_1"), val = tensor(false)]; tensor qk_7_transpose_y_1 = const()[name = tensor("qk_7_transpose_y_1"), val = tensor(true)]; - tensor k_7 = transpose(perm = k_7_perm_0, x = var_790)[name = tensor("transpose_36")]; - tensor q_7 = transpose(perm = q_7_perm_0, x = var_782)[name = tensor("transpose_37")]; + tensor k_7 = transpose(perm = k_7_perm_0, x = var_836)[name = tensor("transpose_36")]; + tensor q_7 = transpose(perm = q_7_perm_0, x = var_828)[name = tensor("transpose_37")]; tensor qk_7 = matmul(transpose_x = qk_7_transpose_x_1, transpose_y = qk_7_transpose_y_1, x = q_7, y = k_7)[name = tensor("qk_7")]; - tensor var_806 = const()[name = tensor("op_806"), val = tensor([3, 1])]; - tensor var_807 = reshape(shape = var_806, x = sqrt_s_t_7)[name = tensor("op_807")]; - tensor M_7 = real_div(x = encoder__causal_mask, y = var_807)[name = tensor("M_7")]; - tensor var_809 = mul(x = qk_7, y = M_7)[name = tensor("op_809")]; + tensor var_852 = const()[name = tensor("op_852"), val = tensor([3, 1])]; + tensor var_853 = reshape(shape = var_852, x = sqrt_s_t_7)[name = tensor("op_853")]; + tensor M_7 = real_div(x = inner_encoder__causal_mask, y = var_853)[name = tensor("M_7")]; + tensor var_855 = mul(x = qk_7, y = M_7)[name = tensor("op_855")]; tensor inner_7_transpose_x_0 = const()[name = tensor("inner_7_transpose_x_0"), val = tensor(false)]; tensor inner_7_transpose_y_0 = const()[name = tensor("inner_7_transpose_y_0"), val = tensor(false)]; - tensor v_7 = transpose(perm = v_7_perm_0, x = var_796)[name = tensor("transpose_35")]; - tensor inner_7 = matmul(transpose_x = inner_7_transpose_x_0, transpose_y = inner_7_transpose_y_0, x = var_809, y = v_7)[name = tensor("inner_7")]; - tensor var_811_transpose_x_0 = const()[name = tensor("op_811_transpose_x_0"), val = tensor(false)]; - tensor var_811_transpose_y_0 = const()[name = tensor("op_811_transpose_y_0"), val = tensor(false)]; - tensor var_811 = matmul(transpose_x = var_811_transpose_x_0, transpose_y = var_811_transpose_y_0, x = q_7, y = prev_kv_7)[name = tensor("op_811")]; - tensor var_812 = real_div(x = sqrt_s0_7, y = sqrt_s_t_7)[name = tensor("op_812")]; - tensor var_813 = const()[name = tensor("op_813"), val = tensor([1, 1, 3, 1])]; - tensor var_814 = reshape(shape = var_813, x = var_812)[name = tensor("op_814")]; - tensor cross_7 = mul(x = var_811, y = var_814)[name = tensor("cross_7")]; + tensor v_7 = transpose(perm = v_7_perm_0, x = var_842)[name = tensor("transpose_35")]; + tensor inner_7 = matmul(transpose_x = inner_7_transpose_x_0, transpose_y = inner_7_transpose_y_0, x = var_855, y = v_7)[name = tensor("inner_7")]; + tensor var_857_transpose_x_0 = const()[name = tensor("op_857_transpose_x_0"), val = tensor(false)]; + tensor var_857_transpose_y_0 = const()[name = tensor("op_857_transpose_y_0"), val = tensor(false)]; + tensor var_857 = matmul(transpose_x = var_857_transpose_x_0, transpose_y = var_857_transpose_y_0, x = q_7, y = prev_kv_7)[name = tensor("op_857")]; + tensor var_858 = real_div(x = sqrt_s0_7, y = sqrt_s_t_7)[name = tensor("op_858")]; + tensor var_859 = const()[name = tensor("op_859"), val = tensor([1, 1, 3, 1])]; + tensor var_860 = reshape(shape = var_859, x = var_858)[name = tensor("op_860")]; + tensor cross_7 = mul(x = var_857, y = var_860)[name = tensor("cross_7")]; tensor out_19 = add(x = inner_7, y = cross_7)[name = tensor("out_19")]; - tensor var_817 = mul(x = prev_kv_7, y = sqrt_s0_7)[name = tensor("op_817")]; - tensor var_819_transpose_x_1 = const()[name = tensor("op_819_transpose_x_1"), val = tensor(true)]; - tensor var_819_transpose_y_1 = const()[name = tensor("op_819_transpose_y_1"), val = tensor(false)]; - tensor var_819 = matmul(transpose_x = var_819_transpose_x_1, transpose_y = var_819_transpose_y_1, x = k_7, y = v_7)[name = tensor("op_819")]; - tensor new_kv_unnorm_7 = add(x = var_817, y = var_819)[name = tensor("new_kv_unnorm_7")]; - tensor var_821 = const()[name = tensor("op_821"), val = tensor(0x1.8p+1)]; - tensor new_scale_7 = add(x = prev_scale_7, y = var_821)[name = tensor("new_scale_7")]; - tensor var_823 = sqrt(x = new_scale_7)[name = tensor("op_823")]; - tensor nkv_1 = real_div(x = new_kv_unnorm_7, y = var_823)[name = tensor("nkv_1")]; - tensor var_825_perm_0 = const()[name = tensor("op_825_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_863 = mul(x = prev_kv_7, y = sqrt_s0_7)[name = tensor("op_863")]; + tensor var_865_transpose_x_1 = const()[name = tensor("op_865_transpose_x_1"), val = tensor(true)]; + tensor var_865_transpose_y_1 = const()[name = tensor("op_865_transpose_y_1"), val = tensor(false)]; + tensor var_865 = matmul(transpose_x = var_865_transpose_x_1, transpose_y = var_865_transpose_y_1, x = k_7, y = v_7)[name = tensor("op_865")]; + tensor new_kv_unnorm_7 = add(x = var_863, y = var_865)[name = tensor("new_kv_unnorm_7")]; + tensor var_867 = const()[name = tensor("op_867"), val = tensor(0x1.8p+1)]; + tensor new_scale_7 = add(x = prev_scale_7, y = var_867)[name = tensor("new_scale_7")]; + tensor var_869 = sqrt(x = new_scale_7)[name = tensor("op_869")]; + tensor nkv_1 = real_div(x = new_kv_unnorm_7, y = var_869)[name = tensor("nkv_1")]; + tensor var_871_perm_0 = const()[name = tensor("op_871_perm_0"), val = tensor([0, 2, 1, 3])]; tensor out_21_axes_0 = const()[name = tensor("out_21_axes_0"), val = tensor([-1])]; - tensor var_825 = transpose(perm = var_825_perm_0, x = out_19)[name = tensor("transpose_34")]; - tensor out_21 = layer_norm(axes = out_21_axes_0, epsilon = var_17, x = var_825)[name = tensor("out_21")]; - tensor var_829 = const()[name = tensor("op_829"), val = tensor([1, 3, 256])]; - tensor out_23 = reshape(shape = var_829, x = out_21)[name = tensor("out_23")]; - tensor var_831 = silu(x = input_137)[name = tensor("op_831")]; - tensor input_139 = mul(x = var_831, y = out_23)[name = tensor("input_139")]; - tensor ret_out_7 = linear(bias = encoder_out_proj_3_bias, weight = encoder_out_proj_3_weight, x = input_139)[name = tensor("linear_34")]; - tensor x_21 = add(x = input_133, y = ret_out_7)[name = tensor("x_21")]; + tensor var_871 = transpose(perm = var_871_perm_0, x = out_19)[name = tensor("transpose_34")]; + tensor out_21 = layer_norm(axes = out_21_axes_0, epsilon = var_64, x = var_871)[name = tensor("out_21")]; + tensor var_875 = const()[name = tensor("op_875"), val = tensor([1, 3, 256])]; + tensor out_23 = reshape(shape = var_875, x = out_21)[name = tensor("out_23")]; + tensor var_877 = silu(x = input_139)[name = tensor("op_877")]; + tensor input_141 = mul(x = var_877, y = out_23)[name = tensor("input_141")]; + tensor ret_out_7 = linear(bias = inner_encoder_out_proj_3_bias, weight = inner_encoder_out_proj_3_weight, x = input_141)[name = tensor("linear_34")]; + tensor x_21 = add(x = input_135, y = ret_out_7)[name = tensor("x_21")]; tensor window_25_begin_0 = const()[name = tensor("window_25_begin_0"), val = tensor([3, 0, 0, 0])]; tensor window_25_end_0 = const()[name = tensor("window_25_end_0"), val = tensor([4, 1, 16, 256])]; tensor window_25_end_mask_0 = const()[name = tensor("window_25_end_mask_0"), val = tensor([false, true, true, true])]; tensor window_25_squeeze_mask_0 = const()[name = tensor("window_25_squeeze_mask_0"), val = tensor([true, false, false, false])]; tensor window_25 = slice_by_index(begin = window_25_begin_0, end = window_25_end_0, end_mask = window_25_end_mask_0, squeeze_mask = window_25_squeeze_mask_0, x = enc_conv_cache)[name = tensor("window_25")]; - tensor var_839_begin_0 = const()[name = tensor("op_839_begin_0"), val = tensor([0, 0, 0])]; - tensor var_839_end_0 = const()[name = tensor("op_839_end_0"), val = tensor([1, 1, 256])]; - tensor var_839_end_mask_0 = const()[name = tensor("op_839_end_mask_0"), val = tensor([true, false, true])]; - tensor var_839 = slice_by_index(begin = var_839_begin_0, end = var_839_end_0, end_mask = var_839_end_mask_0, x = x_21)[name = tensor("op_839")]; - tensor var_842_begin_0 = const()[name = tensor("op_842_begin_0"), val = tensor([0, 1, 0])]; - tensor var_842_end_0 = const()[name = tensor("op_842_end_0"), val = tensor([1, 16, 256])]; - tensor var_842_end_mask_0 = const()[name = tensor("op_842_end_mask_0"), val = tensor([true, true, true])]; - tensor var_842 = slice_by_index(begin = var_842_begin_0, end = var_842_end_0, end_mask = var_842_end_mask_0, x = window_25)[name = tensor("op_842")]; + tensor var_885_begin_0 = const()[name = tensor("op_885_begin_0"), val = tensor([0, 0, 0])]; + tensor var_885_end_0 = const()[name = tensor("op_885_end_0"), val = tensor([1, 1, 256])]; + tensor var_885_end_mask_0 = const()[name = tensor("op_885_end_mask_0"), val = tensor([true, false, true])]; + tensor var_885 = slice_by_index(begin = var_885_begin_0, end = var_885_end_0, end_mask = var_885_end_mask_0, x = x_21)[name = tensor("op_885")]; + tensor var_888_begin_0 = const()[name = tensor("op_888_begin_0"), val = tensor([0, 1, 0])]; + tensor var_888_end_0 = const()[name = tensor("op_888_end_0"), val = tensor([1, 16, 256])]; + tensor var_888_end_mask_0 = const()[name = tensor("op_888_end_mask_0"), val = tensor([true, true, true])]; + tensor var_888 = slice_by_index(begin = var_888_begin_0, end = var_888_end_0, end_mask = var_888_end_mask_0, x = window_25)[name = tensor("op_888")]; tensor window_27_interleave_0 = const()[name = tensor("window_27_interleave_0"), val = tensor(false)]; - tensor window_27 = concat(axis = var_26, interleave = window_27_interleave_0, values = (var_842, var_839))[name = tensor("window_27")]; - tensor var_847_begin_0 = const()[name = tensor("op_847_begin_0"), val = tensor([0, 1, 0])]; - tensor var_847_end_0 = const()[name = tensor("op_847_end_0"), val = tensor([1, 2, 256])]; - tensor var_847_end_mask_0 = const()[name = tensor("op_847_end_mask_0"), val = tensor([true, false, true])]; - tensor var_847 = slice_by_index(begin = var_847_begin_0, end = var_847_end_0, end_mask = var_847_end_mask_0, x = x_21)[name = tensor("op_847")]; - tensor var_850_begin_0 = const()[name = tensor("op_850_begin_0"), val = tensor([0, 1, 0])]; - tensor var_850_end_0 = const()[name = tensor("op_850_end_0"), val = tensor([1, 16, 256])]; - tensor var_850_end_mask_0 = const()[name = tensor("op_850_end_mask_0"), val = tensor([true, true, true])]; - tensor var_850 = slice_by_index(begin = var_850_begin_0, end = var_850_end_0, end_mask = var_850_end_mask_0, x = window_27)[name = tensor("op_850")]; + tensor window_27 = concat(axis = var_72, interleave = window_27_interleave_0, values = (var_888, var_885))[name = tensor("window_27")]; + tensor var_893_begin_0 = const()[name = tensor("op_893_begin_0"), val = tensor([0, 1, 0])]; + tensor var_893_end_0 = const()[name = tensor("op_893_end_0"), val = tensor([1, 2, 256])]; + tensor var_893_end_mask_0 = const()[name = tensor("op_893_end_mask_0"), val = tensor([true, false, true])]; + tensor var_893 = slice_by_index(begin = var_893_begin_0, end = var_893_end_0, end_mask = var_893_end_mask_0, x = x_21)[name = tensor("op_893")]; + tensor var_896_begin_0 = const()[name = tensor("op_896_begin_0"), val = tensor([0, 1, 0])]; + tensor var_896_end_0 = const()[name = tensor("op_896_end_0"), val = tensor([1, 16, 256])]; + tensor var_896_end_mask_0 = const()[name = tensor("op_896_end_mask_0"), val = tensor([true, true, true])]; + tensor var_896 = slice_by_index(begin = var_896_begin_0, end = var_896_end_0, end_mask = var_896_end_mask_0, x = window_27)[name = tensor("op_896")]; tensor window_29_interleave_0 = const()[name = tensor("window_29_interleave_0"), val = tensor(false)]; - tensor window_29 = concat(axis = var_26, interleave = window_29_interleave_0, values = (var_850, var_847))[name = tensor("window_29")]; - tensor var_855_begin_0 = const()[name = tensor("op_855_begin_0"), val = tensor([0, 2, 0])]; - tensor var_855_end_0 = const()[name = tensor("op_855_end_0"), val = tensor([1, 1, 256])]; - tensor var_855_end_mask_0 = const()[name = tensor("op_855_end_mask_0"), val = tensor([true, true, true])]; - tensor var_855 = slice_by_index(begin = var_855_begin_0, end = var_855_end_0, end_mask = var_855_end_mask_0, x = x_21)[name = tensor("op_855")]; - tensor var_858_begin_0 = const()[name = tensor("op_858_begin_0"), val = tensor([0, 1, 0])]; - tensor var_858_end_0 = const()[name = tensor("op_858_end_0"), val = tensor([1, 16, 256])]; - tensor var_858_end_mask_0 = const()[name = tensor("op_858_end_mask_0"), val = tensor([true, true, true])]; - tensor var_858 = slice_by_index(begin = var_858_begin_0, end = var_858_end_0, end_mask = var_858_end_mask_0, x = window_29)[name = tensor("op_858")]; + tensor window_29 = concat(axis = var_72, interleave = window_29_interleave_0, values = (var_896, var_893))[name = tensor("window_29")]; + tensor var_901_begin_0 = const()[name = tensor("op_901_begin_0"), val = tensor([0, 2, 0])]; + tensor var_901_end_0 = const()[name = tensor("op_901_end_0"), val = tensor([1, 1, 256])]; + tensor var_901_end_mask_0 = const()[name = tensor("op_901_end_mask_0"), val = tensor([true, true, true])]; + tensor var_901 = slice_by_index(begin = var_901_begin_0, end = var_901_end_0, end_mask = var_901_end_mask_0, x = x_21)[name = tensor("op_901")]; + tensor var_904_begin_0 = const()[name = tensor("op_904_begin_0"), val = tensor([0, 1, 0])]; + tensor var_904_end_0 = const()[name = tensor("op_904_end_0"), val = tensor([1, 16, 256])]; + tensor var_904_end_mask_0 = const()[name = tensor("op_904_end_mask_0"), val = tensor([true, true, true])]; + tensor var_904 = slice_by_index(begin = var_904_begin_0, end = var_904_end_0, end_mask = var_904_end_mask_0, x = window_29)[name = tensor("op_904")]; tensor window_interleave_0 = const()[name = tensor("window_interleave_0"), val = tensor(false)]; - tensor window = concat(axis = var_26, interleave = window_interleave_0, values = (var_858, var_855))[name = tensor("window")]; - tensor input_141_interleave_0 = const()[name = tensor("input_141_interleave_0"), val = tensor(false)]; - tensor input_141 = concat(axis = var_23, interleave = input_141_interleave_0, values = (window_27, window_29, window))[name = tensor("input_141")]; + tensor window = concat(axis = var_72, interleave = window_interleave_0, values = (var_904, var_901))[name = tensor("window")]; + tensor input_143_interleave_0 = const()[name = tensor("input_143_interleave_0"), val = tensor(false)]; + tensor input_143 = concat(axis = var_59, interleave = input_143_interleave_0, values = (window_27, window_29, window))[name = tensor("input_143")]; tensor x_23_axes_0 = const()[name = tensor("x_23_axes_0"), val = tensor([-1])]; - tensor x_23 = layer_norm(axes = x_23_axes_0, beta = encoder_conv_module_3_sequential_0_bias, epsilon = var_28, gamma = encoder_conv_module_3_sequential_0_weight, x = input_141)[name = tensor("x_23")]; - tensor input_143_perm_0 = const()[name = tensor("input_143_perm_0"), val = tensor([0, 2, 1])]; + tensor x_23 = layer_norm(axes = x_23_axes_0, beta = inner_encoder_conv_module_3_sequential_0_bias, epsilon = var_56, gamma = inner_encoder_conv_module_3_sequential_0_weight, x = input_143)[name = tensor("x_23")]; + tensor input_145_perm_0 = const()[name = tensor("input_145_perm_0"), val = tensor([0, 2, 1])]; tensor inputs_33_pad_type_0 = const()[name = tensor("inputs_33_pad_type_0"), val = tensor("valid")]; tensor inputs_33_strides_0 = const()[name = tensor("inputs_33_strides_0"), val = tensor([1])]; tensor inputs_33_pad_0 = const()[name = tensor("inputs_33_pad_0"), val = tensor([0, 0])]; tensor inputs_33_dilations_0 = const()[name = tensor("inputs_33_dilations_0"), val = tensor([1])]; tensor inputs_33_groups_0 = const()[name = tensor("inputs_33_groups_0"), val = tensor(1)]; - tensor input_143 = transpose(perm = input_143_perm_0, x = x_23)[name = tensor("transpose_33")]; - tensor inputs_33 = conv(bias = encoder_conv_module_3_sequential_2_conv_bias, dilations = inputs_33_dilations_0, groups = inputs_33_groups_0, pad = inputs_33_pad_0, pad_type = inputs_33_pad_type_0, strides = inputs_33_strides_0, weight = encoder_conv_module_3_sequential_2_conv_weight, x = input_143)[name = tensor("inputs_33")]; - tensor var_883_split_sizes_0 = const()[name = tensor("op_883_split_sizes_0"), val = tensor([256, 256])]; - tensor var_883_axis_0 = const()[name = tensor("op_883_axis_0"), val = tensor(1)]; - tensor var_883_0, tensor var_883_1 = split(axis = var_883_axis_0, split_sizes = var_883_split_sizes_0, x = inputs_33)[name = tensor("op_883")]; - tensor var_885 = sigmoid(x = var_883_1)[name = tensor("op_885")]; - tensor inputs_35 = mul(x = var_883_0, y = var_885)[name = tensor("inputs_35")]; + tensor input_145 = transpose(perm = input_145_perm_0, x = x_23)[name = tensor("transpose_33")]; + tensor inputs_33 = conv(bias = inner_encoder_conv_module_3_sequential_2_conv_bias, dilations = inputs_33_dilations_0, groups = inputs_33_groups_0, pad = inputs_33_pad_0, pad_type = inputs_33_pad_type_0, strides = inputs_33_strides_0, weight = inner_encoder_conv_module_3_sequential_2_conv_weight, x = input_145)[name = tensor("inputs_33")]; + tensor var_929_split_sizes_0 = const()[name = tensor("op_929_split_sizes_0"), val = tensor([256, 256])]; + tensor var_929_axis_0 = const()[name = tensor("op_929_axis_0"), val = tensor(1)]; + tensor var_929_0, tensor var_929_1 = split(axis = var_929_axis_0, split_sizes = var_929_split_sizes_0, x = inputs_33)[name = tensor("op_929")]; + tensor var_931 = sigmoid(x = var_929_1)[name = tensor("op_931")]; + tensor inputs_35 = mul(x = var_929_0, y = var_931)[name = tensor("inputs_35")]; tensor outputs_aug_pad_type_0 = const()[name = tensor("outputs_aug_pad_type_0"), val = tensor("custom")]; tensor outputs_aug_pad_0 = const()[name = tensor("outputs_aug_pad_0"), val = tensor([15, 15])]; tensor outputs_aug_groups_0 = const()[name = tensor("outputs_aug_groups_0"), val = tensor(256)]; tensor outputs_aug_strides_0 = const()[name = tensor("outputs_aug_strides_0"), val = tensor([1])]; tensor outputs_aug_dilations_0 = const()[name = tensor("outputs_aug_dilations_0"), val = tensor([1])]; - tensor outputs_aug = conv(dilations = outputs_aug_dilations_0, groups = outputs_aug_groups_0, pad = outputs_aug_pad_0, pad_type = outputs_aug_pad_type_0, strides = outputs_aug_strides_0, weight = encoder_conv_module_3_sequential_4_conv_weight, x = inputs_35)[name = tensor("outputs_aug")]; - tensor input_145_begin_0 = const()[name = tensor("input_145_begin_0"), val = tensor([0, 0, 0])]; - tensor input_145_end_0 = const()[name = tensor("input_145_end_0"), val = tensor([3, 256, 16])]; - tensor input_145_end_mask_0 = const()[name = tensor("input_145_end_mask_0"), val = tensor([true, true, false])]; - tensor input_145 = slice_by_index(begin = input_145_begin_0, end = input_145_end_0, end_mask = input_145_end_mask_0, x = outputs_aug)[name = tensor("input_145")]; - tensor inputs_37 = batch_norm(beta = encoder_conv_module_3_sequential_5_bias, epsilon = var_28, gamma = encoder_conv_module_3_sequential_5_weight, mean = encoder_conv_module_3_sequential_5_running_mean, variance = encoder_conv_module_3_sequential_5_running_var, x = input_145)[name = tensor("inputs_37")]; - tensor input_147 = silu(x = inputs_37)[name = tensor("input_147")]; - tensor input_149_pad_type_0 = const()[name = tensor("input_149_pad_type_0"), val = tensor("valid")]; - tensor input_149_strides_0 = const()[name = tensor("input_149_strides_0"), val = tensor([1])]; - tensor input_149_pad_0 = const()[name = tensor("input_149_pad_0"), val = tensor([0, 0])]; - tensor input_149_dilations_0 = const()[name = tensor("input_149_dilations_0"), val = tensor([1])]; - tensor input_149_groups_0 = const()[name = tensor("input_149_groups_0"), val = tensor(1)]; - tensor input_149 = conv(bias = encoder_conv_module_3_sequential_7_conv_bias, dilations = input_149_dilations_0, groups = input_149_groups_0, pad = input_149_pad_0, pad_type = input_149_pad_type_0, strides = input_149_strides_0, weight = encoder_conv_module_3_sequential_7_conv_weight, x = input_147)[name = tensor("input_149")]; + tensor outputs_aug = conv(dilations = outputs_aug_dilations_0, groups = outputs_aug_groups_0, pad = outputs_aug_pad_0, pad_type = outputs_aug_pad_type_0, strides = outputs_aug_strides_0, weight = inner_encoder_conv_module_3_sequential_4_conv_weight, x = inputs_35)[name = tensor("outputs_aug")]; + tensor input_147_begin_0 = const()[name = tensor("input_147_begin_0"), val = tensor([0, 0, 0])]; + tensor input_147_end_0 = const()[name = tensor("input_147_end_0"), val = tensor([3, 256, 16])]; + tensor input_147_end_mask_0 = const()[name = tensor("input_147_end_mask_0"), val = tensor([true, true, false])]; + tensor input_147 = slice_by_index(begin = input_147_begin_0, end = input_147_end_0, end_mask = input_147_end_mask_0, x = outputs_aug)[name = tensor("input_147")]; + tensor inputs_37 = batch_norm(beta = inner_encoder_conv_module_3_sequential_5_bias, epsilon = var_56, gamma = inner_encoder_conv_module_3_sequential_5_weight, mean = inner_encoder_conv_module_3_sequential_5_running_mean, variance = inner_encoder_conv_module_3_sequential_5_running_var, x = input_147)[name = tensor("inputs_37")]; + tensor input_149 = silu(x = inputs_37)[name = tensor("input_149")]; + tensor input_151_pad_type_0 = const()[name = tensor("input_151_pad_type_0"), val = tensor("valid")]; + tensor input_151_strides_0 = const()[name = tensor("input_151_strides_0"), val = tensor([1])]; + tensor input_151_pad_0 = const()[name = tensor("input_151_pad_0"), val = tensor([0, 0])]; + tensor input_151_dilations_0 = const()[name = tensor("input_151_dilations_0"), val = tensor([1])]; + tensor input_151_groups_0 = const()[name = tensor("input_151_groups_0"), val = tensor(1)]; + tensor input_151 = conv(bias = inner_encoder_conv_module_3_sequential_7_conv_bias, dilations = input_151_dilations_0, groups = input_151_groups_0, pad = input_151_pad_0, pad_type = input_151_pad_type_0, strides = input_151_strides_0, weight = inner_encoder_conv_module_3_sequential_7_conv_weight, x = input_149)[name = tensor("input_151")]; tensor conv_out_7_perm_0 = const()[name = tensor("conv_out_7_perm_0"), val = tensor([0, 2, 1])]; - tensor var_916_begin_0 = const()[name = tensor("op_916_begin_0"), val = tensor([0, -1, 0])]; - tensor var_916_end_0 = const()[name = tensor("op_916_end_0"), val = tensor([3, 16, 256])]; - tensor var_916_end_mask_0 = const()[name = tensor("op_916_end_mask_0"), val = tensor([true, true, true])]; - tensor conv_out_7 = transpose(perm = conv_out_7_perm_0, x = input_149)[name = tensor("transpose_32")]; - tensor var_916 = slice_by_index(begin = var_916_begin_0, end = var_916_end_0, end_mask = var_916_end_mask_0, x = conv_out_7)[name = tensor("op_916")]; - tensor var_918_perm_0 = const()[name = tensor("op_918_perm_0"), val = tensor([1, 0, 2])]; - tensor var_918 = transpose(perm = var_918_perm_0, x = var_916)[name = tensor("transpose_31")]; - tensor input_151 = add(x = x_21, y = var_918)[name = tensor("input_151")]; - tensor input_153_axes_0 = const()[name = tensor("input_153_axes_0"), val = tensor([-1])]; - tensor input_153 = layer_norm(axes = input_153_axes_0, beta = encoder_ffn2_3_module_sequential_0_bias, epsilon = var_28, gamma = encoder_ffn2_3_module_sequential_0_weight, x = input_151)[name = tensor("input_153")]; - tensor inputs = linear(bias = encoder_ffn2_3_module_sequential_1_linear_bias, weight = encoder_ffn2_3_module_sequential_1_linear_weight, x = input_153)[name = tensor("linear_35")]; - tensor input_155 = silu(x = inputs)[name = tensor("input_155")]; - tensor input_159 = linear(bias = encoder_ffn2_3_module_sequential_4_linear_bias, weight = encoder_ffn2_3_module_sequential_4_linear_weight, x = input_155)[name = tensor("linear_36")]; - tensor var_941 = const()[name = tensor("op_941"), val = tensor(0x1p-1)]; - tensor var_942 = mul(x = input_159, y = var_941)[name = tensor("op_942")]; - tensor input_161 = add(x = var_942, y = input_151)[name = tensor("input_161")]; + tensor var_962_begin_0 = const()[name = tensor("op_962_begin_0"), val = tensor([0, -1, 0])]; + tensor var_962_end_0 = const()[name = tensor("op_962_end_0"), val = tensor([3, 16, 256])]; + tensor var_962_end_mask_0 = const()[name = tensor("op_962_end_mask_0"), val = tensor([true, true, true])]; + tensor conv_out_7 = transpose(perm = conv_out_7_perm_0, x = input_151)[name = tensor("transpose_32")]; + tensor var_962 = slice_by_index(begin = var_962_begin_0, end = var_962_end_0, end_mask = var_962_end_mask_0, x = conv_out_7)[name = tensor("op_962")]; + tensor var_964_perm_0 = const()[name = tensor("op_964_perm_0"), val = tensor([1, 0, 2])]; + tensor var_964 = transpose(perm = var_964_perm_0, x = var_962)[name = tensor("transpose_31")]; + tensor input_153 = add(x = x_21, y = var_964)[name = tensor("input_153")]; + tensor input_155_axes_0 = const()[name = tensor("input_155_axes_0"), val = tensor([-1])]; + tensor input_155 = layer_norm(axes = input_155_axes_0, beta = inner_encoder_ffn2_3_module_sequential_0_bias, epsilon = var_56, gamma = inner_encoder_ffn2_3_module_sequential_0_weight, x = input_153)[name = tensor("input_155")]; + tensor inputs = linear(bias = inner_encoder_ffn2_3_module_sequential_1_linear_bias, weight = inner_encoder_ffn2_3_module_sequential_1_linear_weight, x = input_155)[name = tensor("linear_35")]; + tensor input_157 = silu(x = inputs)[name = tensor("input_157")]; + tensor input_161 = linear(bias = inner_encoder_ffn2_3_module_sequential_4_linear_bias, weight = inner_encoder_ffn2_3_module_sequential_4_linear_weight, x = input_157)[name = tensor("linear_36")]; + tensor var_987 = const()[name = tensor("op_987"), val = tensor(0x1p-1)]; + tensor var_988 = mul(x = input_161, y = var_987)[name = tensor("op_988")]; + tensor input_163 = add(x = var_988, y = input_153)[name = tensor("input_163")]; tensor x_25_axes_0 = const()[name = tensor("x_25_axes_0"), val = tensor([-1])]; - tensor x_25 = layer_norm(axes = x_25_axes_0, beta = encoder_layer_norm_3_bias, epsilon = var_28, gamma = encoder_layer_norm_3_weight, x = input_161)[name = tensor("x_25")]; + tensor x_25 = layer_norm(axes = x_25_axes_0, beta = inner_encoder_layer_norm_3_bias, epsilon = var_56, gamma = inner_encoder_layer_norm_3_weight, x = input_163)[name = tensor("x_25")]; tensor x_ct_perm_0 = const()[name = tensor("x_ct_perm_0"), val = tensor([0, 2, 1])]; tensor cat_interleave_0 = const()[name = tensor("cat_interleave_0"), val = tensor(false)]; tensor x_ct = transpose(perm = x_ct_perm_0, x = x_25)[name = tensor("transpose_30")]; - tensor cat = concat(axis = var_20, interleave = cat_interleave_0, values = (cnn_window, x_ct))[name = tensor("cat")]; + tensor cat = concat(axis = var_61, interleave = cat_interleave_0, values = (cnn_window, x_ct))[name = tensor("cat")]; tensor conv_out_pad_type_0 = const()[name = tensor("conv_out_pad_type_0"), val = tensor("valid")]; tensor conv_out_strides_0 = const()[name = tensor("conv_out_strides_0"), val = tensor([1])]; tensor conv_out_pad_0 = const()[name = tensor("conv_out_pad_0"), val = tensor([0, 0])]; tensor conv_out_dilations_0 = const()[name = tensor("conv_out_dilations_0"), val = tensor([1])]; tensor conv_out_groups_0 = const()[name = tensor("conv_out_groups_0"), val = tensor(1)]; - tensor conv_out = conv(bias = encoder_cnn_bias, dilations = conv_out_dilations_0, groups = conv_out_groups_0, pad = conv_out_pad_0, pad_type = conv_out_pad_type_0, strides = conv_out_strides_0, weight = encoder_cnn_weight, x = cat)[name = tensor("conv_out")]; - tensor var_960_begin_0 = const()[name = tensor("op_960_begin_0"), val = tensor([0, 0, 3])]; - tensor var_960_end_0 = const()[name = tensor("op_960_end_0"), val = tensor([1, 256, 21])]; - tensor var_960_end_mask_0 = const()[name = tensor("op_960_end_mask_0"), val = tensor([true, true, true])]; - tensor cnn_window_new = slice_by_index(begin = var_960_begin_0, end = var_960_end_0, end_mask = var_960_end_mask_0, x = cat)[name = tensor("op_960")]; - tensor input_163_perm_0 = const()[name = tensor("input_163_perm_0"), val = tensor([0, 2, 1])]; - tensor var_962 = const()[name = tensor("op_962"), val = tensor([-1])]; - tensor input_163 = transpose(perm = input_163_perm_0, x = conv_out)[name = tensor("transpose_29")]; - tensor var_963 = reduce_l2_norm(axes = var_962, keep_dims = var_29, x = input_163)[name = tensor("op_963")]; + tensor conv_out = conv(bias = inner_encoder_cnn_bias, dilations = conv_out_dilations_0, groups = conv_out_groups_0, pad = conv_out_pad_0, pad_type = conv_out_pad_type_0, strides = conv_out_strides_0, weight = inner_encoder_cnn_weight, x = cat)[name = tensor("conv_out")]; + tensor var_1006_begin_0 = const()[name = tensor("op_1006_begin_0"), val = tensor([0, 0, 3])]; + tensor var_1006_end_0 = const()[name = tensor("op_1006_end_0"), val = tensor([1, 256, 21])]; + tensor var_1006_end_mask_0 = const()[name = tensor("op_1006_end_mask_0"), val = tensor([true, true, true])]; + tensor cnn_window_new = slice_by_index(begin = var_1006_begin_0, end = var_1006_end_0, end_mask = var_1006_end_mask_0, x = cat)[name = tensor("op_1006")]; + tensor input_165_perm_0 = const()[name = tensor("input_165_perm_0"), val = tensor([0, 2, 1])]; + tensor var_1008 = const()[name = tensor("op_1008"), val = tensor([-1])]; + tensor input_165 = transpose(perm = input_165_perm_0, x = conv_out)[name = tensor("transpose_29")]; + tensor var_1009 = reduce_l2_norm(axes = var_1008, keep_dims = var_55, x = input_165)[name = tensor("op_1009")]; tensor const_12 = const()[name = tensor("const_12"), val = tensor(0x1.fffffep+127)]; - tensor clip_0 = clip(alpha = var_11, beta = const_12, x = var_963)[name = tensor("clip_0")]; - tensor emb = real_div(x = input_163, y = clip_0)[name = tensor("emb")]; - tensor var_967_axis_0 = const()[name = tensor("op_967_axis_0"), val = tensor(0)]; - tensor enc_kv_new = stack(axis = var_967_axis_0, values = (var_206, var_412, var_618, nkv_1))[name = tensor("op_967")]; - tensor var_969_axis_0 = const()[name = tensor("op_969_axis_0"), val = tensor(0)]; - tensor enc_scale_new = stack(axis = var_969_axis_0, values = (new_scale_1, new_scale_3, new_scale_5, new_scale_7))[name = tensor("op_969")]; - tensor var_971_axis_0 = const()[name = tensor("op_971_axis_0"), val = tensor(0)]; - tensor enc_conv_cache_new = stack(axis = var_971_axis_0, values = (window_7, window_15, window_23, window))[name = tensor("op_971")]; - tensor var_980 = const()[name = tensor("op_980"), val = tensor(0x1.5798eep-27)]; - tensor var_985 = const()[name = tensor("op_985"), val = tensor(0x1.4f8b58p-17)]; - tensor var_987 = const()[name = tensor("op_987"), val = tensor(0x1.0c6f7ap-20)]; - tensor var_988 = const()[name = tensor("op_988"), val = tensor(true)]; - tensor var_990 = const()[name = tensor("op_990"), val = tensor(0x1p+0)]; - tensor var_994 = const()[name = tensor("op_994"), val = tensor(-1)]; - tensor var_1000 = const()[name = tensor("op_1000"), val = tensor(0)]; + tensor clip_0 = clip(alpha = var_69, beta = const_12, x = var_1009)[name = tensor("clip_0")]; + tensor emb = real_div(x = input_165, y = clip_0)[name = tensor("emb")]; + tensor var_1013_axis_0 = const()[name = tensor("op_1013_axis_0"), val = tensor(0)]; + tensor enc_kv_new = stack(axis = var_1013_axis_0, values = (var_252, var_458, var_664, nkv_1))[name = tensor("op_1013")]; + tensor var_1015_axis_0 = const()[name = tensor("op_1015_axis_0"), val = tensor(0)]; + tensor enc_scale_new = stack(axis = var_1015_axis_0, values = (new_scale_1, new_scale_3, new_scale_5, new_scale_7))[name = tensor("op_1015")]; + tensor var_1017_axis_0 = const()[name = tensor("op_1017_axis_0"), val = tensor(0)]; + tensor enc_conv_cache_new = stack(axis = var_1017_axis_0, values = (window_7, window_15, window_23, window))[name = tensor("op_1017")]; tensor pos = const()[name = tensor("pos"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(44395328)))]; - tensor var_1062_axes_0 = const()[name = tensor("op_1062_axes_0"), val = tensor([2])]; - tensor var_1062 = expand_dims(axes = var_1062_axes_0, x = emb)[name = tensor("op_1062")]; + tensor var_1085_axes_0 = const()[name = tensor("op_1085_axes_0"), val = tensor([2])]; + tensor var_1085 = expand_dims(axes = var_1085_axes_0, x = emb)[name = tensor("op_1085")]; tensor emb_exp_reps_0 = const()[name = tensor("emb_exp_reps_0"), val = tensor([1, 1, 12, 1])]; - tensor emb_exp = tile(reps = emb_exp_reps_0, x = var_1062)[name = tensor("emb_exp")]; - tensor input_165_interleave_0 = const()[name = tensor("input_165_interleave_0"), val = tensor(false)]; - tensor input_165 = concat(axis = var_994, interleave = input_165_interleave_0, values = (emb_exp, pos))[name = tensor("input_165")]; - tensor x_27 = linear(bias = decoder_convert_bias, weight = decoder_convert_weight, x = input_165)[name = tensor("linear_37")]; - tensor var_1070_perm_0 = const()[name = tensor("op_1070_perm_0"), val = tensor([0, 2, 1, 3])]; - tensor var_1074 = const()[name = tensor("op_1074"), val = tensor([12, 3, 256])]; - tensor var_1070 = transpose(perm = var_1070_perm_0, x = x_27)[name = tensor("transpose_28")]; - tensor x_29 = reshape(shape = var_1074, x = var_1070)[name = tensor("x_29")]; + tensor emb_exp = tile(reps = emb_exp_reps_0, x = var_1085)[name = tensor("emb_exp")]; + tensor input_167_interleave_0 = const()[name = tensor("input_167_interleave_0"), val = tensor(false)]; + tensor input_167 = concat(axis = var_62, interleave = input_167_interleave_0, values = (emb_exp, pos))[name = tensor("input_167")]; + tensor x_27 = linear(bias = inner_decoder_convert_bias, weight = inner_decoder_convert_weight, x = input_167)[name = tensor("linear_37")]; + tensor var_1093_perm_0 = const()[name = tensor("op_1093_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_1097 = const()[name = tensor("op_1097"), val = tensor([12, 3, 256])]; + tensor var_1093 = transpose(perm = var_1093_perm_0, x = x_27)[name = tensor("transpose_28")]; + tensor x_29 = reshape(shape = var_1097, x = var_1093)[name = tensor("x_29")]; tensor prev_kv_9_begin_0 = const()[name = tensor("prev_kv_9_begin_0"), val = tensor([0, 0, 0, 0, 0])]; tensor prev_kv_9_end_0 = const()[name = tensor("prev_kv_9_end_0"), val = tensor([1, 12, 4, 64, 64])]; tensor prev_kv_9_end_mask_0 = const()[name = tensor("prev_kv_9_end_mask_0"), val = tensor([false, true, true, true, true])]; @@ -943,132 +954,132 @@ program(1.0) tensor prev_scale_9_end_mask_0 = const()[name = tensor("prev_scale_9_end_mask_0"), val = tensor([false, true])]; tensor prev_scale_9_squeeze_mask_0 = const()[name = tensor("prev_scale_9_squeeze_mask_0"), val = tensor([true, false])]; tensor prev_scale_9 = slice_by_index(begin = prev_scale_9_begin_0, end = prev_scale_9_end_0, end_mask = prev_scale_9_end_mask_0, squeeze_mask = prev_scale_9_squeeze_mask_0, x = dec_scale)[name = tensor("prev_scale_9")]; - tensor var_1082 = linear(bias = decoder_q_proj_0_bias, weight = decoder_q_proj_0_weight, x = x_29)[name = tensor("linear_38")]; - tensor var_1083 = const()[name = tensor("op_1083"), val = tensor([12, 3, 4, 64])]; - tensor var_1084 = reshape(shape = var_1083, x = var_1082)[name = tensor("op_1084")]; + tensor var_1105 = linear(bias = inner_decoder_q_proj_0_bias, weight = inner_decoder_q_proj_0_weight, x = x_29)[name = tensor("linear_38")]; + tensor var_1106 = const()[name = tensor("op_1106"), val = tensor([12, 3, 4, 64])]; + tensor var_1107 = reshape(shape = var_1106, x = var_1105)[name = tensor("op_1107")]; tensor q_9_perm_0 = const()[name = tensor("q_9_perm_0"), val = tensor([0, 2, 1, 3])]; - tensor var_1088 = linear(bias = decoder_k_proj_0_bias, weight = decoder_k_proj_0_weight, x = x_29)[name = tensor("linear_39")]; - tensor var_1089 = const()[name = tensor("op_1089"), val = tensor(0x1p-3)]; - tensor var_1090 = mul(x = var_1088, y = var_1089)[name = tensor("op_1090")]; - tensor var_1091 = const()[name = tensor("op_1091"), val = tensor([12, 3, 4, 64])]; - tensor var_1092 = reshape(shape = var_1091, x = var_1090)[name = tensor("op_1092")]; + tensor var_1111 = linear(bias = inner_decoder_k_proj_0_bias, weight = inner_decoder_k_proj_0_weight, x = x_29)[name = tensor("linear_39")]; + tensor var_1112 = const()[name = tensor("op_1112"), val = tensor(0x1p-3)]; + tensor var_1113 = mul(x = var_1111, y = var_1112)[name = tensor("op_1113")]; + tensor var_1114 = const()[name = tensor("op_1114"), val = tensor([12, 3, 4, 64])]; + tensor var_1115 = reshape(shape = var_1114, x = var_1113)[name = tensor("op_1115")]; tensor k_9_perm_0 = const()[name = tensor("k_9_perm_0"), val = tensor([0, 2, 1, 3])]; - tensor var_1096 = linear(bias = decoder_v_proj_0_bias, weight = decoder_v_proj_0_weight, x = x_29)[name = tensor("linear_40")]; - tensor var_1097 = const()[name = tensor("op_1097"), val = tensor([12, 3, 4, 64])]; - tensor var_1098 = reshape(shape = var_1097, x = var_1096)[name = tensor("op_1098")]; + tensor var_1119 = linear(bias = inner_decoder_v_proj_0_bias, weight = inner_decoder_v_proj_0_weight, x = x_29)[name = tensor("linear_40")]; + tensor var_1120 = const()[name = tensor("op_1120"), val = tensor([12, 3, 4, 64])]; + tensor var_1121 = reshape(shape = var_1120, x = var_1119)[name = tensor("op_1121")]; tensor v_9_perm_0 = const()[name = tensor("v_9_perm_0"), val = tensor([0, 2, 1, 3])]; - tensor input_169 = linear(bias = decoder_g_proj_0_bias, weight = decoder_g_proj_0_weight, x = x_29)[name = tensor("linear_41")]; + tensor input_171 = linear(bias = inner_decoder_g_proj_0_bias, weight = inner_decoder_g_proj_0_weight, x = x_29)[name = tensor("linear_41")]; tensor cumsum_mask_1_exclusive_0 = const()[name = tensor("cumsum_mask_1_exclusive_0"), val = tensor(false)]; tensor cumsum_mask_1_reverse_0 = const()[name = tensor("cumsum_mask_1_reverse_0"), val = tensor(false)]; - tensor cumsum_mask_1 = cumsum(axis = var_1000, exclusive = cumsum_mask_1_exclusive_0, reverse = cumsum_mask_1_reverse_0, x = valid_mask)[name = tensor("cumsum_mask_1")]; + tensor cumsum_mask_1 = cumsum(axis = var_59, exclusive = cumsum_mask_1_exclusive_0, reverse = cumsum_mask_1_reverse_0, x = valid_mask)[name = tensor("cumsum_mask_1")]; tensor sqrt_s0_9 = sqrt(x = prev_scale_9)[name = tensor("sqrt_s0_9")]; tensor s_t_9 = add(x = prev_scale_9, y = cumsum_mask_1)[name = tensor("s_t_9")]; tensor const_20 = const()[name = tensor("const_20"), val = tensor(0x1.fffffep+127)]; - tensor clip_1 = clip(alpha = var_990, beta = const_20, x = s_t_9)[name = tensor("clip_1")]; + tensor clip_1 = clip(alpha = var_49, beta = const_20, x = s_t_9)[name = tensor("clip_1")]; tensor sqrt_s_t_9 = sqrt(x = clip_1)[name = tensor("sqrt_s_t_9")]; tensor qk_9_transpose_x_1 = const()[name = tensor("qk_9_transpose_x_1"), val = tensor(false)]; tensor qk_9_transpose_y_1 = const()[name = tensor("qk_9_transpose_y_1"), val = tensor(true)]; - tensor k_9 = transpose(perm = k_9_perm_0, x = var_1092)[name = tensor("transpose_26")]; - tensor q_9 = transpose(perm = q_9_perm_0, x = var_1084)[name = tensor("transpose_27")]; + tensor k_9 = transpose(perm = k_9_perm_0, x = var_1115)[name = tensor("transpose_26")]; + tensor q_9 = transpose(perm = q_9_perm_0, x = var_1107)[name = tensor("transpose_27")]; tensor qk_9 = matmul(transpose_x = qk_9_transpose_x_1, transpose_y = qk_9_transpose_y_1, x = q_9, y = k_9)[name = tensor("qk_9")]; - tensor var_1110 = const()[name = tensor("op_1110"), val = tensor([1, 3])]; - tensor var_1111 = reshape(shape = var_1110, x = valid_mask)[name = tensor("op_1111")]; - tensor causal_with_valid_1 = mul(x = decoder__causal_mask, y = var_1111)[name = tensor("causal_with_valid_1")]; - tensor var_1113 = const()[name = tensor("op_1113"), val = tensor([3, 1])]; - tensor var_1114 = reshape(shape = var_1113, x = sqrt_s_t_9)[name = tensor("op_1114")]; - tensor M_9 = real_div(x = causal_with_valid_1, y = var_1114)[name = tensor("M_9")]; - tensor var_1116 = mul(x = qk_9, y = M_9)[name = tensor("op_1116")]; + tensor var_1133 = const()[name = tensor("op_1133"), val = tensor([1, 3])]; + tensor var_1134 = reshape(shape = var_1133, x = valid_mask)[name = tensor("op_1134")]; + tensor causal_with_valid_1 = mul(x = inner_decoder__causal_mask, y = var_1134)[name = tensor("causal_with_valid_1")]; + tensor var_1136 = const()[name = tensor("op_1136"), val = tensor([3, 1])]; + tensor var_1137 = reshape(shape = var_1136, x = sqrt_s_t_9)[name = tensor("op_1137")]; + tensor M_9 = real_div(x = causal_with_valid_1, y = var_1137)[name = tensor("M_9")]; + tensor var_1139 = mul(x = qk_9, y = M_9)[name = tensor("op_1139")]; tensor inner_9_transpose_x_0 = const()[name = tensor("inner_9_transpose_x_0"), val = tensor(false)]; tensor inner_9_transpose_y_0 = const()[name = tensor("inner_9_transpose_y_0"), val = tensor(false)]; - tensor v_9 = transpose(perm = v_9_perm_0, x = var_1098)[name = tensor("transpose_25")]; - tensor inner_9 = matmul(transpose_x = inner_9_transpose_x_0, transpose_y = inner_9_transpose_y_0, x = var_1116, y = v_9)[name = tensor("inner_9")]; - tensor var_1118_transpose_x_0 = const()[name = tensor("op_1118_transpose_x_0"), val = tensor(false)]; - tensor var_1118_transpose_y_0 = const()[name = tensor("op_1118_transpose_y_0"), val = tensor(false)]; - tensor var_1118 = matmul(transpose_x = var_1118_transpose_x_0, transpose_y = var_1118_transpose_y_0, x = q_9, y = prev_kv_9)[name = tensor("op_1118")]; - tensor var_1119 = real_div(x = sqrt_s0_9, y = sqrt_s_t_9)[name = tensor("op_1119")]; - tensor var_1120 = const()[name = tensor("op_1120"), val = tensor([1, 1, 3, 1])]; - tensor var_1121 = reshape(shape = var_1120, x = var_1119)[name = tensor("op_1121")]; - tensor cross_9 = mul(x = var_1118, y = var_1121)[name = tensor("cross_9")]; + tensor v_9 = transpose(perm = v_9_perm_0, x = var_1121)[name = tensor("transpose_25")]; + tensor inner_9 = matmul(transpose_x = inner_9_transpose_x_0, transpose_y = inner_9_transpose_y_0, x = var_1139, y = v_9)[name = tensor("inner_9")]; + tensor var_1141_transpose_x_0 = const()[name = tensor("op_1141_transpose_x_0"), val = tensor(false)]; + tensor var_1141_transpose_y_0 = const()[name = tensor("op_1141_transpose_y_0"), val = tensor(false)]; + tensor var_1141 = matmul(transpose_x = var_1141_transpose_x_0, transpose_y = var_1141_transpose_y_0, x = q_9, y = prev_kv_9)[name = tensor("op_1141")]; + tensor var_1142 = real_div(x = sqrt_s0_9, y = sqrt_s_t_9)[name = tensor("op_1142")]; + tensor var_1143 = const()[name = tensor("op_1143"), val = tensor([1, 1, 3, 1])]; + tensor var_1144 = reshape(shape = var_1143, x = var_1142)[name = tensor("op_1144")]; + tensor cross_9 = mul(x = var_1141, y = var_1144)[name = tensor("cross_9")]; tensor out_25 = add(x = inner_9, y = cross_9)[name = tensor("out_25")]; - tensor var_1124 = const()[name = tensor("op_1124"), val = tensor([1, 1, 3, 1])]; - tensor var_1125 = reshape(shape = var_1124, x = valid_mask)[name = tensor("op_1125")]; - tensor v_masked_1 = mul(x = v_9, y = var_1125)[name = tensor("v_masked_1")]; - tensor var_1127 = mul(x = prev_kv_9, y = sqrt_s0_9)[name = tensor("op_1127")]; - tensor var_1129_transpose_x_1 = const()[name = tensor("op_1129_transpose_x_1"), val = tensor(true)]; - tensor var_1129_transpose_y_1 = const()[name = tensor("op_1129_transpose_y_1"), val = tensor(false)]; - tensor var_1129 = matmul(transpose_x = var_1129_transpose_x_1, transpose_y = var_1129_transpose_y_1, x = k_9, y = v_masked_1)[name = tensor("op_1129")]; - tensor new_kv_unnorm_9 = add(x = var_1127, y = var_1129)[name = tensor("new_kv_unnorm_9")]; - tensor var_1131_keep_dims_0 = const()[name = tensor("op_1131_keep_dims_0"), val = tensor(false)]; - tensor var_1131 = reduce_sum(keep_dims = var_1131_keep_dims_0, x = valid_mask)[name = tensor("op_1131")]; - tensor var_1132 = const()[name = tensor("op_1132"), val = tensor([1])]; - tensor var_1133 = reshape(shape = var_1132, x = var_1131)[name = tensor("op_1133")]; - tensor new_scale_9 = add(x = prev_scale_9, y = var_1133)[name = tensor("new_scale_9")]; + tensor var_1147 = const()[name = tensor("op_1147"), val = tensor([1, 1, 3, 1])]; + tensor var_1148 = reshape(shape = var_1147, x = valid_mask)[name = tensor("op_1148")]; + tensor v_masked_1 = mul(x = v_9, y = var_1148)[name = tensor("v_masked_1")]; + tensor var_1150 = mul(x = prev_kv_9, y = sqrt_s0_9)[name = tensor("op_1150")]; + tensor var_1152_transpose_x_1 = const()[name = tensor("op_1152_transpose_x_1"), val = tensor(true)]; + tensor var_1152_transpose_y_1 = const()[name = tensor("op_1152_transpose_y_1"), val = tensor(false)]; + tensor var_1152 = matmul(transpose_x = var_1152_transpose_x_1, transpose_y = var_1152_transpose_y_1, x = k_9, y = v_masked_1)[name = tensor("op_1152")]; + tensor new_kv_unnorm_9 = add(x = var_1150, y = var_1152)[name = tensor("new_kv_unnorm_9")]; + tensor var_1154_keep_dims_0 = const()[name = tensor("op_1154_keep_dims_0"), val = tensor(false)]; + tensor var_1154 = reduce_sum(keep_dims = var_1154_keep_dims_0, x = valid_mask)[name = tensor("op_1154")]; + tensor var_1155 = const()[name = tensor("op_1155"), val = tensor([1])]; + tensor var_1156 = reshape(shape = var_1155, x = var_1154)[name = tensor("op_1156")]; + tensor new_scale_9 = add(x = prev_scale_9, y = var_1156)[name = tensor("new_scale_9")]; tensor const_21 = const()[name = tensor("const_21"), val = tensor(0x1.fffffep+127)]; - tensor clip_2 = clip(alpha = var_990, beta = const_21, x = new_scale_9)[name = tensor("clip_2")]; + tensor clip_2 = clip(alpha = var_49, beta = const_21, x = new_scale_9)[name = tensor("clip_2")]; tensor sqrt_new_scale_1 = sqrt(x = clip_2)[name = tensor("sqrt_new_scale_1")]; - tensor var_1137 = real_div(x = new_kv_unnorm_9, y = sqrt_new_scale_1)[name = tensor("op_1137")]; - tensor var_1138_perm_0 = const()[name = tensor("op_1138_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_1160 = real_div(x = new_kv_unnorm_9, y = sqrt_new_scale_1)[name = tensor("op_1160")]; + tensor var_1161_perm_0 = const()[name = tensor("op_1161_perm_0"), val = tensor([0, 2, 1, 3])]; tensor out_27_axes_0 = const()[name = tensor("out_27_axes_0"), val = tensor([-1])]; - tensor var_1138 = transpose(perm = var_1138_perm_0, x = out_25)[name = tensor("transpose_24")]; - tensor out_27 = layer_norm(axes = out_27_axes_0, epsilon = var_987, x = var_1138)[name = tensor("out_27")]; - tensor var_1142 = const()[name = tensor("op_1142"), val = tensor([12, 3, 256])]; - tensor out_29 = reshape(shape = var_1142, x = out_27)[name = tensor("out_29")]; - tensor var_1144 = silu(x = input_169)[name = tensor("op_1144")]; - tensor input_171 = mul(x = var_1144, y = out_29)[name = tensor("input_171")]; - tensor ret_out_9 = linear(bias = decoder_out_proj_0_bias, weight = decoder_out_proj_0_weight, x = input_171)[name = tensor("linear_42")]; - tensor input_173 = add(x = x_29, y = ret_out_9)[name = tensor("input_173")]; + tensor var_1161 = transpose(perm = var_1161_perm_0, x = out_25)[name = tensor("transpose_24")]; + tensor out_27 = layer_norm(axes = out_27_axes_0, epsilon = var_64, x = var_1161)[name = tensor("out_27")]; + tensor var_1165 = const()[name = tensor("op_1165"), val = tensor([12, 3, 256])]; + tensor out_29 = reshape(shape = var_1165, x = out_27)[name = tensor("out_29")]; + tensor var_1167 = silu(x = input_171)[name = tensor("op_1167")]; + tensor input_173 = mul(x = var_1167, y = out_29)[name = tensor("input_173")]; + tensor ret_out_9 = linear(bias = inner_decoder_out_proj_0_bias, weight = inner_decoder_out_proj_0_weight, x = input_173)[name = tensor("linear_42")]; + tensor input_175 = add(x = x_29, y = ret_out_9)[name = tensor("input_175")]; tensor xt_1_axes_0 = const()[name = tensor("xt_1_axes_0"), val = tensor([-1])]; - tensor xt_1 = layer_norm(axes = xt_1_axes_0, beta = decoder_norm11_0_bias, epsilon = var_985, gamma = decoder_norm11_0_weight, x = input_173)[name = tensor("xt_1")]; - tensor var_1154 = const()[name = tensor("op_1154"), val = tensor([1, 12, 3, 256])]; - tensor var_1155 = reshape(shape = var_1154, x = xt_1)[name = tensor("op_1155")]; - tensor var_1156_perm_0 = const()[name = tensor("op_1156_perm_0"), val = tensor([0, 2, 1, 3])]; - tensor var_1159 = const()[name = tensor("op_1159"), val = tensor([3, 12, 256])]; - tensor var_1156 = transpose(perm = var_1156_perm_0, x = var_1155)[name = tensor("transpose_23")]; - tensor query_1 = reshape(shape = var_1159, x = var_1156)[name = tensor("query_1")]; + tensor xt_1 = layer_norm(axes = xt_1_axes_0, beta = inner_decoder_norm11_0_bias, epsilon = var_56, gamma = inner_decoder_norm11_0_weight, x = input_175)[name = tensor("xt_1")]; + tensor var_1177 = const()[name = tensor("op_1177"), val = tensor([1, 12, 3, 256])]; + tensor var_1178 = reshape(shape = var_1177, x = xt_1)[name = tensor("op_1178")]; + tensor var_1179_perm_0 = const()[name = tensor("op_1179_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_1182 = const()[name = tensor("op_1182"), val = tensor([3, 12, 256])]; + tensor var_1179 = transpose(perm = var_1179_perm_0, x = var_1178)[name = tensor("transpose_23")]; + tensor query_1 = reshape(shape = var_1182, x = var_1179)[name = tensor("query_1")]; tensor query_3_perm_0 = const()[name = tensor("query_3_perm_0"), val = tensor([1, 0, 2])]; tensor query_3 = transpose(perm = query_3_perm_0, x = query_1)[name = tensor("transpose_22")]; - tensor var_1182 = linear(bias = decoder_self_attn2_0_in_proj_bias, weight = decoder_self_attn2_0_in_proj_weight, x = query_3)[name = tensor("linear_43")]; + tensor var_1205 = linear(bias = inner_decoder_self_attn2_0_in_proj_bias, weight = inner_decoder_self_attn2_0_in_proj_weight, x = query_3)[name = tensor("linear_43")]; tensor concat_1 = const()[name = tensor("concat_1"), val = tensor([12, 3, 3, 256])]; - tensor var_1184 = reshape(shape = concat_1, x = var_1182)[name = tensor("op_1184")]; - tensor var_1185_axes_0 = const()[name = tensor("op_1185_axes_0"), val = tensor([0])]; - tensor var_1185 = expand_dims(axes = var_1185_axes_0, x = var_1184)[name = tensor("op_1185")]; - tensor var_1186_perm_0 = const()[name = tensor("op_1186_perm_0"), val = tensor([-2, 1, 2, 0, 4])]; - tensor var_1187_axes_0 = const()[name = tensor("op_1187_axes_0"), val = tensor([-2])]; - tensor var_1186 = transpose(perm = var_1186_perm_0, x = var_1185)[name = tensor("transpose_21")]; - tensor var_1187 = squeeze(axes = var_1187_axes_0, x = var_1186)[name = tensor("op_1187")]; + tensor var_1207 = reshape(shape = concat_1, x = var_1205)[name = tensor("op_1207")]; + tensor var_1208_axes_0 = const()[name = tensor("op_1208_axes_0"), val = tensor([0])]; + tensor var_1208 = expand_dims(axes = var_1208_axes_0, x = var_1207)[name = tensor("op_1208")]; + tensor var_1209_perm_0 = const()[name = tensor("op_1209_perm_0"), val = tensor([-2, 1, 2, 0, 4])]; + tensor var_1210_axes_0 = const()[name = tensor("op_1210_axes_0"), val = tensor([-2])]; + tensor var_1209 = transpose(perm = var_1209_perm_0, x = var_1208)[name = tensor("transpose_21")]; + tensor var_1210 = squeeze(axes = var_1210_axes_0, x = var_1209)[name = tensor("op_1210")]; tensor q_11_begin_0 = const()[name = tensor("q_11_begin_0"), val = tensor([0, 0, 0, 0])]; tensor q_11_end_0 = const()[name = tensor("q_11_end_0"), val = tensor([1, 12, 3, 256])]; tensor q_11_end_mask_0 = const()[name = tensor("q_11_end_mask_0"), val = tensor([false, true, true, true])]; tensor q_11_squeeze_mask_0 = const()[name = tensor("q_11_squeeze_mask_0"), val = tensor([true, false, false, false])]; - tensor q_11 = slice_by_index(begin = q_11_begin_0, end = q_11_end_0, end_mask = q_11_end_mask_0, squeeze_mask = q_11_squeeze_mask_0, x = var_1187)[name = tensor("q_11")]; + tensor q_11 = slice_by_index(begin = q_11_begin_0, end = q_11_end_0, end_mask = q_11_end_mask_0, squeeze_mask = q_11_squeeze_mask_0, x = var_1210)[name = tensor("q_11")]; tensor k_11_begin_0 = const()[name = tensor("k_11_begin_0"), val = tensor([1, 0, 0, 0])]; tensor k_11_end_0 = const()[name = tensor("k_11_end_0"), val = tensor([2, 12, 3, 256])]; tensor k_11_end_mask_0 = const()[name = tensor("k_11_end_mask_0"), val = tensor([false, true, true, true])]; tensor k_11_squeeze_mask_0 = const()[name = tensor("k_11_squeeze_mask_0"), val = tensor([true, false, false, false])]; - tensor k_11 = slice_by_index(begin = k_11_begin_0, end = k_11_end_0, end_mask = k_11_end_mask_0, squeeze_mask = k_11_squeeze_mask_0, x = var_1187)[name = tensor("k_11")]; + tensor k_11 = slice_by_index(begin = k_11_begin_0, end = k_11_end_0, end_mask = k_11_end_mask_0, squeeze_mask = k_11_squeeze_mask_0, x = var_1210)[name = tensor("k_11")]; tensor v_11_begin_0 = const()[name = tensor("v_11_begin_0"), val = tensor([2, 0, 0, 0])]; tensor v_11_end_0 = const()[name = tensor("v_11_end_0"), val = tensor([3, 12, 3, 256])]; tensor v_11_end_mask_0 = const()[name = tensor("v_11_end_mask_0"), val = tensor([false, true, true, true])]; tensor v_11_squeeze_mask_0 = const()[name = tensor("v_11_squeeze_mask_0"), val = tensor([true, false, false, false])]; - tensor v_11 = slice_by_index(begin = v_11_begin_0, end = v_11_end_0, end_mask = v_11_end_mask_0, squeeze_mask = v_11_squeeze_mask_0, x = var_1187)[name = tensor("v_11")]; - tensor var_1195 = const()[name = tensor("op_1195"), val = tensor([12, 12, 64])]; - tensor var_1196 = reshape(shape = var_1195, x = q_11)[name = tensor("op_1196")]; + tensor v_11 = slice_by_index(begin = v_11_begin_0, end = v_11_end_0, end_mask = v_11_end_mask_0, squeeze_mask = v_11_squeeze_mask_0, x = var_1210)[name = tensor("v_11")]; + tensor var_1218 = const()[name = tensor("op_1218"), val = tensor([12, 12, 64])]; + tensor var_1219 = reshape(shape = var_1218, x = q_11)[name = tensor("op_1219")]; tensor q_13_perm_0 = const()[name = tensor("q_13_perm_0"), val = tensor([1, 0, 2])]; - tensor var_1202 = const()[name = tensor("op_1202"), val = tensor([12, 12, 64])]; - tensor var_1203 = reshape(shape = var_1202, x = k_11)[name = tensor("op_1203")]; + tensor var_1225 = const()[name = tensor("op_1225"), val = tensor([12, 12, 64])]; + tensor var_1226 = reshape(shape = var_1225, x = k_11)[name = tensor("op_1226")]; tensor k_13_perm_0 = const()[name = tensor("k_13_perm_0"), val = tensor([1, 0, 2])]; - tensor var_1209 = const()[name = tensor("op_1209"), val = tensor([12, 12, 64])]; - tensor var_1210 = reshape(shape = var_1209, x = v_11)[name = tensor("op_1210")]; + tensor var_1232 = const()[name = tensor("op_1232"), val = tensor([12, 12, 64])]; + tensor var_1233 = reshape(shape = var_1232, x = v_11)[name = tensor("op_1233")]; tensor v_13_perm_0 = const()[name = tensor("v_13_perm_0"), val = tensor([1, 0, 2])]; - tensor var_1213 = const()[name = tensor("op_1213"), val = tensor([3, 4, 12, 64])]; - tensor q_13 = transpose(perm = q_13_perm_0, x = var_1196)[name = tensor("transpose_20")]; - tensor q_15 = reshape(shape = var_1213, x = q_13)[name = tensor("q_15")]; - tensor var_1215 = const()[name = tensor("op_1215"), val = tensor([3, 4, 12, 64])]; - tensor k_13 = transpose(perm = k_13_perm_0, x = var_1203)[name = tensor("transpose_19")]; - tensor k_15 = reshape(shape = var_1215, x = k_13)[name = tensor("k_15")]; - tensor var_1217 = const()[name = tensor("op_1217"), val = tensor([3, 4, 12, 64])]; - tensor v_13 = transpose(perm = v_13_perm_0, x = var_1210)[name = tensor("transpose_18")]; - tensor v_15 = reshape(shape = var_1217, x = v_13)[name = tensor("v_15")]; + tensor var_1236 = const()[name = tensor("op_1236"), val = tensor([3, 4, 12, 64])]; + tensor q_13 = transpose(perm = q_13_perm_0, x = var_1219)[name = tensor("transpose_20")]; + tensor q_15 = reshape(shape = var_1236, x = q_13)[name = tensor("q_15")]; + tensor var_1238 = const()[name = tensor("op_1238"), val = tensor([3, 4, 12, 64])]; + tensor k_13 = transpose(perm = k_13_perm_0, x = var_1226)[name = tensor("transpose_19")]; + tensor k_15 = reshape(shape = var_1238, x = k_13)[name = tensor("k_15")]; + tensor var_1240 = const()[name = tensor("op_1240"), val = tensor([3, 4, 12, 64])]; + tensor v_13 = transpose(perm = v_13_perm_0, x = var_1233)[name = tensor("transpose_18")]; + tensor v_15 = reshape(shape = var_1240, x = v_13)[name = tensor("v_15")]; tensor mul_1_y_0 = const()[name = tensor("mul_1_y_0"), val = tensor(0x1p-3)]; tensor mul_1 = mul(x = q_15, y = mul_1_y_0)[name = tensor("mul_1")]; tensor matmul_0_transpose_y_0 = const()[name = tensor("matmul_0_transpose_y_0"), val = tensor(true)]; @@ -1079,30 +1090,30 @@ program(1.0) tensor attn_output_1_transpose_x_0 = const()[name = tensor("attn_output_1_transpose_x_0"), val = tensor(false)]; tensor attn_output_1_transpose_y_0 = const()[name = tensor("attn_output_1_transpose_y_0"), val = tensor(false)]; tensor attn_output_1 = matmul(transpose_x = attn_output_1_transpose_x_0, transpose_y = attn_output_1_transpose_y_0, x = softmax_0, y = v_15)[name = tensor("attn_output_1")]; - tensor var_1220 = const()[name = tensor("op_1220"), val = tensor([2, 0, 1, 3])]; - tensor var_1225 = const()[name = tensor("op_1225"), val = tensor([36, 256])]; - tensor var_1221 = transpose(perm = var_1220, x = attn_output_1)[name = tensor("transpose_17")]; - tensor attn_output_3 = reshape(shape = var_1225, x = var_1221)[name = tensor("attn_output_3")]; - tensor attn_output_5 = linear(bias = decoder_self_attn2_0_out_proj_bias, weight = decoder_self_attn2_0_out_proj_weight, x = attn_output_3)[name = tensor("linear_44")]; - tensor var_1229 = const()[name = tensor("op_1229"), val = tensor([12, 3, 256])]; - tensor attn_output_7 = reshape(shape = var_1229, x = attn_output_5)[name = tensor("attn_output_7")]; + tensor var_1243 = const()[name = tensor("op_1243"), val = tensor([2, 0, 1, 3])]; + tensor var_1248 = const()[name = tensor("op_1248"), val = tensor([36, 256])]; + tensor var_1244 = transpose(perm = var_1243, x = attn_output_1)[name = tensor("transpose_17")]; + tensor attn_output_3 = reshape(shape = var_1248, x = var_1244)[name = tensor("attn_output_3")]; + tensor attn_output_5 = linear(bias = inner_decoder_self_attn2_0_out_proj_bias, weight = inner_decoder_self_attn2_0_out_proj_weight, x = attn_output_3)[name = tensor("linear_44")]; + tensor var_1252 = const()[name = tensor("op_1252"), val = tensor([12, 3, 256])]; + tensor attn_output_7 = reshape(shape = var_1252, x = attn_output_5)[name = tensor("attn_output_7")]; tensor sa2_out_1_perm_0 = const()[name = tensor("sa2_out_1_perm_0"), val = tensor([1, 0, 2])]; tensor sa2_out_1 = transpose(perm = sa2_out_1_perm_0, x = attn_output_7)[name = tensor("transpose_16")]; - tensor input_175 = add(x = query_1, y = sa2_out_1)[name = tensor("input_175")]; - tensor input_177_axes_0 = const()[name = tensor("input_177_axes_0"), val = tensor([-1])]; - tensor input_177 = layer_norm(axes = input_177_axes_0, beta = decoder_norm21_0_bias, epsilon = var_985, gamma = decoder_norm21_0_weight, x = input_175)[name = tensor("input_177")]; - tensor input_179 = linear(bias = decoder_linear1_0_bias, weight = decoder_linear1_0_weight, x = input_177)[name = tensor("linear_45")]; - tensor input_181 = relu(x = input_179)[name = tensor("input_181")]; - tensor ffn_out_1 = linear(bias = decoder_linear2_0_bias, weight = decoder_linear2_0_weight, x = input_181)[name = tensor("linear_46")]; - tensor input_183 = add(x = input_177, y = ffn_out_1)[name = tensor("input_183")]; + tensor input_177 = add(x = query_1, y = sa2_out_1)[name = tensor("input_177")]; + tensor input_179_axes_0 = const()[name = tensor("input_179_axes_0"), val = tensor([-1])]; + tensor input_179 = layer_norm(axes = input_179_axes_0, beta = inner_decoder_norm21_0_bias, epsilon = var_56, gamma = inner_decoder_norm21_0_weight, x = input_177)[name = tensor("input_179")]; + tensor input_181 = linear(bias = inner_decoder_linear1_0_bias, weight = inner_decoder_linear1_0_weight, x = input_179)[name = tensor("linear_45")]; + tensor input_183 = relu(x = input_181)[name = tensor("input_183")]; + tensor ffn_out_1 = linear(bias = inner_decoder_linear2_0_bias, weight = inner_decoder_linear2_0_weight, x = input_183)[name = tensor("linear_46")]; + tensor input_185 = add(x = input_179, y = ffn_out_1)[name = tensor("input_185")]; tensor xt_3_axes_0 = const()[name = tensor("xt_3_axes_0"), val = tensor([-1])]; - tensor xt_3 = layer_norm(axes = xt_3_axes_0, beta = decoder_norm22_0_bias, epsilon = var_985, gamma = decoder_norm22_0_weight, x = input_183)[name = tensor("xt_3")]; - tensor var_1249 = const()[name = tensor("op_1249"), val = tensor([1, 3, 12, 256])]; - tensor x_31 = reshape(shape = var_1249, x = xt_3)[name = tensor("x_31")]; - tensor var_1251_perm_0 = const()[name = tensor("op_1251_perm_0"), val = tensor([0, 2, 1, 3])]; - tensor var_1255 = const()[name = tensor("op_1255"), val = tensor([12, 3, 256])]; - tensor var_1251 = transpose(perm = var_1251_perm_0, x = x_31)[name = tensor("transpose_15")]; - tensor x = reshape(shape = var_1255, x = var_1251)[name = tensor("x")]; + tensor xt_3 = layer_norm(axes = xt_3_axes_0, beta = inner_decoder_norm22_0_bias, epsilon = var_56, gamma = inner_decoder_norm22_0_weight, x = input_185)[name = tensor("xt_3")]; + tensor var_1272 = const()[name = tensor("op_1272"), val = tensor([1, 3, 12, 256])]; + tensor x_31 = reshape(shape = var_1272, x = xt_3)[name = tensor("x_31")]; + tensor var_1274_perm_0 = const()[name = tensor("op_1274_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_1278 = const()[name = tensor("op_1278"), val = tensor([12, 3, 256])]; + tensor var_1274 = transpose(perm = var_1274_perm_0, x = x_31)[name = tensor("transpose_15")]; + tensor x = reshape(shape = var_1278, x = var_1274)[name = tensor("x")]; tensor prev_kv_begin_0 = const()[name = tensor("prev_kv_begin_0"), val = tensor([1, 0, 0, 0, 0])]; tensor prev_kv_end_0 = const()[name = tensor("prev_kv_end_0"), val = tensor([2, 12, 4, 64, 64])]; tensor prev_kv_end_mask_0 = const()[name = tensor("prev_kv_end_mask_0"), val = tensor([false, true, true, true, true])]; @@ -1113,120 +1124,120 @@ program(1.0) tensor prev_scale_end_mask_0 = const()[name = tensor("prev_scale_end_mask_0"), val = tensor([false, true])]; tensor prev_scale_squeeze_mask_0 = const()[name = tensor("prev_scale_squeeze_mask_0"), val = tensor([true, false])]; tensor prev_scale = slice_by_index(begin = prev_scale_begin_0, end = prev_scale_end_0, end_mask = prev_scale_end_mask_0, squeeze_mask = prev_scale_squeeze_mask_0, x = dec_scale)[name = tensor("prev_scale")]; - tensor var_1263 = linear(bias = decoder_q_proj_1_bias, weight = decoder_q_proj_1_weight, x = x)[name = tensor("linear_47")]; - tensor var_1264 = const()[name = tensor("op_1264"), val = tensor([12, 3, 4, 64])]; - tensor var_1265 = reshape(shape = var_1264, x = var_1263)[name = tensor("op_1265")]; + tensor var_1286 = linear(bias = inner_decoder_q_proj_1_bias, weight = inner_decoder_q_proj_1_weight, x = x)[name = tensor("linear_47")]; + tensor var_1287 = const()[name = tensor("op_1287"), val = tensor([12, 3, 4, 64])]; + tensor var_1288 = reshape(shape = var_1287, x = var_1286)[name = tensor("op_1288")]; tensor q_17_perm_0 = const()[name = tensor("q_17_perm_0"), val = tensor([0, 2, 1, 3])]; - tensor var_1269 = linear(bias = decoder_k_proj_1_bias, weight = decoder_k_proj_1_weight, x = x)[name = tensor("linear_48")]; - tensor var_1270 = const()[name = tensor("op_1270"), val = tensor(0x1p-3)]; - tensor var_1271 = mul(x = var_1269, y = var_1270)[name = tensor("op_1271")]; - tensor var_1272 = const()[name = tensor("op_1272"), val = tensor([12, 3, 4, 64])]; - tensor var_1273 = reshape(shape = var_1272, x = var_1271)[name = tensor("op_1273")]; + tensor var_1292 = linear(bias = inner_decoder_k_proj_1_bias, weight = inner_decoder_k_proj_1_weight, x = x)[name = tensor("linear_48")]; + tensor var_1293 = const()[name = tensor("op_1293"), val = tensor(0x1p-3)]; + tensor var_1294 = mul(x = var_1292, y = var_1293)[name = tensor("op_1294")]; + tensor var_1295 = const()[name = tensor("op_1295"), val = tensor([12, 3, 4, 64])]; + tensor var_1296 = reshape(shape = var_1295, x = var_1294)[name = tensor("op_1296")]; tensor k_17_perm_0 = const()[name = tensor("k_17_perm_0"), val = tensor([0, 2, 1, 3])]; - tensor var_1277 = linear(bias = decoder_v_proj_1_bias, weight = decoder_v_proj_1_weight, x = x)[name = tensor("linear_49")]; - tensor var_1278 = const()[name = tensor("op_1278"), val = tensor([12, 3, 4, 64])]; - tensor var_1279 = reshape(shape = var_1278, x = var_1277)[name = tensor("op_1279")]; + tensor var_1300 = linear(bias = inner_decoder_v_proj_1_bias, weight = inner_decoder_v_proj_1_weight, x = x)[name = tensor("linear_49")]; + tensor var_1301 = const()[name = tensor("op_1301"), val = tensor([12, 3, 4, 64])]; + tensor var_1302 = reshape(shape = var_1301, x = var_1300)[name = tensor("op_1302")]; tensor v_17_perm_0 = const()[name = tensor("v_17_perm_0"), val = tensor([0, 2, 1, 3])]; - tensor input_187 = linear(bias = decoder_g_proj_1_bias, weight = decoder_g_proj_1_weight, x = x)[name = tensor("linear_50")]; + tensor input_189 = linear(bias = inner_decoder_g_proj_1_bias, weight = inner_decoder_g_proj_1_weight, x = x)[name = tensor("linear_50")]; tensor sqrt_s0 = sqrt(x = prev_scale)[name = tensor("sqrt_s0")]; tensor s_t = add(x = prev_scale, y = cumsum_mask_1)[name = tensor("s_t")]; tensor const_32 = const()[name = tensor("const_32"), val = tensor(0x1.fffffep+127)]; - tensor clip_3 = clip(alpha = var_990, beta = const_32, x = s_t)[name = tensor("clip_3")]; + tensor clip_3 = clip(alpha = var_49, beta = const_32, x = s_t)[name = tensor("clip_3")]; tensor sqrt_s_t = sqrt(x = clip_3)[name = tensor("sqrt_s_t")]; tensor qk_transpose_x_1 = const()[name = tensor("qk_transpose_x_1"), val = tensor(false)]; tensor qk_transpose_y_1 = const()[name = tensor("qk_transpose_y_1"), val = tensor(true)]; - tensor k_17 = transpose(perm = k_17_perm_0, x = var_1273)[name = tensor("transpose_13")]; - tensor q_17 = transpose(perm = q_17_perm_0, x = var_1265)[name = tensor("transpose_14")]; + tensor k_17 = transpose(perm = k_17_perm_0, x = var_1296)[name = tensor("transpose_13")]; + tensor q_17 = transpose(perm = q_17_perm_0, x = var_1288)[name = tensor("transpose_14")]; tensor qk = matmul(transpose_x = qk_transpose_x_1, transpose_y = qk_transpose_y_1, x = q_17, y = k_17)[name = tensor("qk")]; - tensor var_1294 = const()[name = tensor("op_1294"), val = tensor([3, 1])]; - tensor var_1295 = reshape(shape = var_1294, x = sqrt_s_t)[name = tensor("op_1295")]; - tensor M = real_div(x = causal_with_valid_1, y = var_1295)[name = tensor("M")]; - tensor var_1297 = mul(x = qk, y = M)[name = tensor("op_1297")]; - tensor inner_transpose_x_0 = const()[name = tensor("inner_transpose_x_0"), val = tensor(false)]; - tensor inner_transpose_y_0 = const()[name = tensor("inner_transpose_y_0"), val = tensor(false)]; - tensor v_17 = transpose(perm = v_17_perm_0, x = var_1279)[name = tensor("transpose_12")]; - tensor inner = matmul(transpose_x = inner_transpose_x_0, transpose_y = inner_transpose_y_0, x = var_1297, y = v_17)[name = tensor("inner")]; - tensor var_1299_transpose_x_0 = const()[name = tensor("op_1299_transpose_x_0"), val = tensor(false)]; - tensor var_1299_transpose_y_0 = const()[name = tensor("op_1299_transpose_y_0"), val = tensor(false)]; - tensor var_1299 = matmul(transpose_x = var_1299_transpose_x_0, transpose_y = var_1299_transpose_y_0, x = q_17, y = prev_kv)[name = tensor("op_1299")]; - tensor var_1300 = real_div(x = sqrt_s0, y = sqrt_s_t)[name = tensor("op_1300")]; - tensor var_1301 = const()[name = tensor("op_1301"), val = tensor([1, 1, 3, 1])]; - tensor var_1302 = reshape(shape = var_1301, x = var_1300)[name = tensor("op_1302")]; - tensor cross = mul(x = var_1299, y = var_1302)[name = tensor("cross")]; - tensor out_31 = add(x = inner, y = cross)[name = tensor("out_31")]; - tensor v_masked = mul(x = v_17, y = var_1125)[name = tensor("v_masked")]; - tensor var_1308 = mul(x = prev_kv, y = sqrt_s0)[name = tensor("op_1308")]; - tensor var_1310_transpose_x_1 = const()[name = tensor("op_1310_transpose_x_1"), val = tensor(true)]; - tensor var_1310_transpose_y_1 = const()[name = tensor("op_1310_transpose_y_1"), val = tensor(false)]; - tensor var_1310 = matmul(transpose_x = var_1310_transpose_x_1, transpose_y = var_1310_transpose_y_1, x = k_17, y = v_masked)[name = tensor("op_1310")]; - tensor new_kv_unnorm = add(x = var_1308, y = var_1310)[name = tensor("new_kv_unnorm")]; - tensor new_scale = add(x = prev_scale, y = var_1133)[name = tensor("new_scale")]; + tensor var_1317 = const()[name = tensor("op_1317"), val = tensor([3, 1])]; + tensor var_1318 = reshape(shape = var_1317, x = sqrt_s_t)[name = tensor("op_1318")]; + tensor M = real_div(x = causal_with_valid_1, y = var_1318)[name = tensor("M")]; + tensor var_1320 = mul(x = qk, y = M)[name = tensor("op_1320")]; + tensor inner_11_transpose_x_0 = const()[name = tensor("inner_11_transpose_x_0"), val = tensor(false)]; + tensor inner_11_transpose_y_0 = const()[name = tensor("inner_11_transpose_y_0"), val = tensor(false)]; + tensor v_17 = transpose(perm = v_17_perm_0, x = var_1302)[name = tensor("transpose_12")]; + tensor inner_11 = matmul(transpose_x = inner_11_transpose_x_0, transpose_y = inner_11_transpose_y_0, x = var_1320, y = v_17)[name = tensor("inner_11")]; + tensor var_1322_transpose_x_0 = const()[name = tensor("op_1322_transpose_x_0"), val = tensor(false)]; + tensor var_1322_transpose_y_0 = const()[name = tensor("op_1322_transpose_y_0"), val = tensor(false)]; + tensor var_1322 = matmul(transpose_x = var_1322_transpose_x_0, transpose_y = var_1322_transpose_y_0, x = q_17, y = prev_kv)[name = tensor("op_1322")]; + tensor var_1323 = real_div(x = sqrt_s0, y = sqrt_s_t)[name = tensor("op_1323")]; + tensor var_1324 = const()[name = tensor("op_1324"), val = tensor([1, 1, 3, 1])]; + tensor var_1325 = reshape(shape = var_1324, x = var_1323)[name = tensor("op_1325")]; + tensor cross = mul(x = var_1322, y = var_1325)[name = tensor("cross")]; + tensor out_31 = add(x = inner_11, y = cross)[name = tensor("out_31")]; + tensor v_masked = mul(x = v_17, y = var_1148)[name = tensor("v_masked")]; + tensor var_1331 = mul(x = prev_kv, y = sqrt_s0)[name = tensor("op_1331")]; + tensor var_1333_transpose_x_1 = const()[name = tensor("op_1333_transpose_x_1"), val = tensor(true)]; + tensor var_1333_transpose_y_1 = const()[name = tensor("op_1333_transpose_y_1"), val = tensor(false)]; + tensor var_1333 = matmul(transpose_x = var_1333_transpose_x_1, transpose_y = var_1333_transpose_y_1, x = k_17, y = v_masked)[name = tensor("op_1333")]; + tensor new_kv_unnorm = add(x = var_1331, y = var_1333)[name = tensor("new_kv_unnorm")]; + tensor new_scale = add(x = prev_scale, y = var_1156)[name = tensor("new_scale")]; tensor const_33 = const()[name = tensor("const_33"), val = tensor(0x1.fffffep+127)]; - tensor clip_4 = clip(alpha = var_990, beta = const_33, x = new_scale)[name = tensor("clip_4")]; + tensor clip_4 = clip(alpha = var_49, beta = const_33, x = new_scale)[name = tensor("clip_4")]; tensor sqrt_new_scale = sqrt(x = clip_4)[name = tensor("sqrt_new_scale")]; tensor nkv = real_div(x = new_kv_unnorm, y = sqrt_new_scale)[name = tensor("nkv")]; - tensor var_1319_perm_0 = const()[name = tensor("op_1319_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_1342_perm_0 = const()[name = tensor("op_1342_perm_0"), val = tensor([0, 2, 1, 3])]; tensor out_33_axes_0 = const()[name = tensor("out_33_axes_0"), val = tensor([-1])]; - tensor var_1319 = transpose(perm = var_1319_perm_0, x = out_31)[name = tensor("transpose_11")]; - tensor out_33 = layer_norm(axes = out_33_axes_0, epsilon = var_987, x = var_1319)[name = tensor("out_33")]; - tensor var_1323 = const()[name = tensor("op_1323"), val = tensor([12, 3, 256])]; - tensor out = reshape(shape = var_1323, x = out_33)[name = tensor("out")]; - tensor var_1325 = silu(x = input_187)[name = tensor("op_1325")]; - tensor input_189 = mul(x = var_1325, y = out)[name = tensor("input_189")]; - tensor ret_out = linear(bias = decoder_out_proj_1_bias, weight = decoder_out_proj_1_weight, x = input_189)[name = tensor("linear_51")]; - tensor input_191 = add(x = x, y = ret_out)[name = tensor("input_191")]; + tensor var_1342 = transpose(perm = var_1342_perm_0, x = out_31)[name = tensor("transpose_11")]; + tensor out_33 = layer_norm(axes = out_33_axes_0, epsilon = var_64, x = var_1342)[name = tensor("out_33")]; + tensor var_1346 = const()[name = tensor("op_1346"), val = tensor([12, 3, 256])]; + tensor out = reshape(shape = var_1346, x = out_33)[name = tensor("out")]; + tensor var_1348 = silu(x = input_189)[name = tensor("op_1348")]; + tensor input_191 = mul(x = var_1348, y = out)[name = tensor("input_191")]; + tensor ret_out = linear(bias = inner_decoder_out_proj_1_bias, weight = inner_decoder_out_proj_1_weight, x = input_191)[name = tensor("linear_51")]; + tensor input_193 = add(x = x, y = ret_out)[name = tensor("input_193")]; tensor xt_5_axes_0 = const()[name = tensor("xt_5_axes_0"), val = tensor([-1])]; - tensor xt_5 = layer_norm(axes = xt_5_axes_0, beta = decoder_norm11_1_bias, epsilon = var_985, gamma = decoder_norm11_1_weight, x = input_191)[name = tensor("xt_5")]; - tensor var_1335 = const()[name = tensor("op_1335"), val = tensor([1, 12, 3, 256])]; - tensor var_1336 = reshape(shape = var_1335, x = xt_5)[name = tensor("op_1336")]; - tensor var_1337_perm_0 = const()[name = tensor("op_1337_perm_0"), val = tensor([0, 2, 1, 3])]; - tensor var_1340 = const()[name = tensor("op_1340"), val = tensor([3, 12, 256])]; - tensor var_1337 = transpose(perm = var_1337_perm_0, x = var_1336)[name = tensor("transpose_10")]; - tensor query_5 = reshape(shape = var_1340, x = var_1337)[name = tensor("query_5")]; + tensor xt_5 = layer_norm(axes = xt_5_axes_0, beta = inner_decoder_norm11_1_bias, epsilon = var_56, gamma = inner_decoder_norm11_1_weight, x = input_193)[name = tensor("xt_5")]; + tensor var_1358 = const()[name = tensor("op_1358"), val = tensor([1, 12, 3, 256])]; + tensor var_1359 = reshape(shape = var_1358, x = xt_5)[name = tensor("op_1359")]; + tensor var_1360_perm_0 = const()[name = tensor("op_1360_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_1363 = const()[name = tensor("op_1363"), val = tensor([3, 12, 256])]; + tensor var_1360 = transpose(perm = var_1360_perm_0, x = var_1359)[name = tensor("transpose_10")]; + tensor query_5 = reshape(shape = var_1363, x = var_1360)[name = tensor("query_5")]; tensor query_perm_0 = const()[name = tensor("query_perm_0"), val = tensor([1, 0, 2])]; tensor query = transpose(perm = query_perm_0, x = query_5)[name = tensor("transpose_9")]; - tensor var_1363 = linear(bias = decoder_self_attn2_1_in_proj_bias, weight = decoder_self_attn2_1_in_proj_weight, x = query)[name = tensor("linear_52")]; + tensor var_1386 = linear(bias = inner_decoder_self_attn2_1_in_proj_bias, weight = inner_decoder_self_attn2_1_in_proj_weight, x = query)[name = tensor("linear_52")]; tensor concat_2 = const()[name = tensor("concat_2"), val = tensor([12, 3, 3, 256])]; - tensor var_1365 = reshape(shape = concat_2, x = var_1363)[name = tensor("op_1365")]; - tensor var_1366_axes_0 = const()[name = tensor("op_1366_axes_0"), val = tensor([0])]; - tensor var_1366 = expand_dims(axes = var_1366_axes_0, x = var_1365)[name = tensor("op_1366")]; - tensor var_1367_perm_0 = const()[name = tensor("op_1367_perm_0"), val = tensor([-2, 1, 2, 0, 4])]; - tensor var_1368_axes_0 = const()[name = tensor("op_1368_axes_0"), val = tensor([-2])]; - tensor var_1367 = transpose(perm = var_1367_perm_0, x = var_1366)[name = tensor("transpose_8")]; - tensor var_1368 = squeeze(axes = var_1368_axes_0, x = var_1367)[name = tensor("op_1368")]; + tensor var_1388 = reshape(shape = concat_2, x = var_1386)[name = tensor("op_1388")]; + tensor var_1389_axes_0 = const()[name = tensor("op_1389_axes_0"), val = tensor([0])]; + tensor var_1389 = expand_dims(axes = var_1389_axes_0, x = var_1388)[name = tensor("op_1389")]; + tensor var_1390_perm_0 = const()[name = tensor("op_1390_perm_0"), val = tensor([-2, 1, 2, 0, 4])]; + tensor var_1391_axes_0 = const()[name = tensor("op_1391_axes_0"), val = tensor([-2])]; + tensor var_1390 = transpose(perm = var_1390_perm_0, x = var_1389)[name = tensor("transpose_8")]; + tensor var_1391 = squeeze(axes = var_1391_axes_0, x = var_1390)[name = tensor("op_1391")]; tensor q_19_begin_0 = const()[name = tensor("q_19_begin_0"), val = tensor([0, 0, 0, 0])]; tensor q_19_end_0 = const()[name = tensor("q_19_end_0"), val = tensor([1, 12, 3, 256])]; tensor q_19_end_mask_0 = const()[name = tensor("q_19_end_mask_0"), val = tensor([false, true, true, true])]; tensor q_19_squeeze_mask_0 = const()[name = tensor("q_19_squeeze_mask_0"), val = tensor([true, false, false, false])]; - tensor q_19 = slice_by_index(begin = q_19_begin_0, end = q_19_end_0, end_mask = q_19_end_mask_0, squeeze_mask = q_19_squeeze_mask_0, x = var_1368)[name = tensor("q_19")]; + tensor q_19 = slice_by_index(begin = q_19_begin_0, end = q_19_end_0, end_mask = q_19_end_mask_0, squeeze_mask = q_19_squeeze_mask_0, x = var_1391)[name = tensor("q_19")]; tensor k_19_begin_0 = const()[name = tensor("k_19_begin_0"), val = tensor([1, 0, 0, 0])]; tensor k_19_end_0 = const()[name = tensor("k_19_end_0"), val = tensor([2, 12, 3, 256])]; tensor k_19_end_mask_0 = const()[name = tensor("k_19_end_mask_0"), val = tensor([false, true, true, true])]; tensor k_19_squeeze_mask_0 = const()[name = tensor("k_19_squeeze_mask_0"), val = tensor([true, false, false, false])]; - tensor k_19 = slice_by_index(begin = k_19_begin_0, end = k_19_end_0, end_mask = k_19_end_mask_0, squeeze_mask = k_19_squeeze_mask_0, x = var_1368)[name = tensor("k_19")]; + tensor k_19 = slice_by_index(begin = k_19_begin_0, end = k_19_end_0, end_mask = k_19_end_mask_0, squeeze_mask = k_19_squeeze_mask_0, x = var_1391)[name = tensor("k_19")]; tensor v_19_begin_0 = const()[name = tensor("v_19_begin_0"), val = tensor([2, 0, 0, 0])]; tensor v_19_end_0 = const()[name = tensor("v_19_end_0"), val = tensor([3, 12, 3, 256])]; tensor v_19_end_mask_0 = const()[name = tensor("v_19_end_mask_0"), val = tensor([false, true, true, true])]; tensor v_19_squeeze_mask_0 = const()[name = tensor("v_19_squeeze_mask_0"), val = tensor([true, false, false, false])]; - tensor v_19 = slice_by_index(begin = v_19_begin_0, end = v_19_end_0, end_mask = v_19_end_mask_0, squeeze_mask = v_19_squeeze_mask_0, x = var_1368)[name = tensor("v_19")]; - tensor var_1376 = const()[name = tensor("op_1376"), val = tensor([12, 12, 64])]; - tensor var_1377 = reshape(shape = var_1376, x = q_19)[name = tensor("op_1377")]; + tensor v_19 = slice_by_index(begin = v_19_begin_0, end = v_19_end_0, end_mask = v_19_end_mask_0, squeeze_mask = v_19_squeeze_mask_0, x = var_1391)[name = tensor("v_19")]; + tensor var_1399 = const()[name = tensor("op_1399"), val = tensor([12, 12, 64])]; + tensor var_1400 = reshape(shape = var_1399, x = q_19)[name = tensor("op_1400")]; tensor q_21_perm_0 = const()[name = tensor("q_21_perm_0"), val = tensor([1, 0, 2])]; - tensor var_1383 = const()[name = tensor("op_1383"), val = tensor([12, 12, 64])]; - tensor var_1384 = reshape(shape = var_1383, x = k_19)[name = tensor("op_1384")]; + tensor var_1406 = const()[name = tensor("op_1406"), val = tensor([12, 12, 64])]; + tensor var_1407 = reshape(shape = var_1406, x = k_19)[name = tensor("op_1407")]; tensor k_21_perm_0 = const()[name = tensor("k_21_perm_0"), val = tensor([1, 0, 2])]; - tensor var_1390 = const()[name = tensor("op_1390"), val = tensor([12, 12, 64])]; - tensor var_1391 = reshape(shape = var_1390, x = v_19)[name = tensor("op_1391")]; + tensor var_1413 = const()[name = tensor("op_1413"), val = tensor([12, 12, 64])]; + tensor var_1414 = reshape(shape = var_1413, x = v_19)[name = tensor("op_1414")]; tensor v_21_perm_0 = const()[name = tensor("v_21_perm_0"), val = tensor([1, 0, 2])]; - tensor var_1394 = const()[name = tensor("op_1394"), val = tensor([3, 4, 12, 64])]; - tensor q_21 = transpose(perm = q_21_perm_0, x = var_1377)[name = tensor("transpose_7")]; - tensor q = reshape(shape = var_1394, x = q_21)[name = tensor("q")]; - tensor var_1396 = const()[name = tensor("op_1396"), val = tensor([3, 4, 12, 64])]; - tensor k_21 = transpose(perm = k_21_perm_0, x = var_1384)[name = tensor("transpose_6")]; - tensor k = reshape(shape = var_1396, x = k_21)[name = tensor("k")]; - tensor var_1398 = const()[name = tensor("op_1398"), val = tensor([3, 4, 12, 64])]; - tensor v_21 = transpose(perm = v_21_perm_0, x = var_1391)[name = tensor("transpose_5")]; - tensor v = reshape(shape = var_1398, x = v_21)[name = tensor("v")]; + tensor var_1417 = const()[name = tensor("op_1417"), val = tensor([3, 4, 12, 64])]; + tensor q_21 = transpose(perm = q_21_perm_0, x = var_1400)[name = tensor("transpose_7")]; + tensor q = reshape(shape = var_1417, x = q_21)[name = tensor("q")]; + tensor var_1419 = const()[name = tensor("op_1419"), val = tensor([3, 4, 12, 64])]; + tensor k_21 = transpose(perm = k_21_perm_0, x = var_1407)[name = tensor("transpose_6")]; + tensor k = reshape(shape = var_1419, x = k_21)[name = tensor("k")]; + tensor var_1421 = const()[name = tensor("op_1421"), val = tensor([3, 4, 12, 64])]; + tensor v_21 = transpose(perm = v_21_perm_0, x = var_1414)[name = tensor("transpose_5")]; + tensor v = reshape(shape = var_1421, x = v_21)[name = tensor("v")]; tensor mul_3_y_0 = const()[name = tensor("mul_3_y_0"), val = tensor(0x1p-3)]; tensor mul_3 = mul(x = q, y = mul_3_y_0)[name = tensor("mul_3")]; tensor matmul_1_transpose_y_0 = const()[name = tensor("matmul_1_transpose_y_0"), val = tensor(true)]; @@ -1237,36 +1248,36 @@ program(1.0) tensor attn_output_9_transpose_x_0 = const()[name = tensor("attn_output_9_transpose_x_0"), val = tensor(false)]; tensor attn_output_9_transpose_y_0 = const()[name = tensor("attn_output_9_transpose_y_0"), val = tensor(false)]; tensor attn_output_9 = matmul(transpose_x = attn_output_9_transpose_x_0, transpose_y = attn_output_9_transpose_y_0, x = softmax_1, y = v)[name = tensor("attn_output_9")]; - tensor var_1401 = const()[name = tensor("op_1401"), val = tensor([2, 0, 1, 3])]; - tensor var_1406 = const()[name = tensor("op_1406"), val = tensor([36, 256])]; - tensor var_1402 = transpose(perm = var_1401, x = attn_output_9)[name = tensor("transpose_4")]; - tensor attn_output_11 = reshape(shape = var_1406, x = var_1402)[name = tensor("attn_output_11")]; - tensor attn_output_13 = linear(bias = decoder_self_attn2_1_out_proj_bias, weight = decoder_self_attn2_1_out_proj_weight, x = attn_output_11)[name = tensor("linear_53")]; - tensor var_1410 = const()[name = tensor("op_1410"), val = tensor([12, 3, 256])]; - tensor attn_output = reshape(shape = var_1410, x = attn_output_13)[name = tensor("attn_output")]; + tensor var_1424 = const()[name = tensor("op_1424"), val = tensor([2, 0, 1, 3])]; + tensor var_1429 = const()[name = tensor("op_1429"), val = tensor([36, 256])]; + tensor var_1425 = transpose(perm = var_1424, x = attn_output_9)[name = tensor("transpose_4")]; + tensor attn_output_11 = reshape(shape = var_1429, x = var_1425)[name = tensor("attn_output_11")]; + tensor attn_output_13 = linear(bias = inner_decoder_self_attn2_1_out_proj_bias, weight = inner_decoder_self_attn2_1_out_proj_weight, x = attn_output_11)[name = tensor("linear_53")]; + tensor var_1433 = const()[name = tensor("op_1433"), val = tensor([12, 3, 256])]; + tensor attn_output = reshape(shape = var_1433, x = attn_output_13)[name = tensor("attn_output")]; tensor sa2_out_perm_0 = const()[name = tensor("sa2_out_perm_0"), val = tensor([1, 0, 2])]; tensor sa2_out = transpose(perm = sa2_out_perm_0, x = attn_output)[name = tensor("transpose_3")]; - tensor input_193 = add(x = query_5, y = sa2_out)[name = tensor("input_193")]; - tensor input_195_axes_0 = const()[name = tensor("input_195_axes_0"), val = tensor([-1])]; - tensor input_195 = layer_norm(axes = input_195_axes_0, beta = decoder_norm21_1_bias, epsilon = var_985, gamma = decoder_norm21_1_weight, x = input_193)[name = tensor("input_195")]; - tensor input_197 = linear(bias = decoder_linear1_1_bias, weight = decoder_linear1_1_weight, x = input_195)[name = tensor("linear_54")]; - tensor input_199 = relu(x = input_197)[name = tensor("input_199")]; - tensor ffn_out = linear(bias = decoder_linear2_1_bias, weight = decoder_linear2_1_weight, x = input_199)[name = tensor("linear_55")]; - tensor input_201 = add(x = input_195, y = ffn_out)[name = tensor("input_201")]; + tensor input_195 = add(x = query_5, y = sa2_out)[name = tensor("input_195")]; + tensor input_197_axes_0 = const()[name = tensor("input_197_axes_0"), val = tensor([-1])]; + tensor input_197 = layer_norm(axes = input_197_axes_0, beta = inner_decoder_norm21_1_bias, epsilon = var_56, gamma = inner_decoder_norm21_1_weight, x = input_195)[name = tensor("input_197")]; + tensor input_199 = linear(bias = inner_decoder_linear1_1_bias, weight = inner_decoder_linear1_1_weight, x = input_197)[name = tensor("linear_54")]; + tensor input_201 = relu(x = input_199)[name = tensor("input_201")]; + tensor ffn_out = linear(bias = inner_decoder_linear2_1_bias, weight = inner_decoder_linear2_1_weight, x = input_201)[name = tensor("linear_55")]; + tensor input_203 = add(x = input_197, y = ffn_out)[name = tensor("input_203")]; tensor xt_axes_0 = const()[name = tensor("xt_axes_0"), val = tensor([-1])]; - tensor xt = layer_norm(axes = xt_axes_0, beta = decoder_norm22_1_bias, epsilon = var_985, gamma = decoder_norm22_1_weight, x = input_201)[name = tensor("xt")]; - tensor var_1430 = const()[name = tensor("op_1430"), val = tensor([1, 3, 12, 256])]; - tensor input = reshape(shape = var_1430, x = xt)[name = tensor("input")]; - tensor var_1432 = const()[name = tensor("op_1432"), val = tensor([-1])]; - tensor var_1433 = reduce_l2_norm(axes = var_1432, keep_dims = var_988, x = input)[name = tensor("op_1433")]; + tensor xt = layer_norm(axes = xt_axes_0, beta = inner_decoder_norm22_1_bias, epsilon = var_56, gamma = inner_decoder_norm22_1_weight, x = input_203)[name = tensor("xt")]; + tensor var_1453 = const()[name = tensor("op_1453"), val = tensor([1, 3, 12, 256])]; + tensor input = reshape(shape = var_1453, x = xt)[name = tensor("input")]; + tensor var_1455 = const()[name = tensor("op_1455"), val = tensor([-1])]; + tensor var_1456 = reduce_l2_norm(axes = var_1455, keep_dims = var_55, x = input)[name = tensor("op_1456")]; tensor const_42 = const()[name = tensor("const_42"), val = tensor(0x1.fffffep+127)]; - tensor clip_5 = clip(alpha = var_980, beta = const_42, x = var_1433)[name = tensor("clip_5")]; - tensor var_1435 = real_div(x = input, y = clip_5)[name = tensor("op_1435")]; + tensor clip_5 = clip(alpha = var_69, beta = const_42, x = var_1456)[name = tensor("clip_5")]; + tensor var_1458 = real_div(x = input, y = clip_5)[name = tensor("op_1458")]; tensor concat_6 = const()[name = tensor("concat_6"), val = tensor([3, 1, 256])]; tensor reshape_0 = reshape(shape = concat_6, x = emb)[name = tensor("reshape_0")]; tensor transpose_1_perm_0 = const()[name = tensor("transpose_1_perm_0"), val = tensor([0, 1, 3, 2])]; tensor concat_7 = const()[name = tensor("concat_7"), val = tensor([3, 256, 12])]; - tensor transpose_1 = transpose(perm = transpose_1_perm_0, x = var_1435)[name = tensor("transpose_2")]; + tensor transpose_1 = transpose(perm = transpose_1_perm_0, x = var_1458)[name = tensor("transpose_2")]; tensor reshape_1 = reshape(shape = concat_7, x = transpose_1)[name = tensor("reshape_1")]; tensor matmul_2_transpose_x_0 = const()[name = tensor("matmul_2_transpose_x_0"), val = tensor(false)]; tensor matmul_2_transpose_y_0 = const()[name = tensor("matmul_2_transpose_y_0"), val = tensor(false)]; @@ -1277,10 +1288,10 @@ program(1.0) tensor output_end_0 = const()[name = tensor("output_end_0"), val = tensor([1, 3, 11])]; tensor output_end_mask_0 = const()[name = tensor("output_end_mask_0"), val = tensor([true, true, false])]; tensor output = slice_by_index(begin = output_begin_0, end = output_end_0, end_mask = output_end_mask_0, x = reshape_2)[name = tensor("output")]; - tensor probs = sigmoid(x = output)[name = tensor("op_1439")]; - tensor var_1441_axis_0 = const()[name = tensor("op_1441_axis_0"), val = tensor(0)]; - tensor dec_kv_new = stack(axis = var_1441_axis_0, values = (var_1137, nkv))[name = tensor("op_1441")]; - tensor var_1443_axis_0 = const()[name = tensor("op_1443_axis_0"), val = tensor(0)]; - tensor dec_scale_new = stack(axis = var_1443_axis_0, values = (new_scale_9, new_scale))[name = tensor("op_1443")]; + tensor probs = sigmoid(x = output)[name = tensor("op_1462")]; + tensor var_1464_axis_0 = const()[name = tensor("op_1464_axis_0"), val = tensor(0)]; + tensor dec_kv_new = stack(axis = var_1464_axis_0, values = (var_1160, nkv))[name = tensor("op_1464")]; + tensor var_1466_axis_0 = const()[name = tensor("op_1466_axis_0"), val = tensor(0)]; + tensor dec_scale_new = stack(axis = var_1466_axis_0, values = (new_scale_9, new_scale))[name = tensor("op_1466")]; } -> (probs, enc_kv_new, enc_scale_new, enc_conv_cache_new, cnn_window_new, dec_kv_new, dec_scale_new); } \ No newline at end of file diff --git a/optimized/dih3/300ms/ls_eend_dih3_300ms.mlpackage/Data/com.apple.CoreML/model.mlmodel b/optimized/dih3/300ms/ls_eend_dih3_300ms.mlpackage/Data/com.apple.CoreML/model.mlmodel index d17264963ac53e18b1094bfefb69b97ef1395517..d68918dc375d2c7c3af14336bad337a916fb7fd2 100644 --- a/optimized/dih3/300ms/ls_eend_dih3_300ms.mlpackage/Data/com.apple.CoreML/model.mlmodel +++ b/optimized/dih3/300ms/ls_eend_dih3_300ms.mlpackage/Data/com.apple.CoreML/model.mlmodel @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:9986be1fe8f5022bafee300a2d8b49effe1540945ddb6ae9eaab94821982c313 -size 185469 +oid sha256:50cd667bc954b3b7ebc4888f43fd165ccee40cb8ed18429ed49740e418099933 +size 191014 diff --git a/optimized/dih3/300ms/ls_eend_dih3_300ms.mlpackage/Manifest.json b/optimized/dih3/300ms/ls_eend_dih3_300ms.mlpackage/Manifest.json index d174cfd424c10645eda95ea853dbe518047303b0..c3ffeb208da36539e3188bc5e845c45adf972d18 100644 --- a/optimized/dih3/300ms/ls_eend_dih3_300ms.mlpackage/Manifest.json +++ b/optimized/dih3/300ms/ls_eend_dih3_300ms.mlpackage/Manifest.json @@ -1,18 +1,18 @@ { "fileFormatVersion": "1.0.0", "itemInfoEntries": { - "5F03C3B0-3340-4DD2-B75B-F78D5596775F": { - "author": "com.apple.CoreML", - "description": "CoreML Model Weights", - "name": "weights", - "path": "com.apple.CoreML/weights" - }, - "F681008A-C1D4-4681-92CB-73FB966EF8F3": { + "0454F758-53A0-42B0-9756-169BE37A9D36": { "author": "com.apple.CoreML", "description": "CoreML Model Specification", "name": "model.mlmodel", "path": "com.apple.CoreML/model.mlmodel" + }, + "AE39AE0A-40A8-4BF2-AD07-BF3EC2B3068E": { + "author": "com.apple.CoreML", + "description": "CoreML Model Weights", + "name": "weights", + "path": "com.apple.CoreML/weights" } }, - "rootModelIdentifier": "F681008A-C1D4-4681-92CB-73FB966EF8F3" + "rootModelIdentifier": "0454F758-53A0-42B0-9756-169BE37A9D36" } diff --git a/optimized/dih3/400ms/ls_eend_dih3_400ms.mlmodelc/analytics/coremldata.bin b/optimized/dih3/400ms/ls_eend_dih3_400ms.mlmodelc/analytics/coremldata.bin index 71fa07880e9abdfe454605d65a2e99de84ec0b8c..9bd6bbbc32040f0faedbd1ff3b74fb49b5c05998 100644 --- a/optimized/dih3/400ms/ls_eend_dih3_400ms.mlmodelc/analytics/coremldata.bin +++ b/optimized/dih3/400ms/ls_eend_dih3_400ms.mlmodelc/analytics/coremldata.bin @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:88851cb773f9d1d6b36189e9c262dcc461c0d873a69a1f1dcf4087f48c4dc60b +oid sha256:3addfcd6bc7150ebc79d1c7db27c7efa955489919122c583720e067b219299cc size 243 diff --git a/optimized/dih3/400ms/ls_eend_dih3_400ms.mlmodelc/coremldata.bin b/optimized/dih3/400ms/ls_eend_dih3_400ms.mlmodelc/coremldata.bin index 18f01273657c24b054f38d6280b09122f7253d2a..a5260a611a9affcac436c67af4d6a8db2d0621fa 100644 --- a/optimized/dih3/400ms/ls_eend_dih3_400ms.mlmodelc/coremldata.bin +++ b/optimized/dih3/400ms/ls_eend_dih3_400ms.mlmodelc/coremldata.bin @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:606fe839f02edeb537028ed20f1abdc784cf5a18cef58ae9ff1a2783540da01b -size 1310 +oid sha256:f4df96ad5df011033bd68dce218b1d654cbb0cb49cbb3ee9dea54e7fa31130ac +size 1413 diff --git a/optimized/dih3/400ms/ls_eend_dih3_400ms.mlmodelc/metadata.json b/optimized/dih3/400ms/ls_eend_dih3_400ms.mlmodelc/metadata.json index cba5d30a2ad995503a2fcbad94a71f5afd760118..5dc93bf1d80f7012ba070e1a4aefee4d175ae4f1 100644 --- a/optimized/dih3/400ms/ls_eend_dih3_400ms.mlmodelc/metadata.json +++ b/optimized/dih3/400ms/ls_eend_dih3_400ms.mlmodelc/metadata.json @@ -1,7 +1,7 @@ [ { "metadataOutputVersion" : "3.0", - "shortDescription" : "LS-EEND DIHARD III streaming diarizer (pipeline, T=4, max_speakers=10)", + "shortDescription" : "LS-EEND DIHARD III streaming diarizer (pipeline, T=4, max_speakers=10, layout=raw_mel, cu=all)", "outputSchema" : [ { "hasShapeFlexibility" : "0", @@ -81,7 +81,7 @@ "specificationVersion" : 8, "mlProgramOperationTypeHistogram" : { "Ios16.reduceL2Norm" : 2, - "Ios17.reshape" : 68, + "Ios17.reshape" : 69, "Ios16.softmax" : 2, "Ios17.matmul" : 29, "Ios17.transpose" : 57, @@ -89,7 +89,7 @@ "Ios17.expandDims" : 3, "Ios17.add" : 46, "Ios16.sigmoid" : 5, - "Ios17.sliceByIndex" : 64, + "Ios17.sliceByIndex" : 68, "Tile" : 1, "Ios16.reduceSum" : 1, "Ios17.squeeze" : 2, @@ -101,7 +101,7 @@ "Ios16.silu" : 18, "Ios17.realDiv" : 20, "Ios17.linear" : 56, - "Stack" : 5, + "Stack" : 6, "Ios17.concat" : 22, "Ios16.relu" : 2, "Ios16.cumsum" : 1, @@ -128,9 +128,9 @@ "hasShapeFlexibility" : "0", "isOptional" : "0", "dataType" : "Float32", - "formattedType" : "MultiArray (Float32 1 × 4 × 345)", + "formattedType" : "MultiArray (Float32 1 × 45 × 23)", "shortDescription" : "", - "shape" : "[1, 4, 345]", + "shape" : "[1, 45, 23]", "name" : "features", "type" : "MultiArray" }, @@ -206,8 +206,8 @@ } ], "userDefinedMetadata" : { - "com.github.apple.coremltools.conversion_date" : "2026-04-16", - "config" : "{\"model_name\": \"dih3\", \"model_label\": \"DIHARD III\", \"variant\": \"pipeline\", \"chunk_size\": 4, \"step_duration_ms\": 400, \"frame_duration_ms\": 100, \"frame_duration_seconds\": 0.1, \"max_speakers\": 10, \"max_nspks\": 12, \"n_units\": 256, \"n_heads\": 4, \"enc_n_layers\": 4, \"dec_n_layers\": 2, \"conv_kernel_size\": 16, \"conv_delay\": 9, \"sample_rate\": 8000, \"win_length\": 200, \"hop_length\": 80, \"n_mels\": 23, \"context_size\": 7, \"subsampling\": 10, \"feat_type\": \"logmel23_cummn\", \"pure_roll\": true}", + "com.github.apple.coremltools.conversion_date" : "2026-04-18", + "config" : "{\"model_name\": \"dih3\", \"model_label\": \"DIHARD III\", \"variant\": \"pipeline\", \"chunk_size\": 4, \"step_duration_ms\": 400, \"frame_duration_ms\": 100, \"frame_duration_seconds\": 0.1, \"max_speakers\": 10, \"max_nspks\": 12, \"n_units\": 256, \"n_heads\": 4, \"enc_n_layers\": 4, \"dec_n_layers\": 2, \"conv_kernel_size\": 16, \"conv_delay\": 9, \"sample_rate\": 8000, \"win_length\": 200, \"hop_length\": 80, \"n_mels\": 23, \"context_size\": 7, \"subsampling\": 10, \"feat_type\": \"logmel23_cummn\", \"pure_roll\": true, \"input_layout\": \"raw_mel\", \"compute_units_export\": \"all\", \"raw_mel_length\": 45}", "com.github.apple.coremltools.source" : "torch==2.6.0", "com.github.apple.coremltools.version" : "9.0", "com.github.apple.coremltools.source_dialect" : "TorchScript" diff --git a/optimized/dih3/400ms/ls_eend_dih3_400ms.mlmodelc/model.mil b/optimized/dih3/400ms/ls_eend_dih3_400ms.mlmodelc/model.mil index 7135f6328b831d1038e863b851b7563d75bba5c3..546c200804eb017fad873f0d353fbf1c38982ea7 100644 --- a/optimized/dih3/400ms/ls_eend_dih3_400ms.mlmodelc/model.mil +++ b/optimized/dih3/400ms/ls_eend_dih3_400ms.mlmodelc/model.mil @@ -1,234 +1,256 @@ program(1.0) [buildInfo = dict, tensor>({{"coremlc-component-MIL", "3510.2.1"}, {"coremlc-version", "3500.32.1"}, {"coremltools-component-torch", "2.6.0"}, {"coremltools-source-dialect", "TorchScript"}, {"coremltools-version", "9.0"}})] { - func main(tensor cnn_window, tensor dec_kv, tensor dec_scale, tensor enc_conv_cache, tensor enc_kv, tensor enc_scale, tensor features, tensor valid_mask) { - tensor encoder_cnn_bias = const()[name = tensor("encoder_cnn_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(64)))]; - tensor encoder_cnn_weight = const()[name = tensor("encoder_cnn_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1152)))]; - tensor encoder__causal_mask = const()[name = tensor("encoder__causal_mask"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4981952)))]; - tensor encoder__t_index = const()[name = tensor("encoder__t_index"), val = tensor([0x1p+0, 0x1p+1, 0x1.8p+1, 0x1p+2])]; - tensor encoder_input_projection_linear_bias = const()[name = tensor("encoder_input_projection_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4982080)))]; - tensor encoder_input_projection_linear_weight = const()[name = tensor("encoder_input_projection_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4983168)))]; - tensor encoder_pre_layer_norm_bias = const()[name = tensor("encoder_pre_layer_norm_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(5336512)))]; - tensor encoder_pre_layer_norm_weight = const()[name = tensor("encoder_pre_layer_norm_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(5337600)))]; - tensor encoder_ffn1_0_module_sequential_0_bias = const()[name = tensor("encoder_ffn1_0_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(5338688)))]; - tensor encoder_ffn1_0_module_sequential_0_weight = const()[name = tensor("encoder_ffn1_0_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(5339776)))]; - tensor encoder_ffn1_0_module_sequential_1_linear_bias = const()[name = tensor("encoder_ffn1_0_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(5340864)))]; - tensor encoder_ffn1_0_module_sequential_1_linear_weight = const()[name = tensor("encoder_ffn1_0_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(5345024)))]; - tensor encoder_ffn1_0_module_sequential_4_linear_bias = const()[name = tensor("encoder_ffn1_0_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(6393664)))]; - tensor encoder_ffn1_0_module_sequential_4_linear_weight = const()[name = tensor("encoder_ffn1_0_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(6394752)))]; - tensor encoder_ret_lns_0_bias = const()[name = tensor("encoder_ret_lns_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7443392)))]; - tensor encoder_ret_lns_0_weight = const()[name = tensor("encoder_ret_lns_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7444480)))]; - tensor encoder_q_proj_0_bias = const()[name = tensor("encoder_q_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7445568)))]; - tensor encoder_q_proj_0_weight = const()[name = tensor("encoder_q_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7446656)))]; - tensor encoder_k_proj_0_bias = const()[name = tensor("encoder_k_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7708864)))]; - tensor encoder_k_proj_0_weight = const()[name = tensor("encoder_k_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7709952)))]; - tensor encoder_v_proj_0_bias = const()[name = tensor("encoder_v_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7972160)))]; - tensor encoder_v_proj_0_weight = const()[name = tensor("encoder_v_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7973248)))]; - tensor encoder_g_proj_0_bias = const()[name = tensor("encoder_g_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8235456)))]; - tensor encoder_g_proj_0_weight = const()[name = tensor("encoder_g_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8236544)))]; - tensor encoder_out_proj_0_bias = const()[name = tensor("encoder_out_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8498752)))]; - tensor encoder_out_proj_0_weight = const()[name = tensor("encoder_out_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8499840)))]; - tensor encoder_conv_module_0_sequential_0_bias = const()[name = tensor("encoder_conv_module_0_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8762048)))]; - tensor encoder_conv_module_0_sequential_0_weight = const()[name = tensor("encoder_conv_module_0_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8763136)))]; - tensor encoder_conv_module_0_sequential_2_conv_bias = const()[name = tensor("encoder_conv_module_0_sequential_2_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8764224)))]; - tensor encoder_conv_module_0_sequential_2_conv_weight = const()[name = tensor("encoder_conv_module_0_sequential_2_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8766336)))]; - tensor encoder_conv_module_0_sequential_4_conv_weight = const()[name = tensor("encoder_conv_module_0_sequential_4_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9290688)))]; - tensor encoder_conv_module_0_sequential_5_running_var = const()[name = tensor("encoder_conv_module_0_sequential_5_running_var"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9307136)))]; - tensor encoder_conv_module_0_sequential_5_running_mean = const()[name = tensor("encoder_conv_module_0_sequential_5_running_mean"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9308224)))]; - tensor encoder_conv_module_0_sequential_5_bias = const()[name = tensor("encoder_conv_module_0_sequential_5_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9309312)))]; - tensor encoder_conv_module_0_sequential_5_weight = const()[name = tensor("encoder_conv_module_0_sequential_5_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9310400)))]; - tensor encoder_conv_module_0_sequential_7_conv_bias = const()[name = tensor("encoder_conv_module_0_sequential_7_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9311488)))]; - tensor encoder_conv_module_0_sequential_7_conv_weight = const()[name = tensor("encoder_conv_module_0_sequential_7_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9312576)))]; - tensor encoder_ffn2_0_module_sequential_0_bias = const()[name = tensor("encoder_ffn2_0_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9574784)))]; - tensor encoder_ffn2_0_module_sequential_0_weight = const()[name = tensor("encoder_ffn2_0_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9575872)))]; - tensor encoder_ffn2_0_module_sequential_1_linear_bias = const()[name = tensor("encoder_ffn2_0_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9576960)))]; - tensor encoder_ffn2_0_module_sequential_1_linear_weight = const()[name = tensor("encoder_ffn2_0_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9581120)))]; - tensor encoder_ffn2_0_module_sequential_4_linear_bias = const()[name = tensor("encoder_ffn2_0_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(10629760)))]; - tensor encoder_ffn2_0_module_sequential_4_linear_weight = const()[name = tensor("encoder_ffn2_0_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(10630848)))]; - tensor encoder_layer_norm_0_bias = const()[name = tensor("encoder_layer_norm_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11679488)))]; - tensor encoder_layer_norm_0_weight = const()[name = tensor("encoder_layer_norm_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11680576)))]; - tensor encoder_ffn1_1_module_sequential_0_bias = const()[name = tensor("encoder_ffn1_1_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11681664)))]; - tensor encoder_ffn1_1_module_sequential_0_weight = const()[name = tensor("encoder_ffn1_1_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11682752)))]; - tensor encoder_ffn1_1_module_sequential_1_linear_bias = const()[name = tensor("encoder_ffn1_1_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11683840)))]; - tensor encoder_ffn1_1_module_sequential_1_linear_weight = const()[name = tensor("encoder_ffn1_1_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11688000)))]; - tensor encoder_ffn1_1_module_sequential_4_linear_bias = const()[name = tensor("encoder_ffn1_1_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(12736640)))]; - tensor encoder_ffn1_1_module_sequential_4_linear_weight = const()[name = tensor("encoder_ffn1_1_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(12737728)))]; - tensor encoder_ret_lns_1_bias = const()[name = tensor("encoder_ret_lns_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(13786368)))]; - tensor encoder_ret_lns_1_weight = const()[name = tensor("encoder_ret_lns_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(13787456)))]; - tensor encoder_q_proj_1_bias = const()[name = tensor("encoder_q_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(13788544)))]; - tensor encoder_q_proj_1_weight = const()[name = tensor("encoder_q_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(13789632)))]; - tensor encoder_k_proj_1_bias = const()[name = tensor("encoder_k_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14051840)))]; - tensor encoder_k_proj_1_weight = const()[name = tensor("encoder_k_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14052928)))]; - tensor encoder_v_proj_1_bias = const()[name = tensor("encoder_v_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14315136)))]; - tensor encoder_v_proj_1_weight = const()[name = tensor("encoder_v_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14316224)))]; - tensor encoder_g_proj_1_bias = const()[name = tensor("encoder_g_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14578432)))]; - tensor encoder_g_proj_1_weight = const()[name = tensor("encoder_g_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14579520)))]; - tensor encoder_out_proj_1_bias = const()[name = tensor("encoder_out_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14841728)))]; - tensor encoder_out_proj_1_weight = const()[name = tensor("encoder_out_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14842816)))]; - tensor encoder_conv_module_1_sequential_0_bias = const()[name = tensor("encoder_conv_module_1_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15105024)))]; - tensor encoder_conv_module_1_sequential_0_weight = const()[name = tensor("encoder_conv_module_1_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15106112)))]; - tensor encoder_conv_module_1_sequential_2_conv_bias = const()[name = tensor("encoder_conv_module_1_sequential_2_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15107200)))]; - tensor encoder_conv_module_1_sequential_2_conv_weight = const()[name = tensor("encoder_conv_module_1_sequential_2_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15109312)))]; - tensor encoder_conv_module_1_sequential_4_conv_weight = const()[name = tensor("encoder_conv_module_1_sequential_4_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15633664)))]; - tensor encoder_conv_module_1_sequential_5_running_var = const()[name = tensor("encoder_conv_module_1_sequential_5_running_var"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15650112)))]; - tensor encoder_conv_module_1_sequential_5_running_mean = const()[name = tensor("encoder_conv_module_1_sequential_5_running_mean"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15651200)))]; - tensor encoder_conv_module_1_sequential_5_bias = const()[name = tensor("encoder_conv_module_1_sequential_5_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15652288)))]; - tensor encoder_conv_module_1_sequential_5_weight = const()[name = tensor("encoder_conv_module_1_sequential_5_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15653376)))]; - tensor encoder_conv_module_1_sequential_7_conv_bias = const()[name = tensor("encoder_conv_module_1_sequential_7_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15654464)))]; - tensor encoder_conv_module_1_sequential_7_conv_weight = const()[name = tensor("encoder_conv_module_1_sequential_7_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15655552)))]; - tensor encoder_ffn2_1_module_sequential_0_bias = const()[name = tensor("encoder_ffn2_1_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15917760)))]; - tensor encoder_ffn2_1_module_sequential_0_weight = const()[name = tensor("encoder_ffn2_1_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15918848)))]; - tensor encoder_ffn2_1_module_sequential_1_linear_bias = const()[name = tensor("encoder_ffn2_1_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15919936)))]; - tensor encoder_ffn2_1_module_sequential_1_linear_weight = const()[name = tensor("encoder_ffn2_1_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15924096)))]; - tensor encoder_ffn2_1_module_sequential_4_linear_bias = const()[name = tensor("encoder_ffn2_1_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(16972736)))]; - tensor encoder_ffn2_1_module_sequential_4_linear_weight = const()[name = tensor("encoder_ffn2_1_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(16973824)))]; - tensor encoder_layer_norm_1_bias = const()[name = tensor("encoder_layer_norm_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18022464)))]; - tensor encoder_layer_norm_1_weight = const()[name = tensor("encoder_layer_norm_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18023552)))]; - tensor encoder_ffn1_2_module_sequential_0_bias = const()[name = tensor("encoder_ffn1_2_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18024640)))]; - tensor encoder_ffn1_2_module_sequential_0_weight = const()[name = tensor("encoder_ffn1_2_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18025728)))]; - tensor encoder_ffn1_2_module_sequential_1_linear_bias = const()[name = tensor("encoder_ffn1_2_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18026816)))]; - tensor encoder_ffn1_2_module_sequential_1_linear_weight = const()[name = tensor("encoder_ffn1_2_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18030976)))]; - tensor encoder_ffn1_2_module_sequential_4_linear_bias = const()[name = tensor("encoder_ffn1_2_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(19079616)))]; - tensor encoder_ffn1_2_module_sequential_4_linear_weight = const()[name = tensor("encoder_ffn1_2_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(19080704)))]; - tensor encoder_ret_lns_2_bias = const()[name = tensor("encoder_ret_lns_2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20129344)))]; - tensor encoder_ret_lns_2_weight = const()[name = tensor("encoder_ret_lns_2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20130432)))]; - tensor encoder_q_proj_2_bias = const()[name = tensor("encoder_q_proj_2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20131520)))]; - tensor encoder_q_proj_2_weight = const()[name = tensor("encoder_q_proj_2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20132608)))]; - tensor encoder_k_proj_2_bias = const()[name = tensor("encoder_k_proj_2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20394816)))]; - tensor encoder_k_proj_2_weight = const()[name = tensor("encoder_k_proj_2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20395904)))]; - tensor encoder_v_proj_2_bias = const()[name = tensor("encoder_v_proj_2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20658112)))]; - tensor encoder_v_proj_2_weight = const()[name = tensor("encoder_v_proj_2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20659200)))]; - tensor encoder_g_proj_2_bias = const()[name = tensor("encoder_g_proj_2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20921408)))]; - tensor encoder_g_proj_2_weight = const()[name = tensor("encoder_g_proj_2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20922496)))]; - tensor encoder_out_proj_2_bias = const()[name = tensor("encoder_out_proj_2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21184704)))]; - tensor encoder_out_proj_2_weight = const()[name = tensor("encoder_out_proj_2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21185792)))]; - tensor encoder_conv_module_2_sequential_0_bias = const()[name = tensor("encoder_conv_module_2_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21448000)))]; - tensor encoder_conv_module_2_sequential_0_weight = const()[name = tensor("encoder_conv_module_2_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21449088)))]; - tensor encoder_conv_module_2_sequential_2_conv_bias = const()[name = tensor("encoder_conv_module_2_sequential_2_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21450176)))]; - tensor encoder_conv_module_2_sequential_2_conv_weight = const()[name = tensor("encoder_conv_module_2_sequential_2_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21452288)))]; - tensor encoder_conv_module_2_sequential_4_conv_weight = const()[name = tensor("encoder_conv_module_2_sequential_4_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21976640)))]; - tensor encoder_conv_module_2_sequential_5_running_var = const()[name = tensor("encoder_conv_module_2_sequential_5_running_var"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21993088)))]; - tensor encoder_conv_module_2_sequential_5_running_mean = const()[name = tensor("encoder_conv_module_2_sequential_5_running_mean"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21994176)))]; - tensor encoder_conv_module_2_sequential_5_bias = const()[name = tensor("encoder_conv_module_2_sequential_5_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21995264)))]; - tensor encoder_conv_module_2_sequential_5_weight = const()[name = tensor("encoder_conv_module_2_sequential_5_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21996352)))]; - tensor encoder_conv_module_2_sequential_7_conv_bias = const()[name = tensor("encoder_conv_module_2_sequential_7_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21997440)))]; - tensor encoder_conv_module_2_sequential_7_conv_weight = const()[name = tensor("encoder_conv_module_2_sequential_7_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21998528)))]; - tensor encoder_ffn2_2_module_sequential_0_bias = const()[name = tensor("encoder_ffn2_2_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(22260736)))]; - tensor encoder_ffn2_2_module_sequential_0_weight = const()[name = tensor("encoder_ffn2_2_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(22261824)))]; - tensor encoder_ffn2_2_module_sequential_1_linear_bias = const()[name = tensor("encoder_ffn2_2_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(22262912)))]; - tensor encoder_ffn2_2_module_sequential_1_linear_weight = const()[name = tensor("encoder_ffn2_2_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(22267072)))]; - tensor encoder_ffn2_2_module_sequential_4_linear_bias = const()[name = tensor("encoder_ffn2_2_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(23315712)))]; - tensor encoder_ffn2_2_module_sequential_4_linear_weight = const()[name = tensor("encoder_ffn2_2_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(23316800)))]; - tensor encoder_layer_norm_2_bias = const()[name = tensor("encoder_layer_norm_2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(24365440)))]; - tensor encoder_layer_norm_2_weight = const()[name = tensor("encoder_layer_norm_2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(24366528)))]; - tensor encoder_ffn1_3_module_sequential_0_bias = const()[name = tensor("encoder_ffn1_3_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(24367616)))]; - tensor encoder_ffn1_3_module_sequential_0_weight = const()[name = tensor("encoder_ffn1_3_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(24368704)))]; - tensor encoder_ffn1_3_module_sequential_1_linear_bias = const()[name = tensor("encoder_ffn1_3_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(24369792)))]; - tensor encoder_ffn1_3_module_sequential_1_linear_weight = const()[name = tensor("encoder_ffn1_3_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(24373952)))]; - tensor encoder_ffn1_3_module_sequential_4_linear_bias = const()[name = tensor("encoder_ffn1_3_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(25422592)))]; - tensor encoder_ffn1_3_module_sequential_4_linear_weight = const()[name = tensor("encoder_ffn1_3_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(25423680)))]; - tensor encoder_ret_lns_3_bias = const()[name = tensor("encoder_ret_lns_3_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(26472320)))]; - tensor encoder_ret_lns_3_weight = const()[name = tensor("encoder_ret_lns_3_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(26473408)))]; - tensor encoder_q_proj_3_bias = const()[name = tensor("encoder_q_proj_3_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(26474496)))]; - tensor encoder_q_proj_3_weight = const()[name = tensor("encoder_q_proj_3_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(26475584)))]; - tensor encoder_k_proj_3_bias = const()[name = tensor("encoder_k_proj_3_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(26737792)))]; - tensor encoder_k_proj_3_weight = const()[name = tensor("encoder_k_proj_3_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(26738880)))]; - tensor encoder_v_proj_3_bias = const()[name = tensor("encoder_v_proj_3_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27001088)))]; - tensor encoder_v_proj_3_weight = const()[name = tensor("encoder_v_proj_3_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27002176)))]; - tensor encoder_g_proj_3_bias = const()[name = tensor("encoder_g_proj_3_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27264384)))]; - tensor encoder_g_proj_3_weight = const()[name = tensor("encoder_g_proj_3_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27265472)))]; - tensor encoder_out_proj_3_bias = const()[name = tensor("encoder_out_proj_3_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27527680)))]; - tensor encoder_out_proj_3_weight = const()[name = tensor("encoder_out_proj_3_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27528768)))]; - tensor encoder_conv_module_3_sequential_0_bias = const()[name = tensor("encoder_conv_module_3_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27790976)))]; - tensor encoder_conv_module_3_sequential_0_weight = const()[name = tensor("encoder_conv_module_3_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27792064)))]; - tensor encoder_conv_module_3_sequential_2_conv_bias = const()[name = tensor("encoder_conv_module_3_sequential_2_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27793152)))]; - tensor encoder_conv_module_3_sequential_2_conv_weight = const()[name = tensor("encoder_conv_module_3_sequential_2_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27795264)))]; - tensor encoder_conv_module_3_sequential_4_conv_weight = const()[name = tensor("encoder_conv_module_3_sequential_4_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28319616)))]; - tensor encoder_conv_module_3_sequential_5_running_var = const()[name = tensor("encoder_conv_module_3_sequential_5_running_var"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28336064)))]; - tensor encoder_conv_module_3_sequential_5_running_mean = const()[name = tensor("encoder_conv_module_3_sequential_5_running_mean"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28337152)))]; - tensor encoder_conv_module_3_sequential_5_bias = const()[name = tensor("encoder_conv_module_3_sequential_5_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28338240)))]; - tensor encoder_conv_module_3_sequential_5_weight = const()[name = tensor("encoder_conv_module_3_sequential_5_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28339328)))]; - tensor encoder_conv_module_3_sequential_7_conv_bias = const()[name = tensor("encoder_conv_module_3_sequential_7_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28340416)))]; - tensor encoder_conv_module_3_sequential_7_conv_weight = const()[name = tensor("encoder_conv_module_3_sequential_7_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28341504)))]; - tensor encoder_ffn2_3_module_sequential_0_bias = const()[name = tensor("encoder_ffn2_3_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28603712)))]; - tensor encoder_ffn2_3_module_sequential_0_weight = const()[name = tensor("encoder_ffn2_3_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28604800)))]; - tensor encoder_ffn2_3_module_sequential_1_linear_bias = const()[name = tensor("encoder_ffn2_3_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28605888)))]; - tensor encoder_ffn2_3_module_sequential_1_linear_weight = const()[name = tensor("encoder_ffn2_3_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28610048)))]; - tensor encoder_ffn2_3_module_sequential_4_linear_bias = const()[name = tensor("encoder_ffn2_3_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(29658688)))]; - tensor encoder_ffn2_3_module_sequential_4_linear_weight = const()[name = tensor("encoder_ffn2_3_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(29659776)))]; - tensor encoder_layer_norm_3_bias = const()[name = tensor("encoder_layer_norm_3_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(30708416)))]; - tensor encoder_layer_norm_3_weight = const()[name = tensor("encoder_layer_norm_3_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(30709504)))]; - tensor decoder__causal_mask = const()[name = tensor("decoder__causal_mask"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(30710592)))]; - tensor decoder_convert_bias = const()[name = tensor("decoder_convert_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(30710720)))]; - tensor decoder_convert_weight = const()[name = tensor("decoder_convert_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(30711808)))]; - tensor decoder_q_proj_0_bias = const()[name = tensor("decoder_q_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(31236160)))]; - tensor decoder_q_proj_0_weight = const()[name = tensor("decoder_q_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(31237248)))]; - tensor decoder_k_proj_0_bias = const()[name = tensor("decoder_k_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(31499456)))]; - tensor decoder_k_proj_0_weight = const()[name = tensor("decoder_k_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(31500544)))]; - tensor decoder_v_proj_0_bias = const()[name = tensor("decoder_v_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(31762752)))]; - tensor decoder_v_proj_0_weight = const()[name = tensor("decoder_v_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(31763840)))]; - tensor decoder_g_proj_0_bias = const()[name = tensor("decoder_g_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32026048)))]; - tensor decoder_g_proj_0_weight = const()[name = tensor("decoder_g_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32027136)))]; - tensor decoder_out_proj_0_bias = const()[name = tensor("decoder_out_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32289344)))]; - tensor decoder_out_proj_0_weight = const()[name = tensor("decoder_out_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32290432)))]; - tensor decoder_norm11_0_bias = const()[name = tensor("decoder_norm11_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32552640)))]; - tensor decoder_norm11_0_weight = const()[name = tensor("decoder_norm11_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32553728)))]; - tensor decoder_self_attn2_0_out_proj_bias = const()[name = tensor("decoder_self_attn2_0_out_proj_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32554816)))]; - tensor decoder_self_attn2_0_out_proj_weight = const()[name = tensor("decoder_self_attn2_0_out_proj_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32555904)))]; - tensor decoder_self_attn2_0_in_proj_bias = const()[name = tensor("decoder_self_attn2_0_in_proj_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32818112)))]; - tensor decoder_self_attn2_0_in_proj_weight = const()[name = tensor("decoder_self_attn2_0_in_proj_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32821248)))]; - tensor decoder_norm21_0_bias = const()[name = tensor("decoder_norm21_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(33607744)))]; - tensor decoder_norm21_0_weight = const()[name = tensor("decoder_norm21_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(33608832)))]; - tensor decoder_linear1_0_bias = const()[name = tensor("decoder_linear1_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(33609920)))]; - tensor decoder_linear1_0_weight = const()[name = tensor("decoder_linear1_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(33618176)))]; - tensor decoder_linear2_0_bias = const()[name = tensor("decoder_linear2_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(35715392)))]; - tensor decoder_linear2_0_weight = const()[name = tensor("decoder_linear2_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(35716480)))]; - tensor decoder_norm22_0_bias = const()[name = tensor("decoder_norm22_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(37813696)))]; - tensor decoder_norm22_0_weight = const()[name = tensor("decoder_norm22_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(37814784)))]; - tensor decoder_q_proj_1_bias = const()[name = tensor("decoder_q_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(37815872)))]; - tensor decoder_q_proj_1_weight = const()[name = tensor("decoder_q_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(37816960)))]; - tensor decoder_k_proj_1_bias = const()[name = tensor("decoder_k_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38079168)))]; - tensor decoder_k_proj_1_weight = const()[name = tensor("decoder_k_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38080256)))]; - tensor decoder_v_proj_1_bias = const()[name = tensor("decoder_v_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38342464)))]; - tensor decoder_v_proj_1_weight = const()[name = tensor("decoder_v_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38343552)))]; - tensor decoder_g_proj_1_bias = const()[name = tensor("decoder_g_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38605760)))]; - tensor decoder_g_proj_1_weight = const()[name = tensor("decoder_g_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38606848)))]; - tensor decoder_out_proj_1_bias = const()[name = tensor("decoder_out_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38869056)))]; - tensor decoder_out_proj_1_weight = const()[name = tensor("decoder_out_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38870144)))]; - tensor decoder_norm11_1_bias = const()[name = tensor("decoder_norm11_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(39132352)))]; - tensor decoder_norm11_1_weight = const()[name = tensor("decoder_norm11_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(39133440)))]; - tensor decoder_self_attn2_1_out_proj_bias = const()[name = tensor("decoder_self_attn2_1_out_proj_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(39134528)))]; - tensor decoder_self_attn2_1_out_proj_weight = const()[name = tensor("decoder_self_attn2_1_out_proj_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(39135616)))]; - tensor decoder_self_attn2_1_in_proj_bias = const()[name = tensor("decoder_self_attn2_1_in_proj_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(39397824)))]; - tensor decoder_self_attn2_1_in_proj_weight = const()[name = tensor("decoder_self_attn2_1_in_proj_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(39400960)))]; - tensor decoder_norm21_1_bias = const()[name = tensor("decoder_norm21_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(40187456)))]; - tensor decoder_norm21_1_weight = const()[name = tensor("decoder_norm21_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(40188544)))]; - tensor decoder_linear1_1_bias = const()[name = tensor("decoder_linear1_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(40189632)))]; - tensor decoder_linear1_1_weight = const()[name = tensor("decoder_linear1_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(40197888)))]; - tensor decoder_linear2_1_bias = const()[name = tensor("decoder_linear2_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(42295104)))]; - tensor decoder_linear2_1_weight = const()[name = tensor("decoder_linear2_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(42296192)))]; - tensor decoder_norm22_1_bias = const()[name = tensor("decoder_norm22_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(44393408)))]; - tensor decoder_norm22_1_weight = const()[name = tensor("decoder_norm22_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(44394496)))]; - tensor var_11 = const()[name = tensor("op_11"), val = tensor(0x1.5798eep-27)]; - tensor var_17 = const()[name = tensor("op_17"), val = tensor(0x1.0c6f7ap-20)]; - tensor var_20 = const()[name = tensor("op_20"), val = tensor(2)]; - tensor var_23 = const()[name = tensor("op_23"), val = tensor(0)]; - tensor var_26 = const()[name = tensor("op_26"), val = tensor(1)]; - tensor var_28 = const()[name = tensor("op_28"), val = tensor(0x1.4f8b58p-17)]; - tensor var_29 = const()[name = tensor("op_29"), val = tensor(true)]; - tensor input_1 = linear(bias = encoder_input_projection_linear_bias, weight = encoder_input_projection_linear_weight, x = features)[name = tensor("linear_0")]; - tensor input_3_axes_0 = const()[name = tensor("input_3_axes_0"), val = tensor([-1])]; - tensor input_3 = layer_norm(axes = input_3_axes_0, beta = encoder_pre_layer_norm_bias, epsilon = var_28, gamma = encoder_pre_layer_norm_weight, x = input_1)[name = tensor("input_3")]; + func main(tensor cnn_window, tensor dec_kv, tensor dec_scale, tensor enc_conv_cache, tensor enc_kv, tensor enc_scale, tensor features, tensor valid_mask) { + tensor inner_encoder_cnn_bias = const()[name = tensor("inner_encoder_cnn_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(64)))]; + tensor inner_encoder_cnn_weight = const()[name = tensor("inner_encoder_cnn_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1152)))]; + tensor inner_encoder__causal_mask = const()[name = tensor("inner_encoder__causal_mask"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4981952)))]; + tensor inner_encoder__t_index = const()[name = tensor("inner_encoder__t_index"), val = tensor([0x1p+0, 0x1p+1, 0x1.8p+1, 0x1p+2])]; + tensor inner_encoder_input_projection_linear_bias = const()[name = tensor("inner_encoder_input_projection_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4982080)))]; + tensor inner_encoder_input_projection_linear_weight = const()[name = tensor("inner_encoder_input_projection_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4983168)))]; + tensor inner_encoder_pre_layer_norm_bias = const()[name = tensor("inner_encoder_pre_layer_norm_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(5336512)))]; + tensor inner_encoder_pre_layer_norm_weight = const()[name = tensor("inner_encoder_pre_layer_norm_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(5337600)))]; + tensor inner_encoder_ffn1_0_module_sequential_0_bias = const()[name = tensor("inner_encoder_ffn1_0_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(5338688)))]; + tensor inner_encoder_ffn1_0_module_sequential_0_weight = const()[name = tensor("inner_encoder_ffn1_0_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(5339776)))]; + tensor inner_encoder_ffn1_0_module_sequential_1_linear_bias = const()[name = tensor("inner_encoder_ffn1_0_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(5340864)))]; + tensor inner_encoder_ffn1_0_module_sequential_1_linear_weight = const()[name = tensor("inner_encoder_ffn1_0_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(5345024)))]; + tensor inner_encoder_ffn1_0_module_sequential_4_linear_bias = const()[name = tensor("inner_encoder_ffn1_0_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(6393664)))]; + tensor inner_encoder_ffn1_0_module_sequential_4_linear_weight = const()[name = tensor("inner_encoder_ffn1_0_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(6394752)))]; + tensor inner_encoder_ret_lns_0_bias = const()[name = tensor("inner_encoder_ret_lns_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7443392)))]; + tensor inner_encoder_ret_lns_0_weight = const()[name = tensor("inner_encoder_ret_lns_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7444480)))]; + tensor inner_encoder_q_proj_0_bias = const()[name = tensor("inner_encoder_q_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7445568)))]; + tensor inner_encoder_q_proj_0_weight = const()[name = tensor("inner_encoder_q_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7446656)))]; + tensor inner_encoder_k_proj_0_bias = const()[name = tensor("inner_encoder_k_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7708864)))]; + tensor inner_encoder_k_proj_0_weight = const()[name = tensor("inner_encoder_k_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7709952)))]; + tensor inner_encoder_v_proj_0_bias = const()[name = tensor("inner_encoder_v_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7972160)))]; + tensor inner_encoder_v_proj_0_weight = const()[name = tensor("inner_encoder_v_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7973248)))]; + tensor inner_encoder_g_proj_0_bias = const()[name = tensor("inner_encoder_g_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8235456)))]; + tensor inner_encoder_g_proj_0_weight = const()[name = tensor("inner_encoder_g_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8236544)))]; + tensor inner_encoder_out_proj_0_bias = const()[name = tensor("inner_encoder_out_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8498752)))]; + tensor inner_encoder_out_proj_0_weight = const()[name = tensor("inner_encoder_out_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8499840)))]; + tensor inner_encoder_conv_module_0_sequential_0_bias = const()[name = tensor("inner_encoder_conv_module_0_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8762048)))]; + tensor inner_encoder_conv_module_0_sequential_0_weight = const()[name = tensor("inner_encoder_conv_module_0_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8763136)))]; + tensor inner_encoder_conv_module_0_sequential_2_conv_bias = const()[name = tensor("inner_encoder_conv_module_0_sequential_2_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8764224)))]; + tensor inner_encoder_conv_module_0_sequential_2_conv_weight = const()[name = tensor("inner_encoder_conv_module_0_sequential_2_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8766336)))]; + tensor inner_encoder_conv_module_0_sequential_4_conv_weight = const()[name = tensor("inner_encoder_conv_module_0_sequential_4_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9290688)))]; + tensor inner_encoder_conv_module_0_sequential_5_running_var = const()[name = tensor("inner_encoder_conv_module_0_sequential_5_running_var"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9307136)))]; + tensor inner_encoder_conv_module_0_sequential_5_running_mean = const()[name = tensor("inner_encoder_conv_module_0_sequential_5_running_mean"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9308224)))]; + tensor inner_encoder_conv_module_0_sequential_5_bias = const()[name = tensor("inner_encoder_conv_module_0_sequential_5_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9309312)))]; + tensor inner_encoder_conv_module_0_sequential_5_weight = const()[name = tensor("inner_encoder_conv_module_0_sequential_5_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9310400)))]; + tensor inner_encoder_conv_module_0_sequential_7_conv_bias = const()[name = tensor("inner_encoder_conv_module_0_sequential_7_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9311488)))]; + tensor inner_encoder_conv_module_0_sequential_7_conv_weight = const()[name = tensor("inner_encoder_conv_module_0_sequential_7_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9312576)))]; + tensor inner_encoder_ffn2_0_module_sequential_0_bias = const()[name = tensor("inner_encoder_ffn2_0_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9574784)))]; + tensor inner_encoder_ffn2_0_module_sequential_0_weight = const()[name = tensor("inner_encoder_ffn2_0_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9575872)))]; + tensor inner_encoder_ffn2_0_module_sequential_1_linear_bias = const()[name = tensor("inner_encoder_ffn2_0_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9576960)))]; + tensor inner_encoder_ffn2_0_module_sequential_1_linear_weight = const()[name = tensor("inner_encoder_ffn2_0_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9581120)))]; + tensor inner_encoder_ffn2_0_module_sequential_4_linear_bias = const()[name = tensor("inner_encoder_ffn2_0_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(10629760)))]; + tensor inner_encoder_ffn2_0_module_sequential_4_linear_weight = const()[name = tensor("inner_encoder_ffn2_0_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(10630848)))]; + tensor inner_encoder_layer_norm_0_bias = const()[name = tensor("inner_encoder_layer_norm_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11679488)))]; + tensor inner_encoder_layer_norm_0_weight = const()[name = tensor("inner_encoder_layer_norm_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11680576)))]; + tensor inner_encoder_ffn1_1_module_sequential_0_bias = const()[name = tensor("inner_encoder_ffn1_1_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11681664)))]; + tensor inner_encoder_ffn1_1_module_sequential_0_weight = const()[name = tensor("inner_encoder_ffn1_1_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11682752)))]; + tensor inner_encoder_ffn1_1_module_sequential_1_linear_bias = const()[name = tensor("inner_encoder_ffn1_1_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11683840)))]; + tensor inner_encoder_ffn1_1_module_sequential_1_linear_weight = const()[name = tensor("inner_encoder_ffn1_1_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11688000)))]; + tensor inner_encoder_ffn1_1_module_sequential_4_linear_bias = const()[name = tensor("inner_encoder_ffn1_1_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(12736640)))]; + tensor inner_encoder_ffn1_1_module_sequential_4_linear_weight = const()[name = tensor("inner_encoder_ffn1_1_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(12737728)))]; + tensor inner_encoder_ret_lns_1_bias = const()[name = tensor("inner_encoder_ret_lns_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(13786368)))]; + tensor inner_encoder_ret_lns_1_weight = const()[name = tensor("inner_encoder_ret_lns_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(13787456)))]; + tensor inner_encoder_q_proj_1_bias = const()[name = tensor("inner_encoder_q_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(13788544)))]; + tensor inner_encoder_q_proj_1_weight = const()[name = tensor("inner_encoder_q_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(13789632)))]; + tensor inner_encoder_k_proj_1_bias = const()[name = tensor("inner_encoder_k_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14051840)))]; + tensor inner_encoder_k_proj_1_weight = const()[name = tensor("inner_encoder_k_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14052928)))]; + tensor inner_encoder_v_proj_1_bias = const()[name = tensor("inner_encoder_v_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14315136)))]; + tensor inner_encoder_v_proj_1_weight = const()[name = tensor("inner_encoder_v_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14316224)))]; + tensor inner_encoder_g_proj_1_bias = const()[name = tensor("inner_encoder_g_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14578432)))]; + tensor inner_encoder_g_proj_1_weight = const()[name = tensor("inner_encoder_g_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14579520)))]; + tensor inner_encoder_out_proj_1_bias = const()[name = tensor("inner_encoder_out_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14841728)))]; + tensor inner_encoder_out_proj_1_weight = const()[name = tensor("inner_encoder_out_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14842816)))]; + tensor inner_encoder_conv_module_1_sequential_0_bias = const()[name = tensor("inner_encoder_conv_module_1_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15105024)))]; + tensor inner_encoder_conv_module_1_sequential_0_weight = const()[name = tensor("inner_encoder_conv_module_1_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15106112)))]; + tensor inner_encoder_conv_module_1_sequential_2_conv_bias = const()[name = tensor("inner_encoder_conv_module_1_sequential_2_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15107200)))]; + tensor inner_encoder_conv_module_1_sequential_2_conv_weight = const()[name = tensor("inner_encoder_conv_module_1_sequential_2_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15109312)))]; + tensor inner_encoder_conv_module_1_sequential_4_conv_weight = const()[name = tensor("inner_encoder_conv_module_1_sequential_4_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15633664)))]; + tensor inner_encoder_conv_module_1_sequential_5_running_var = const()[name = tensor("inner_encoder_conv_module_1_sequential_5_running_var"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15650112)))]; + tensor inner_encoder_conv_module_1_sequential_5_running_mean = const()[name = tensor("inner_encoder_conv_module_1_sequential_5_running_mean"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15651200)))]; + tensor inner_encoder_conv_module_1_sequential_5_bias = const()[name = tensor("inner_encoder_conv_module_1_sequential_5_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15652288)))]; + tensor inner_encoder_conv_module_1_sequential_5_weight = const()[name = tensor("inner_encoder_conv_module_1_sequential_5_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15653376)))]; + tensor inner_encoder_conv_module_1_sequential_7_conv_bias = const()[name = tensor("inner_encoder_conv_module_1_sequential_7_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15654464)))]; + tensor inner_encoder_conv_module_1_sequential_7_conv_weight = const()[name = tensor("inner_encoder_conv_module_1_sequential_7_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15655552)))]; + tensor inner_encoder_ffn2_1_module_sequential_0_bias = const()[name = tensor("inner_encoder_ffn2_1_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15917760)))]; + tensor inner_encoder_ffn2_1_module_sequential_0_weight = const()[name = tensor("inner_encoder_ffn2_1_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15918848)))]; + tensor inner_encoder_ffn2_1_module_sequential_1_linear_bias = const()[name = tensor("inner_encoder_ffn2_1_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15919936)))]; + tensor inner_encoder_ffn2_1_module_sequential_1_linear_weight = const()[name = tensor("inner_encoder_ffn2_1_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15924096)))]; + tensor inner_encoder_ffn2_1_module_sequential_4_linear_bias = const()[name = tensor("inner_encoder_ffn2_1_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(16972736)))]; + tensor inner_encoder_ffn2_1_module_sequential_4_linear_weight = const()[name = tensor("inner_encoder_ffn2_1_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(16973824)))]; + tensor inner_encoder_layer_norm_1_bias = const()[name = tensor("inner_encoder_layer_norm_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18022464)))]; + tensor inner_encoder_layer_norm_1_weight = const()[name = tensor("inner_encoder_layer_norm_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18023552)))]; + tensor inner_encoder_ffn1_2_module_sequential_0_bias = const()[name = tensor("inner_encoder_ffn1_2_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18024640)))]; + tensor inner_encoder_ffn1_2_module_sequential_0_weight = const()[name = tensor("inner_encoder_ffn1_2_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18025728)))]; + tensor inner_encoder_ffn1_2_module_sequential_1_linear_bias = const()[name = tensor("inner_encoder_ffn1_2_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18026816)))]; + tensor inner_encoder_ffn1_2_module_sequential_1_linear_weight = const()[name = tensor("inner_encoder_ffn1_2_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18030976)))]; + tensor inner_encoder_ffn1_2_module_sequential_4_linear_bias = const()[name = tensor("inner_encoder_ffn1_2_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(19079616)))]; + tensor inner_encoder_ffn1_2_module_sequential_4_linear_weight = const()[name = tensor("inner_encoder_ffn1_2_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(19080704)))]; + tensor inner_encoder_ret_lns_2_bias = const()[name = tensor("inner_encoder_ret_lns_2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20129344)))]; + tensor inner_encoder_ret_lns_2_weight = const()[name = tensor("inner_encoder_ret_lns_2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20130432)))]; + tensor inner_encoder_q_proj_2_bias = const()[name = tensor("inner_encoder_q_proj_2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20131520)))]; + tensor inner_encoder_q_proj_2_weight = const()[name = tensor("inner_encoder_q_proj_2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20132608)))]; + tensor inner_encoder_k_proj_2_bias = const()[name = tensor("inner_encoder_k_proj_2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20394816)))]; + tensor inner_encoder_k_proj_2_weight = const()[name = tensor("inner_encoder_k_proj_2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20395904)))]; + tensor inner_encoder_v_proj_2_bias = const()[name = tensor("inner_encoder_v_proj_2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20658112)))]; + tensor inner_encoder_v_proj_2_weight = const()[name = tensor("inner_encoder_v_proj_2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20659200)))]; + tensor inner_encoder_g_proj_2_bias = const()[name = tensor("inner_encoder_g_proj_2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20921408)))]; + tensor inner_encoder_g_proj_2_weight = const()[name = tensor("inner_encoder_g_proj_2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20922496)))]; + tensor inner_encoder_out_proj_2_bias = const()[name = tensor("inner_encoder_out_proj_2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21184704)))]; + tensor inner_encoder_out_proj_2_weight = const()[name = tensor("inner_encoder_out_proj_2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21185792)))]; + tensor inner_encoder_conv_module_2_sequential_0_bias = const()[name = tensor("inner_encoder_conv_module_2_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21448000)))]; + tensor inner_encoder_conv_module_2_sequential_0_weight = const()[name = tensor("inner_encoder_conv_module_2_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21449088)))]; + tensor inner_encoder_conv_module_2_sequential_2_conv_bias = const()[name = tensor("inner_encoder_conv_module_2_sequential_2_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21450176)))]; + tensor inner_encoder_conv_module_2_sequential_2_conv_weight = const()[name = tensor("inner_encoder_conv_module_2_sequential_2_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21452288)))]; + tensor inner_encoder_conv_module_2_sequential_4_conv_weight = const()[name = tensor("inner_encoder_conv_module_2_sequential_4_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21976640)))]; + tensor inner_encoder_conv_module_2_sequential_5_running_var = const()[name = tensor("inner_encoder_conv_module_2_sequential_5_running_var"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21993088)))]; + tensor inner_encoder_conv_module_2_sequential_5_running_mean = const()[name = tensor("inner_encoder_conv_module_2_sequential_5_running_mean"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21994176)))]; + tensor inner_encoder_conv_module_2_sequential_5_bias = const()[name = tensor("inner_encoder_conv_module_2_sequential_5_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21995264)))]; + tensor inner_encoder_conv_module_2_sequential_5_weight = const()[name = tensor("inner_encoder_conv_module_2_sequential_5_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21996352)))]; + tensor inner_encoder_conv_module_2_sequential_7_conv_bias = const()[name = tensor("inner_encoder_conv_module_2_sequential_7_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21997440)))]; + tensor inner_encoder_conv_module_2_sequential_7_conv_weight = const()[name = tensor("inner_encoder_conv_module_2_sequential_7_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21998528)))]; + tensor inner_encoder_ffn2_2_module_sequential_0_bias = const()[name = tensor("inner_encoder_ffn2_2_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(22260736)))]; + tensor inner_encoder_ffn2_2_module_sequential_0_weight = const()[name = tensor("inner_encoder_ffn2_2_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(22261824)))]; + tensor inner_encoder_ffn2_2_module_sequential_1_linear_bias = const()[name = tensor("inner_encoder_ffn2_2_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(22262912)))]; + tensor inner_encoder_ffn2_2_module_sequential_1_linear_weight = const()[name = tensor("inner_encoder_ffn2_2_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(22267072)))]; + tensor inner_encoder_ffn2_2_module_sequential_4_linear_bias = const()[name = tensor("inner_encoder_ffn2_2_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(23315712)))]; + tensor inner_encoder_ffn2_2_module_sequential_4_linear_weight = const()[name = tensor("inner_encoder_ffn2_2_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(23316800)))]; + tensor inner_encoder_layer_norm_2_bias = const()[name = tensor("inner_encoder_layer_norm_2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(24365440)))]; + tensor inner_encoder_layer_norm_2_weight = const()[name = tensor("inner_encoder_layer_norm_2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(24366528)))]; + tensor inner_encoder_ffn1_3_module_sequential_0_bias = const()[name = tensor("inner_encoder_ffn1_3_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(24367616)))]; + tensor inner_encoder_ffn1_3_module_sequential_0_weight = const()[name = tensor("inner_encoder_ffn1_3_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(24368704)))]; + tensor inner_encoder_ffn1_3_module_sequential_1_linear_bias = const()[name = tensor("inner_encoder_ffn1_3_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(24369792)))]; + tensor inner_encoder_ffn1_3_module_sequential_1_linear_weight = const()[name = tensor("inner_encoder_ffn1_3_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(24373952)))]; + tensor inner_encoder_ffn1_3_module_sequential_4_linear_bias = const()[name = tensor("inner_encoder_ffn1_3_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(25422592)))]; + tensor inner_encoder_ffn1_3_module_sequential_4_linear_weight = const()[name = tensor("inner_encoder_ffn1_3_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(25423680)))]; + tensor inner_encoder_ret_lns_3_bias = const()[name = tensor("inner_encoder_ret_lns_3_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(26472320)))]; + tensor inner_encoder_ret_lns_3_weight = const()[name = tensor("inner_encoder_ret_lns_3_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(26473408)))]; + tensor inner_encoder_q_proj_3_bias = const()[name = tensor("inner_encoder_q_proj_3_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(26474496)))]; + tensor inner_encoder_q_proj_3_weight = const()[name = tensor("inner_encoder_q_proj_3_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(26475584)))]; + tensor inner_encoder_k_proj_3_bias = const()[name = tensor("inner_encoder_k_proj_3_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(26737792)))]; + tensor inner_encoder_k_proj_3_weight = const()[name = tensor("inner_encoder_k_proj_3_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(26738880)))]; + tensor inner_encoder_v_proj_3_bias = const()[name = tensor("inner_encoder_v_proj_3_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27001088)))]; + tensor inner_encoder_v_proj_3_weight = const()[name = tensor("inner_encoder_v_proj_3_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27002176)))]; + tensor inner_encoder_g_proj_3_bias = const()[name = tensor("inner_encoder_g_proj_3_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27264384)))]; + tensor inner_encoder_g_proj_3_weight = const()[name = tensor("inner_encoder_g_proj_3_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27265472)))]; + tensor inner_encoder_out_proj_3_bias = const()[name = tensor("inner_encoder_out_proj_3_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27527680)))]; + tensor inner_encoder_out_proj_3_weight = const()[name = tensor("inner_encoder_out_proj_3_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27528768)))]; + tensor inner_encoder_conv_module_3_sequential_0_bias = const()[name = tensor("inner_encoder_conv_module_3_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27790976)))]; + tensor inner_encoder_conv_module_3_sequential_0_weight = const()[name = tensor("inner_encoder_conv_module_3_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27792064)))]; + tensor inner_encoder_conv_module_3_sequential_2_conv_bias = const()[name = tensor("inner_encoder_conv_module_3_sequential_2_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27793152)))]; + tensor inner_encoder_conv_module_3_sequential_2_conv_weight = const()[name = tensor("inner_encoder_conv_module_3_sequential_2_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27795264)))]; + tensor inner_encoder_conv_module_3_sequential_4_conv_weight = const()[name = tensor("inner_encoder_conv_module_3_sequential_4_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28319616)))]; + tensor inner_encoder_conv_module_3_sequential_5_running_var = const()[name = tensor("inner_encoder_conv_module_3_sequential_5_running_var"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28336064)))]; + tensor inner_encoder_conv_module_3_sequential_5_running_mean = const()[name = tensor("inner_encoder_conv_module_3_sequential_5_running_mean"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28337152)))]; + tensor inner_encoder_conv_module_3_sequential_5_bias = const()[name = tensor("inner_encoder_conv_module_3_sequential_5_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28338240)))]; + tensor inner_encoder_conv_module_3_sequential_5_weight = const()[name = tensor("inner_encoder_conv_module_3_sequential_5_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28339328)))]; + tensor inner_encoder_conv_module_3_sequential_7_conv_bias = const()[name = tensor("inner_encoder_conv_module_3_sequential_7_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28340416)))]; + tensor inner_encoder_conv_module_3_sequential_7_conv_weight = const()[name = tensor("inner_encoder_conv_module_3_sequential_7_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28341504)))]; + tensor inner_encoder_ffn2_3_module_sequential_0_bias = const()[name = tensor("inner_encoder_ffn2_3_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28603712)))]; + tensor inner_encoder_ffn2_3_module_sequential_0_weight = const()[name = tensor("inner_encoder_ffn2_3_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28604800)))]; + tensor inner_encoder_ffn2_3_module_sequential_1_linear_bias = const()[name = tensor("inner_encoder_ffn2_3_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28605888)))]; + tensor inner_encoder_ffn2_3_module_sequential_1_linear_weight = const()[name = tensor("inner_encoder_ffn2_3_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28610048)))]; + tensor inner_encoder_ffn2_3_module_sequential_4_linear_bias = const()[name = tensor("inner_encoder_ffn2_3_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(29658688)))]; + tensor inner_encoder_ffn2_3_module_sequential_4_linear_weight = const()[name = tensor("inner_encoder_ffn2_3_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(29659776)))]; + tensor inner_encoder_layer_norm_3_bias = const()[name = tensor("inner_encoder_layer_norm_3_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(30708416)))]; + tensor inner_encoder_layer_norm_3_weight = const()[name = tensor("inner_encoder_layer_norm_3_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(30709504)))]; + tensor inner_decoder__causal_mask = const()[name = tensor("inner_decoder__causal_mask"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(30710592)))]; + tensor inner_decoder_convert_bias = const()[name = tensor("inner_decoder_convert_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(30710720)))]; + tensor inner_decoder_convert_weight = const()[name = tensor("inner_decoder_convert_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(30711808)))]; + tensor inner_decoder_q_proj_0_bias = const()[name = tensor("inner_decoder_q_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(31236160)))]; + tensor inner_decoder_q_proj_0_weight = const()[name = tensor("inner_decoder_q_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(31237248)))]; + tensor inner_decoder_k_proj_0_bias = const()[name = tensor("inner_decoder_k_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(31499456)))]; + tensor inner_decoder_k_proj_0_weight = const()[name = tensor("inner_decoder_k_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(31500544)))]; + tensor inner_decoder_v_proj_0_bias = const()[name = tensor("inner_decoder_v_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(31762752)))]; + tensor inner_decoder_v_proj_0_weight = const()[name = tensor("inner_decoder_v_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(31763840)))]; + tensor inner_decoder_g_proj_0_bias = const()[name = tensor("inner_decoder_g_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32026048)))]; + tensor inner_decoder_g_proj_0_weight = const()[name = tensor("inner_decoder_g_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32027136)))]; + tensor inner_decoder_out_proj_0_bias = const()[name = tensor("inner_decoder_out_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32289344)))]; + tensor inner_decoder_out_proj_0_weight = const()[name = tensor("inner_decoder_out_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32290432)))]; + tensor inner_decoder_norm11_0_bias = const()[name = tensor("inner_decoder_norm11_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32552640)))]; + tensor inner_decoder_norm11_0_weight = const()[name = tensor("inner_decoder_norm11_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32553728)))]; + tensor inner_decoder_self_attn2_0_out_proj_bias = const()[name = tensor("inner_decoder_self_attn2_0_out_proj_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32554816)))]; + tensor inner_decoder_self_attn2_0_out_proj_weight = const()[name = tensor("inner_decoder_self_attn2_0_out_proj_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32555904)))]; + tensor inner_decoder_self_attn2_0_in_proj_bias = const()[name = tensor("inner_decoder_self_attn2_0_in_proj_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32818112)))]; + tensor inner_decoder_self_attn2_0_in_proj_weight = const()[name = tensor("inner_decoder_self_attn2_0_in_proj_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32821248)))]; + tensor inner_decoder_norm21_0_bias = const()[name = tensor("inner_decoder_norm21_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(33607744)))]; + tensor inner_decoder_norm21_0_weight = const()[name = tensor("inner_decoder_norm21_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(33608832)))]; + tensor inner_decoder_linear1_0_bias = const()[name = tensor("inner_decoder_linear1_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(33609920)))]; + tensor inner_decoder_linear1_0_weight = const()[name = tensor("inner_decoder_linear1_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(33618176)))]; + tensor inner_decoder_linear2_0_bias = const()[name = tensor("inner_decoder_linear2_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(35715392)))]; + tensor inner_decoder_linear2_0_weight = const()[name = tensor("inner_decoder_linear2_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(35716480)))]; + tensor inner_decoder_norm22_0_bias = const()[name = tensor("inner_decoder_norm22_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(37813696)))]; + tensor inner_decoder_norm22_0_weight = const()[name = tensor("inner_decoder_norm22_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(37814784)))]; + tensor inner_decoder_q_proj_1_bias = const()[name = tensor("inner_decoder_q_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(37815872)))]; + tensor inner_decoder_q_proj_1_weight = const()[name = tensor("inner_decoder_q_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(37816960)))]; + tensor inner_decoder_k_proj_1_bias = const()[name = tensor("inner_decoder_k_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38079168)))]; + tensor inner_decoder_k_proj_1_weight = const()[name = tensor("inner_decoder_k_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38080256)))]; + tensor inner_decoder_v_proj_1_bias = const()[name = tensor("inner_decoder_v_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38342464)))]; + tensor inner_decoder_v_proj_1_weight = const()[name = tensor("inner_decoder_v_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38343552)))]; + tensor inner_decoder_g_proj_1_bias = const()[name = tensor("inner_decoder_g_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38605760)))]; + tensor inner_decoder_g_proj_1_weight = const()[name = tensor("inner_decoder_g_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38606848)))]; + tensor inner_decoder_out_proj_1_bias = const()[name = tensor("inner_decoder_out_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38869056)))]; + tensor inner_decoder_out_proj_1_weight = const()[name = tensor("inner_decoder_out_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38870144)))]; + tensor inner_decoder_norm11_1_bias = const()[name = tensor("inner_decoder_norm11_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(39132352)))]; + tensor inner_decoder_norm11_1_weight = const()[name = tensor("inner_decoder_norm11_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(39133440)))]; + tensor inner_decoder_self_attn2_1_out_proj_bias = const()[name = tensor("inner_decoder_self_attn2_1_out_proj_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(39134528)))]; + tensor inner_decoder_self_attn2_1_out_proj_weight = const()[name = tensor("inner_decoder_self_attn2_1_out_proj_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(39135616)))]; + tensor inner_decoder_self_attn2_1_in_proj_bias = const()[name = tensor("inner_decoder_self_attn2_1_in_proj_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(39397824)))]; + tensor inner_decoder_self_attn2_1_in_proj_weight = const()[name = tensor("inner_decoder_self_attn2_1_in_proj_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(39400960)))]; + tensor inner_decoder_norm21_1_bias = const()[name = tensor("inner_decoder_norm21_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(40187456)))]; + tensor inner_decoder_norm21_1_weight = const()[name = tensor("inner_decoder_norm21_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(40188544)))]; + tensor inner_decoder_linear1_1_bias = const()[name = tensor("inner_decoder_linear1_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(40189632)))]; + tensor inner_decoder_linear1_1_weight = const()[name = tensor("inner_decoder_linear1_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(40197888)))]; + tensor inner_decoder_linear2_1_bias = const()[name = tensor("inner_decoder_linear2_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(42295104)))]; + tensor inner_decoder_linear2_1_weight = const()[name = tensor("inner_decoder_linear2_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(42296192)))]; + tensor inner_decoder_norm22_1_bias = const()[name = tensor("inner_decoder_norm22_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(44393408)))]; + tensor inner_decoder_norm22_1_weight = const()[name = tensor("inner_decoder_norm22_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(44394496)))]; + tensor var_19_begin_0 = const()[name = tensor("op_19_begin_0"), val = tensor([0, 0, 0])]; + tensor var_19_end_0 = const()[name = tensor("op_19_end_0"), val = tensor([1, 15, 23])]; + tensor var_19_end_mask_0 = const()[name = tensor("op_19_end_mask_0"), val = tensor([true, false, true])]; + tensor var_19 = slice_by_index(begin = var_19_begin_0, end = var_19_end_0, end_mask = var_19_end_mask_0, x = features)[name = tensor("op_19")]; + tensor var_29_begin_0 = const()[name = tensor("op_29_begin_0"), val = tensor([0, 10, 0])]; + tensor var_29_end_0 = const()[name = tensor("op_29_end_0"), val = tensor([1, 25, 23])]; + tensor var_29_end_mask_0 = const()[name = tensor("op_29_end_mask_0"), val = tensor([true, false, true])]; + tensor var_29 = slice_by_index(begin = var_29_begin_0, end = var_29_end_0, end_mask = var_29_end_mask_0, x = features)[name = tensor("op_29")]; + tensor var_39_begin_0 = const()[name = tensor("op_39_begin_0"), val = tensor([0, 20, 0])]; + tensor var_39_end_0 = const()[name = tensor("op_39_end_0"), val = tensor([1, 35, 23])]; + tensor var_39_end_mask_0 = const()[name = tensor("op_39_end_mask_0"), val = tensor([true, false, true])]; + tensor var_39 = slice_by_index(begin = var_39_begin_0, end = var_39_end_0, end_mask = var_39_end_mask_0, x = features)[name = tensor("op_39")]; + tensor var_49_begin_0 = const()[name = tensor("op_49_begin_0"), val = tensor([0, 30, 0])]; + tensor var_49_end_0 = const()[name = tensor("op_49_end_0"), val = tensor([1, 1, 23])]; + tensor var_49_end_mask_0 = const()[name = tensor("op_49_end_mask_0"), val = tensor([true, true, true])]; + tensor var_49 = slice_by_index(begin = var_49_begin_0, end = var_49_end_0, end_mask = var_49_end_mask_0, x = features)[name = tensor("op_49")]; + tensor stacked_axis_0 = const()[name = tensor("stacked_axis_0"), val = tensor(1)]; + tensor stacked = stack(axis = stacked_axis_0, values = (var_19, var_29, var_39, var_49))[name = tensor("stacked")]; + tensor var_56 = const()[name = tensor("op_56"), val = tensor([1, 4, 345])]; + tensor input_1 = reshape(shape = var_56, x = stacked)[name = tensor("input_1")]; + tensor var_59 = const()[name = tensor("op_59"), val = tensor(0x1p+0)]; + tensor var_65 = const()[name = tensor("op_65"), val = tensor(true)]; + tensor var_66 = const()[name = tensor("op_66"), val = tensor(0x1.4f8b58p-17)]; + tensor var_69 = const()[name = tensor("op_69"), val = tensor(0)]; + tensor var_71 = const()[name = tensor("op_71"), val = tensor(2)]; + tensor var_72 = const()[name = tensor("op_72"), val = tensor(-1)]; + tensor var_74 = const()[name = tensor("op_74"), val = tensor(0x1.0c6f7ap-20)]; + tensor var_79 = const()[name = tensor("op_79"), val = tensor(0x1.5798eep-27)]; + tensor var_82 = const()[name = tensor("op_82"), val = tensor(1)]; + tensor input_3 = linear(bias = inner_encoder_input_projection_linear_bias, weight = inner_encoder_input_projection_linear_weight, x = input_1)[name = tensor("linear_0")]; tensor input_5_axes_0 = const()[name = tensor("input_5_axes_0"), val = tensor([-1])]; - tensor input_5 = layer_norm(axes = input_5_axes_0, beta = encoder_ffn1_0_module_sequential_0_bias, epsilon = var_28, gamma = encoder_ffn1_0_module_sequential_0_weight, x = input_3)[name = tensor("input_5")]; - tensor inputs_1 = linear(bias = encoder_ffn1_0_module_sequential_1_linear_bias, weight = encoder_ffn1_0_module_sequential_1_linear_weight, x = input_5)[name = tensor("linear_1")]; - tensor input_7 = silu(x = inputs_1)[name = tensor("input_7")]; - tensor input_11 = linear(bias = encoder_ffn1_0_module_sequential_4_linear_bias, weight = encoder_ffn1_0_module_sequential_4_linear_weight, x = input_7)[name = tensor("linear_2")]; - tensor var_147 = const()[name = tensor("op_147"), val = tensor(0x1p-1)]; - tensor var_148 = mul(x = input_11, y = var_147)[name = tensor("op_148")]; - tensor input_13 = add(x = var_148, y = input_3)[name = tensor("input_13")]; + tensor input_5 = layer_norm(axes = input_5_axes_0, beta = inner_encoder_pre_layer_norm_bias, epsilon = var_66, gamma = inner_encoder_pre_layer_norm_weight, x = input_3)[name = tensor("input_5")]; + tensor input_7_axes_0 = const()[name = tensor("input_7_axes_0"), val = tensor([-1])]; + tensor input_7 = layer_norm(axes = input_7_axes_0, beta = inner_encoder_ffn1_0_module_sequential_0_bias, epsilon = var_66, gamma = inner_encoder_ffn1_0_module_sequential_0_weight, x = input_5)[name = tensor("input_7")]; + tensor inputs_1 = linear(bias = inner_encoder_ffn1_0_module_sequential_1_linear_bias, weight = inner_encoder_ffn1_0_module_sequential_1_linear_weight, x = input_7)[name = tensor("linear_1")]; + tensor input_9 = silu(x = inputs_1)[name = tensor("input_9")]; + tensor input_13 = linear(bias = inner_encoder_ffn1_0_module_sequential_4_linear_bias, weight = inner_encoder_ffn1_0_module_sequential_4_linear_weight, x = input_9)[name = tensor("linear_2")]; + tensor var_203 = const()[name = tensor("op_203"), val = tensor(0x1p-1)]; + tensor var_204 = mul(x = input_13, y = var_203)[name = tensor("op_204")]; + tensor input_15 = add(x = var_204, y = input_5)[name = tensor("input_15")]; tensor x_1_axes_0 = const()[name = tensor("x_1_axes_0"), val = tensor([-1])]; - tensor x_1 = layer_norm(axes = x_1_axes_0, beta = encoder_ret_lns_0_bias, epsilon = var_28, gamma = encoder_ret_lns_0_weight, x = input_13)[name = tensor("x_1")]; + tensor x_1 = layer_norm(axes = x_1_axes_0, beta = inner_encoder_ret_lns_0_bias, epsilon = var_66, gamma = inner_encoder_ret_lns_0_weight, x = input_15)[name = tensor("x_1")]; tensor prev_kv_1_begin_0 = const()[name = tensor("prev_kv_1_begin_0"), val = tensor([0, 0, 0, 0, 0])]; tensor prev_kv_1_end_0 = const()[name = tensor("prev_kv_1_end_0"), val = tensor([1, 1, 4, 64, 64])]; tensor prev_kv_1_end_mask_0 = const()[name = tensor("prev_kv_1_end_mask_0"), val = tensor([false, true, true, true, true])]; @@ -239,173 +261,173 @@ program(1.0) tensor prev_scale_1_end_mask_0 = const()[name = tensor("prev_scale_1_end_mask_0"), val = tensor([false, true])]; tensor prev_scale_1_squeeze_mask_0 = const()[name = tensor("prev_scale_1_squeeze_mask_0"), val = tensor([true, false])]; tensor prev_scale_1 = slice_by_index(begin = prev_scale_1_begin_0, end = prev_scale_1_end_0, end_mask = prev_scale_1_end_mask_0, squeeze_mask = prev_scale_1_squeeze_mask_0, x = enc_scale)[name = tensor("prev_scale_1")]; - tensor var_162 = linear(bias = encoder_q_proj_0_bias, weight = encoder_q_proj_0_weight, x = x_1)[name = tensor("linear_3")]; - tensor var_163 = const()[name = tensor("op_163"), val = tensor([1, 4, 4, 64])]; - tensor var_164 = reshape(shape = var_163, x = var_162)[name = tensor("op_164")]; + tensor var_218 = linear(bias = inner_encoder_q_proj_0_bias, weight = inner_encoder_q_proj_0_weight, x = x_1)[name = tensor("linear_3")]; + tensor var_219 = const()[name = tensor("op_219"), val = tensor([1, 4, 4, 64])]; + tensor var_220 = reshape(shape = var_219, x = var_218)[name = tensor("op_220")]; tensor q_1_perm_0 = const()[name = tensor("q_1_perm_0"), val = tensor([0, 2, 1, 3])]; - tensor var_168 = linear(bias = encoder_k_proj_0_bias, weight = encoder_k_proj_0_weight, x = x_1)[name = tensor("linear_4")]; - tensor var_169 = const()[name = tensor("op_169"), val = tensor(0x1p-3)]; - tensor var_170 = mul(x = var_168, y = var_169)[name = tensor("op_170")]; - tensor var_171 = const()[name = tensor("op_171"), val = tensor([1, 4, 4, 64])]; - tensor var_172 = reshape(shape = var_171, x = var_170)[name = tensor("op_172")]; + tensor var_224 = linear(bias = inner_encoder_k_proj_0_bias, weight = inner_encoder_k_proj_0_weight, x = x_1)[name = tensor("linear_4")]; + tensor var_225 = const()[name = tensor("op_225"), val = tensor(0x1p-3)]; + tensor var_226 = mul(x = var_224, y = var_225)[name = tensor("op_226")]; + tensor var_227 = const()[name = tensor("op_227"), val = tensor([1, 4, 4, 64])]; + tensor var_228 = reshape(shape = var_227, x = var_226)[name = tensor("op_228")]; tensor k_1_perm_0 = const()[name = tensor("k_1_perm_0"), val = tensor([0, 2, 1, 3])]; - tensor var_176 = linear(bias = encoder_v_proj_0_bias, weight = encoder_v_proj_0_weight, x = x_1)[name = tensor("linear_5")]; - tensor var_177 = const()[name = tensor("op_177"), val = tensor([1, 4, 4, 64])]; - tensor var_178 = reshape(shape = var_177, x = var_176)[name = tensor("op_178")]; + tensor var_232 = linear(bias = inner_encoder_v_proj_0_bias, weight = inner_encoder_v_proj_0_weight, x = x_1)[name = tensor("linear_5")]; + tensor var_233 = const()[name = tensor("op_233"), val = tensor([1, 4, 4, 64])]; + tensor var_234 = reshape(shape = var_233, x = var_232)[name = tensor("op_234")]; tensor v_1_perm_0 = const()[name = tensor("v_1_perm_0"), val = tensor([0, 2, 1, 3])]; - tensor input_17 = linear(bias = encoder_g_proj_0_bias, weight = encoder_g_proj_0_weight, x = x_1)[name = tensor("linear_6")]; + tensor input_19 = linear(bias = inner_encoder_g_proj_0_bias, weight = inner_encoder_g_proj_0_weight, x = x_1)[name = tensor("linear_6")]; tensor sqrt_s0_1 = sqrt(x = prev_scale_1)[name = tensor("sqrt_s0_1")]; - tensor s_t_1 = add(x = prev_scale_1, y = encoder__t_index)[name = tensor("s_t_1")]; + tensor s_t_1 = add(x = prev_scale_1, y = inner_encoder__t_index)[name = tensor("s_t_1")]; tensor sqrt_s_t_1 = sqrt(x = s_t_1)[name = tensor("sqrt_s_t_1")]; tensor qk_1_transpose_x_1 = const()[name = tensor("qk_1_transpose_x_1"), val = tensor(false)]; tensor qk_1_transpose_y_1 = const()[name = tensor("qk_1_transpose_y_1"), val = tensor(true)]; - tensor k_1 = transpose(perm = k_1_perm_0, x = var_172)[name = tensor("transpose_57")]; - tensor q_1 = transpose(perm = q_1_perm_0, x = var_164)[name = tensor("transpose_58")]; + tensor k_1 = transpose(perm = k_1_perm_0, x = var_228)[name = tensor("transpose_57")]; + tensor q_1 = transpose(perm = q_1_perm_0, x = var_220)[name = tensor("transpose_58")]; tensor qk_1 = matmul(transpose_x = qk_1_transpose_x_1, transpose_y = qk_1_transpose_y_1, x = q_1, y = k_1)[name = tensor("qk_1")]; - tensor var_188 = const()[name = tensor("op_188"), val = tensor([4, 1])]; - tensor var_189 = reshape(shape = var_188, x = sqrt_s_t_1)[name = tensor("op_189")]; - tensor M_1 = real_div(x = encoder__causal_mask, y = var_189)[name = tensor("M_1")]; - tensor var_191 = mul(x = qk_1, y = M_1)[name = tensor("op_191")]; + tensor var_244 = const()[name = tensor("op_244"), val = tensor([4, 1])]; + tensor var_245 = reshape(shape = var_244, x = sqrt_s_t_1)[name = tensor("op_245")]; + tensor M_1 = real_div(x = inner_encoder__causal_mask, y = var_245)[name = tensor("M_1")]; + tensor var_247 = mul(x = qk_1, y = M_1)[name = tensor("op_247")]; tensor inner_1_transpose_x_0 = const()[name = tensor("inner_1_transpose_x_0"), val = tensor(false)]; tensor inner_1_transpose_y_0 = const()[name = tensor("inner_1_transpose_y_0"), val = tensor(false)]; - tensor v_1 = transpose(perm = v_1_perm_0, x = var_178)[name = tensor("transpose_56")]; - tensor inner_1 = matmul(transpose_x = inner_1_transpose_x_0, transpose_y = inner_1_transpose_y_0, x = var_191, y = v_1)[name = tensor("inner_1")]; - tensor var_193_transpose_x_0 = const()[name = tensor("op_193_transpose_x_0"), val = tensor(false)]; - tensor var_193_transpose_y_0 = const()[name = tensor("op_193_transpose_y_0"), val = tensor(false)]; - tensor var_193 = matmul(transpose_x = var_193_transpose_x_0, transpose_y = var_193_transpose_y_0, x = q_1, y = prev_kv_1)[name = tensor("op_193")]; - tensor var_194 = real_div(x = sqrt_s0_1, y = sqrt_s_t_1)[name = tensor("op_194")]; - tensor var_195 = const()[name = tensor("op_195"), val = tensor([1, 1, 4, 1])]; - tensor var_196 = reshape(shape = var_195, x = var_194)[name = tensor("op_196")]; - tensor cross_1 = mul(x = var_193, y = var_196)[name = tensor("cross_1")]; + tensor v_1 = transpose(perm = v_1_perm_0, x = var_234)[name = tensor("transpose_56")]; + tensor inner_1 = matmul(transpose_x = inner_1_transpose_x_0, transpose_y = inner_1_transpose_y_0, x = var_247, y = v_1)[name = tensor("inner_1")]; + tensor var_249_transpose_x_0 = const()[name = tensor("op_249_transpose_x_0"), val = tensor(false)]; + tensor var_249_transpose_y_0 = const()[name = tensor("op_249_transpose_y_0"), val = tensor(false)]; + tensor var_249 = matmul(transpose_x = var_249_transpose_x_0, transpose_y = var_249_transpose_y_0, x = q_1, y = prev_kv_1)[name = tensor("op_249")]; + tensor var_250 = real_div(x = sqrt_s0_1, y = sqrt_s_t_1)[name = tensor("op_250")]; + tensor var_251 = const()[name = tensor("op_251"), val = tensor([1, 1, 4, 1])]; + tensor var_252 = reshape(shape = var_251, x = var_250)[name = tensor("op_252")]; + tensor cross_1 = mul(x = var_249, y = var_252)[name = tensor("cross_1")]; tensor out_1 = add(x = inner_1, y = cross_1)[name = tensor("out_1")]; - tensor var_199 = mul(x = prev_kv_1, y = sqrt_s0_1)[name = tensor("op_199")]; - tensor var_201_transpose_x_1 = const()[name = tensor("op_201_transpose_x_1"), val = tensor(true)]; - tensor var_201_transpose_y_1 = const()[name = tensor("op_201_transpose_y_1"), val = tensor(false)]; - tensor var_201 = matmul(transpose_x = var_201_transpose_x_1, transpose_y = var_201_transpose_y_1, x = k_1, y = v_1)[name = tensor("op_201")]; - tensor new_kv_unnorm_1 = add(x = var_199, y = var_201)[name = tensor("new_kv_unnorm_1")]; - tensor var_203 = const()[name = tensor("op_203"), val = tensor(0x1p+2)]; - tensor new_scale_1 = add(x = prev_scale_1, y = var_203)[name = tensor("new_scale_1")]; - tensor var_205 = sqrt(x = new_scale_1)[name = tensor("op_205")]; - tensor var_206 = real_div(x = new_kv_unnorm_1, y = var_205)[name = tensor("op_206")]; - tensor var_207_perm_0 = const()[name = tensor("op_207_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_255 = mul(x = prev_kv_1, y = sqrt_s0_1)[name = tensor("op_255")]; + tensor var_257_transpose_x_1 = const()[name = tensor("op_257_transpose_x_1"), val = tensor(true)]; + tensor var_257_transpose_y_1 = const()[name = tensor("op_257_transpose_y_1"), val = tensor(false)]; + tensor var_257 = matmul(transpose_x = var_257_transpose_x_1, transpose_y = var_257_transpose_y_1, x = k_1, y = v_1)[name = tensor("op_257")]; + tensor new_kv_unnorm_1 = add(x = var_255, y = var_257)[name = tensor("new_kv_unnorm_1")]; + tensor var_259 = const()[name = tensor("op_259"), val = tensor(0x1p+2)]; + tensor new_scale_1 = add(x = prev_scale_1, y = var_259)[name = tensor("new_scale_1")]; + tensor var_261 = sqrt(x = new_scale_1)[name = tensor("op_261")]; + tensor var_262 = real_div(x = new_kv_unnorm_1, y = var_261)[name = tensor("op_262")]; + tensor var_263_perm_0 = const()[name = tensor("op_263_perm_0"), val = tensor([0, 2, 1, 3])]; tensor out_3_axes_0 = const()[name = tensor("out_3_axes_0"), val = tensor([-1])]; - tensor var_207 = transpose(perm = var_207_perm_0, x = out_1)[name = tensor("transpose_55")]; - tensor out_3 = layer_norm(axes = out_3_axes_0, epsilon = var_17, x = var_207)[name = tensor("out_3")]; - tensor var_211 = const()[name = tensor("op_211"), val = tensor([1, 4, 256])]; - tensor out_5 = reshape(shape = var_211, x = out_3)[name = tensor("out_5")]; - tensor var_213 = silu(x = input_17)[name = tensor("op_213")]; - tensor input_19 = mul(x = var_213, y = out_5)[name = tensor("input_19")]; - tensor ret_out_1 = linear(bias = encoder_out_proj_0_bias, weight = encoder_out_proj_0_weight, x = input_19)[name = tensor("linear_7")]; - tensor x_3 = add(x = input_13, y = ret_out_1)[name = tensor("x_3")]; + tensor var_263 = transpose(perm = var_263_perm_0, x = out_1)[name = tensor("transpose_55")]; + tensor out_3 = layer_norm(axes = out_3_axes_0, epsilon = var_74, x = var_263)[name = tensor("out_3")]; + tensor var_267 = const()[name = tensor("op_267"), val = tensor([1, 4, 256])]; + tensor out_5 = reshape(shape = var_267, x = out_3)[name = tensor("out_5")]; + tensor var_269 = silu(x = input_19)[name = tensor("op_269")]; + tensor input_21 = mul(x = var_269, y = out_5)[name = tensor("input_21")]; + tensor ret_out_1 = linear(bias = inner_encoder_out_proj_0_bias, weight = inner_encoder_out_proj_0_weight, x = input_21)[name = tensor("linear_7")]; + tensor x_3 = add(x = input_15, y = ret_out_1)[name = tensor("x_3")]; tensor window_1_begin_0 = const()[name = tensor("window_1_begin_0"), val = tensor([0, 0, 0, 0])]; tensor window_1_end_0 = const()[name = tensor("window_1_end_0"), val = tensor([1, 1, 16, 256])]; tensor window_1_end_mask_0 = const()[name = tensor("window_1_end_mask_0"), val = tensor([false, true, true, true])]; tensor window_1_squeeze_mask_0 = const()[name = tensor("window_1_squeeze_mask_0"), val = tensor([true, false, false, false])]; tensor window_1 = slice_by_index(begin = window_1_begin_0, end = window_1_end_0, end_mask = window_1_end_mask_0, squeeze_mask = window_1_squeeze_mask_0, x = enc_conv_cache)[name = tensor("window_1")]; - tensor var_221_begin_0 = const()[name = tensor("op_221_begin_0"), val = tensor([0, 0, 0])]; - tensor var_221_end_0 = const()[name = tensor("op_221_end_0"), val = tensor([1, 1, 256])]; - tensor var_221_end_mask_0 = const()[name = tensor("op_221_end_mask_0"), val = tensor([true, false, true])]; - tensor var_221 = slice_by_index(begin = var_221_begin_0, end = var_221_end_0, end_mask = var_221_end_mask_0, x = x_3)[name = tensor("op_221")]; - tensor var_224_begin_0 = const()[name = tensor("op_224_begin_0"), val = tensor([0, 1, 0])]; - tensor var_224_end_0 = const()[name = tensor("op_224_end_0"), val = tensor([1, 16, 256])]; - tensor var_224_end_mask_0 = const()[name = tensor("op_224_end_mask_0"), val = tensor([true, true, true])]; - tensor var_224 = slice_by_index(begin = var_224_begin_0, end = var_224_end_0, end_mask = var_224_end_mask_0, x = window_1)[name = tensor("op_224")]; + tensor var_277_begin_0 = const()[name = tensor("op_277_begin_0"), val = tensor([0, 0, 0])]; + tensor var_277_end_0 = const()[name = tensor("op_277_end_0"), val = tensor([1, 1, 256])]; + tensor var_277_end_mask_0 = const()[name = tensor("op_277_end_mask_0"), val = tensor([true, false, true])]; + tensor var_277 = slice_by_index(begin = var_277_begin_0, end = var_277_end_0, end_mask = var_277_end_mask_0, x = x_3)[name = tensor("op_277")]; + tensor var_280_begin_0 = const()[name = tensor("op_280_begin_0"), val = tensor([0, 1, 0])]; + tensor var_280_end_0 = const()[name = tensor("op_280_end_0"), val = tensor([1, 16, 256])]; + tensor var_280_end_mask_0 = const()[name = tensor("op_280_end_mask_0"), val = tensor([true, true, true])]; + tensor var_280 = slice_by_index(begin = var_280_begin_0, end = var_280_end_0, end_mask = var_280_end_mask_0, x = window_1)[name = tensor("op_280")]; tensor window_3_interleave_0 = const()[name = tensor("window_3_interleave_0"), val = tensor(false)]; - tensor window_3 = concat(axis = var_26, interleave = window_3_interleave_0, values = (var_224, var_221))[name = tensor("window_3")]; - tensor var_229_begin_0 = const()[name = tensor("op_229_begin_0"), val = tensor([0, 1, 0])]; - tensor var_229_end_0 = const()[name = tensor("op_229_end_0"), val = tensor([1, 2, 256])]; - tensor var_229_end_mask_0 = const()[name = tensor("op_229_end_mask_0"), val = tensor([true, false, true])]; - tensor var_229 = slice_by_index(begin = var_229_begin_0, end = var_229_end_0, end_mask = var_229_end_mask_0, x = x_3)[name = tensor("op_229")]; - tensor var_232_begin_0 = const()[name = tensor("op_232_begin_0"), val = tensor([0, 1, 0])]; - tensor var_232_end_0 = const()[name = tensor("op_232_end_0"), val = tensor([1, 16, 256])]; - tensor var_232_end_mask_0 = const()[name = tensor("op_232_end_mask_0"), val = tensor([true, true, true])]; - tensor var_232 = slice_by_index(begin = var_232_begin_0, end = var_232_end_0, end_mask = var_232_end_mask_0, x = window_3)[name = tensor("op_232")]; + tensor window_3 = concat(axis = var_82, interleave = window_3_interleave_0, values = (var_280, var_277))[name = tensor("window_3")]; + tensor var_285_begin_0 = const()[name = tensor("op_285_begin_0"), val = tensor([0, 1, 0])]; + tensor var_285_end_0 = const()[name = tensor("op_285_end_0"), val = tensor([1, 2, 256])]; + tensor var_285_end_mask_0 = const()[name = tensor("op_285_end_mask_0"), val = tensor([true, false, true])]; + tensor var_285 = slice_by_index(begin = var_285_begin_0, end = var_285_end_0, end_mask = var_285_end_mask_0, x = x_3)[name = tensor("op_285")]; + tensor var_288_begin_0 = const()[name = tensor("op_288_begin_0"), val = tensor([0, 1, 0])]; + tensor var_288_end_0 = const()[name = tensor("op_288_end_0"), val = tensor([1, 16, 256])]; + tensor var_288_end_mask_0 = const()[name = tensor("op_288_end_mask_0"), val = tensor([true, true, true])]; + tensor var_288 = slice_by_index(begin = var_288_begin_0, end = var_288_end_0, end_mask = var_288_end_mask_0, x = window_3)[name = tensor("op_288")]; tensor window_5_interleave_0 = const()[name = tensor("window_5_interleave_0"), val = tensor(false)]; - tensor window_5 = concat(axis = var_26, interleave = window_5_interleave_0, values = (var_232, var_229))[name = tensor("window_5")]; - tensor var_237_begin_0 = const()[name = tensor("op_237_begin_0"), val = tensor([0, 2, 0])]; - tensor var_237_end_0 = const()[name = tensor("op_237_end_0"), val = tensor([1, 3, 256])]; - tensor var_237_end_mask_0 = const()[name = tensor("op_237_end_mask_0"), val = tensor([true, false, true])]; - tensor var_237 = slice_by_index(begin = var_237_begin_0, end = var_237_end_0, end_mask = var_237_end_mask_0, x = x_3)[name = tensor("op_237")]; - tensor var_240_begin_0 = const()[name = tensor("op_240_begin_0"), val = tensor([0, 1, 0])]; - tensor var_240_end_0 = const()[name = tensor("op_240_end_0"), val = tensor([1, 16, 256])]; - tensor var_240_end_mask_0 = const()[name = tensor("op_240_end_mask_0"), val = tensor([true, true, true])]; - tensor var_240 = slice_by_index(begin = var_240_begin_0, end = var_240_end_0, end_mask = var_240_end_mask_0, x = window_5)[name = tensor("op_240")]; + tensor window_5 = concat(axis = var_82, interleave = window_5_interleave_0, values = (var_288, var_285))[name = tensor("window_5")]; + tensor var_293_begin_0 = const()[name = tensor("op_293_begin_0"), val = tensor([0, 2, 0])]; + tensor var_293_end_0 = const()[name = tensor("op_293_end_0"), val = tensor([1, 3, 256])]; + tensor var_293_end_mask_0 = const()[name = tensor("op_293_end_mask_0"), val = tensor([true, false, true])]; + tensor var_293 = slice_by_index(begin = var_293_begin_0, end = var_293_end_0, end_mask = var_293_end_mask_0, x = x_3)[name = tensor("op_293")]; + tensor var_296_begin_0 = const()[name = tensor("op_296_begin_0"), val = tensor([0, 1, 0])]; + tensor var_296_end_0 = const()[name = tensor("op_296_end_0"), val = tensor([1, 16, 256])]; + tensor var_296_end_mask_0 = const()[name = tensor("op_296_end_mask_0"), val = tensor([true, true, true])]; + tensor var_296 = slice_by_index(begin = var_296_begin_0, end = var_296_end_0, end_mask = var_296_end_mask_0, x = window_5)[name = tensor("op_296")]; tensor window_7_interleave_0 = const()[name = tensor("window_7_interleave_0"), val = tensor(false)]; - tensor window_7 = concat(axis = var_26, interleave = window_7_interleave_0, values = (var_240, var_237))[name = tensor("window_7")]; - tensor var_245_begin_0 = const()[name = tensor("op_245_begin_0"), val = tensor([0, 3, 0])]; - tensor var_245_end_0 = const()[name = tensor("op_245_end_0"), val = tensor([1, 1, 256])]; - tensor var_245_end_mask_0 = const()[name = tensor("op_245_end_mask_0"), val = tensor([true, true, true])]; - tensor var_245 = slice_by_index(begin = var_245_begin_0, end = var_245_end_0, end_mask = var_245_end_mask_0, x = x_3)[name = tensor("op_245")]; - tensor var_248_begin_0 = const()[name = tensor("op_248_begin_0"), val = tensor([0, 1, 0])]; - tensor var_248_end_0 = const()[name = tensor("op_248_end_0"), val = tensor([1, 16, 256])]; - tensor var_248_end_mask_0 = const()[name = tensor("op_248_end_mask_0"), val = tensor([true, true, true])]; - tensor var_248 = slice_by_index(begin = var_248_begin_0, end = var_248_end_0, end_mask = var_248_end_mask_0, x = window_7)[name = tensor("op_248")]; + tensor window_7 = concat(axis = var_82, interleave = window_7_interleave_0, values = (var_296, var_293))[name = tensor("window_7")]; + tensor var_301_begin_0 = const()[name = tensor("op_301_begin_0"), val = tensor([0, 3, 0])]; + tensor var_301_end_0 = const()[name = tensor("op_301_end_0"), val = tensor([1, 1, 256])]; + tensor var_301_end_mask_0 = const()[name = tensor("op_301_end_mask_0"), val = tensor([true, true, true])]; + tensor var_301 = slice_by_index(begin = var_301_begin_0, end = var_301_end_0, end_mask = var_301_end_mask_0, x = x_3)[name = tensor("op_301")]; + tensor var_304_begin_0 = const()[name = tensor("op_304_begin_0"), val = tensor([0, 1, 0])]; + tensor var_304_end_0 = const()[name = tensor("op_304_end_0"), val = tensor([1, 16, 256])]; + tensor var_304_end_mask_0 = const()[name = tensor("op_304_end_mask_0"), val = tensor([true, true, true])]; + tensor var_304 = slice_by_index(begin = var_304_begin_0, end = var_304_end_0, end_mask = var_304_end_mask_0, x = window_7)[name = tensor("op_304")]; tensor window_9_interleave_0 = const()[name = tensor("window_9_interleave_0"), val = tensor(false)]; - tensor window_9 = concat(axis = var_26, interleave = window_9_interleave_0, values = (var_248, var_245))[name = tensor("window_9")]; - tensor input_21_interleave_0 = const()[name = tensor("input_21_interleave_0"), val = tensor(false)]; - tensor input_21 = concat(axis = var_23, interleave = input_21_interleave_0, values = (window_3, window_5, window_7, window_9))[name = tensor("input_21")]; + tensor window_9 = concat(axis = var_82, interleave = window_9_interleave_0, values = (var_304, var_301))[name = tensor("window_9")]; + tensor input_23_interleave_0 = const()[name = tensor("input_23_interleave_0"), val = tensor(false)]; + tensor input_23 = concat(axis = var_69, interleave = input_23_interleave_0, values = (window_3, window_5, window_7, window_9))[name = tensor("input_23")]; tensor x_5_axes_0 = const()[name = tensor("x_5_axes_0"), val = tensor([-1])]; - tensor x_5 = layer_norm(axes = x_5_axes_0, beta = encoder_conv_module_0_sequential_0_bias, epsilon = var_28, gamma = encoder_conv_module_0_sequential_0_weight, x = input_21)[name = tensor("x_5")]; - tensor input_23_perm_0 = const()[name = tensor("input_23_perm_0"), val = tensor([0, 2, 1])]; + tensor x_5 = layer_norm(axes = x_5_axes_0, beta = inner_encoder_conv_module_0_sequential_0_bias, epsilon = var_66, gamma = inner_encoder_conv_module_0_sequential_0_weight, x = input_23)[name = tensor("x_5")]; + tensor input_25_perm_0 = const()[name = tensor("input_25_perm_0"), val = tensor([0, 2, 1])]; tensor inputs_3_pad_type_0 = const()[name = tensor("inputs_3_pad_type_0"), val = tensor("valid")]; tensor inputs_3_strides_0 = const()[name = tensor("inputs_3_strides_0"), val = tensor([1])]; tensor inputs_3_pad_0 = const()[name = tensor("inputs_3_pad_0"), val = tensor([0, 0])]; tensor inputs_3_dilations_0 = const()[name = tensor("inputs_3_dilations_0"), val = tensor([1])]; tensor inputs_3_groups_0 = const()[name = tensor("inputs_3_groups_0"), val = tensor(1)]; - tensor input_23 = transpose(perm = input_23_perm_0, x = x_5)[name = tensor("transpose_54")]; - tensor inputs_3 = conv(bias = encoder_conv_module_0_sequential_2_conv_bias, dilations = inputs_3_dilations_0, groups = inputs_3_groups_0, pad = inputs_3_pad_0, pad_type = inputs_3_pad_type_0, strides = inputs_3_strides_0, weight = encoder_conv_module_0_sequential_2_conv_weight, x = input_23)[name = tensor("inputs_3")]; - tensor var_273_split_sizes_0 = const()[name = tensor("op_273_split_sizes_0"), val = tensor([256, 256])]; - tensor var_273_axis_0 = const()[name = tensor("op_273_axis_0"), val = tensor(1)]; - tensor var_273_0, tensor var_273_1 = split(axis = var_273_axis_0, split_sizes = var_273_split_sizes_0, x = inputs_3)[name = tensor("op_273")]; - tensor var_275 = sigmoid(x = var_273_1)[name = tensor("op_275")]; - tensor inputs_5 = mul(x = var_273_0, y = var_275)[name = tensor("inputs_5")]; + tensor input_25 = transpose(perm = input_25_perm_0, x = x_5)[name = tensor("transpose_54")]; + tensor inputs_3 = conv(bias = inner_encoder_conv_module_0_sequential_2_conv_bias, dilations = inputs_3_dilations_0, groups = inputs_3_groups_0, pad = inputs_3_pad_0, pad_type = inputs_3_pad_type_0, strides = inputs_3_strides_0, weight = inner_encoder_conv_module_0_sequential_2_conv_weight, x = input_25)[name = tensor("inputs_3")]; + tensor var_329_split_sizes_0 = const()[name = tensor("op_329_split_sizes_0"), val = tensor([256, 256])]; + tensor var_329_axis_0 = const()[name = tensor("op_329_axis_0"), val = tensor(1)]; + tensor var_329_0, tensor var_329_1 = split(axis = var_329_axis_0, split_sizes = var_329_split_sizes_0, x = inputs_3)[name = tensor("op_329")]; + tensor var_331 = sigmoid(x = var_329_1)[name = tensor("op_331")]; + tensor inputs_5 = mul(x = var_329_0, y = var_331)[name = tensor("inputs_5")]; tensor outputs_aug_1_pad_type_0 = const()[name = tensor("outputs_aug_1_pad_type_0"), val = tensor("custom")]; tensor outputs_aug_1_pad_0 = const()[name = tensor("outputs_aug_1_pad_0"), val = tensor([15, 15])]; tensor outputs_aug_1_groups_0 = const()[name = tensor("outputs_aug_1_groups_0"), val = tensor(256)]; tensor outputs_aug_1_strides_0 = const()[name = tensor("outputs_aug_1_strides_0"), val = tensor([1])]; tensor outputs_aug_1_dilations_0 = const()[name = tensor("outputs_aug_1_dilations_0"), val = tensor([1])]; - tensor outputs_aug_1 = conv(dilations = outputs_aug_1_dilations_0, groups = outputs_aug_1_groups_0, pad = outputs_aug_1_pad_0, pad_type = outputs_aug_1_pad_type_0, strides = outputs_aug_1_strides_0, weight = encoder_conv_module_0_sequential_4_conv_weight, x = inputs_5)[name = tensor("outputs_aug_1")]; - tensor input_25_begin_0 = const()[name = tensor("input_25_begin_0"), val = tensor([0, 0, 0])]; - tensor input_25_end_0 = const()[name = tensor("input_25_end_0"), val = tensor([4, 256, 16])]; - tensor input_25_end_mask_0 = const()[name = tensor("input_25_end_mask_0"), val = tensor([true, true, false])]; - tensor input_25 = slice_by_index(begin = input_25_begin_0, end = input_25_end_0, end_mask = input_25_end_mask_0, x = outputs_aug_1)[name = tensor("input_25")]; - tensor inputs_7 = batch_norm(beta = encoder_conv_module_0_sequential_5_bias, epsilon = var_28, gamma = encoder_conv_module_0_sequential_5_weight, mean = encoder_conv_module_0_sequential_5_running_mean, variance = encoder_conv_module_0_sequential_5_running_var, x = input_25)[name = tensor("inputs_7")]; - tensor input_27 = silu(x = inputs_7)[name = tensor("input_27")]; - tensor input_29_pad_type_0 = const()[name = tensor("input_29_pad_type_0"), val = tensor("valid")]; - tensor input_29_strides_0 = const()[name = tensor("input_29_strides_0"), val = tensor([1])]; - tensor input_29_pad_0 = const()[name = tensor("input_29_pad_0"), val = tensor([0, 0])]; - tensor input_29_dilations_0 = const()[name = tensor("input_29_dilations_0"), val = tensor([1])]; - tensor input_29_groups_0 = const()[name = tensor("input_29_groups_0"), val = tensor(1)]; - tensor input_29 = conv(bias = encoder_conv_module_0_sequential_7_conv_bias, dilations = input_29_dilations_0, groups = input_29_groups_0, pad = input_29_pad_0, pad_type = input_29_pad_type_0, strides = input_29_strides_0, weight = encoder_conv_module_0_sequential_7_conv_weight, x = input_27)[name = tensor("input_29")]; + tensor outputs_aug_1 = conv(dilations = outputs_aug_1_dilations_0, groups = outputs_aug_1_groups_0, pad = outputs_aug_1_pad_0, pad_type = outputs_aug_1_pad_type_0, strides = outputs_aug_1_strides_0, weight = inner_encoder_conv_module_0_sequential_4_conv_weight, x = inputs_5)[name = tensor("outputs_aug_1")]; + tensor input_27_begin_0 = const()[name = tensor("input_27_begin_0"), val = tensor([0, 0, 0])]; + tensor input_27_end_0 = const()[name = tensor("input_27_end_0"), val = tensor([4, 256, 16])]; + tensor input_27_end_mask_0 = const()[name = tensor("input_27_end_mask_0"), val = tensor([true, true, false])]; + tensor input_27 = slice_by_index(begin = input_27_begin_0, end = input_27_end_0, end_mask = input_27_end_mask_0, x = outputs_aug_1)[name = tensor("input_27")]; + tensor inputs_7 = batch_norm(beta = inner_encoder_conv_module_0_sequential_5_bias, epsilon = var_66, gamma = inner_encoder_conv_module_0_sequential_5_weight, mean = inner_encoder_conv_module_0_sequential_5_running_mean, variance = inner_encoder_conv_module_0_sequential_5_running_var, x = input_27)[name = tensor("inputs_7")]; + tensor input_29 = silu(x = inputs_7)[name = tensor("input_29")]; + tensor input_31_pad_type_0 = const()[name = tensor("input_31_pad_type_0"), val = tensor("valid")]; + tensor input_31_strides_0 = const()[name = tensor("input_31_strides_0"), val = tensor([1])]; + tensor input_31_pad_0 = const()[name = tensor("input_31_pad_0"), val = tensor([0, 0])]; + tensor input_31_dilations_0 = const()[name = tensor("input_31_dilations_0"), val = tensor([1])]; + tensor input_31_groups_0 = const()[name = tensor("input_31_groups_0"), val = tensor(1)]; + tensor input_31 = conv(bias = inner_encoder_conv_module_0_sequential_7_conv_bias, dilations = input_31_dilations_0, groups = input_31_groups_0, pad = input_31_pad_0, pad_type = input_31_pad_type_0, strides = input_31_strides_0, weight = inner_encoder_conv_module_0_sequential_7_conv_weight, x = input_29)[name = tensor("input_31")]; tensor conv_out_1_perm_0 = const()[name = tensor("conv_out_1_perm_0"), val = tensor([0, 2, 1])]; - tensor var_306_begin_0 = const()[name = tensor("op_306_begin_0"), val = tensor([0, -1, 0])]; - tensor var_306_end_0 = const()[name = tensor("op_306_end_0"), val = tensor([4, 16, 256])]; - tensor var_306_end_mask_0 = const()[name = tensor("op_306_end_mask_0"), val = tensor([true, true, true])]; - tensor conv_out_1 = transpose(perm = conv_out_1_perm_0, x = input_29)[name = tensor("transpose_53")]; - tensor var_306 = slice_by_index(begin = var_306_begin_0, end = var_306_end_0, end_mask = var_306_end_mask_0, x = conv_out_1)[name = tensor("op_306")]; - tensor var_308_perm_0 = const()[name = tensor("op_308_perm_0"), val = tensor([1, 0, 2])]; - tensor var_308 = transpose(perm = var_308_perm_0, x = var_306)[name = tensor("transpose_52")]; - tensor input_31 = add(x = x_3, y = var_308)[name = tensor("input_31")]; - tensor input_33_axes_0 = const()[name = tensor("input_33_axes_0"), val = tensor([-1])]; - tensor input_33 = layer_norm(axes = input_33_axes_0, beta = encoder_ffn2_0_module_sequential_0_bias, epsilon = var_28, gamma = encoder_ffn2_0_module_sequential_0_weight, x = input_31)[name = tensor("input_33")]; - tensor inputs_9 = linear(bias = encoder_ffn2_0_module_sequential_1_linear_bias, weight = encoder_ffn2_0_module_sequential_1_linear_weight, x = input_33)[name = tensor("linear_8")]; - tensor input_35 = silu(x = inputs_9)[name = tensor("input_35")]; - tensor input_39 = linear(bias = encoder_ffn2_0_module_sequential_4_linear_bias, weight = encoder_ffn2_0_module_sequential_4_linear_weight, x = input_35)[name = tensor("linear_9")]; - tensor var_331 = const()[name = tensor("op_331"), val = tensor(0x1p-1)]; - tensor var_332 = mul(x = input_39, y = var_331)[name = tensor("op_332")]; - tensor input_41 = add(x = var_332, y = input_31)[name = tensor("input_41")]; - tensor input_43_axes_0 = const()[name = tensor("input_43_axes_0"), val = tensor([-1])]; - tensor input_43 = layer_norm(axes = input_43_axes_0, beta = encoder_layer_norm_0_bias, epsilon = var_28, gamma = encoder_layer_norm_0_weight, x = input_41)[name = tensor("input_43")]; + tensor var_362_begin_0 = const()[name = tensor("op_362_begin_0"), val = tensor([0, -1, 0])]; + tensor var_362_end_0 = const()[name = tensor("op_362_end_0"), val = tensor([4, 16, 256])]; + tensor var_362_end_mask_0 = const()[name = tensor("op_362_end_mask_0"), val = tensor([true, true, true])]; + tensor conv_out_1 = transpose(perm = conv_out_1_perm_0, x = input_31)[name = tensor("transpose_53")]; + tensor var_362 = slice_by_index(begin = var_362_begin_0, end = var_362_end_0, end_mask = var_362_end_mask_0, x = conv_out_1)[name = tensor("op_362")]; + tensor var_364_perm_0 = const()[name = tensor("op_364_perm_0"), val = tensor([1, 0, 2])]; + tensor var_364 = transpose(perm = var_364_perm_0, x = var_362)[name = tensor("transpose_52")]; + tensor input_33 = add(x = x_3, y = var_364)[name = tensor("input_33")]; + tensor input_35_axes_0 = const()[name = tensor("input_35_axes_0"), val = tensor([-1])]; + tensor input_35 = layer_norm(axes = input_35_axes_0, beta = inner_encoder_ffn2_0_module_sequential_0_bias, epsilon = var_66, gamma = inner_encoder_ffn2_0_module_sequential_0_weight, x = input_33)[name = tensor("input_35")]; + tensor inputs_9 = linear(bias = inner_encoder_ffn2_0_module_sequential_1_linear_bias, weight = inner_encoder_ffn2_0_module_sequential_1_linear_weight, x = input_35)[name = tensor("linear_8")]; + tensor input_37 = silu(x = inputs_9)[name = tensor("input_37")]; + tensor input_41 = linear(bias = inner_encoder_ffn2_0_module_sequential_4_linear_bias, weight = inner_encoder_ffn2_0_module_sequential_4_linear_weight, x = input_37)[name = tensor("linear_9")]; + tensor var_387 = const()[name = tensor("op_387"), val = tensor(0x1p-1)]; + tensor var_388 = mul(x = input_41, y = var_387)[name = tensor("op_388")]; + tensor input_43 = add(x = var_388, y = input_33)[name = tensor("input_43")]; tensor input_45_axes_0 = const()[name = tensor("input_45_axes_0"), val = tensor([-1])]; - tensor input_45 = layer_norm(axes = input_45_axes_0, beta = encoder_ffn1_1_module_sequential_0_bias, epsilon = var_28, gamma = encoder_ffn1_1_module_sequential_0_weight, x = input_43)[name = tensor("input_45")]; - tensor inputs_11 = linear(bias = encoder_ffn1_1_module_sequential_1_linear_bias, weight = encoder_ffn1_1_module_sequential_1_linear_weight, x = input_45)[name = tensor("linear_10")]; - tensor input_47 = silu(x = inputs_11)[name = tensor("input_47")]; - tensor input_51 = linear(bias = encoder_ffn1_1_module_sequential_4_linear_bias, weight = encoder_ffn1_1_module_sequential_4_linear_weight, x = input_47)[name = tensor("linear_11")]; - tensor var_361 = const()[name = tensor("op_361"), val = tensor(0x1p-1)]; - tensor var_362 = mul(x = input_51, y = var_361)[name = tensor("op_362")]; - tensor input_53 = add(x = var_362, y = input_43)[name = tensor("input_53")]; + tensor input_45 = layer_norm(axes = input_45_axes_0, beta = inner_encoder_layer_norm_0_bias, epsilon = var_66, gamma = inner_encoder_layer_norm_0_weight, x = input_43)[name = tensor("input_45")]; + tensor input_47_axes_0 = const()[name = tensor("input_47_axes_0"), val = tensor([-1])]; + tensor input_47 = layer_norm(axes = input_47_axes_0, beta = inner_encoder_ffn1_1_module_sequential_0_bias, epsilon = var_66, gamma = inner_encoder_ffn1_1_module_sequential_0_weight, x = input_45)[name = tensor("input_47")]; + tensor inputs_11 = linear(bias = inner_encoder_ffn1_1_module_sequential_1_linear_bias, weight = inner_encoder_ffn1_1_module_sequential_1_linear_weight, x = input_47)[name = tensor("linear_10")]; + tensor input_49 = silu(x = inputs_11)[name = tensor("input_49")]; + tensor input_53 = linear(bias = inner_encoder_ffn1_1_module_sequential_4_linear_bias, weight = inner_encoder_ffn1_1_module_sequential_4_linear_weight, x = input_49)[name = tensor("linear_11")]; + tensor var_417 = const()[name = tensor("op_417"), val = tensor(0x1p-1)]; + tensor var_418 = mul(x = input_53, y = var_417)[name = tensor("op_418")]; + tensor input_55 = add(x = var_418, y = input_45)[name = tensor("input_55")]; tensor x_7_axes_0 = const()[name = tensor("x_7_axes_0"), val = tensor([-1])]; - tensor x_7 = layer_norm(axes = x_7_axes_0, beta = encoder_ret_lns_1_bias, epsilon = var_28, gamma = encoder_ret_lns_1_weight, x = input_53)[name = tensor("x_7")]; + tensor x_7 = layer_norm(axes = x_7_axes_0, beta = inner_encoder_ret_lns_1_bias, epsilon = var_66, gamma = inner_encoder_ret_lns_1_weight, x = input_55)[name = tensor("x_7")]; tensor prev_kv_3_begin_0 = const()[name = tensor("prev_kv_3_begin_0"), val = tensor([1, 0, 0, 0, 0])]; tensor prev_kv_3_end_0 = const()[name = tensor("prev_kv_3_end_0"), val = tensor([2, 1, 4, 64, 64])]; tensor prev_kv_3_end_mask_0 = const()[name = tensor("prev_kv_3_end_mask_0"), val = tensor([false, true, true, true, true])]; @@ -416,173 +438,173 @@ program(1.0) tensor prev_scale_3_end_mask_0 = const()[name = tensor("prev_scale_3_end_mask_0"), val = tensor([false, true])]; tensor prev_scale_3_squeeze_mask_0 = const()[name = tensor("prev_scale_3_squeeze_mask_0"), val = tensor([true, false])]; tensor prev_scale_3 = slice_by_index(begin = prev_scale_3_begin_0, end = prev_scale_3_end_0, end_mask = prev_scale_3_end_mask_0, squeeze_mask = prev_scale_3_squeeze_mask_0, x = enc_scale)[name = tensor("prev_scale_3")]; - tensor var_376 = linear(bias = encoder_q_proj_1_bias, weight = encoder_q_proj_1_weight, x = x_7)[name = tensor("linear_12")]; - tensor var_377 = const()[name = tensor("op_377"), val = tensor([1, 4, 4, 64])]; - tensor var_378 = reshape(shape = var_377, x = var_376)[name = tensor("op_378")]; + tensor var_432 = linear(bias = inner_encoder_q_proj_1_bias, weight = inner_encoder_q_proj_1_weight, x = x_7)[name = tensor("linear_12")]; + tensor var_433 = const()[name = tensor("op_433"), val = tensor([1, 4, 4, 64])]; + tensor var_434 = reshape(shape = var_433, x = var_432)[name = tensor("op_434")]; tensor q_3_perm_0 = const()[name = tensor("q_3_perm_0"), val = tensor([0, 2, 1, 3])]; - tensor var_382 = linear(bias = encoder_k_proj_1_bias, weight = encoder_k_proj_1_weight, x = x_7)[name = tensor("linear_13")]; - tensor var_383 = const()[name = tensor("op_383"), val = tensor(0x1p-3)]; - tensor var_384 = mul(x = var_382, y = var_383)[name = tensor("op_384")]; - tensor var_385 = const()[name = tensor("op_385"), val = tensor([1, 4, 4, 64])]; - tensor var_386 = reshape(shape = var_385, x = var_384)[name = tensor("op_386")]; + tensor var_438 = linear(bias = inner_encoder_k_proj_1_bias, weight = inner_encoder_k_proj_1_weight, x = x_7)[name = tensor("linear_13")]; + tensor var_439 = const()[name = tensor("op_439"), val = tensor(0x1p-3)]; + tensor var_440 = mul(x = var_438, y = var_439)[name = tensor("op_440")]; + tensor var_441 = const()[name = tensor("op_441"), val = tensor([1, 4, 4, 64])]; + tensor var_442 = reshape(shape = var_441, x = var_440)[name = tensor("op_442")]; tensor k_3_perm_0 = const()[name = tensor("k_3_perm_0"), val = tensor([0, 2, 1, 3])]; - tensor var_390 = linear(bias = encoder_v_proj_1_bias, weight = encoder_v_proj_1_weight, x = x_7)[name = tensor("linear_14")]; - tensor var_391 = const()[name = tensor("op_391"), val = tensor([1, 4, 4, 64])]; - tensor var_392 = reshape(shape = var_391, x = var_390)[name = tensor("op_392")]; + tensor var_446 = linear(bias = inner_encoder_v_proj_1_bias, weight = inner_encoder_v_proj_1_weight, x = x_7)[name = tensor("linear_14")]; + tensor var_447 = const()[name = tensor("op_447"), val = tensor([1, 4, 4, 64])]; + tensor var_448 = reshape(shape = var_447, x = var_446)[name = tensor("op_448")]; tensor v_3_perm_0 = const()[name = tensor("v_3_perm_0"), val = tensor([0, 2, 1, 3])]; - tensor input_57 = linear(bias = encoder_g_proj_1_bias, weight = encoder_g_proj_1_weight, x = x_7)[name = tensor("linear_15")]; + tensor input_59 = linear(bias = inner_encoder_g_proj_1_bias, weight = inner_encoder_g_proj_1_weight, x = x_7)[name = tensor("linear_15")]; tensor sqrt_s0_3 = sqrt(x = prev_scale_3)[name = tensor("sqrt_s0_3")]; - tensor s_t_3 = add(x = prev_scale_3, y = encoder__t_index)[name = tensor("s_t_3")]; + tensor s_t_3 = add(x = prev_scale_3, y = inner_encoder__t_index)[name = tensor("s_t_3")]; tensor sqrt_s_t_3 = sqrt(x = s_t_3)[name = tensor("sqrt_s_t_3")]; tensor qk_3_transpose_x_1 = const()[name = tensor("qk_3_transpose_x_1"), val = tensor(false)]; tensor qk_3_transpose_y_1 = const()[name = tensor("qk_3_transpose_y_1"), val = tensor(true)]; - tensor k_3 = transpose(perm = k_3_perm_0, x = var_386)[name = tensor("transpose_50")]; - tensor q_3 = transpose(perm = q_3_perm_0, x = var_378)[name = tensor("transpose_51")]; + tensor k_3 = transpose(perm = k_3_perm_0, x = var_442)[name = tensor("transpose_50")]; + tensor q_3 = transpose(perm = q_3_perm_0, x = var_434)[name = tensor("transpose_51")]; tensor qk_3 = matmul(transpose_x = qk_3_transpose_x_1, transpose_y = qk_3_transpose_y_1, x = q_3, y = k_3)[name = tensor("qk_3")]; - tensor var_402 = const()[name = tensor("op_402"), val = tensor([4, 1])]; - tensor var_403 = reshape(shape = var_402, x = sqrt_s_t_3)[name = tensor("op_403")]; - tensor M_3 = real_div(x = encoder__causal_mask, y = var_403)[name = tensor("M_3")]; - tensor var_405 = mul(x = qk_3, y = M_3)[name = tensor("op_405")]; + tensor var_458 = const()[name = tensor("op_458"), val = tensor([4, 1])]; + tensor var_459 = reshape(shape = var_458, x = sqrt_s_t_3)[name = tensor("op_459")]; + tensor M_3 = real_div(x = inner_encoder__causal_mask, y = var_459)[name = tensor("M_3")]; + tensor var_461 = mul(x = qk_3, y = M_3)[name = tensor("op_461")]; tensor inner_3_transpose_x_0 = const()[name = tensor("inner_3_transpose_x_0"), val = tensor(false)]; tensor inner_3_transpose_y_0 = const()[name = tensor("inner_3_transpose_y_0"), val = tensor(false)]; - tensor v_3 = transpose(perm = v_3_perm_0, x = var_392)[name = tensor("transpose_49")]; - tensor inner_3 = matmul(transpose_x = inner_3_transpose_x_0, transpose_y = inner_3_transpose_y_0, x = var_405, y = v_3)[name = tensor("inner_3")]; - tensor var_407_transpose_x_0 = const()[name = tensor("op_407_transpose_x_0"), val = tensor(false)]; - tensor var_407_transpose_y_0 = const()[name = tensor("op_407_transpose_y_0"), val = tensor(false)]; - tensor var_407 = matmul(transpose_x = var_407_transpose_x_0, transpose_y = var_407_transpose_y_0, x = q_3, y = prev_kv_3)[name = tensor("op_407")]; - tensor var_408 = real_div(x = sqrt_s0_3, y = sqrt_s_t_3)[name = tensor("op_408")]; - tensor var_409 = const()[name = tensor("op_409"), val = tensor([1, 1, 4, 1])]; - tensor var_410 = reshape(shape = var_409, x = var_408)[name = tensor("op_410")]; - tensor cross_3 = mul(x = var_407, y = var_410)[name = tensor("cross_3")]; + tensor v_3 = transpose(perm = v_3_perm_0, x = var_448)[name = tensor("transpose_49")]; + tensor inner_3 = matmul(transpose_x = inner_3_transpose_x_0, transpose_y = inner_3_transpose_y_0, x = var_461, y = v_3)[name = tensor("inner_3")]; + tensor var_463_transpose_x_0 = const()[name = tensor("op_463_transpose_x_0"), val = tensor(false)]; + tensor var_463_transpose_y_0 = const()[name = tensor("op_463_transpose_y_0"), val = tensor(false)]; + tensor var_463 = matmul(transpose_x = var_463_transpose_x_0, transpose_y = var_463_transpose_y_0, x = q_3, y = prev_kv_3)[name = tensor("op_463")]; + tensor var_464 = real_div(x = sqrt_s0_3, y = sqrt_s_t_3)[name = tensor("op_464")]; + tensor var_465 = const()[name = tensor("op_465"), val = tensor([1, 1, 4, 1])]; + tensor var_466 = reshape(shape = var_465, x = var_464)[name = tensor("op_466")]; + tensor cross_3 = mul(x = var_463, y = var_466)[name = tensor("cross_3")]; tensor out_7 = add(x = inner_3, y = cross_3)[name = tensor("out_7")]; - tensor var_413 = mul(x = prev_kv_3, y = sqrt_s0_3)[name = tensor("op_413")]; - tensor var_415_transpose_x_1 = const()[name = tensor("op_415_transpose_x_1"), val = tensor(true)]; - tensor var_415_transpose_y_1 = const()[name = tensor("op_415_transpose_y_1"), val = tensor(false)]; - tensor var_415 = matmul(transpose_x = var_415_transpose_x_1, transpose_y = var_415_transpose_y_1, x = k_3, y = v_3)[name = tensor("op_415")]; - tensor new_kv_unnorm_3 = add(x = var_413, y = var_415)[name = tensor("new_kv_unnorm_3")]; - tensor var_417 = const()[name = tensor("op_417"), val = tensor(0x1p+2)]; - tensor new_scale_3 = add(x = prev_scale_3, y = var_417)[name = tensor("new_scale_3")]; - tensor var_419 = sqrt(x = new_scale_3)[name = tensor("op_419")]; - tensor var_420 = real_div(x = new_kv_unnorm_3, y = var_419)[name = tensor("op_420")]; - tensor var_421_perm_0 = const()[name = tensor("op_421_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_469 = mul(x = prev_kv_3, y = sqrt_s0_3)[name = tensor("op_469")]; + tensor var_471_transpose_x_1 = const()[name = tensor("op_471_transpose_x_1"), val = tensor(true)]; + tensor var_471_transpose_y_1 = const()[name = tensor("op_471_transpose_y_1"), val = tensor(false)]; + tensor var_471 = matmul(transpose_x = var_471_transpose_x_1, transpose_y = var_471_transpose_y_1, x = k_3, y = v_3)[name = tensor("op_471")]; + tensor new_kv_unnorm_3 = add(x = var_469, y = var_471)[name = tensor("new_kv_unnorm_3")]; + tensor var_473 = const()[name = tensor("op_473"), val = tensor(0x1p+2)]; + tensor new_scale_3 = add(x = prev_scale_3, y = var_473)[name = tensor("new_scale_3")]; + tensor var_475 = sqrt(x = new_scale_3)[name = tensor("op_475")]; + tensor var_476 = real_div(x = new_kv_unnorm_3, y = var_475)[name = tensor("op_476")]; + tensor var_477_perm_0 = const()[name = tensor("op_477_perm_0"), val = tensor([0, 2, 1, 3])]; tensor out_9_axes_0 = const()[name = tensor("out_9_axes_0"), val = tensor([-1])]; - tensor var_421 = transpose(perm = var_421_perm_0, x = out_7)[name = tensor("transpose_48")]; - tensor out_9 = layer_norm(axes = out_9_axes_0, epsilon = var_17, x = var_421)[name = tensor("out_9")]; - tensor var_425 = const()[name = tensor("op_425"), val = tensor([1, 4, 256])]; - tensor out_11 = reshape(shape = var_425, x = out_9)[name = tensor("out_11")]; - tensor var_427 = silu(x = input_57)[name = tensor("op_427")]; - tensor input_59 = mul(x = var_427, y = out_11)[name = tensor("input_59")]; - tensor ret_out_3 = linear(bias = encoder_out_proj_1_bias, weight = encoder_out_proj_1_weight, x = input_59)[name = tensor("linear_16")]; - tensor x_9 = add(x = input_53, y = ret_out_3)[name = tensor("x_9")]; + tensor var_477 = transpose(perm = var_477_perm_0, x = out_7)[name = tensor("transpose_48")]; + tensor out_9 = layer_norm(axes = out_9_axes_0, epsilon = var_74, x = var_477)[name = tensor("out_9")]; + tensor var_481 = const()[name = tensor("op_481"), val = tensor([1, 4, 256])]; + tensor out_11 = reshape(shape = var_481, x = out_9)[name = tensor("out_11")]; + tensor var_483 = silu(x = input_59)[name = tensor("op_483")]; + tensor input_61 = mul(x = var_483, y = out_11)[name = tensor("input_61")]; + tensor ret_out_3 = linear(bias = inner_encoder_out_proj_1_bias, weight = inner_encoder_out_proj_1_weight, x = input_61)[name = tensor("linear_16")]; + tensor x_9 = add(x = input_55, y = ret_out_3)[name = tensor("x_9")]; tensor window_11_begin_0 = const()[name = tensor("window_11_begin_0"), val = tensor([1, 0, 0, 0])]; tensor window_11_end_0 = const()[name = tensor("window_11_end_0"), val = tensor([2, 1, 16, 256])]; tensor window_11_end_mask_0 = const()[name = tensor("window_11_end_mask_0"), val = tensor([false, true, true, true])]; tensor window_11_squeeze_mask_0 = const()[name = tensor("window_11_squeeze_mask_0"), val = tensor([true, false, false, false])]; tensor window_11 = slice_by_index(begin = window_11_begin_0, end = window_11_end_0, end_mask = window_11_end_mask_0, squeeze_mask = window_11_squeeze_mask_0, x = enc_conv_cache)[name = tensor("window_11")]; - tensor var_435_begin_0 = const()[name = tensor("op_435_begin_0"), val = tensor([0, 0, 0])]; - tensor var_435_end_0 = const()[name = tensor("op_435_end_0"), val = tensor([1, 1, 256])]; - tensor var_435_end_mask_0 = const()[name = tensor("op_435_end_mask_0"), val = tensor([true, false, true])]; - tensor var_435 = slice_by_index(begin = var_435_begin_0, end = var_435_end_0, end_mask = var_435_end_mask_0, x = x_9)[name = tensor("op_435")]; - tensor var_438_begin_0 = const()[name = tensor("op_438_begin_0"), val = tensor([0, 1, 0])]; - tensor var_438_end_0 = const()[name = tensor("op_438_end_0"), val = tensor([1, 16, 256])]; - tensor var_438_end_mask_0 = const()[name = tensor("op_438_end_mask_0"), val = tensor([true, true, true])]; - tensor var_438 = slice_by_index(begin = var_438_begin_0, end = var_438_end_0, end_mask = var_438_end_mask_0, x = window_11)[name = tensor("op_438")]; + tensor var_491_begin_0 = const()[name = tensor("op_491_begin_0"), val = tensor([0, 0, 0])]; + tensor var_491_end_0 = const()[name = tensor("op_491_end_0"), val = tensor([1, 1, 256])]; + tensor var_491_end_mask_0 = const()[name = tensor("op_491_end_mask_0"), val = tensor([true, false, true])]; + tensor var_491 = slice_by_index(begin = var_491_begin_0, end = var_491_end_0, end_mask = var_491_end_mask_0, x = x_9)[name = tensor("op_491")]; + tensor var_494_begin_0 = const()[name = tensor("op_494_begin_0"), val = tensor([0, 1, 0])]; + tensor var_494_end_0 = const()[name = tensor("op_494_end_0"), val = tensor([1, 16, 256])]; + tensor var_494_end_mask_0 = const()[name = tensor("op_494_end_mask_0"), val = tensor([true, true, true])]; + tensor var_494 = slice_by_index(begin = var_494_begin_0, end = var_494_end_0, end_mask = var_494_end_mask_0, x = window_11)[name = tensor("op_494")]; tensor window_13_interleave_0 = const()[name = tensor("window_13_interleave_0"), val = tensor(false)]; - tensor window_13 = concat(axis = var_26, interleave = window_13_interleave_0, values = (var_438, var_435))[name = tensor("window_13")]; - tensor var_443_begin_0 = const()[name = tensor("op_443_begin_0"), val = tensor([0, 1, 0])]; - tensor var_443_end_0 = const()[name = tensor("op_443_end_0"), val = tensor([1, 2, 256])]; - tensor var_443_end_mask_0 = const()[name = tensor("op_443_end_mask_0"), val = tensor([true, false, true])]; - tensor var_443 = slice_by_index(begin = var_443_begin_0, end = var_443_end_0, end_mask = var_443_end_mask_0, x = x_9)[name = tensor("op_443")]; - tensor var_446_begin_0 = const()[name = tensor("op_446_begin_0"), val = tensor([0, 1, 0])]; - tensor var_446_end_0 = const()[name = tensor("op_446_end_0"), val = tensor([1, 16, 256])]; - tensor var_446_end_mask_0 = const()[name = tensor("op_446_end_mask_0"), val = tensor([true, true, true])]; - tensor var_446 = slice_by_index(begin = var_446_begin_0, end = var_446_end_0, end_mask = var_446_end_mask_0, x = window_13)[name = tensor("op_446")]; + tensor window_13 = concat(axis = var_82, interleave = window_13_interleave_0, values = (var_494, var_491))[name = tensor("window_13")]; + tensor var_499_begin_0 = const()[name = tensor("op_499_begin_0"), val = tensor([0, 1, 0])]; + tensor var_499_end_0 = const()[name = tensor("op_499_end_0"), val = tensor([1, 2, 256])]; + tensor var_499_end_mask_0 = const()[name = tensor("op_499_end_mask_0"), val = tensor([true, false, true])]; + tensor var_499 = slice_by_index(begin = var_499_begin_0, end = var_499_end_0, end_mask = var_499_end_mask_0, x = x_9)[name = tensor("op_499")]; + tensor var_502_begin_0 = const()[name = tensor("op_502_begin_0"), val = tensor([0, 1, 0])]; + tensor var_502_end_0 = const()[name = tensor("op_502_end_0"), val = tensor([1, 16, 256])]; + tensor var_502_end_mask_0 = const()[name = tensor("op_502_end_mask_0"), val = tensor([true, true, true])]; + tensor var_502 = slice_by_index(begin = var_502_begin_0, end = var_502_end_0, end_mask = var_502_end_mask_0, x = window_13)[name = tensor("op_502")]; tensor window_15_interleave_0 = const()[name = tensor("window_15_interleave_0"), val = tensor(false)]; - tensor window_15 = concat(axis = var_26, interleave = window_15_interleave_0, values = (var_446, var_443))[name = tensor("window_15")]; - tensor var_451_begin_0 = const()[name = tensor("op_451_begin_0"), val = tensor([0, 2, 0])]; - tensor var_451_end_0 = const()[name = tensor("op_451_end_0"), val = tensor([1, 3, 256])]; - tensor var_451_end_mask_0 = const()[name = tensor("op_451_end_mask_0"), val = tensor([true, false, true])]; - tensor var_451 = slice_by_index(begin = var_451_begin_0, end = var_451_end_0, end_mask = var_451_end_mask_0, x = x_9)[name = tensor("op_451")]; - tensor var_454_begin_0 = const()[name = tensor("op_454_begin_0"), val = tensor([0, 1, 0])]; - tensor var_454_end_0 = const()[name = tensor("op_454_end_0"), val = tensor([1, 16, 256])]; - tensor var_454_end_mask_0 = const()[name = tensor("op_454_end_mask_0"), val = tensor([true, true, true])]; - tensor var_454 = slice_by_index(begin = var_454_begin_0, end = var_454_end_0, end_mask = var_454_end_mask_0, x = window_15)[name = tensor("op_454")]; + tensor window_15 = concat(axis = var_82, interleave = window_15_interleave_0, values = (var_502, var_499))[name = tensor("window_15")]; + tensor var_507_begin_0 = const()[name = tensor("op_507_begin_0"), val = tensor([0, 2, 0])]; + tensor var_507_end_0 = const()[name = tensor("op_507_end_0"), val = tensor([1, 3, 256])]; + tensor var_507_end_mask_0 = const()[name = tensor("op_507_end_mask_0"), val = tensor([true, false, true])]; + tensor var_507 = slice_by_index(begin = var_507_begin_0, end = var_507_end_0, end_mask = var_507_end_mask_0, x = x_9)[name = tensor("op_507")]; + tensor var_510_begin_0 = const()[name = tensor("op_510_begin_0"), val = tensor([0, 1, 0])]; + tensor var_510_end_0 = const()[name = tensor("op_510_end_0"), val = tensor([1, 16, 256])]; + tensor var_510_end_mask_0 = const()[name = tensor("op_510_end_mask_0"), val = tensor([true, true, true])]; + tensor var_510 = slice_by_index(begin = var_510_begin_0, end = var_510_end_0, end_mask = var_510_end_mask_0, x = window_15)[name = tensor("op_510")]; tensor window_17_interleave_0 = const()[name = tensor("window_17_interleave_0"), val = tensor(false)]; - tensor window_17 = concat(axis = var_26, interleave = window_17_interleave_0, values = (var_454, var_451))[name = tensor("window_17")]; - tensor var_459_begin_0 = const()[name = tensor("op_459_begin_0"), val = tensor([0, 3, 0])]; - tensor var_459_end_0 = const()[name = tensor("op_459_end_0"), val = tensor([1, 1, 256])]; - tensor var_459_end_mask_0 = const()[name = tensor("op_459_end_mask_0"), val = tensor([true, true, true])]; - tensor var_459 = slice_by_index(begin = var_459_begin_0, end = var_459_end_0, end_mask = var_459_end_mask_0, x = x_9)[name = tensor("op_459")]; - tensor var_462_begin_0 = const()[name = tensor("op_462_begin_0"), val = tensor([0, 1, 0])]; - tensor var_462_end_0 = const()[name = tensor("op_462_end_0"), val = tensor([1, 16, 256])]; - tensor var_462_end_mask_0 = const()[name = tensor("op_462_end_mask_0"), val = tensor([true, true, true])]; - tensor var_462 = slice_by_index(begin = var_462_begin_0, end = var_462_end_0, end_mask = var_462_end_mask_0, x = window_17)[name = tensor("op_462")]; + tensor window_17 = concat(axis = var_82, interleave = window_17_interleave_0, values = (var_510, var_507))[name = tensor("window_17")]; + tensor var_515_begin_0 = const()[name = tensor("op_515_begin_0"), val = tensor([0, 3, 0])]; + tensor var_515_end_0 = const()[name = tensor("op_515_end_0"), val = tensor([1, 1, 256])]; + tensor var_515_end_mask_0 = const()[name = tensor("op_515_end_mask_0"), val = tensor([true, true, true])]; + tensor var_515 = slice_by_index(begin = var_515_begin_0, end = var_515_end_0, end_mask = var_515_end_mask_0, x = x_9)[name = tensor("op_515")]; + tensor var_518_begin_0 = const()[name = tensor("op_518_begin_0"), val = tensor([0, 1, 0])]; + tensor var_518_end_0 = const()[name = tensor("op_518_end_0"), val = tensor([1, 16, 256])]; + tensor var_518_end_mask_0 = const()[name = tensor("op_518_end_mask_0"), val = tensor([true, true, true])]; + tensor var_518 = slice_by_index(begin = var_518_begin_0, end = var_518_end_0, end_mask = var_518_end_mask_0, x = window_17)[name = tensor("op_518")]; tensor window_19_interleave_0 = const()[name = tensor("window_19_interleave_0"), val = tensor(false)]; - tensor window_19 = concat(axis = var_26, interleave = window_19_interleave_0, values = (var_462, var_459))[name = tensor("window_19")]; - tensor input_61_interleave_0 = const()[name = tensor("input_61_interleave_0"), val = tensor(false)]; - tensor input_61 = concat(axis = var_23, interleave = input_61_interleave_0, values = (window_13, window_15, window_17, window_19))[name = tensor("input_61")]; + tensor window_19 = concat(axis = var_82, interleave = window_19_interleave_0, values = (var_518, var_515))[name = tensor("window_19")]; + tensor input_63_interleave_0 = const()[name = tensor("input_63_interleave_0"), val = tensor(false)]; + tensor input_63 = concat(axis = var_69, interleave = input_63_interleave_0, values = (window_13, window_15, window_17, window_19))[name = tensor("input_63")]; tensor x_11_axes_0 = const()[name = tensor("x_11_axes_0"), val = tensor([-1])]; - tensor x_11 = layer_norm(axes = x_11_axes_0, beta = encoder_conv_module_1_sequential_0_bias, epsilon = var_28, gamma = encoder_conv_module_1_sequential_0_weight, x = input_61)[name = tensor("x_11")]; - tensor input_63_perm_0 = const()[name = tensor("input_63_perm_0"), val = tensor([0, 2, 1])]; + tensor x_11 = layer_norm(axes = x_11_axes_0, beta = inner_encoder_conv_module_1_sequential_0_bias, epsilon = var_66, gamma = inner_encoder_conv_module_1_sequential_0_weight, x = input_63)[name = tensor("x_11")]; + tensor input_65_perm_0 = const()[name = tensor("input_65_perm_0"), val = tensor([0, 2, 1])]; tensor inputs_13_pad_type_0 = const()[name = tensor("inputs_13_pad_type_0"), val = tensor("valid")]; tensor inputs_13_strides_0 = const()[name = tensor("inputs_13_strides_0"), val = tensor([1])]; tensor inputs_13_pad_0 = const()[name = tensor("inputs_13_pad_0"), val = tensor([0, 0])]; tensor inputs_13_dilations_0 = const()[name = tensor("inputs_13_dilations_0"), val = tensor([1])]; tensor inputs_13_groups_0 = const()[name = tensor("inputs_13_groups_0"), val = tensor(1)]; - tensor input_63 = transpose(perm = input_63_perm_0, x = x_11)[name = tensor("transpose_47")]; - tensor inputs_13 = conv(bias = encoder_conv_module_1_sequential_2_conv_bias, dilations = inputs_13_dilations_0, groups = inputs_13_groups_0, pad = inputs_13_pad_0, pad_type = inputs_13_pad_type_0, strides = inputs_13_strides_0, weight = encoder_conv_module_1_sequential_2_conv_weight, x = input_63)[name = tensor("inputs_13")]; - tensor var_487_split_sizes_0 = const()[name = tensor("op_487_split_sizes_0"), val = tensor([256, 256])]; - tensor var_487_axis_0 = const()[name = tensor("op_487_axis_0"), val = tensor(1)]; - tensor var_487_0, tensor var_487_1 = split(axis = var_487_axis_0, split_sizes = var_487_split_sizes_0, x = inputs_13)[name = tensor("op_487")]; - tensor var_489 = sigmoid(x = var_487_1)[name = tensor("op_489")]; - tensor inputs_15 = mul(x = var_487_0, y = var_489)[name = tensor("inputs_15")]; + tensor input_65 = transpose(perm = input_65_perm_0, x = x_11)[name = tensor("transpose_47")]; + tensor inputs_13 = conv(bias = inner_encoder_conv_module_1_sequential_2_conv_bias, dilations = inputs_13_dilations_0, groups = inputs_13_groups_0, pad = inputs_13_pad_0, pad_type = inputs_13_pad_type_0, strides = inputs_13_strides_0, weight = inner_encoder_conv_module_1_sequential_2_conv_weight, x = input_65)[name = tensor("inputs_13")]; + tensor var_543_split_sizes_0 = const()[name = tensor("op_543_split_sizes_0"), val = tensor([256, 256])]; + tensor var_543_axis_0 = const()[name = tensor("op_543_axis_0"), val = tensor(1)]; + tensor var_543_0, tensor var_543_1 = split(axis = var_543_axis_0, split_sizes = var_543_split_sizes_0, x = inputs_13)[name = tensor("op_543")]; + tensor var_545 = sigmoid(x = var_543_1)[name = tensor("op_545")]; + tensor inputs_15 = mul(x = var_543_0, y = var_545)[name = tensor("inputs_15")]; tensor outputs_aug_3_pad_type_0 = const()[name = tensor("outputs_aug_3_pad_type_0"), val = tensor("custom")]; tensor outputs_aug_3_pad_0 = const()[name = tensor("outputs_aug_3_pad_0"), val = tensor([15, 15])]; tensor outputs_aug_3_groups_0 = const()[name = tensor("outputs_aug_3_groups_0"), val = tensor(256)]; tensor outputs_aug_3_strides_0 = const()[name = tensor("outputs_aug_3_strides_0"), val = tensor([1])]; tensor outputs_aug_3_dilations_0 = const()[name = tensor("outputs_aug_3_dilations_0"), val = tensor([1])]; - tensor outputs_aug_3 = conv(dilations = outputs_aug_3_dilations_0, groups = outputs_aug_3_groups_0, pad = outputs_aug_3_pad_0, pad_type = outputs_aug_3_pad_type_0, strides = outputs_aug_3_strides_0, weight = encoder_conv_module_1_sequential_4_conv_weight, x = inputs_15)[name = tensor("outputs_aug_3")]; - tensor input_65_begin_0 = const()[name = tensor("input_65_begin_0"), val = tensor([0, 0, 0])]; - tensor input_65_end_0 = const()[name = tensor("input_65_end_0"), val = tensor([4, 256, 16])]; - tensor input_65_end_mask_0 = const()[name = tensor("input_65_end_mask_0"), val = tensor([true, true, false])]; - tensor input_65 = slice_by_index(begin = input_65_begin_0, end = input_65_end_0, end_mask = input_65_end_mask_0, x = outputs_aug_3)[name = tensor("input_65")]; - tensor inputs_17 = batch_norm(beta = encoder_conv_module_1_sequential_5_bias, epsilon = var_28, gamma = encoder_conv_module_1_sequential_5_weight, mean = encoder_conv_module_1_sequential_5_running_mean, variance = encoder_conv_module_1_sequential_5_running_var, x = input_65)[name = tensor("inputs_17")]; - tensor input_67 = silu(x = inputs_17)[name = tensor("input_67")]; - tensor input_69_pad_type_0 = const()[name = tensor("input_69_pad_type_0"), val = tensor("valid")]; - tensor input_69_strides_0 = const()[name = tensor("input_69_strides_0"), val = tensor([1])]; - tensor input_69_pad_0 = const()[name = tensor("input_69_pad_0"), val = tensor([0, 0])]; - tensor input_69_dilations_0 = const()[name = tensor("input_69_dilations_0"), val = tensor([1])]; - tensor input_69_groups_0 = const()[name = tensor("input_69_groups_0"), val = tensor(1)]; - tensor input_69 = conv(bias = encoder_conv_module_1_sequential_7_conv_bias, dilations = input_69_dilations_0, groups = input_69_groups_0, pad = input_69_pad_0, pad_type = input_69_pad_type_0, strides = input_69_strides_0, weight = encoder_conv_module_1_sequential_7_conv_weight, x = input_67)[name = tensor("input_69")]; + tensor outputs_aug_3 = conv(dilations = outputs_aug_3_dilations_0, groups = outputs_aug_3_groups_0, pad = outputs_aug_3_pad_0, pad_type = outputs_aug_3_pad_type_0, strides = outputs_aug_3_strides_0, weight = inner_encoder_conv_module_1_sequential_4_conv_weight, x = inputs_15)[name = tensor("outputs_aug_3")]; + tensor input_67_begin_0 = const()[name = tensor("input_67_begin_0"), val = tensor([0, 0, 0])]; + tensor input_67_end_0 = const()[name = tensor("input_67_end_0"), val = tensor([4, 256, 16])]; + tensor input_67_end_mask_0 = const()[name = tensor("input_67_end_mask_0"), val = tensor([true, true, false])]; + tensor input_67 = slice_by_index(begin = input_67_begin_0, end = input_67_end_0, end_mask = input_67_end_mask_0, x = outputs_aug_3)[name = tensor("input_67")]; + tensor inputs_17 = batch_norm(beta = inner_encoder_conv_module_1_sequential_5_bias, epsilon = var_66, gamma = inner_encoder_conv_module_1_sequential_5_weight, mean = inner_encoder_conv_module_1_sequential_5_running_mean, variance = inner_encoder_conv_module_1_sequential_5_running_var, x = input_67)[name = tensor("inputs_17")]; + tensor input_69 = silu(x = inputs_17)[name = tensor("input_69")]; + tensor input_71_pad_type_0 = const()[name = tensor("input_71_pad_type_0"), val = tensor("valid")]; + tensor input_71_strides_0 = const()[name = tensor("input_71_strides_0"), val = tensor([1])]; + tensor input_71_pad_0 = const()[name = tensor("input_71_pad_0"), val = tensor([0, 0])]; + tensor input_71_dilations_0 = const()[name = tensor("input_71_dilations_0"), val = tensor([1])]; + tensor input_71_groups_0 = const()[name = tensor("input_71_groups_0"), val = tensor(1)]; + tensor input_71 = conv(bias = inner_encoder_conv_module_1_sequential_7_conv_bias, dilations = input_71_dilations_0, groups = input_71_groups_0, pad = input_71_pad_0, pad_type = input_71_pad_type_0, strides = input_71_strides_0, weight = inner_encoder_conv_module_1_sequential_7_conv_weight, x = input_69)[name = tensor("input_71")]; tensor conv_out_3_perm_0 = const()[name = tensor("conv_out_3_perm_0"), val = tensor([0, 2, 1])]; - tensor var_520_begin_0 = const()[name = tensor("op_520_begin_0"), val = tensor([0, -1, 0])]; - tensor var_520_end_0 = const()[name = tensor("op_520_end_0"), val = tensor([4, 16, 256])]; - tensor var_520_end_mask_0 = const()[name = tensor("op_520_end_mask_0"), val = tensor([true, true, true])]; - tensor conv_out_3 = transpose(perm = conv_out_3_perm_0, x = input_69)[name = tensor("transpose_46")]; - tensor var_520 = slice_by_index(begin = var_520_begin_0, end = var_520_end_0, end_mask = var_520_end_mask_0, x = conv_out_3)[name = tensor("op_520")]; - tensor var_522_perm_0 = const()[name = tensor("op_522_perm_0"), val = tensor([1, 0, 2])]; - tensor var_522 = transpose(perm = var_522_perm_0, x = var_520)[name = tensor("transpose_45")]; - tensor input_71 = add(x = x_9, y = var_522)[name = tensor("input_71")]; - tensor input_73_axes_0 = const()[name = tensor("input_73_axes_0"), val = tensor([-1])]; - tensor input_73 = layer_norm(axes = input_73_axes_0, beta = encoder_ffn2_1_module_sequential_0_bias, epsilon = var_28, gamma = encoder_ffn2_1_module_sequential_0_weight, x = input_71)[name = tensor("input_73")]; - tensor inputs_19 = linear(bias = encoder_ffn2_1_module_sequential_1_linear_bias, weight = encoder_ffn2_1_module_sequential_1_linear_weight, x = input_73)[name = tensor("linear_17")]; - tensor input_75 = silu(x = inputs_19)[name = tensor("input_75")]; - tensor input_79 = linear(bias = encoder_ffn2_1_module_sequential_4_linear_bias, weight = encoder_ffn2_1_module_sequential_4_linear_weight, x = input_75)[name = tensor("linear_18")]; - tensor var_545 = const()[name = tensor("op_545"), val = tensor(0x1p-1)]; - tensor var_546 = mul(x = input_79, y = var_545)[name = tensor("op_546")]; - tensor input_81 = add(x = var_546, y = input_71)[name = tensor("input_81")]; - tensor input_83_axes_0 = const()[name = tensor("input_83_axes_0"), val = tensor([-1])]; - tensor input_83 = layer_norm(axes = input_83_axes_0, beta = encoder_layer_norm_1_bias, epsilon = var_28, gamma = encoder_layer_norm_1_weight, x = input_81)[name = tensor("input_83")]; + tensor var_576_begin_0 = const()[name = tensor("op_576_begin_0"), val = tensor([0, -1, 0])]; + tensor var_576_end_0 = const()[name = tensor("op_576_end_0"), val = tensor([4, 16, 256])]; + tensor var_576_end_mask_0 = const()[name = tensor("op_576_end_mask_0"), val = tensor([true, true, true])]; + tensor conv_out_3 = transpose(perm = conv_out_3_perm_0, x = input_71)[name = tensor("transpose_46")]; + tensor var_576 = slice_by_index(begin = var_576_begin_0, end = var_576_end_0, end_mask = var_576_end_mask_0, x = conv_out_3)[name = tensor("op_576")]; + tensor var_578_perm_0 = const()[name = tensor("op_578_perm_0"), val = tensor([1, 0, 2])]; + tensor var_578 = transpose(perm = var_578_perm_0, x = var_576)[name = tensor("transpose_45")]; + tensor input_73 = add(x = x_9, y = var_578)[name = tensor("input_73")]; + tensor input_75_axes_0 = const()[name = tensor("input_75_axes_0"), val = tensor([-1])]; + tensor input_75 = layer_norm(axes = input_75_axes_0, beta = inner_encoder_ffn2_1_module_sequential_0_bias, epsilon = var_66, gamma = inner_encoder_ffn2_1_module_sequential_0_weight, x = input_73)[name = tensor("input_75")]; + tensor inputs_19 = linear(bias = inner_encoder_ffn2_1_module_sequential_1_linear_bias, weight = inner_encoder_ffn2_1_module_sequential_1_linear_weight, x = input_75)[name = tensor("linear_17")]; + tensor input_77 = silu(x = inputs_19)[name = tensor("input_77")]; + tensor input_81 = linear(bias = inner_encoder_ffn2_1_module_sequential_4_linear_bias, weight = inner_encoder_ffn2_1_module_sequential_4_linear_weight, x = input_77)[name = tensor("linear_18")]; + tensor var_601 = const()[name = tensor("op_601"), val = tensor(0x1p-1)]; + tensor var_602 = mul(x = input_81, y = var_601)[name = tensor("op_602")]; + tensor input_83 = add(x = var_602, y = input_73)[name = tensor("input_83")]; tensor input_85_axes_0 = const()[name = tensor("input_85_axes_0"), val = tensor([-1])]; - tensor input_85 = layer_norm(axes = input_85_axes_0, beta = encoder_ffn1_2_module_sequential_0_bias, epsilon = var_28, gamma = encoder_ffn1_2_module_sequential_0_weight, x = input_83)[name = tensor("input_85")]; - tensor inputs_21 = linear(bias = encoder_ffn1_2_module_sequential_1_linear_bias, weight = encoder_ffn1_2_module_sequential_1_linear_weight, x = input_85)[name = tensor("linear_19")]; - tensor input_87 = silu(x = inputs_21)[name = tensor("input_87")]; - tensor input_91 = linear(bias = encoder_ffn1_2_module_sequential_4_linear_bias, weight = encoder_ffn1_2_module_sequential_4_linear_weight, x = input_87)[name = tensor("linear_20")]; - tensor var_575 = const()[name = tensor("op_575"), val = tensor(0x1p-1)]; - tensor var_576 = mul(x = input_91, y = var_575)[name = tensor("op_576")]; - tensor input_93 = add(x = var_576, y = input_83)[name = tensor("input_93")]; + tensor input_85 = layer_norm(axes = input_85_axes_0, beta = inner_encoder_layer_norm_1_bias, epsilon = var_66, gamma = inner_encoder_layer_norm_1_weight, x = input_83)[name = tensor("input_85")]; + tensor input_87_axes_0 = const()[name = tensor("input_87_axes_0"), val = tensor([-1])]; + tensor input_87 = layer_norm(axes = input_87_axes_0, beta = inner_encoder_ffn1_2_module_sequential_0_bias, epsilon = var_66, gamma = inner_encoder_ffn1_2_module_sequential_0_weight, x = input_85)[name = tensor("input_87")]; + tensor inputs_21 = linear(bias = inner_encoder_ffn1_2_module_sequential_1_linear_bias, weight = inner_encoder_ffn1_2_module_sequential_1_linear_weight, x = input_87)[name = tensor("linear_19")]; + tensor input_89 = silu(x = inputs_21)[name = tensor("input_89")]; + tensor input_93 = linear(bias = inner_encoder_ffn1_2_module_sequential_4_linear_bias, weight = inner_encoder_ffn1_2_module_sequential_4_linear_weight, x = input_89)[name = tensor("linear_20")]; + tensor var_631 = const()[name = tensor("op_631"), val = tensor(0x1p-1)]; + tensor var_632 = mul(x = input_93, y = var_631)[name = tensor("op_632")]; + tensor input_95 = add(x = var_632, y = input_85)[name = tensor("input_95")]; tensor x_13_axes_0 = const()[name = tensor("x_13_axes_0"), val = tensor([-1])]; - tensor x_13 = layer_norm(axes = x_13_axes_0, beta = encoder_ret_lns_2_bias, epsilon = var_28, gamma = encoder_ret_lns_2_weight, x = input_93)[name = tensor("x_13")]; + tensor x_13 = layer_norm(axes = x_13_axes_0, beta = inner_encoder_ret_lns_2_bias, epsilon = var_66, gamma = inner_encoder_ret_lns_2_weight, x = input_95)[name = tensor("x_13")]; tensor prev_kv_5_begin_0 = const()[name = tensor("prev_kv_5_begin_0"), val = tensor([2, 0, 0, 0, 0])]; tensor prev_kv_5_end_0 = const()[name = tensor("prev_kv_5_end_0"), val = tensor([3, 1, 4, 64, 64])]; tensor prev_kv_5_end_mask_0 = const()[name = tensor("prev_kv_5_end_mask_0"), val = tensor([false, true, true, true, true])]; @@ -593,173 +615,173 @@ program(1.0) tensor prev_scale_5_end_mask_0 = const()[name = tensor("prev_scale_5_end_mask_0"), val = tensor([false, true])]; tensor prev_scale_5_squeeze_mask_0 = const()[name = tensor("prev_scale_5_squeeze_mask_0"), val = tensor([true, false])]; tensor prev_scale_5 = slice_by_index(begin = prev_scale_5_begin_0, end = prev_scale_5_end_0, end_mask = prev_scale_5_end_mask_0, squeeze_mask = prev_scale_5_squeeze_mask_0, x = enc_scale)[name = tensor("prev_scale_5")]; - tensor var_590 = linear(bias = encoder_q_proj_2_bias, weight = encoder_q_proj_2_weight, x = x_13)[name = tensor("linear_21")]; - tensor var_591 = const()[name = tensor("op_591"), val = tensor([1, 4, 4, 64])]; - tensor var_592 = reshape(shape = var_591, x = var_590)[name = tensor("op_592")]; + tensor var_646 = linear(bias = inner_encoder_q_proj_2_bias, weight = inner_encoder_q_proj_2_weight, x = x_13)[name = tensor("linear_21")]; + tensor var_647 = const()[name = tensor("op_647"), val = tensor([1, 4, 4, 64])]; + tensor var_648 = reshape(shape = var_647, x = var_646)[name = tensor("op_648")]; tensor q_5_perm_0 = const()[name = tensor("q_5_perm_0"), val = tensor([0, 2, 1, 3])]; - tensor var_596 = linear(bias = encoder_k_proj_2_bias, weight = encoder_k_proj_2_weight, x = x_13)[name = tensor("linear_22")]; - tensor var_597 = const()[name = tensor("op_597"), val = tensor(0x1p-3)]; - tensor var_598 = mul(x = var_596, y = var_597)[name = tensor("op_598")]; - tensor var_599 = const()[name = tensor("op_599"), val = tensor([1, 4, 4, 64])]; - tensor var_600 = reshape(shape = var_599, x = var_598)[name = tensor("op_600")]; + tensor var_652 = linear(bias = inner_encoder_k_proj_2_bias, weight = inner_encoder_k_proj_2_weight, x = x_13)[name = tensor("linear_22")]; + tensor var_653 = const()[name = tensor("op_653"), val = tensor(0x1p-3)]; + tensor var_654 = mul(x = var_652, y = var_653)[name = tensor("op_654")]; + tensor var_655 = const()[name = tensor("op_655"), val = tensor([1, 4, 4, 64])]; + tensor var_656 = reshape(shape = var_655, x = var_654)[name = tensor("op_656")]; tensor k_5_perm_0 = const()[name = tensor("k_5_perm_0"), val = tensor([0, 2, 1, 3])]; - tensor var_604 = linear(bias = encoder_v_proj_2_bias, weight = encoder_v_proj_2_weight, x = x_13)[name = tensor("linear_23")]; - tensor var_605 = const()[name = tensor("op_605"), val = tensor([1, 4, 4, 64])]; - tensor var_606 = reshape(shape = var_605, x = var_604)[name = tensor("op_606")]; + tensor var_660 = linear(bias = inner_encoder_v_proj_2_bias, weight = inner_encoder_v_proj_2_weight, x = x_13)[name = tensor("linear_23")]; + tensor var_661 = const()[name = tensor("op_661"), val = tensor([1, 4, 4, 64])]; + tensor var_662 = reshape(shape = var_661, x = var_660)[name = tensor("op_662")]; tensor v_5_perm_0 = const()[name = tensor("v_5_perm_0"), val = tensor([0, 2, 1, 3])]; - tensor input_97 = linear(bias = encoder_g_proj_2_bias, weight = encoder_g_proj_2_weight, x = x_13)[name = tensor("linear_24")]; + tensor input_99 = linear(bias = inner_encoder_g_proj_2_bias, weight = inner_encoder_g_proj_2_weight, x = x_13)[name = tensor("linear_24")]; tensor sqrt_s0_5 = sqrt(x = prev_scale_5)[name = tensor("sqrt_s0_5")]; - tensor s_t_5 = add(x = prev_scale_5, y = encoder__t_index)[name = tensor("s_t_5")]; + tensor s_t_5 = add(x = prev_scale_5, y = inner_encoder__t_index)[name = tensor("s_t_5")]; tensor sqrt_s_t_5 = sqrt(x = s_t_5)[name = tensor("sqrt_s_t_5")]; tensor qk_5_transpose_x_1 = const()[name = tensor("qk_5_transpose_x_1"), val = tensor(false)]; tensor qk_5_transpose_y_1 = const()[name = tensor("qk_5_transpose_y_1"), val = tensor(true)]; - tensor k_5 = transpose(perm = k_5_perm_0, x = var_600)[name = tensor("transpose_43")]; - tensor q_5 = transpose(perm = q_5_perm_0, x = var_592)[name = tensor("transpose_44")]; + tensor k_5 = transpose(perm = k_5_perm_0, x = var_656)[name = tensor("transpose_43")]; + tensor q_5 = transpose(perm = q_5_perm_0, x = var_648)[name = tensor("transpose_44")]; tensor qk_5 = matmul(transpose_x = qk_5_transpose_x_1, transpose_y = qk_5_transpose_y_1, x = q_5, y = k_5)[name = tensor("qk_5")]; - tensor var_616 = const()[name = tensor("op_616"), val = tensor([4, 1])]; - tensor var_617 = reshape(shape = var_616, x = sqrt_s_t_5)[name = tensor("op_617")]; - tensor M_5 = real_div(x = encoder__causal_mask, y = var_617)[name = tensor("M_5")]; - tensor var_619 = mul(x = qk_5, y = M_5)[name = tensor("op_619")]; + tensor var_672 = const()[name = tensor("op_672"), val = tensor([4, 1])]; + tensor var_673 = reshape(shape = var_672, x = sqrt_s_t_5)[name = tensor("op_673")]; + tensor M_5 = real_div(x = inner_encoder__causal_mask, y = var_673)[name = tensor("M_5")]; + tensor var_675 = mul(x = qk_5, y = M_5)[name = tensor("op_675")]; tensor inner_5_transpose_x_0 = const()[name = tensor("inner_5_transpose_x_0"), val = tensor(false)]; tensor inner_5_transpose_y_0 = const()[name = tensor("inner_5_transpose_y_0"), val = tensor(false)]; - tensor v_5 = transpose(perm = v_5_perm_0, x = var_606)[name = tensor("transpose_42")]; - tensor inner_5 = matmul(transpose_x = inner_5_transpose_x_0, transpose_y = inner_5_transpose_y_0, x = var_619, y = v_5)[name = tensor("inner_5")]; - tensor var_621_transpose_x_0 = const()[name = tensor("op_621_transpose_x_0"), val = tensor(false)]; - tensor var_621_transpose_y_0 = const()[name = tensor("op_621_transpose_y_0"), val = tensor(false)]; - tensor var_621 = matmul(transpose_x = var_621_transpose_x_0, transpose_y = var_621_transpose_y_0, x = q_5, y = prev_kv_5)[name = tensor("op_621")]; - tensor var_622 = real_div(x = sqrt_s0_5, y = sqrt_s_t_5)[name = tensor("op_622")]; - tensor var_623 = const()[name = tensor("op_623"), val = tensor([1, 1, 4, 1])]; - tensor var_624 = reshape(shape = var_623, x = var_622)[name = tensor("op_624")]; - tensor cross_5 = mul(x = var_621, y = var_624)[name = tensor("cross_5")]; + tensor v_5 = transpose(perm = v_5_perm_0, x = var_662)[name = tensor("transpose_42")]; + tensor inner_5 = matmul(transpose_x = inner_5_transpose_x_0, transpose_y = inner_5_transpose_y_0, x = var_675, y = v_5)[name = tensor("inner_5")]; + tensor var_677_transpose_x_0 = const()[name = tensor("op_677_transpose_x_0"), val = tensor(false)]; + tensor var_677_transpose_y_0 = const()[name = tensor("op_677_transpose_y_0"), val = tensor(false)]; + tensor var_677 = matmul(transpose_x = var_677_transpose_x_0, transpose_y = var_677_transpose_y_0, x = q_5, y = prev_kv_5)[name = tensor("op_677")]; + tensor var_678 = real_div(x = sqrt_s0_5, y = sqrt_s_t_5)[name = tensor("op_678")]; + tensor var_679 = const()[name = tensor("op_679"), val = tensor([1, 1, 4, 1])]; + tensor var_680 = reshape(shape = var_679, x = var_678)[name = tensor("op_680")]; + tensor cross_5 = mul(x = var_677, y = var_680)[name = tensor("cross_5")]; tensor out_13 = add(x = inner_5, y = cross_5)[name = tensor("out_13")]; - tensor var_627 = mul(x = prev_kv_5, y = sqrt_s0_5)[name = tensor("op_627")]; - tensor var_629_transpose_x_1 = const()[name = tensor("op_629_transpose_x_1"), val = tensor(true)]; - tensor var_629_transpose_y_1 = const()[name = tensor("op_629_transpose_y_1"), val = tensor(false)]; - tensor var_629 = matmul(transpose_x = var_629_transpose_x_1, transpose_y = var_629_transpose_y_1, x = k_5, y = v_5)[name = tensor("op_629")]; - tensor new_kv_unnorm_5 = add(x = var_627, y = var_629)[name = tensor("new_kv_unnorm_5")]; - tensor var_631 = const()[name = tensor("op_631"), val = tensor(0x1p+2)]; - tensor new_scale_5 = add(x = prev_scale_5, y = var_631)[name = tensor("new_scale_5")]; - tensor var_633 = sqrt(x = new_scale_5)[name = tensor("op_633")]; - tensor var_634 = real_div(x = new_kv_unnorm_5, y = var_633)[name = tensor("op_634")]; - tensor var_635_perm_0 = const()[name = tensor("op_635_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_683 = mul(x = prev_kv_5, y = sqrt_s0_5)[name = tensor("op_683")]; + tensor var_685_transpose_x_1 = const()[name = tensor("op_685_transpose_x_1"), val = tensor(true)]; + tensor var_685_transpose_y_1 = const()[name = tensor("op_685_transpose_y_1"), val = tensor(false)]; + tensor var_685 = matmul(transpose_x = var_685_transpose_x_1, transpose_y = var_685_transpose_y_1, x = k_5, y = v_5)[name = tensor("op_685")]; + tensor new_kv_unnorm_5 = add(x = var_683, y = var_685)[name = tensor("new_kv_unnorm_5")]; + tensor var_687 = const()[name = tensor("op_687"), val = tensor(0x1p+2)]; + tensor new_scale_5 = add(x = prev_scale_5, y = var_687)[name = tensor("new_scale_5")]; + tensor var_689 = sqrt(x = new_scale_5)[name = tensor("op_689")]; + tensor var_690 = real_div(x = new_kv_unnorm_5, y = var_689)[name = tensor("op_690")]; + tensor var_691_perm_0 = const()[name = tensor("op_691_perm_0"), val = tensor([0, 2, 1, 3])]; tensor out_15_axes_0 = const()[name = tensor("out_15_axes_0"), val = tensor([-1])]; - tensor var_635 = transpose(perm = var_635_perm_0, x = out_13)[name = tensor("transpose_41")]; - tensor out_15 = layer_norm(axes = out_15_axes_0, epsilon = var_17, x = var_635)[name = tensor("out_15")]; - tensor var_639 = const()[name = tensor("op_639"), val = tensor([1, 4, 256])]; - tensor out_17 = reshape(shape = var_639, x = out_15)[name = tensor("out_17")]; - tensor var_641 = silu(x = input_97)[name = tensor("op_641")]; - tensor input_99 = mul(x = var_641, y = out_17)[name = tensor("input_99")]; - tensor ret_out_5 = linear(bias = encoder_out_proj_2_bias, weight = encoder_out_proj_2_weight, x = input_99)[name = tensor("linear_25")]; - tensor x_15 = add(x = input_93, y = ret_out_5)[name = tensor("x_15")]; + tensor var_691 = transpose(perm = var_691_perm_0, x = out_13)[name = tensor("transpose_41")]; + tensor out_15 = layer_norm(axes = out_15_axes_0, epsilon = var_74, x = var_691)[name = tensor("out_15")]; + tensor var_695 = const()[name = tensor("op_695"), val = tensor([1, 4, 256])]; + tensor out_17 = reshape(shape = var_695, x = out_15)[name = tensor("out_17")]; + tensor var_697 = silu(x = input_99)[name = tensor("op_697")]; + tensor input_101 = mul(x = var_697, y = out_17)[name = tensor("input_101")]; + tensor ret_out_5 = linear(bias = inner_encoder_out_proj_2_bias, weight = inner_encoder_out_proj_2_weight, x = input_101)[name = tensor("linear_25")]; + tensor x_15 = add(x = input_95, y = ret_out_5)[name = tensor("x_15")]; tensor window_21_begin_0 = const()[name = tensor("window_21_begin_0"), val = tensor([2, 0, 0, 0])]; tensor window_21_end_0 = const()[name = tensor("window_21_end_0"), val = tensor([3, 1, 16, 256])]; tensor window_21_end_mask_0 = const()[name = tensor("window_21_end_mask_0"), val = tensor([false, true, true, true])]; tensor window_21_squeeze_mask_0 = const()[name = tensor("window_21_squeeze_mask_0"), val = tensor([true, false, false, false])]; tensor window_21 = slice_by_index(begin = window_21_begin_0, end = window_21_end_0, end_mask = window_21_end_mask_0, squeeze_mask = window_21_squeeze_mask_0, x = enc_conv_cache)[name = tensor("window_21")]; - tensor var_649_begin_0 = const()[name = tensor("op_649_begin_0"), val = tensor([0, 0, 0])]; - tensor var_649_end_0 = const()[name = tensor("op_649_end_0"), val = tensor([1, 1, 256])]; - tensor var_649_end_mask_0 = const()[name = tensor("op_649_end_mask_0"), val = tensor([true, false, true])]; - tensor var_649 = slice_by_index(begin = var_649_begin_0, end = var_649_end_0, end_mask = var_649_end_mask_0, x = x_15)[name = tensor("op_649")]; - tensor var_652_begin_0 = const()[name = tensor("op_652_begin_0"), val = tensor([0, 1, 0])]; - tensor var_652_end_0 = const()[name = tensor("op_652_end_0"), val = tensor([1, 16, 256])]; - tensor var_652_end_mask_0 = const()[name = tensor("op_652_end_mask_0"), val = tensor([true, true, true])]; - tensor var_652 = slice_by_index(begin = var_652_begin_0, end = var_652_end_0, end_mask = var_652_end_mask_0, x = window_21)[name = tensor("op_652")]; + tensor var_705_begin_0 = const()[name = tensor("op_705_begin_0"), val = tensor([0, 0, 0])]; + tensor var_705_end_0 = const()[name = tensor("op_705_end_0"), val = tensor([1, 1, 256])]; + tensor var_705_end_mask_0 = const()[name = tensor("op_705_end_mask_0"), val = tensor([true, false, true])]; + tensor var_705 = slice_by_index(begin = var_705_begin_0, end = var_705_end_0, end_mask = var_705_end_mask_0, x = x_15)[name = tensor("op_705")]; + tensor var_708_begin_0 = const()[name = tensor("op_708_begin_0"), val = tensor([0, 1, 0])]; + tensor var_708_end_0 = const()[name = tensor("op_708_end_0"), val = tensor([1, 16, 256])]; + tensor var_708_end_mask_0 = const()[name = tensor("op_708_end_mask_0"), val = tensor([true, true, true])]; + tensor var_708 = slice_by_index(begin = var_708_begin_0, end = var_708_end_0, end_mask = var_708_end_mask_0, x = window_21)[name = tensor("op_708")]; tensor window_23_interleave_0 = const()[name = tensor("window_23_interleave_0"), val = tensor(false)]; - tensor window_23 = concat(axis = var_26, interleave = window_23_interleave_0, values = (var_652, var_649))[name = tensor("window_23")]; - tensor var_657_begin_0 = const()[name = tensor("op_657_begin_0"), val = tensor([0, 1, 0])]; - tensor var_657_end_0 = const()[name = tensor("op_657_end_0"), val = tensor([1, 2, 256])]; - tensor var_657_end_mask_0 = const()[name = tensor("op_657_end_mask_0"), val = tensor([true, false, true])]; - tensor var_657 = slice_by_index(begin = var_657_begin_0, end = var_657_end_0, end_mask = var_657_end_mask_0, x = x_15)[name = tensor("op_657")]; - tensor var_660_begin_0 = const()[name = tensor("op_660_begin_0"), val = tensor([0, 1, 0])]; - tensor var_660_end_0 = const()[name = tensor("op_660_end_0"), val = tensor([1, 16, 256])]; - tensor var_660_end_mask_0 = const()[name = tensor("op_660_end_mask_0"), val = tensor([true, true, true])]; - tensor var_660 = slice_by_index(begin = var_660_begin_0, end = var_660_end_0, end_mask = var_660_end_mask_0, x = window_23)[name = tensor("op_660")]; + tensor window_23 = concat(axis = var_82, interleave = window_23_interleave_0, values = (var_708, var_705))[name = tensor("window_23")]; + tensor var_713_begin_0 = const()[name = tensor("op_713_begin_0"), val = tensor([0, 1, 0])]; + tensor var_713_end_0 = const()[name = tensor("op_713_end_0"), val = tensor([1, 2, 256])]; + tensor var_713_end_mask_0 = const()[name = tensor("op_713_end_mask_0"), val = tensor([true, false, true])]; + tensor var_713 = slice_by_index(begin = var_713_begin_0, end = var_713_end_0, end_mask = var_713_end_mask_0, x = x_15)[name = tensor("op_713")]; + tensor var_716_begin_0 = const()[name = tensor("op_716_begin_0"), val = tensor([0, 1, 0])]; + tensor var_716_end_0 = const()[name = tensor("op_716_end_0"), val = tensor([1, 16, 256])]; + tensor var_716_end_mask_0 = const()[name = tensor("op_716_end_mask_0"), val = tensor([true, true, true])]; + tensor var_716 = slice_by_index(begin = var_716_begin_0, end = var_716_end_0, end_mask = var_716_end_mask_0, x = window_23)[name = tensor("op_716")]; tensor window_25_interleave_0 = const()[name = tensor("window_25_interleave_0"), val = tensor(false)]; - tensor window_25 = concat(axis = var_26, interleave = window_25_interleave_0, values = (var_660, var_657))[name = tensor("window_25")]; - tensor var_665_begin_0 = const()[name = tensor("op_665_begin_0"), val = tensor([0, 2, 0])]; - tensor var_665_end_0 = const()[name = tensor("op_665_end_0"), val = tensor([1, 3, 256])]; - tensor var_665_end_mask_0 = const()[name = tensor("op_665_end_mask_0"), val = tensor([true, false, true])]; - tensor var_665 = slice_by_index(begin = var_665_begin_0, end = var_665_end_0, end_mask = var_665_end_mask_0, x = x_15)[name = tensor("op_665")]; - tensor var_668_begin_0 = const()[name = tensor("op_668_begin_0"), val = tensor([0, 1, 0])]; - tensor var_668_end_0 = const()[name = tensor("op_668_end_0"), val = tensor([1, 16, 256])]; - tensor var_668_end_mask_0 = const()[name = tensor("op_668_end_mask_0"), val = tensor([true, true, true])]; - tensor var_668 = slice_by_index(begin = var_668_begin_0, end = var_668_end_0, end_mask = var_668_end_mask_0, x = window_25)[name = tensor("op_668")]; + tensor window_25 = concat(axis = var_82, interleave = window_25_interleave_0, values = (var_716, var_713))[name = tensor("window_25")]; + tensor var_721_begin_0 = const()[name = tensor("op_721_begin_0"), val = tensor([0, 2, 0])]; + tensor var_721_end_0 = const()[name = tensor("op_721_end_0"), val = tensor([1, 3, 256])]; + tensor var_721_end_mask_0 = const()[name = tensor("op_721_end_mask_0"), val = tensor([true, false, true])]; + tensor var_721 = slice_by_index(begin = var_721_begin_0, end = var_721_end_0, end_mask = var_721_end_mask_0, x = x_15)[name = tensor("op_721")]; + tensor var_724_begin_0 = const()[name = tensor("op_724_begin_0"), val = tensor([0, 1, 0])]; + tensor var_724_end_0 = const()[name = tensor("op_724_end_0"), val = tensor([1, 16, 256])]; + tensor var_724_end_mask_0 = const()[name = tensor("op_724_end_mask_0"), val = tensor([true, true, true])]; + tensor var_724 = slice_by_index(begin = var_724_begin_0, end = var_724_end_0, end_mask = var_724_end_mask_0, x = window_25)[name = tensor("op_724")]; tensor window_27_interleave_0 = const()[name = tensor("window_27_interleave_0"), val = tensor(false)]; - tensor window_27 = concat(axis = var_26, interleave = window_27_interleave_0, values = (var_668, var_665))[name = tensor("window_27")]; - tensor var_673_begin_0 = const()[name = tensor("op_673_begin_0"), val = tensor([0, 3, 0])]; - tensor var_673_end_0 = const()[name = tensor("op_673_end_0"), val = tensor([1, 1, 256])]; - tensor var_673_end_mask_0 = const()[name = tensor("op_673_end_mask_0"), val = tensor([true, true, true])]; - tensor var_673 = slice_by_index(begin = var_673_begin_0, end = var_673_end_0, end_mask = var_673_end_mask_0, x = x_15)[name = tensor("op_673")]; - tensor var_676_begin_0 = const()[name = tensor("op_676_begin_0"), val = tensor([0, 1, 0])]; - tensor var_676_end_0 = const()[name = tensor("op_676_end_0"), val = tensor([1, 16, 256])]; - tensor var_676_end_mask_0 = const()[name = tensor("op_676_end_mask_0"), val = tensor([true, true, true])]; - tensor var_676 = slice_by_index(begin = var_676_begin_0, end = var_676_end_0, end_mask = var_676_end_mask_0, x = window_27)[name = tensor("op_676")]; + tensor window_27 = concat(axis = var_82, interleave = window_27_interleave_0, values = (var_724, var_721))[name = tensor("window_27")]; + tensor var_729_begin_0 = const()[name = tensor("op_729_begin_0"), val = tensor([0, 3, 0])]; + tensor var_729_end_0 = const()[name = tensor("op_729_end_0"), val = tensor([1, 1, 256])]; + tensor var_729_end_mask_0 = const()[name = tensor("op_729_end_mask_0"), val = tensor([true, true, true])]; + tensor var_729 = slice_by_index(begin = var_729_begin_0, end = var_729_end_0, end_mask = var_729_end_mask_0, x = x_15)[name = tensor("op_729")]; + tensor var_732_begin_0 = const()[name = tensor("op_732_begin_0"), val = tensor([0, 1, 0])]; + tensor var_732_end_0 = const()[name = tensor("op_732_end_0"), val = tensor([1, 16, 256])]; + tensor var_732_end_mask_0 = const()[name = tensor("op_732_end_mask_0"), val = tensor([true, true, true])]; + tensor var_732 = slice_by_index(begin = var_732_begin_0, end = var_732_end_0, end_mask = var_732_end_mask_0, x = window_27)[name = tensor("op_732")]; tensor window_29_interleave_0 = const()[name = tensor("window_29_interleave_0"), val = tensor(false)]; - tensor window_29 = concat(axis = var_26, interleave = window_29_interleave_0, values = (var_676, var_673))[name = tensor("window_29")]; - tensor input_101_interleave_0 = const()[name = tensor("input_101_interleave_0"), val = tensor(false)]; - tensor input_101 = concat(axis = var_23, interleave = input_101_interleave_0, values = (window_23, window_25, window_27, window_29))[name = tensor("input_101")]; + tensor window_29 = concat(axis = var_82, interleave = window_29_interleave_0, values = (var_732, var_729))[name = tensor("window_29")]; + tensor input_103_interleave_0 = const()[name = tensor("input_103_interleave_0"), val = tensor(false)]; + tensor input_103 = concat(axis = var_69, interleave = input_103_interleave_0, values = (window_23, window_25, window_27, window_29))[name = tensor("input_103")]; tensor x_17_axes_0 = const()[name = tensor("x_17_axes_0"), val = tensor([-1])]; - tensor x_17 = layer_norm(axes = x_17_axes_0, beta = encoder_conv_module_2_sequential_0_bias, epsilon = var_28, gamma = encoder_conv_module_2_sequential_0_weight, x = input_101)[name = tensor("x_17")]; - tensor input_103_perm_0 = const()[name = tensor("input_103_perm_0"), val = tensor([0, 2, 1])]; + tensor x_17 = layer_norm(axes = x_17_axes_0, beta = inner_encoder_conv_module_2_sequential_0_bias, epsilon = var_66, gamma = inner_encoder_conv_module_2_sequential_0_weight, x = input_103)[name = tensor("x_17")]; + tensor input_105_perm_0 = const()[name = tensor("input_105_perm_0"), val = tensor([0, 2, 1])]; tensor inputs_23_pad_type_0 = const()[name = tensor("inputs_23_pad_type_0"), val = tensor("valid")]; tensor inputs_23_strides_0 = const()[name = tensor("inputs_23_strides_0"), val = tensor([1])]; tensor inputs_23_pad_0 = const()[name = tensor("inputs_23_pad_0"), val = tensor([0, 0])]; tensor inputs_23_dilations_0 = const()[name = tensor("inputs_23_dilations_0"), val = tensor([1])]; tensor inputs_23_groups_0 = const()[name = tensor("inputs_23_groups_0"), val = tensor(1)]; - tensor input_103 = transpose(perm = input_103_perm_0, x = x_17)[name = tensor("transpose_40")]; - tensor inputs_23 = conv(bias = encoder_conv_module_2_sequential_2_conv_bias, dilations = inputs_23_dilations_0, groups = inputs_23_groups_0, pad = inputs_23_pad_0, pad_type = inputs_23_pad_type_0, strides = inputs_23_strides_0, weight = encoder_conv_module_2_sequential_2_conv_weight, x = input_103)[name = tensor("inputs_23")]; - tensor var_701_split_sizes_0 = const()[name = tensor("op_701_split_sizes_0"), val = tensor([256, 256])]; - tensor var_701_axis_0 = const()[name = tensor("op_701_axis_0"), val = tensor(1)]; - tensor var_701_0, tensor var_701_1 = split(axis = var_701_axis_0, split_sizes = var_701_split_sizes_0, x = inputs_23)[name = tensor("op_701")]; - tensor var_703 = sigmoid(x = var_701_1)[name = tensor("op_703")]; - tensor inputs_25 = mul(x = var_701_0, y = var_703)[name = tensor("inputs_25")]; + tensor input_105 = transpose(perm = input_105_perm_0, x = x_17)[name = tensor("transpose_40")]; + tensor inputs_23 = conv(bias = inner_encoder_conv_module_2_sequential_2_conv_bias, dilations = inputs_23_dilations_0, groups = inputs_23_groups_0, pad = inputs_23_pad_0, pad_type = inputs_23_pad_type_0, strides = inputs_23_strides_0, weight = inner_encoder_conv_module_2_sequential_2_conv_weight, x = input_105)[name = tensor("inputs_23")]; + tensor var_757_split_sizes_0 = const()[name = tensor("op_757_split_sizes_0"), val = tensor([256, 256])]; + tensor var_757_axis_0 = const()[name = tensor("op_757_axis_0"), val = tensor(1)]; + tensor var_757_0, tensor var_757_1 = split(axis = var_757_axis_0, split_sizes = var_757_split_sizes_0, x = inputs_23)[name = tensor("op_757")]; + tensor var_759 = sigmoid(x = var_757_1)[name = tensor("op_759")]; + tensor inputs_25 = mul(x = var_757_0, y = var_759)[name = tensor("inputs_25")]; tensor outputs_aug_5_pad_type_0 = const()[name = tensor("outputs_aug_5_pad_type_0"), val = tensor("custom")]; tensor outputs_aug_5_pad_0 = const()[name = tensor("outputs_aug_5_pad_0"), val = tensor([15, 15])]; tensor outputs_aug_5_groups_0 = const()[name = tensor("outputs_aug_5_groups_0"), val = tensor(256)]; tensor outputs_aug_5_strides_0 = const()[name = tensor("outputs_aug_5_strides_0"), val = tensor([1])]; tensor outputs_aug_5_dilations_0 = const()[name = tensor("outputs_aug_5_dilations_0"), val = tensor([1])]; - tensor outputs_aug_5 = conv(dilations = outputs_aug_5_dilations_0, groups = outputs_aug_5_groups_0, pad = outputs_aug_5_pad_0, pad_type = outputs_aug_5_pad_type_0, strides = outputs_aug_5_strides_0, weight = encoder_conv_module_2_sequential_4_conv_weight, x = inputs_25)[name = tensor("outputs_aug_5")]; - tensor input_105_begin_0 = const()[name = tensor("input_105_begin_0"), val = tensor([0, 0, 0])]; - tensor input_105_end_0 = const()[name = tensor("input_105_end_0"), val = tensor([4, 256, 16])]; - tensor input_105_end_mask_0 = const()[name = tensor("input_105_end_mask_0"), val = tensor([true, true, false])]; - tensor input_105 = slice_by_index(begin = input_105_begin_0, end = input_105_end_0, end_mask = input_105_end_mask_0, x = outputs_aug_5)[name = tensor("input_105")]; - tensor inputs_27 = batch_norm(beta = encoder_conv_module_2_sequential_5_bias, epsilon = var_28, gamma = encoder_conv_module_2_sequential_5_weight, mean = encoder_conv_module_2_sequential_5_running_mean, variance = encoder_conv_module_2_sequential_5_running_var, x = input_105)[name = tensor("inputs_27")]; - tensor input_107 = silu(x = inputs_27)[name = tensor("input_107")]; - tensor input_109_pad_type_0 = const()[name = tensor("input_109_pad_type_0"), val = tensor("valid")]; - tensor input_109_strides_0 = const()[name = tensor("input_109_strides_0"), val = tensor([1])]; - tensor input_109_pad_0 = const()[name = tensor("input_109_pad_0"), val = tensor([0, 0])]; - tensor input_109_dilations_0 = const()[name = tensor("input_109_dilations_0"), val = tensor([1])]; - tensor input_109_groups_0 = const()[name = tensor("input_109_groups_0"), val = tensor(1)]; - tensor input_109 = conv(bias = encoder_conv_module_2_sequential_7_conv_bias, dilations = input_109_dilations_0, groups = input_109_groups_0, pad = input_109_pad_0, pad_type = input_109_pad_type_0, strides = input_109_strides_0, weight = encoder_conv_module_2_sequential_7_conv_weight, x = input_107)[name = tensor("input_109")]; + tensor outputs_aug_5 = conv(dilations = outputs_aug_5_dilations_0, groups = outputs_aug_5_groups_0, pad = outputs_aug_5_pad_0, pad_type = outputs_aug_5_pad_type_0, strides = outputs_aug_5_strides_0, weight = inner_encoder_conv_module_2_sequential_4_conv_weight, x = inputs_25)[name = tensor("outputs_aug_5")]; + tensor input_107_begin_0 = const()[name = tensor("input_107_begin_0"), val = tensor([0, 0, 0])]; + tensor input_107_end_0 = const()[name = tensor("input_107_end_0"), val = tensor([4, 256, 16])]; + tensor input_107_end_mask_0 = const()[name = tensor("input_107_end_mask_0"), val = tensor([true, true, false])]; + tensor input_107 = slice_by_index(begin = input_107_begin_0, end = input_107_end_0, end_mask = input_107_end_mask_0, x = outputs_aug_5)[name = tensor("input_107")]; + tensor inputs_27 = batch_norm(beta = inner_encoder_conv_module_2_sequential_5_bias, epsilon = var_66, gamma = inner_encoder_conv_module_2_sequential_5_weight, mean = inner_encoder_conv_module_2_sequential_5_running_mean, variance = inner_encoder_conv_module_2_sequential_5_running_var, x = input_107)[name = tensor("inputs_27")]; + tensor input_109 = silu(x = inputs_27)[name = tensor("input_109")]; + tensor input_111_pad_type_0 = const()[name = tensor("input_111_pad_type_0"), val = tensor("valid")]; + tensor input_111_strides_0 = const()[name = tensor("input_111_strides_0"), val = tensor([1])]; + tensor input_111_pad_0 = const()[name = tensor("input_111_pad_0"), val = tensor([0, 0])]; + tensor input_111_dilations_0 = const()[name = tensor("input_111_dilations_0"), val = tensor([1])]; + tensor input_111_groups_0 = const()[name = tensor("input_111_groups_0"), val = tensor(1)]; + tensor input_111 = conv(bias = inner_encoder_conv_module_2_sequential_7_conv_bias, dilations = input_111_dilations_0, groups = input_111_groups_0, pad = input_111_pad_0, pad_type = input_111_pad_type_0, strides = input_111_strides_0, weight = inner_encoder_conv_module_2_sequential_7_conv_weight, x = input_109)[name = tensor("input_111")]; tensor conv_out_5_perm_0 = const()[name = tensor("conv_out_5_perm_0"), val = tensor([0, 2, 1])]; - tensor var_734_begin_0 = const()[name = tensor("op_734_begin_0"), val = tensor([0, -1, 0])]; - tensor var_734_end_0 = const()[name = tensor("op_734_end_0"), val = tensor([4, 16, 256])]; - tensor var_734_end_mask_0 = const()[name = tensor("op_734_end_mask_0"), val = tensor([true, true, true])]; - tensor conv_out_5 = transpose(perm = conv_out_5_perm_0, x = input_109)[name = tensor("transpose_39")]; - tensor var_734 = slice_by_index(begin = var_734_begin_0, end = var_734_end_0, end_mask = var_734_end_mask_0, x = conv_out_5)[name = tensor("op_734")]; - tensor var_736_perm_0 = const()[name = tensor("op_736_perm_0"), val = tensor([1, 0, 2])]; - tensor var_736 = transpose(perm = var_736_perm_0, x = var_734)[name = tensor("transpose_38")]; - tensor input_111 = add(x = x_15, y = var_736)[name = tensor("input_111")]; - tensor input_113_axes_0 = const()[name = tensor("input_113_axes_0"), val = tensor([-1])]; - tensor input_113 = layer_norm(axes = input_113_axes_0, beta = encoder_ffn2_2_module_sequential_0_bias, epsilon = var_28, gamma = encoder_ffn2_2_module_sequential_0_weight, x = input_111)[name = tensor("input_113")]; - tensor inputs_29 = linear(bias = encoder_ffn2_2_module_sequential_1_linear_bias, weight = encoder_ffn2_2_module_sequential_1_linear_weight, x = input_113)[name = tensor("linear_26")]; - tensor input_115 = silu(x = inputs_29)[name = tensor("input_115")]; - tensor input_119 = linear(bias = encoder_ffn2_2_module_sequential_4_linear_bias, weight = encoder_ffn2_2_module_sequential_4_linear_weight, x = input_115)[name = tensor("linear_27")]; - tensor var_759 = const()[name = tensor("op_759"), val = tensor(0x1p-1)]; - tensor var_760 = mul(x = input_119, y = var_759)[name = tensor("op_760")]; - tensor input_121 = add(x = var_760, y = input_111)[name = tensor("input_121")]; - tensor input_123_axes_0 = const()[name = tensor("input_123_axes_0"), val = tensor([-1])]; - tensor input_123 = layer_norm(axes = input_123_axes_0, beta = encoder_layer_norm_2_bias, epsilon = var_28, gamma = encoder_layer_norm_2_weight, x = input_121)[name = tensor("input_123")]; + tensor var_790_begin_0 = const()[name = tensor("op_790_begin_0"), val = tensor([0, -1, 0])]; + tensor var_790_end_0 = const()[name = tensor("op_790_end_0"), val = tensor([4, 16, 256])]; + tensor var_790_end_mask_0 = const()[name = tensor("op_790_end_mask_0"), val = tensor([true, true, true])]; + tensor conv_out_5 = transpose(perm = conv_out_5_perm_0, x = input_111)[name = tensor("transpose_39")]; + tensor var_790 = slice_by_index(begin = var_790_begin_0, end = var_790_end_0, end_mask = var_790_end_mask_0, x = conv_out_5)[name = tensor("op_790")]; + tensor var_792_perm_0 = const()[name = tensor("op_792_perm_0"), val = tensor([1, 0, 2])]; + tensor var_792 = transpose(perm = var_792_perm_0, x = var_790)[name = tensor("transpose_38")]; + tensor input_113 = add(x = x_15, y = var_792)[name = tensor("input_113")]; + tensor input_115_axes_0 = const()[name = tensor("input_115_axes_0"), val = tensor([-1])]; + tensor input_115 = layer_norm(axes = input_115_axes_0, beta = inner_encoder_ffn2_2_module_sequential_0_bias, epsilon = var_66, gamma = inner_encoder_ffn2_2_module_sequential_0_weight, x = input_113)[name = tensor("input_115")]; + tensor inputs_29 = linear(bias = inner_encoder_ffn2_2_module_sequential_1_linear_bias, weight = inner_encoder_ffn2_2_module_sequential_1_linear_weight, x = input_115)[name = tensor("linear_26")]; + tensor input_117 = silu(x = inputs_29)[name = tensor("input_117")]; + tensor input_121 = linear(bias = inner_encoder_ffn2_2_module_sequential_4_linear_bias, weight = inner_encoder_ffn2_2_module_sequential_4_linear_weight, x = input_117)[name = tensor("linear_27")]; + tensor var_815 = const()[name = tensor("op_815"), val = tensor(0x1p-1)]; + tensor var_816 = mul(x = input_121, y = var_815)[name = tensor("op_816")]; + tensor input_123 = add(x = var_816, y = input_113)[name = tensor("input_123")]; tensor input_125_axes_0 = const()[name = tensor("input_125_axes_0"), val = tensor([-1])]; - tensor input_125 = layer_norm(axes = input_125_axes_0, beta = encoder_ffn1_3_module_sequential_0_bias, epsilon = var_28, gamma = encoder_ffn1_3_module_sequential_0_weight, x = input_123)[name = tensor("input_125")]; - tensor inputs_31 = linear(bias = encoder_ffn1_3_module_sequential_1_linear_bias, weight = encoder_ffn1_3_module_sequential_1_linear_weight, x = input_125)[name = tensor("linear_28")]; - tensor input_127 = silu(x = inputs_31)[name = tensor("input_127")]; - tensor input_131 = linear(bias = encoder_ffn1_3_module_sequential_4_linear_bias, weight = encoder_ffn1_3_module_sequential_4_linear_weight, x = input_127)[name = tensor("linear_29")]; - tensor var_789 = const()[name = tensor("op_789"), val = tensor(0x1p-1)]; - tensor var_790 = mul(x = input_131, y = var_789)[name = tensor("op_790")]; - tensor input_133 = add(x = var_790, y = input_123)[name = tensor("input_133")]; + tensor input_125 = layer_norm(axes = input_125_axes_0, beta = inner_encoder_layer_norm_2_bias, epsilon = var_66, gamma = inner_encoder_layer_norm_2_weight, x = input_123)[name = tensor("input_125")]; + tensor input_127_axes_0 = const()[name = tensor("input_127_axes_0"), val = tensor([-1])]; + tensor input_127 = layer_norm(axes = input_127_axes_0, beta = inner_encoder_ffn1_3_module_sequential_0_bias, epsilon = var_66, gamma = inner_encoder_ffn1_3_module_sequential_0_weight, x = input_125)[name = tensor("input_127")]; + tensor inputs_31 = linear(bias = inner_encoder_ffn1_3_module_sequential_1_linear_bias, weight = inner_encoder_ffn1_3_module_sequential_1_linear_weight, x = input_127)[name = tensor("linear_28")]; + tensor input_129 = silu(x = inputs_31)[name = tensor("input_129")]; + tensor input_133 = linear(bias = inner_encoder_ffn1_3_module_sequential_4_linear_bias, weight = inner_encoder_ffn1_3_module_sequential_4_linear_weight, x = input_129)[name = tensor("linear_29")]; + tensor var_845 = const()[name = tensor("op_845"), val = tensor(0x1p-1)]; + tensor var_846 = mul(x = input_133, y = var_845)[name = tensor("op_846")]; + tensor input_135 = add(x = var_846, y = input_125)[name = tensor("input_135")]; tensor x_19_axes_0 = const()[name = tensor("x_19_axes_0"), val = tensor([-1])]; - tensor x_19 = layer_norm(axes = x_19_axes_0, beta = encoder_ret_lns_3_bias, epsilon = var_28, gamma = encoder_ret_lns_3_weight, x = input_133)[name = tensor("x_19")]; + tensor x_19 = layer_norm(axes = x_19_axes_0, beta = inner_encoder_ret_lns_3_bias, epsilon = var_66, gamma = inner_encoder_ret_lns_3_weight, x = input_135)[name = tensor("x_19")]; tensor prev_kv_7_begin_0 = const()[name = tensor("prev_kv_7_begin_0"), val = tensor([3, 0, 0, 0, 0])]; tensor prev_kv_7_end_0 = const()[name = tensor("prev_kv_7_end_0"), val = tensor([4, 1, 4, 64, 64])]; tensor prev_kv_7_end_mask_0 = const()[name = tensor("prev_kv_7_end_mask_0"), val = tensor([false, true, true, true, true])]; @@ -770,209 +792,202 @@ program(1.0) tensor prev_scale_7_end_mask_0 = const()[name = tensor("prev_scale_7_end_mask_0"), val = tensor([false, true])]; tensor prev_scale_7_squeeze_mask_0 = const()[name = tensor("prev_scale_7_squeeze_mask_0"), val = tensor([true, false])]; tensor prev_scale_7 = slice_by_index(begin = prev_scale_7_begin_0, end = prev_scale_7_end_0, end_mask = prev_scale_7_end_mask_0, squeeze_mask = prev_scale_7_squeeze_mask_0, x = enc_scale)[name = tensor("prev_scale_7")]; - tensor var_804 = linear(bias = encoder_q_proj_3_bias, weight = encoder_q_proj_3_weight, x = x_19)[name = tensor("linear_30")]; - tensor var_805 = const()[name = tensor("op_805"), val = tensor([1, 4, 4, 64])]; - tensor var_806 = reshape(shape = var_805, x = var_804)[name = tensor("op_806")]; + tensor var_860 = linear(bias = inner_encoder_q_proj_3_bias, weight = inner_encoder_q_proj_3_weight, x = x_19)[name = tensor("linear_30")]; + tensor var_861 = const()[name = tensor("op_861"), val = tensor([1, 4, 4, 64])]; + tensor var_862 = reshape(shape = var_861, x = var_860)[name = tensor("op_862")]; tensor q_7_perm_0 = const()[name = tensor("q_7_perm_0"), val = tensor([0, 2, 1, 3])]; - tensor var_810 = linear(bias = encoder_k_proj_3_bias, weight = encoder_k_proj_3_weight, x = x_19)[name = tensor("linear_31")]; - tensor var_811 = const()[name = tensor("op_811"), val = tensor(0x1p-3)]; - tensor var_812 = mul(x = var_810, y = var_811)[name = tensor("op_812")]; - tensor var_813 = const()[name = tensor("op_813"), val = tensor([1, 4, 4, 64])]; - tensor var_814 = reshape(shape = var_813, x = var_812)[name = tensor("op_814")]; + tensor var_866 = linear(bias = inner_encoder_k_proj_3_bias, weight = inner_encoder_k_proj_3_weight, x = x_19)[name = tensor("linear_31")]; + tensor var_867 = const()[name = tensor("op_867"), val = tensor(0x1p-3)]; + tensor var_868 = mul(x = var_866, y = var_867)[name = tensor("op_868")]; + tensor var_869 = const()[name = tensor("op_869"), val = tensor([1, 4, 4, 64])]; + tensor var_870 = reshape(shape = var_869, x = var_868)[name = tensor("op_870")]; tensor k_7_perm_0 = const()[name = tensor("k_7_perm_0"), val = tensor([0, 2, 1, 3])]; - tensor var_818 = linear(bias = encoder_v_proj_3_bias, weight = encoder_v_proj_3_weight, x = x_19)[name = tensor("linear_32")]; - tensor var_819 = const()[name = tensor("op_819"), val = tensor([1, 4, 4, 64])]; - tensor var_820 = reshape(shape = var_819, x = var_818)[name = tensor("op_820")]; + tensor var_874 = linear(bias = inner_encoder_v_proj_3_bias, weight = inner_encoder_v_proj_3_weight, x = x_19)[name = tensor("linear_32")]; + tensor var_875 = const()[name = tensor("op_875"), val = tensor([1, 4, 4, 64])]; + tensor var_876 = reshape(shape = var_875, x = var_874)[name = tensor("op_876")]; tensor v_7_perm_0 = const()[name = tensor("v_7_perm_0"), val = tensor([0, 2, 1, 3])]; - tensor input_137 = linear(bias = encoder_g_proj_3_bias, weight = encoder_g_proj_3_weight, x = x_19)[name = tensor("linear_33")]; + tensor input_139 = linear(bias = inner_encoder_g_proj_3_bias, weight = inner_encoder_g_proj_3_weight, x = x_19)[name = tensor("linear_33")]; tensor sqrt_s0_7 = sqrt(x = prev_scale_7)[name = tensor("sqrt_s0_7")]; - tensor s_t_7 = add(x = prev_scale_7, y = encoder__t_index)[name = tensor("s_t_7")]; + tensor s_t_7 = add(x = prev_scale_7, y = inner_encoder__t_index)[name = tensor("s_t_7")]; tensor sqrt_s_t_7 = sqrt(x = s_t_7)[name = tensor("sqrt_s_t_7")]; tensor qk_7_transpose_x_1 = const()[name = tensor("qk_7_transpose_x_1"), val = tensor(false)]; tensor qk_7_transpose_y_1 = const()[name = tensor("qk_7_transpose_y_1"), val = tensor(true)]; - tensor k_7 = transpose(perm = k_7_perm_0, x = var_814)[name = tensor("transpose_36")]; - tensor q_7 = transpose(perm = q_7_perm_0, x = var_806)[name = tensor("transpose_37")]; + tensor k_7 = transpose(perm = k_7_perm_0, x = var_870)[name = tensor("transpose_36")]; + tensor q_7 = transpose(perm = q_7_perm_0, x = var_862)[name = tensor("transpose_37")]; tensor qk_7 = matmul(transpose_x = qk_7_transpose_x_1, transpose_y = qk_7_transpose_y_1, x = q_7, y = k_7)[name = tensor("qk_7")]; - tensor var_830 = const()[name = tensor("op_830"), val = tensor([4, 1])]; - tensor var_831 = reshape(shape = var_830, x = sqrt_s_t_7)[name = tensor("op_831")]; - tensor M_7 = real_div(x = encoder__causal_mask, y = var_831)[name = tensor("M_7")]; - tensor var_833 = mul(x = qk_7, y = M_7)[name = tensor("op_833")]; + tensor var_886 = const()[name = tensor("op_886"), val = tensor([4, 1])]; + tensor var_887 = reshape(shape = var_886, x = sqrt_s_t_7)[name = tensor("op_887")]; + tensor M_7 = real_div(x = inner_encoder__causal_mask, y = var_887)[name = tensor("M_7")]; + tensor var_889 = mul(x = qk_7, y = M_7)[name = tensor("op_889")]; tensor inner_7_transpose_x_0 = const()[name = tensor("inner_7_transpose_x_0"), val = tensor(false)]; tensor inner_7_transpose_y_0 = const()[name = tensor("inner_7_transpose_y_0"), val = tensor(false)]; - tensor v_7 = transpose(perm = v_7_perm_0, x = var_820)[name = tensor("transpose_35")]; - tensor inner_7 = matmul(transpose_x = inner_7_transpose_x_0, transpose_y = inner_7_transpose_y_0, x = var_833, y = v_7)[name = tensor("inner_7")]; - tensor var_835_transpose_x_0 = const()[name = tensor("op_835_transpose_x_0"), val = tensor(false)]; - tensor var_835_transpose_y_0 = const()[name = tensor("op_835_transpose_y_0"), val = tensor(false)]; - tensor var_835 = matmul(transpose_x = var_835_transpose_x_0, transpose_y = var_835_transpose_y_0, x = q_7, y = prev_kv_7)[name = tensor("op_835")]; - tensor var_836 = real_div(x = sqrt_s0_7, y = sqrt_s_t_7)[name = tensor("op_836")]; - tensor var_837 = const()[name = tensor("op_837"), val = tensor([1, 1, 4, 1])]; - tensor var_838 = reshape(shape = var_837, x = var_836)[name = tensor("op_838")]; - tensor cross_7 = mul(x = var_835, y = var_838)[name = tensor("cross_7")]; + tensor v_7 = transpose(perm = v_7_perm_0, x = var_876)[name = tensor("transpose_35")]; + tensor inner_7 = matmul(transpose_x = inner_7_transpose_x_0, transpose_y = inner_7_transpose_y_0, x = var_889, y = v_7)[name = tensor("inner_7")]; + tensor var_891_transpose_x_0 = const()[name = tensor("op_891_transpose_x_0"), val = tensor(false)]; + tensor var_891_transpose_y_0 = const()[name = tensor("op_891_transpose_y_0"), val = tensor(false)]; + tensor var_891 = matmul(transpose_x = var_891_transpose_x_0, transpose_y = var_891_transpose_y_0, x = q_7, y = prev_kv_7)[name = tensor("op_891")]; + tensor var_892 = real_div(x = sqrt_s0_7, y = sqrt_s_t_7)[name = tensor("op_892")]; + tensor var_893 = const()[name = tensor("op_893"), val = tensor([1, 1, 4, 1])]; + tensor var_894 = reshape(shape = var_893, x = var_892)[name = tensor("op_894")]; + tensor cross_7 = mul(x = var_891, y = var_894)[name = tensor("cross_7")]; tensor out_19 = add(x = inner_7, y = cross_7)[name = tensor("out_19")]; - tensor var_841 = mul(x = prev_kv_7, y = sqrt_s0_7)[name = tensor("op_841")]; - tensor var_843_transpose_x_1 = const()[name = tensor("op_843_transpose_x_1"), val = tensor(true)]; - tensor var_843_transpose_y_1 = const()[name = tensor("op_843_transpose_y_1"), val = tensor(false)]; - tensor var_843 = matmul(transpose_x = var_843_transpose_x_1, transpose_y = var_843_transpose_y_1, x = k_7, y = v_7)[name = tensor("op_843")]; - tensor new_kv_unnorm_7 = add(x = var_841, y = var_843)[name = tensor("new_kv_unnorm_7")]; - tensor var_845 = const()[name = tensor("op_845"), val = tensor(0x1p+2)]; - tensor new_scale_7 = add(x = prev_scale_7, y = var_845)[name = tensor("new_scale_7")]; - tensor var_847 = sqrt(x = new_scale_7)[name = tensor("op_847")]; - tensor nkv_1 = real_div(x = new_kv_unnorm_7, y = var_847)[name = tensor("nkv_1")]; - tensor var_849_perm_0 = const()[name = tensor("op_849_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_897 = mul(x = prev_kv_7, y = sqrt_s0_7)[name = tensor("op_897")]; + tensor var_899_transpose_x_1 = const()[name = tensor("op_899_transpose_x_1"), val = tensor(true)]; + tensor var_899_transpose_y_1 = const()[name = tensor("op_899_transpose_y_1"), val = tensor(false)]; + tensor var_899 = matmul(transpose_x = var_899_transpose_x_1, transpose_y = var_899_transpose_y_1, x = k_7, y = v_7)[name = tensor("op_899")]; + tensor new_kv_unnorm_7 = add(x = var_897, y = var_899)[name = tensor("new_kv_unnorm_7")]; + tensor var_901 = const()[name = tensor("op_901"), val = tensor(0x1p+2)]; + tensor new_scale_7 = add(x = prev_scale_7, y = var_901)[name = tensor("new_scale_7")]; + tensor var_903 = sqrt(x = new_scale_7)[name = tensor("op_903")]; + tensor nkv_1 = real_div(x = new_kv_unnorm_7, y = var_903)[name = tensor("nkv_1")]; + tensor var_905_perm_0 = const()[name = tensor("op_905_perm_0"), val = tensor([0, 2, 1, 3])]; tensor out_21_axes_0 = const()[name = tensor("out_21_axes_0"), val = tensor([-1])]; - tensor var_849 = transpose(perm = var_849_perm_0, x = out_19)[name = tensor("transpose_34")]; - tensor out_21 = layer_norm(axes = out_21_axes_0, epsilon = var_17, x = var_849)[name = tensor("out_21")]; - tensor var_853 = const()[name = tensor("op_853"), val = tensor([1, 4, 256])]; - tensor out_23 = reshape(shape = var_853, x = out_21)[name = tensor("out_23")]; - tensor var_855 = silu(x = input_137)[name = tensor("op_855")]; - tensor input_139 = mul(x = var_855, y = out_23)[name = tensor("input_139")]; - tensor ret_out_7 = linear(bias = encoder_out_proj_3_bias, weight = encoder_out_proj_3_weight, x = input_139)[name = tensor("linear_34")]; - tensor x_21 = add(x = input_133, y = ret_out_7)[name = tensor("x_21")]; + tensor var_905 = transpose(perm = var_905_perm_0, x = out_19)[name = tensor("transpose_34")]; + tensor out_21 = layer_norm(axes = out_21_axes_0, epsilon = var_74, x = var_905)[name = tensor("out_21")]; + tensor var_909 = const()[name = tensor("op_909"), val = tensor([1, 4, 256])]; + tensor out_23 = reshape(shape = var_909, x = out_21)[name = tensor("out_23")]; + tensor var_911 = silu(x = input_139)[name = tensor("op_911")]; + tensor input_141 = mul(x = var_911, y = out_23)[name = tensor("input_141")]; + tensor ret_out_7 = linear(bias = inner_encoder_out_proj_3_bias, weight = inner_encoder_out_proj_3_weight, x = input_141)[name = tensor("linear_34")]; + tensor x_21 = add(x = input_135, y = ret_out_7)[name = tensor("x_21")]; tensor window_31_begin_0 = const()[name = tensor("window_31_begin_0"), val = tensor([3, 0, 0, 0])]; tensor window_31_end_0 = const()[name = tensor("window_31_end_0"), val = tensor([4, 1, 16, 256])]; tensor window_31_end_mask_0 = const()[name = tensor("window_31_end_mask_0"), val = tensor([false, true, true, true])]; tensor window_31_squeeze_mask_0 = const()[name = tensor("window_31_squeeze_mask_0"), val = tensor([true, false, false, false])]; tensor window_31 = slice_by_index(begin = window_31_begin_0, end = window_31_end_0, end_mask = window_31_end_mask_0, squeeze_mask = window_31_squeeze_mask_0, x = enc_conv_cache)[name = tensor("window_31")]; - tensor var_863_begin_0 = const()[name = tensor("op_863_begin_0"), val = tensor([0, 0, 0])]; - tensor var_863_end_0 = const()[name = tensor("op_863_end_0"), val = tensor([1, 1, 256])]; - tensor var_863_end_mask_0 = const()[name = tensor("op_863_end_mask_0"), val = tensor([true, false, true])]; - tensor var_863 = slice_by_index(begin = var_863_begin_0, end = var_863_end_0, end_mask = var_863_end_mask_0, x = x_21)[name = tensor("op_863")]; - tensor var_866_begin_0 = const()[name = tensor("op_866_begin_0"), val = tensor([0, 1, 0])]; - tensor var_866_end_0 = const()[name = tensor("op_866_end_0"), val = tensor([1, 16, 256])]; - tensor var_866_end_mask_0 = const()[name = tensor("op_866_end_mask_0"), val = tensor([true, true, true])]; - tensor var_866 = slice_by_index(begin = var_866_begin_0, end = var_866_end_0, end_mask = var_866_end_mask_0, x = window_31)[name = tensor("op_866")]; + tensor var_919_begin_0 = const()[name = tensor("op_919_begin_0"), val = tensor([0, 0, 0])]; + tensor var_919_end_0 = const()[name = tensor("op_919_end_0"), val = tensor([1, 1, 256])]; + tensor var_919_end_mask_0 = const()[name = tensor("op_919_end_mask_0"), val = tensor([true, false, true])]; + tensor var_919 = slice_by_index(begin = var_919_begin_0, end = var_919_end_0, end_mask = var_919_end_mask_0, x = x_21)[name = tensor("op_919")]; + tensor var_922_begin_0 = const()[name = tensor("op_922_begin_0"), val = tensor([0, 1, 0])]; + tensor var_922_end_0 = const()[name = tensor("op_922_end_0"), val = tensor([1, 16, 256])]; + tensor var_922_end_mask_0 = const()[name = tensor("op_922_end_mask_0"), val = tensor([true, true, true])]; + tensor var_922 = slice_by_index(begin = var_922_begin_0, end = var_922_end_0, end_mask = var_922_end_mask_0, x = window_31)[name = tensor("op_922")]; tensor window_33_interleave_0 = const()[name = tensor("window_33_interleave_0"), val = tensor(false)]; - tensor window_33 = concat(axis = var_26, interleave = window_33_interleave_0, values = (var_866, var_863))[name = tensor("window_33")]; - tensor var_871_begin_0 = const()[name = tensor("op_871_begin_0"), val = tensor([0, 1, 0])]; - tensor var_871_end_0 = const()[name = tensor("op_871_end_0"), val = tensor([1, 2, 256])]; - tensor var_871_end_mask_0 = const()[name = tensor("op_871_end_mask_0"), val = tensor([true, false, true])]; - tensor var_871 = slice_by_index(begin = var_871_begin_0, end = var_871_end_0, end_mask = var_871_end_mask_0, x = x_21)[name = tensor("op_871")]; - tensor var_874_begin_0 = const()[name = tensor("op_874_begin_0"), val = tensor([0, 1, 0])]; - tensor var_874_end_0 = const()[name = tensor("op_874_end_0"), val = tensor([1, 16, 256])]; - tensor var_874_end_mask_0 = const()[name = tensor("op_874_end_mask_0"), val = tensor([true, true, true])]; - tensor var_874 = slice_by_index(begin = var_874_begin_0, end = var_874_end_0, end_mask = var_874_end_mask_0, x = window_33)[name = tensor("op_874")]; + tensor window_33 = concat(axis = var_82, interleave = window_33_interleave_0, values = (var_922, var_919))[name = tensor("window_33")]; + tensor var_927_begin_0 = const()[name = tensor("op_927_begin_0"), val = tensor([0, 1, 0])]; + tensor var_927_end_0 = const()[name = tensor("op_927_end_0"), val = tensor([1, 2, 256])]; + tensor var_927_end_mask_0 = const()[name = tensor("op_927_end_mask_0"), val = tensor([true, false, true])]; + tensor var_927 = slice_by_index(begin = var_927_begin_0, end = var_927_end_0, end_mask = var_927_end_mask_0, x = x_21)[name = tensor("op_927")]; + tensor var_930_begin_0 = const()[name = tensor("op_930_begin_0"), val = tensor([0, 1, 0])]; + tensor var_930_end_0 = const()[name = tensor("op_930_end_0"), val = tensor([1, 16, 256])]; + tensor var_930_end_mask_0 = const()[name = tensor("op_930_end_mask_0"), val = tensor([true, true, true])]; + tensor var_930 = slice_by_index(begin = var_930_begin_0, end = var_930_end_0, end_mask = var_930_end_mask_0, x = window_33)[name = tensor("op_930")]; tensor window_35_interleave_0 = const()[name = tensor("window_35_interleave_0"), val = tensor(false)]; - tensor window_35 = concat(axis = var_26, interleave = window_35_interleave_0, values = (var_874, var_871))[name = tensor("window_35")]; - tensor var_879_begin_0 = const()[name = tensor("op_879_begin_0"), val = tensor([0, 2, 0])]; - tensor var_879_end_0 = const()[name = tensor("op_879_end_0"), val = tensor([1, 3, 256])]; - tensor var_879_end_mask_0 = const()[name = tensor("op_879_end_mask_0"), val = tensor([true, false, true])]; - tensor var_879 = slice_by_index(begin = var_879_begin_0, end = var_879_end_0, end_mask = var_879_end_mask_0, x = x_21)[name = tensor("op_879")]; - tensor var_882_begin_0 = const()[name = tensor("op_882_begin_0"), val = tensor([0, 1, 0])]; - tensor var_882_end_0 = const()[name = tensor("op_882_end_0"), val = tensor([1, 16, 256])]; - tensor var_882_end_mask_0 = const()[name = tensor("op_882_end_mask_0"), val = tensor([true, true, true])]; - tensor var_882 = slice_by_index(begin = var_882_begin_0, end = var_882_end_0, end_mask = var_882_end_mask_0, x = window_35)[name = tensor("op_882")]; + tensor window_35 = concat(axis = var_82, interleave = window_35_interleave_0, values = (var_930, var_927))[name = tensor("window_35")]; + tensor var_935_begin_0 = const()[name = tensor("op_935_begin_0"), val = tensor([0, 2, 0])]; + tensor var_935_end_0 = const()[name = tensor("op_935_end_0"), val = tensor([1, 3, 256])]; + tensor var_935_end_mask_0 = const()[name = tensor("op_935_end_mask_0"), val = tensor([true, false, true])]; + tensor var_935 = slice_by_index(begin = var_935_begin_0, end = var_935_end_0, end_mask = var_935_end_mask_0, x = x_21)[name = tensor("op_935")]; + tensor var_938_begin_0 = const()[name = tensor("op_938_begin_0"), val = tensor([0, 1, 0])]; + tensor var_938_end_0 = const()[name = tensor("op_938_end_0"), val = tensor([1, 16, 256])]; + tensor var_938_end_mask_0 = const()[name = tensor("op_938_end_mask_0"), val = tensor([true, true, true])]; + tensor var_938 = slice_by_index(begin = var_938_begin_0, end = var_938_end_0, end_mask = var_938_end_mask_0, x = window_35)[name = tensor("op_938")]; tensor window_37_interleave_0 = const()[name = tensor("window_37_interleave_0"), val = tensor(false)]; - tensor window_37 = concat(axis = var_26, interleave = window_37_interleave_0, values = (var_882, var_879))[name = tensor("window_37")]; - tensor var_887_begin_0 = const()[name = tensor("op_887_begin_0"), val = tensor([0, 3, 0])]; - tensor var_887_end_0 = const()[name = tensor("op_887_end_0"), val = tensor([1, 1, 256])]; - tensor var_887_end_mask_0 = const()[name = tensor("op_887_end_mask_0"), val = tensor([true, true, true])]; - tensor var_887 = slice_by_index(begin = var_887_begin_0, end = var_887_end_0, end_mask = var_887_end_mask_0, x = x_21)[name = tensor("op_887")]; - tensor var_890_begin_0 = const()[name = tensor("op_890_begin_0"), val = tensor([0, 1, 0])]; - tensor var_890_end_0 = const()[name = tensor("op_890_end_0"), val = tensor([1, 16, 256])]; - tensor var_890_end_mask_0 = const()[name = tensor("op_890_end_mask_0"), val = tensor([true, true, true])]; - tensor var_890 = slice_by_index(begin = var_890_begin_0, end = var_890_end_0, end_mask = var_890_end_mask_0, x = window_37)[name = tensor("op_890")]; + tensor window_37 = concat(axis = var_82, interleave = window_37_interleave_0, values = (var_938, var_935))[name = tensor("window_37")]; + tensor var_943_begin_0 = const()[name = tensor("op_943_begin_0"), val = tensor([0, 3, 0])]; + tensor var_943_end_0 = const()[name = tensor("op_943_end_0"), val = tensor([1, 1, 256])]; + tensor var_943_end_mask_0 = const()[name = tensor("op_943_end_mask_0"), val = tensor([true, true, true])]; + tensor var_943 = slice_by_index(begin = var_943_begin_0, end = var_943_end_0, end_mask = var_943_end_mask_0, x = x_21)[name = tensor("op_943")]; + tensor var_946_begin_0 = const()[name = tensor("op_946_begin_0"), val = tensor([0, 1, 0])]; + tensor var_946_end_0 = const()[name = tensor("op_946_end_0"), val = tensor([1, 16, 256])]; + tensor var_946_end_mask_0 = const()[name = tensor("op_946_end_mask_0"), val = tensor([true, true, true])]; + tensor var_946 = slice_by_index(begin = var_946_begin_0, end = var_946_end_0, end_mask = var_946_end_mask_0, x = window_37)[name = tensor("op_946")]; tensor window_interleave_0 = const()[name = tensor("window_interleave_0"), val = tensor(false)]; - tensor window = concat(axis = var_26, interleave = window_interleave_0, values = (var_890, var_887))[name = tensor("window")]; - tensor input_141_interleave_0 = const()[name = tensor("input_141_interleave_0"), val = tensor(false)]; - tensor input_141 = concat(axis = var_23, interleave = input_141_interleave_0, values = (window_33, window_35, window_37, window))[name = tensor("input_141")]; + tensor window = concat(axis = var_82, interleave = window_interleave_0, values = (var_946, var_943))[name = tensor("window")]; + tensor input_143_interleave_0 = const()[name = tensor("input_143_interleave_0"), val = tensor(false)]; + tensor input_143 = concat(axis = var_69, interleave = input_143_interleave_0, values = (window_33, window_35, window_37, window))[name = tensor("input_143")]; tensor x_23_axes_0 = const()[name = tensor("x_23_axes_0"), val = tensor([-1])]; - tensor x_23 = layer_norm(axes = x_23_axes_0, beta = encoder_conv_module_3_sequential_0_bias, epsilon = var_28, gamma = encoder_conv_module_3_sequential_0_weight, x = input_141)[name = tensor("x_23")]; - tensor input_143_perm_0 = const()[name = tensor("input_143_perm_0"), val = tensor([0, 2, 1])]; + tensor x_23 = layer_norm(axes = x_23_axes_0, beta = inner_encoder_conv_module_3_sequential_0_bias, epsilon = var_66, gamma = inner_encoder_conv_module_3_sequential_0_weight, x = input_143)[name = tensor("x_23")]; + tensor input_145_perm_0 = const()[name = tensor("input_145_perm_0"), val = tensor([0, 2, 1])]; tensor inputs_33_pad_type_0 = const()[name = tensor("inputs_33_pad_type_0"), val = tensor("valid")]; tensor inputs_33_strides_0 = const()[name = tensor("inputs_33_strides_0"), val = tensor([1])]; tensor inputs_33_pad_0 = const()[name = tensor("inputs_33_pad_0"), val = tensor([0, 0])]; tensor inputs_33_dilations_0 = const()[name = tensor("inputs_33_dilations_0"), val = tensor([1])]; tensor inputs_33_groups_0 = const()[name = tensor("inputs_33_groups_0"), val = tensor(1)]; - tensor input_143 = transpose(perm = input_143_perm_0, x = x_23)[name = tensor("transpose_33")]; - tensor inputs_33 = conv(bias = encoder_conv_module_3_sequential_2_conv_bias, dilations = inputs_33_dilations_0, groups = inputs_33_groups_0, pad = inputs_33_pad_0, pad_type = inputs_33_pad_type_0, strides = inputs_33_strides_0, weight = encoder_conv_module_3_sequential_2_conv_weight, x = input_143)[name = tensor("inputs_33")]; - tensor var_915_split_sizes_0 = const()[name = tensor("op_915_split_sizes_0"), val = tensor([256, 256])]; - tensor var_915_axis_0 = const()[name = tensor("op_915_axis_0"), val = tensor(1)]; - tensor var_915_0, tensor var_915_1 = split(axis = var_915_axis_0, split_sizes = var_915_split_sizes_0, x = inputs_33)[name = tensor("op_915")]; - tensor var_917 = sigmoid(x = var_915_1)[name = tensor("op_917")]; - tensor inputs_35 = mul(x = var_915_0, y = var_917)[name = tensor("inputs_35")]; + tensor input_145 = transpose(perm = input_145_perm_0, x = x_23)[name = tensor("transpose_33")]; + tensor inputs_33 = conv(bias = inner_encoder_conv_module_3_sequential_2_conv_bias, dilations = inputs_33_dilations_0, groups = inputs_33_groups_0, pad = inputs_33_pad_0, pad_type = inputs_33_pad_type_0, strides = inputs_33_strides_0, weight = inner_encoder_conv_module_3_sequential_2_conv_weight, x = input_145)[name = tensor("inputs_33")]; + tensor var_971_split_sizes_0 = const()[name = tensor("op_971_split_sizes_0"), val = tensor([256, 256])]; + tensor var_971_axis_0 = const()[name = tensor("op_971_axis_0"), val = tensor(1)]; + tensor var_971_0, tensor var_971_1 = split(axis = var_971_axis_0, split_sizes = var_971_split_sizes_0, x = inputs_33)[name = tensor("op_971")]; + tensor var_973 = sigmoid(x = var_971_1)[name = tensor("op_973")]; + tensor inputs_35 = mul(x = var_971_0, y = var_973)[name = tensor("inputs_35")]; tensor outputs_aug_pad_type_0 = const()[name = tensor("outputs_aug_pad_type_0"), val = tensor("custom")]; tensor outputs_aug_pad_0 = const()[name = tensor("outputs_aug_pad_0"), val = tensor([15, 15])]; tensor outputs_aug_groups_0 = const()[name = tensor("outputs_aug_groups_0"), val = tensor(256)]; tensor outputs_aug_strides_0 = const()[name = tensor("outputs_aug_strides_0"), val = tensor([1])]; tensor outputs_aug_dilations_0 = const()[name = tensor("outputs_aug_dilations_0"), val = tensor([1])]; - tensor outputs_aug = conv(dilations = outputs_aug_dilations_0, groups = outputs_aug_groups_0, pad = outputs_aug_pad_0, pad_type = outputs_aug_pad_type_0, strides = outputs_aug_strides_0, weight = encoder_conv_module_3_sequential_4_conv_weight, x = inputs_35)[name = tensor("outputs_aug")]; - tensor input_145_begin_0 = const()[name = tensor("input_145_begin_0"), val = tensor([0, 0, 0])]; - tensor input_145_end_0 = const()[name = tensor("input_145_end_0"), val = tensor([4, 256, 16])]; - tensor input_145_end_mask_0 = const()[name = tensor("input_145_end_mask_0"), val = tensor([true, true, false])]; - tensor input_145 = slice_by_index(begin = input_145_begin_0, end = input_145_end_0, end_mask = input_145_end_mask_0, x = outputs_aug)[name = tensor("input_145")]; - tensor inputs_37 = batch_norm(beta = encoder_conv_module_3_sequential_5_bias, epsilon = var_28, gamma = encoder_conv_module_3_sequential_5_weight, mean = encoder_conv_module_3_sequential_5_running_mean, variance = encoder_conv_module_3_sequential_5_running_var, x = input_145)[name = tensor("inputs_37")]; - tensor input_147 = silu(x = inputs_37)[name = tensor("input_147")]; - tensor input_149_pad_type_0 = const()[name = tensor("input_149_pad_type_0"), val = tensor("valid")]; - tensor input_149_strides_0 = const()[name = tensor("input_149_strides_0"), val = tensor([1])]; - tensor input_149_pad_0 = const()[name = tensor("input_149_pad_0"), val = tensor([0, 0])]; - tensor input_149_dilations_0 = const()[name = tensor("input_149_dilations_0"), val = tensor([1])]; - tensor input_149_groups_0 = const()[name = tensor("input_149_groups_0"), val = tensor(1)]; - tensor input_149 = conv(bias = encoder_conv_module_3_sequential_7_conv_bias, dilations = input_149_dilations_0, groups = input_149_groups_0, pad = input_149_pad_0, pad_type = input_149_pad_type_0, strides = input_149_strides_0, weight = encoder_conv_module_3_sequential_7_conv_weight, x = input_147)[name = tensor("input_149")]; + tensor outputs_aug = conv(dilations = outputs_aug_dilations_0, groups = outputs_aug_groups_0, pad = outputs_aug_pad_0, pad_type = outputs_aug_pad_type_0, strides = outputs_aug_strides_0, weight = inner_encoder_conv_module_3_sequential_4_conv_weight, x = inputs_35)[name = tensor("outputs_aug")]; + tensor input_147_begin_0 = const()[name = tensor("input_147_begin_0"), val = tensor([0, 0, 0])]; + tensor input_147_end_0 = const()[name = tensor("input_147_end_0"), val = tensor([4, 256, 16])]; + tensor input_147_end_mask_0 = const()[name = tensor("input_147_end_mask_0"), val = tensor([true, true, false])]; + tensor input_147 = slice_by_index(begin = input_147_begin_0, end = input_147_end_0, end_mask = input_147_end_mask_0, x = outputs_aug)[name = tensor("input_147")]; + tensor inputs_37 = batch_norm(beta = inner_encoder_conv_module_3_sequential_5_bias, epsilon = var_66, gamma = inner_encoder_conv_module_3_sequential_5_weight, mean = inner_encoder_conv_module_3_sequential_5_running_mean, variance = inner_encoder_conv_module_3_sequential_5_running_var, x = input_147)[name = tensor("inputs_37")]; + tensor input_149 = silu(x = inputs_37)[name = tensor("input_149")]; + tensor input_151_pad_type_0 = const()[name = tensor("input_151_pad_type_0"), val = tensor("valid")]; + tensor input_151_strides_0 = const()[name = tensor("input_151_strides_0"), val = tensor([1])]; + tensor input_151_pad_0 = const()[name = tensor("input_151_pad_0"), val = tensor([0, 0])]; + tensor input_151_dilations_0 = const()[name = tensor("input_151_dilations_0"), val = tensor([1])]; + tensor input_151_groups_0 = const()[name = tensor("input_151_groups_0"), val = tensor(1)]; + tensor input_151 = conv(bias = inner_encoder_conv_module_3_sequential_7_conv_bias, dilations = input_151_dilations_0, groups = input_151_groups_0, pad = input_151_pad_0, pad_type = input_151_pad_type_0, strides = input_151_strides_0, weight = inner_encoder_conv_module_3_sequential_7_conv_weight, x = input_149)[name = tensor("input_151")]; tensor conv_out_7_perm_0 = const()[name = tensor("conv_out_7_perm_0"), val = tensor([0, 2, 1])]; - tensor var_948_begin_0 = const()[name = tensor("op_948_begin_0"), val = tensor([0, -1, 0])]; - tensor var_948_end_0 = const()[name = tensor("op_948_end_0"), val = tensor([4, 16, 256])]; - tensor var_948_end_mask_0 = const()[name = tensor("op_948_end_mask_0"), val = tensor([true, true, true])]; - tensor conv_out_7 = transpose(perm = conv_out_7_perm_0, x = input_149)[name = tensor("transpose_32")]; - tensor var_948 = slice_by_index(begin = var_948_begin_0, end = var_948_end_0, end_mask = var_948_end_mask_0, x = conv_out_7)[name = tensor("op_948")]; - tensor var_950_perm_0 = const()[name = tensor("op_950_perm_0"), val = tensor([1, 0, 2])]; - tensor var_950 = transpose(perm = var_950_perm_0, x = var_948)[name = tensor("transpose_31")]; - tensor input_151 = add(x = x_21, y = var_950)[name = tensor("input_151")]; - tensor input_153_axes_0 = const()[name = tensor("input_153_axes_0"), val = tensor([-1])]; - tensor input_153 = layer_norm(axes = input_153_axes_0, beta = encoder_ffn2_3_module_sequential_0_bias, epsilon = var_28, gamma = encoder_ffn2_3_module_sequential_0_weight, x = input_151)[name = tensor("input_153")]; - tensor inputs = linear(bias = encoder_ffn2_3_module_sequential_1_linear_bias, weight = encoder_ffn2_3_module_sequential_1_linear_weight, x = input_153)[name = tensor("linear_35")]; - tensor input_155 = silu(x = inputs)[name = tensor("input_155")]; - tensor input_159 = linear(bias = encoder_ffn2_3_module_sequential_4_linear_bias, weight = encoder_ffn2_3_module_sequential_4_linear_weight, x = input_155)[name = tensor("linear_36")]; - tensor var_973 = const()[name = tensor("op_973"), val = tensor(0x1p-1)]; - tensor var_974 = mul(x = input_159, y = var_973)[name = tensor("op_974")]; - tensor input_161 = add(x = var_974, y = input_151)[name = tensor("input_161")]; + tensor var_1004_begin_0 = const()[name = tensor("op_1004_begin_0"), val = tensor([0, -1, 0])]; + tensor var_1004_end_0 = const()[name = tensor("op_1004_end_0"), val = tensor([4, 16, 256])]; + tensor var_1004_end_mask_0 = const()[name = tensor("op_1004_end_mask_0"), val = tensor([true, true, true])]; + tensor conv_out_7 = transpose(perm = conv_out_7_perm_0, x = input_151)[name = tensor("transpose_32")]; + tensor var_1004 = slice_by_index(begin = var_1004_begin_0, end = var_1004_end_0, end_mask = var_1004_end_mask_0, x = conv_out_7)[name = tensor("op_1004")]; + tensor var_1006_perm_0 = const()[name = tensor("op_1006_perm_0"), val = tensor([1, 0, 2])]; + tensor var_1006 = transpose(perm = var_1006_perm_0, x = var_1004)[name = tensor("transpose_31")]; + tensor input_153 = add(x = x_21, y = var_1006)[name = tensor("input_153")]; + tensor input_155_axes_0 = const()[name = tensor("input_155_axes_0"), val = tensor([-1])]; + tensor input_155 = layer_norm(axes = input_155_axes_0, beta = inner_encoder_ffn2_3_module_sequential_0_bias, epsilon = var_66, gamma = inner_encoder_ffn2_3_module_sequential_0_weight, x = input_153)[name = tensor("input_155")]; + tensor inputs = linear(bias = inner_encoder_ffn2_3_module_sequential_1_linear_bias, weight = inner_encoder_ffn2_3_module_sequential_1_linear_weight, x = input_155)[name = tensor("linear_35")]; + tensor input_157 = silu(x = inputs)[name = tensor("input_157")]; + tensor input_161 = linear(bias = inner_encoder_ffn2_3_module_sequential_4_linear_bias, weight = inner_encoder_ffn2_3_module_sequential_4_linear_weight, x = input_157)[name = tensor("linear_36")]; + tensor var_1029 = const()[name = tensor("op_1029"), val = tensor(0x1p-1)]; + tensor var_1030 = mul(x = input_161, y = var_1029)[name = tensor("op_1030")]; + tensor input_163 = add(x = var_1030, y = input_153)[name = tensor("input_163")]; tensor x_25_axes_0 = const()[name = tensor("x_25_axes_0"), val = tensor([-1])]; - tensor x_25 = layer_norm(axes = x_25_axes_0, beta = encoder_layer_norm_3_bias, epsilon = var_28, gamma = encoder_layer_norm_3_weight, x = input_161)[name = tensor("x_25")]; + tensor x_25 = layer_norm(axes = x_25_axes_0, beta = inner_encoder_layer_norm_3_bias, epsilon = var_66, gamma = inner_encoder_layer_norm_3_weight, x = input_163)[name = tensor("x_25")]; tensor x_ct_perm_0 = const()[name = tensor("x_ct_perm_0"), val = tensor([0, 2, 1])]; tensor cat_interleave_0 = const()[name = tensor("cat_interleave_0"), val = tensor(false)]; tensor x_ct = transpose(perm = x_ct_perm_0, x = x_25)[name = tensor("transpose_30")]; - tensor cat = concat(axis = var_20, interleave = cat_interleave_0, values = (cnn_window, x_ct))[name = tensor("cat")]; + tensor cat = concat(axis = var_71, interleave = cat_interleave_0, values = (cnn_window, x_ct))[name = tensor("cat")]; tensor conv_out_pad_type_0 = const()[name = tensor("conv_out_pad_type_0"), val = tensor("valid")]; tensor conv_out_strides_0 = const()[name = tensor("conv_out_strides_0"), val = tensor([1])]; tensor conv_out_pad_0 = const()[name = tensor("conv_out_pad_0"), val = tensor([0, 0])]; tensor conv_out_dilations_0 = const()[name = tensor("conv_out_dilations_0"), val = tensor([1])]; tensor conv_out_groups_0 = const()[name = tensor("conv_out_groups_0"), val = tensor(1)]; - tensor conv_out = conv(bias = encoder_cnn_bias, dilations = conv_out_dilations_0, groups = conv_out_groups_0, pad = conv_out_pad_0, pad_type = conv_out_pad_type_0, strides = conv_out_strides_0, weight = encoder_cnn_weight, x = cat)[name = tensor("conv_out")]; - tensor var_992_begin_0 = const()[name = tensor("op_992_begin_0"), val = tensor([0, 0, 4])]; - tensor var_992_end_0 = const()[name = tensor("op_992_end_0"), val = tensor([1, 256, 22])]; - tensor var_992_end_mask_0 = const()[name = tensor("op_992_end_mask_0"), val = tensor([true, true, true])]; - tensor cnn_window_new = slice_by_index(begin = var_992_begin_0, end = var_992_end_0, end_mask = var_992_end_mask_0, x = cat)[name = tensor("op_992")]; - tensor input_163_perm_0 = const()[name = tensor("input_163_perm_0"), val = tensor([0, 2, 1])]; - tensor var_994 = const()[name = tensor("op_994"), val = tensor([-1])]; - tensor input_163 = transpose(perm = input_163_perm_0, x = conv_out)[name = tensor("transpose_29")]; - tensor var_995 = reduce_l2_norm(axes = var_994, keep_dims = var_29, x = input_163)[name = tensor("op_995")]; + tensor conv_out = conv(bias = inner_encoder_cnn_bias, dilations = conv_out_dilations_0, groups = conv_out_groups_0, pad = conv_out_pad_0, pad_type = conv_out_pad_type_0, strides = conv_out_strides_0, weight = inner_encoder_cnn_weight, x = cat)[name = tensor("conv_out")]; + tensor var_1048_begin_0 = const()[name = tensor("op_1048_begin_0"), val = tensor([0, 0, 4])]; + tensor var_1048_end_0 = const()[name = tensor("op_1048_end_0"), val = tensor([1, 256, 22])]; + tensor var_1048_end_mask_0 = const()[name = tensor("op_1048_end_mask_0"), val = tensor([true, true, true])]; + tensor cnn_window_new = slice_by_index(begin = var_1048_begin_0, end = var_1048_end_0, end_mask = var_1048_end_mask_0, x = cat)[name = tensor("op_1048")]; + tensor input_165_perm_0 = const()[name = tensor("input_165_perm_0"), val = tensor([0, 2, 1])]; + tensor var_1050 = const()[name = tensor("op_1050"), val = tensor([-1])]; + tensor input_165 = transpose(perm = input_165_perm_0, x = conv_out)[name = tensor("transpose_29")]; + tensor var_1051 = reduce_l2_norm(axes = var_1050, keep_dims = var_65, x = input_165)[name = tensor("op_1051")]; tensor const_12 = const()[name = tensor("const_12"), val = tensor(0x1.fffffep+127)]; - tensor clip_0 = clip(alpha = var_11, beta = const_12, x = var_995)[name = tensor("clip_0")]; - tensor emb = real_div(x = input_163, y = clip_0)[name = tensor("emb")]; - tensor var_999_axis_0 = const()[name = tensor("op_999_axis_0"), val = tensor(0)]; - tensor enc_kv_new = stack(axis = var_999_axis_0, values = (var_206, var_420, var_634, nkv_1))[name = tensor("op_999")]; - tensor var_1001_axis_0 = const()[name = tensor("op_1001_axis_0"), val = tensor(0)]; - tensor enc_scale_new = stack(axis = var_1001_axis_0, values = (new_scale_1, new_scale_3, new_scale_5, new_scale_7))[name = tensor("op_1001")]; - tensor var_1003_axis_0 = const()[name = tensor("op_1003_axis_0"), val = tensor(0)]; - tensor enc_conv_cache_new = stack(axis = var_1003_axis_0, values = (window_9, window_19, window_29, window))[name = tensor("op_1003")]; - tensor var_1012 = const()[name = tensor("op_1012"), val = tensor(0x1.5798eep-27)]; - tensor var_1017 = const()[name = tensor("op_1017"), val = tensor(0x1.4f8b58p-17)]; - tensor var_1019 = const()[name = tensor("op_1019"), val = tensor(0x1.0c6f7ap-20)]; - tensor var_1020 = const()[name = tensor("op_1020"), val = tensor(true)]; - tensor var_1022 = const()[name = tensor("op_1022"), val = tensor(0x1p+0)]; - tensor var_1026 = const()[name = tensor("op_1026"), val = tensor(-1)]; - tensor var_1032 = const()[name = tensor("op_1032"), val = tensor(0)]; + tensor clip_0 = clip(alpha = var_79, beta = const_12, x = var_1051)[name = tensor("clip_0")]; + tensor emb = real_div(x = input_165, y = clip_0)[name = tensor("emb")]; + tensor var_1055_axis_0 = const()[name = tensor("op_1055_axis_0"), val = tensor(0)]; + tensor enc_kv_new = stack(axis = var_1055_axis_0, values = (var_262, var_476, var_690, nkv_1))[name = tensor("op_1055")]; + tensor var_1057_axis_0 = const()[name = tensor("op_1057_axis_0"), val = tensor(0)]; + tensor enc_scale_new = stack(axis = var_1057_axis_0, values = (new_scale_1, new_scale_3, new_scale_5, new_scale_7))[name = tensor("op_1057")]; + tensor var_1059_axis_0 = const()[name = tensor("op_1059_axis_0"), val = tensor(0)]; + tensor enc_conv_cache_new = stack(axis = var_1059_axis_0, values = (window_9, window_19, window_29, window))[name = tensor("op_1059")]; tensor pos = const()[name = tensor("pos"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(44395584)))]; - tensor var_1094_axes_0 = const()[name = tensor("op_1094_axes_0"), val = tensor([2])]; - tensor var_1094 = expand_dims(axes = var_1094_axes_0, x = emb)[name = tensor("op_1094")]; + tensor var_1127_axes_0 = const()[name = tensor("op_1127_axes_0"), val = tensor([2])]; + tensor var_1127 = expand_dims(axes = var_1127_axes_0, x = emb)[name = tensor("op_1127")]; tensor emb_exp_reps_0 = const()[name = tensor("emb_exp_reps_0"), val = tensor([1, 1, 12, 1])]; - tensor emb_exp = tile(reps = emb_exp_reps_0, x = var_1094)[name = tensor("emb_exp")]; - tensor input_165_interleave_0 = const()[name = tensor("input_165_interleave_0"), val = tensor(false)]; - tensor input_165 = concat(axis = var_1026, interleave = input_165_interleave_0, values = (emb_exp, pos))[name = tensor("input_165")]; - tensor x_27 = linear(bias = decoder_convert_bias, weight = decoder_convert_weight, x = input_165)[name = tensor("linear_37")]; - tensor var_1102_perm_0 = const()[name = tensor("op_1102_perm_0"), val = tensor([0, 2, 1, 3])]; - tensor var_1106 = const()[name = tensor("op_1106"), val = tensor([12, 4, 256])]; - tensor var_1102 = transpose(perm = var_1102_perm_0, x = x_27)[name = tensor("transpose_28")]; - tensor x_29 = reshape(shape = var_1106, x = var_1102)[name = tensor("x_29")]; + tensor emb_exp = tile(reps = emb_exp_reps_0, x = var_1127)[name = tensor("emb_exp")]; + tensor input_167_interleave_0 = const()[name = tensor("input_167_interleave_0"), val = tensor(false)]; + tensor input_167 = concat(axis = var_72, interleave = input_167_interleave_0, values = (emb_exp, pos))[name = tensor("input_167")]; + tensor x_27 = linear(bias = inner_decoder_convert_bias, weight = inner_decoder_convert_weight, x = input_167)[name = tensor("linear_37")]; + tensor var_1135_perm_0 = const()[name = tensor("op_1135_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_1139 = const()[name = tensor("op_1139"), val = tensor([12, 4, 256])]; + tensor var_1135 = transpose(perm = var_1135_perm_0, x = x_27)[name = tensor("transpose_28")]; + tensor x_29 = reshape(shape = var_1139, x = var_1135)[name = tensor("x_29")]; tensor prev_kv_9_begin_0 = const()[name = tensor("prev_kv_9_begin_0"), val = tensor([0, 0, 0, 0, 0])]; tensor prev_kv_9_end_0 = const()[name = tensor("prev_kv_9_end_0"), val = tensor([1, 12, 4, 64, 64])]; tensor prev_kv_9_end_mask_0 = const()[name = tensor("prev_kv_9_end_mask_0"), val = tensor([false, true, true, true, true])]; @@ -983,132 +998,132 @@ program(1.0) tensor prev_scale_9_end_mask_0 = const()[name = tensor("prev_scale_9_end_mask_0"), val = tensor([false, true])]; tensor prev_scale_9_squeeze_mask_0 = const()[name = tensor("prev_scale_9_squeeze_mask_0"), val = tensor([true, false])]; tensor prev_scale_9 = slice_by_index(begin = prev_scale_9_begin_0, end = prev_scale_9_end_0, end_mask = prev_scale_9_end_mask_0, squeeze_mask = prev_scale_9_squeeze_mask_0, x = dec_scale)[name = tensor("prev_scale_9")]; - tensor var_1114 = linear(bias = decoder_q_proj_0_bias, weight = decoder_q_proj_0_weight, x = x_29)[name = tensor("linear_38")]; - tensor var_1115 = const()[name = tensor("op_1115"), val = tensor([12, 4, 4, 64])]; - tensor var_1116 = reshape(shape = var_1115, x = var_1114)[name = tensor("op_1116")]; + tensor var_1147 = linear(bias = inner_decoder_q_proj_0_bias, weight = inner_decoder_q_proj_0_weight, x = x_29)[name = tensor("linear_38")]; + tensor var_1148 = const()[name = tensor("op_1148"), val = tensor([12, 4, 4, 64])]; + tensor var_1149 = reshape(shape = var_1148, x = var_1147)[name = tensor("op_1149")]; tensor q_9_perm_0 = const()[name = tensor("q_9_perm_0"), val = tensor([0, 2, 1, 3])]; - tensor var_1120 = linear(bias = decoder_k_proj_0_bias, weight = decoder_k_proj_0_weight, x = x_29)[name = tensor("linear_39")]; - tensor var_1121 = const()[name = tensor("op_1121"), val = tensor(0x1p-3)]; - tensor var_1122 = mul(x = var_1120, y = var_1121)[name = tensor("op_1122")]; - tensor var_1123 = const()[name = tensor("op_1123"), val = tensor([12, 4, 4, 64])]; - tensor var_1124 = reshape(shape = var_1123, x = var_1122)[name = tensor("op_1124")]; + tensor var_1153 = linear(bias = inner_decoder_k_proj_0_bias, weight = inner_decoder_k_proj_0_weight, x = x_29)[name = tensor("linear_39")]; + tensor var_1154 = const()[name = tensor("op_1154"), val = tensor(0x1p-3)]; + tensor var_1155 = mul(x = var_1153, y = var_1154)[name = tensor("op_1155")]; + tensor var_1156 = const()[name = tensor("op_1156"), val = tensor([12, 4, 4, 64])]; + tensor var_1157 = reshape(shape = var_1156, x = var_1155)[name = tensor("op_1157")]; tensor k_9_perm_0 = const()[name = tensor("k_9_perm_0"), val = tensor([0, 2, 1, 3])]; - tensor var_1128 = linear(bias = decoder_v_proj_0_bias, weight = decoder_v_proj_0_weight, x = x_29)[name = tensor("linear_40")]; - tensor var_1129 = const()[name = tensor("op_1129"), val = tensor([12, 4, 4, 64])]; - tensor var_1130 = reshape(shape = var_1129, x = var_1128)[name = tensor("op_1130")]; + tensor var_1161 = linear(bias = inner_decoder_v_proj_0_bias, weight = inner_decoder_v_proj_0_weight, x = x_29)[name = tensor("linear_40")]; + tensor var_1162 = const()[name = tensor("op_1162"), val = tensor([12, 4, 4, 64])]; + tensor var_1163 = reshape(shape = var_1162, x = var_1161)[name = tensor("op_1163")]; tensor v_9_perm_0 = const()[name = tensor("v_9_perm_0"), val = tensor([0, 2, 1, 3])]; - tensor input_169 = linear(bias = decoder_g_proj_0_bias, weight = decoder_g_proj_0_weight, x = x_29)[name = tensor("linear_41")]; + tensor input_171 = linear(bias = inner_decoder_g_proj_0_bias, weight = inner_decoder_g_proj_0_weight, x = x_29)[name = tensor("linear_41")]; tensor cumsum_mask_1_exclusive_0 = const()[name = tensor("cumsum_mask_1_exclusive_0"), val = tensor(false)]; tensor cumsum_mask_1_reverse_0 = const()[name = tensor("cumsum_mask_1_reverse_0"), val = tensor(false)]; - tensor cumsum_mask_1 = cumsum(axis = var_1032, exclusive = cumsum_mask_1_exclusive_0, reverse = cumsum_mask_1_reverse_0, x = valid_mask)[name = tensor("cumsum_mask_1")]; + tensor cumsum_mask_1 = cumsum(axis = var_69, exclusive = cumsum_mask_1_exclusive_0, reverse = cumsum_mask_1_reverse_0, x = valid_mask)[name = tensor("cumsum_mask_1")]; tensor sqrt_s0_9 = sqrt(x = prev_scale_9)[name = tensor("sqrt_s0_9")]; tensor s_t_9 = add(x = prev_scale_9, y = cumsum_mask_1)[name = tensor("s_t_9")]; tensor const_20 = const()[name = tensor("const_20"), val = tensor(0x1.fffffep+127)]; - tensor clip_1 = clip(alpha = var_1022, beta = const_20, x = s_t_9)[name = tensor("clip_1")]; + tensor clip_1 = clip(alpha = var_59, beta = const_20, x = s_t_9)[name = tensor("clip_1")]; tensor sqrt_s_t_9 = sqrt(x = clip_1)[name = tensor("sqrt_s_t_9")]; tensor qk_9_transpose_x_1 = const()[name = tensor("qk_9_transpose_x_1"), val = tensor(false)]; tensor qk_9_transpose_y_1 = const()[name = tensor("qk_9_transpose_y_1"), val = tensor(true)]; - tensor k_9 = transpose(perm = k_9_perm_0, x = var_1124)[name = tensor("transpose_26")]; - tensor q_9 = transpose(perm = q_9_perm_0, x = var_1116)[name = tensor("transpose_27")]; + tensor k_9 = transpose(perm = k_9_perm_0, x = var_1157)[name = tensor("transpose_26")]; + tensor q_9 = transpose(perm = q_9_perm_0, x = var_1149)[name = tensor("transpose_27")]; tensor qk_9 = matmul(transpose_x = qk_9_transpose_x_1, transpose_y = qk_9_transpose_y_1, x = q_9, y = k_9)[name = tensor("qk_9")]; - tensor var_1142 = const()[name = tensor("op_1142"), val = tensor([1, 4])]; - tensor var_1143 = reshape(shape = var_1142, x = valid_mask)[name = tensor("op_1143")]; - tensor causal_with_valid_1 = mul(x = decoder__causal_mask, y = var_1143)[name = tensor("causal_with_valid_1")]; - tensor var_1145 = const()[name = tensor("op_1145"), val = tensor([4, 1])]; - tensor var_1146 = reshape(shape = var_1145, x = sqrt_s_t_9)[name = tensor("op_1146")]; - tensor M_9 = real_div(x = causal_with_valid_1, y = var_1146)[name = tensor("M_9")]; - tensor var_1148 = mul(x = qk_9, y = M_9)[name = tensor("op_1148")]; + tensor var_1175 = const()[name = tensor("op_1175"), val = tensor([1, 4])]; + tensor var_1176 = reshape(shape = var_1175, x = valid_mask)[name = tensor("op_1176")]; + tensor causal_with_valid_1 = mul(x = inner_decoder__causal_mask, y = var_1176)[name = tensor("causal_with_valid_1")]; + tensor var_1178 = const()[name = tensor("op_1178"), val = tensor([4, 1])]; + tensor var_1179 = reshape(shape = var_1178, x = sqrt_s_t_9)[name = tensor("op_1179")]; + tensor M_9 = real_div(x = causal_with_valid_1, y = var_1179)[name = tensor("M_9")]; + tensor var_1181 = mul(x = qk_9, y = M_9)[name = tensor("op_1181")]; tensor inner_9_transpose_x_0 = const()[name = tensor("inner_9_transpose_x_0"), val = tensor(false)]; tensor inner_9_transpose_y_0 = const()[name = tensor("inner_9_transpose_y_0"), val = tensor(false)]; - tensor v_9 = transpose(perm = v_9_perm_0, x = var_1130)[name = tensor("transpose_25")]; - tensor inner_9 = matmul(transpose_x = inner_9_transpose_x_0, transpose_y = inner_9_transpose_y_0, x = var_1148, y = v_9)[name = tensor("inner_9")]; - tensor var_1150_transpose_x_0 = const()[name = tensor("op_1150_transpose_x_0"), val = tensor(false)]; - tensor var_1150_transpose_y_0 = const()[name = tensor("op_1150_transpose_y_0"), val = tensor(false)]; - tensor var_1150 = matmul(transpose_x = var_1150_transpose_x_0, transpose_y = var_1150_transpose_y_0, x = q_9, y = prev_kv_9)[name = tensor("op_1150")]; - tensor var_1151 = real_div(x = sqrt_s0_9, y = sqrt_s_t_9)[name = tensor("op_1151")]; - tensor var_1152 = const()[name = tensor("op_1152"), val = tensor([1, 1, 4, 1])]; - tensor var_1153 = reshape(shape = var_1152, x = var_1151)[name = tensor("op_1153")]; - tensor cross_9 = mul(x = var_1150, y = var_1153)[name = tensor("cross_9")]; + tensor v_9 = transpose(perm = v_9_perm_0, x = var_1163)[name = tensor("transpose_25")]; + tensor inner_9 = matmul(transpose_x = inner_9_transpose_x_0, transpose_y = inner_9_transpose_y_0, x = var_1181, y = v_9)[name = tensor("inner_9")]; + tensor var_1183_transpose_x_0 = const()[name = tensor("op_1183_transpose_x_0"), val = tensor(false)]; + tensor var_1183_transpose_y_0 = const()[name = tensor("op_1183_transpose_y_0"), val = tensor(false)]; + tensor var_1183 = matmul(transpose_x = var_1183_transpose_x_0, transpose_y = var_1183_transpose_y_0, x = q_9, y = prev_kv_9)[name = tensor("op_1183")]; + tensor var_1184 = real_div(x = sqrt_s0_9, y = sqrt_s_t_9)[name = tensor("op_1184")]; + tensor var_1185 = const()[name = tensor("op_1185"), val = tensor([1, 1, 4, 1])]; + tensor var_1186 = reshape(shape = var_1185, x = var_1184)[name = tensor("op_1186")]; + tensor cross_9 = mul(x = var_1183, y = var_1186)[name = tensor("cross_9")]; tensor out_25 = add(x = inner_9, y = cross_9)[name = tensor("out_25")]; - tensor var_1156 = const()[name = tensor("op_1156"), val = tensor([1, 1, 4, 1])]; - tensor var_1157 = reshape(shape = var_1156, x = valid_mask)[name = tensor("op_1157")]; - tensor v_masked_1 = mul(x = v_9, y = var_1157)[name = tensor("v_masked_1")]; - tensor var_1159 = mul(x = prev_kv_9, y = sqrt_s0_9)[name = tensor("op_1159")]; - tensor var_1161_transpose_x_1 = const()[name = tensor("op_1161_transpose_x_1"), val = tensor(true)]; - tensor var_1161_transpose_y_1 = const()[name = tensor("op_1161_transpose_y_1"), val = tensor(false)]; - tensor var_1161 = matmul(transpose_x = var_1161_transpose_x_1, transpose_y = var_1161_transpose_y_1, x = k_9, y = v_masked_1)[name = tensor("op_1161")]; - tensor new_kv_unnorm_9 = add(x = var_1159, y = var_1161)[name = tensor("new_kv_unnorm_9")]; - tensor var_1163_keep_dims_0 = const()[name = tensor("op_1163_keep_dims_0"), val = tensor(false)]; - tensor var_1163 = reduce_sum(keep_dims = var_1163_keep_dims_0, x = valid_mask)[name = tensor("op_1163")]; - tensor var_1164 = const()[name = tensor("op_1164"), val = tensor([1])]; - tensor var_1165 = reshape(shape = var_1164, x = var_1163)[name = tensor("op_1165")]; - tensor new_scale_9 = add(x = prev_scale_9, y = var_1165)[name = tensor("new_scale_9")]; + tensor var_1189 = const()[name = tensor("op_1189"), val = tensor([1, 1, 4, 1])]; + tensor var_1190 = reshape(shape = var_1189, x = valid_mask)[name = tensor("op_1190")]; + tensor v_masked_1 = mul(x = v_9, y = var_1190)[name = tensor("v_masked_1")]; + tensor var_1192 = mul(x = prev_kv_9, y = sqrt_s0_9)[name = tensor("op_1192")]; + tensor var_1194_transpose_x_1 = const()[name = tensor("op_1194_transpose_x_1"), val = tensor(true)]; + tensor var_1194_transpose_y_1 = const()[name = tensor("op_1194_transpose_y_1"), val = tensor(false)]; + tensor var_1194 = matmul(transpose_x = var_1194_transpose_x_1, transpose_y = var_1194_transpose_y_1, x = k_9, y = v_masked_1)[name = tensor("op_1194")]; + tensor new_kv_unnorm_9 = add(x = var_1192, y = var_1194)[name = tensor("new_kv_unnorm_9")]; + tensor var_1196_keep_dims_0 = const()[name = tensor("op_1196_keep_dims_0"), val = tensor(false)]; + tensor var_1196 = reduce_sum(keep_dims = var_1196_keep_dims_0, x = valid_mask)[name = tensor("op_1196")]; + tensor var_1197 = const()[name = tensor("op_1197"), val = tensor([1])]; + tensor var_1198 = reshape(shape = var_1197, x = var_1196)[name = tensor("op_1198")]; + tensor new_scale_9 = add(x = prev_scale_9, y = var_1198)[name = tensor("new_scale_9")]; tensor const_21 = const()[name = tensor("const_21"), val = tensor(0x1.fffffep+127)]; - tensor clip_2 = clip(alpha = var_1022, beta = const_21, x = new_scale_9)[name = tensor("clip_2")]; + tensor clip_2 = clip(alpha = var_59, beta = const_21, x = new_scale_9)[name = tensor("clip_2")]; tensor sqrt_new_scale_1 = sqrt(x = clip_2)[name = tensor("sqrt_new_scale_1")]; - tensor var_1169 = real_div(x = new_kv_unnorm_9, y = sqrt_new_scale_1)[name = tensor("op_1169")]; - tensor var_1170_perm_0 = const()[name = tensor("op_1170_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_1202 = real_div(x = new_kv_unnorm_9, y = sqrt_new_scale_1)[name = tensor("op_1202")]; + tensor var_1203_perm_0 = const()[name = tensor("op_1203_perm_0"), val = tensor([0, 2, 1, 3])]; tensor out_27_axes_0 = const()[name = tensor("out_27_axes_0"), val = tensor([-1])]; - tensor var_1170 = transpose(perm = var_1170_perm_0, x = out_25)[name = tensor("transpose_24")]; - tensor out_27 = layer_norm(axes = out_27_axes_0, epsilon = var_1019, x = var_1170)[name = tensor("out_27")]; - tensor var_1174 = const()[name = tensor("op_1174"), val = tensor([12, 4, 256])]; - tensor out_29 = reshape(shape = var_1174, x = out_27)[name = tensor("out_29")]; - tensor var_1176 = silu(x = input_169)[name = tensor("op_1176")]; - tensor input_171 = mul(x = var_1176, y = out_29)[name = tensor("input_171")]; - tensor ret_out_9 = linear(bias = decoder_out_proj_0_bias, weight = decoder_out_proj_0_weight, x = input_171)[name = tensor("linear_42")]; - tensor input_173 = add(x = x_29, y = ret_out_9)[name = tensor("input_173")]; + tensor var_1203 = transpose(perm = var_1203_perm_0, x = out_25)[name = tensor("transpose_24")]; + tensor out_27 = layer_norm(axes = out_27_axes_0, epsilon = var_74, x = var_1203)[name = tensor("out_27")]; + tensor var_1207 = const()[name = tensor("op_1207"), val = tensor([12, 4, 256])]; + tensor out_29 = reshape(shape = var_1207, x = out_27)[name = tensor("out_29")]; + tensor var_1209 = silu(x = input_171)[name = tensor("op_1209")]; + tensor input_173 = mul(x = var_1209, y = out_29)[name = tensor("input_173")]; + tensor ret_out_9 = linear(bias = inner_decoder_out_proj_0_bias, weight = inner_decoder_out_proj_0_weight, x = input_173)[name = tensor("linear_42")]; + tensor input_175 = add(x = x_29, y = ret_out_9)[name = tensor("input_175")]; tensor xt_1_axes_0 = const()[name = tensor("xt_1_axes_0"), val = tensor([-1])]; - tensor xt_1 = layer_norm(axes = xt_1_axes_0, beta = decoder_norm11_0_bias, epsilon = var_1017, gamma = decoder_norm11_0_weight, x = input_173)[name = tensor("xt_1")]; - tensor var_1186 = const()[name = tensor("op_1186"), val = tensor([1, 12, 4, 256])]; - tensor var_1187 = reshape(shape = var_1186, x = xt_1)[name = tensor("op_1187")]; - tensor var_1188_perm_0 = const()[name = tensor("op_1188_perm_0"), val = tensor([0, 2, 1, 3])]; - tensor var_1191 = const()[name = tensor("op_1191"), val = tensor([4, 12, 256])]; - tensor var_1188 = transpose(perm = var_1188_perm_0, x = var_1187)[name = tensor("transpose_23")]; - tensor query_1 = reshape(shape = var_1191, x = var_1188)[name = tensor("query_1")]; + tensor xt_1 = layer_norm(axes = xt_1_axes_0, beta = inner_decoder_norm11_0_bias, epsilon = var_66, gamma = inner_decoder_norm11_0_weight, x = input_175)[name = tensor("xt_1")]; + tensor var_1219 = const()[name = tensor("op_1219"), val = tensor([1, 12, 4, 256])]; + tensor var_1220 = reshape(shape = var_1219, x = xt_1)[name = tensor("op_1220")]; + tensor var_1221_perm_0 = const()[name = tensor("op_1221_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_1224 = const()[name = tensor("op_1224"), val = tensor([4, 12, 256])]; + tensor var_1221 = transpose(perm = var_1221_perm_0, x = var_1220)[name = tensor("transpose_23")]; + tensor query_1 = reshape(shape = var_1224, x = var_1221)[name = tensor("query_1")]; tensor query_3_perm_0 = const()[name = tensor("query_3_perm_0"), val = tensor([1, 0, 2])]; tensor query_3 = transpose(perm = query_3_perm_0, x = query_1)[name = tensor("transpose_22")]; - tensor var_1214 = linear(bias = decoder_self_attn2_0_in_proj_bias, weight = decoder_self_attn2_0_in_proj_weight, x = query_3)[name = tensor("linear_43")]; + tensor var_1247 = linear(bias = inner_decoder_self_attn2_0_in_proj_bias, weight = inner_decoder_self_attn2_0_in_proj_weight, x = query_3)[name = tensor("linear_43")]; tensor concat_1 = const()[name = tensor("concat_1"), val = tensor([12, 4, 3, 256])]; - tensor var_1216 = reshape(shape = concat_1, x = var_1214)[name = tensor("op_1216")]; - tensor var_1217_axes_0 = const()[name = tensor("op_1217_axes_0"), val = tensor([0])]; - tensor var_1217 = expand_dims(axes = var_1217_axes_0, x = var_1216)[name = tensor("op_1217")]; - tensor var_1218_perm_0 = const()[name = tensor("op_1218_perm_0"), val = tensor([-2, 1, 2, 0, 4])]; - tensor var_1219_axes_0 = const()[name = tensor("op_1219_axes_0"), val = tensor([-2])]; - tensor var_1218 = transpose(perm = var_1218_perm_0, x = var_1217)[name = tensor("transpose_21")]; - tensor var_1219 = squeeze(axes = var_1219_axes_0, x = var_1218)[name = tensor("op_1219")]; + tensor var_1249 = reshape(shape = concat_1, x = var_1247)[name = tensor("op_1249")]; + tensor var_1250_axes_0 = const()[name = tensor("op_1250_axes_0"), val = tensor([0])]; + tensor var_1250 = expand_dims(axes = var_1250_axes_0, x = var_1249)[name = tensor("op_1250")]; + tensor var_1251_perm_0 = const()[name = tensor("op_1251_perm_0"), val = tensor([-2, 1, 2, 0, 4])]; + tensor var_1252_axes_0 = const()[name = tensor("op_1252_axes_0"), val = tensor([-2])]; + tensor var_1251 = transpose(perm = var_1251_perm_0, x = var_1250)[name = tensor("transpose_21")]; + tensor var_1252 = squeeze(axes = var_1252_axes_0, x = var_1251)[name = tensor("op_1252")]; tensor q_11_begin_0 = const()[name = tensor("q_11_begin_0"), val = tensor([0, 0, 0, 0])]; tensor q_11_end_0 = const()[name = tensor("q_11_end_0"), val = tensor([1, 12, 4, 256])]; tensor q_11_end_mask_0 = const()[name = tensor("q_11_end_mask_0"), val = tensor([false, true, true, true])]; tensor q_11_squeeze_mask_0 = const()[name = tensor("q_11_squeeze_mask_0"), val = tensor([true, false, false, false])]; - tensor q_11 = slice_by_index(begin = q_11_begin_0, end = q_11_end_0, end_mask = q_11_end_mask_0, squeeze_mask = q_11_squeeze_mask_0, x = var_1219)[name = tensor("q_11")]; + tensor q_11 = slice_by_index(begin = q_11_begin_0, end = q_11_end_0, end_mask = q_11_end_mask_0, squeeze_mask = q_11_squeeze_mask_0, x = var_1252)[name = tensor("q_11")]; tensor k_11_begin_0 = const()[name = tensor("k_11_begin_0"), val = tensor([1, 0, 0, 0])]; tensor k_11_end_0 = const()[name = tensor("k_11_end_0"), val = tensor([2, 12, 4, 256])]; tensor k_11_end_mask_0 = const()[name = tensor("k_11_end_mask_0"), val = tensor([false, true, true, true])]; tensor k_11_squeeze_mask_0 = const()[name = tensor("k_11_squeeze_mask_0"), val = tensor([true, false, false, false])]; - tensor k_11 = slice_by_index(begin = k_11_begin_0, end = k_11_end_0, end_mask = k_11_end_mask_0, squeeze_mask = k_11_squeeze_mask_0, x = var_1219)[name = tensor("k_11")]; + tensor k_11 = slice_by_index(begin = k_11_begin_0, end = k_11_end_0, end_mask = k_11_end_mask_0, squeeze_mask = k_11_squeeze_mask_0, x = var_1252)[name = tensor("k_11")]; tensor v_11_begin_0 = const()[name = tensor("v_11_begin_0"), val = tensor([2, 0, 0, 0])]; tensor v_11_end_0 = const()[name = tensor("v_11_end_0"), val = tensor([3, 12, 4, 256])]; tensor v_11_end_mask_0 = const()[name = tensor("v_11_end_mask_0"), val = tensor([false, true, true, true])]; tensor v_11_squeeze_mask_0 = const()[name = tensor("v_11_squeeze_mask_0"), val = tensor([true, false, false, false])]; - tensor v_11 = slice_by_index(begin = v_11_begin_0, end = v_11_end_0, end_mask = v_11_end_mask_0, squeeze_mask = v_11_squeeze_mask_0, x = var_1219)[name = tensor("v_11")]; - tensor var_1227 = const()[name = tensor("op_1227"), val = tensor([12, 16, 64])]; - tensor var_1228 = reshape(shape = var_1227, x = q_11)[name = tensor("op_1228")]; + tensor v_11 = slice_by_index(begin = v_11_begin_0, end = v_11_end_0, end_mask = v_11_end_mask_0, squeeze_mask = v_11_squeeze_mask_0, x = var_1252)[name = tensor("v_11")]; + tensor var_1260 = const()[name = tensor("op_1260"), val = tensor([12, 16, 64])]; + tensor var_1261 = reshape(shape = var_1260, x = q_11)[name = tensor("op_1261")]; tensor q_13_perm_0 = const()[name = tensor("q_13_perm_0"), val = tensor([1, 0, 2])]; - tensor var_1234 = const()[name = tensor("op_1234"), val = tensor([12, 16, 64])]; - tensor var_1235 = reshape(shape = var_1234, x = k_11)[name = tensor("op_1235")]; + tensor var_1267 = const()[name = tensor("op_1267"), val = tensor([12, 16, 64])]; + tensor var_1268 = reshape(shape = var_1267, x = k_11)[name = tensor("op_1268")]; tensor k_13_perm_0 = const()[name = tensor("k_13_perm_0"), val = tensor([1, 0, 2])]; - tensor var_1241 = const()[name = tensor("op_1241"), val = tensor([12, 16, 64])]; - tensor var_1242 = reshape(shape = var_1241, x = v_11)[name = tensor("op_1242")]; + tensor var_1274 = const()[name = tensor("op_1274"), val = tensor([12, 16, 64])]; + tensor var_1275 = reshape(shape = var_1274, x = v_11)[name = tensor("op_1275")]; tensor v_13_perm_0 = const()[name = tensor("v_13_perm_0"), val = tensor([1, 0, 2])]; - tensor var_1245 = const()[name = tensor("op_1245"), val = tensor([4, 4, 12, 64])]; - tensor q_13 = transpose(perm = q_13_perm_0, x = var_1228)[name = tensor("transpose_20")]; - tensor q_15 = reshape(shape = var_1245, x = q_13)[name = tensor("q_15")]; - tensor var_1247 = const()[name = tensor("op_1247"), val = tensor([4, 4, 12, 64])]; - tensor k_13 = transpose(perm = k_13_perm_0, x = var_1235)[name = tensor("transpose_19")]; - tensor k_15 = reshape(shape = var_1247, x = k_13)[name = tensor("k_15")]; - tensor var_1249 = const()[name = tensor("op_1249"), val = tensor([4, 4, 12, 64])]; - tensor v_13 = transpose(perm = v_13_perm_0, x = var_1242)[name = tensor("transpose_18")]; - tensor v_15 = reshape(shape = var_1249, x = v_13)[name = tensor("v_15")]; + tensor var_1278 = const()[name = tensor("op_1278"), val = tensor([4, 4, 12, 64])]; + tensor q_13 = transpose(perm = q_13_perm_0, x = var_1261)[name = tensor("transpose_20")]; + tensor q_15 = reshape(shape = var_1278, x = q_13)[name = tensor("q_15")]; + tensor var_1280 = const()[name = tensor("op_1280"), val = tensor([4, 4, 12, 64])]; + tensor k_13 = transpose(perm = k_13_perm_0, x = var_1268)[name = tensor("transpose_19")]; + tensor k_15 = reshape(shape = var_1280, x = k_13)[name = tensor("k_15")]; + tensor var_1282 = const()[name = tensor("op_1282"), val = tensor([4, 4, 12, 64])]; + tensor v_13 = transpose(perm = v_13_perm_0, x = var_1275)[name = tensor("transpose_18")]; + tensor v_15 = reshape(shape = var_1282, x = v_13)[name = tensor("v_15")]; tensor mul_1_y_0 = const()[name = tensor("mul_1_y_0"), val = tensor(0x1p-3)]; tensor mul_1 = mul(x = q_15, y = mul_1_y_0)[name = tensor("mul_1")]; tensor matmul_0_transpose_y_0 = const()[name = tensor("matmul_0_transpose_y_0"), val = tensor(true)]; @@ -1119,30 +1134,30 @@ program(1.0) tensor attn_output_1_transpose_x_0 = const()[name = tensor("attn_output_1_transpose_x_0"), val = tensor(false)]; tensor attn_output_1_transpose_y_0 = const()[name = tensor("attn_output_1_transpose_y_0"), val = tensor(false)]; tensor attn_output_1 = matmul(transpose_x = attn_output_1_transpose_x_0, transpose_y = attn_output_1_transpose_y_0, x = softmax_0, y = v_15)[name = tensor("attn_output_1")]; - tensor var_1252 = const()[name = tensor("op_1252"), val = tensor([2, 0, 1, 3])]; - tensor var_1257 = const()[name = tensor("op_1257"), val = tensor([48, 256])]; - tensor var_1253 = transpose(perm = var_1252, x = attn_output_1)[name = tensor("transpose_17")]; - tensor attn_output_3 = reshape(shape = var_1257, x = var_1253)[name = tensor("attn_output_3")]; - tensor attn_output_5 = linear(bias = decoder_self_attn2_0_out_proj_bias, weight = decoder_self_attn2_0_out_proj_weight, x = attn_output_3)[name = tensor("linear_44")]; - tensor var_1261 = const()[name = tensor("op_1261"), val = tensor([12, 4, 256])]; - tensor attn_output_7 = reshape(shape = var_1261, x = attn_output_5)[name = tensor("attn_output_7")]; + tensor var_1285 = const()[name = tensor("op_1285"), val = tensor([2, 0, 1, 3])]; + tensor var_1290 = const()[name = tensor("op_1290"), val = tensor([48, 256])]; + tensor var_1286 = transpose(perm = var_1285, x = attn_output_1)[name = tensor("transpose_17")]; + tensor attn_output_3 = reshape(shape = var_1290, x = var_1286)[name = tensor("attn_output_3")]; + tensor attn_output_5 = linear(bias = inner_decoder_self_attn2_0_out_proj_bias, weight = inner_decoder_self_attn2_0_out_proj_weight, x = attn_output_3)[name = tensor("linear_44")]; + tensor var_1294 = const()[name = tensor("op_1294"), val = tensor([12, 4, 256])]; + tensor attn_output_7 = reshape(shape = var_1294, x = attn_output_5)[name = tensor("attn_output_7")]; tensor sa2_out_1_perm_0 = const()[name = tensor("sa2_out_1_perm_0"), val = tensor([1, 0, 2])]; tensor sa2_out_1 = transpose(perm = sa2_out_1_perm_0, x = attn_output_7)[name = tensor("transpose_16")]; - tensor input_175 = add(x = query_1, y = sa2_out_1)[name = tensor("input_175")]; - tensor input_177_axes_0 = const()[name = tensor("input_177_axes_0"), val = tensor([-1])]; - tensor input_177 = layer_norm(axes = input_177_axes_0, beta = decoder_norm21_0_bias, epsilon = var_1017, gamma = decoder_norm21_0_weight, x = input_175)[name = tensor("input_177")]; - tensor input_179 = linear(bias = decoder_linear1_0_bias, weight = decoder_linear1_0_weight, x = input_177)[name = tensor("linear_45")]; - tensor input_181 = relu(x = input_179)[name = tensor("input_181")]; - tensor ffn_out_1 = linear(bias = decoder_linear2_0_bias, weight = decoder_linear2_0_weight, x = input_181)[name = tensor("linear_46")]; - tensor input_183 = add(x = input_177, y = ffn_out_1)[name = tensor("input_183")]; + tensor input_177 = add(x = query_1, y = sa2_out_1)[name = tensor("input_177")]; + tensor input_179_axes_0 = const()[name = tensor("input_179_axes_0"), val = tensor([-1])]; + tensor input_179 = layer_norm(axes = input_179_axes_0, beta = inner_decoder_norm21_0_bias, epsilon = var_66, gamma = inner_decoder_norm21_0_weight, x = input_177)[name = tensor("input_179")]; + tensor input_181 = linear(bias = inner_decoder_linear1_0_bias, weight = inner_decoder_linear1_0_weight, x = input_179)[name = tensor("linear_45")]; + tensor input_183 = relu(x = input_181)[name = tensor("input_183")]; + tensor ffn_out_1 = linear(bias = inner_decoder_linear2_0_bias, weight = inner_decoder_linear2_0_weight, x = input_183)[name = tensor("linear_46")]; + tensor input_185 = add(x = input_179, y = ffn_out_1)[name = tensor("input_185")]; tensor xt_3_axes_0 = const()[name = tensor("xt_3_axes_0"), val = tensor([-1])]; - tensor xt_3 = layer_norm(axes = xt_3_axes_0, beta = decoder_norm22_0_bias, epsilon = var_1017, gamma = decoder_norm22_0_weight, x = input_183)[name = tensor("xt_3")]; - tensor var_1281 = const()[name = tensor("op_1281"), val = tensor([1, 4, 12, 256])]; - tensor x_31 = reshape(shape = var_1281, x = xt_3)[name = tensor("x_31")]; - tensor var_1283_perm_0 = const()[name = tensor("op_1283_perm_0"), val = tensor([0, 2, 1, 3])]; - tensor var_1287 = const()[name = tensor("op_1287"), val = tensor([12, 4, 256])]; - tensor var_1283 = transpose(perm = var_1283_perm_0, x = x_31)[name = tensor("transpose_15")]; - tensor x = reshape(shape = var_1287, x = var_1283)[name = tensor("x")]; + tensor xt_3 = layer_norm(axes = xt_3_axes_0, beta = inner_decoder_norm22_0_bias, epsilon = var_66, gamma = inner_decoder_norm22_0_weight, x = input_185)[name = tensor("xt_3")]; + tensor var_1314 = const()[name = tensor("op_1314"), val = tensor([1, 4, 12, 256])]; + tensor x_31 = reshape(shape = var_1314, x = xt_3)[name = tensor("x_31")]; + tensor var_1316_perm_0 = const()[name = tensor("op_1316_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_1320 = const()[name = tensor("op_1320"), val = tensor([12, 4, 256])]; + tensor var_1316 = transpose(perm = var_1316_perm_0, x = x_31)[name = tensor("transpose_15")]; + tensor x = reshape(shape = var_1320, x = var_1316)[name = tensor("x")]; tensor prev_kv_begin_0 = const()[name = tensor("prev_kv_begin_0"), val = tensor([1, 0, 0, 0, 0])]; tensor prev_kv_end_0 = const()[name = tensor("prev_kv_end_0"), val = tensor([2, 12, 4, 64, 64])]; tensor prev_kv_end_mask_0 = const()[name = tensor("prev_kv_end_mask_0"), val = tensor([false, true, true, true, true])]; @@ -1153,120 +1168,120 @@ program(1.0) tensor prev_scale_end_mask_0 = const()[name = tensor("prev_scale_end_mask_0"), val = tensor([false, true])]; tensor prev_scale_squeeze_mask_0 = const()[name = tensor("prev_scale_squeeze_mask_0"), val = tensor([true, false])]; tensor prev_scale = slice_by_index(begin = prev_scale_begin_0, end = prev_scale_end_0, end_mask = prev_scale_end_mask_0, squeeze_mask = prev_scale_squeeze_mask_0, x = dec_scale)[name = tensor("prev_scale")]; - tensor var_1295 = linear(bias = decoder_q_proj_1_bias, weight = decoder_q_proj_1_weight, x = x)[name = tensor("linear_47")]; - tensor var_1296 = const()[name = tensor("op_1296"), val = tensor([12, 4, 4, 64])]; - tensor var_1297 = reshape(shape = var_1296, x = var_1295)[name = tensor("op_1297")]; + tensor var_1328 = linear(bias = inner_decoder_q_proj_1_bias, weight = inner_decoder_q_proj_1_weight, x = x)[name = tensor("linear_47")]; + tensor var_1329 = const()[name = tensor("op_1329"), val = tensor([12, 4, 4, 64])]; + tensor var_1330 = reshape(shape = var_1329, x = var_1328)[name = tensor("op_1330")]; tensor q_17_perm_0 = const()[name = tensor("q_17_perm_0"), val = tensor([0, 2, 1, 3])]; - tensor var_1301 = linear(bias = decoder_k_proj_1_bias, weight = decoder_k_proj_1_weight, x = x)[name = tensor("linear_48")]; - tensor var_1302 = const()[name = tensor("op_1302"), val = tensor(0x1p-3)]; - tensor var_1303 = mul(x = var_1301, y = var_1302)[name = tensor("op_1303")]; - tensor var_1304 = const()[name = tensor("op_1304"), val = tensor([12, 4, 4, 64])]; - tensor var_1305 = reshape(shape = var_1304, x = var_1303)[name = tensor("op_1305")]; + tensor var_1334 = linear(bias = inner_decoder_k_proj_1_bias, weight = inner_decoder_k_proj_1_weight, x = x)[name = tensor("linear_48")]; + tensor var_1335 = const()[name = tensor("op_1335"), val = tensor(0x1p-3)]; + tensor var_1336 = mul(x = var_1334, y = var_1335)[name = tensor("op_1336")]; + tensor var_1337 = const()[name = tensor("op_1337"), val = tensor([12, 4, 4, 64])]; + tensor var_1338 = reshape(shape = var_1337, x = var_1336)[name = tensor("op_1338")]; tensor k_17_perm_0 = const()[name = tensor("k_17_perm_0"), val = tensor([0, 2, 1, 3])]; - tensor var_1309 = linear(bias = decoder_v_proj_1_bias, weight = decoder_v_proj_1_weight, x = x)[name = tensor("linear_49")]; - tensor var_1310 = const()[name = tensor("op_1310"), val = tensor([12, 4, 4, 64])]; - tensor var_1311 = reshape(shape = var_1310, x = var_1309)[name = tensor("op_1311")]; + tensor var_1342 = linear(bias = inner_decoder_v_proj_1_bias, weight = inner_decoder_v_proj_1_weight, x = x)[name = tensor("linear_49")]; + tensor var_1343 = const()[name = tensor("op_1343"), val = tensor([12, 4, 4, 64])]; + tensor var_1344 = reshape(shape = var_1343, x = var_1342)[name = tensor("op_1344")]; tensor v_17_perm_0 = const()[name = tensor("v_17_perm_0"), val = tensor([0, 2, 1, 3])]; - tensor input_187 = linear(bias = decoder_g_proj_1_bias, weight = decoder_g_proj_1_weight, x = x)[name = tensor("linear_50")]; + tensor input_189 = linear(bias = inner_decoder_g_proj_1_bias, weight = inner_decoder_g_proj_1_weight, x = x)[name = tensor("linear_50")]; tensor sqrt_s0 = sqrt(x = prev_scale)[name = tensor("sqrt_s0")]; tensor s_t = add(x = prev_scale, y = cumsum_mask_1)[name = tensor("s_t")]; tensor const_32 = const()[name = tensor("const_32"), val = tensor(0x1.fffffep+127)]; - tensor clip_3 = clip(alpha = var_1022, beta = const_32, x = s_t)[name = tensor("clip_3")]; + tensor clip_3 = clip(alpha = var_59, beta = const_32, x = s_t)[name = tensor("clip_3")]; tensor sqrt_s_t = sqrt(x = clip_3)[name = tensor("sqrt_s_t")]; tensor qk_transpose_x_1 = const()[name = tensor("qk_transpose_x_1"), val = tensor(false)]; tensor qk_transpose_y_1 = const()[name = tensor("qk_transpose_y_1"), val = tensor(true)]; - tensor k_17 = transpose(perm = k_17_perm_0, x = var_1305)[name = tensor("transpose_13")]; - tensor q_17 = transpose(perm = q_17_perm_0, x = var_1297)[name = tensor("transpose_14")]; + tensor k_17 = transpose(perm = k_17_perm_0, x = var_1338)[name = tensor("transpose_13")]; + tensor q_17 = transpose(perm = q_17_perm_0, x = var_1330)[name = tensor("transpose_14")]; tensor qk = matmul(transpose_x = qk_transpose_x_1, transpose_y = qk_transpose_y_1, x = q_17, y = k_17)[name = tensor("qk")]; - tensor var_1326 = const()[name = tensor("op_1326"), val = tensor([4, 1])]; - tensor var_1327 = reshape(shape = var_1326, x = sqrt_s_t)[name = tensor("op_1327")]; - tensor M = real_div(x = causal_with_valid_1, y = var_1327)[name = tensor("M")]; - tensor var_1329 = mul(x = qk, y = M)[name = tensor("op_1329")]; - tensor inner_transpose_x_0 = const()[name = tensor("inner_transpose_x_0"), val = tensor(false)]; - tensor inner_transpose_y_0 = const()[name = tensor("inner_transpose_y_0"), val = tensor(false)]; - tensor v_17 = transpose(perm = v_17_perm_0, x = var_1311)[name = tensor("transpose_12")]; - tensor inner = matmul(transpose_x = inner_transpose_x_0, transpose_y = inner_transpose_y_0, x = var_1329, y = v_17)[name = tensor("inner")]; - tensor var_1331_transpose_x_0 = const()[name = tensor("op_1331_transpose_x_0"), val = tensor(false)]; - tensor var_1331_transpose_y_0 = const()[name = tensor("op_1331_transpose_y_0"), val = tensor(false)]; - tensor var_1331 = matmul(transpose_x = var_1331_transpose_x_0, transpose_y = var_1331_transpose_y_0, x = q_17, y = prev_kv)[name = tensor("op_1331")]; - tensor var_1332 = real_div(x = sqrt_s0, y = sqrt_s_t)[name = tensor("op_1332")]; - tensor var_1333 = const()[name = tensor("op_1333"), val = tensor([1, 1, 4, 1])]; - tensor var_1334 = reshape(shape = var_1333, x = var_1332)[name = tensor("op_1334")]; - tensor cross = mul(x = var_1331, y = var_1334)[name = tensor("cross")]; - tensor out_31 = add(x = inner, y = cross)[name = tensor("out_31")]; - tensor v_masked = mul(x = v_17, y = var_1157)[name = tensor("v_masked")]; - tensor var_1340 = mul(x = prev_kv, y = sqrt_s0)[name = tensor("op_1340")]; - tensor var_1342_transpose_x_1 = const()[name = tensor("op_1342_transpose_x_1"), val = tensor(true)]; - tensor var_1342_transpose_y_1 = const()[name = tensor("op_1342_transpose_y_1"), val = tensor(false)]; - tensor var_1342 = matmul(transpose_x = var_1342_transpose_x_1, transpose_y = var_1342_transpose_y_1, x = k_17, y = v_masked)[name = tensor("op_1342")]; - tensor new_kv_unnorm = add(x = var_1340, y = var_1342)[name = tensor("new_kv_unnorm")]; - tensor new_scale = add(x = prev_scale, y = var_1165)[name = tensor("new_scale")]; + tensor var_1359 = const()[name = tensor("op_1359"), val = tensor([4, 1])]; + tensor var_1360 = reshape(shape = var_1359, x = sqrt_s_t)[name = tensor("op_1360")]; + tensor M = real_div(x = causal_with_valid_1, y = var_1360)[name = tensor("M")]; + tensor var_1362 = mul(x = qk, y = M)[name = tensor("op_1362")]; + tensor inner_11_transpose_x_0 = const()[name = tensor("inner_11_transpose_x_0"), val = tensor(false)]; + tensor inner_11_transpose_y_0 = const()[name = tensor("inner_11_transpose_y_0"), val = tensor(false)]; + tensor v_17 = transpose(perm = v_17_perm_0, x = var_1344)[name = tensor("transpose_12")]; + tensor inner_11 = matmul(transpose_x = inner_11_transpose_x_0, transpose_y = inner_11_transpose_y_0, x = var_1362, y = v_17)[name = tensor("inner_11")]; + tensor var_1364_transpose_x_0 = const()[name = tensor("op_1364_transpose_x_0"), val = tensor(false)]; + tensor var_1364_transpose_y_0 = const()[name = tensor("op_1364_transpose_y_0"), val = tensor(false)]; + tensor var_1364 = matmul(transpose_x = var_1364_transpose_x_0, transpose_y = var_1364_transpose_y_0, x = q_17, y = prev_kv)[name = tensor("op_1364")]; + tensor var_1365 = real_div(x = sqrt_s0, y = sqrt_s_t)[name = tensor("op_1365")]; + tensor var_1366 = const()[name = tensor("op_1366"), val = tensor([1, 1, 4, 1])]; + tensor var_1367 = reshape(shape = var_1366, x = var_1365)[name = tensor("op_1367")]; + tensor cross = mul(x = var_1364, y = var_1367)[name = tensor("cross")]; + tensor out_31 = add(x = inner_11, y = cross)[name = tensor("out_31")]; + tensor v_masked = mul(x = v_17, y = var_1190)[name = tensor("v_masked")]; + tensor var_1373 = mul(x = prev_kv, y = sqrt_s0)[name = tensor("op_1373")]; + tensor var_1375_transpose_x_1 = const()[name = tensor("op_1375_transpose_x_1"), val = tensor(true)]; + tensor var_1375_transpose_y_1 = const()[name = tensor("op_1375_transpose_y_1"), val = tensor(false)]; + tensor var_1375 = matmul(transpose_x = var_1375_transpose_x_1, transpose_y = var_1375_transpose_y_1, x = k_17, y = v_masked)[name = tensor("op_1375")]; + tensor new_kv_unnorm = add(x = var_1373, y = var_1375)[name = tensor("new_kv_unnorm")]; + tensor new_scale = add(x = prev_scale, y = var_1198)[name = tensor("new_scale")]; tensor const_33 = const()[name = tensor("const_33"), val = tensor(0x1.fffffep+127)]; - tensor clip_4 = clip(alpha = var_1022, beta = const_33, x = new_scale)[name = tensor("clip_4")]; + tensor clip_4 = clip(alpha = var_59, beta = const_33, x = new_scale)[name = tensor("clip_4")]; tensor sqrt_new_scale = sqrt(x = clip_4)[name = tensor("sqrt_new_scale")]; tensor nkv = real_div(x = new_kv_unnorm, y = sqrt_new_scale)[name = tensor("nkv")]; - tensor var_1351_perm_0 = const()[name = tensor("op_1351_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_1384_perm_0 = const()[name = tensor("op_1384_perm_0"), val = tensor([0, 2, 1, 3])]; tensor out_33_axes_0 = const()[name = tensor("out_33_axes_0"), val = tensor([-1])]; - tensor var_1351 = transpose(perm = var_1351_perm_0, x = out_31)[name = tensor("transpose_11")]; - tensor out_33 = layer_norm(axes = out_33_axes_0, epsilon = var_1019, x = var_1351)[name = tensor("out_33")]; - tensor var_1355 = const()[name = tensor("op_1355"), val = tensor([12, 4, 256])]; - tensor out = reshape(shape = var_1355, x = out_33)[name = tensor("out")]; - tensor var_1357 = silu(x = input_187)[name = tensor("op_1357")]; - tensor input_189 = mul(x = var_1357, y = out)[name = tensor("input_189")]; - tensor ret_out = linear(bias = decoder_out_proj_1_bias, weight = decoder_out_proj_1_weight, x = input_189)[name = tensor("linear_51")]; - tensor input_191 = add(x = x, y = ret_out)[name = tensor("input_191")]; + tensor var_1384 = transpose(perm = var_1384_perm_0, x = out_31)[name = tensor("transpose_11")]; + tensor out_33 = layer_norm(axes = out_33_axes_0, epsilon = var_74, x = var_1384)[name = tensor("out_33")]; + tensor var_1388 = const()[name = tensor("op_1388"), val = tensor([12, 4, 256])]; + tensor out = reshape(shape = var_1388, x = out_33)[name = tensor("out")]; + tensor var_1390 = silu(x = input_189)[name = tensor("op_1390")]; + tensor input_191 = mul(x = var_1390, y = out)[name = tensor("input_191")]; + tensor ret_out = linear(bias = inner_decoder_out_proj_1_bias, weight = inner_decoder_out_proj_1_weight, x = input_191)[name = tensor("linear_51")]; + tensor input_193 = add(x = x, y = ret_out)[name = tensor("input_193")]; tensor xt_5_axes_0 = const()[name = tensor("xt_5_axes_0"), val = tensor([-1])]; - tensor xt_5 = layer_norm(axes = xt_5_axes_0, beta = decoder_norm11_1_bias, epsilon = var_1017, gamma = decoder_norm11_1_weight, x = input_191)[name = tensor("xt_5")]; - tensor var_1367 = const()[name = tensor("op_1367"), val = tensor([1, 12, 4, 256])]; - tensor var_1368 = reshape(shape = var_1367, x = xt_5)[name = tensor("op_1368")]; - tensor var_1369_perm_0 = const()[name = tensor("op_1369_perm_0"), val = tensor([0, 2, 1, 3])]; - tensor var_1372 = const()[name = tensor("op_1372"), val = tensor([4, 12, 256])]; - tensor var_1369 = transpose(perm = var_1369_perm_0, x = var_1368)[name = tensor("transpose_10")]; - tensor query_5 = reshape(shape = var_1372, x = var_1369)[name = tensor("query_5")]; + tensor xt_5 = layer_norm(axes = xt_5_axes_0, beta = inner_decoder_norm11_1_bias, epsilon = var_66, gamma = inner_decoder_norm11_1_weight, x = input_193)[name = tensor("xt_5")]; + tensor var_1400 = const()[name = tensor("op_1400"), val = tensor([1, 12, 4, 256])]; + tensor var_1401 = reshape(shape = var_1400, x = xt_5)[name = tensor("op_1401")]; + tensor var_1402_perm_0 = const()[name = tensor("op_1402_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_1405 = const()[name = tensor("op_1405"), val = tensor([4, 12, 256])]; + tensor var_1402 = transpose(perm = var_1402_perm_0, x = var_1401)[name = tensor("transpose_10")]; + tensor query_5 = reshape(shape = var_1405, x = var_1402)[name = tensor("query_5")]; tensor query_perm_0 = const()[name = tensor("query_perm_0"), val = tensor([1, 0, 2])]; tensor query = transpose(perm = query_perm_0, x = query_5)[name = tensor("transpose_9")]; - tensor var_1395 = linear(bias = decoder_self_attn2_1_in_proj_bias, weight = decoder_self_attn2_1_in_proj_weight, x = query)[name = tensor("linear_52")]; + tensor var_1428 = linear(bias = inner_decoder_self_attn2_1_in_proj_bias, weight = inner_decoder_self_attn2_1_in_proj_weight, x = query)[name = tensor("linear_52")]; tensor concat_2 = const()[name = tensor("concat_2"), val = tensor([12, 4, 3, 256])]; - tensor var_1397 = reshape(shape = concat_2, x = var_1395)[name = tensor("op_1397")]; - tensor var_1398_axes_0 = const()[name = tensor("op_1398_axes_0"), val = tensor([0])]; - tensor var_1398 = expand_dims(axes = var_1398_axes_0, x = var_1397)[name = tensor("op_1398")]; - tensor var_1399_perm_0 = const()[name = tensor("op_1399_perm_0"), val = tensor([-2, 1, 2, 0, 4])]; - tensor var_1400_axes_0 = const()[name = tensor("op_1400_axes_0"), val = tensor([-2])]; - tensor var_1399 = transpose(perm = var_1399_perm_0, x = var_1398)[name = tensor("transpose_8")]; - tensor var_1400 = squeeze(axes = var_1400_axes_0, x = var_1399)[name = tensor("op_1400")]; + tensor var_1430 = reshape(shape = concat_2, x = var_1428)[name = tensor("op_1430")]; + tensor var_1431_axes_0 = const()[name = tensor("op_1431_axes_0"), val = tensor([0])]; + tensor var_1431 = expand_dims(axes = var_1431_axes_0, x = var_1430)[name = tensor("op_1431")]; + tensor var_1432_perm_0 = const()[name = tensor("op_1432_perm_0"), val = tensor([-2, 1, 2, 0, 4])]; + tensor var_1433_axes_0 = const()[name = tensor("op_1433_axes_0"), val = tensor([-2])]; + tensor var_1432 = transpose(perm = var_1432_perm_0, x = var_1431)[name = tensor("transpose_8")]; + tensor var_1433 = squeeze(axes = var_1433_axes_0, x = var_1432)[name = tensor("op_1433")]; tensor q_19_begin_0 = const()[name = tensor("q_19_begin_0"), val = tensor([0, 0, 0, 0])]; tensor q_19_end_0 = const()[name = tensor("q_19_end_0"), val = tensor([1, 12, 4, 256])]; tensor q_19_end_mask_0 = const()[name = tensor("q_19_end_mask_0"), val = tensor([false, true, true, true])]; tensor q_19_squeeze_mask_0 = const()[name = tensor("q_19_squeeze_mask_0"), val = tensor([true, false, false, false])]; - tensor q_19 = slice_by_index(begin = q_19_begin_0, end = q_19_end_0, end_mask = q_19_end_mask_0, squeeze_mask = q_19_squeeze_mask_0, x = var_1400)[name = tensor("q_19")]; + tensor q_19 = slice_by_index(begin = q_19_begin_0, end = q_19_end_0, end_mask = q_19_end_mask_0, squeeze_mask = q_19_squeeze_mask_0, x = var_1433)[name = tensor("q_19")]; tensor k_19_begin_0 = const()[name = tensor("k_19_begin_0"), val = tensor([1, 0, 0, 0])]; tensor k_19_end_0 = const()[name = tensor("k_19_end_0"), val = tensor([2, 12, 4, 256])]; tensor k_19_end_mask_0 = const()[name = tensor("k_19_end_mask_0"), val = tensor([false, true, true, true])]; tensor k_19_squeeze_mask_0 = const()[name = tensor("k_19_squeeze_mask_0"), val = tensor([true, false, false, false])]; - tensor k_19 = slice_by_index(begin = k_19_begin_0, end = k_19_end_0, end_mask = k_19_end_mask_0, squeeze_mask = k_19_squeeze_mask_0, x = var_1400)[name = tensor("k_19")]; + tensor k_19 = slice_by_index(begin = k_19_begin_0, end = k_19_end_0, end_mask = k_19_end_mask_0, squeeze_mask = k_19_squeeze_mask_0, x = var_1433)[name = tensor("k_19")]; tensor v_19_begin_0 = const()[name = tensor("v_19_begin_0"), val = tensor([2, 0, 0, 0])]; tensor v_19_end_0 = const()[name = tensor("v_19_end_0"), val = tensor([3, 12, 4, 256])]; tensor v_19_end_mask_0 = const()[name = tensor("v_19_end_mask_0"), val = tensor([false, true, true, true])]; tensor v_19_squeeze_mask_0 = const()[name = tensor("v_19_squeeze_mask_0"), val = tensor([true, false, false, false])]; - tensor v_19 = slice_by_index(begin = v_19_begin_0, end = v_19_end_0, end_mask = v_19_end_mask_0, squeeze_mask = v_19_squeeze_mask_0, x = var_1400)[name = tensor("v_19")]; - tensor var_1408 = const()[name = tensor("op_1408"), val = tensor([12, 16, 64])]; - tensor var_1409 = reshape(shape = var_1408, x = q_19)[name = tensor("op_1409")]; + tensor v_19 = slice_by_index(begin = v_19_begin_0, end = v_19_end_0, end_mask = v_19_end_mask_0, squeeze_mask = v_19_squeeze_mask_0, x = var_1433)[name = tensor("v_19")]; + tensor var_1441 = const()[name = tensor("op_1441"), val = tensor([12, 16, 64])]; + tensor var_1442 = reshape(shape = var_1441, x = q_19)[name = tensor("op_1442")]; tensor q_21_perm_0 = const()[name = tensor("q_21_perm_0"), val = tensor([1, 0, 2])]; - tensor var_1415 = const()[name = tensor("op_1415"), val = tensor([12, 16, 64])]; - tensor var_1416 = reshape(shape = var_1415, x = k_19)[name = tensor("op_1416")]; + tensor var_1448 = const()[name = tensor("op_1448"), val = tensor([12, 16, 64])]; + tensor var_1449 = reshape(shape = var_1448, x = k_19)[name = tensor("op_1449")]; tensor k_21_perm_0 = const()[name = tensor("k_21_perm_0"), val = tensor([1, 0, 2])]; - tensor var_1422 = const()[name = tensor("op_1422"), val = tensor([12, 16, 64])]; - tensor var_1423 = reshape(shape = var_1422, x = v_19)[name = tensor("op_1423")]; + tensor var_1455 = const()[name = tensor("op_1455"), val = tensor([12, 16, 64])]; + tensor var_1456 = reshape(shape = var_1455, x = v_19)[name = tensor("op_1456")]; tensor v_21_perm_0 = const()[name = tensor("v_21_perm_0"), val = tensor([1, 0, 2])]; - tensor var_1426 = const()[name = tensor("op_1426"), val = tensor([4, 4, 12, 64])]; - tensor q_21 = transpose(perm = q_21_perm_0, x = var_1409)[name = tensor("transpose_7")]; - tensor q = reshape(shape = var_1426, x = q_21)[name = tensor("q")]; - tensor var_1428 = const()[name = tensor("op_1428"), val = tensor([4, 4, 12, 64])]; - tensor k_21 = transpose(perm = k_21_perm_0, x = var_1416)[name = tensor("transpose_6")]; - tensor k = reshape(shape = var_1428, x = k_21)[name = tensor("k")]; - tensor var_1430 = const()[name = tensor("op_1430"), val = tensor([4, 4, 12, 64])]; - tensor v_21 = transpose(perm = v_21_perm_0, x = var_1423)[name = tensor("transpose_5")]; - tensor v = reshape(shape = var_1430, x = v_21)[name = tensor("v")]; + tensor var_1459 = const()[name = tensor("op_1459"), val = tensor([4, 4, 12, 64])]; + tensor q_21 = transpose(perm = q_21_perm_0, x = var_1442)[name = tensor("transpose_7")]; + tensor q = reshape(shape = var_1459, x = q_21)[name = tensor("q")]; + tensor var_1461 = const()[name = tensor("op_1461"), val = tensor([4, 4, 12, 64])]; + tensor k_21 = transpose(perm = k_21_perm_0, x = var_1449)[name = tensor("transpose_6")]; + tensor k = reshape(shape = var_1461, x = k_21)[name = tensor("k")]; + tensor var_1463 = const()[name = tensor("op_1463"), val = tensor([4, 4, 12, 64])]; + tensor v_21 = transpose(perm = v_21_perm_0, x = var_1456)[name = tensor("transpose_5")]; + tensor v = reshape(shape = var_1463, x = v_21)[name = tensor("v")]; tensor mul_3_y_0 = const()[name = tensor("mul_3_y_0"), val = tensor(0x1p-3)]; tensor mul_3 = mul(x = q, y = mul_3_y_0)[name = tensor("mul_3")]; tensor matmul_1_transpose_y_0 = const()[name = tensor("matmul_1_transpose_y_0"), val = tensor(true)]; @@ -1277,36 +1292,36 @@ program(1.0) tensor attn_output_9_transpose_x_0 = const()[name = tensor("attn_output_9_transpose_x_0"), val = tensor(false)]; tensor attn_output_9_transpose_y_0 = const()[name = tensor("attn_output_9_transpose_y_0"), val = tensor(false)]; tensor attn_output_9 = matmul(transpose_x = attn_output_9_transpose_x_0, transpose_y = attn_output_9_transpose_y_0, x = softmax_1, y = v)[name = tensor("attn_output_9")]; - tensor var_1433 = const()[name = tensor("op_1433"), val = tensor([2, 0, 1, 3])]; - tensor var_1438 = const()[name = tensor("op_1438"), val = tensor([48, 256])]; - tensor var_1434 = transpose(perm = var_1433, x = attn_output_9)[name = tensor("transpose_4")]; - tensor attn_output_11 = reshape(shape = var_1438, x = var_1434)[name = tensor("attn_output_11")]; - tensor attn_output_13 = linear(bias = decoder_self_attn2_1_out_proj_bias, weight = decoder_self_attn2_1_out_proj_weight, x = attn_output_11)[name = tensor("linear_53")]; - tensor var_1442 = const()[name = tensor("op_1442"), val = tensor([12, 4, 256])]; - tensor attn_output = reshape(shape = var_1442, x = attn_output_13)[name = tensor("attn_output")]; + tensor var_1466 = const()[name = tensor("op_1466"), val = tensor([2, 0, 1, 3])]; + tensor var_1471 = const()[name = tensor("op_1471"), val = tensor([48, 256])]; + tensor var_1467 = transpose(perm = var_1466, x = attn_output_9)[name = tensor("transpose_4")]; + tensor attn_output_11 = reshape(shape = var_1471, x = var_1467)[name = tensor("attn_output_11")]; + tensor attn_output_13 = linear(bias = inner_decoder_self_attn2_1_out_proj_bias, weight = inner_decoder_self_attn2_1_out_proj_weight, x = attn_output_11)[name = tensor("linear_53")]; + tensor var_1475 = const()[name = tensor("op_1475"), val = tensor([12, 4, 256])]; + tensor attn_output = reshape(shape = var_1475, x = attn_output_13)[name = tensor("attn_output")]; tensor sa2_out_perm_0 = const()[name = tensor("sa2_out_perm_0"), val = tensor([1, 0, 2])]; tensor sa2_out = transpose(perm = sa2_out_perm_0, x = attn_output)[name = tensor("transpose_3")]; - tensor input_193 = add(x = query_5, y = sa2_out)[name = tensor("input_193")]; - tensor input_195_axes_0 = const()[name = tensor("input_195_axes_0"), val = tensor([-1])]; - tensor input_195 = layer_norm(axes = input_195_axes_0, beta = decoder_norm21_1_bias, epsilon = var_1017, gamma = decoder_norm21_1_weight, x = input_193)[name = tensor("input_195")]; - tensor input_197 = linear(bias = decoder_linear1_1_bias, weight = decoder_linear1_1_weight, x = input_195)[name = tensor("linear_54")]; - tensor input_199 = relu(x = input_197)[name = tensor("input_199")]; - tensor ffn_out = linear(bias = decoder_linear2_1_bias, weight = decoder_linear2_1_weight, x = input_199)[name = tensor("linear_55")]; - tensor input_201 = add(x = input_195, y = ffn_out)[name = tensor("input_201")]; + tensor input_195 = add(x = query_5, y = sa2_out)[name = tensor("input_195")]; + tensor input_197_axes_0 = const()[name = tensor("input_197_axes_0"), val = tensor([-1])]; + tensor input_197 = layer_norm(axes = input_197_axes_0, beta = inner_decoder_norm21_1_bias, epsilon = var_66, gamma = inner_decoder_norm21_1_weight, x = input_195)[name = tensor("input_197")]; + tensor input_199 = linear(bias = inner_decoder_linear1_1_bias, weight = inner_decoder_linear1_1_weight, x = input_197)[name = tensor("linear_54")]; + tensor input_201 = relu(x = input_199)[name = tensor("input_201")]; + tensor ffn_out = linear(bias = inner_decoder_linear2_1_bias, weight = inner_decoder_linear2_1_weight, x = input_201)[name = tensor("linear_55")]; + tensor input_203 = add(x = input_197, y = ffn_out)[name = tensor("input_203")]; tensor xt_axes_0 = const()[name = tensor("xt_axes_0"), val = tensor([-1])]; - tensor xt = layer_norm(axes = xt_axes_0, beta = decoder_norm22_1_bias, epsilon = var_1017, gamma = decoder_norm22_1_weight, x = input_201)[name = tensor("xt")]; - tensor var_1462 = const()[name = tensor("op_1462"), val = tensor([1, 4, 12, 256])]; - tensor input = reshape(shape = var_1462, x = xt)[name = tensor("input")]; - tensor var_1464 = const()[name = tensor("op_1464"), val = tensor([-1])]; - tensor var_1465 = reduce_l2_norm(axes = var_1464, keep_dims = var_1020, x = input)[name = tensor("op_1465")]; + tensor xt = layer_norm(axes = xt_axes_0, beta = inner_decoder_norm22_1_bias, epsilon = var_66, gamma = inner_decoder_norm22_1_weight, x = input_203)[name = tensor("xt")]; + tensor var_1495 = const()[name = tensor("op_1495"), val = tensor([1, 4, 12, 256])]; + tensor input = reshape(shape = var_1495, x = xt)[name = tensor("input")]; + tensor var_1497 = const()[name = tensor("op_1497"), val = tensor([-1])]; + tensor var_1498 = reduce_l2_norm(axes = var_1497, keep_dims = var_65, x = input)[name = tensor("op_1498")]; tensor const_42 = const()[name = tensor("const_42"), val = tensor(0x1.fffffep+127)]; - tensor clip_5 = clip(alpha = var_1012, beta = const_42, x = var_1465)[name = tensor("clip_5")]; - tensor var_1467 = real_div(x = input, y = clip_5)[name = tensor("op_1467")]; + tensor clip_5 = clip(alpha = var_79, beta = const_42, x = var_1498)[name = tensor("clip_5")]; + tensor var_1500 = real_div(x = input, y = clip_5)[name = tensor("op_1500")]; tensor concat_6 = const()[name = tensor("concat_6"), val = tensor([4, 1, 256])]; tensor reshape_0 = reshape(shape = concat_6, x = emb)[name = tensor("reshape_0")]; tensor transpose_1_perm_0 = const()[name = tensor("transpose_1_perm_0"), val = tensor([0, 1, 3, 2])]; tensor concat_7 = const()[name = tensor("concat_7"), val = tensor([4, 256, 12])]; - tensor transpose_1 = transpose(perm = transpose_1_perm_0, x = var_1467)[name = tensor("transpose_2")]; + tensor transpose_1 = transpose(perm = transpose_1_perm_0, x = var_1500)[name = tensor("transpose_2")]; tensor reshape_1 = reshape(shape = concat_7, x = transpose_1)[name = tensor("reshape_1")]; tensor matmul_2_transpose_x_0 = const()[name = tensor("matmul_2_transpose_x_0"), val = tensor(false)]; tensor matmul_2_transpose_y_0 = const()[name = tensor("matmul_2_transpose_y_0"), val = tensor(false)]; @@ -1317,10 +1332,10 @@ program(1.0) tensor output_end_0 = const()[name = tensor("output_end_0"), val = tensor([1, 4, 11])]; tensor output_end_mask_0 = const()[name = tensor("output_end_mask_0"), val = tensor([true, true, false])]; tensor output = slice_by_index(begin = output_begin_0, end = output_end_0, end_mask = output_end_mask_0, x = reshape_2)[name = tensor("output")]; - tensor probs = sigmoid(x = output)[name = tensor("op_1471")]; - tensor var_1473_axis_0 = const()[name = tensor("op_1473_axis_0"), val = tensor(0)]; - tensor dec_kv_new = stack(axis = var_1473_axis_0, values = (var_1169, nkv))[name = tensor("op_1473")]; - tensor var_1475_axis_0 = const()[name = tensor("op_1475_axis_0"), val = tensor(0)]; - tensor dec_scale_new = stack(axis = var_1475_axis_0, values = (new_scale_9, new_scale))[name = tensor("op_1475")]; + tensor probs = sigmoid(x = output)[name = tensor("op_1504")]; + tensor var_1506_axis_0 = const()[name = tensor("op_1506_axis_0"), val = tensor(0)]; + tensor dec_kv_new = stack(axis = var_1506_axis_0, values = (var_1202, nkv))[name = tensor("op_1506")]; + tensor var_1508_axis_0 = const()[name = tensor("op_1508_axis_0"), val = tensor(0)]; + tensor dec_scale_new = stack(axis = var_1508_axis_0, values = (new_scale_9, new_scale))[name = tensor("op_1508")]; } -> (probs, enc_kv_new, enc_scale_new, enc_conv_cache_new, cnn_window_new, dec_kv_new, dec_scale_new); } \ No newline at end of file diff --git a/optimized/dih3/400ms/ls_eend_dih3_400ms.mlpackage/Data/com.apple.CoreML/model.mlmodel b/optimized/dih3/400ms/ls_eend_dih3_400ms.mlpackage/Data/com.apple.CoreML/model.mlmodel index 966abc59c27168f5003194361dd833d1665658b1..ea36f705b04c49693e5c48fc1441db62573a3b14 100644 --- a/optimized/dih3/400ms/ls_eend_dih3_400ms.mlpackage/Data/com.apple.CoreML/model.mlmodel +++ b/optimized/dih3/400ms/ls_eend_dih3_400ms.mlpackage/Data/com.apple.CoreML/model.mlmodel @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:000abbd1df11bb0943f0adfab67500eee955b12db8f10bdc9369e7e808b09657 -size 191053 +oid sha256:d7ca94092df462cdab87116ed8e295dc81895635f82bbf9f00a73e4b8816b11d +size 197125 diff --git a/optimized/dih3/400ms/ls_eend_dih3_400ms.mlpackage/Manifest.json b/optimized/dih3/400ms/ls_eend_dih3_400ms.mlpackage/Manifest.json index 7f043626a70d0c5eccc20cae217fcc5767d6d198..d123250cff4561b4464284ef420ac06e692da2ef 100644 --- a/optimized/dih3/400ms/ls_eend_dih3_400ms.mlpackage/Manifest.json +++ b/optimized/dih3/400ms/ls_eend_dih3_400ms.mlpackage/Manifest.json @@ -1,18 +1,18 @@ { "fileFormatVersion": "1.0.0", "itemInfoEntries": { - "7E6A9ABB-631C-4011-875E-660D5DCE8EFC": { - "author": "com.apple.CoreML", - "description": "CoreML Model Specification", - "name": "model.mlmodel", - "path": "com.apple.CoreML/model.mlmodel" - }, - "A7E5EDC0-82DC-425C-A42A-618F7B2317F7": { + "17D6D145-F4A1-42F1-BE9D-64E3EE4EB6F2": { "author": "com.apple.CoreML", "description": "CoreML Model Weights", "name": "weights", "path": "com.apple.CoreML/weights" + }, + "B19871A9-DBD2-40BC-AB61-069157BF4EFD": { + "author": "com.apple.CoreML", + "description": "CoreML Model Specification", + "name": "model.mlmodel", + "path": "com.apple.CoreML/model.mlmodel" } }, - "rootModelIdentifier": "7E6A9ABB-631C-4011-875E-660D5DCE8EFC" + "rootModelIdentifier": "B19871A9-DBD2-40BC-AB61-069157BF4EFD" } diff --git a/optimized/dih3/500ms/ls_eend_dih3_500ms.mlmodelc/analytics/coremldata.bin b/optimized/dih3/500ms/ls_eend_dih3_500ms.mlmodelc/analytics/coremldata.bin index 4d38c0949a58bc566d5e4794910ebd4ae04a7a27..2b2834505010012c170e9a17ae542b2bca78c3c8 100644 --- a/optimized/dih3/500ms/ls_eend_dih3_500ms.mlmodelc/analytics/coremldata.bin +++ b/optimized/dih3/500ms/ls_eend_dih3_500ms.mlmodelc/analytics/coremldata.bin @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:4f8e24abfe7f695ffad28e2adc4c3072705f7bf9cf80102cc782f8e933e72f27 +oid sha256:0d0577390ca35f2d7e42a1e75c3d4275127b262c37750a48fec0ac1d711ab2ff size 243 diff --git a/optimized/dih3/500ms/ls_eend_dih3_500ms.mlmodelc/coremldata.bin b/optimized/dih3/500ms/ls_eend_dih3_500ms.mlmodelc/coremldata.bin index 8915b742fe103cc63e86c1d10e5b105b8647f78d..dd62f2e097c2cc7ef122683ba8079a5bcc35ec5e 100644 --- a/optimized/dih3/500ms/ls_eend_dih3_500ms.mlmodelc/coremldata.bin +++ b/optimized/dih3/500ms/ls_eend_dih3_500ms.mlmodelc/coremldata.bin @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:b9da2e560b5f0570964bbcfe58e5520dfcb2b29af2679608619c52ae68309c02 -size 1310 +oid sha256:3cc2c2502a1cfbc4ccf456695ae9130fb7f0dbfb992c52a2ebd8f20654744133 +size 1413 diff --git a/optimized/dih3/500ms/ls_eend_dih3_500ms.mlmodelc/metadata.json b/optimized/dih3/500ms/ls_eend_dih3_500ms.mlmodelc/metadata.json index bbd1303a91af367bee74f4f22f06f98437f409e1..d982b6126dab58eb13d536236f47070ffa68742e 100644 --- a/optimized/dih3/500ms/ls_eend_dih3_500ms.mlmodelc/metadata.json +++ b/optimized/dih3/500ms/ls_eend_dih3_500ms.mlmodelc/metadata.json @@ -1,7 +1,7 @@ [ { "metadataOutputVersion" : "3.0", - "shortDescription" : "LS-EEND DIHARD III streaming diarizer (pipeline, T=5, max_speakers=10)", + "shortDescription" : "LS-EEND DIHARD III streaming diarizer (pipeline, T=5, max_speakers=10, layout=raw_mel, cu=all)", "outputSchema" : [ { "hasShapeFlexibility" : "0", @@ -81,7 +81,7 @@ "specificationVersion" : 8, "mlProgramOperationTypeHistogram" : { "Ios16.reduceL2Norm" : 2, - "Ios17.reshape" : 68, + "Ios17.reshape" : 69, "Ios16.softmax" : 2, "Ios17.matmul" : 29, "Ios17.transpose" : 57, @@ -89,7 +89,7 @@ "Ios17.expandDims" : 3, "Ios17.add" : 46, "Ios16.sigmoid" : 5, - "Ios17.sliceByIndex" : 72, + "Ios17.sliceByIndex" : 77, "Tile" : 1, "Ios16.reduceSum" : 1, "Ios17.squeeze" : 2, @@ -101,7 +101,7 @@ "Ios16.silu" : 18, "Ios17.realDiv" : 20, "Ios17.linear" : 56, - "Stack" : 5, + "Stack" : 6, "Ios17.concat" : 26, "Ios16.relu" : 2, "Ios16.cumsum" : 1, @@ -128,9 +128,9 @@ "hasShapeFlexibility" : "0", "isOptional" : "0", "dataType" : "Float32", - "formattedType" : "MultiArray (Float32 1 × 5 × 345)", + "formattedType" : "MultiArray (Float32 1 × 55 × 23)", "shortDescription" : "", - "shape" : "[1, 5, 345]", + "shape" : "[1, 55, 23]", "name" : "features", "type" : "MultiArray" }, @@ -206,8 +206,8 @@ } ], "userDefinedMetadata" : { - "com.github.apple.coremltools.conversion_date" : "2026-04-16", - "config" : "{\"model_name\": \"dih3\", \"model_label\": \"DIHARD III\", \"variant\": \"pipeline\", \"chunk_size\": 5, \"step_duration_ms\": 500, \"frame_duration_ms\": 100, \"frame_duration_seconds\": 0.1, \"max_speakers\": 10, \"max_nspks\": 12, \"n_units\": 256, \"n_heads\": 4, \"enc_n_layers\": 4, \"dec_n_layers\": 2, \"conv_kernel_size\": 16, \"conv_delay\": 9, \"sample_rate\": 8000, \"win_length\": 200, \"hop_length\": 80, \"n_mels\": 23, \"context_size\": 7, \"subsampling\": 10, \"feat_type\": \"logmel23_cummn\", \"pure_roll\": true}", + "com.github.apple.coremltools.conversion_date" : "2026-04-18", + "config" : "{\"model_name\": \"dih3\", \"model_label\": \"DIHARD III\", \"variant\": \"pipeline\", \"chunk_size\": 5, \"step_duration_ms\": 500, \"frame_duration_ms\": 100, \"frame_duration_seconds\": 0.1, \"max_speakers\": 10, \"max_nspks\": 12, \"n_units\": 256, \"n_heads\": 4, \"enc_n_layers\": 4, \"dec_n_layers\": 2, \"conv_kernel_size\": 16, \"conv_delay\": 9, \"sample_rate\": 8000, \"win_length\": 200, \"hop_length\": 80, \"n_mels\": 23, \"context_size\": 7, \"subsampling\": 10, \"feat_type\": \"logmel23_cummn\", \"pure_roll\": true, \"input_layout\": \"raw_mel\", \"compute_units_export\": \"all\", \"raw_mel_length\": 55}", "com.github.apple.coremltools.source" : "torch==2.6.0", "com.github.apple.coremltools.version" : "9.0", "com.github.apple.coremltools.source_dialect" : "TorchScript" diff --git a/optimized/dih3/500ms/ls_eend_dih3_500ms.mlmodelc/model.mil b/optimized/dih3/500ms/ls_eend_dih3_500ms.mlmodelc/model.mil index d06d8cb5807c4c0b54c3e07f38ba5655c925422a..ab59896474305668d95443aae5b48ae6a2df960c 100644 --- a/optimized/dih3/500ms/ls_eend_dih3_500ms.mlmodelc/model.mil +++ b/optimized/dih3/500ms/ls_eend_dih3_500ms.mlmodelc/model.mil @@ -1,234 +1,260 @@ program(1.0) [buildInfo = dict, tensor>({{"coremlc-component-MIL", "3510.2.1"}, {"coremlc-version", "3500.32.1"}, {"coremltools-component-torch", "2.6.0"}, {"coremltools-source-dialect", "TorchScript"}, {"coremltools-version", "9.0"}})] { - func main(tensor cnn_window, tensor dec_kv, tensor dec_scale, tensor enc_conv_cache, tensor enc_kv, tensor enc_scale, tensor features, tensor valid_mask) { - tensor encoder_cnn_bias = const()[name = tensor("encoder_cnn_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(64)))]; - tensor encoder_cnn_weight = const()[name = tensor("encoder_cnn_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1152)))]; - tensor encoder__causal_mask = const()[name = tensor("encoder__causal_mask"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4981952)))]; - tensor encoder__t_index = const()[name = tensor("encoder__t_index"), val = tensor([0x1p+0, 0x1p+1, 0x1.8p+1, 0x1p+2, 0x1.4p+2])]; - tensor encoder_input_projection_linear_bias = const()[name = tensor("encoder_input_projection_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4982144)))]; - tensor encoder_input_projection_linear_weight = const()[name = tensor("encoder_input_projection_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4983232)))]; - tensor encoder_pre_layer_norm_bias = const()[name = tensor("encoder_pre_layer_norm_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(5336576)))]; - tensor encoder_pre_layer_norm_weight = const()[name = tensor("encoder_pre_layer_norm_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(5337664)))]; - tensor encoder_ffn1_0_module_sequential_0_bias = const()[name = tensor("encoder_ffn1_0_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(5338752)))]; - tensor encoder_ffn1_0_module_sequential_0_weight = const()[name = tensor("encoder_ffn1_0_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(5339840)))]; - tensor encoder_ffn1_0_module_sequential_1_linear_bias = const()[name = tensor("encoder_ffn1_0_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(5340928)))]; - tensor encoder_ffn1_0_module_sequential_1_linear_weight = const()[name = tensor("encoder_ffn1_0_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(5345088)))]; - tensor encoder_ffn1_0_module_sequential_4_linear_bias = const()[name = tensor("encoder_ffn1_0_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(6393728)))]; - tensor encoder_ffn1_0_module_sequential_4_linear_weight = const()[name = tensor("encoder_ffn1_0_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(6394816)))]; - tensor encoder_ret_lns_0_bias = const()[name = tensor("encoder_ret_lns_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7443456)))]; - tensor encoder_ret_lns_0_weight = const()[name = tensor("encoder_ret_lns_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7444544)))]; - tensor encoder_q_proj_0_bias = const()[name = tensor("encoder_q_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7445632)))]; - tensor encoder_q_proj_0_weight = const()[name = tensor("encoder_q_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7446720)))]; - tensor encoder_k_proj_0_bias = const()[name = tensor("encoder_k_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7708928)))]; - tensor encoder_k_proj_0_weight = const()[name = tensor("encoder_k_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7710016)))]; - tensor encoder_v_proj_0_bias = const()[name = tensor("encoder_v_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7972224)))]; - tensor encoder_v_proj_0_weight = const()[name = tensor("encoder_v_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7973312)))]; - tensor encoder_g_proj_0_bias = const()[name = tensor("encoder_g_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8235520)))]; - tensor encoder_g_proj_0_weight = const()[name = tensor("encoder_g_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8236608)))]; - tensor encoder_out_proj_0_bias = const()[name = tensor("encoder_out_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8498816)))]; - tensor encoder_out_proj_0_weight = const()[name = tensor("encoder_out_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8499904)))]; - tensor encoder_conv_module_0_sequential_0_bias = const()[name = tensor("encoder_conv_module_0_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8762112)))]; - tensor encoder_conv_module_0_sequential_0_weight = const()[name = tensor("encoder_conv_module_0_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8763200)))]; - tensor encoder_conv_module_0_sequential_2_conv_bias = const()[name = tensor("encoder_conv_module_0_sequential_2_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8764288)))]; - tensor encoder_conv_module_0_sequential_2_conv_weight = const()[name = tensor("encoder_conv_module_0_sequential_2_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8766400)))]; - tensor encoder_conv_module_0_sequential_4_conv_weight = const()[name = tensor("encoder_conv_module_0_sequential_4_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9290752)))]; - tensor encoder_conv_module_0_sequential_5_running_var = const()[name = tensor("encoder_conv_module_0_sequential_5_running_var"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9307200)))]; - tensor encoder_conv_module_0_sequential_5_running_mean = const()[name = tensor("encoder_conv_module_0_sequential_5_running_mean"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9308288)))]; - tensor encoder_conv_module_0_sequential_5_bias = const()[name = tensor("encoder_conv_module_0_sequential_5_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9309376)))]; - tensor encoder_conv_module_0_sequential_5_weight = const()[name = tensor("encoder_conv_module_0_sequential_5_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9310464)))]; - tensor encoder_conv_module_0_sequential_7_conv_bias = const()[name = tensor("encoder_conv_module_0_sequential_7_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9311552)))]; - tensor encoder_conv_module_0_sequential_7_conv_weight = const()[name = tensor("encoder_conv_module_0_sequential_7_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9312640)))]; - tensor encoder_ffn2_0_module_sequential_0_bias = const()[name = tensor("encoder_ffn2_0_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9574848)))]; - tensor encoder_ffn2_0_module_sequential_0_weight = const()[name = tensor("encoder_ffn2_0_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9575936)))]; - tensor encoder_ffn2_0_module_sequential_1_linear_bias = const()[name = tensor("encoder_ffn2_0_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9577024)))]; - tensor encoder_ffn2_0_module_sequential_1_linear_weight = const()[name = tensor("encoder_ffn2_0_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9581184)))]; - tensor encoder_ffn2_0_module_sequential_4_linear_bias = const()[name = tensor("encoder_ffn2_0_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(10629824)))]; - tensor encoder_ffn2_0_module_sequential_4_linear_weight = const()[name = tensor("encoder_ffn2_0_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(10630912)))]; - tensor encoder_layer_norm_0_bias = const()[name = tensor("encoder_layer_norm_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11679552)))]; - tensor encoder_layer_norm_0_weight = const()[name = tensor("encoder_layer_norm_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11680640)))]; - tensor encoder_ffn1_1_module_sequential_0_bias = const()[name = tensor("encoder_ffn1_1_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11681728)))]; - tensor encoder_ffn1_1_module_sequential_0_weight = const()[name = tensor("encoder_ffn1_1_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11682816)))]; - tensor encoder_ffn1_1_module_sequential_1_linear_bias = const()[name = tensor("encoder_ffn1_1_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11683904)))]; - tensor encoder_ffn1_1_module_sequential_1_linear_weight = const()[name = tensor("encoder_ffn1_1_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11688064)))]; - tensor encoder_ffn1_1_module_sequential_4_linear_bias = const()[name = tensor("encoder_ffn1_1_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(12736704)))]; - tensor encoder_ffn1_1_module_sequential_4_linear_weight = const()[name = tensor("encoder_ffn1_1_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(12737792)))]; - tensor encoder_ret_lns_1_bias = const()[name = tensor("encoder_ret_lns_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(13786432)))]; - tensor encoder_ret_lns_1_weight = const()[name = tensor("encoder_ret_lns_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(13787520)))]; - tensor encoder_q_proj_1_bias = const()[name = tensor("encoder_q_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(13788608)))]; - tensor encoder_q_proj_1_weight = const()[name = tensor("encoder_q_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(13789696)))]; - tensor encoder_k_proj_1_bias = const()[name = tensor("encoder_k_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14051904)))]; - tensor encoder_k_proj_1_weight = const()[name = tensor("encoder_k_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14052992)))]; - tensor encoder_v_proj_1_bias = const()[name = tensor("encoder_v_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14315200)))]; - tensor encoder_v_proj_1_weight = const()[name = tensor("encoder_v_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14316288)))]; - tensor encoder_g_proj_1_bias = const()[name = tensor("encoder_g_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14578496)))]; - tensor encoder_g_proj_1_weight = const()[name = tensor("encoder_g_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14579584)))]; - tensor encoder_out_proj_1_bias = const()[name = tensor("encoder_out_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14841792)))]; - tensor encoder_out_proj_1_weight = const()[name = tensor("encoder_out_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14842880)))]; - tensor encoder_conv_module_1_sequential_0_bias = const()[name = tensor("encoder_conv_module_1_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15105088)))]; - tensor encoder_conv_module_1_sequential_0_weight = const()[name = tensor("encoder_conv_module_1_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15106176)))]; - tensor encoder_conv_module_1_sequential_2_conv_bias = const()[name = tensor("encoder_conv_module_1_sequential_2_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15107264)))]; - tensor encoder_conv_module_1_sequential_2_conv_weight = const()[name = tensor("encoder_conv_module_1_sequential_2_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15109376)))]; - tensor encoder_conv_module_1_sequential_4_conv_weight = const()[name = tensor("encoder_conv_module_1_sequential_4_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15633728)))]; - tensor encoder_conv_module_1_sequential_5_running_var = const()[name = tensor("encoder_conv_module_1_sequential_5_running_var"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15650176)))]; - tensor encoder_conv_module_1_sequential_5_running_mean = const()[name = tensor("encoder_conv_module_1_sequential_5_running_mean"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15651264)))]; - tensor encoder_conv_module_1_sequential_5_bias = const()[name = tensor("encoder_conv_module_1_sequential_5_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15652352)))]; - tensor encoder_conv_module_1_sequential_5_weight = const()[name = tensor("encoder_conv_module_1_sequential_5_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15653440)))]; - tensor encoder_conv_module_1_sequential_7_conv_bias = const()[name = tensor("encoder_conv_module_1_sequential_7_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15654528)))]; - tensor encoder_conv_module_1_sequential_7_conv_weight = const()[name = tensor("encoder_conv_module_1_sequential_7_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15655616)))]; - tensor encoder_ffn2_1_module_sequential_0_bias = const()[name = tensor("encoder_ffn2_1_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15917824)))]; - tensor encoder_ffn2_1_module_sequential_0_weight = const()[name = tensor("encoder_ffn2_1_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15918912)))]; - tensor encoder_ffn2_1_module_sequential_1_linear_bias = const()[name = tensor("encoder_ffn2_1_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15920000)))]; - tensor encoder_ffn2_1_module_sequential_1_linear_weight = const()[name = tensor("encoder_ffn2_1_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15924160)))]; - tensor encoder_ffn2_1_module_sequential_4_linear_bias = const()[name = tensor("encoder_ffn2_1_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(16972800)))]; - tensor encoder_ffn2_1_module_sequential_4_linear_weight = const()[name = tensor("encoder_ffn2_1_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(16973888)))]; - tensor encoder_layer_norm_1_bias = const()[name = tensor("encoder_layer_norm_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18022528)))]; - tensor encoder_layer_norm_1_weight = const()[name = tensor("encoder_layer_norm_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18023616)))]; - tensor encoder_ffn1_2_module_sequential_0_bias = const()[name = tensor("encoder_ffn1_2_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18024704)))]; - tensor encoder_ffn1_2_module_sequential_0_weight = const()[name = tensor("encoder_ffn1_2_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18025792)))]; - tensor encoder_ffn1_2_module_sequential_1_linear_bias = const()[name = tensor("encoder_ffn1_2_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18026880)))]; - tensor encoder_ffn1_2_module_sequential_1_linear_weight = const()[name = tensor("encoder_ffn1_2_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18031040)))]; - tensor encoder_ffn1_2_module_sequential_4_linear_bias = const()[name = tensor("encoder_ffn1_2_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(19079680)))]; - tensor encoder_ffn1_2_module_sequential_4_linear_weight = const()[name = tensor("encoder_ffn1_2_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(19080768)))]; - tensor encoder_ret_lns_2_bias = const()[name = tensor("encoder_ret_lns_2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20129408)))]; - tensor encoder_ret_lns_2_weight = const()[name = tensor("encoder_ret_lns_2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20130496)))]; - tensor encoder_q_proj_2_bias = const()[name = tensor("encoder_q_proj_2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20131584)))]; - tensor encoder_q_proj_2_weight = const()[name = tensor("encoder_q_proj_2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20132672)))]; - tensor encoder_k_proj_2_bias = const()[name = tensor("encoder_k_proj_2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20394880)))]; - tensor encoder_k_proj_2_weight = const()[name = tensor("encoder_k_proj_2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20395968)))]; - tensor encoder_v_proj_2_bias = const()[name = tensor("encoder_v_proj_2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20658176)))]; - tensor encoder_v_proj_2_weight = const()[name = tensor("encoder_v_proj_2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20659264)))]; - tensor encoder_g_proj_2_bias = const()[name = tensor("encoder_g_proj_2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20921472)))]; - tensor encoder_g_proj_2_weight = const()[name = tensor("encoder_g_proj_2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20922560)))]; - tensor encoder_out_proj_2_bias = const()[name = tensor("encoder_out_proj_2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21184768)))]; - tensor encoder_out_proj_2_weight = const()[name = tensor("encoder_out_proj_2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21185856)))]; - tensor encoder_conv_module_2_sequential_0_bias = const()[name = tensor("encoder_conv_module_2_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21448064)))]; - tensor encoder_conv_module_2_sequential_0_weight = const()[name = tensor("encoder_conv_module_2_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21449152)))]; - tensor encoder_conv_module_2_sequential_2_conv_bias = const()[name = tensor("encoder_conv_module_2_sequential_2_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21450240)))]; - tensor encoder_conv_module_2_sequential_2_conv_weight = const()[name = tensor("encoder_conv_module_2_sequential_2_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21452352)))]; - tensor encoder_conv_module_2_sequential_4_conv_weight = const()[name = tensor("encoder_conv_module_2_sequential_4_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21976704)))]; - tensor encoder_conv_module_2_sequential_5_running_var = const()[name = tensor("encoder_conv_module_2_sequential_5_running_var"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21993152)))]; - tensor encoder_conv_module_2_sequential_5_running_mean = const()[name = tensor("encoder_conv_module_2_sequential_5_running_mean"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21994240)))]; - tensor encoder_conv_module_2_sequential_5_bias = const()[name = tensor("encoder_conv_module_2_sequential_5_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21995328)))]; - tensor encoder_conv_module_2_sequential_5_weight = const()[name = tensor("encoder_conv_module_2_sequential_5_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21996416)))]; - tensor encoder_conv_module_2_sequential_7_conv_bias = const()[name = tensor("encoder_conv_module_2_sequential_7_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21997504)))]; - tensor encoder_conv_module_2_sequential_7_conv_weight = const()[name = tensor("encoder_conv_module_2_sequential_7_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21998592)))]; - tensor encoder_ffn2_2_module_sequential_0_bias = const()[name = tensor("encoder_ffn2_2_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(22260800)))]; - tensor encoder_ffn2_2_module_sequential_0_weight = const()[name = tensor("encoder_ffn2_2_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(22261888)))]; - tensor encoder_ffn2_2_module_sequential_1_linear_bias = const()[name = tensor("encoder_ffn2_2_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(22262976)))]; - tensor encoder_ffn2_2_module_sequential_1_linear_weight = const()[name = tensor("encoder_ffn2_2_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(22267136)))]; - tensor encoder_ffn2_2_module_sequential_4_linear_bias = const()[name = tensor("encoder_ffn2_2_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(23315776)))]; - tensor encoder_ffn2_2_module_sequential_4_linear_weight = const()[name = tensor("encoder_ffn2_2_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(23316864)))]; - tensor encoder_layer_norm_2_bias = const()[name = tensor("encoder_layer_norm_2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(24365504)))]; - tensor encoder_layer_norm_2_weight = const()[name = tensor("encoder_layer_norm_2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(24366592)))]; - tensor encoder_ffn1_3_module_sequential_0_bias = const()[name = tensor("encoder_ffn1_3_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(24367680)))]; - tensor encoder_ffn1_3_module_sequential_0_weight = const()[name = tensor("encoder_ffn1_3_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(24368768)))]; - tensor encoder_ffn1_3_module_sequential_1_linear_bias = const()[name = tensor("encoder_ffn1_3_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(24369856)))]; - tensor encoder_ffn1_3_module_sequential_1_linear_weight = const()[name = tensor("encoder_ffn1_3_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(24374016)))]; - tensor encoder_ffn1_3_module_sequential_4_linear_bias = const()[name = tensor("encoder_ffn1_3_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(25422656)))]; - tensor encoder_ffn1_3_module_sequential_4_linear_weight = const()[name = tensor("encoder_ffn1_3_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(25423744)))]; - tensor encoder_ret_lns_3_bias = const()[name = tensor("encoder_ret_lns_3_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(26472384)))]; - tensor encoder_ret_lns_3_weight = const()[name = tensor("encoder_ret_lns_3_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(26473472)))]; - tensor encoder_q_proj_3_bias = const()[name = tensor("encoder_q_proj_3_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(26474560)))]; - tensor encoder_q_proj_3_weight = const()[name = tensor("encoder_q_proj_3_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(26475648)))]; - tensor encoder_k_proj_3_bias = const()[name = tensor("encoder_k_proj_3_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(26737856)))]; - tensor encoder_k_proj_3_weight = const()[name = tensor("encoder_k_proj_3_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(26738944)))]; - tensor encoder_v_proj_3_bias = const()[name = tensor("encoder_v_proj_3_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27001152)))]; - tensor encoder_v_proj_3_weight = const()[name = tensor("encoder_v_proj_3_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27002240)))]; - tensor encoder_g_proj_3_bias = const()[name = tensor("encoder_g_proj_3_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27264448)))]; - tensor encoder_g_proj_3_weight = const()[name = tensor("encoder_g_proj_3_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27265536)))]; - tensor encoder_out_proj_3_bias = const()[name = tensor("encoder_out_proj_3_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27527744)))]; - tensor encoder_out_proj_3_weight = const()[name = tensor("encoder_out_proj_3_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27528832)))]; - tensor encoder_conv_module_3_sequential_0_bias = const()[name = tensor("encoder_conv_module_3_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27791040)))]; - tensor encoder_conv_module_3_sequential_0_weight = const()[name = tensor("encoder_conv_module_3_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27792128)))]; - tensor encoder_conv_module_3_sequential_2_conv_bias = const()[name = tensor("encoder_conv_module_3_sequential_2_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27793216)))]; - tensor encoder_conv_module_3_sequential_2_conv_weight = const()[name = tensor("encoder_conv_module_3_sequential_2_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27795328)))]; - tensor encoder_conv_module_3_sequential_4_conv_weight = const()[name = tensor("encoder_conv_module_3_sequential_4_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28319680)))]; - tensor encoder_conv_module_3_sequential_5_running_var = const()[name = tensor("encoder_conv_module_3_sequential_5_running_var"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28336128)))]; - tensor encoder_conv_module_3_sequential_5_running_mean = const()[name = tensor("encoder_conv_module_3_sequential_5_running_mean"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28337216)))]; - tensor encoder_conv_module_3_sequential_5_bias = const()[name = tensor("encoder_conv_module_3_sequential_5_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28338304)))]; - tensor encoder_conv_module_3_sequential_5_weight = const()[name = tensor("encoder_conv_module_3_sequential_5_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28339392)))]; - tensor encoder_conv_module_3_sequential_7_conv_bias = const()[name = tensor("encoder_conv_module_3_sequential_7_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28340480)))]; - tensor encoder_conv_module_3_sequential_7_conv_weight = const()[name = tensor("encoder_conv_module_3_sequential_7_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28341568)))]; - tensor encoder_ffn2_3_module_sequential_0_bias = const()[name = tensor("encoder_ffn2_3_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28603776)))]; - tensor encoder_ffn2_3_module_sequential_0_weight = const()[name = tensor("encoder_ffn2_3_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28604864)))]; - tensor encoder_ffn2_3_module_sequential_1_linear_bias = const()[name = tensor("encoder_ffn2_3_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28605952)))]; - tensor encoder_ffn2_3_module_sequential_1_linear_weight = const()[name = tensor("encoder_ffn2_3_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28610112)))]; - tensor encoder_ffn2_3_module_sequential_4_linear_bias = const()[name = tensor("encoder_ffn2_3_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(29658752)))]; - tensor encoder_ffn2_3_module_sequential_4_linear_weight = const()[name = tensor("encoder_ffn2_3_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(29659840)))]; - tensor encoder_layer_norm_3_bias = const()[name = tensor("encoder_layer_norm_3_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(30708480)))]; - tensor encoder_layer_norm_3_weight = const()[name = tensor("encoder_layer_norm_3_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(30709568)))]; - tensor decoder__causal_mask = const()[name = tensor("decoder__causal_mask"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(30710656)))]; - tensor decoder_convert_bias = const()[name = tensor("decoder_convert_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(30710848)))]; - tensor decoder_convert_weight = const()[name = tensor("decoder_convert_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(30711936)))]; - tensor decoder_q_proj_0_bias = const()[name = tensor("decoder_q_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(31236288)))]; - tensor decoder_q_proj_0_weight = const()[name = tensor("decoder_q_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(31237376)))]; - tensor decoder_k_proj_0_bias = const()[name = tensor("decoder_k_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(31499584)))]; - tensor decoder_k_proj_0_weight = const()[name = tensor("decoder_k_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(31500672)))]; - tensor decoder_v_proj_0_bias = const()[name = tensor("decoder_v_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(31762880)))]; - tensor decoder_v_proj_0_weight = const()[name = tensor("decoder_v_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(31763968)))]; - tensor decoder_g_proj_0_bias = const()[name = tensor("decoder_g_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32026176)))]; - tensor decoder_g_proj_0_weight = const()[name = tensor("decoder_g_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32027264)))]; - tensor decoder_out_proj_0_bias = const()[name = tensor("decoder_out_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32289472)))]; - tensor decoder_out_proj_0_weight = const()[name = tensor("decoder_out_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32290560)))]; - tensor decoder_norm11_0_bias = const()[name = tensor("decoder_norm11_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32552768)))]; - tensor decoder_norm11_0_weight = const()[name = tensor("decoder_norm11_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32553856)))]; - tensor decoder_self_attn2_0_out_proj_bias = const()[name = tensor("decoder_self_attn2_0_out_proj_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32554944)))]; - tensor decoder_self_attn2_0_out_proj_weight = const()[name = tensor("decoder_self_attn2_0_out_proj_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32556032)))]; - tensor decoder_self_attn2_0_in_proj_bias = const()[name = tensor("decoder_self_attn2_0_in_proj_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32818240)))]; - tensor decoder_self_attn2_0_in_proj_weight = const()[name = tensor("decoder_self_attn2_0_in_proj_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32821376)))]; - tensor decoder_norm21_0_bias = const()[name = tensor("decoder_norm21_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(33607872)))]; - tensor decoder_norm21_0_weight = const()[name = tensor("decoder_norm21_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(33608960)))]; - tensor decoder_linear1_0_bias = const()[name = tensor("decoder_linear1_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(33610048)))]; - tensor decoder_linear1_0_weight = const()[name = tensor("decoder_linear1_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(33618304)))]; - tensor decoder_linear2_0_bias = const()[name = tensor("decoder_linear2_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(35715520)))]; - tensor decoder_linear2_0_weight = const()[name = tensor("decoder_linear2_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(35716608)))]; - tensor decoder_norm22_0_bias = const()[name = tensor("decoder_norm22_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(37813824)))]; - tensor decoder_norm22_0_weight = const()[name = tensor("decoder_norm22_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(37814912)))]; - tensor decoder_q_proj_1_bias = const()[name = tensor("decoder_q_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(37816000)))]; - tensor decoder_q_proj_1_weight = const()[name = tensor("decoder_q_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(37817088)))]; - tensor decoder_k_proj_1_bias = const()[name = tensor("decoder_k_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38079296)))]; - tensor decoder_k_proj_1_weight = const()[name = tensor("decoder_k_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38080384)))]; - tensor decoder_v_proj_1_bias = const()[name = tensor("decoder_v_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38342592)))]; - tensor decoder_v_proj_1_weight = const()[name = tensor("decoder_v_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38343680)))]; - tensor decoder_g_proj_1_bias = const()[name = tensor("decoder_g_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38605888)))]; - tensor decoder_g_proj_1_weight = const()[name = tensor("decoder_g_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38606976)))]; - tensor decoder_out_proj_1_bias = const()[name = tensor("decoder_out_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38869184)))]; - tensor decoder_out_proj_1_weight = const()[name = tensor("decoder_out_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38870272)))]; - tensor decoder_norm11_1_bias = const()[name = tensor("decoder_norm11_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(39132480)))]; - tensor decoder_norm11_1_weight = const()[name = tensor("decoder_norm11_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(39133568)))]; - tensor decoder_self_attn2_1_out_proj_bias = const()[name = tensor("decoder_self_attn2_1_out_proj_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(39134656)))]; - tensor decoder_self_attn2_1_out_proj_weight = const()[name = tensor("decoder_self_attn2_1_out_proj_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(39135744)))]; - tensor decoder_self_attn2_1_in_proj_bias = const()[name = tensor("decoder_self_attn2_1_in_proj_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(39397952)))]; - tensor decoder_self_attn2_1_in_proj_weight = const()[name = tensor("decoder_self_attn2_1_in_proj_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(39401088)))]; - tensor decoder_norm21_1_bias = const()[name = tensor("decoder_norm21_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(40187584)))]; - tensor decoder_norm21_1_weight = const()[name = tensor("decoder_norm21_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(40188672)))]; - tensor decoder_linear1_1_bias = const()[name = tensor("decoder_linear1_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(40189760)))]; - tensor decoder_linear1_1_weight = const()[name = tensor("decoder_linear1_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(40198016)))]; - tensor decoder_linear2_1_bias = const()[name = tensor("decoder_linear2_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(42295232)))]; - tensor decoder_linear2_1_weight = const()[name = tensor("decoder_linear2_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(42296320)))]; - tensor decoder_norm22_1_bias = const()[name = tensor("decoder_norm22_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(44393536)))]; - tensor decoder_norm22_1_weight = const()[name = tensor("decoder_norm22_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(44394624)))]; - tensor var_11 = const()[name = tensor("op_11"), val = tensor(0x1.5798eep-27)]; - tensor var_18 = const()[name = tensor("op_18"), val = tensor(0x1.0c6f7ap-20)]; - tensor var_21 = const()[name = tensor("op_21"), val = tensor(2)]; - tensor var_24 = const()[name = tensor("op_24"), val = tensor(0)]; - tensor var_27 = const()[name = tensor("op_27"), val = tensor(1)]; - tensor var_29 = const()[name = tensor("op_29"), val = tensor(0x1.4f8b58p-17)]; - tensor var_30 = const()[name = tensor("op_30"), val = tensor(true)]; - tensor input_1 = linear(bias = encoder_input_projection_linear_bias, weight = encoder_input_projection_linear_weight, x = features)[name = tensor("linear_0")]; - tensor input_3_axes_0 = const()[name = tensor("input_3_axes_0"), val = tensor([-1])]; - tensor input_3 = layer_norm(axes = input_3_axes_0, beta = encoder_pre_layer_norm_bias, epsilon = var_29, gamma = encoder_pre_layer_norm_weight, x = input_1)[name = tensor("input_3")]; + func main(tensor cnn_window, tensor dec_kv, tensor dec_scale, tensor enc_conv_cache, tensor enc_kv, tensor enc_scale, tensor features, tensor valid_mask) { + tensor inner_encoder_cnn_bias = const()[name = tensor("inner_encoder_cnn_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(64)))]; + tensor inner_encoder_cnn_weight = const()[name = tensor("inner_encoder_cnn_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1152)))]; + tensor inner_encoder__causal_mask = const()[name = tensor("inner_encoder__causal_mask"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4981952)))]; + tensor inner_encoder__t_index = const()[name = tensor("inner_encoder__t_index"), val = tensor([0x1p+0, 0x1p+1, 0x1.8p+1, 0x1p+2, 0x1.4p+2])]; + tensor inner_encoder_input_projection_linear_bias = const()[name = tensor("inner_encoder_input_projection_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4982144)))]; + tensor inner_encoder_input_projection_linear_weight = const()[name = tensor("inner_encoder_input_projection_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4983232)))]; + tensor inner_encoder_pre_layer_norm_bias = const()[name = tensor("inner_encoder_pre_layer_norm_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(5336576)))]; + tensor inner_encoder_pre_layer_norm_weight = const()[name = tensor("inner_encoder_pre_layer_norm_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(5337664)))]; + tensor inner_encoder_ffn1_0_module_sequential_0_bias = const()[name = tensor("inner_encoder_ffn1_0_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(5338752)))]; + tensor inner_encoder_ffn1_0_module_sequential_0_weight = const()[name = tensor("inner_encoder_ffn1_0_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(5339840)))]; + tensor inner_encoder_ffn1_0_module_sequential_1_linear_bias = const()[name = tensor("inner_encoder_ffn1_0_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(5340928)))]; + tensor inner_encoder_ffn1_0_module_sequential_1_linear_weight = const()[name = tensor("inner_encoder_ffn1_0_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(5345088)))]; + tensor inner_encoder_ffn1_0_module_sequential_4_linear_bias = const()[name = tensor("inner_encoder_ffn1_0_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(6393728)))]; + tensor inner_encoder_ffn1_0_module_sequential_4_linear_weight = const()[name = tensor("inner_encoder_ffn1_0_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(6394816)))]; + tensor inner_encoder_ret_lns_0_bias = const()[name = tensor("inner_encoder_ret_lns_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7443456)))]; + tensor inner_encoder_ret_lns_0_weight = const()[name = tensor("inner_encoder_ret_lns_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7444544)))]; + tensor inner_encoder_q_proj_0_bias = const()[name = tensor("inner_encoder_q_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7445632)))]; + tensor inner_encoder_q_proj_0_weight = const()[name = tensor("inner_encoder_q_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7446720)))]; + tensor inner_encoder_k_proj_0_bias = const()[name = tensor("inner_encoder_k_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7708928)))]; + tensor inner_encoder_k_proj_0_weight = const()[name = tensor("inner_encoder_k_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7710016)))]; + tensor inner_encoder_v_proj_0_bias = const()[name = tensor("inner_encoder_v_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7972224)))]; + tensor inner_encoder_v_proj_0_weight = const()[name = tensor("inner_encoder_v_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7973312)))]; + tensor inner_encoder_g_proj_0_bias = const()[name = tensor("inner_encoder_g_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8235520)))]; + tensor inner_encoder_g_proj_0_weight = const()[name = tensor("inner_encoder_g_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8236608)))]; + tensor inner_encoder_out_proj_0_bias = const()[name = tensor("inner_encoder_out_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8498816)))]; + tensor inner_encoder_out_proj_0_weight = const()[name = tensor("inner_encoder_out_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8499904)))]; + tensor inner_encoder_conv_module_0_sequential_0_bias = const()[name = tensor("inner_encoder_conv_module_0_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8762112)))]; + tensor inner_encoder_conv_module_0_sequential_0_weight = const()[name = tensor("inner_encoder_conv_module_0_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8763200)))]; + tensor inner_encoder_conv_module_0_sequential_2_conv_bias = const()[name = tensor("inner_encoder_conv_module_0_sequential_2_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8764288)))]; + tensor inner_encoder_conv_module_0_sequential_2_conv_weight = const()[name = tensor("inner_encoder_conv_module_0_sequential_2_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8766400)))]; + tensor inner_encoder_conv_module_0_sequential_4_conv_weight = const()[name = tensor("inner_encoder_conv_module_0_sequential_4_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9290752)))]; + tensor inner_encoder_conv_module_0_sequential_5_running_var = const()[name = tensor("inner_encoder_conv_module_0_sequential_5_running_var"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9307200)))]; + tensor inner_encoder_conv_module_0_sequential_5_running_mean = const()[name = tensor("inner_encoder_conv_module_0_sequential_5_running_mean"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9308288)))]; + tensor inner_encoder_conv_module_0_sequential_5_bias = const()[name = tensor("inner_encoder_conv_module_0_sequential_5_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9309376)))]; + tensor inner_encoder_conv_module_0_sequential_5_weight = const()[name = tensor("inner_encoder_conv_module_0_sequential_5_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9310464)))]; + tensor inner_encoder_conv_module_0_sequential_7_conv_bias = const()[name = tensor("inner_encoder_conv_module_0_sequential_7_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9311552)))]; + tensor inner_encoder_conv_module_0_sequential_7_conv_weight = const()[name = tensor("inner_encoder_conv_module_0_sequential_7_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9312640)))]; + tensor inner_encoder_ffn2_0_module_sequential_0_bias = const()[name = tensor("inner_encoder_ffn2_0_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9574848)))]; + tensor inner_encoder_ffn2_0_module_sequential_0_weight = const()[name = tensor("inner_encoder_ffn2_0_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9575936)))]; + tensor inner_encoder_ffn2_0_module_sequential_1_linear_bias = const()[name = tensor("inner_encoder_ffn2_0_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9577024)))]; + tensor inner_encoder_ffn2_0_module_sequential_1_linear_weight = const()[name = tensor("inner_encoder_ffn2_0_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9581184)))]; + tensor inner_encoder_ffn2_0_module_sequential_4_linear_bias = const()[name = tensor("inner_encoder_ffn2_0_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(10629824)))]; + tensor inner_encoder_ffn2_0_module_sequential_4_linear_weight = const()[name = tensor("inner_encoder_ffn2_0_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(10630912)))]; + tensor inner_encoder_layer_norm_0_bias = const()[name = tensor("inner_encoder_layer_norm_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11679552)))]; + tensor inner_encoder_layer_norm_0_weight = const()[name = tensor("inner_encoder_layer_norm_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11680640)))]; + tensor inner_encoder_ffn1_1_module_sequential_0_bias = const()[name = tensor("inner_encoder_ffn1_1_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11681728)))]; + tensor inner_encoder_ffn1_1_module_sequential_0_weight = const()[name = tensor("inner_encoder_ffn1_1_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11682816)))]; + tensor inner_encoder_ffn1_1_module_sequential_1_linear_bias = const()[name = tensor("inner_encoder_ffn1_1_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11683904)))]; + tensor inner_encoder_ffn1_1_module_sequential_1_linear_weight = const()[name = tensor("inner_encoder_ffn1_1_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11688064)))]; + tensor inner_encoder_ffn1_1_module_sequential_4_linear_bias = const()[name = tensor("inner_encoder_ffn1_1_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(12736704)))]; + tensor inner_encoder_ffn1_1_module_sequential_4_linear_weight = const()[name = tensor("inner_encoder_ffn1_1_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(12737792)))]; + tensor inner_encoder_ret_lns_1_bias = const()[name = tensor("inner_encoder_ret_lns_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(13786432)))]; + tensor inner_encoder_ret_lns_1_weight = const()[name = tensor("inner_encoder_ret_lns_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(13787520)))]; + tensor inner_encoder_q_proj_1_bias = const()[name = tensor("inner_encoder_q_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(13788608)))]; + tensor inner_encoder_q_proj_1_weight = const()[name = tensor("inner_encoder_q_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(13789696)))]; + tensor inner_encoder_k_proj_1_bias = const()[name = tensor("inner_encoder_k_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14051904)))]; + tensor inner_encoder_k_proj_1_weight = const()[name = tensor("inner_encoder_k_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14052992)))]; + tensor inner_encoder_v_proj_1_bias = const()[name = tensor("inner_encoder_v_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14315200)))]; + tensor inner_encoder_v_proj_1_weight = const()[name = tensor("inner_encoder_v_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14316288)))]; + tensor inner_encoder_g_proj_1_bias = const()[name = tensor("inner_encoder_g_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14578496)))]; + tensor inner_encoder_g_proj_1_weight = const()[name = tensor("inner_encoder_g_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14579584)))]; + tensor inner_encoder_out_proj_1_bias = const()[name = tensor("inner_encoder_out_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14841792)))]; + tensor inner_encoder_out_proj_1_weight = const()[name = tensor("inner_encoder_out_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14842880)))]; + tensor inner_encoder_conv_module_1_sequential_0_bias = const()[name = tensor("inner_encoder_conv_module_1_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15105088)))]; + tensor inner_encoder_conv_module_1_sequential_0_weight = const()[name = tensor("inner_encoder_conv_module_1_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15106176)))]; + tensor inner_encoder_conv_module_1_sequential_2_conv_bias = const()[name = tensor("inner_encoder_conv_module_1_sequential_2_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15107264)))]; + tensor inner_encoder_conv_module_1_sequential_2_conv_weight = const()[name = tensor("inner_encoder_conv_module_1_sequential_2_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15109376)))]; + tensor inner_encoder_conv_module_1_sequential_4_conv_weight = const()[name = tensor("inner_encoder_conv_module_1_sequential_4_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15633728)))]; + tensor inner_encoder_conv_module_1_sequential_5_running_var = const()[name = tensor("inner_encoder_conv_module_1_sequential_5_running_var"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15650176)))]; + tensor inner_encoder_conv_module_1_sequential_5_running_mean = const()[name = tensor("inner_encoder_conv_module_1_sequential_5_running_mean"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15651264)))]; + tensor inner_encoder_conv_module_1_sequential_5_bias = const()[name = tensor("inner_encoder_conv_module_1_sequential_5_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15652352)))]; + tensor inner_encoder_conv_module_1_sequential_5_weight = const()[name = tensor("inner_encoder_conv_module_1_sequential_5_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15653440)))]; + tensor inner_encoder_conv_module_1_sequential_7_conv_bias = const()[name = tensor("inner_encoder_conv_module_1_sequential_7_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15654528)))]; + tensor inner_encoder_conv_module_1_sequential_7_conv_weight = const()[name = tensor("inner_encoder_conv_module_1_sequential_7_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15655616)))]; + tensor inner_encoder_ffn2_1_module_sequential_0_bias = const()[name = tensor("inner_encoder_ffn2_1_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15917824)))]; + tensor inner_encoder_ffn2_1_module_sequential_0_weight = const()[name = tensor("inner_encoder_ffn2_1_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15918912)))]; + tensor inner_encoder_ffn2_1_module_sequential_1_linear_bias = const()[name = tensor("inner_encoder_ffn2_1_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15920000)))]; + tensor inner_encoder_ffn2_1_module_sequential_1_linear_weight = const()[name = tensor("inner_encoder_ffn2_1_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15924160)))]; + tensor inner_encoder_ffn2_1_module_sequential_4_linear_bias = const()[name = tensor("inner_encoder_ffn2_1_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(16972800)))]; + tensor inner_encoder_ffn2_1_module_sequential_4_linear_weight = const()[name = tensor("inner_encoder_ffn2_1_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(16973888)))]; + tensor inner_encoder_layer_norm_1_bias = const()[name = tensor("inner_encoder_layer_norm_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18022528)))]; + tensor inner_encoder_layer_norm_1_weight = const()[name = tensor("inner_encoder_layer_norm_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18023616)))]; + tensor inner_encoder_ffn1_2_module_sequential_0_bias = const()[name = tensor("inner_encoder_ffn1_2_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18024704)))]; + tensor inner_encoder_ffn1_2_module_sequential_0_weight = const()[name = tensor("inner_encoder_ffn1_2_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18025792)))]; + tensor inner_encoder_ffn1_2_module_sequential_1_linear_bias = const()[name = tensor("inner_encoder_ffn1_2_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18026880)))]; + tensor inner_encoder_ffn1_2_module_sequential_1_linear_weight = const()[name = tensor("inner_encoder_ffn1_2_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18031040)))]; + tensor inner_encoder_ffn1_2_module_sequential_4_linear_bias = const()[name = tensor("inner_encoder_ffn1_2_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(19079680)))]; + tensor inner_encoder_ffn1_2_module_sequential_4_linear_weight = const()[name = tensor("inner_encoder_ffn1_2_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(19080768)))]; + tensor inner_encoder_ret_lns_2_bias = const()[name = tensor("inner_encoder_ret_lns_2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20129408)))]; + tensor inner_encoder_ret_lns_2_weight = const()[name = tensor("inner_encoder_ret_lns_2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20130496)))]; + tensor inner_encoder_q_proj_2_bias = const()[name = tensor("inner_encoder_q_proj_2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20131584)))]; + tensor inner_encoder_q_proj_2_weight = const()[name = tensor("inner_encoder_q_proj_2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20132672)))]; + tensor inner_encoder_k_proj_2_bias = const()[name = tensor("inner_encoder_k_proj_2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20394880)))]; + tensor inner_encoder_k_proj_2_weight = const()[name = tensor("inner_encoder_k_proj_2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20395968)))]; + tensor inner_encoder_v_proj_2_bias = const()[name = tensor("inner_encoder_v_proj_2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20658176)))]; + tensor inner_encoder_v_proj_2_weight = const()[name = tensor("inner_encoder_v_proj_2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20659264)))]; + tensor inner_encoder_g_proj_2_bias = const()[name = tensor("inner_encoder_g_proj_2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20921472)))]; + tensor inner_encoder_g_proj_2_weight = const()[name = tensor("inner_encoder_g_proj_2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(20922560)))]; + tensor inner_encoder_out_proj_2_bias = const()[name = tensor("inner_encoder_out_proj_2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21184768)))]; + tensor inner_encoder_out_proj_2_weight = const()[name = tensor("inner_encoder_out_proj_2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21185856)))]; + tensor inner_encoder_conv_module_2_sequential_0_bias = const()[name = tensor("inner_encoder_conv_module_2_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21448064)))]; + tensor inner_encoder_conv_module_2_sequential_0_weight = const()[name = tensor("inner_encoder_conv_module_2_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21449152)))]; + tensor inner_encoder_conv_module_2_sequential_2_conv_bias = const()[name = tensor("inner_encoder_conv_module_2_sequential_2_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21450240)))]; + tensor inner_encoder_conv_module_2_sequential_2_conv_weight = const()[name = tensor("inner_encoder_conv_module_2_sequential_2_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21452352)))]; + tensor inner_encoder_conv_module_2_sequential_4_conv_weight = const()[name = tensor("inner_encoder_conv_module_2_sequential_4_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21976704)))]; + tensor inner_encoder_conv_module_2_sequential_5_running_var = const()[name = tensor("inner_encoder_conv_module_2_sequential_5_running_var"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21993152)))]; + tensor inner_encoder_conv_module_2_sequential_5_running_mean = const()[name = tensor("inner_encoder_conv_module_2_sequential_5_running_mean"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21994240)))]; + tensor inner_encoder_conv_module_2_sequential_5_bias = const()[name = tensor("inner_encoder_conv_module_2_sequential_5_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21995328)))]; + tensor inner_encoder_conv_module_2_sequential_5_weight = const()[name = tensor("inner_encoder_conv_module_2_sequential_5_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21996416)))]; + tensor inner_encoder_conv_module_2_sequential_7_conv_bias = const()[name = tensor("inner_encoder_conv_module_2_sequential_7_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21997504)))]; + tensor inner_encoder_conv_module_2_sequential_7_conv_weight = const()[name = tensor("inner_encoder_conv_module_2_sequential_7_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21998592)))]; + tensor inner_encoder_ffn2_2_module_sequential_0_bias = const()[name = tensor("inner_encoder_ffn2_2_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(22260800)))]; + tensor inner_encoder_ffn2_2_module_sequential_0_weight = const()[name = tensor("inner_encoder_ffn2_2_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(22261888)))]; + tensor inner_encoder_ffn2_2_module_sequential_1_linear_bias = const()[name = tensor("inner_encoder_ffn2_2_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(22262976)))]; + tensor inner_encoder_ffn2_2_module_sequential_1_linear_weight = const()[name = tensor("inner_encoder_ffn2_2_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(22267136)))]; + tensor inner_encoder_ffn2_2_module_sequential_4_linear_bias = const()[name = tensor("inner_encoder_ffn2_2_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(23315776)))]; + tensor inner_encoder_ffn2_2_module_sequential_4_linear_weight = const()[name = tensor("inner_encoder_ffn2_2_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(23316864)))]; + tensor inner_encoder_layer_norm_2_bias = const()[name = tensor("inner_encoder_layer_norm_2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(24365504)))]; + tensor inner_encoder_layer_norm_2_weight = const()[name = tensor("inner_encoder_layer_norm_2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(24366592)))]; + tensor inner_encoder_ffn1_3_module_sequential_0_bias = const()[name = tensor("inner_encoder_ffn1_3_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(24367680)))]; + tensor inner_encoder_ffn1_3_module_sequential_0_weight = const()[name = tensor("inner_encoder_ffn1_3_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(24368768)))]; + tensor inner_encoder_ffn1_3_module_sequential_1_linear_bias = const()[name = tensor("inner_encoder_ffn1_3_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(24369856)))]; + tensor inner_encoder_ffn1_3_module_sequential_1_linear_weight = const()[name = tensor("inner_encoder_ffn1_3_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(24374016)))]; + tensor inner_encoder_ffn1_3_module_sequential_4_linear_bias = const()[name = tensor("inner_encoder_ffn1_3_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(25422656)))]; + tensor inner_encoder_ffn1_3_module_sequential_4_linear_weight = const()[name = tensor("inner_encoder_ffn1_3_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(25423744)))]; + tensor inner_encoder_ret_lns_3_bias = const()[name = tensor("inner_encoder_ret_lns_3_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(26472384)))]; + tensor inner_encoder_ret_lns_3_weight = const()[name = tensor("inner_encoder_ret_lns_3_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(26473472)))]; + tensor inner_encoder_q_proj_3_bias = const()[name = tensor("inner_encoder_q_proj_3_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(26474560)))]; + tensor inner_encoder_q_proj_3_weight = const()[name = tensor("inner_encoder_q_proj_3_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(26475648)))]; + tensor inner_encoder_k_proj_3_bias = const()[name = tensor("inner_encoder_k_proj_3_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(26737856)))]; + tensor inner_encoder_k_proj_3_weight = const()[name = tensor("inner_encoder_k_proj_3_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(26738944)))]; + tensor inner_encoder_v_proj_3_bias = const()[name = tensor("inner_encoder_v_proj_3_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27001152)))]; + tensor inner_encoder_v_proj_3_weight = const()[name = tensor("inner_encoder_v_proj_3_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27002240)))]; + tensor inner_encoder_g_proj_3_bias = const()[name = tensor("inner_encoder_g_proj_3_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27264448)))]; + tensor inner_encoder_g_proj_3_weight = const()[name = tensor("inner_encoder_g_proj_3_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27265536)))]; + tensor inner_encoder_out_proj_3_bias = const()[name = tensor("inner_encoder_out_proj_3_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27527744)))]; + tensor inner_encoder_out_proj_3_weight = const()[name = tensor("inner_encoder_out_proj_3_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27528832)))]; + tensor inner_encoder_conv_module_3_sequential_0_bias = const()[name = tensor("inner_encoder_conv_module_3_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27791040)))]; + tensor inner_encoder_conv_module_3_sequential_0_weight = const()[name = tensor("inner_encoder_conv_module_3_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27792128)))]; + tensor inner_encoder_conv_module_3_sequential_2_conv_bias = const()[name = tensor("inner_encoder_conv_module_3_sequential_2_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27793216)))]; + tensor inner_encoder_conv_module_3_sequential_2_conv_weight = const()[name = tensor("inner_encoder_conv_module_3_sequential_2_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(27795328)))]; + tensor inner_encoder_conv_module_3_sequential_4_conv_weight = const()[name = tensor("inner_encoder_conv_module_3_sequential_4_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28319680)))]; + tensor inner_encoder_conv_module_3_sequential_5_running_var = const()[name = tensor("inner_encoder_conv_module_3_sequential_5_running_var"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28336128)))]; + tensor inner_encoder_conv_module_3_sequential_5_running_mean = const()[name = tensor("inner_encoder_conv_module_3_sequential_5_running_mean"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28337216)))]; + tensor inner_encoder_conv_module_3_sequential_5_bias = const()[name = tensor("inner_encoder_conv_module_3_sequential_5_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28338304)))]; + tensor inner_encoder_conv_module_3_sequential_5_weight = const()[name = tensor("inner_encoder_conv_module_3_sequential_5_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28339392)))]; + tensor inner_encoder_conv_module_3_sequential_7_conv_bias = const()[name = tensor("inner_encoder_conv_module_3_sequential_7_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28340480)))]; + tensor inner_encoder_conv_module_3_sequential_7_conv_weight = const()[name = tensor("inner_encoder_conv_module_3_sequential_7_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28341568)))]; + tensor inner_encoder_ffn2_3_module_sequential_0_bias = const()[name = tensor("inner_encoder_ffn2_3_module_sequential_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28603776)))]; + tensor inner_encoder_ffn2_3_module_sequential_0_weight = const()[name = tensor("inner_encoder_ffn2_3_module_sequential_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28604864)))]; + tensor inner_encoder_ffn2_3_module_sequential_1_linear_bias = const()[name = tensor("inner_encoder_ffn2_3_module_sequential_1_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28605952)))]; + tensor inner_encoder_ffn2_3_module_sequential_1_linear_weight = const()[name = tensor("inner_encoder_ffn2_3_module_sequential_1_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28610112)))]; + tensor inner_encoder_ffn2_3_module_sequential_4_linear_bias = const()[name = tensor("inner_encoder_ffn2_3_module_sequential_4_linear_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(29658752)))]; + tensor inner_encoder_ffn2_3_module_sequential_4_linear_weight = const()[name = tensor("inner_encoder_ffn2_3_module_sequential_4_linear_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(29659840)))]; + tensor inner_encoder_layer_norm_3_bias = const()[name = tensor("inner_encoder_layer_norm_3_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(30708480)))]; + tensor inner_encoder_layer_norm_3_weight = const()[name = tensor("inner_encoder_layer_norm_3_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(30709568)))]; + tensor inner_decoder__causal_mask = const()[name = tensor("inner_decoder__causal_mask"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(30710656)))]; + tensor inner_decoder_convert_bias = const()[name = tensor("inner_decoder_convert_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(30710848)))]; + tensor inner_decoder_convert_weight = const()[name = tensor("inner_decoder_convert_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(30711936)))]; + tensor inner_decoder_q_proj_0_bias = const()[name = tensor("inner_decoder_q_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(31236288)))]; + tensor inner_decoder_q_proj_0_weight = const()[name = tensor("inner_decoder_q_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(31237376)))]; + tensor inner_decoder_k_proj_0_bias = const()[name = tensor("inner_decoder_k_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(31499584)))]; + tensor inner_decoder_k_proj_0_weight = const()[name = tensor("inner_decoder_k_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(31500672)))]; + tensor inner_decoder_v_proj_0_bias = const()[name = tensor("inner_decoder_v_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(31762880)))]; + tensor inner_decoder_v_proj_0_weight = const()[name = tensor("inner_decoder_v_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(31763968)))]; + tensor inner_decoder_g_proj_0_bias = const()[name = tensor("inner_decoder_g_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32026176)))]; + tensor inner_decoder_g_proj_0_weight = const()[name = tensor("inner_decoder_g_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32027264)))]; + tensor inner_decoder_out_proj_0_bias = const()[name = tensor("inner_decoder_out_proj_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32289472)))]; + tensor inner_decoder_out_proj_0_weight = const()[name = tensor("inner_decoder_out_proj_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32290560)))]; + tensor inner_decoder_norm11_0_bias = const()[name = tensor("inner_decoder_norm11_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32552768)))]; + tensor inner_decoder_norm11_0_weight = const()[name = tensor("inner_decoder_norm11_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32553856)))]; + tensor inner_decoder_self_attn2_0_out_proj_bias = const()[name = tensor("inner_decoder_self_attn2_0_out_proj_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32554944)))]; + tensor inner_decoder_self_attn2_0_out_proj_weight = const()[name = tensor("inner_decoder_self_attn2_0_out_proj_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32556032)))]; + tensor inner_decoder_self_attn2_0_in_proj_bias = const()[name = tensor("inner_decoder_self_attn2_0_in_proj_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32818240)))]; + tensor inner_decoder_self_attn2_0_in_proj_weight = const()[name = tensor("inner_decoder_self_attn2_0_in_proj_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32821376)))]; + tensor inner_decoder_norm21_0_bias = const()[name = tensor("inner_decoder_norm21_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(33607872)))]; + tensor inner_decoder_norm21_0_weight = const()[name = tensor("inner_decoder_norm21_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(33608960)))]; + tensor inner_decoder_linear1_0_bias = const()[name = tensor("inner_decoder_linear1_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(33610048)))]; + tensor inner_decoder_linear1_0_weight = const()[name = tensor("inner_decoder_linear1_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(33618304)))]; + tensor inner_decoder_linear2_0_bias = const()[name = tensor("inner_decoder_linear2_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(35715520)))]; + tensor inner_decoder_linear2_0_weight = const()[name = tensor("inner_decoder_linear2_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(35716608)))]; + tensor inner_decoder_norm22_0_bias = const()[name = tensor("inner_decoder_norm22_0_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(37813824)))]; + tensor inner_decoder_norm22_0_weight = const()[name = tensor("inner_decoder_norm22_0_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(37814912)))]; + tensor inner_decoder_q_proj_1_bias = const()[name = tensor("inner_decoder_q_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(37816000)))]; + tensor inner_decoder_q_proj_1_weight = const()[name = tensor("inner_decoder_q_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(37817088)))]; + tensor inner_decoder_k_proj_1_bias = const()[name = tensor("inner_decoder_k_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38079296)))]; + tensor inner_decoder_k_proj_1_weight = const()[name = tensor("inner_decoder_k_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38080384)))]; + tensor inner_decoder_v_proj_1_bias = const()[name = tensor("inner_decoder_v_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38342592)))]; + tensor inner_decoder_v_proj_1_weight = const()[name = tensor("inner_decoder_v_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38343680)))]; + tensor inner_decoder_g_proj_1_bias = const()[name = tensor("inner_decoder_g_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38605888)))]; + tensor inner_decoder_g_proj_1_weight = const()[name = tensor("inner_decoder_g_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38606976)))]; + tensor inner_decoder_out_proj_1_bias = const()[name = tensor("inner_decoder_out_proj_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38869184)))]; + tensor inner_decoder_out_proj_1_weight = const()[name = tensor("inner_decoder_out_proj_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38870272)))]; + tensor inner_decoder_norm11_1_bias = const()[name = tensor("inner_decoder_norm11_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(39132480)))]; + tensor inner_decoder_norm11_1_weight = const()[name = tensor("inner_decoder_norm11_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(39133568)))]; + tensor inner_decoder_self_attn2_1_out_proj_bias = const()[name = tensor("inner_decoder_self_attn2_1_out_proj_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(39134656)))]; + tensor inner_decoder_self_attn2_1_out_proj_weight = const()[name = tensor("inner_decoder_self_attn2_1_out_proj_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(39135744)))]; + tensor inner_decoder_self_attn2_1_in_proj_bias = const()[name = tensor("inner_decoder_self_attn2_1_in_proj_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(39397952)))]; + tensor inner_decoder_self_attn2_1_in_proj_weight = const()[name = tensor("inner_decoder_self_attn2_1_in_proj_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(39401088)))]; + tensor inner_decoder_norm21_1_bias = const()[name = tensor("inner_decoder_norm21_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(40187584)))]; + tensor inner_decoder_norm21_1_weight = const()[name = tensor("inner_decoder_norm21_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(40188672)))]; + tensor inner_decoder_linear1_1_bias = const()[name = tensor("inner_decoder_linear1_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(40189760)))]; + tensor inner_decoder_linear1_1_weight = const()[name = tensor("inner_decoder_linear1_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(40198016)))]; + tensor inner_decoder_linear2_1_bias = const()[name = tensor("inner_decoder_linear2_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(42295232)))]; + tensor inner_decoder_linear2_1_weight = const()[name = tensor("inner_decoder_linear2_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(42296320)))]; + tensor inner_decoder_norm22_1_bias = const()[name = tensor("inner_decoder_norm22_1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(44393536)))]; + tensor inner_decoder_norm22_1_weight = const()[name = tensor("inner_decoder_norm22_1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(44394624)))]; + tensor var_19_begin_0 = const()[name = tensor("op_19_begin_0"), val = tensor([0, 0, 0])]; + tensor var_19_end_0 = const()[name = tensor("op_19_end_0"), val = tensor([1, 15, 23])]; + tensor var_19_end_mask_0 = const()[name = tensor("op_19_end_mask_0"), val = tensor([true, false, true])]; + tensor var_19 = slice_by_index(begin = var_19_begin_0, end = var_19_end_0, end_mask = var_19_end_mask_0, x = features)[name = tensor("op_19")]; + tensor var_29_begin_0 = const()[name = tensor("op_29_begin_0"), val = tensor([0, 10, 0])]; + tensor var_29_end_0 = const()[name = tensor("op_29_end_0"), val = tensor([1, 25, 23])]; + tensor var_29_end_mask_0 = const()[name = tensor("op_29_end_mask_0"), val = tensor([true, false, true])]; + tensor var_29 = slice_by_index(begin = var_29_begin_0, end = var_29_end_0, end_mask = var_29_end_mask_0, x = features)[name = tensor("op_29")]; + tensor var_39_begin_0 = const()[name = tensor("op_39_begin_0"), val = tensor([0, 20, 0])]; + tensor var_39_end_0 = const()[name = tensor("op_39_end_0"), val = tensor([1, 35, 23])]; + tensor var_39_end_mask_0 = const()[name = tensor("op_39_end_mask_0"), val = tensor([true, false, true])]; + tensor var_39 = slice_by_index(begin = var_39_begin_0, end = var_39_end_0, end_mask = var_39_end_mask_0, x = features)[name = tensor("op_39")]; + tensor var_49_begin_0 = const()[name = tensor("op_49_begin_0"), val = tensor([0, 30, 0])]; + tensor var_49_end_0 = const()[name = tensor("op_49_end_0"), val = tensor([1, 45, 23])]; + tensor var_49_end_mask_0 = const()[name = tensor("op_49_end_mask_0"), val = tensor([true, false, true])]; + tensor var_49 = slice_by_index(begin = var_49_begin_0, end = var_49_end_0, end_mask = var_49_end_mask_0, x = features)[name = tensor("op_49")]; + tensor var_59_begin_0 = const()[name = tensor("op_59_begin_0"), val = tensor([0, 40, 0])]; + tensor var_59_end_0 = const()[name = tensor("op_59_end_0"), val = tensor([1, 1, 23])]; + tensor var_59_end_mask_0 = const()[name = tensor("op_59_end_mask_0"), val = tensor([true, true, true])]; + tensor var_59 = slice_by_index(begin = var_59_begin_0, end = var_59_end_0, end_mask = var_59_end_mask_0, x = features)[name = tensor("op_59")]; + tensor stacked_axis_0 = const()[name = tensor("stacked_axis_0"), val = tensor(1)]; + tensor stacked = stack(axis = stacked_axis_0, values = (var_19, var_29, var_39, var_49, var_59))[name = tensor("stacked")]; + tensor var_66 = const()[name = tensor("op_66"), val = tensor([1, 5, 345])]; + tensor input_1 = reshape(shape = var_66, x = stacked)[name = tensor("input_1")]; + tensor var_69 = const()[name = tensor("op_69"), val = tensor(0x1p+0)]; + tensor var_75 = const()[name = tensor("op_75"), val = tensor(true)]; + tensor var_76 = const()[name = tensor("op_76"), val = tensor(0x1.4f8b58p-17)]; + tensor var_79 = const()[name = tensor("op_79"), val = tensor(0)]; + tensor var_81 = const()[name = tensor("op_81"), val = tensor(2)]; + tensor var_82 = const()[name = tensor("op_82"), val = tensor(-1)]; + tensor var_84 = const()[name = tensor("op_84"), val = tensor(0x1.0c6f7ap-20)]; + tensor var_90 = const()[name = tensor("op_90"), val = tensor(0x1.5798eep-27)]; + tensor var_93 = const()[name = tensor("op_93"), val = tensor(1)]; + tensor input_3 = linear(bias = inner_encoder_input_projection_linear_bias, weight = inner_encoder_input_projection_linear_weight, x = input_1)[name = tensor("linear_0")]; tensor input_5_axes_0 = const()[name = tensor("input_5_axes_0"), val = tensor([-1])]; - tensor input_5 = layer_norm(axes = input_5_axes_0, beta = encoder_ffn1_0_module_sequential_0_bias, epsilon = var_29, gamma = encoder_ffn1_0_module_sequential_0_weight, x = input_3)[name = tensor("input_5")]; - tensor inputs_1 = linear(bias = encoder_ffn1_0_module_sequential_1_linear_bias, weight = encoder_ffn1_0_module_sequential_1_linear_weight, x = input_5)[name = tensor("linear_1")]; - tensor input_7 = silu(x = inputs_1)[name = tensor("input_7")]; - tensor input_11 = linear(bias = encoder_ffn1_0_module_sequential_4_linear_bias, weight = encoder_ffn1_0_module_sequential_4_linear_weight, x = input_7)[name = tensor("linear_2")]; - tensor var_148 = const()[name = tensor("op_148"), val = tensor(0x1p-1)]; - tensor var_149 = mul(x = input_11, y = var_148)[name = tensor("op_149")]; - tensor input_13 = add(x = var_149, y = input_3)[name = tensor("input_13")]; + tensor input_5 = layer_norm(axes = input_5_axes_0, beta = inner_encoder_pre_layer_norm_bias, epsilon = var_76, gamma = inner_encoder_pre_layer_norm_weight, x = input_3)[name = tensor("input_5")]; + tensor input_7_axes_0 = const()[name = tensor("input_7_axes_0"), val = tensor([-1])]; + tensor input_7 = layer_norm(axes = input_7_axes_0, beta = inner_encoder_ffn1_0_module_sequential_0_bias, epsilon = var_76, gamma = inner_encoder_ffn1_0_module_sequential_0_weight, x = input_5)[name = tensor("input_7")]; + tensor inputs_1 = linear(bias = inner_encoder_ffn1_0_module_sequential_1_linear_bias, weight = inner_encoder_ffn1_0_module_sequential_1_linear_weight, x = input_7)[name = tensor("linear_1")]; + tensor input_9 = silu(x = inputs_1)[name = tensor("input_9")]; + tensor input_13 = linear(bias = inner_encoder_ffn1_0_module_sequential_4_linear_bias, weight = inner_encoder_ffn1_0_module_sequential_4_linear_weight, x = input_9)[name = tensor("linear_2")]; + tensor var_214 = const()[name = tensor("op_214"), val = tensor(0x1p-1)]; + tensor var_215 = mul(x = input_13, y = var_214)[name = tensor("op_215")]; + tensor input_15 = add(x = var_215, y = input_5)[name = tensor("input_15")]; tensor x_1_axes_0 = const()[name = tensor("x_1_axes_0"), val = tensor([-1])]; - tensor x_1 = layer_norm(axes = x_1_axes_0, beta = encoder_ret_lns_0_bias, epsilon = var_29, gamma = encoder_ret_lns_0_weight, x = input_13)[name = tensor("x_1")]; + tensor x_1 = layer_norm(axes = x_1_axes_0, beta = inner_encoder_ret_lns_0_bias, epsilon = var_76, gamma = inner_encoder_ret_lns_0_weight, x = input_15)[name = tensor("x_1")]; tensor prev_kv_1_begin_0 = const()[name = tensor("prev_kv_1_begin_0"), val = tensor([0, 0, 0, 0, 0])]; tensor prev_kv_1_end_0 = const()[name = tensor("prev_kv_1_end_0"), val = tensor([1, 1, 4, 64, 64])]; tensor prev_kv_1_end_mask_0 = const()[name = tensor("prev_kv_1_end_mask_0"), val = tensor([false, true, true, true, true])]; @@ -239,183 +265,183 @@ program(1.0) tensor prev_scale_1_end_mask_0 = const()[name = tensor("prev_scale_1_end_mask_0"), val = tensor([false, true])]; tensor prev_scale_1_squeeze_mask_0 = const()[name = tensor("prev_scale_1_squeeze_mask_0"), val = tensor([true, false])]; tensor prev_scale_1 = slice_by_index(begin = prev_scale_1_begin_0, end = prev_scale_1_end_0, end_mask = prev_scale_1_end_mask_0, squeeze_mask = prev_scale_1_squeeze_mask_0, x = enc_scale)[name = tensor("prev_scale_1")]; - tensor var_163 = linear(bias = encoder_q_proj_0_bias, weight = encoder_q_proj_0_weight, x = x_1)[name = tensor("linear_3")]; - tensor var_164 = const()[name = tensor("op_164"), val = tensor([1, 5, 4, 64])]; - tensor var_165 = reshape(shape = var_164, x = var_163)[name = tensor("op_165")]; + tensor var_229 = linear(bias = inner_encoder_q_proj_0_bias, weight = inner_encoder_q_proj_0_weight, x = x_1)[name = tensor("linear_3")]; + tensor var_230 = const()[name = tensor("op_230"), val = tensor([1, 5, 4, 64])]; + tensor var_231 = reshape(shape = var_230, x = var_229)[name = tensor("op_231")]; tensor q_1_perm_0 = const()[name = tensor("q_1_perm_0"), val = tensor([0, 2, 1, 3])]; - tensor var_169 = linear(bias = encoder_k_proj_0_bias, weight = encoder_k_proj_0_weight, x = x_1)[name = tensor("linear_4")]; - tensor var_170 = const()[name = tensor("op_170"), val = tensor(0x1p-3)]; - tensor var_171 = mul(x = var_169, y = var_170)[name = tensor("op_171")]; - tensor var_172 = const()[name = tensor("op_172"), val = tensor([1, 5, 4, 64])]; - tensor var_173 = reshape(shape = var_172, x = var_171)[name = tensor("op_173")]; + tensor var_235 = linear(bias = inner_encoder_k_proj_0_bias, weight = inner_encoder_k_proj_0_weight, x = x_1)[name = tensor("linear_4")]; + tensor var_236 = const()[name = tensor("op_236"), val = tensor(0x1p-3)]; + tensor var_237 = mul(x = var_235, y = var_236)[name = tensor("op_237")]; + tensor var_238 = const()[name = tensor("op_238"), val = tensor([1, 5, 4, 64])]; + tensor var_239 = reshape(shape = var_238, x = var_237)[name = tensor("op_239")]; tensor k_1_perm_0 = const()[name = tensor("k_1_perm_0"), val = tensor([0, 2, 1, 3])]; - tensor var_177 = linear(bias = encoder_v_proj_0_bias, weight = encoder_v_proj_0_weight, x = x_1)[name = tensor("linear_5")]; - tensor var_178 = const()[name = tensor("op_178"), val = tensor([1, 5, 4, 64])]; - tensor var_179 = reshape(shape = var_178, x = var_177)[name = tensor("op_179")]; + tensor var_243 = linear(bias = inner_encoder_v_proj_0_bias, weight = inner_encoder_v_proj_0_weight, x = x_1)[name = tensor("linear_5")]; + tensor var_244 = const()[name = tensor("op_244"), val = tensor([1, 5, 4, 64])]; + tensor var_245 = reshape(shape = var_244, x = var_243)[name = tensor("op_245")]; tensor v_1_perm_0 = const()[name = tensor("v_1_perm_0"), val = tensor([0, 2, 1, 3])]; - tensor input_17 = linear(bias = encoder_g_proj_0_bias, weight = encoder_g_proj_0_weight, x = x_1)[name = tensor("linear_6")]; + tensor input_19 = linear(bias = inner_encoder_g_proj_0_bias, weight = inner_encoder_g_proj_0_weight, x = x_1)[name = tensor("linear_6")]; tensor sqrt_s0_1 = sqrt(x = prev_scale_1)[name = tensor("sqrt_s0_1")]; - tensor s_t_1 = add(x = prev_scale_1, y = encoder__t_index)[name = tensor("s_t_1")]; + tensor s_t_1 = add(x = prev_scale_1, y = inner_encoder__t_index)[name = tensor("s_t_1")]; tensor sqrt_s_t_1 = sqrt(x = s_t_1)[name = tensor("sqrt_s_t_1")]; tensor qk_1_transpose_x_1 = const()[name = tensor("qk_1_transpose_x_1"), val = tensor(false)]; tensor qk_1_transpose_y_1 = const()[name = tensor("qk_1_transpose_y_1"), val = tensor(true)]; - tensor k_1 = transpose(perm = k_1_perm_0, x = var_173)[name = tensor("transpose_57")]; - tensor q_1 = transpose(perm = q_1_perm_0, x = var_165)[name = tensor("transpose_58")]; + tensor k_1 = transpose(perm = k_1_perm_0, x = var_239)[name = tensor("transpose_57")]; + tensor q_1 = transpose(perm = q_1_perm_0, x = var_231)[name = tensor("transpose_58")]; tensor qk_1 = matmul(transpose_x = qk_1_transpose_x_1, transpose_y = qk_1_transpose_y_1, x = q_1, y = k_1)[name = tensor("qk_1")]; - tensor var_189 = const()[name = tensor("op_189"), val = tensor([5, 1])]; - tensor var_190 = reshape(shape = var_189, x = sqrt_s_t_1)[name = tensor("op_190")]; - tensor M_1 = real_div(x = encoder__causal_mask, y = var_190)[name = tensor("M_1")]; - tensor var_192 = mul(x = qk_1, y = M_1)[name = tensor("op_192")]; + tensor var_255 = const()[name = tensor("op_255"), val = tensor([5, 1])]; + tensor var_256 = reshape(shape = var_255, x = sqrt_s_t_1)[name = tensor("op_256")]; + tensor M_1 = real_div(x = inner_encoder__causal_mask, y = var_256)[name = tensor("M_1")]; + tensor var_258 = mul(x = qk_1, y = M_1)[name = tensor("op_258")]; tensor inner_1_transpose_x_0 = const()[name = tensor("inner_1_transpose_x_0"), val = tensor(false)]; tensor inner_1_transpose_y_0 = const()[name = tensor("inner_1_transpose_y_0"), val = tensor(false)]; - tensor v_1 = transpose(perm = v_1_perm_0, x = var_179)[name = tensor("transpose_56")]; - tensor inner_1 = matmul(transpose_x = inner_1_transpose_x_0, transpose_y = inner_1_transpose_y_0, x = var_192, y = v_1)[name = tensor("inner_1")]; - tensor var_194_transpose_x_0 = const()[name = tensor("op_194_transpose_x_0"), val = tensor(false)]; - tensor var_194_transpose_y_0 = const()[name = tensor("op_194_transpose_y_0"), val = tensor(false)]; - tensor var_194 = matmul(transpose_x = var_194_transpose_x_0, transpose_y = var_194_transpose_y_0, x = q_1, y = prev_kv_1)[name = tensor("op_194")]; - tensor var_195 = real_div(x = sqrt_s0_1, y = sqrt_s_t_1)[name = tensor("op_195")]; - tensor var_196 = const()[name = tensor("op_196"), val = tensor([1, 1, 5, 1])]; - tensor var_197 = reshape(shape = var_196, x = var_195)[name = tensor("op_197")]; - tensor cross_1 = mul(x = var_194, y = var_197)[name = tensor("cross_1")]; + tensor v_1 = transpose(perm = v_1_perm_0, x = var_245)[name = tensor("transpose_56")]; + tensor inner_1 = matmul(transpose_x = inner_1_transpose_x_0, transpose_y = inner_1_transpose_y_0, x = var_258, y = v_1)[name = tensor("inner_1")]; + tensor var_260_transpose_x_0 = const()[name = tensor("op_260_transpose_x_0"), val = tensor(false)]; + tensor var_260_transpose_y_0 = const()[name = tensor("op_260_transpose_y_0"), val = tensor(false)]; + tensor var_260 = matmul(transpose_x = var_260_transpose_x_0, transpose_y = var_260_transpose_y_0, x = q_1, y = prev_kv_1)[name = tensor("op_260")]; + tensor var_261 = real_div(x = sqrt_s0_1, y = sqrt_s_t_1)[name = tensor("op_261")]; + tensor var_262 = const()[name = tensor("op_262"), val = tensor([1, 1, 5, 1])]; + tensor var_263 = reshape(shape = var_262, x = var_261)[name = tensor("op_263")]; + tensor cross_1 = mul(x = var_260, y = var_263)[name = tensor("cross_1")]; tensor out_1 = add(x = inner_1, y = cross_1)[name = tensor("out_1")]; - tensor var_200 = mul(x = prev_kv_1, y = sqrt_s0_1)[name = tensor("op_200")]; - tensor var_202_transpose_x_1 = const()[name = tensor("op_202_transpose_x_1"), val = tensor(true)]; - tensor var_202_transpose_y_1 = const()[name = tensor("op_202_transpose_y_1"), val = tensor(false)]; - tensor var_202 = matmul(transpose_x = var_202_transpose_x_1, transpose_y = var_202_transpose_y_1, x = k_1, y = v_1)[name = tensor("op_202")]; - tensor new_kv_unnorm_1 = add(x = var_200, y = var_202)[name = tensor("new_kv_unnorm_1")]; - tensor var_204 = const()[name = tensor("op_204"), val = tensor(0x1.4p+2)]; - tensor new_scale_1 = add(x = prev_scale_1, y = var_204)[name = tensor("new_scale_1")]; - tensor var_206 = sqrt(x = new_scale_1)[name = tensor("op_206")]; - tensor var_207 = real_div(x = new_kv_unnorm_1, y = var_206)[name = tensor("op_207")]; - tensor var_208_perm_0 = const()[name = tensor("op_208_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_266 = mul(x = prev_kv_1, y = sqrt_s0_1)[name = tensor("op_266")]; + tensor var_268_transpose_x_1 = const()[name = tensor("op_268_transpose_x_1"), val = tensor(true)]; + tensor var_268_transpose_y_1 = const()[name = tensor("op_268_transpose_y_1"), val = tensor(false)]; + tensor var_268 = matmul(transpose_x = var_268_transpose_x_1, transpose_y = var_268_transpose_y_1, x = k_1, y = v_1)[name = tensor("op_268")]; + tensor new_kv_unnorm_1 = add(x = var_266, y = var_268)[name = tensor("new_kv_unnorm_1")]; + tensor var_270 = const()[name = tensor("op_270"), val = tensor(0x1.4p+2)]; + tensor new_scale_1 = add(x = prev_scale_1, y = var_270)[name = tensor("new_scale_1")]; + tensor var_272 = sqrt(x = new_scale_1)[name = tensor("op_272")]; + tensor var_273 = real_div(x = new_kv_unnorm_1, y = var_272)[name = tensor("op_273")]; + tensor var_274_perm_0 = const()[name = tensor("op_274_perm_0"), val = tensor([0, 2, 1, 3])]; tensor out_3_axes_0 = const()[name = tensor("out_3_axes_0"), val = tensor([-1])]; - tensor var_208 = transpose(perm = var_208_perm_0, x = out_1)[name = tensor("transpose_55")]; - tensor out_3 = layer_norm(axes = out_3_axes_0, epsilon = var_18, x = var_208)[name = tensor("out_3")]; - tensor var_212 = const()[name = tensor("op_212"), val = tensor([1, 5, 256])]; - tensor out_5 = reshape(shape = var_212, x = out_3)[name = tensor("out_5")]; - tensor var_214 = silu(x = input_17)[name = tensor("op_214")]; - tensor input_19 = mul(x = var_214, y = out_5)[name = tensor("input_19")]; - tensor ret_out_1 = linear(bias = encoder_out_proj_0_bias, weight = encoder_out_proj_0_weight, x = input_19)[name = tensor("linear_7")]; - tensor x_3 = add(x = input_13, y = ret_out_1)[name = tensor("x_3")]; + tensor var_274 = transpose(perm = var_274_perm_0, x = out_1)[name = tensor("transpose_55")]; + tensor out_3 = layer_norm(axes = out_3_axes_0, epsilon = var_84, x = var_274)[name = tensor("out_3")]; + tensor var_278 = const()[name = tensor("op_278"), val = tensor([1, 5, 256])]; + tensor out_5 = reshape(shape = var_278, x = out_3)[name = tensor("out_5")]; + tensor var_280 = silu(x = input_19)[name = tensor("op_280")]; + tensor input_21 = mul(x = var_280, y = out_5)[name = tensor("input_21")]; + tensor ret_out_1 = linear(bias = inner_encoder_out_proj_0_bias, weight = inner_encoder_out_proj_0_weight, x = input_21)[name = tensor("linear_7")]; + tensor x_3 = add(x = input_15, y = ret_out_1)[name = tensor("x_3")]; tensor window_1_begin_0 = const()[name = tensor("window_1_begin_0"), val = tensor([0, 0, 0, 0])]; tensor window_1_end_0 = const()[name = tensor("window_1_end_0"), val = tensor([1, 1, 16, 256])]; tensor window_1_end_mask_0 = const()[name = tensor("window_1_end_mask_0"), val = tensor([false, true, true, true])]; tensor window_1_squeeze_mask_0 = const()[name = tensor("window_1_squeeze_mask_0"), val = tensor([true, false, false, false])]; tensor window_1 = slice_by_index(begin = window_1_begin_0, end = window_1_end_0, end_mask = window_1_end_mask_0, squeeze_mask = window_1_squeeze_mask_0, x = enc_conv_cache)[name = tensor("window_1")]; - tensor var_222_begin_0 = const()[name = tensor("op_222_begin_0"), val = tensor([0, 0, 0])]; - tensor var_222_end_0 = const()[name = tensor("op_222_end_0"), val = tensor([1, 1, 256])]; - tensor var_222_end_mask_0 = const()[name = tensor("op_222_end_mask_0"), val = tensor([true, false, true])]; - tensor var_222 = slice_by_index(begin = var_222_begin_0, end = var_222_end_0, end_mask = var_222_end_mask_0, x = x_3)[name = tensor("op_222")]; - tensor var_225_begin_0 = const()[name = tensor("op_225_begin_0"), val = tensor([0, 1, 0])]; - tensor var_225_end_0 = const()[name = tensor("op_225_end_0"), val = tensor([1, 16, 256])]; - tensor var_225_end_mask_0 = const()[name = tensor("op_225_end_mask_0"), val = tensor([true, true, true])]; - tensor var_225 = slice_by_index(begin = var_225_begin_0, end = var_225_end_0, end_mask = var_225_end_mask_0, x = window_1)[name = tensor("op_225")]; + tensor var_288_begin_0 = const()[name = tensor("op_288_begin_0"), val = tensor([0, 0, 0])]; + tensor var_288_end_0 = const()[name = tensor("op_288_end_0"), val = tensor([1, 1, 256])]; + tensor var_288_end_mask_0 = const()[name = tensor("op_288_end_mask_0"), val = tensor([true, false, true])]; + tensor var_288 = slice_by_index(begin = var_288_begin_0, end = var_288_end_0, end_mask = var_288_end_mask_0, x = x_3)[name = tensor("op_288")]; + tensor var_291_begin_0 = const()[name = tensor("op_291_begin_0"), val = tensor([0, 1, 0])]; + tensor var_291_end_0 = const()[name = tensor("op_291_end_0"), val = tensor([1, 16, 256])]; + tensor var_291_end_mask_0 = const()[name = tensor("op_291_end_mask_0"), val = tensor([true, true, true])]; + tensor var_291 = slice_by_index(begin = var_291_begin_0, end = var_291_end_0, end_mask = var_291_end_mask_0, x = window_1)[name = tensor("op_291")]; tensor window_3_interleave_0 = const()[name = tensor("window_3_interleave_0"), val = tensor(false)]; - tensor window_3 = concat(axis = var_27, interleave = window_3_interleave_0, values = (var_225, var_222))[name = tensor("window_3")]; - tensor var_230_begin_0 = const()[name = tensor("op_230_begin_0"), val = tensor([0, 1, 0])]; - tensor var_230_end_0 = const()[name = tensor("op_230_end_0"), val = tensor([1, 2, 256])]; - tensor var_230_end_mask_0 = const()[name = tensor("op_230_end_mask_0"), val = tensor([true, false, true])]; - tensor var_230 = slice_by_index(begin = var_230_begin_0, end = var_230_end_0, end_mask = var_230_end_mask_0, x = x_3)[name = tensor("op_230")]; - tensor var_233_begin_0 = const()[name = tensor("op_233_begin_0"), val = tensor([0, 1, 0])]; - tensor var_233_end_0 = const()[name = tensor("op_233_end_0"), val = tensor([1, 16, 256])]; - tensor var_233_end_mask_0 = const()[name = tensor("op_233_end_mask_0"), val = tensor([true, true, true])]; - tensor var_233 = slice_by_index(begin = var_233_begin_0, end = var_233_end_0, end_mask = var_233_end_mask_0, x = window_3)[name = tensor("op_233")]; + tensor window_3 = concat(axis = var_93, interleave = window_3_interleave_0, values = (var_291, var_288))[name = tensor("window_3")]; + tensor var_296_begin_0 = const()[name = tensor("op_296_begin_0"), val = tensor([0, 1, 0])]; + tensor var_296_end_0 = const()[name = tensor("op_296_end_0"), val = tensor([1, 2, 256])]; + tensor var_296_end_mask_0 = const()[name = tensor("op_296_end_mask_0"), val = tensor([true, false, true])]; + tensor var_296 = slice_by_index(begin = var_296_begin_0, end = var_296_end_0, end_mask = var_296_end_mask_0, x = x_3)[name = tensor("op_296")]; + tensor var_299_begin_0 = const()[name = tensor("op_299_begin_0"), val = tensor([0, 1, 0])]; + tensor var_299_end_0 = const()[name = tensor("op_299_end_0"), val = tensor([1, 16, 256])]; + tensor var_299_end_mask_0 = const()[name = tensor("op_299_end_mask_0"), val = tensor([true, true, true])]; + tensor var_299 = slice_by_index(begin = var_299_begin_0, end = var_299_end_0, end_mask = var_299_end_mask_0, x = window_3)[name = tensor("op_299")]; tensor window_5_interleave_0 = const()[name = tensor("window_5_interleave_0"), val = tensor(false)]; - tensor window_5 = concat(axis = var_27, interleave = window_5_interleave_0, values = (var_233, var_230))[name = tensor("window_5")]; - tensor var_238_begin_0 = const()[name = tensor("op_238_begin_0"), val = tensor([0, 2, 0])]; - tensor var_238_end_0 = const()[name = tensor("op_238_end_0"), val = tensor([1, 3, 256])]; - tensor var_238_end_mask_0 = const()[name = tensor("op_238_end_mask_0"), val = tensor([true, false, true])]; - tensor var_238 = slice_by_index(begin = var_238_begin_0, end = var_238_end_0, end_mask = var_238_end_mask_0, x = x_3)[name = tensor("op_238")]; - tensor var_241_begin_0 = const()[name = tensor("op_241_begin_0"), val = tensor([0, 1, 0])]; - tensor var_241_end_0 = const()[name = tensor("op_241_end_0"), val = tensor([1, 16, 256])]; - tensor var_241_end_mask_0 = const()[name = tensor("op_241_end_mask_0"), val = tensor([true, true, true])]; - tensor var_241 = slice_by_index(begin = var_241_begin_0, end = var_241_end_0, end_mask = var_241_end_mask_0, x = window_5)[name = tensor("op_241")]; + tensor window_5 = concat(axis = var_93, interleave = window_5_interleave_0, values = (var_299, var_296))[name = tensor("window_5")]; + tensor var_304_begin_0 = const()[name = tensor("op_304_begin_0"), val = tensor([0, 2, 0])]; + tensor var_304_end_0 = const()[name = tensor("op_304_end_0"), val = tensor([1, 3, 256])]; + tensor var_304_end_mask_0 = const()[name = tensor("op_304_end_mask_0"), val = tensor([true, false, true])]; + tensor var_304 = slice_by_index(begin = var_304_begin_0, end = var_304_end_0, end_mask = var_304_end_mask_0, x = x_3)[name = tensor("op_304")]; + tensor var_307_begin_0 = const()[name = tensor("op_307_begin_0"), val = tensor([0, 1, 0])]; + tensor var_307_end_0 = const()[name = tensor("op_307_end_0"), val = tensor([1, 16, 256])]; + tensor var_307_end_mask_0 = const()[name = tensor("op_307_end_mask_0"), val = tensor([true, true, true])]; + tensor var_307 = slice_by_index(begin = var_307_begin_0, end = var_307_end_0, end_mask = var_307_end_mask_0, x = window_5)[name = tensor("op_307")]; tensor window_7_interleave_0 = const()[name = tensor("window_7_interleave_0"), val = tensor(false)]; - tensor window_7 = concat(axis = var_27, interleave = window_7_interleave_0, values = (var_241, var_238))[name = tensor("window_7")]; - tensor var_246_begin_0 = const()[name = tensor("op_246_begin_0"), val = tensor([0, 3, 0])]; - tensor var_246_end_0 = const()[name = tensor("op_246_end_0"), val = tensor([1, 4, 256])]; - tensor var_246_end_mask_0 = const()[name = tensor("op_246_end_mask_0"), val = tensor([true, false, true])]; - tensor var_246 = slice_by_index(begin = var_246_begin_0, end = var_246_end_0, end_mask = var_246_end_mask_0, x = x_3)[name = tensor("op_246")]; - tensor var_249_begin_0 = const()[name = tensor("op_249_begin_0"), val = tensor([0, 1, 0])]; - tensor var_249_end_0 = const()[name = tensor("op_249_end_0"), val = tensor([1, 16, 256])]; - tensor var_249_end_mask_0 = const()[name = tensor("op_249_end_mask_0"), val = tensor([true, true, true])]; - tensor var_249 = slice_by_index(begin = var_249_begin_0, end = var_249_end_0, end_mask = var_249_end_mask_0, x = window_7)[name = tensor("op_249")]; + tensor window_7 = concat(axis = var_93, interleave = window_7_interleave_0, values = (var_307, var_304))[name = tensor("window_7")]; + tensor var_312_begin_0 = const()[name = tensor("op_312_begin_0"), val = tensor([0, 3, 0])]; + tensor var_312_end_0 = const()[name = tensor("op_312_end_0"), val = tensor([1, 4, 256])]; + tensor var_312_end_mask_0 = const()[name = tensor("op_312_end_mask_0"), val = tensor([true, false, true])]; + tensor var_312 = slice_by_index(begin = var_312_begin_0, end = var_312_end_0, end_mask = var_312_end_mask_0, x = x_3)[name = tensor("op_312")]; + tensor var_315_begin_0 = const()[name = tensor("op_315_begin_0"), val = tensor([0, 1, 0])]; + tensor var_315_end_0 = const()[name = tensor("op_315_end_0"), val = tensor([1, 16, 256])]; + tensor var_315_end_mask_0 = const()[name = tensor("op_315_end_mask_0"), val = tensor([true, true, true])]; + tensor var_315 = slice_by_index(begin = var_315_begin_0, end = var_315_end_0, end_mask = var_315_end_mask_0, x = window_7)[name = tensor("op_315")]; tensor window_9_interleave_0 = const()[name = tensor("window_9_interleave_0"), val = tensor(false)]; - tensor window_9 = concat(axis = var_27, interleave = window_9_interleave_0, values = (var_249, var_246))[name = tensor("window_9")]; - tensor var_254_begin_0 = const()[name = tensor("op_254_begin_0"), val = tensor([0, 4, 0])]; - tensor var_254_end_0 = const()[name = tensor("op_254_end_0"), val = tensor([1, 1, 256])]; - tensor var_254_end_mask_0 = const()[name = tensor("op_254_end_mask_0"), val = tensor([true, true, true])]; - tensor var_254 = slice_by_index(begin = var_254_begin_0, end = var_254_end_0, end_mask = var_254_end_mask_0, x = x_3)[name = tensor("op_254")]; - tensor var_257_begin_0 = const()[name = tensor("op_257_begin_0"), val = tensor([0, 1, 0])]; - tensor var_257_end_0 = const()[name = tensor("op_257_end_0"), val = tensor([1, 16, 256])]; - tensor var_257_end_mask_0 = const()[name = tensor("op_257_end_mask_0"), val = tensor([true, true, true])]; - tensor var_257 = slice_by_index(begin = var_257_begin_0, end = var_257_end_0, end_mask = var_257_end_mask_0, x = window_9)[name = tensor("op_257")]; + tensor window_9 = concat(axis = var_93, interleave = window_9_interleave_0, values = (var_315, var_312))[name = tensor("window_9")]; + tensor var_320_begin_0 = const()[name = tensor("op_320_begin_0"), val = tensor([0, 4, 0])]; + tensor var_320_end_0 = const()[name = tensor("op_320_end_0"), val = tensor([1, 1, 256])]; + tensor var_320_end_mask_0 = const()[name = tensor("op_320_end_mask_0"), val = tensor([true, true, true])]; + tensor var_320 = slice_by_index(begin = var_320_begin_0, end = var_320_end_0, end_mask = var_320_end_mask_0, x = x_3)[name = tensor("op_320")]; + tensor var_323_begin_0 = const()[name = tensor("op_323_begin_0"), val = tensor([0, 1, 0])]; + tensor var_323_end_0 = const()[name = tensor("op_323_end_0"), val = tensor([1, 16, 256])]; + tensor var_323_end_mask_0 = const()[name = tensor("op_323_end_mask_0"), val = tensor([true, true, true])]; + tensor var_323 = slice_by_index(begin = var_323_begin_0, end = var_323_end_0, end_mask = var_323_end_mask_0, x = window_9)[name = tensor("op_323")]; tensor window_11_interleave_0 = const()[name = tensor("window_11_interleave_0"), val = tensor(false)]; - tensor window_11 = concat(axis = var_27, interleave = window_11_interleave_0, values = (var_257, var_254))[name = tensor("window_11")]; - tensor input_21_interleave_0 = const()[name = tensor("input_21_interleave_0"), val = tensor(false)]; - tensor input_21 = concat(axis = var_24, interleave = input_21_interleave_0, values = (window_3, window_5, window_7, window_9, window_11))[name = tensor("input_21")]; + tensor window_11 = concat(axis = var_93, interleave = window_11_interleave_0, values = (var_323, var_320))[name = tensor("window_11")]; + tensor input_23_interleave_0 = const()[name = tensor("input_23_interleave_0"), val = tensor(false)]; + tensor input_23 = concat(axis = var_79, interleave = input_23_interleave_0, values = (window_3, window_5, window_7, window_9, window_11))[name = tensor("input_23")]; tensor x_5_axes_0 = const()[name = tensor("x_5_axes_0"), val = tensor([-1])]; - tensor x_5 = layer_norm(axes = x_5_axes_0, beta = encoder_conv_module_0_sequential_0_bias, epsilon = var_29, gamma = encoder_conv_module_0_sequential_0_weight, x = input_21)[name = tensor("x_5")]; - tensor input_23_perm_0 = const()[name = tensor("input_23_perm_0"), val = tensor([0, 2, 1])]; + tensor x_5 = layer_norm(axes = x_5_axes_0, beta = inner_encoder_conv_module_0_sequential_0_bias, epsilon = var_76, gamma = inner_encoder_conv_module_0_sequential_0_weight, x = input_23)[name = tensor("x_5")]; + tensor input_25_perm_0 = const()[name = tensor("input_25_perm_0"), val = tensor([0, 2, 1])]; tensor inputs_3_pad_type_0 = const()[name = tensor("inputs_3_pad_type_0"), val = tensor("valid")]; tensor inputs_3_strides_0 = const()[name = tensor("inputs_3_strides_0"), val = tensor([1])]; tensor inputs_3_pad_0 = const()[name = tensor("inputs_3_pad_0"), val = tensor([0, 0])]; tensor inputs_3_dilations_0 = const()[name = tensor("inputs_3_dilations_0"), val = tensor([1])]; tensor inputs_3_groups_0 = const()[name = tensor("inputs_3_groups_0"), val = tensor(1)]; - tensor input_23 = transpose(perm = input_23_perm_0, x = x_5)[name = tensor("transpose_54")]; - tensor inputs_3 = conv(bias = encoder_conv_module_0_sequential_2_conv_bias, dilations = inputs_3_dilations_0, groups = inputs_3_groups_0, pad = inputs_3_pad_0, pad_type = inputs_3_pad_type_0, strides = inputs_3_strides_0, weight = encoder_conv_module_0_sequential_2_conv_weight, x = input_23)[name = tensor("inputs_3")]; - tensor var_282_split_sizes_0 = const()[name = tensor("op_282_split_sizes_0"), val = tensor([256, 256])]; - tensor var_282_axis_0 = const()[name = tensor("op_282_axis_0"), val = tensor(1)]; - tensor var_282_0, tensor var_282_1 = split(axis = var_282_axis_0, split_sizes = var_282_split_sizes_0, x = inputs_3)[name = tensor("op_282")]; - tensor var_284 = sigmoid(x = var_282_1)[name = tensor("op_284")]; - tensor inputs_5 = mul(x = var_282_0, y = var_284)[name = tensor("inputs_5")]; + tensor input_25 = transpose(perm = input_25_perm_0, x = x_5)[name = tensor("transpose_54")]; + tensor inputs_3 = conv(bias = inner_encoder_conv_module_0_sequential_2_conv_bias, dilations = inputs_3_dilations_0, groups = inputs_3_groups_0, pad = inputs_3_pad_0, pad_type = inputs_3_pad_type_0, strides = inputs_3_strides_0, weight = inner_encoder_conv_module_0_sequential_2_conv_weight, x = input_25)[name = tensor("inputs_3")]; + tensor var_348_split_sizes_0 = const()[name = tensor("op_348_split_sizes_0"), val = tensor([256, 256])]; + tensor var_348_axis_0 = const()[name = tensor("op_348_axis_0"), val = tensor(1)]; + tensor var_348_0, tensor var_348_1 = split(axis = var_348_axis_0, split_sizes = var_348_split_sizes_0, x = inputs_3)[name = tensor("op_348")]; + tensor var_350 = sigmoid(x = var_348_1)[name = tensor("op_350")]; + tensor inputs_5 = mul(x = var_348_0, y = var_350)[name = tensor("inputs_5")]; tensor outputs_aug_1_pad_type_0 = const()[name = tensor("outputs_aug_1_pad_type_0"), val = tensor("custom")]; tensor outputs_aug_1_pad_0 = const()[name = tensor("outputs_aug_1_pad_0"), val = tensor([15, 15])]; tensor outputs_aug_1_groups_0 = const()[name = tensor("outputs_aug_1_groups_0"), val = tensor(256)]; tensor outputs_aug_1_strides_0 = const()[name = tensor("outputs_aug_1_strides_0"), val = tensor([1])]; tensor outputs_aug_1_dilations_0 = const()[name = tensor("outputs_aug_1_dilations_0"), val = tensor([1])]; - tensor outputs_aug_1 = conv(dilations = outputs_aug_1_dilations_0, groups = outputs_aug_1_groups_0, pad = outputs_aug_1_pad_0, pad_type = outputs_aug_1_pad_type_0, strides = outputs_aug_1_strides_0, weight = encoder_conv_module_0_sequential_4_conv_weight, x = inputs_5)[name = tensor("outputs_aug_1")]; - tensor input_25_begin_0 = const()[name = tensor("input_25_begin_0"), val = tensor([0, 0, 0])]; - tensor input_25_end_0 = const()[name = tensor("input_25_end_0"), val = tensor([5, 256, 16])]; - tensor input_25_end_mask_0 = const()[name = tensor("input_25_end_mask_0"), val = tensor([true, true, false])]; - tensor input_25 = slice_by_index(begin = input_25_begin_0, end = input_25_end_0, end_mask = input_25_end_mask_0, x = outputs_aug_1)[name = tensor("input_25")]; - tensor inputs_7 = batch_norm(beta = encoder_conv_module_0_sequential_5_bias, epsilon = var_29, gamma = encoder_conv_module_0_sequential_5_weight, mean = encoder_conv_module_0_sequential_5_running_mean, variance = encoder_conv_module_0_sequential_5_running_var, x = input_25)[name = tensor("inputs_7")]; - tensor input_27 = silu(x = inputs_7)[name = tensor("input_27")]; - tensor input_29_pad_type_0 = const()[name = tensor("input_29_pad_type_0"), val = tensor("valid")]; - tensor input_29_strides_0 = const()[name = tensor("input_29_strides_0"), val = tensor([1])]; - tensor input_29_pad_0 = const()[name = tensor("input_29_pad_0"), val = tensor([0, 0])]; - tensor input_29_dilations_0 = const()[name = tensor("input_29_dilations_0"), val = tensor([1])]; - tensor input_29_groups_0 = const()[name = tensor("input_29_groups_0"), val = tensor(1)]; - tensor input_29 = conv(bias = encoder_conv_module_0_sequential_7_conv_bias, dilations = input_29_dilations_0, groups = input_29_groups_0, pad = input_29_pad_0, pad_type = input_29_pad_type_0, strides = input_29_strides_0, weight = encoder_conv_module_0_sequential_7_conv_weight, x = input_27)[name = tensor("input_29")]; + tensor outputs_aug_1 = conv(dilations = outputs_aug_1_dilations_0, groups = outputs_aug_1_groups_0, pad = outputs_aug_1_pad_0, pad_type = outputs_aug_1_pad_type_0, strides = outputs_aug_1_strides_0, weight = inner_encoder_conv_module_0_sequential_4_conv_weight, x = inputs_5)[name = tensor("outputs_aug_1")]; + tensor input_27_begin_0 = const()[name = tensor("input_27_begin_0"), val = tensor([0, 0, 0])]; + tensor input_27_end_0 = const()[name = tensor("input_27_end_0"), val = tensor([5, 256, 16])]; + tensor input_27_end_mask_0 = const()[name = tensor("input_27_end_mask_0"), val = tensor([true, true, false])]; + tensor input_27 = slice_by_index(begin = input_27_begin_0, end = input_27_end_0, end_mask = input_27_end_mask_0, x = outputs_aug_1)[name = tensor("input_27")]; + tensor inputs_7 = batch_norm(beta = inner_encoder_conv_module_0_sequential_5_bias, epsilon = var_76, gamma = inner_encoder_conv_module_0_sequential_5_weight, mean = inner_encoder_conv_module_0_sequential_5_running_mean, variance = inner_encoder_conv_module_0_sequential_5_running_var, x = input_27)[name = tensor("inputs_7")]; + tensor input_29 = silu(x = inputs_7)[name = tensor("input_29")]; + tensor input_31_pad_type_0 = const()[name = tensor("input_31_pad_type_0"), val = tensor("valid")]; + tensor input_31_strides_0 = const()[name = tensor("input_31_strides_0"), val = tensor([1])]; + tensor input_31_pad_0 = const()[name = tensor("input_31_pad_0"), val = tensor([0, 0])]; + tensor input_31_dilations_0 = const()[name = tensor("input_31_dilations_0"), val = tensor([1])]; + tensor input_31_groups_0 = const()[name = tensor("input_31_groups_0"), val = tensor(1)]; + tensor input_31 = conv(bias = inner_encoder_conv_module_0_sequential_7_conv_bias, dilations = input_31_dilations_0, groups = input_31_groups_0, pad = input_31_pad_0, pad_type = input_31_pad_type_0, strides = input_31_strides_0, weight = inner_encoder_conv_module_0_sequential_7_conv_weight, x = input_29)[name = tensor("input_31")]; tensor conv_out_1_perm_0 = const()[name = tensor("conv_out_1_perm_0"), val = tensor([0, 2, 1])]; - tensor var_315_begin_0 = const()[name = tensor("op_315_begin_0"), val = tensor([0, -1, 0])]; - tensor var_315_end_0 = const()[name = tensor("op_315_end_0"), val = tensor([5, 16, 256])]; - tensor var_315_end_mask_0 = const()[name = tensor("op_315_end_mask_0"), val = tensor([true, true, true])]; - tensor conv_out_1 = transpose(perm = conv_out_1_perm_0, x = input_29)[name = tensor("transpose_53")]; - tensor var_315 = slice_by_index(begin = var_315_begin_0, end = var_315_end_0, end_mask = var_315_end_mask_0, x = conv_out_1)[name = tensor("op_315")]; - tensor var_317_perm_0 = const()[name = tensor("op_317_perm_0"), val = tensor([1, 0, 2])]; - tensor var_317 = transpose(perm = var_317_perm_0, x = var_315)[name = tensor("transpose_52")]; - tensor input_31 = add(x = x_3, y = var_317)[name = tensor("input_31")]; - tensor input_33_axes_0 = const()[name = tensor("input_33_axes_0"), val = tensor([-1])]; - tensor input_33 = layer_norm(axes = input_33_axes_0, beta = encoder_ffn2_0_module_sequential_0_bias, epsilon = var_29, gamma = encoder_ffn2_0_module_sequential_0_weight, x = input_31)[name = tensor("input_33")]; - tensor inputs_9 = linear(bias = encoder_ffn2_0_module_sequential_1_linear_bias, weight = encoder_ffn2_0_module_sequential_1_linear_weight, x = input_33)[name = tensor("linear_8")]; - tensor input_35 = silu(x = inputs_9)[name = tensor("input_35")]; - tensor input_39 = linear(bias = encoder_ffn2_0_module_sequential_4_linear_bias, weight = encoder_ffn2_0_module_sequential_4_linear_weight, x = input_35)[name = tensor("linear_9")]; - tensor var_340 = const()[name = tensor("op_340"), val = tensor(0x1p-1)]; - tensor var_341 = mul(x = input_39, y = var_340)[name = tensor("op_341")]; - tensor input_41 = add(x = var_341, y = input_31)[name = tensor("input_41")]; - tensor input_43_axes_0 = const()[name = tensor("input_43_axes_0"), val = tensor([-1])]; - tensor input_43 = layer_norm(axes = input_43_axes_0, beta = encoder_layer_norm_0_bias, epsilon = var_29, gamma = encoder_layer_norm_0_weight, x = input_41)[name = tensor("input_43")]; + tensor var_381_begin_0 = const()[name = tensor("op_381_begin_0"), val = tensor([0, -1, 0])]; + tensor var_381_end_0 = const()[name = tensor("op_381_end_0"), val = tensor([5, 16, 256])]; + tensor var_381_end_mask_0 = const()[name = tensor("op_381_end_mask_0"), val = tensor([true, true, true])]; + tensor conv_out_1 = transpose(perm = conv_out_1_perm_0, x = input_31)[name = tensor("transpose_53")]; + tensor var_381 = slice_by_index(begin = var_381_begin_0, end = var_381_end_0, end_mask = var_381_end_mask_0, x = conv_out_1)[name = tensor("op_381")]; + tensor var_383_perm_0 = const()[name = tensor("op_383_perm_0"), val = tensor([1, 0, 2])]; + tensor var_383 = transpose(perm = var_383_perm_0, x = var_381)[name = tensor("transpose_52")]; + tensor input_33 = add(x = x_3, y = var_383)[name = tensor("input_33")]; + tensor input_35_axes_0 = const()[name = tensor("input_35_axes_0"), val = tensor([-1])]; + tensor input_35 = layer_norm(axes = input_35_axes_0, beta = inner_encoder_ffn2_0_module_sequential_0_bias, epsilon = var_76, gamma = inner_encoder_ffn2_0_module_sequential_0_weight, x = input_33)[name = tensor("input_35")]; + tensor inputs_9 = linear(bias = inner_encoder_ffn2_0_module_sequential_1_linear_bias, weight = inner_encoder_ffn2_0_module_sequential_1_linear_weight, x = input_35)[name = tensor("linear_8")]; + tensor input_37 = silu(x = inputs_9)[name = tensor("input_37")]; + tensor input_41 = linear(bias = inner_encoder_ffn2_0_module_sequential_4_linear_bias, weight = inner_encoder_ffn2_0_module_sequential_4_linear_weight, x = input_37)[name = tensor("linear_9")]; + tensor var_406 = const()[name = tensor("op_406"), val = tensor(0x1p-1)]; + tensor var_407 = mul(x = input_41, y = var_406)[name = tensor("op_407")]; + tensor input_43 = add(x = var_407, y = input_33)[name = tensor("input_43")]; tensor input_45_axes_0 = const()[name = tensor("input_45_axes_0"), val = tensor([-1])]; - tensor input_45 = layer_norm(axes = input_45_axes_0, beta = encoder_ffn1_1_module_sequential_0_bias, epsilon = var_29, gamma = encoder_ffn1_1_module_sequential_0_weight, x = input_43)[name = tensor("input_45")]; - tensor inputs_11 = linear(bias = encoder_ffn1_1_module_sequential_1_linear_bias, weight = encoder_ffn1_1_module_sequential_1_linear_weight, x = input_45)[name = tensor("linear_10")]; - tensor input_47 = silu(x = inputs_11)[name = tensor("input_47")]; - tensor input_51 = linear(bias = encoder_ffn1_1_module_sequential_4_linear_bias, weight = encoder_ffn1_1_module_sequential_4_linear_weight, x = input_47)[name = tensor("linear_11")]; - tensor var_370 = const()[name = tensor("op_370"), val = tensor(0x1p-1)]; - tensor var_371 = mul(x = input_51, y = var_370)[name = tensor("op_371")]; - tensor input_53 = add(x = var_371, y = input_43)[name = tensor("input_53")]; + tensor input_45 = layer_norm(axes = input_45_axes_0, beta = inner_encoder_layer_norm_0_bias, epsilon = var_76, gamma = inner_encoder_layer_norm_0_weight, x = input_43)[name = tensor("input_45")]; + tensor input_47_axes_0 = const()[name = tensor("input_47_axes_0"), val = tensor([-1])]; + tensor input_47 = layer_norm(axes = input_47_axes_0, beta = inner_encoder_ffn1_1_module_sequential_0_bias, epsilon = var_76, gamma = inner_encoder_ffn1_1_module_sequential_0_weight, x = input_45)[name = tensor("input_47")]; + tensor inputs_11 = linear(bias = inner_encoder_ffn1_1_module_sequential_1_linear_bias, weight = inner_encoder_ffn1_1_module_sequential_1_linear_weight, x = input_47)[name = tensor("linear_10")]; + tensor input_49 = silu(x = inputs_11)[name = tensor("input_49")]; + tensor input_53 = linear(bias = inner_encoder_ffn1_1_module_sequential_4_linear_bias, weight = inner_encoder_ffn1_1_module_sequential_4_linear_weight, x = input_49)[name = tensor("linear_11")]; + tensor var_436 = const()[name = tensor("op_436"), val = tensor(0x1p-1)]; + tensor var_437 = mul(x = input_53, y = var_436)[name = tensor("op_437")]; + tensor input_55 = add(x = var_437, y = input_45)[name = tensor("input_55")]; tensor x_7_axes_0 = const()[name = tensor("x_7_axes_0"), val = tensor([-1])]; - tensor x_7 = layer_norm(axes = x_7_axes_0, beta = encoder_ret_lns_1_bias, epsilon = var_29, gamma = encoder_ret_lns_1_weight, x = input_53)[name = tensor("x_7")]; + tensor x_7 = layer_norm(axes = x_7_axes_0, beta = inner_encoder_ret_lns_1_bias, epsilon = var_76, gamma = inner_encoder_ret_lns_1_weight, x = input_55)[name = tensor("x_7")]; tensor prev_kv_3_begin_0 = const()[name = tensor("prev_kv_3_begin_0"), val = tensor([1, 0, 0, 0, 0])]; tensor prev_kv_3_end_0 = const()[name = tensor("prev_kv_3_end_0"), val = tensor([2, 1, 4, 64, 64])]; tensor prev_kv_3_end_mask_0 = const()[name = tensor("prev_kv_3_end_mask_0"), val = tensor([false, true, true, true, true])]; @@ -426,183 +452,183 @@ program(1.0) tensor prev_scale_3_end_mask_0 = const()[name = tensor("prev_scale_3_end_mask_0"), val = tensor([false, true])]; tensor prev_scale_3_squeeze_mask_0 = const()[name = tensor("prev_scale_3_squeeze_mask_0"), val = tensor([true, false])]; tensor prev_scale_3 = slice_by_index(begin = prev_scale_3_begin_0, end = prev_scale_3_end_0, end_mask = prev_scale_3_end_mask_0, squeeze_mask = prev_scale_3_squeeze_mask_0, x = enc_scale)[name = tensor("prev_scale_3")]; - tensor var_385 = linear(bias = encoder_q_proj_1_bias, weight = encoder_q_proj_1_weight, x = x_7)[name = tensor("linear_12")]; - tensor var_386 = const()[name = tensor("op_386"), val = tensor([1, 5, 4, 64])]; - tensor var_387 = reshape(shape = var_386, x = var_385)[name = tensor("op_387")]; + tensor var_451 = linear(bias = inner_encoder_q_proj_1_bias, weight = inner_encoder_q_proj_1_weight, x = x_7)[name = tensor("linear_12")]; + tensor var_452 = const()[name = tensor("op_452"), val = tensor([1, 5, 4, 64])]; + tensor var_453 = reshape(shape = var_452, x = var_451)[name = tensor("op_453")]; tensor q_3_perm_0 = const()[name = tensor("q_3_perm_0"), val = tensor([0, 2, 1, 3])]; - tensor var_391 = linear(bias = encoder_k_proj_1_bias, weight = encoder_k_proj_1_weight, x = x_7)[name = tensor("linear_13")]; - tensor var_392 = const()[name = tensor("op_392"), val = tensor(0x1p-3)]; - tensor var_393 = mul(x = var_391, y = var_392)[name = tensor("op_393")]; - tensor var_394 = const()[name = tensor("op_394"), val = tensor([1, 5, 4, 64])]; - tensor var_395 = reshape(shape = var_394, x = var_393)[name = tensor("op_395")]; + tensor var_457 = linear(bias = inner_encoder_k_proj_1_bias, weight = inner_encoder_k_proj_1_weight, x = x_7)[name = tensor("linear_13")]; + tensor var_458 = const()[name = tensor("op_458"), val = tensor(0x1p-3)]; + tensor var_459 = mul(x = var_457, y = var_458)[name = tensor("op_459")]; + tensor var_460 = const()[name = tensor("op_460"), val = tensor([1, 5, 4, 64])]; + tensor var_461 = reshape(shape = var_460, x = var_459)[name = tensor("op_461")]; tensor k_3_perm_0 = const()[name = tensor("k_3_perm_0"), val = tensor([0, 2, 1, 3])]; - tensor var_399 = linear(bias = encoder_v_proj_1_bias, weight = encoder_v_proj_1_weight, x = x_7)[name = tensor("linear_14")]; - tensor var_400 = const()[name = tensor("op_400"), val = tensor([1, 5, 4, 64])]; - tensor var_401 = reshape(shape = var_400, x = var_399)[name = tensor("op_401")]; + tensor var_465 = linear(bias = inner_encoder_v_proj_1_bias, weight = inner_encoder_v_proj_1_weight, x = x_7)[name = tensor("linear_14")]; + tensor var_466 = const()[name = tensor("op_466"), val = tensor([1, 5, 4, 64])]; + tensor var_467 = reshape(shape = var_466, x = var_465)[name = tensor("op_467")]; tensor v_3_perm_0 = const()[name = tensor("v_3_perm_0"), val = tensor([0, 2, 1, 3])]; - tensor input_57 = linear(bias = encoder_g_proj_1_bias, weight = encoder_g_proj_1_weight, x = x_7)[name = tensor("linear_15")]; + tensor input_59 = linear(bias = inner_encoder_g_proj_1_bias, weight = inner_encoder_g_proj_1_weight, x = x_7)[name = tensor("linear_15")]; tensor sqrt_s0_3 = sqrt(x = prev_scale_3)[name = tensor("sqrt_s0_3")]; - tensor s_t_3 = add(x = prev_scale_3, y = encoder__t_index)[name = tensor("s_t_3")]; + tensor s_t_3 = add(x = prev_scale_3, y = inner_encoder__t_index)[name = tensor("s_t_3")]; tensor sqrt_s_t_3 = sqrt(x = s_t_3)[name = tensor("sqrt_s_t_3")]; tensor qk_3_transpose_x_1 = const()[name = tensor("qk_3_transpose_x_1"), val = tensor(false)]; tensor qk_3_transpose_y_1 = const()[name = tensor("qk_3_transpose_y_1"), val = tensor(true)]; - tensor k_3 = transpose(perm = k_3_perm_0, x = var_395)[name = tensor("transpose_50")]; - tensor q_3 = transpose(perm = q_3_perm_0, x = var_387)[name = tensor("transpose_51")]; + tensor k_3 = transpose(perm = k_3_perm_0, x = var_461)[name = tensor("transpose_50")]; + tensor q_3 = transpose(perm = q_3_perm_0, x = var_453)[name = tensor("transpose_51")]; tensor qk_3 = matmul(transpose_x = qk_3_transpose_x_1, transpose_y = qk_3_transpose_y_1, x = q_3, y = k_3)[name = tensor("qk_3")]; - tensor var_411 = const()[name = tensor("op_411"), val = tensor([5, 1])]; - tensor var_412 = reshape(shape = var_411, x = sqrt_s_t_3)[name = tensor("op_412")]; - tensor M_3 = real_div(x = encoder__causal_mask, y = var_412)[name = tensor("M_3")]; - tensor var_414 = mul(x = qk_3, y = M_3)[name = tensor("op_414")]; + tensor var_477 = const()[name = tensor("op_477"), val = tensor([5, 1])]; + tensor var_478 = reshape(shape = var_477, x = sqrt_s_t_3)[name = tensor("op_478")]; + tensor M_3 = real_div(x = inner_encoder__causal_mask, y = var_478)[name = tensor("M_3")]; + tensor var_480 = mul(x = qk_3, y = M_3)[name = tensor("op_480")]; tensor inner_3_transpose_x_0 = const()[name = tensor("inner_3_transpose_x_0"), val = tensor(false)]; tensor inner_3_transpose_y_0 = const()[name = tensor("inner_3_transpose_y_0"), val = tensor(false)]; - tensor v_3 = transpose(perm = v_3_perm_0, x = var_401)[name = tensor("transpose_49")]; - tensor inner_3 = matmul(transpose_x = inner_3_transpose_x_0, transpose_y = inner_3_transpose_y_0, x = var_414, y = v_3)[name = tensor("inner_3")]; - tensor var_416_transpose_x_0 = const()[name = tensor("op_416_transpose_x_0"), val = tensor(false)]; - tensor var_416_transpose_y_0 = const()[name = tensor("op_416_transpose_y_0"), val = tensor(false)]; - tensor var_416 = matmul(transpose_x = var_416_transpose_x_0, transpose_y = var_416_transpose_y_0, x = q_3, y = prev_kv_3)[name = tensor("op_416")]; - tensor var_417 = real_div(x = sqrt_s0_3, y = sqrt_s_t_3)[name = tensor("op_417")]; - tensor var_418 = const()[name = tensor("op_418"), val = tensor([1, 1, 5, 1])]; - tensor var_419 = reshape(shape = var_418, x = var_417)[name = tensor("op_419")]; - tensor cross_3 = mul(x = var_416, y = var_419)[name = tensor("cross_3")]; + tensor v_3 = transpose(perm = v_3_perm_0, x = var_467)[name = tensor("transpose_49")]; + tensor inner_3 = matmul(transpose_x = inner_3_transpose_x_0, transpose_y = inner_3_transpose_y_0, x = var_480, y = v_3)[name = tensor("inner_3")]; + tensor var_482_transpose_x_0 = const()[name = tensor("op_482_transpose_x_0"), val = tensor(false)]; + tensor var_482_transpose_y_0 = const()[name = tensor("op_482_transpose_y_0"), val = tensor(false)]; + tensor var_482 = matmul(transpose_x = var_482_transpose_x_0, transpose_y = var_482_transpose_y_0, x = q_3, y = prev_kv_3)[name = tensor("op_482")]; + tensor var_483 = real_div(x = sqrt_s0_3, y = sqrt_s_t_3)[name = tensor("op_483")]; + tensor var_484 = const()[name = tensor("op_484"), val = tensor([1, 1, 5, 1])]; + tensor var_485 = reshape(shape = var_484, x = var_483)[name = tensor("op_485")]; + tensor cross_3 = mul(x = var_482, y = var_485)[name = tensor("cross_3")]; tensor out_7 = add(x = inner_3, y = cross_3)[name = tensor("out_7")]; - tensor var_422 = mul(x = prev_kv_3, y = sqrt_s0_3)[name = tensor("op_422")]; - tensor var_424_transpose_x_1 = const()[name = tensor("op_424_transpose_x_1"), val = tensor(true)]; - tensor var_424_transpose_y_1 = const()[name = tensor("op_424_transpose_y_1"), val = tensor(false)]; - tensor var_424 = matmul(transpose_x = var_424_transpose_x_1, transpose_y = var_424_transpose_y_1, x = k_3, y = v_3)[name = tensor("op_424")]; - tensor new_kv_unnorm_3 = add(x = var_422, y = var_424)[name = tensor("new_kv_unnorm_3")]; - tensor var_426 = const()[name = tensor("op_426"), val = tensor(0x1.4p+2)]; - tensor new_scale_3 = add(x = prev_scale_3, y = var_426)[name = tensor("new_scale_3")]; - tensor var_428 = sqrt(x = new_scale_3)[name = tensor("op_428")]; - tensor var_429 = real_div(x = new_kv_unnorm_3, y = var_428)[name = tensor("op_429")]; - tensor var_430_perm_0 = const()[name = tensor("op_430_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_488 = mul(x = prev_kv_3, y = sqrt_s0_3)[name = tensor("op_488")]; + tensor var_490_transpose_x_1 = const()[name = tensor("op_490_transpose_x_1"), val = tensor(true)]; + tensor var_490_transpose_y_1 = const()[name = tensor("op_490_transpose_y_1"), val = tensor(false)]; + tensor var_490 = matmul(transpose_x = var_490_transpose_x_1, transpose_y = var_490_transpose_y_1, x = k_3, y = v_3)[name = tensor("op_490")]; + tensor new_kv_unnorm_3 = add(x = var_488, y = var_490)[name = tensor("new_kv_unnorm_3")]; + tensor var_492 = const()[name = tensor("op_492"), val = tensor(0x1.4p+2)]; + tensor new_scale_3 = add(x = prev_scale_3, y = var_492)[name = tensor("new_scale_3")]; + tensor var_494 = sqrt(x = new_scale_3)[name = tensor("op_494")]; + tensor var_495 = real_div(x = new_kv_unnorm_3, y = var_494)[name = tensor("op_495")]; + tensor var_496_perm_0 = const()[name = tensor("op_496_perm_0"), val = tensor([0, 2, 1, 3])]; tensor out_9_axes_0 = const()[name = tensor("out_9_axes_0"), val = tensor([-1])]; - tensor var_430 = transpose(perm = var_430_perm_0, x = out_7)[name = tensor("transpose_48")]; - tensor out_9 = layer_norm(axes = out_9_axes_0, epsilon = var_18, x = var_430)[name = tensor("out_9")]; - tensor var_434 = const()[name = tensor("op_434"), val = tensor([1, 5, 256])]; - tensor out_11 = reshape(shape = var_434, x = out_9)[name = tensor("out_11")]; - tensor var_436 = silu(x = input_57)[name = tensor("op_436")]; - tensor input_59 = mul(x = var_436, y = out_11)[name = tensor("input_59")]; - tensor ret_out_3 = linear(bias = encoder_out_proj_1_bias, weight = encoder_out_proj_1_weight, x = input_59)[name = tensor("linear_16")]; - tensor x_9 = add(x = input_53, y = ret_out_3)[name = tensor("x_9")]; + tensor var_496 = transpose(perm = var_496_perm_0, x = out_7)[name = tensor("transpose_48")]; + tensor out_9 = layer_norm(axes = out_9_axes_0, epsilon = var_84, x = var_496)[name = tensor("out_9")]; + tensor var_500 = const()[name = tensor("op_500"), val = tensor([1, 5, 256])]; + tensor out_11 = reshape(shape = var_500, x = out_9)[name = tensor("out_11")]; + tensor var_502 = silu(x = input_59)[name = tensor("op_502")]; + tensor input_61 = mul(x = var_502, y = out_11)[name = tensor("input_61")]; + tensor ret_out_3 = linear(bias = inner_encoder_out_proj_1_bias, weight = inner_encoder_out_proj_1_weight, x = input_61)[name = tensor("linear_16")]; + tensor x_9 = add(x = input_55, y = ret_out_3)[name = tensor("x_9")]; tensor window_13_begin_0 = const()[name = tensor("window_13_begin_0"), val = tensor([1, 0, 0, 0])]; tensor window_13_end_0 = const()[name = tensor("window_13_end_0"), val = tensor([2, 1, 16, 256])]; tensor window_13_end_mask_0 = const()[name = tensor("window_13_end_mask_0"), val = tensor([false, true, true, true])]; tensor window_13_squeeze_mask_0 = const()[name = tensor("window_13_squeeze_mask_0"), val = tensor([true, false, false, false])]; tensor window_13 = slice_by_index(begin = window_13_begin_0, end = window_13_end_0, end_mask = window_13_end_mask_0, squeeze_mask = window_13_squeeze_mask_0, x = enc_conv_cache)[name = tensor("window_13")]; - tensor var_444_begin_0 = const()[name = tensor("op_444_begin_0"), val = tensor([0, 0, 0])]; - tensor var_444_end_0 = const()[name = tensor("op_444_end_0"), val = tensor([1, 1, 256])]; - tensor var_444_end_mask_0 = const()[name = tensor("op_444_end_mask_0"), val = tensor([true, false, true])]; - tensor var_444 = slice_by_index(begin = var_444_begin_0, end = var_444_end_0, end_mask = var_444_end_mask_0, x = x_9)[name = tensor("op_444")]; - tensor var_447_begin_0 = const()[name = tensor("op_447_begin_0"), val = tensor([0, 1, 0])]; - tensor var_447_end_0 = const()[name = tensor("op_447_end_0"), val = tensor([1, 16, 256])]; - tensor var_447_end_mask_0 = const()[name = tensor("op_447_end_mask_0"), val = tensor([true, true, true])]; - tensor var_447 = slice_by_index(begin = var_447_begin_0, end = var_447_end_0, end_mask = var_447_end_mask_0, x = window_13)[name = tensor("op_447")]; + tensor var_510_begin_0 = const()[name = tensor("op_510_begin_0"), val = tensor([0, 0, 0])]; + tensor var_510_end_0 = const()[name = tensor("op_510_end_0"), val = tensor([1, 1, 256])]; + tensor var_510_end_mask_0 = const()[name = tensor("op_510_end_mask_0"), val = tensor([true, false, true])]; + tensor var_510 = slice_by_index(begin = var_510_begin_0, end = var_510_end_0, end_mask = var_510_end_mask_0, x = x_9)[name = tensor("op_510")]; + tensor var_513_begin_0 = const()[name = tensor("op_513_begin_0"), val = tensor([0, 1, 0])]; + tensor var_513_end_0 = const()[name = tensor("op_513_end_0"), val = tensor([1, 16, 256])]; + tensor var_513_end_mask_0 = const()[name = tensor("op_513_end_mask_0"), val = tensor([true, true, true])]; + tensor var_513 = slice_by_index(begin = var_513_begin_0, end = var_513_end_0, end_mask = var_513_end_mask_0, x = window_13)[name = tensor("op_513")]; tensor window_15_interleave_0 = const()[name = tensor("window_15_interleave_0"), val = tensor(false)]; - tensor window_15 = concat(axis = var_27, interleave = window_15_interleave_0, values = (var_447, var_444))[name = tensor("window_15")]; - tensor var_452_begin_0 = const()[name = tensor("op_452_begin_0"), val = tensor([0, 1, 0])]; - tensor var_452_end_0 = const()[name = tensor("op_452_end_0"), val = tensor([1, 2, 256])]; - tensor var_452_end_mask_0 = const()[name = tensor("op_452_end_mask_0"), val = tensor([true, false, true])]; - tensor var_452 = slice_by_index(begin = var_452_begin_0, end = var_452_end_0, end_mask = var_452_end_mask_0, x = x_9)[name = tensor("op_452")]; - tensor var_455_begin_0 = const()[name = tensor("op_455_begin_0"), val = tensor([0, 1, 0])]; - tensor var_455_end_0 = const()[name = tensor("op_455_end_0"), val = tensor([1, 16, 256])]; - tensor var_455_end_mask_0 = const()[name = tensor("op_455_end_mask_0"), val = tensor([true, true, true])]; - tensor var_455 = slice_by_index(begin = var_455_begin_0, end = var_455_end_0, end_mask = var_455_end_mask_0, x = window_15)[name = tensor("op_455")]; + tensor window_15 = concat(axis = var_93, interleave = window_15_interleave_0, values = (var_513, var_510))[name = tensor("window_15")]; + tensor var_518_begin_0 = const()[name = tensor("op_518_begin_0"), val = tensor([0, 1, 0])]; + tensor var_518_end_0 = const()[name = tensor("op_518_end_0"), val = tensor([1, 2, 256])]; + tensor var_518_end_mask_0 = const()[name = tensor("op_518_end_mask_0"), val = tensor([true, false, true])]; + tensor var_518 = slice_by_index(begin = var_518_begin_0, end = var_518_end_0, end_mask = var_518_end_mask_0, x = x_9)[name = tensor("op_518")]; + tensor var_521_begin_0 = const()[name = tensor("op_521_begin_0"), val = tensor([0, 1, 0])]; + tensor var_521_end_0 = const()[name = tensor("op_521_end_0"), val = tensor([1, 16, 256])]; + tensor var_521_end_mask_0 = const()[name = tensor("op_521_end_mask_0"), val = tensor([true, true, true])]; + tensor var_521 = slice_by_index(begin = var_521_begin_0, end = var_521_end_0, end_mask = var_521_end_mask_0, x = window_15)[name = tensor("op_521")]; tensor window_17_interleave_0 = const()[name = tensor("window_17_interleave_0"), val = tensor(false)]; - tensor window_17 = concat(axis = var_27, interleave = window_17_interleave_0, values = (var_455, var_452))[name = tensor("window_17")]; - tensor var_460_begin_0 = const()[name = tensor("op_460_begin_0"), val = tensor([0, 2, 0])]; - tensor var_460_end_0 = const()[name = tensor("op_460_end_0"), val = tensor([1, 3, 256])]; - tensor var_460_end_mask_0 = const()[name = tensor("op_460_end_mask_0"), val = tensor([true, false, true])]; - tensor var_460 = slice_by_index(begin = var_460_begin_0, end = var_460_end_0, end_mask = var_460_end_mask_0, x = x_9)[name = tensor("op_460")]; - tensor var_463_begin_0 = const()[name = tensor("op_463_begin_0"), val = tensor([0, 1, 0])]; - tensor var_463_end_0 = const()[name = tensor("op_463_end_0"), val = tensor([1, 16, 256])]; - tensor var_463_end_mask_0 = const()[name = tensor("op_463_end_mask_0"), val = tensor([true, true, true])]; - tensor var_463 = slice_by_index(begin = var_463_begin_0, end = var_463_end_0, end_mask = var_463_end_mask_0, x = window_17)[name = tensor("op_463")]; + tensor window_17 = concat(axis = var_93, interleave = window_17_interleave_0, values = (var_521, var_518))[name = tensor("window_17")]; + tensor var_526_begin_0 = const()[name = tensor("op_526_begin_0"), val = tensor([0, 2, 0])]; + tensor var_526_end_0 = const()[name = tensor("op_526_end_0"), val = tensor([1, 3, 256])]; + tensor var_526_end_mask_0 = const()[name = tensor("op_526_end_mask_0"), val = tensor([true, false, true])]; + tensor var_526 = slice_by_index(begin = var_526_begin_0, end = var_526_end_0, end_mask = var_526_end_mask_0, x = x_9)[name = tensor("op_526")]; + tensor var_529_begin_0 = const()[name = tensor("op_529_begin_0"), val = tensor([0, 1, 0])]; + tensor var_529_end_0 = const()[name = tensor("op_529_end_0"), val = tensor([1, 16, 256])]; + tensor var_529_end_mask_0 = const()[name = tensor("op_529_end_mask_0"), val = tensor([true, true, true])]; + tensor var_529 = slice_by_index(begin = var_529_begin_0, end = var_529_end_0, end_mask = var_529_end_mask_0, x = window_17)[name = tensor("op_529")]; tensor window_19_interleave_0 = const()[name = tensor("window_19_interleave_0"), val = tensor(false)]; - tensor window_19 = concat(axis = var_27, interleave = window_19_interleave_0, values = (var_463, var_460))[name = tensor("window_19")]; - tensor var_468_begin_0 = const()[name = tensor("op_468_begin_0"), val = tensor([0, 3, 0])]; - tensor var_468_end_0 = const()[name = tensor("op_468_end_0"), val = tensor([1, 4, 256])]; - tensor var_468_end_mask_0 = const()[name = tensor("op_468_end_mask_0"), val = tensor([true, false, true])]; - tensor var_468 = slice_by_index(begin = var_468_begin_0, end = var_468_end_0, end_mask = var_468_end_mask_0, x = x_9)[name = tensor("op_468")]; - tensor var_471_begin_0 = const()[name = tensor("op_471_begin_0"), val = tensor([0, 1, 0])]; - tensor var_471_end_0 = const()[name = tensor("op_471_end_0"), val = tensor([1, 16, 256])]; - tensor var_471_end_mask_0 = const()[name = tensor("op_471_end_mask_0"), val = tensor([true, true, true])]; - tensor var_471 = slice_by_index(begin = var_471_begin_0, end = var_471_end_0, end_mask = var_471_end_mask_0, x = window_19)[name = tensor("op_471")]; + tensor window_19 = concat(axis = var_93, interleave = window_19_interleave_0, values = (var_529, var_526))[name = tensor("window_19")]; + tensor var_534_begin_0 = const()[name = tensor("op_534_begin_0"), val = tensor([0, 3, 0])]; + tensor var_534_end_0 = const()[name = tensor("op_534_end_0"), val = tensor([1, 4, 256])]; + tensor var_534_end_mask_0 = const()[name = tensor("op_534_end_mask_0"), val = tensor([true, false, true])]; + tensor var_534 = slice_by_index(begin = var_534_begin_0, end = var_534_end_0, end_mask = var_534_end_mask_0, x = x_9)[name = tensor("op_534")]; + tensor var_537_begin_0 = const()[name = tensor("op_537_begin_0"), val = tensor([0, 1, 0])]; + tensor var_537_end_0 = const()[name = tensor("op_537_end_0"), val = tensor([1, 16, 256])]; + tensor var_537_end_mask_0 = const()[name = tensor("op_537_end_mask_0"), val = tensor([true, true, true])]; + tensor var_537 = slice_by_index(begin = var_537_begin_0, end = var_537_end_0, end_mask = var_537_end_mask_0, x = window_19)[name = tensor("op_537")]; tensor window_21_interleave_0 = const()[name = tensor("window_21_interleave_0"), val = tensor(false)]; - tensor window_21 = concat(axis = var_27, interleave = window_21_interleave_0, values = (var_471, var_468))[name = tensor("window_21")]; - tensor var_476_begin_0 = const()[name = tensor("op_476_begin_0"), val = tensor([0, 4, 0])]; - tensor var_476_end_0 = const()[name = tensor("op_476_end_0"), val = tensor([1, 1, 256])]; - tensor var_476_end_mask_0 = const()[name = tensor("op_476_end_mask_0"), val = tensor([true, true, true])]; - tensor var_476 = slice_by_index(begin = var_476_begin_0, end = var_476_end_0, end_mask = var_476_end_mask_0, x = x_9)[name = tensor("op_476")]; - tensor var_479_begin_0 = const()[name = tensor("op_479_begin_0"), val = tensor([0, 1, 0])]; - tensor var_479_end_0 = const()[name = tensor("op_479_end_0"), val = tensor([1, 16, 256])]; - tensor var_479_end_mask_0 = const()[name = tensor("op_479_end_mask_0"), val = tensor([true, true, true])]; - tensor var_479 = slice_by_index(begin = var_479_begin_0, end = var_479_end_0, end_mask = var_479_end_mask_0, x = window_21)[name = tensor("op_479")]; + tensor window_21 = concat(axis = var_93, interleave = window_21_interleave_0, values = (var_537, var_534))[name = tensor("window_21")]; + tensor var_542_begin_0 = const()[name = tensor("op_542_begin_0"), val = tensor([0, 4, 0])]; + tensor var_542_end_0 = const()[name = tensor("op_542_end_0"), val = tensor([1, 1, 256])]; + tensor var_542_end_mask_0 = const()[name = tensor("op_542_end_mask_0"), val = tensor([true, true, true])]; + tensor var_542 = slice_by_index(begin = var_542_begin_0, end = var_542_end_0, end_mask = var_542_end_mask_0, x = x_9)[name = tensor("op_542")]; + tensor var_545_begin_0 = const()[name = tensor("op_545_begin_0"), val = tensor([0, 1, 0])]; + tensor var_545_end_0 = const()[name = tensor("op_545_end_0"), val = tensor([1, 16, 256])]; + tensor var_545_end_mask_0 = const()[name = tensor("op_545_end_mask_0"), val = tensor([true, true, true])]; + tensor var_545 = slice_by_index(begin = var_545_begin_0, end = var_545_end_0, end_mask = var_545_end_mask_0, x = window_21)[name = tensor("op_545")]; tensor window_23_interleave_0 = const()[name = tensor("window_23_interleave_0"), val = tensor(false)]; - tensor window_23 = concat(axis = var_27, interleave = window_23_interleave_0, values = (var_479, var_476))[name = tensor("window_23")]; - tensor input_61_interleave_0 = const()[name = tensor("input_61_interleave_0"), val = tensor(false)]; - tensor input_61 = concat(axis = var_24, interleave = input_61_interleave_0, values = (window_15, window_17, window_19, window_21, window_23))[name = tensor("input_61")]; + tensor window_23 = concat(axis = var_93, interleave = window_23_interleave_0, values = (var_545, var_542))[name = tensor("window_23")]; + tensor input_63_interleave_0 = const()[name = tensor("input_63_interleave_0"), val = tensor(false)]; + tensor input_63 = concat(axis = var_79, interleave = input_63_interleave_0, values = (window_15, window_17, window_19, window_21, window_23))[name = tensor("input_63")]; tensor x_11_axes_0 = const()[name = tensor("x_11_axes_0"), val = tensor([-1])]; - tensor x_11 = layer_norm(axes = x_11_axes_0, beta = encoder_conv_module_1_sequential_0_bias, epsilon = var_29, gamma = encoder_conv_module_1_sequential_0_weight, x = input_61)[name = tensor("x_11")]; - tensor input_63_perm_0 = const()[name = tensor("input_63_perm_0"), val = tensor([0, 2, 1])]; + tensor x_11 = layer_norm(axes = x_11_axes_0, beta = inner_encoder_conv_module_1_sequential_0_bias, epsilon = var_76, gamma = inner_encoder_conv_module_1_sequential_0_weight, x = input_63)[name = tensor("x_11")]; + tensor input_65_perm_0 = const()[name = tensor("input_65_perm_0"), val = tensor([0, 2, 1])]; tensor inputs_13_pad_type_0 = const()[name = tensor("inputs_13_pad_type_0"), val = tensor("valid")]; tensor inputs_13_strides_0 = const()[name = tensor("inputs_13_strides_0"), val = tensor([1])]; tensor inputs_13_pad_0 = const()[name = tensor("inputs_13_pad_0"), val = tensor([0, 0])]; tensor inputs_13_dilations_0 = const()[name = tensor("inputs_13_dilations_0"), val = tensor([1])]; tensor inputs_13_groups_0 = const()[name = tensor("inputs_13_groups_0"), val = tensor(1)]; - tensor input_63 = transpose(perm = input_63_perm_0, x = x_11)[name = tensor("transpose_47")]; - tensor inputs_13 = conv(bias = encoder_conv_module_1_sequential_2_conv_bias, dilations = inputs_13_dilations_0, groups = inputs_13_groups_0, pad = inputs_13_pad_0, pad_type = inputs_13_pad_type_0, strides = inputs_13_strides_0, weight = encoder_conv_module_1_sequential_2_conv_weight, x = input_63)[name = tensor("inputs_13")]; - tensor var_504_split_sizes_0 = const()[name = tensor("op_504_split_sizes_0"), val = tensor([256, 256])]; - tensor var_504_axis_0 = const()[name = tensor("op_504_axis_0"), val = tensor(1)]; - tensor var_504_0, tensor var_504_1 = split(axis = var_504_axis_0, split_sizes = var_504_split_sizes_0, x = inputs_13)[name = tensor("op_504")]; - tensor var_506 = sigmoid(x = var_504_1)[name = tensor("op_506")]; - tensor inputs_15 = mul(x = var_504_0, y = var_506)[name = tensor("inputs_15")]; + tensor input_65 = transpose(perm = input_65_perm_0, x = x_11)[name = tensor("transpose_47")]; + tensor inputs_13 = conv(bias = inner_encoder_conv_module_1_sequential_2_conv_bias, dilations = inputs_13_dilations_0, groups = inputs_13_groups_0, pad = inputs_13_pad_0, pad_type = inputs_13_pad_type_0, strides = inputs_13_strides_0, weight = inner_encoder_conv_module_1_sequential_2_conv_weight, x = input_65)[name = tensor("inputs_13")]; + tensor var_570_split_sizes_0 = const()[name = tensor("op_570_split_sizes_0"), val = tensor([256, 256])]; + tensor var_570_axis_0 = const()[name = tensor("op_570_axis_0"), val = tensor(1)]; + tensor var_570_0, tensor var_570_1 = split(axis = var_570_axis_0, split_sizes = var_570_split_sizes_0, x = inputs_13)[name = tensor("op_570")]; + tensor var_572 = sigmoid(x = var_570_1)[name = tensor("op_572")]; + tensor inputs_15 = mul(x = var_570_0, y = var_572)[name = tensor("inputs_15")]; tensor outputs_aug_3_pad_type_0 = const()[name = tensor("outputs_aug_3_pad_type_0"), val = tensor("custom")]; tensor outputs_aug_3_pad_0 = const()[name = tensor("outputs_aug_3_pad_0"), val = tensor([15, 15])]; tensor outputs_aug_3_groups_0 = const()[name = tensor("outputs_aug_3_groups_0"), val = tensor(256)]; tensor outputs_aug_3_strides_0 = const()[name = tensor("outputs_aug_3_strides_0"), val = tensor([1])]; tensor outputs_aug_3_dilations_0 = const()[name = tensor("outputs_aug_3_dilations_0"), val = tensor([1])]; - tensor outputs_aug_3 = conv(dilations = outputs_aug_3_dilations_0, groups = outputs_aug_3_groups_0, pad = outputs_aug_3_pad_0, pad_type = outputs_aug_3_pad_type_0, strides = outputs_aug_3_strides_0, weight = encoder_conv_module_1_sequential_4_conv_weight, x = inputs_15)[name = tensor("outputs_aug_3")]; - tensor input_65_begin_0 = const()[name = tensor("input_65_begin_0"), val = tensor([0, 0, 0])]; - tensor input_65_end_0 = const()[name = tensor("input_65_end_0"), val = tensor([5, 256, 16])]; - tensor input_65_end_mask_0 = const()[name = tensor("input_65_end_mask_0"), val = tensor([true, true, false])]; - tensor input_65 = slice_by_index(begin = input_65_begin_0, end = input_65_end_0, end_mask = input_65_end_mask_0, x = outputs_aug_3)[name = tensor("input_65")]; - tensor inputs_17 = batch_norm(beta = encoder_conv_module_1_sequential_5_bias, epsilon = var_29, gamma = encoder_conv_module_1_sequential_5_weight, mean = encoder_conv_module_1_sequential_5_running_mean, variance = encoder_conv_module_1_sequential_5_running_var, x = input_65)[name = tensor("inputs_17")]; - tensor input_67 = silu(x = inputs_17)[name = tensor("input_67")]; - tensor input_69_pad_type_0 = const()[name = tensor("input_69_pad_type_0"), val = tensor("valid")]; - tensor input_69_strides_0 = const()[name = tensor("input_69_strides_0"), val = tensor([1])]; - tensor input_69_pad_0 = const()[name = tensor("input_69_pad_0"), val = tensor([0, 0])]; - tensor input_69_dilations_0 = const()[name = tensor("input_69_dilations_0"), val = tensor([1])]; - tensor input_69_groups_0 = const()[name = tensor("input_69_groups_0"), val = tensor(1)]; - tensor input_69 = conv(bias = encoder_conv_module_1_sequential_7_conv_bias, dilations = input_69_dilations_0, groups = input_69_groups_0, pad = input_69_pad_0, pad_type = input_69_pad_type_0, strides = input_69_strides_0, weight = encoder_conv_module_1_sequential_7_conv_weight, x = input_67)[name = tensor("input_69")]; + tensor outputs_aug_3 = conv(dilations = outputs_aug_3_dilations_0, groups = outputs_aug_3_groups_0, pad = outputs_aug_3_pad_0, pad_type = outputs_aug_3_pad_type_0, strides = outputs_aug_3_strides_0, weight = inner_encoder_conv_module_1_sequential_4_conv_weight, x = inputs_15)[name = tensor("outputs_aug_3")]; + tensor input_67_begin_0 = const()[name = tensor("input_67_begin_0"), val = tensor([0, 0, 0])]; + tensor input_67_end_0 = const()[name = tensor("input_67_end_0"), val = tensor([5, 256, 16])]; + tensor input_67_end_mask_0 = const()[name = tensor("input_67_end_mask_0"), val = tensor([true, true, false])]; + tensor input_67 = slice_by_index(begin = input_67_begin_0, end = input_67_end_0, end_mask = input_67_end_mask_0, x = outputs_aug_3)[name = tensor("input_67")]; + tensor inputs_17 = batch_norm(beta = inner_encoder_conv_module_1_sequential_5_bias, epsilon = var_76, gamma = inner_encoder_conv_module_1_sequential_5_weight, mean = inner_encoder_conv_module_1_sequential_5_running_mean, variance = inner_encoder_conv_module_1_sequential_5_running_var, x = input_67)[name = tensor("inputs_17")]; + tensor input_69 = silu(x = inputs_17)[name = tensor("input_69")]; + tensor input_71_pad_type_0 = const()[name = tensor("input_71_pad_type_0"), val = tensor("valid")]; + tensor input_71_strides_0 = const()[name = tensor("input_71_strides_0"), val = tensor([1])]; + tensor input_71_pad_0 = const()[name = tensor("input_71_pad_0"), val = tensor([0, 0])]; + tensor input_71_dilations_0 = const()[name = tensor("input_71_dilations_0"), val = tensor([1])]; + tensor input_71_groups_0 = const()[name = tensor("input_71_groups_0"), val = tensor(1)]; + tensor input_71 = conv(bias = inner_encoder_conv_module_1_sequential_7_conv_bias, dilations = input_71_dilations_0, groups = input_71_groups_0, pad = input_71_pad_0, pad_type = input_71_pad_type_0, strides = input_71_strides_0, weight = inner_encoder_conv_module_1_sequential_7_conv_weight, x = input_69)[name = tensor("input_71")]; tensor conv_out_3_perm_0 = const()[name = tensor("conv_out_3_perm_0"), val = tensor([0, 2, 1])]; - tensor var_537_begin_0 = const()[name = tensor("op_537_begin_0"), val = tensor([0, -1, 0])]; - tensor var_537_end_0 = const()[name = tensor("op_537_end_0"), val = tensor([5, 16, 256])]; - tensor var_537_end_mask_0 = const()[name = tensor("op_537_end_mask_0"), val = tensor([true, true, true])]; - tensor conv_out_3 = transpose(perm = conv_out_3_perm_0, x = input_69)[name = tensor("transpose_46")]; - tensor var_537 = slice_by_index(begin = var_537_begin_0, end = var_537_end_0, end_mask = var_537_end_mask_0, x = conv_out_3)[name = tensor("op_537")]; - tensor var_539_perm_0 = const()[name = tensor("op_539_perm_0"), val = tensor([1, 0, 2])]; - tensor var_539 = transpose(perm = var_539_perm_0, x = var_537)[name = tensor("transpose_45")]; - tensor input_71 = add(x = x_9, y = var_539)[name = tensor("input_71")]; - tensor input_73_axes_0 = const()[name = tensor("input_73_axes_0"), val = tensor([-1])]; - tensor input_73 = layer_norm(axes = input_73_axes_0, beta = encoder_ffn2_1_module_sequential_0_bias, epsilon = var_29, gamma = encoder_ffn2_1_module_sequential_0_weight, x = input_71)[name = tensor("input_73")]; - tensor inputs_19 = linear(bias = encoder_ffn2_1_module_sequential_1_linear_bias, weight = encoder_ffn2_1_module_sequential_1_linear_weight, x = input_73)[name = tensor("linear_17")]; - tensor input_75 = silu(x = inputs_19)[name = tensor("input_75")]; - tensor input_79 = linear(bias = encoder_ffn2_1_module_sequential_4_linear_bias, weight = encoder_ffn2_1_module_sequential_4_linear_weight, x = input_75)[name = tensor("linear_18")]; - tensor var_562 = const()[name = tensor("op_562"), val = tensor(0x1p-1)]; - tensor var_563 = mul(x = input_79, y = var_562)[name = tensor("op_563")]; - tensor input_81 = add(x = var_563, y = input_71)[name = tensor("input_81")]; - tensor input_83_axes_0 = const()[name = tensor("input_83_axes_0"), val = tensor([-1])]; - tensor input_83 = layer_norm(axes = input_83_axes_0, beta = encoder_layer_norm_1_bias, epsilon = var_29, gamma = encoder_layer_norm_1_weight, x = input_81)[name = tensor("input_83")]; + tensor var_603_begin_0 = const()[name = tensor("op_603_begin_0"), val = tensor([0, -1, 0])]; + tensor var_603_end_0 = const()[name = tensor("op_603_end_0"), val = tensor([5, 16, 256])]; + tensor var_603_end_mask_0 = const()[name = tensor("op_603_end_mask_0"), val = tensor([true, true, true])]; + tensor conv_out_3 = transpose(perm = conv_out_3_perm_0, x = input_71)[name = tensor("transpose_46")]; + tensor var_603 = slice_by_index(begin = var_603_begin_0, end = var_603_end_0, end_mask = var_603_end_mask_0, x = conv_out_3)[name = tensor("op_603")]; + tensor var_605_perm_0 = const()[name = tensor("op_605_perm_0"), val = tensor([1, 0, 2])]; + tensor var_605 = transpose(perm = var_605_perm_0, x = var_603)[name = tensor("transpose_45")]; + tensor input_73 = add(x = x_9, y = var_605)[name = tensor("input_73")]; + tensor input_75_axes_0 = const()[name = tensor("input_75_axes_0"), val = tensor([-1])]; + tensor input_75 = layer_norm(axes = input_75_axes_0, beta = inner_encoder_ffn2_1_module_sequential_0_bias, epsilon = var_76, gamma = inner_encoder_ffn2_1_module_sequential_0_weight, x = input_73)[name = tensor("input_75")]; + tensor inputs_19 = linear(bias = inner_encoder_ffn2_1_module_sequential_1_linear_bias, weight = inner_encoder_ffn2_1_module_sequential_1_linear_weight, x = input_75)[name = tensor("linear_17")]; + tensor input_77 = silu(x = inputs_19)[name = tensor("input_77")]; + tensor input_81 = linear(bias = inner_encoder_ffn2_1_module_sequential_4_linear_bias, weight = inner_encoder_ffn2_1_module_sequential_4_linear_weight, x = input_77)[name = tensor("linear_18")]; + tensor var_628 = const()[name = tensor("op_628"), val = tensor(0x1p-1)]; + tensor var_629 = mul(x = input_81, y = var_628)[name = tensor("op_629")]; + tensor input_83 = add(x = var_629, y = input_73)[name = tensor("input_83")]; tensor input_85_axes_0 = const()[name = tensor("input_85_axes_0"), val = tensor([-1])]; - tensor input_85 = layer_norm(axes = input_85_axes_0, beta = encoder_ffn1_2_module_sequential_0_bias, epsilon = var_29, gamma = encoder_ffn1_2_module_sequential_0_weight, x = input_83)[name = tensor("input_85")]; - tensor inputs_21 = linear(bias = encoder_ffn1_2_module_sequential_1_linear_bias, weight = encoder_ffn1_2_module_sequential_1_linear_weight, x = input_85)[name = tensor("linear_19")]; - tensor input_87 = silu(x = inputs_21)[name = tensor("input_87")]; - tensor input_91 = linear(bias = encoder_ffn1_2_module_sequential_4_linear_bias, weight = encoder_ffn1_2_module_sequential_4_linear_weight, x = input_87)[name = tensor("linear_20")]; - tensor var_592 = const()[name = tensor("op_592"), val = tensor(0x1p-1)]; - tensor var_593 = mul(x = input_91, y = var_592)[name = tensor("op_593")]; - tensor input_93 = add(x = var_593, y = input_83)[name = tensor("input_93")]; + tensor input_85 = layer_norm(axes = input_85_axes_0, beta = inner_encoder_layer_norm_1_bias, epsilon = var_76, gamma = inner_encoder_layer_norm_1_weight, x = input_83)[name = tensor("input_85")]; + tensor input_87_axes_0 = const()[name = tensor("input_87_axes_0"), val = tensor([-1])]; + tensor input_87 = layer_norm(axes = input_87_axes_0, beta = inner_encoder_ffn1_2_module_sequential_0_bias, epsilon = var_76, gamma = inner_encoder_ffn1_2_module_sequential_0_weight, x = input_85)[name = tensor("input_87")]; + tensor inputs_21 = linear(bias = inner_encoder_ffn1_2_module_sequential_1_linear_bias, weight = inner_encoder_ffn1_2_module_sequential_1_linear_weight, x = input_87)[name = tensor("linear_19")]; + tensor input_89 = silu(x = inputs_21)[name = tensor("input_89")]; + tensor input_93 = linear(bias = inner_encoder_ffn1_2_module_sequential_4_linear_bias, weight = inner_encoder_ffn1_2_module_sequential_4_linear_weight, x = input_89)[name = tensor("linear_20")]; + tensor var_658 = const()[name = tensor("op_658"), val = tensor(0x1p-1)]; + tensor var_659 = mul(x = input_93, y = var_658)[name = tensor("op_659")]; + tensor input_95 = add(x = var_659, y = input_85)[name = tensor("input_95")]; tensor x_13_axes_0 = const()[name = tensor("x_13_axes_0"), val = tensor([-1])]; - tensor x_13 = layer_norm(axes = x_13_axes_0, beta = encoder_ret_lns_2_bias, epsilon = var_29, gamma = encoder_ret_lns_2_weight, x = input_93)[name = tensor("x_13")]; + tensor x_13 = layer_norm(axes = x_13_axes_0, beta = inner_encoder_ret_lns_2_bias, epsilon = var_76, gamma = inner_encoder_ret_lns_2_weight, x = input_95)[name = tensor("x_13")]; tensor prev_kv_5_begin_0 = const()[name = tensor("prev_kv_5_begin_0"), val = tensor([2, 0, 0, 0, 0])]; tensor prev_kv_5_end_0 = const()[name = tensor("prev_kv_5_end_0"), val = tensor([3, 1, 4, 64, 64])]; tensor prev_kv_5_end_mask_0 = const()[name = tensor("prev_kv_5_end_mask_0"), val = tensor([false, true, true, true, true])]; @@ -613,183 +639,183 @@ program(1.0) tensor prev_scale_5_end_mask_0 = const()[name = tensor("prev_scale_5_end_mask_0"), val = tensor([false, true])]; tensor prev_scale_5_squeeze_mask_0 = const()[name = tensor("prev_scale_5_squeeze_mask_0"), val = tensor([true, false])]; tensor prev_scale_5 = slice_by_index(begin = prev_scale_5_begin_0, end = prev_scale_5_end_0, end_mask = prev_scale_5_end_mask_0, squeeze_mask = prev_scale_5_squeeze_mask_0, x = enc_scale)[name = tensor("prev_scale_5")]; - tensor var_607 = linear(bias = encoder_q_proj_2_bias, weight = encoder_q_proj_2_weight, x = x_13)[name = tensor("linear_21")]; - tensor var_608 = const()[name = tensor("op_608"), val = tensor([1, 5, 4, 64])]; - tensor var_609 = reshape(shape = var_608, x = var_607)[name = tensor("op_609")]; + tensor var_673 = linear(bias = inner_encoder_q_proj_2_bias, weight = inner_encoder_q_proj_2_weight, x = x_13)[name = tensor("linear_21")]; + tensor var_674 = const()[name = tensor("op_674"), val = tensor([1, 5, 4, 64])]; + tensor var_675 = reshape(shape = var_674, x = var_673)[name = tensor("op_675")]; tensor q_5_perm_0 = const()[name = tensor("q_5_perm_0"), val = tensor([0, 2, 1, 3])]; - tensor var_613 = linear(bias = encoder_k_proj_2_bias, weight = encoder_k_proj_2_weight, x = x_13)[name = tensor("linear_22")]; - tensor var_614 = const()[name = tensor("op_614"), val = tensor(0x1p-3)]; - tensor var_615 = mul(x = var_613, y = var_614)[name = tensor("op_615")]; - tensor var_616 = const()[name = tensor("op_616"), val = tensor([1, 5, 4, 64])]; - tensor var_617 = reshape(shape = var_616, x = var_615)[name = tensor("op_617")]; + tensor var_679 = linear(bias = inner_encoder_k_proj_2_bias, weight = inner_encoder_k_proj_2_weight, x = x_13)[name = tensor("linear_22")]; + tensor var_680 = const()[name = tensor("op_680"), val = tensor(0x1p-3)]; + tensor var_681 = mul(x = var_679, y = var_680)[name = tensor("op_681")]; + tensor var_682 = const()[name = tensor("op_682"), val = tensor([1, 5, 4, 64])]; + tensor var_683 = reshape(shape = var_682, x = var_681)[name = tensor("op_683")]; tensor k_5_perm_0 = const()[name = tensor("k_5_perm_0"), val = tensor([0, 2, 1, 3])]; - tensor var_621 = linear(bias = encoder_v_proj_2_bias, weight = encoder_v_proj_2_weight, x = x_13)[name = tensor("linear_23")]; - tensor var_622 = const()[name = tensor("op_622"), val = tensor([1, 5, 4, 64])]; - tensor var_623 = reshape(shape = var_622, x = var_621)[name = tensor("op_623")]; + tensor var_687 = linear(bias = inner_encoder_v_proj_2_bias, weight = inner_encoder_v_proj_2_weight, x = x_13)[name = tensor("linear_23")]; + tensor var_688 = const()[name = tensor("op_688"), val = tensor([1, 5, 4, 64])]; + tensor var_689 = reshape(shape = var_688, x = var_687)[name = tensor("op_689")]; tensor v_5_perm_0 = const()[name = tensor("v_5_perm_0"), val = tensor([0, 2, 1, 3])]; - tensor input_97 = linear(bias = encoder_g_proj_2_bias, weight = encoder_g_proj_2_weight, x = x_13)[name = tensor("linear_24")]; + tensor input_99 = linear(bias = inner_encoder_g_proj_2_bias, weight = inner_encoder_g_proj_2_weight, x = x_13)[name = tensor("linear_24")]; tensor sqrt_s0_5 = sqrt(x = prev_scale_5)[name = tensor("sqrt_s0_5")]; - tensor s_t_5 = add(x = prev_scale_5, y = encoder__t_index)[name = tensor("s_t_5")]; + tensor s_t_5 = add(x = prev_scale_5, y = inner_encoder__t_index)[name = tensor("s_t_5")]; tensor sqrt_s_t_5 = sqrt(x = s_t_5)[name = tensor("sqrt_s_t_5")]; tensor qk_5_transpose_x_1 = const()[name = tensor("qk_5_transpose_x_1"), val = tensor(false)]; tensor qk_5_transpose_y_1 = const()[name = tensor("qk_5_transpose_y_1"), val = tensor(true)]; - tensor k_5 = transpose(perm = k_5_perm_0, x = var_617)[name = tensor("transpose_43")]; - tensor q_5 = transpose(perm = q_5_perm_0, x = var_609)[name = tensor("transpose_44")]; + tensor k_5 = transpose(perm = k_5_perm_0, x = var_683)[name = tensor("transpose_43")]; + tensor q_5 = transpose(perm = q_5_perm_0, x = var_675)[name = tensor("transpose_44")]; tensor qk_5 = matmul(transpose_x = qk_5_transpose_x_1, transpose_y = qk_5_transpose_y_1, x = q_5, y = k_5)[name = tensor("qk_5")]; - tensor var_633 = const()[name = tensor("op_633"), val = tensor([5, 1])]; - tensor var_634 = reshape(shape = var_633, x = sqrt_s_t_5)[name = tensor("op_634")]; - tensor M_5 = real_div(x = encoder__causal_mask, y = var_634)[name = tensor("M_5")]; - tensor var_636 = mul(x = qk_5, y = M_5)[name = tensor("op_636")]; + tensor var_699 = const()[name = tensor("op_699"), val = tensor([5, 1])]; + tensor var_700 = reshape(shape = var_699, x = sqrt_s_t_5)[name = tensor("op_700")]; + tensor M_5 = real_div(x = inner_encoder__causal_mask, y = var_700)[name = tensor("M_5")]; + tensor var_702 = mul(x = qk_5, y = M_5)[name = tensor("op_702")]; tensor inner_5_transpose_x_0 = const()[name = tensor("inner_5_transpose_x_0"), val = tensor(false)]; tensor inner_5_transpose_y_0 = const()[name = tensor("inner_5_transpose_y_0"), val = tensor(false)]; - tensor v_5 = transpose(perm = v_5_perm_0, x = var_623)[name = tensor("transpose_42")]; - tensor inner_5 = matmul(transpose_x = inner_5_transpose_x_0, transpose_y = inner_5_transpose_y_0, x = var_636, y = v_5)[name = tensor("inner_5")]; - tensor var_638_transpose_x_0 = const()[name = tensor("op_638_transpose_x_0"), val = tensor(false)]; - tensor var_638_transpose_y_0 = const()[name = tensor("op_638_transpose_y_0"), val = tensor(false)]; - tensor var_638 = matmul(transpose_x = var_638_transpose_x_0, transpose_y = var_638_transpose_y_0, x = q_5, y = prev_kv_5)[name = tensor("op_638")]; - tensor var_639 = real_div(x = sqrt_s0_5, y = sqrt_s_t_5)[name = tensor("op_639")]; - tensor var_640 = const()[name = tensor("op_640"), val = tensor([1, 1, 5, 1])]; - tensor var_641 = reshape(shape = var_640, x = var_639)[name = tensor("op_641")]; - tensor cross_5 = mul(x = var_638, y = var_641)[name = tensor("cross_5")]; + tensor v_5 = transpose(perm = v_5_perm_0, x = var_689)[name = tensor("transpose_42")]; + tensor inner_5 = matmul(transpose_x = inner_5_transpose_x_0, transpose_y = inner_5_transpose_y_0, x = var_702, y = v_5)[name = tensor("inner_5")]; + tensor var_704_transpose_x_0 = const()[name = tensor("op_704_transpose_x_0"), val = tensor(false)]; + tensor var_704_transpose_y_0 = const()[name = tensor("op_704_transpose_y_0"), val = tensor(false)]; + tensor var_704 = matmul(transpose_x = var_704_transpose_x_0, transpose_y = var_704_transpose_y_0, x = q_5, y = prev_kv_5)[name = tensor("op_704")]; + tensor var_705 = real_div(x = sqrt_s0_5, y = sqrt_s_t_5)[name = tensor("op_705")]; + tensor var_706 = const()[name = tensor("op_706"), val = tensor([1, 1, 5, 1])]; + tensor var_707 = reshape(shape = var_706, x = var_705)[name = tensor("op_707")]; + tensor cross_5 = mul(x = var_704, y = var_707)[name = tensor("cross_5")]; tensor out_13 = add(x = inner_5, y = cross_5)[name = tensor("out_13")]; - tensor var_644 = mul(x = prev_kv_5, y = sqrt_s0_5)[name = tensor("op_644")]; - tensor var_646_transpose_x_1 = const()[name = tensor("op_646_transpose_x_1"), val = tensor(true)]; - tensor var_646_transpose_y_1 = const()[name = tensor("op_646_transpose_y_1"), val = tensor(false)]; - tensor var_646 = matmul(transpose_x = var_646_transpose_x_1, transpose_y = var_646_transpose_y_1, x = k_5, y = v_5)[name = tensor("op_646")]; - tensor new_kv_unnorm_5 = add(x = var_644, y = var_646)[name = tensor("new_kv_unnorm_5")]; - tensor var_648 = const()[name = tensor("op_648"), val = tensor(0x1.4p+2)]; - tensor new_scale_5 = add(x = prev_scale_5, y = var_648)[name = tensor("new_scale_5")]; - tensor var_650 = sqrt(x = new_scale_5)[name = tensor("op_650")]; - tensor var_651 = real_div(x = new_kv_unnorm_5, y = var_650)[name = tensor("op_651")]; - tensor var_652_perm_0 = const()[name = tensor("op_652_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_710 = mul(x = prev_kv_5, y = sqrt_s0_5)[name = tensor("op_710")]; + tensor var_712_transpose_x_1 = const()[name = tensor("op_712_transpose_x_1"), val = tensor(true)]; + tensor var_712_transpose_y_1 = const()[name = tensor("op_712_transpose_y_1"), val = tensor(false)]; + tensor var_712 = matmul(transpose_x = var_712_transpose_x_1, transpose_y = var_712_transpose_y_1, x = k_5, y = v_5)[name = tensor("op_712")]; + tensor new_kv_unnorm_5 = add(x = var_710, y = var_712)[name = tensor("new_kv_unnorm_5")]; + tensor var_714 = const()[name = tensor("op_714"), val = tensor(0x1.4p+2)]; + tensor new_scale_5 = add(x = prev_scale_5, y = var_714)[name = tensor("new_scale_5")]; + tensor var_716 = sqrt(x = new_scale_5)[name = tensor("op_716")]; + tensor var_717 = real_div(x = new_kv_unnorm_5, y = var_716)[name = tensor("op_717")]; + tensor var_718_perm_0 = const()[name = tensor("op_718_perm_0"), val = tensor([0, 2, 1, 3])]; tensor out_15_axes_0 = const()[name = tensor("out_15_axes_0"), val = tensor([-1])]; - tensor var_652 = transpose(perm = var_652_perm_0, x = out_13)[name = tensor("transpose_41")]; - tensor out_15 = layer_norm(axes = out_15_axes_0, epsilon = var_18, x = var_652)[name = tensor("out_15")]; - tensor var_656 = const()[name = tensor("op_656"), val = tensor([1, 5, 256])]; - tensor out_17 = reshape(shape = var_656, x = out_15)[name = tensor("out_17")]; - tensor var_658 = silu(x = input_97)[name = tensor("op_658")]; - tensor input_99 = mul(x = var_658, y = out_17)[name = tensor("input_99")]; - tensor ret_out_5 = linear(bias = encoder_out_proj_2_bias, weight = encoder_out_proj_2_weight, x = input_99)[name = tensor("linear_25")]; - tensor x_15 = add(x = input_93, y = ret_out_5)[name = tensor("x_15")]; + tensor var_718 = transpose(perm = var_718_perm_0, x = out_13)[name = tensor("transpose_41")]; + tensor out_15 = layer_norm(axes = out_15_axes_0, epsilon = var_84, x = var_718)[name = tensor("out_15")]; + tensor var_722 = const()[name = tensor("op_722"), val = tensor([1, 5, 256])]; + tensor out_17 = reshape(shape = var_722, x = out_15)[name = tensor("out_17")]; + tensor var_724 = silu(x = input_99)[name = tensor("op_724")]; + tensor input_101 = mul(x = var_724, y = out_17)[name = tensor("input_101")]; + tensor ret_out_5 = linear(bias = inner_encoder_out_proj_2_bias, weight = inner_encoder_out_proj_2_weight, x = input_101)[name = tensor("linear_25")]; + tensor x_15 = add(x = input_95, y = ret_out_5)[name = tensor("x_15")]; tensor window_25_begin_0 = const()[name = tensor("window_25_begin_0"), val = tensor([2, 0, 0, 0])]; tensor window_25_end_0 = const()[name = tensor("window_25_end_0"), val = tensor([3, 1, 16, 256])]; tensor window_25_end_mask_0 = const()[name = tensor("window_25_end_mask_0"), val = tensor([false, true, true, true])]; tensor window_25_squeeze_mask_0 = const()[name = tensor("window_25_squeeze_mask_0"), val = tensor([true, false, false, false])]; tensor window_25 = slice_by_index(begin = window_25_begin_0, end = window_25_end_0, end_mask = window_25_end_mask_0, squeeze_mask = window_25_squeeze_mask_0, x = enc_conv_cache)[name = tensor("window_25")]; - tensor var_666_begin_0 = const()[name = tensor("op_666_begin_0"), val = tensor([0, 0, 0])]; - tensor var_666_end_0 = const()[name = tensor("op_666_end_0"), val = tensor([1, 1, 256])]; - tensor var_666_end_mask_0 = const()[name = tensor("op_666_end_mask_0"), val = tensor([true, false, true])]; - tensor var_666 = slice_by_index(begin = var_666_begin_0, end = var_666_end_0, end_mask = var_666_end_mask_0, x = x_15)[name = tensor("op_666")]; - tensor var_669_begin_0 = const()[name = tensor("op_669_begin_0"), val = tensor([0, 1, 0])]; - tensor var_669_end_0 = const()[name = tensor("op_669_end_0"), val = tensor([1, 16, 256])]; - tensor var_669_end_mask_0 = const()[name = tensor("op_669_end_mask_0"), val = tensor([true, true, true])]; - tensor var_669 = slice_by_index(begin = var_669_begin_0, end = var_669_end_0, end_mask = var_669_end_mask_0, x = window_25)[name = tensor("op_669")]; + tensor var_732_begin_0 = const()[name = tensor("op_732_begin_0"), val = tensor([0, 0, 0])]; + tensor var_732_end_0 = const()[name = tensor("op_732_end_0"), val = tensor([1, 1, 256])]; + tensor var_732_end_mask_0 = const()[name = tensor("op_732_end_mask_0"), val = tensor([true, false, true])]; + tensor var_732 = slice_by_index(begin = var_732_begin_0, end = var_732_end_0, end_mask = var_732_end_mask_0, x = x_15)[name = tensor("op_732")]; + tensor var_735_begin_0 = const()[name = tensor("op_735_begin_0"), val = tensor([0, 1, 0])]; + tensor var_735_end_0 = const()[name = tensor("op_735_end_0"), val = tensor([1, 16, 256])]; + tensor var_735_end_mask_0 = const()[name = tensor("op_735_end_mask_0"), val = tensor([true, true, true])]; + tensor var_735 = slice_by_index(begin = var_735_begin_0, end = var_735_end_0, end_mask = var_735_end_mask_0, x = window_25)[name = tensor("op_735")]; tensor window_27_interleave_0 = const()[name = tensor("window_27_interleave_0"), val = tensor(false)]; - tensor window_27 = concat(axis = var_27, interleave = window_27_interleave_0, values = (var_669, var_666))[name = tensor("window_27")]; - tensor var_674_begin_0 = const()[name = tensor("op_674_begin_0"), val = tensor([0, 1, 0])]; - tensor var_674_end_0 = const()[name = tensor("op_674_end_0"), val = tensor([1, 2, 256])]; - tensor var_674_end_mask_0 = const()[name = tensor("op_674_end_mask_0"), val = tensor([true, false, true])]; - tensor var_674 = slice_by_index(begin = var_674_begin_0, end = var_674_end_0, end_mask = var_674_end_mask_0, x = x_15)[name = tensor("op_674")]; - tensor var_677_begin_0 = const()[name = tensor("op_677_begin_0"), val = tensor([0, 1, 0])]; - tensor var_677_end_0 = const()[name = tensor("op_677_end_0"), val = tensor([1, 16, 256])]; - tensor var_677_end_mask_0 = const()[name = tensor("op_677_end_mask_0"), val = tensor([true, true, true])]; - tensor var_677 = slice_by_index(begin = var_677_begin_0, end = var_677_end_0, end_mask = var_677_end_mask_0, x = window_27)[name = tensor("op_677")]; + tensor window_27 = concat(axis = var_93, interleave = window_27_interleave_0, values = (var_735, var_732))[name = tensor("window_27")]; + tensor var_740_begin_0 = const()[name = tensor("op_740_begin_0"), val = tensor([0, 1, 0])]; + tensor var_740_end_0 = const()[name = tensor("op_740_end_0"), val = tensor([1, 2, 256])]; + tensor var_740_end_mask_0 = const()[name = tensor("op_740_end_mask_0"), val = tensor([true, false, true])]; + tensor var_740 = slice_by_index(begin = var_740_begin_0, end = var_740_end_0, end_mask = var_740_end_mask_0, x = x_15)[name = tensor("op_740")]; + tensor var_743_begin_0 = const()[name = tensor("op_743_begin_0"), val = tensor([0, 1, 0])]; + tensor var_743_end_0 = const()[name = tensor("op_743_end_0"), val = tensor([1, 16, 256])]; + tensor var_743_end_mask_0 = const()[name = tensor("op_743_end_mask_0"), val = tensor([true, true, true])]; + tensor var_743 = slice_by_index(begin = var_743_begin_0, end = var_743_end_0, end_mask = var_743_end_mask_0, x = window_27)[name = tensor("op_743")]; tensor window_29_interleave_0 = const()[name = tensor("window_29_interleave_0"), val = tensor(false)]; - tensor window_29 = concat(axis = var_27, interleave = window_29_interleave_0, values = (var_677, var_674))[name = tensor("window_29")]; - tensor var_682_begin_0 = const()[name = tensor("op_682_begin_0"), val = tensor([0, 2, 0])]; - tensor var_682_end_0 = const()[name = tensor("op_682_end_0"), val = tensor([1, 3, 256])]; - tensor var_682_end_mask_0 = const()[name = tensor("op_682_end_mask_0"), val = tensor([true, false, true])]; - tensor var_682 = slice_by_index(begin = var_682_begin_0, end = var_682_end_0, end_mask = var_682_end_mask_0, x = x_15)[name = tensor("op_682")]; - tensor var_685_begin_0 = const()[name = tensor("op_685_begin_0"), val = tensor([0, 1, 0])]; - tensor var_685_end_0 = const()[name = tensor("op_685_end_0"), val = tensor([1, 16, 256])]; - tensor var_685_end_mask_0 = const()[name = tensor("op_685_end_mask_0"), val = tensor([true, true, true])]; - tensor var_685 = slice_by_index(begin = var_685_begin_0, end = var_685_end_0, end_mask = var_685_end_mask_0, x = window_29)[name = tensor("op_685")]; + tensor window_29 = concat(axis = var_93, interleave = window_29_interleave_0, values = (var_743, var_740))[name = tensor("window_29")]; + tensor var_748_begin_0 = const()[name = tensor("op_748_begin_0"), val = tensor([0, 2, 0])]; + tensor var_748_end_0 = const()[name = tensor("op_748_end_0"), val = tensor([1, 3, 256])]; + tensor var_748_end_mask_0 = const()[name = tensor("op_748_end_mask_0"), val = tensor([true, false, true])]; + tensor var_748 = slice_by_index(begin = var_748_begin_0, end = var_748_end_0, end_mask = var_748_end_mask_0, x = x_15)[name = tensor("op_748")]; + tensor var_751_begin_0 = const()[name = tensor("op_751_begin_0"), val = tensor([0, 1, 0])]; + tensor var_751_end_0 = const()[name = tensor("op_751_end_0"), val = tensor([1, 16, 256])]; + tensor var_751_end_mask_0 = const()[name = tensor("op_751_end_mask_0"), val = tensor([true, true, true])]; + tensor var_751 = slice_by_index(begin = var_751_begin_0, end = var_751_end_0, end_mask = var_751_end_mask_0, x = window_29)[name = tensor("op_751")]; tensor window_31_interleave_0 = const()[name = tensor("window_31_interleave_0"), val = tensor(false)]; - tensor window_31 = concat(axis = var_27, interleave = window_31_interleave_0, values = (var_685, var_682))[name = tensor("window_31")]; - tensor var_690_begin_0 = const()[name = tensor("op_690_begin_0"), val = tensor([0, 3, 0])]; - tensor var_690_end_0 = const()[name = tensor("op_690_end_0"), val = tensor([1, 4, 256])]; - tensor var_690_end_mask_0 = const()[name = tensor("op_690_end_mask_0"), val = tensor([true, false, true])]; - tensor var_690 = slice_by_index(begin = var_690_begin_0, end = var_690_end_0, end_mask = var_690_end_mask_0, x = x_15)[name = tensor("op_690")]; - tensor var_693_begin_0 = const()[name = tensor("op_693_begin_0"), val = tensor([0, 1, 0])]; - tensor var_693_end_0 = const()[name = tensor("op_693_end_0"), val = tensor([1, 16, 256])]; - tensor var_693_end_mask_0 = const()[name = tensor("op_693_end_mask_0"), val = tensor([true, true, true])]; - tensor var_693 = slice_by_index(begin = var_693_begin_0, end = var_693_end_0, end_mask = var_693_end_mask_0, x = window_31)[name = tensor("op_693")]; + tensor window_31 = concat(axis = var_93, interleave = window_31_interleave_0, values = (var_751, var_748))[name = tensor("window_31")]; + tensor var_756_begin_0 = const()[name = tensor("op_756_begin_0"), val = tensor([0, 3, 0])]; + tensor var_756_end_0 = const()[name = tensor("op_756_end_0"), val = tensor([1, 4, 256])]; + tensor var_756_end_mask_0 = const()[name = tensor("op_756_end_mask_0"), val = tensor([true, false, true])]; + tensor var_756 = slice_by_index(begin = var_756_begin_0, end = var_756_end_0, end_mask = var_756_end_mask_0, x = x_15)[name = tensor("op_756")]; + tensor var_759_begin_0 = const()[name = tensor("op_759_begin_0"), val = tensor([0, 1, 0])]; + tensor var_759_end_0 = const()[name = tensor("op_759_end_0"), val = tensor([1, 16, 256])]; + tensor var_759_end_mask_0 = const()[name = tensor("op_759_end_mask_0"), val = tensor([true, true, true])]; + tensor var_759 = slice_by_index(begin = var_759_begin_0, end = var_759_end_0, end_mask = var_759_end_mask_0, x = window_31)[name = tensor("op_759")]; tensor window_33_interleave_0 = const()[name = tensor("window_33_interleave_0"), val = tensor(false)]; - tensor window_33 = concat(axis = var_27, interleave = window_33_interleave_0, values = (var_693, var_690))[name = tensor("window_33")]; - tensor var_698_begin_0 = const()[name = tensor("op_698_begin_0"), val = tensor([0, 4, 0])]; - tensor var_698_end_0 = const()[name = tensor("op_698_end_0"), val = tensor([1, 1, 256])]; - tensor var_698_end_mask_0 = const()[name = tensor("op_698_end_mask_0"), val = tensor([true, true, true])]; - tensor var_698 = slice_by_index(begin = var_698_begin_0, end = var_698_end_0, end_mask = var_698_end_mask_0, x = x_15)[name = tensor("op_698")]; - tensor var_701_begin_0 = const()[name = tensor("op_701_begin_0"), val = tensor([0, 1, 0])]; - tensor var_701_end_0 = const()[name = tensor("op_701_end_0"), val = tensor([1, 16, 256])]; - tensor var_701_end_mask_0 = const()[name = tensor("op_701_end_mask_0"), val = tensor([true, true, true])]; - tensor var_701 = slice_by_index(begin = var_701_begin_0, end = var_701_end_0, end_mask = var_701_end_mask_0, x = window_33)[name = tensor("op_701")]; + tensor window_33 = concat(axis = var_93, interleave = window_33_interleave_0, values = (var_759, var_756))[name = tensor("window_33")]; + tensor var_764_begin_0 = const()[name = tensor("op_764_begin_0"), val = tensor([0, 4, 0])]; + tensor var_764_end_0 = const()[name = tensor("op_764_end_0"), val = tensor([1, 1, 256])]; + tensor var_764_end_mask_0 = const()[name = tensor("op_764_end_mask_0"), val = tensor([true, true, true])]; + tensor var_764 = slice_by_index(begin = var_764_begin_0, end = var_764_end_0, end_mask = var_764_end_mask_0, x = x_15)[name = tensor("op_764")]; + tensor var_767_begin_0 = const()[name = tensor("op_767_begin_0"), val = tensor([0, 1, 0])]; + tensor var_767_end_0 = const()[name = tensor("op_767_end_0"), val = tensor([1, 16, 256])]; + tensor var_767_end_mask_0 = const()[name = tensor("op_767_end_mask_0"), val = tensor([true, true, true])]; + tensor var_767 = slice_by_index(begin = var_767_begin_0, end = var_767_end_0, end_mask = var_767_end_mask_0, x = window_33)[name = tensor("op_767")]; tensor window_35_interleave_0 = const()[name = tensor("window_35_interleave_0"), val = tensor(false)]; - tensor window_35 = concat(axis = var_27, interleave = window_35_interleave_0, values = (var_701, var_698))[name = tensor("window_35")]; - tensor input_101_interleave_0 = const()[name = tensor("input_101_interleave_0"), val = tensor(false)]; - tensor input_101 = concat(axis = var_24, interleave = input_101_interleave_0, values = (window_27, window_29, window_31, window_33, window_35))[name = tensor("input_101")]; + tensor window_35 = concat(axis = var_93, interleave = window_35_interleave_0, values = (var_767, var_764))[name = tensor("window_35")]; + tensor input_103_interleave_0 = const()[name = tensor("input_103_interleave_0"), val = tensor(false)]; + tensor input_103 = concat(axis = var_79, interleave = input_103_interleave_0, values = (window_27, window_29, window_31, window_33, window_35))[name = tensor("input_103")]; tensor x_17_axes_0 = const()[name = tensor("x_17_axes_0"), val = tensor([-1])]; - tensor x_17 = layer_norm(axes = x_17_axes_0, beta = encoder_conv_module_2_sequential_0_bias, epsilon = var_29, gamma = encoder_conv_module_2_sequential_0_weight, x = input_101)[name = tensor("x_17")]; - tensor input_103_perm_0 = const()[name = tensor("input_103_perm_0"), val = tensor([0, 2, 1])]; + tensor x_17 = layer_norm(axes = x_17_axes_0, beta = inner_encoder_conv_module_2_sequential_0_bias, epsilon = var_76, gamma = inner_encoder_conv_module_2_sequential_0_weight, x = input_103)[name = tensor("x_17")]; + tensor input_105_perm_0 = const()[name = tensor("input_105_perm_0"), val = tensor([0, 2, 1])]; tensor inputs_23_pad_type_0 = const()[name = tensor("inputs_23_pad_type_0"), val = tensor("valid")]; tensor inputs_23_strides_0 = const()[name = tensor("inputs_23_strides_0"), val = tensor([1])]; tensor inputs_23_pad_0 = const()[name = tensor("inputs_23_pad_0"), val = tensor([0, 0])]; tensor inputs_23_dilations_0 = const()[name = tensor("inputs_23_dilations_0"), val = tensor([1])]; tensor inputs_23_groups_0 = const()[name = tensor("inputs_23_groups_0"), val = tensor(1)]; - tensor input_103 = transpose(perm = input_103_perm_0, x = x_17)[name = tensor("transpose_40")]; - tensor inputs_23 = conv(bias = encoder_conv_module_2_sequential_2_conv_bias, dilations = inputs_23_dilations_0, groups = inputs_23_groups_0, pad = inputs_23_pad_0, pad_type = inputs_23_pad_type_0, strides = inputs_23_strides_0, weight = encoder_conv_module_2_sequential_2_conv_weight, x = input_103)[name = tensor("inputs_23")]; - tensor var_726_split_sizes_0 = const()[name = tensor("op_726_split_sizes_0"), val = tensor([256, 256])]; - tensor var_726_axis_0 = const()[name = tensor("op_726_axis_0"), val = tensor(1)]; - tensor var_726_0, tensor var_726_1 = split(axis = var_726_axis_0, split_sizes = var_726_split_sizes_0, x = inputs_23)[name = tensor("op_726")]; - tensor var_728 = sigmoid(x = var_726_1)[name = tensor("op_728")]; - tensor inputs_25 = mul(x = var_726_0, y = var_728)[name = tensor("inputs_25")]; + tensor input_105 = transpose(perm = input_105_perm_0, x = x_17)[name = tensor("transpose_40")]; + tensor inputs_23 = conv(bias = inner_encoder_conv_module_2_sequential_2_conv_bias, dilations = inputs_23_dilations_0, groups = inputs_23_groups_0, pad = inputs_23_pad_0, pad_type = inputs_23_pad_type_0, strides = inputs_23_strides_0, weight = inner_encoder_conv_module_2_sequential_2_conv_weight, x = input_105)[name = tensor("inputs_23")]; + tensor var_792_split_sizes_0 = const()[name = tensor("op_792_split_sizes_0"), val = tensor([256, 256])]; + tensor var_792_axis_0 = const()[name = tensor("op_792_axis_0"), val = tensor(1)]; + tensor var_792_0, tensor var_792_1 = split(axis = var_792_axis_0, split_sizes = var_792_split_sizes_0, x = inputs_23)[name = tensor("op_792")]; + tensor var_794 = sigmoid(x = var_792_1)[name = tensor("op_794")]; + tensor inputs_25 = mul(x = var_792_0, y = var_794)[name = tensor("inputs_25")]; tensor outputs_aug_5_pad_type_0 = const()[name = tensor("outputs_aug_5_pad_type_0"), val = tensor("custom")]; tensor outputs_aug_5_pad_0 = const()[name = tensor("outputs_aug_5_pad_0"), val = tensor([15, 15])]; tensor outputs_aug_5_groups_0 = const()[name = tensor("outputs_aug_5_groups_0"), val = tensor(256)]; tensor outputs_aug_5_strides_0 = const()[name = tensor("outputs_aug_5_strides_0"), val = tensor([1])]; tensor outputs_aug_5_dilations_0 = const()[name = tensor("outputs_aug_5_dilations_0"), val = tensor([1])]; - tensor outputs_aug_5 = conv(dilations = outputs_aug_5_dilations_0, groups = outputs_aug_5_groups_0, pad = outputs_aug_5_pad_0, pad_type = outputs_aug_5_pad_type_0, strides = outputs_aug_5_strides_0, weight = encoder_conv_module_2_sequential_4_conv_weight, x = inputs_25)[name = tensor("outputs_aug_5")]; - tensor input_105_begin_0 = const()[name = tensor("input_105_begin_0"), val = tensor([0, 0, 0])]; - tensor input_105_end_0 = const()[name = tensor("input_105_end_0"), val = tensor([5, 256, 16])]; - tensor input_105_end_mask_0 = const()[name = tensor("input_105_end_mask_0"), val = tensor([true, true, false])]; - tensor input_105 = slice_by_index(begin = input_105_begin_0, end = input_105_end_0, end_mask = input_105_end_mask_0, x = outputs_aug_5)[name = tensor("input_105")]; - tensor inputs_27 = batch_norm(beta = encoder_conv_module_2_sequential_5_bias, epsilon = var_29, gamma = encoder_conv_module_2_sequential_5_weight, mean = encoder_conv_module_2_sequential_5_running_mean, variance = encoder_conv_module_2_sequential_5_running_var, x = input_105)[name = tensor("inputs_27")]; - tensor input_107 = silu(x = inputs_27)[name = tensor("input_107")]; - tensor input_109_pad_type_0 = const()[name = tensor("input_109_pad_type_0"), val = tensor("valid")]; - tensor input_109_strides_0 = const()[name = tensor("input_109_strides_0"), val = tensor([1])]; - tensor input_109_pad_0 = const()[name = tensor("input_109_pad_0"), val = tensor([0, 0])]; - tensor input_109_dilations_0 = const()[name = tensor("input_109_dilations_0"), val = tensor([1])]; - tensor input_109_groups_0 = const()[name = tensor("input_109_groups_0"), val = tensor(1)]; - tensor input_109 = conv(bias = encoder_conv_module_2_sequential_7_conv_bias, dilations = input_109_dilations_0, groups = input_109_groups_0, pad = input_109_pad_0, pad_type = input_109_pad_type_0, strides = input_109_strides_0, weight = encoder_conv_module_2_sequential_7_conv_weight, x = input_107)[name = tensor("input_109")]; + tensor outputs_aug_5 = conv(dilations = outputs_aug_5_dilations_0, groups = outputs_aug_5_groups_0, pad = outputs_aug_5_pad_0, pad_type = outputs_aug_5_pad_type_0, strides = outputs_aug_5_strides_0, weight = inner_encoder_conv_module_2_sequential_4_conv_weight, x = inputs_25)[name = tensor("outputs_aug_5")]; + tensor input_107_begin_0 = const()[name = tensor("input_107_begin_0"), val = tensor([0, 0, 0])]; + tensor input_107_end_0 = const()[name = tensor("input_107_end_0"), val = tensor([5, 256, 16])]; + tensor input_107_end_mask_0 = const()[name = tensor("input_107_end_mask_0"), val = tensor([true, true, false])]; + tensor input_107 = slice_by_index(begin = input_107_begin_0, end = input_107_end_0, end_mask = input_107_end_mask_0, x = outputs_aug_5)[name = tensor("input_107")]; + tensor inputs_27 = batch_norm(beta = inner_encoder_conv_module_2_sequential_5_bias, epsilon = var_76, gamma = inner_encoder_conv_module_2_sequential_5_weight, mean = inner_encoder_conv_module_2_sequential_5_running_mean, variance = inner_encoder_conv_module_2_sequential_5_running_var, x = input_107)[name = tensor("inputs_27")]; + tensor input_109 = silu(x = inputs_27)[name = tensor("input_109")]; + tensor input_111_pad_type_0 = const()[name = tensor("input_111_pad_type_0"), val = tensor("valid")]; + tensor input_111_strides_0 = const()[name = tensor("input_111_strides_0"), val = tensor([1])]; + tensor input_111_pad_0 = const()[name = tensor("input_111_pad_0"), val = tensor([0, 0])]; + tensor input_111_dilations_0 = const()[name = tensor("input_111_dilations_0"), val = tensor([1])]; + tensor input_111_groups_0 = const()[name = tensor("input_111_groups_0"), val = tensor(1)]; + tensor input_111 = conv(bias = inner_encoder_conv_module_2_sequential_7_conv_bias, dilations = input_111_dilations_0, groups = input_111_groups_0, pad = input_111_pad_0, pad_type = input_111_pad_type_0, strides = input_111_strides_0, weight = inner_encoder_conv_module_2_sequential_7_conv_weight, x = input_109)[name = tensor("input_111")]; tensor conv_out_5_perm_0 = const()[name = tensor("conv_out_5_perm_0"), val = tensor([0, 2, 1])]; - tensor var_759_begin_0 = const()[name = tensor("op_759_begin_0"), val = tensor([0, -1, 0])]; - tensor var_759_end_0 = const()[name = tensor("op_759_end_0"), val = tensor([5, 16, 256])]; - tensor var_759_end_mask_0 = const()[name = tensor("op_759_end_mask_0"), val = tensor([true, true, true])]; - tensor conv_out_5 = transpose(perm = conv_out_5_perm_0, x = input_109)[name = tensor("transpose_39")]; - tensor var_759 = slice_by_index(begin = var_759_begin_0, end = var_759_end_0, end_mask = var_759_end_mask_0, x = conv_out_5)[name = tensor("op_759")]; - tensor var_761_perm_0 = const()[name = tensor("op_761_perm_0"), val = tensor([1, 0, 2])]; - tensor var_761 = transpose(perm = var_761_perm_0, x = var_759)[name = tensor("transpose_38")]; - tensor input_111 = add(x = x_15, y = var_761)[name = tensor("input_111")]; - tensor input_113_axes_0 = const()[name = tensor("input_113_axes_0"), val = tensor([-1])]; - tensor input_113 = layer_norm(axes = input_113_axes_0, beta = encoder_ffn2_2_module_sequential_0_bias, epsilon = var_29, gamma = encoder_ffn2_2_module_sequential_0_weight, x = input_111)[name = tensor("input_113")]; - tensor inputs_29 = linear(bias = encoder_ffn2_2_module_sequential_1_linear_bias, weight = encoder_ffn2_2_module_sequential_1_linear_weight, x = input_113)[name = tensor("linear_26")]; - tensor input_115 = silu(x = inputs_29)[name = tensor("input_115")]; - tensor input_119 = linear(bias = encoder_ffn2_2_module_sequential_4_linear_bias, weight = encoder_ffn2_2_module_sequential_4_linear_weight, x = input_115)[name = tensor("linear_27")]; - tensor var_784 = const()[name = tensor("op_784"), val = tensor(0x1p-1)]; - tensor var_785 = mul(x = input_119, y = var_784)[name = tensor("op_785")]; - tensor input_121 = add(x = var_785, y = input_111)[name = tensor("input_121")]; - tensor input_123_axes_0 = const()[name = tensor("input_123_axes_0"), val = tensor([-1])]; - tensor input_123 = layer_norm(axes = input_123_axes_0, beta = encoder_layer_norm_2_bias, epsilon = var_29, gamma = encoder_layer_norm_2_weight, x = input_121)[name = tensor("input_123")]; + tensor var_825_begin_0 = const()[name = tensor("op_825_begin_0"), val = tensor([0, -1, 0])]; + tensor var_825_end_0 = const()[name = tensor("op_825_end_0"), val = tensor([5, 16, 256])]; + tensor var_825_end_mask_0 = const()[name = tensor("op_825_end_mask_0"), val = tensor([true, true, true])]; + tensor conv_out_5 = transpose(perm = conv_out_5_perm_0, x = input_111)[name = tensor("transpose_39")]; + tensor var_825 = slice_by_index(begin = var_825_begin_0, end = var_825_end_0, end_mask = var_825_end_mask_0, x = conv_out_5)[name = tensor("op_825")]; + tensor var_827_perm_0 = const()[name = tensor("op_827_perm_0"), val = tensor([1, 0, 2])]; + tensor var_827 = transpose(perm = var_827_perm_0, x = var_825)[name = tensor("transpose_38")]; + tensor input_113 = add(x = x_15, y = var_827)[name = tensor("input_113")]; + tensor input_115_axes_0 = const()[name = tensor("input_115_axes_0"), val = tensor([-1])]; + tensor input_115 = layer_norm(axes = input_115_axes_0, beta = inner_encoder_ffn2_2_module_sequential_0_bias, epsilon = var_76, gamma = inner_encoder_ffn2_2_module_sequential_0_weight, x = input_113)[name = tensor("input_115")]; + tensor inputs_29 = linear(bias = inner_encoder_ffn2_2_module_sequential_1_linear_bias, weight = inner_encoder_ffn2_2_module_sequential_1_linear_weight, x = input_115)[name = tensor("linear_26")]; + tensor input_117 = silu(x = inputs_29)[name = tensor("input_117")]; + tensor input_121 = linear(bias = inner_encoder_ffn2_2_module_sequential_4_linear_bias, weight = inner_encoder_ffn2_2_module_sequential_4_linear_weight, x = input_117)[name = tensor("linear_27")]; + tensor var_850 = const()[name = tensor("op_850"), val = tensor(0x1p-1)]; + tensor var_851 = mul(x = input_121, y = var_850)[name = tensor("op_851")]; + tensor input_123 = add(x = var_851, y = input_113)[name = tensor("input_123")]; tensor input_125_axes_0 = const()[name = tensor("input_125_axes_0"), val = tensor([-1])]; - tensor input_125 = layer_norm(axes = input_125_axes_0, beta = encoder_ffn1_3_module_sequential_0_bias, epsilon = var_29, gamma = encoder_ffn1_3_module_sequential_0_weight, x = input_123)[name = tensor("input_125")]; - tensor inputs_31 = linear(bias = encoder_ffn1_3_module_sequential_1_linear_bias, weight = encoder_ffn1_3_module_sequential_1_linear_weight, x = input_125)[name = tensor("linear_28")]; - tensor input_127 = silu(x = inputs_31)[name = tensor("input_127")]; - tensor input_131 = linear(bias = encoder_ffn1_3_module_sequential_4_linear_bias, weight = encoder_ffn1_3_module_sequential_4_linear_weight, x = input_127)[name = tensor("linear_29")]; - tensor var_814 = const()[name = tensor("op_814"), val = tensor(0x1p-1)]; - tensor var_815 = mul(x = input_131, y = var_814)[name = tensor("op_815")]; - tensor input_133 = add(x = var_815, y = input_123)[name = tensor("input_133")]; + tensor input_125 = layer_norm(axes = input_125_axes_0, beta = inner_encoder_layer_norm_2_bias, epsilon = var_76, gamma = inner_encoder_layer_norm_2_weight, x = input_123)[name = tensor("input_125")]; + tensor input_127_axes_0 = const()[name = tensor("input_127_axes_0"), val = tensor([-1])]; + tensor input_127 = layer_norm(axes = input_127_axes_0, beta = inner_encoder_ffn1_3_module_sequential_0_bias, epsilon = var_76, gamma = inner_encoder_ffn1_3_module_sequential_0_weight, x = input_125)[name = tensor("input_127")]; + tensor inputs_31 = linear(bias = inner_encoder_ffn1_3_module_sequential_1_linear_bias, weight = inner_encoder_ffn1_3_module_sequential_1_linear_weight, x = input_127)[name = tensor("linear_28")]; + tensor input_129 = silu(x = inputs_31)[name = tensor("input_129")]; + tensor input_133 = linear(bias = inner_encoder_ffn1_3_module_sequential_4_linear_bias, weight = inner_encoder_ffn1_3_module_sequential_4_linear_weight, x = input_129)[name = tensor("linear_29")]; + tensor var_880 = const()[name = tensor("op_880"), val = tensor(0x1p-1)]; + tensor var_881 = mul(x = input_133, y = var_880)[name = tensor("op_881")]; + tensor input_135 = add(x = var_881, y = input_125)[name = tensor("input_135")]; tensor x_19_axes_0 = const()[name = tensor("x_19_axes_0"), val = tensor([-1])]; - tensor x_19 = layer_norm(axes = x_19_axes_0, beta = encoder_ret_lns_3_bias, epsilon = var_29, gamma = encoder_ret_lns_3_weight, x = input_133)[name = tensor("x_19")]; + tensor x_19 = layer_norm(axes = x_19_axes_0, beta = inner_encoder_ret_lns_3_bias, epsilon = var_76, gamma = inner_encoder_ret_lns_3_weight, x = input_135)[name = tensor("x_19")]; tensor prev_kv_7_begin_0 = const()[name = tensor("prev_kv_7_begin_0"), val = tensor([3, 0, 0, 0, 0])]; tensor prev_kv_7_end_0 = const()[name = tensor("prev_kv_7_end_0"), val = tensor([4, 1, 4, 64, 64])]; tensor prev_kv_7_end_mask_0 = const()[name = tensor("prev_kv_7_end_mask_0"), val = tensor([false, true, true, true, true])]; @@ -800,219 +826,212 @@ program(1.0) tensor prev_scale_7_end_mask_0 = const()[name = tensor("prev_scale_7_end_mask_0"), val = tensor([false, true])]; tensor prev_scale_7_squeeze_mask_0 = const()[name = tensor("prev_scale_7_squeeze_mask_0"), val = tensor([true, false])]; tensor prev_scale_7 = slice_by_index(begin = prev_scale_7_begin_0, end = prev_scale_7_end_0, end_mask = prev_scale_7_end_mask_0, squeeze_mask = prev_scale_7_squeeze_mask_0, x = enc_scale)[name = tensor("prev_scale_7")]; - tensor var_829 = linear(bias = encoder_q_proj_3_bias, weight = encoder_q_proj_3_weight, x = x_19)[name = tensor("linear_30")]; - tensor var_830 = const()[name = tensor("op_830"), val = tensor([1, 5, 4, 64])]; - tensor var_831 = reshape(shape = var_830, x = var_829)[name = tensor("op_831")]; + tensor var_895 = linear(bias = inner_encoder_q_proj_3_bias, weight = inner_encoder_q_proj_3_weight, x = x_19)[name = tensor("linear_30")]; + tensor var_896 = const()[name = tensor("op_896"), val = tensor([1, 5, 4, 64])]; + tensor var_897 = reshape(shape = var_896, x = var_895)[name = tensor("op_897")]; tensor q_7_perm_0 = const()[name = tensor("q_7_perm_0"), val = tensor([0, 2, 1, 3])]; - tensor var_835 = linear(bias = encoder_k_proj_3_bias, weight = encoder_k_proj_3_weight, x = x_19)[name = tensor("linear_31")]; - tensor var_836 = const()[name = tensor("op_836"), val = tensor(0x1p-3)]; - tensor var_837 = mul(x = var_835, y = var_836)[name = tensor("op_837")]; - tensor var_838 = const()[name = tensor("op_838"), val = tensor([1, 5, 4, 64])]; - tensor var_839 = reshape(shape = var_838, x = var_837)[name = tensor("op_839")]; + tensor var_901 = linear(bias = inner_encoder_k_proj_3_bias, weight = inner_encoder_k_proj_3_weight, x = x_19)[name = tensor("linear_31")]; + tensor var_902 = const()[name = tensor("op_902"), val = tensor(0x1p-3)]; + tensor var_903 = mul(x = var_901, y = var_902)[name = tensor("op_903")]; + tensor var_904 = const()[name = tensor("op_904"), val = tensor([1, 5, 4, 64])]; + tensor var_905 = reshape(shape = var_904, x = var_903)[name = tensor("op_905")]; tensor k_7_perm_0 = const()[name = tensor("k_7_perm_0"), val = tensor([0, 2, 1, 3])]; - tensor var_843 = linear(bias = encoder_v_proj_3_bias, weight = encoder_v_proj_3_weight, x = x_19)[name = tensor("linear_32")]; - tensor var_844 = const()[name = tensor("op_844"), val = tensor([1, 5, 4, 64])]; - tensor var_845 = reshape(shape = var_844, x = var_843)[name = tensor("op_845")]; + tensor var_909 = linear(bias = inner_encoder_v_proj_3_bias, weight = inner_encoder_v_proj_3_weight, x = x_19)[name = tensor("linear_32")]; + tensor var_910 = const()[name = tensor("op_910"), val = tensor([1, 5, 4, 64])]; + tensor var_911 = reshape(shape = var_910, x = var_909)[name = tensor("op_911")]; tensor v_7_perm_0 = const()[name = tensor("v_7_perm_0"), val = tensor([0, 2, 1, 3])]; - tensor input_137 = linear(bias = encoder_g_proj_3_bias, weight = encoder_g_proj_3_weight, x = x_19)[name = tensor("linear_33")]; + tensor input_139 = linear(bias = inner_encoder_g_proj_3_bias, weight = inner_encoder_g_proj_3_weight, x = x_19)[name = tensor("linear_33")]; tensor sqrt_s0_7 = sqrt(x = prev_scale_7)[name = tensor("sqrt_s0_7")]; - tensor s_t_7 = add(x = prev_scale_7, y = encoder__t_index)[name = tensor("s_t_7")]; + tensor s_t_7 = add(x = prev_scale_7, y = inner_encoder__t_index)[name = tensor("s_t_7")]; tensor sqrt_s_t_7 = sqrt(x = s_t_7)[name = tensor("sqrt_s_t_7")]; tensor qk_7_transpose_x_1 = const()[name = tensor("qk_7_transpose_x_1"), val = tensor(false)]; tensor qk_7_transpose_y_1 = const()[name = tensor("qk_7_transpose_y_1"), val = tensor(true)]; - tensor k_7 = transpose(perm = k_7_perm_0, x = var_839)[name = tensor("transpose_36")]; - tensor q_7 = transpose(perm = q_7_perm_0, x = var_831)[name = tensor("transpose_37")]; + tensor k_7 = transpose(perm = k_7_perm_0, x = var_905)[name = tensor("transpose_36")]; + tensor q_7 = transpose(perm = q_7_perm_0, x = var_897)[name = tensor("transpose_37")]; tensor qk_7 = matmul(transpose_x = qk_7_transpose_x_1, transpose_y = qk_7_transpose_y_1, x = q_7, y = k_7)[name = tensor("qk_7")]; - tensor var_855 = const()[name = tensor("op_855"), val = tensor([5, 1])]; - tensor var_856 = reshape(shape = var_855, x = sqrt_s_t_7)[name = tensor("op_856")]; - tensor M_7 = real_div(x = encoder__causal_mask, y = var_856)[name = tensor("M_7")]; - tensor var_858 = mul(x = qk_7, y = M_7)[name = tensor("op_858")]; + tensor var_921 = const()[name = tensor("op_921"), val = tensor([5, 1])]; + tensor var_922 = reshape(shape = var_921, x = sqrt_s_t_7)[name = tensor("op_922")]; + tensor M_7 = real_div(x = inner_encoder__causal_mask, y = var_922)[name = tensor("M_7")]; + tensor var_924 = mul(x = qk_7, y = M_7)[name = tensor("op_924")]; tensor inner_7_transpose_x_0 = const()[name = tensor("inner_7_transpose_x_0"), val = tensor(false)]; tensor inner_7_transpose_y_0 = const()[name = tensor("inner_7_transpose_y_0"), val = tensor(false)]; - tensor v_7 = transpose(perm = v_7_perm_0, x = var_845)[name = tensor("transpose_35")]; - tensor inner_7 = matmul(transpose_x = inner_7_transpose_x_0, transpose_y = inner_7_transpose_y_0, x = var_858, y = v_7)[name = tensor("inner_7")]; - tensor var_860_transpose_x_0 = const()[name = tensor("op_860_transpose_x_0"), val = tensor(false)]; - tensor var_860_transpose_y_0 = const()[name = tensor("op_860_transpose_y_0"), val = tensor(false)]; - tensor var_860 = matmul(transpose_x = var_860_transpose_x_0, transpose_y = var_860_transpose_y_0, x = q_7, y = prev_kv_7)[name = tensor("op_860")]; - tensor var_861 = real_div(x = sqrt_s0_7, y = sqrt_s_t_7)[name = tensor("op_861")]; - tensor var_862 = const()[name = tensor("op_862"), val = tensor([1, 1, 5, 1])]; - tensor var_863 = reshape(shape = var_862, x = var_861)[name = tensor("op_863")]; - tensor cross_7 = mul(x = var_860, y = var_863)[name = tensor("cross_7")]; + tensor v_7 = transpose(perm = v_7_perm_0, x = var_911)[name = tensor("transpose_35")]; + tensor inner_7 = matmul(transpose_x = inner_7_transpose_x_0, transpose_y = inner_7_transpose_y_0, x = var_924, y = v_7)[name = tensor("inner_7")]; + tensor var_926_transpose_x_0 = const()[name = tensor("op_926_transpose_x_0"), val = tensor(false)]; + tensor var_926_transpose_y_0 = const()[name = tensor("op_926_transpose_y_0"), val = tensor(false)]; + tensor var_926 = matmul(transpose_x = var_926_transpose_x_0, transpose_y = var_926_transpose_y_0, x = q_7, y = prev_kv_7)[name = tensor("op_926")]; + tensor var_927 = real_div(x = sqrt_s0_7, y = sqrt_s_t_7)[name = tensor("op_927")]; + tensor var_928 = const()[name = tensor("op_928"), val = tensor([1, 1, 5, 1])]; + tensor var_929 = reshape(shape = var_928, x = var_927)[name = tensor("op_929")]; + tensor cross_7 = mul(x = var_926, y = var_929)[name = tensor("cross_7")]; tensor out_19 = add(x = inner_7, y = cross_7)[name = tensor("out_19")]; - tensor var_866 = mul(x = prev_kv_7, y = sqrt_s0_7)[name = tensor("op_866")]; - tensor var_868_transpose_x_1 = const()[name = tensor("op_868_transpose_x_1"), val = tensor(true)]; - tensor var_868_transpose_y_1 = const()[name = tensor("op_868_transpose_y_1"), val = tensor(false)]; - tensor var_868 = matmul(transpose_x = var_868_transpose_x_1, transpose_y = var_868_transpose_y_1, x = k_7, y = v_7)[name = tensor("op_868")]; - tensor new_kv_unnorm_7 = add(x = var_866, y = var_868)[name = tensor("new_kv_unnorm_7")]; - tensor var_870 = const()[name = tensor("op_870"), val = tensor(0x1.4p+2)]; - tensor new_scale_7 = add(x = prev_scale_7, y = var_870)[name = tensor("new_scale_7")]; - tensor var_872 = sqrt(x = new_scale_7)[name = tensor("op_872")]; - tensor nkv_1 = real_div(x = new_kv_unnorm_7, y = var_872)[name = tensor("nkv_1")]; - tensor var_874_perm_0 = const()[name = tensor("op_874_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_932 = mul(x = prev_kv_7, y = sqrt_s0_7)[name = tensor("op_932")]; + tensor var_934_transpose_x_1 = const()[name = tensor("op_934_transpose_x_1"), val = tensor(true)]; + tensor var_934_transpose_y_1 = const()[name = tensor("op_934_transpose_y_1"), val = tensor(false)]; + tensor var_934 = matmul(transpose_x = var_934_transpose_x_1, transpose_y = var_934_transpose_y_1, x = k_7, y = v_7)[name = tensor("op_934")]; + tensor new_kv_unnorm_7 = add(x = var_932, y = var_934)[name = tensor("new_kv_unnorm_7")]; + tensor var_936 = const()[name = tensor("op_936"), val = tensor(0x1.4p+2)]; + tensor new_scale_7 = add(x = prev_scale_7, y = var_936)[name = tensor("new_scale_7")]; + tensor var_938 = sqrt(x = new_scale_7)[name = tensor("op_938")]; + tensor nkv_1 = real_div(x = new_kv_unnorm_7, y = var_938)[name = tensor("nkv_1")]; + tensor var_940_perm_0 = const()[name = tensor("op_940_perm_0"), val = tensor([0, 2, 1, 3])]; tensor out_21_axes_0 = const()[name = tensor("out_21_axes_0"), val = tensor([-1])]; - tensor var_874 = transpose(perm = var_874_perm_0, x = out_19)[name = tensor("transpose_34")]; - tensor out_21 = layer_norm(axes = out_21_axes_0, epsilon = var_18, x = var_874)[name = tensor("out_21")]; - tensor var_878 = const()[name = tensor("op_878"), val = tensor([1, 5, 256])]; - tensor out_23 = reshape(shape = var_878, x = out_21)[name = tensor("out_23")]; - tensor var_880 = silu(x = input_137)[name = tensor("op_880")]; - tensor input_139 = mul(x = var_880, y = out_23)[name = tensor("input_139")]; - tensor ret_out_7 = linear(bias = encoder_out_proj_3_bias, weight = encoder_out_proj_3_weight, x = input_139)[name = tensor("linear_34")]; - tensor x_21 = add(x = input_133, y = ret_out_7)[name = tensor("x_21")]; + tensor var_940 = transpose(perm = var_940_perm_0, x = out_19)[name = tensor("transpose_34")]; + tensor out_21 = layer_norm(axes = out_21_axes_0, epsilon = var_84, x = var_940)[name = tensor("out_21")]; + tensor var_944 = const()[name = tensor("op_944"), val = tensor([1, 5, 256])]; + tensor out_23 = reshape(shape = var_944, x = out_21)[name = tensor("out_23")]; + tensor var_946 = silu(x = input_139)[name = tensor("op_946")]; + tensor input_141 = mul(x = var_946, y = out_23)[name = tensor("input_141")]; + tensor ret_out_7 = linear(bias = inner_encoder_out_proj_3_bias, weight = inner_encoder_out_proj_3_weight, x = input_141)[name = tensor("linear_34")]; + tensor x_21 = add(x = input_135, y = ret_out_7)[name = tensor("x_21")]; tensor window_37_begin_0 = const()[name = tensor("window_37_begin_0"), val = tensor([3, 0, 0, 0])]; tensor window_37_end_0 = const()[name = tensor("window_37_end_0"), val = tensor([4, 1, 16, 256])]; tensor window_37_end_mask_0 = const()[name = tensor("window_37_end_mask_0"), val = tensor([false, true, true, true])]; tensor window_37_squeeze_mask_0 = const()[name = tensor("window_37_squeeze_mask_0"), val = tensor([true, false, false, false])]; tensor window_37 = slice_by_index(begin = window_37_begin_0, end = window_37_end_0, end_mask = window_37_end_mask_0, squeeze_mask = window_37_squeeze_mask_0, x = enc_conv_cache)[name = tensor("window_37")]; - tensor var_888_begin_0 = const()[name = tensor("op_888_begin_0"), val = tensor([0, 0, 0])]; - tensor var_888_end_0 = const()[name = tensor("op_888_end_0"), val = tensor([1, 1, 256])]; - tensor var_888_end_mask_0 = const()[name = tensor("op_888_end_mask_0"), val = tensor([true, false, true])]; - tensor var_888 = slice_by_index(begin = var_888_begin_0, end = var_888_end_0, end_mask = var_888_end_mask_0, x = x_21)[name = tensor("op_888")]; - tensor var_891_begin_0 = const()[name = tensor("op_891_begin_0"), val = tensor([0, 1, 0])]; - tensor var_891_end_0 = const()[name = tensor("op_891_end_0"), val = tensor([1, 16, 256])]; - tensor var_891_end_mask_0 = const()[name = tensor("op_891_end_mask_0"), val = tensor([true, true, true])]; - tensor var_891 = slice_by_index(begin = var_891_begin_0, end = var_891_end_0, end_mask = var_891_end_mask_0, x = window_37)[name = tensor("op_891")]; + tensor var_954_begin_0 = const()[name = tensor("op_954_begin_0"), val = tensor([0, 0, 0])]; + tensor var_954_end_0 = const()[name = tensor("op_954_end_0"), val = tensor([1, 1, 256])]; + tensor var_954_end_mask_0 = const()[name = tensor("op_954_end_mask_0"), val = tensor([true, false, true])]; + tensor var_954 = slice_by_index(begin = var_954_begin_0, end = var_954_end_0, end_mask = var_954_end_mask_0, x = x_21)[name = tensor("op_954")]; + tensor var_957_begin_0 = const()[name = tensor("op_957_begin_0"), val = tensor([0, 1, 0])]; + tensor var_957_end_0 = const()[name = tensor("op_957_end_0"), val = tensor([1, 16, 256])]; + tensor var_957_end_mask_0 = const()[name = tensor("op_957_end_mask_0"), val = tensor([true, true, true])]; + tensor var_957 = slice_by_index(begin = var_957_begin_0, end = var_957_end_0, end_mask = var_957_end_mask_0, x = window_37)[name = tensor("op_957")]; tensor window_39_interleave_0 = const()[name = tensor("window_39_interleave_0"), val = tensor(false)]; - tensor window_39 = concat(axis = var_27, interleave = window_39_interleave_0, values = (var_891, var_888))[name = tensor("window_39")]; - tensor var_896_begin_0 = const()[name = tensor("op_896_begin_0"), val = tensor([0, 1, 0])]; - tensor var_896_end_0 = const()[name = tensor("op_896_end_0"), val = tensor([1, 2, 256])]; - tensor var_896_end_mask_0 = const()[name = tensor("op_896_end_mask_0"), val = tensor([true, false, true])]; - tensor var_896 = slice_by_index(begin = var_896_begin_0, end = var_896_end_0, end_mask = var_896_end_mask_0, x = x_21)[name = tensor("op_896")]; - tensor var_899_begin_0 = const()[name = tensor("op_899_begin_0"), val = tensor([0, 1, 0])]; - tensor var_899_end_0 = const()[name = tensor("op_899_end_0"), val = tensor([1, 16, 256])]; - tensor var_899_end_mask_0 = const()[name = tensor("op_899_end_mask_0"), val = tensor([true, true, true])]; - tensor var_899 = slice_by_index(begin = var_899_begin_0, end = var_899_end_0, end_mask = var_899_end_mask_0, x = window_39)[name = tensor("op_899")]; + tensor window_39 = concat(axis = var_93, interleave = window_39_interleave_0, values = (var_957, var_954))[name = tensor("window_39")]; + tensor var_962_begin_0 = const()[name = tensor("op_962_begin_0"), val = tensor([0, 1, 0])]; + tensor var_962_end_0 = const()[name = tensor("op_962_end_0"), val = tensor([1, 2, 256])]; + tensor var_962_end_mask_0 = const()[name = tensor("op_962_end_mask_0"), val = tensor([true, false, true])]; + tensor var_962 = slice_by_index(begin = var_962_begin_0, end = var_962_end_0, end_mask = var_962_end_mask_0, x = x_21)[name = tensor("op_962")]; + tensor var_965_begin_0 = const()[name = tensor("op_965_begin_0"), val = tensor([0, 1, 0])]; + tensor var_965_end_0 = const()[name = tensor("op_965_end_0"), val = tensor([1, 16, 256])]; + tensor var_965_end_mask_0 = const()[name = tensor("op_965_end_mask_0"), val = tensor([true, true, true])]; + tensor var_965 = slice_by_index(begin = var_965_begin_0, end = var_965_end_0, end_mask = var_965_end_mask_0, x = window_39)[name = tensor("op_965")]; tensor window_41_interleave_0 = const()[name = tensor("window_41_interleave_0"), val = tensor(false)]; - tensor window_41 = concat(axis = var_27, interleave = window_41_interleave_0, values = (var_899, var_896))[name = tensor("window_41")]; - tensor var_904_begin_0 = const()[name = tensor("op_904_begin_0"), val = tensor([0, 2, 0])]; - tensor var_904_end_0 = const()[name = tensor("op_904_end_0"), val = tensor([1, 3, 256])]; - tensor var_904_end_mask_0 = const()[name = tensor("op_904_end_mask_0"), val = tensor([true, false, true])]; - tensor var_904 = slice_by_index(begin = var_904_begin_0, end = var_904_end_0, end_mask = var_904_end_mask_0, x = x_21)[name = tensor("op_904")]; - tensor var_907_begin_0 = const()[name = tensor("op_907_begin_0"), val = tensor([0, 1, 0])]; - tensor var_907_end_0 = const()[name = tensor("op_907_end_0"), val = tensor([1, 16, 256])]; - tensor var_907_end_mask_0 = const()[name = tensor("op_907_end_mask_0"), val = tensor([true, true, true])]; - tensor var_907 = slice_by_index(begin = var_907_begin_0, end = var_907_end_0, end_mask = var_907_end_mask_0, x = window_41)[name = tensor("op_907")]; + tensor window_41 = concat(axis = var_93, interleave = window_41_interleave_0, values = (var_965, var_962))[name = tensor("window_41")]; + tensor var_970_begin_0 = const()[name = tensor("op_970_begin_0"), val = tensor([0, 2, 0])]; + tensor var_970_end_0 = const()[name = tensor("op_970_end_0"), val = tensor([1, 3, 256])]; + tensor var_970_end_mask_0 = const()[name = tensor("op_970_end_mask_0"), val = tensor([true, false, true])]; + tensor var_970 = slice_by_index(begin = var_970_begin_0, end = var_970_end_0, end_mask = var_970_end_mask_0, x = x_21)[name = tensor("op_970")]; + tensor var_973_begin_0 = const()[name = tensor("op_973_begin_0"), val = tensor([0, 1, 0])]; + tensor var_973_end_0 = const()[name = tensor("op_973_end_0"), val = tensor([1, 16, 256])]; + tensor var_973_end_mask_0 = const()[name = tensor("op_973_end_mask_0"), val = tensor([true, true, true])]; + tensor var_973 = slice_by_index(begin = var_973_begin_0, end = var_973_end_0, end_mask = var_973_end_mask_0, x = window_41)[name = tensor("op_973")]; tensor window_43_interleave_0 = const()[name = tensor("window_43_interleave_0"), val = tensor(false)]; - tensor window_43 = concat(axis = var_27, interleave = window_43_interleave_0, values = (var_907, var_904))[name = tensor("window_43")]; - tensor var_912_begin_0 = const()[name = tensor("op_912_begin_0"), val = tensor([0, 3, 0])]; - tensor var_912_end_0 = const()[name = tensor("op_912_end_0"), val = tensor([1, 4, 256])]; - tensor var_912_end_mask_0 = const()[name = tensor("op_912_end_mask_0"), val = tensor([true, false, true])]; - tensor var_912 = slice_by_index(begin = var_912_begin_0, end = var_912_end_0, end_mask = var_912_end_mask_0, x = x_21)[name = tensor("op_912")]; - tensor var_915_begin_0 = const()[name = tensor("op_915_begin_0"), val = tensor([0, 1, 0])]; - tensor var_915_end_0 = const()[name = tensor("op_915_end_0"), val = tensor([1, 16, 256])]; - tensor var_915_end_mask_0 = const()[name = tensor("op_915_end_mask_0"), val = tensor([true, true, true])]; - tensor var_915 = slice_by_index(begin = var_915_begin_0, end = var_915_end_0, end_mask = var_915_end_mask_0, x = window_43)[name = tensor("op_915")]; + tensor window_43 = concat(axis = var_93, interleave = window_43_interleave_0, values = (var_973, var_970))[name = tensor("window_43")]; + tensor var_978_begin_0 = const()[name = tensor("op_978_begin_0"), val = tensor([0, 3, 0])]; + tensor var_978_end_0 = const()[name = tensor("op_978_end_0"), val = tensor([1, 4, 256])]; + tensor var_978_end_mask_0 = const()[name = tensor("op_978_end_mask_0"), val = tensor([true, false, true])]; + tensor var_978 = slice_by_index(begin = var_978_begin_0, end = var_978_end_0, end_mask = var_978_end_mask_0, x = x_21)[name = tensor("op_978")]; + tensor var_981_begin_0 = const()[name = tensor("op_981_begin_0"), val = tensor([0, 1, 0])]; + tensor var_981_end_0 = const()[name = tensor("op_981_end_0"), val = tensor([1, 16, 256])]; + tensor var_981_end_mask_0 = const()[name = tensor("op_981_end_mask_0"), val = tensor([true, true, true])]; + tensor var_981 = slice_by_index(begin = var_981_begin_0, end = var_981_end_0, end_mask = var_981_end_mask_0, x = window_43)[name = tensor("op_981")]; tensor window_45_interleave_0 = const()[name = tensor("window_45_interleave_0"), val = tensor(false)]; - tensor window_45 = concat(axis = var_27, interleave = window_45_interleave_0, values = (var_915, var_912))[name = tensor("window_45")]; - tensor var_920_begin_0 = const()[name = tensor("op_920_begin_0"), val = tensor([0, 4, 0])]; - tensor var_920_end_0 = const()[name = tensor("op_920_end_0"), val = tensor([1, 1, 256])]; - tensor var_920_end_mask_0 = const()[name = tensor("op_920_end_mask_0"), val = tensor([true, true, true])]; - tensor var_920 = slice_by_index(begin = var_920_begin_0, end = var_920_end_0, end_mask = var_920_end_mask_0, x = x_21)[name = tensor("op_920")]; - tensor var_923_begin_0 = const()[name = tensor("op_923_begin_0"), val = tensor([0, 1, 0])]; - tensor var_923_end_0 = const()[name = tensor("op_923_end_0"), val = tensor([1, 16, 256])]; - tensor var_923_end_mask_0 = const()[name = tensor("op_923_end_mask_0"), val = tensor([true, true, true])]; - tensor var_923 = slice_by_index(begin = var_923_begin_0, end = var_923_end_0, end_mask = var_923_end_mask_0, x = window_45)[name = tensor("op_923")]; + tensor window_45 = concat(axis = var_93, interleave = window_45_interleave_0, values = (var_981, var_978))[name = tensor("window_45")]; + tensor var_986_begin_0 = const()[name = tensor("op_986_begin_0"), val = tensor([0, 4, 0])]; + tensor var_986_end_0 = const()[name = tensor("op_986_end_0"), val = tensor([1, 1, 256])]; + tensor var_986_end_mask_0 = const()[name = tensor("op_986_end_mask_0"), val = tensor([true, true, true])]; + tensor var_986 = slice_by_index(begin = var_986_begin_0, end = var_986_end_0, end_mask = var_986_end_mask_0, x = x_21)[name = tensor("op_986")]; + tensor var_989_begin_0 = const()[name = tensor("op_989_begin_0"), val = tensor([0, 1, 0])]; + tensor var_989_end_0 = const()[name = tensor("op_989_end_0"), val = tensor([1, 16, 256])]; + tensor var_989_end_mask_0 = const()[name = tensor("op_989_end_mask_0"), val = tensor([true, true, true])]; + tensor var_989 = slice_by_index(begin = var_989_begin_0, end = var_989_end_0, end_mask = var_989_end_mask_0, x = window_45)[name = tensor("op_989")]; tensor window_interleave_0 = const()[name = tensor("window_interleave_0"), val = tensor(false)]; - tensor window = concat(axis = var_27, interleave = window_interleave_0, values = (var_923, var_920))[name = tensor("window")]; - tensor input_141_interleave_0 = const()[name = tensor("input_141_interleave_0"), val = tensor(false)]; - tensor input_141 = concat(axis = var_24, interleave = input_141_interleave_0, values = (window_39, window_41, window_43, window_45, window))[name = tensor("input_141")]; + tensor window = concat(axis = var_93, interleave = window_interleave_0, values = (var_989, var_986))[name = tensor("window")]; + tensor input_143_interleave_0 = const()[name = tensor("input_143_interleave_0"), val = tensor(false)]; + tensor input_143 = concat(axis = var_79, interleave = input_143_interleave_0, values = (window_39, window_41, window_43, window_45, window))[name = tensor("input_143")]; tensor x_23_axes_0 = const()[name = tensor("x_23_axes_0"), val = tensor([-1])]; - tensor x_23 = layer_norm(axes = x_23_axes_0, beta = encoder_conv_module_3_sequential_0_bias, epsilon = var_29, gamma = encoder_conv_module_3_sequential_0_weight, x = input_141)[name = tensor("x_23")]; - tensor input_143_perm_0 = const()[name = tensor("input_143_perm_0"), val = tensor([0, 2, 1])]; + tensor x_23 = layer_norm(axes = x_23_axes_0, beta = inner_encoder_conv_module_3_sequential_0_bias, epsilon = var_76, gamma = inner_encoder_conv_module_3_sequential_0_weight, x = input_143)[name = tensor("x_23")]; + tensor input_145_perm_0 = const()[name = tensor("input_145_perm_0"), val = tensor([0, 2, 1])]; tensor inputs_33_pad_type_0 = const()[name = tensor("inputs_33_pad_type_0"), val = tensor("valid")]; tensor inputs_33_strides_0 = const()[name = tensor("inputs_33_strides_0"), val = tensor([1])]; tensor inputs_33_pad_0 = const()[name = tensor("inputs_33_pad_0"), val = tensor([0, 0])]; tensor inputs_33_dilations_0 = const()[name = tensor("inputs_33_dilations_0"), val = tensor([1])]; tensor inputs_33_groups_0 = const()[name = tensor("inputs_33_groups_0"), val = tensor(1)]; - tensor input_143 = transpose(perm = input_143_perm_0, x = x_23)[name = tensor("transpose_33")]; - tensor inputs_33 = conv(bias = encoder_conv_module_3_sequential_2_conv_bias, dilations = inputs_33_dilations_0, groups = inputs_33_groups_0, pad = inputs_33_pad_0, pad_type = inputs_33_pad_type_0, strides = inputs_33_strides_0, weight = encoder_conv_module_3_sequential_2_conv_weight, x = input_143)[name = tensor("inputs_33")]; - tensor var_948_split_sizes_0 = const()[name = tensor("op_948_split_sizes_0"), val = tensor([256, 256])]; - tensor var_948_axis_0 = const()[name = tensor("op_948_axis_0"), val = tensor(1)]; - tensor var_948_0, tensor var_948_1 = split(axis = var_948_axis_0, split_sizes = var_948_split_sizes_0, x = inputs_33)[name = tensor("op_948")]; - tensor var_950 = sigmoid(x = var_948_1)[name = tensor("op_950")]; - tensor inputs_35 = mul(x = var_948_0, y = var_950)[name = tensor("inputs_35")]; + tensor input_145 = transpose(perm = input_145_perm_0, x = x_23)[name = tensor("transpose_33")]; + tensor inputs_33 = conv(bias = inner_encoder_conv_module_3_sequential_2_conv_bias, dilations = inputs_33_dilations_0, groups = inputs_33_groups_0, pad = inputs_33_pad_0, pad_type = inputs_33_pad_type_0, strides = inputs_33_strides_0, weight = inner_encoder_conv_module_3_sequential_2_conv_weight, x = input_145)[name = tensor("inputs_33")]; + tensor var_1014_split_sizes_0 = const()[name = tensor("op_1014_split_sizes_0"), val = tensor([256, 256])]; + tensor var_1014_axis_0 = const()[name = tensor("op_1014_axis_0"), val = tensor(1)]; + tensor var_1014_0, tensor var_1014_1 = split(axis = var_1014_axis_0, split_sizes = var_1014_split_sizes_0, x = inputs_33)[name = tensor("op_1014")]; + tensor var_1016 = sigmoid(x = var_1014_1)[name = tensor("op_1016")]; + tensor inputs_35 = mul(x = var_1014_0, y = var_1016)[name = tensor("inputs_35")]; tensor outputs_aug_pad_type_0 = const()[name = tensor("outputs_aug_pad_type_0"), val = tensor("custom")]; tensor outputs_aug_pad_0 = const()[name = tensor("outputs_aug_pad_0"), val = tensor([15, 15])]; tensor outputs_aug_groups_0 = const()[name = tensor("outputs_aug_groups_0"), val = tensor(256)]; tensor outputs_aug_strides_0 = const()[name = tensor("outputs_aug_strides_0"), val = tensor([1])]; tensor outputs_aug_dilations_0 = const()[name = tensor("outputs_aug_dilations_0"), val = tensor([1])]; - tensor outputs_aug = conv(dilations = outputs_aug_dilations_0, groups = outputs_aug_groups_0, pad = outputs_aug_pad_0, pad_type = outputs_aug_pad_type_0, strides = outputs_aug_strides_0, weight = encoder_conv_module_3_sequential_4_conv_weight, x = inputs_35)[name = tensor("outputs_aug")]; - tensor input_145_begin_0 = const()[name = tensor("input_145_begin_0"), val = tensor([0, 0, 0])]; - tensor input_145_end_0 = const()[name = tensor("input_145_end_0"), val = tensor([5, 256, 16])]; - tensor input_145_end_mask_0 = const()[name = tensor("input_145_end_mask_0"), val = tensor([true, true, false])]; - tensor input_145 = slice_by_index(begin = input_145_begin_0, end = input_145_end_0, end_mask = input_145_end_mask_0, x = outputs_aug)[name = tensor("input_145")]; - tensor inputs_37 = batch_norm(beta = encoder_conv_module_3_sequential_5_bias, epsilon = var_29, gamma = encoder_conv_module_3_sequential_5_weight, mean = encoder_conv_module_3_sequential_5_running_mean, variance = encoder_conv_module_3_sequential_5_running_var, x = input_145)[name = tensor("inputs_37")]; - tensor input_147 = silu(x = inputs_37)[name = tensor("input_147")]; - tensor input_149_pad_type_0 = const()[name = tensor("input_149_pad_type_0"), val = tensor("valid")]; - tensor input_149_strides_0 = const()[name = tensor("input_149_strides_0"), val = tensor([1])]; - tensor input_149_pad_0 = const()[name = tensor("input_149_pad_0"), val = tensor([0, 0])]; - tensor input_149_dilations_0 = const()[name = tensor("input_149_dilations_0"), val = tensor([1])]; - tensor input_149_groups_0 = const()[name = tensor("input_149_groups_0"), val = tensor(1)]; - tensor input_149 = conv(bias = encoder_conv_module_3_sequential_7_conv_bias, dilations = input_149_dilations_0, groups = input_149_groups_0, pad = input_149_pad_0, pad_type = input_149_pad_type_0, strides = input_149_strides_0, weight = encoder_conv_module_3_sequential_7_conv_weight, x = input_147)[name = tensor("input_149")]; + tensor outputs_aug = conv(dilations = outputs_aug_dilations_0, groups = outputs_aug_groups_0, pad = outputs_aug_pad_0, pad_type = outputs_aug_pad_type_0, strides = outputs_aug_strides_0, weight = inner_encoder_conv_module_3_sequential_4_conv_weight, x = inputs_35)[name = tensor("outputs_aug")]; + tensor input_147_begin_0 = const()[name = tensor("input_147_begin_0"), val = tensor([0, 0, 0])]; + tensor input_147_end_0 = const()[name = tensor("input_147_end_0"), val = tensor([5, 256, 16])]; + tensor input_147_end_mask_0 = const()[name = tensor("input_147_end_mask_0"), val = tensor([true, true, false])]; + tensor input_147 = slice_by_index(begin = input_147_begin_0, end = input_147_end_0, end_mask = input_147_end_mask_0, x = outputs_aug)[name = tensor("input_147")]; + tensor inputs_37 = batch_norm(beta = inner_encoder_conv_module_3_sequential_5_bias, epsilon = var_76, gamma = inner_encoder_conv_module_3_sequential_5_weight, mean = inner_encoder_conv_module_3_sequential_5_running_mean, variance = inner_encoder_conv_module_3_sequential_5_running_var, x = input_147)[name = tensor("inputs_37")]; + tensor input_149 = silu(x = inputs_37)[name = tensor("input_149")]; + tensor input_151_pad_type_0 = const()[name = tensor("input_151_pad_type_0"), val = tensor("valid")]; + tensor input_151_strides_0 = const()[name = tensor("input_151_strides_0"), val = tensor([1])]; + tensor input_151_pad_0 = const()[name = tensor("input_151_pad_0"), val = tensor([0, 0])]; + tensor input_151_dilations_0 = const()[name = tensor("input_151_dilations_0"), val = tensor([1])]; + tensor input_151_groups_0 = const()[name = tensor("input_151_groups_0"), val = tensor(1)]; + tensor input_151 = conv(bias = inner_encoder_conv_module_3_sequential_7_conv_bias, dilations = input_151_dilations_0, groups = input_151_groups_0, pad = input_151_pad_0, pad_type = input_151_pad_type_0, strides = input_151_strides_0, weight = inner_encoder_conv_module_3_sequential_7_conv_weight, x = input_149)[name = tensor("input_151")]; tensor conv_out_7_perm_0 = const()[name = tensor("conv_out_7_perm_0"), val = tensor([0, 2, 1])]; - tensor var_981_begin_0 = const()[name = tensor("op_981_begin_0"), val = tensor([0, -1, 0])]; - tensor var_981_end_0 = const()[name = tensor("op_981_end_0"), val = tensor([5, 16, 256])]; - tensor var_981_end_mask_0 = const()[name = tensor("op_981_end_mask_0"), val = tensor([true, true, true])]; - tensor conv_out_7 = transpose(perm = conv_out_7_perm_0, x = input_149)[name = tensor("transpose_32")]; - tensor var_981 = slice_by_index(begin = var_981_begin_0, end = var_981_end_0, end_mask = var_981_end_mask_0, x = conv_out_7)[name = tensor("op_981")]; - tensor var_983_perm_0 = const()[name = tensor("op_983_perm_0"), val = tensor([1, 0, 2])]; - tensor var_983 = transpose(perm = var_983_perm_0, x = var_981)[name = tensor("transpose_31")]; - tensor input_151 = add(x = x_21, y = var_983)[name = tensor("input_151")]; - tensor input_153_axes_0 = const()[name = tensor("input_153_axes_0"), val = tensor([-1])]; - tensor input_153 = layer_norm(axes = input_153_axes_0, beta = encoder_ffn2_3_module_sequential_0_bias, epsilon = var_29, gamma = encoder_ffn2_3_module_sequential_0_weight, x = input_151)[name = tensor("input_153")]; - tensor inputs = linear(bias = encoder_ffn2_3_module_sequential_1_linear_bias, weight = encoder_ffn2_3_module_sequential_1_linear_weight, x = input_153)[name = tensor("linear_35")]; - tensor input_155 = silu(x = inputs)[name = tensor("input_155")]; - tensor input_159 = linear(bias = encoder_ffn2_3_module_sequential_4_linear_bias, weight = encoder_ffn2_3_module_sequential_4_linear_weight, x = input_155)[name = tensor("linear_36")]; - tensor var_1006 = const()[name = tensor("op_1006"), val = tensor(0x1p-1)]; - tensor var_1007 = mul(x = input_159, y = var_1006)[name = tensor("op_1007")]; - tensor input_161 = add(x = var_1007, y = input_151)[name = tensor("input_161")]; + tensor var_1047_begin_0 = const()[name = tensor("op_1047_begin_0"), val = tensor([0, -1, 0])]; + tensor var_1047_end_0 = const()[name = tensor("op_1047_end_0"), val = tensor([5, 16, 256])]; + tensor var_1047_end_mask_0 = const()[name = tensor("op_1047_end_mask_0"), val = tensor([true, true, true])]; + tensor conv_out_7 = transpose(perm = conv_out_7_perm_0, x = input_151)[name = tensor("transpose_32")]; + tensor var_1047 = slice_by_index(begin = var_1047_begin_0, end = var_1047_end_0, end_mask = var_1047_end_mask_0, x = conv_out_7)[name = tensor("op_1047")]; + tensor var_1049_perm_0 = const()[name = tensor("op_1049_perm_0"), val = tensor([1, 0, 2])]; + tensor var_1049 = transpose(perm = var_1049_perm_0, x = var_1047)[name = tensor("transpose_31")]; + tensor input_153 = add(x = x_21, y = var_1049)[name = tensor("input_153")]; + tensor input_155_axes_0 = const()[name = tensor("input_155_axes_0"), val = tensor([-1])]; + tensor input_155 = layer_norm(axes = input_155_axes_0, beta = inner_encoder_ffn2_3_module_sequential_0_bias, epsilon = var_76, gamma = inner_encoder_ffn2_3_module_sequential_0_weight, x = input_153)[name = tensor("input_155")]; + tensor inputs = linear(bias = inner_encoder_ffn2_3_module_sequential_1_linear_bias, weight = inner_encoder_ffn2_3_module_sequential_1_linear_weight, x = input_155)[name = tensor("linear_35")]; + tensor input_157 = silu(x = inputs)[name = tensor("input_157")]; + tensor input_161 = linear(bias = inner_encoder_ffn2_3_module_sequential_4_linear_bias, weight = inner_encoder_ffn2_3_module_sequential_4_linear_weight, x = input_157)[name = tensor("linear_36")]; + tensor var_1072 = const()[name = tensor("op_1072"), val = tensor(0x1p-1)]; + tensor var_1073 = mul(x = input_161, y = var_1072)[name = tensor("op_1073")]; + tensor input_163 = add(x = var_1073, y = input_153)[name = tensor("input_163")]; tensor x_25_axes_0 = const()[name = tensor("x_25_axes_0"), val = tensor([-1])]; - tensor x_25 = layer_norm(axes = x_25_axes_0, beta = encoder_layer_norm_3_bias, epsilon = var_29, gamma = encoder_layer_norm_3_weight, x = input_161)[name = tensor("x_25")]; + tensor x_25 = layer_norm(axes = x_25_axes_0, beta = inner_encoder_layer_norm_3_bias, epsilon = var_76, gamma = inner_encoder_layer_norm_3_weight, x = input_163)[name = tensor("x_25")]; tensor x_ct_perm_0 = const()[name = tensor("x_ct_perm_0"), val = tensor([0, 2, 1])]; tensor cat_interleave_0 = const()[name = tensor("cat_interleave_0"), val = tensor(false)]; tensor x_ct = transpose(perm = x_ct_perm_0, x = x_25)[name = tensor("transpose_30")]; - tensor cat = concat(axis = var_21, interleave = cat_interleave_0, values = (cnn_window, x_ct))[name = tensor("cat")]; + tensor cat = concat(axis = var_81, interleave = cat_interleave_0, values = (cnn_window, x_ct))[name = tensor("cat")]; tensor conv_out_pad_type_0 = const()[name = tensor("conv_out_pad_type_0"), val = tensor("valid")]; tensor conv_out_strides_0 = const()[name = tensor("conv_out_strides_0"), val = tensor([1])]; tensor conv_out_pad_0 = const()[name = tensor("conv_out_pad_0"), val = tensor([0, 0])]; tensor conv_out_dilations_0 = const()[name = tensor("conv_out_dilations_0"), val = tensor([1])]; tensor conv_out_groups_0 = const()[name = tensor("conv_out_groups_0"), val = tensor(1)]; - tensor conv_out = conv(bias = encoder_cnn_bias, dilations = conv_out_dilations_0, groups = conv_out_groups_0, pad = conv_out_pad_0, pad_type = conv_out_pad_type_0, strides = conv_out_strides_0, weight = encoder_cnn_weight, x = cat)[name = tensor("conv_out")]; - tensor var_1025_begin_0 = const()[name = tensor("op_1025_begin_0"), val = tensor([0, 0, 5])]; - tensor var_1025_end_0 = const()[name = tensor("op_1025_end_0"), val = tensor([1, 256, 23])]; - tensor var_1025_end_mask_0 = const()[name = tensor("op_1025_end_mask_0"), val = tensor([true, true, true])]; - tensor cnn_window_new = slice_by_index(begin = var_1025_begin_0, end = var_1025_end_0, end_mask = var_1025_end_mask_0, x = cat)[name = tensor("op_1025")]; - tensor input_163_perm_0 = const()[name = tensor("input_163_perm_0"), val = tensor([0, 2, 1])]; - tensor var_1027 = const()[name = tensor("op_1027"), val = tensor([-1])]; - tensor input_163 = transpose(perm = input_163_perm_0, x = conv_out)[name = tensor("transpose_29")]; - tensor var_1028 = reduce_l2_norm(axes = var_1027, keep_dims = var_30, x = input_163)[name = tensor("op_1028")]; + tensor conv_out = conv(bias = inner_encoder_cnn_bias, dilations = conv_out_dilations_0, groups = conv_out_groups_0, pad = conv_out_pad_0, pad_type = conv_out_pad_type_0, strides = conv_out_strides_0, weight = inner_encoder_cnn_weight, x = cat)[name = tensor("conv_out")]; + tensor var_1091_begin_0 = const()[name = tensor("op_1091_begin_0"), val = tensor([0, 0, 5])]; + tensor var_1091_end_0 = const()[name = tensor("op_1091_end_0"), val = tensor([1, 256, 23])]; + tensor var_1091_end_mask_0 = const()[name = tensor("op_1091_end_mask_0"), val = tensor([true, true, true])]; + tensor cnn_window_new = slice_by_index(begin = var_1091_begin_0, end = var_1091_end_0, end_mask = var_1091_end_mask_0, x = cat)[name = tensor("op_1091")]; + tensor input_165_perm_0 = const()[name = tensor("input_165_perm_0"), val = tensor([0, 2, 1])]; + tensor var_1093 = const()[name = tensor("op_1093"), val = tensor([-1])]; + tensor input_165 = transpose(perm = input_165_perm_0, x = conv_out)[name = tensor("transpose_29")]; + tensor var_1094 = reduce_l2_norm(axes = var_1093, keep_dims = var_75, x = input_165)[name = tensor("op_1094")]; tensor const_12 = const()[name = tensor("const_12"), val = tensor(0x1.fffffep+127)]; - tensor clip_0 = clip(alpha = var_11, beta = const_12, x = var_1028)[name = tensor("clip_0")]; - tensor emb = real_div(x = input_163, y = clip_0)[name = tensor("emb")]; - tensor var_1032_axis_0 = const()[name = tensor("op_1032_axis_0"), val = tensor(0)]; - tensor enc_kv_new = stack(axis = var_1032_axis_0, values = (var_207, var_429, var_651, nkv_1))[name = tensor("op_1032")]; - tensor var_1034_axis_0 = const()[name = tensor("op_1034_axis_0"), val = tensor(0)]; - tensor enc_scale_new = stack(axis = var_1034_axis_0, values = (new_scale_1, new_scale_3, new_scale_5, new_scale_7))[name = tensor("op_1034")]; - tensor var_1036_axis_0 = const()[name = tensor("op_1036_axis_0"), val = tensor(0)]; - tensor enc_conv_cache_new = stack(axis = var_1036_axis_0, values = (window_11, window_23, window_35, window))[name = tensor("op_1036")]; - tensor var_1045 = const()[name = tensor("op_1045"), val = tensor(0x1.5798eep-27)]; - tensor var_1050 = const()[name = tensor("op_1050"), val = tensor(0x1.4f8b58p-17)]; - tensor var_1052 = const()[name = tensor("op_1052"), val = tensor(0x1.0c6f7ap-20)]; - tensor var_1053 = const()[name = tensor("op_1053"), val = tensor(true)]; - tensor var_1055 = const()[name = tensor("op_1055"), val = tensor(0x1p+0)]; - tensor var_1059 = const()[name = tensor("op_1059"), val = tensor(-1)]; - tensor var_1065 = const()[name = tensor("op_1065"), val = tensor(0)]; + tensor clip_0 = clip(alpha = var_90, beta = const_12, x = var_1094)[name = tensor("clip_0")]; + tensor emb = real_div(x = input_165, y = clip_0)[name = tensor("emb")]; + tensor var_1098_axis_0 = const()[name = tensor("op_1098_axis_0"), val = tensor(0)]; + tensor enc_kv_new = stack(axis = var_1098_axis_0, values = (var_273, var_495, var_717, nkv_1))[name = tensor("op_1098")]; + tensor var_1100_axis_0 = const()[name = tensor("op_1100_axis_0"), val = tensor(0)]; + tensor enc_scale_new = stack(axis = var_1100_axis_0, values = (new_scale_1, new_scale_3, new_scale_5, new_scale_7))[name = tensor("op_1100")]; + tensor var_1102_axis_0 = const()[name = tensor("op_1102_axis_0"), val = tensor(0)]; + tensor enc_conv_cache_new = stack(axis = var_1102_axis_0, values = (window_11, window_23, window_35, window))[name = tensor("op_1102")]; tensor pos = const()[name = tensor("pos"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(44395712)))]; - tensor var_1127_axes_0 = const()[name = tensor("op_1127_axes_0"), val = tensor([2])]; - tensor var_1127 = expand_dims(axes = var_1127_axes_0, x = emb)[name = tensor("op_1127")]; + tensor var_1170_axes_0 = const()[name = tensor("op_1170_axes_0"), val = tensor([2])]; + tensor var_1170 = expand_dims(axes = var_1170_axes_0, x = emb)[name = tensor("op_1170")]; tensor emb_exp_reps_0 = const()[name = tensor("emb_exp_reps_0"), val = tensor([1, 1, 12, 1])]; - tensor emb_exp = tile(reps = emb_exp_reps_0, x = var_1127)[name = tensor("emb_exp")]; - tensor input_165_interleave_0 = const()[name = tensor("input_165_interleave_0"), val = tensor(false)]; - tensor input_165 = concat(axis = var_1059, interleave = input_165_interleave_0, values = (emb_exp, pos))[name = tensor("input_165")]; - tensor x_27 = linear(bias = decoder_convert_bias, weight = decoder_convert_weight, x = input_165)[name = tensor("linear_37")]; - tensor var_1135_perm_0 = const()[name = tensor("op_1135_perm_0"), val = tensor([0, 2, 1, 3])]; - tensor var_1139 = const()[name = tensor("op_1139"), val = tensor([12, 5, 256])]; - tensor var_1135 = transpose(perm = var_1135_perm_0, x = x_27)[name = tensor("transpose_28")]; - tensor x_29 = reshape(shape = var_1139, x = var_1135)[name = tensor("x_29")]; + tensor emb_exp = tile(reps = emb_exp_reps_0, x = var_1170)[name = tensor("emb_exp")]; + tensor input_167_interleave_0 = const()[name = tensor("input_167_interleave_0"), val = tensor(false)]; + tensor input_167 = concat(axis = var_82, interleave = input_167_interleave_0, values = (emb_exp, pos))[name = tensor("input_167")]; + tensor x_27 = linear(bias = inner_decoder_convert_bias, weight = inner_decoder_convert_weight, x = input_167)[name = tensor("linear_37")]; + tensor var_1178_perm_0 = const()[name = tensor("op_1178_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_1182 = const()[name = tensor("op_1182"), val = tensor([12, 5, 256])]; + tensor var_1178 = transpose(perm = var_1178_perm_0, x = x_27)[name = tensor("transpose_28")]; + tensor x_29 = reshape(shape = var_1182, x = var_1178)[name = tensor("x_29")]; tensor prev_kv_9_begin_0 = const()[name = tensor("prev_kv_9_begin_0"), val = tensor([0, 0, 0, 0, 0])]; tensor prev_kv_9_end_0 = const()[name = tensor("prev_kv_9_end_0"), val = tensor([1, 12, 4, 64, 64])]; tensor prev_kv_9_end_mask_0 = const()[name = tensor("prev_kv_9_end_mask_0"), val = tensor([false, true, true, true, true])]; @@ -1023,132 +1042,132 @@ program(1.0) tensor prev_scale_9_end_mask_0 = const()[name = tensor("prev_scale_9_end_mask_0"), val = tensor([false, true])]; tensor prev_scale_9_squeeze_mask_0 = const()[name = tensor("prev_scale_9_squeeze_mask_0"), val = tensor([true, false])]; tensor prev_scale_9 = slice_by_index(begin = prev_scale_9_begin_0, end = prev_scale_9_end_0, end_mask = prev_scale_9_end_mask_0, squeeze_mask = prev_scale_9_squeeze_mask_0, x = dec_scale)[name = tensor("prev_scale_9")]; - tensor var_1147 = linear(bias = decoder_q_proj_0_bias, weight = decoder_q_proj_0_weight, x = x_29)[name = tensor("linear_38")]; - tensor var_1148 = const()[name = tensor("op_1148"), val = tensor([12, 5, 4, 64])]; - tensor var_1149 = reshape(shape = var_1148, x = var_1147)[name = tensor("op_1149")]; + tensor var_1190 = linear(bias = inner_decoder_q_proj_0_bias, weight = inner_decoder_q_proj_0_weight, x = x_29)[name = tensor("linear_38")]; + tensor var_1191 = const()[name = tensor("op_1191"), val = tensor([12, 5, 4, 64])]; + tensor var_1192 = reshape(shape = var_1191, x = var_1190)[name = tensor("op_1192")]; tensor q_9_perm_0 = const()[name = tensor("q_9_perm_0"), val = tensor([0, 2, 1, 3])]; - tensor var_1153 = linear(bias = decoder_k_proj_0_bias, weight = decoder_k_proj_0_weight, x = x_29)[name = tensor("linear_39")]; - tensor var_1154 = const()[name = tensor("op_1154"), val = tensor(0x1p-3)]; - tensor var_1155 = mul(x = var_1153, y = var_1154)[name = tensor("op_1155")]; - tensor var_1156 = const()[name = tensor("op_1156"), val = tensor([12, 5, 4, 64])]; - tensor var_1157 = reshape(shape = var_1156, x = var_1155)[name = tensor("op_1157")]; + tensor var_1196 = linear(bias = inner_decoder_k_proj_0_bias, weight = inner_decoder_k_proj_0_weight, x = x_29)[name = tensor("linear_39")]; + tensor var_1197 = const()[name = tensor("op_1197"), val = tensor(0x1p-3)]; + tensor var_1198 = mul(x = var_1196, y = var_1197)[name = tensor("op_1198")]; + tensor var_1199 = const()[name = tensor("op_1199"), val = tensor([12, 5, 4, 64])]; + tensor var_1200 = reshape(shape = var_1199, x = var_1198)[name = tensor("op_1200")]; tensor k_9_perm_0 = const()[name = tensor("k_9_perm_0"), val = tensor([0, 2, 1, 3])]; - tensor var_1161 = linear(bias = decoder_v_proj_0_bias, weight = decoder_v_proj_0_weight, x = x_29)[name = tensor("linear_40")]; - tensor var_1162 = const()[name = tensor("op_1162"), val = tensor([12, 5, 4, 64])]; - tensor var_1163 = reshape(shape = var_1162, x = var_1161)[name = tensor("op_1163")]; + tensor var_1204 = linear(bias = inner_decoder_v_proj_0_bias, weight = inner_decoder_v_proj_0_weight, x = x_29)[name = tensor("linear_40")]; + tensor var_1205 = const()[name = tensor("op_1205"), val = tensor([12, 5, 4, 64])]; + tensor var_1206 = reshape(shape = var_1205, x = var_1204)[name = tensor("op_1206")]; tensor v_9_perm_0 = const()[name = tensor("v_9_perm_0"), val = tensor([0, 2, 1, 3])]; - tensor input_169 = linear(bias = decoder_g_proj_0_bias, weight = decoder_g_proj_0_weight, x = x_29)[name = tensor("linear_41")]; + tensor input_171 = linear(bias = inner_decoder_g_proj_0_bias, weight = inner_decoder_g_proj_0_weight, x = x_29)[name = tensor("linear_41")]; tensor cumsum_mask_1_exclusive_0 = const()[name = tensor("cumsum_mask_1_exclusive_0"), val = tensor(false)]; tensor cumsum_mask_1_reverse_0 = const()[name = tensor("cumsum_mask_1_reverse_0"), val = tensor(false)]; - tensor cumsum_mask_1 = cumsum(axis = var_1065, exclusive = cumsum_mask_1_exclusive_0, reverse = cumsum_mask_1_reverse_0, x = valid_mask)[name = tensor("cumsum_mask_1")]; + tensor cumsum_mask_1 = cumsum(axis = var_79, exclusive = cumsum_mask_1_exclusive_0, reverse = cumsum_mask_1_reverse_0, x = valid_mask)[name = tensor("cumsum_mask_1")]; tensor sqrt_s0_9 = sqrt(x = prev_scale_9)[name = tensor("sqrt_s0_9")]; tensor s_t_9 = add(x = prev_scale_9, y = cumsum_mask_1)[name = tensor("s_t_9")]; tensor const_20 = const()[name = tensor("const_20"), val = tensor(0x1.fffffep+127)]; - tensor clip_1 = clip(alpha = var_1055, beta = const_20, x = s_t_9)[name = tensor("clip_1")]; + tensor clip_1 = clip(alpha = var_69, beta = const_20, x = s_t_9)[name = tensor("clip_1")]; tensor sqrt_s_t_9 = sqrt(x = clip_1)[name = tensor("sqrt_s_t_9")]; tensor qk_9_transpose_x_1 = const()[name = tensor("qk_9_transpose_x_1"), val = tensor(false)]; tensor qk_9_transpose_y_1 = const()[name = tensor("qk_9_transpose_y_1"), val = tensor(true)]; - tensor k_9 = transpose(perm = k_9_perm_0, x = var_1157)[name = tensor("transpose_26")]; - tensor q_9 = transpose(perm = q_9_perm_0, x = var_1149)[name = tensor("transpose_27")]; + tensor k_9 = transpose(perm = k_9_perm_0, x = var_1200)[name = tensor("transpose_26")]; + tensor q_9 = transpose(perm = q_9_perm_0, x = var_1192)[name = tensor("transpose_27")]; tensor qk_9 = matmul(transpose_x = qk_9_transpose_x_1, transpose_y = qk_9_transpose_y_1, x = q_9, y = k_9)[name = tensor("qk_9")]; - tensor var_1175 = const()[name = tensor("op_1175"), val = tensor([1, 5])]; - tensor var_1176 = reshape(shape = var_1175, x = valid_mask)[name = tensor("op_1176")]; - tensor causal_with_valid_1 = mul(x = decoder__causal_mask, y = var_1176)[name = tensor("causal_with_valid_1")]; - tensor var_1178 = const()[name = tensor("op_1178"), val = tensor([5, 1])]; - tensor var_1179 = reshape(shape = var_1178, x = sqrt_s_t_9)[name = tensor("op_1179")]; - tensor M_9 = real_div(x = causal_with_valid_1, y = var_1179)[name = tensor("M_9")]; - tensor var_1181 = mul(x = qk_9, y = M_9)[name = tensor("op_1181")]; + tensor var_1218 = const()[name = tensor("op_1218"), val = tensor([1, 5])]; + tensor var_1219 = reshape(shape = var_1218, x = valid_mask)[name = tensor("op_1219")]; + tensor causal_with_valid_1 = mul(x = inner_decoder__causal_mask, y = var_1219)[name = tensor("causal_with_valid_1")]; + tensor var_1221 = const()[name = tensor("op_1221"), val = tensor([5, 1])]; + tensor var_1222 = reshape(shape = var_1221, x = sqrt_s_t_9)[name = tensor("op_1222")]; + tensor M_9 = real_div(x = causal_with_valid_1, y = var_1222)[name = tensor("M_9")]; + tensor var_1224 = mul(x = qk_9, y = M_9)[name = tensor("op_1224")]; tensor inner_9_transpose_x_0 = const()[name = tensor("inner_9_transpose_x_0"), val = tensor(false)]; tensor inner_9_transpose_y_0 = const()[name = tensor("inner_9_transpose_y_0"), val = tensor(false)]; - tensor v_9 = transpose(perm = v_9_perm_0, x = var_1163)[name = tensor("transpose_25")]; - tensor inner_9 = matmul(transpose_x = inner_9_transpose_x_0, transpose_y = inner_9_transpose_y_0, x = var_1181, y = v_9)[name = tensor("inner_9")]; - tensor var_1183_transpose_x_0 = const()[name = tensor("op_1183_transpose_x_0"), val = tensor(false)]; - tensor var_1183_transpose_y_0 = const()[name = tensor("op_1183_transpose_y_0"), val = tensor(false)]; - tensor var_1183 = matmul(transpose_x = var_1183_transpose_x_0, transpose_y = var_1183_transpose_y_0, x = q_9, y = prev_kv_9)[name = tensor("op_1183")]; - tensor var_1184 = real_div(x = sqrt_s0_9, y = sqrt_s_t_9)[name = tensor("op_1184")]; - tensor var_1185 = const()[name = tensor("op_1185"), val = tensor([1, 1, 5, 1])]; - tensor var_1186 = reshape(shape = var_1185, x = var_1184)[name = tensor("op_1186")]; - tensor cross_9 = mul(x = var_1183, y = var_1186)[name = tensor("cross_9")]; + tensor v_9 = transpose(perm = v_9_perm_0, x = var_1206)[name = tensor("transpose_25")]; + tensor inner_9 = matmul(transpose_x = inner_9_transpose_x_0, transpose_y = inner_9_transpose_y_0, x = var_1224, y = v_9)[name = tensor("inner_9")]; + tensor var_1226_transpose_x_0 = const()[name = tensor("op_1226_transpose_x_0"), val = tensor(false)]; + tensor var_1226_transpose_y_0 = const()[name = tensor("op_1226_transpose_y_0"), val = tensor(false)]; + tensor var_1226 = matmul(transpose_x = var_1226_transpose_x_0, transpose_y = var_1226_transpose_y_0, x = q_9, y = prev_kv_9)[name = tensor("op_1226")]; + tensor var_1227 = real_div(x = sqrt_s0_9, y = sqrt_s_t_9)[name = tensor("op_1227")]; + tensor var_1228 = const()[name = tensor("op_1228"), val = tensor([1, 1, 5, 1])]; + tensor var_1229 = reshape(shape = var_1228, x = var_1227)[name = tensor("op_1229")]; + tensor cross_9 = mul(x = var_1226, y = var_1229)[name = tensor("cross_9")]; tensor out_25 = add(x = inner_9, y = cross_9)[name = tensor("out_25")]; - tensor var_1189 = const()[name = tensor("op_1189"), val = tensor([1, 1, 5, 1])]; - tensor var_1190 = reshape(shape = var_1189, x = valid_mask)[name = tensor("op_1190")]; - tensor v_masked_1 = mul(x = v_9, y = var_1190)[name = tensor("v_masked_1")]; - tensor var_1192 = mul(x = prev_kv_9, y = sqrt_s0_9)[name = tensor("op_1192")]; - tensor var_1194_transpose_x_1 = const()[name = tensor("op_1194_transpose_x_1"), val = tensor(true)]; - tensor var_1194_transpose_y_1 = const()[name = tensor("op_1194_transpose_y_1"), val = tensor(false)]; - tensor var_1194 = matmul(transpose_x = var_1194_transpose_x_1, transpose_y = var_1194_transpose_y_1, x = k_9, y = v_masked_1)[name = tensor("op_1194")]; - tensor new_kv_unnorm_9 = add(x = var_1192, y = var_1194)[name = tensor("new_kv_unnorm_9")]; - tensor var_1196_keep_dims_0 = const()[name = tensor("op_1196_keep_dims_0"), val = tensor(false)]; - tensor var_1196 = reduce_sum(keep_dims = var_1196_keep_dims_0, x = valid_mask)[name = tensor("op_1196")]; - tensor var_1197 = const()[name = tensor("op_1197"), val = tensor([1])]; - tensor var_1198 = reshape(shape = var_1197, x = var_1196)[name = tensor("op_1198")]; - tensor new_scale_9 = add(x = prev_scale_9, y = var_1198)[name = tensor("new_scale_9")]; + tensor var_1232 = const()[name = tensor("op_1232"), val = tensor([1, 1, 5, 1])]; + tensor var_1233 = reshape(shape = var_1232, x = valid_mask)[name = tensor("op_1233")]; + tensor v_masked_1 = mul(x = v_9, y = var_1233)[name = tensor("v_masked_1")]; + tensor var_1235 = mul(x = prev_kv_9, y = sqrt_s0_9)[name = tensor("op_1235")]; + tensor var_1237_transpose_x_1 = const()[name = tensor("op_1237_transpose_x_1"), val = tensor(true)]; + tensor var_1237_transpose_y_1 = const()[name = tensor("op_1237_transpose_y_1"), val = tensor(false)]; + tensor var_1237 = matmul(transpose_x = var_1237_transpose_x_1, transpose_y = var_1237_transpose_y_1, x = k_9, y = v_masked_1)[name = tensor("op_1237")]; + tensor new_kv_unnorm_9 = add(x = var_1235, y = var_1237)[name = tensor("new_kv_unnorm_9")]; + tensor var_1239_keep_dims_0 = const()[name = tensor("op_1239_keep_dims_0"), val = tensor(false)]; + tensor var_1239 = reduce_sum(keep_dims = var_1239_keep_dims_0, x = valid_mask)[name = tensor("op_1239")]; + tensor var_1240 = const()[name = tensor("op_1240"), val = tensor([1])]; + tensor var_1241 = reshape(shape = var_1240, x = var_1239)[name = tensor("op_1241")]; + tensor new_scale_9 = add(x = prev_scale_9, y = var_1241)[name = tensor("new_scale_9")]; tensor const_21 = const()[name = tensor("const_21"), val = tensor(0x1.fffffep+127)]; - tensor clip_2 = clip(alpha = var_1055, beta = const_21, x = new_scale_9)[name = tensor("clip_2")]; + tensor clip_2 = clip(alpha = var_69, beta = const_21, x = new_scale_9)[name = tensor("clip_2")]; tensor sqrt_new_scale_1 = sqrt(x = clip_2)[name = tensor("sqrt_new_scale_1")]; - tensor var_1202 = real_div(x = new_kv_unnorm_9, y = sqrt_new_scale_1)[name = tensor("op_1202")]; - tensor var_1203_perm_0 = const()[name = tensor("op_1203_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_1245 = real_div(x = new_kv_unnorm_9, y = sqrt_new_scale_1)[name = tensor("op_1245")]; + tensor var_1246_perm_0 = const()[name = tensor("op_1246_perm_0"), val = tensor([0, 2, 1, 3])]; tensor out_27_axes_0 = const()[name = tensor("out_27_axes_0"), val = tensor([-1])]; - tensor var_1203 = transpose(perm = var_1203_perm_0, x = out_25)[name = tensor("transpose_24")]; - tensor out_27 = layer_norm(axes = out_27_axes_0, epsilon = var_1052, x = var_1203)[name = tensor("out_27")]; - tensor var_1207 = const()[name = tensor("op_1207"), val = tensor([12, 5, 256])]; - tensor out_29 = reshape(shape = var_1207, x = out_27)[name = tensor("out_29")]; - tensor var_1209 = silu(x = input_169)[name = tensor("op_1209")]; - tensor input_171 = mul(x = var_1209, y = out_29)[name = tensor("input_171")]; - tensor ret_out_9 = linear(bias = decoder_out_proj_0_bias, weight = decoder_out_proj_0_weight, x = input_171)[name = tensor("linear_42")]; - tensor input_173 = add(x = x_29, y = ret_out_9)[name = tensor("input_173")]; + tensor var_1246 = transpose(perm = var_1246_perm_0, x = out_25)[name = tensor("transpose_24")]; + tensor out_27 = layer_norm(axes = out_27_axes_0, epsilon = var_84, x = var_1246)[name = tensor("out_27")]; + tensor var_1250 = const()[name = tensor("op_1250"), val = tensor([12, 5, 256])]; + tensor out_29 = reshape(shape = var_1250, x = out_27)[name = tensor("out_29")]; + tensor var_1252 = silu(x = input_171)[name = tensor("op_1252")]; + tensor input_173 = mul(x = var_1252, y = out_29)[name = tensor("input_173")]; + tensor ret_out_9 = linear(bias = inner_decoder_out_proj_0_bias, weight = inner_decoder_out_proj_0_weight, x = input_173)[name = tensor("linear_42")]; + tensor input_175 = add(x = x_29, y = ret_out_9)[name = tensor("input_175")]; tensor xt_1_axes_0 = const()[name = tensor("xt_1_axes_0"), val = tensor([-1])]; - tensor xt_1 = layer_norm(axes = xt_1_axes_0, beta = decoder_norm11_0_bias, epsilon = var_1050, gamma = decoder_norm11_0_weight, x = input_173)[name = tensor("xt_1")]; - tensor var_1219 = const()[name = tensor("op_1219"), val = tensor([1, 12, 5, 256])]; - tensor var_1220 = reshape(shape = var_1219, x = xt_1)[name = tensor("op_1220")]; - tensor var_1221_perm_0 = const()[name = tensor("op_1221_perm_0"), val = tensor([0, 2, 1, 3])]; - tensor var_1224 = const()[name = tensor("op_1224"), val = tensor([5, 12, 256])]; - tensor var_1221 = transpose(perm = var_1221_perm_0, x = var_1220)[name = tensor("transpose_23")]; - tensor query_1 = reshape(shape = var_1224, x = var_1221)[name = tensor("query_1")]; + tensor xt_1 = layer_norm(axes = xt_1_axes_0, beta = inner_decoder_norm11_0_bias, epsilon = var_76, gamma = inner_decoder_norm11_0_weight, x = input_175)[name = tensor("xt_1")]; + tensor var_1262 = const()[name = tensor("op_1262"), val = tensor([1, 12, 5, 256])]; + tensor var_1263 = reshape(shape = var_1262, x = xt_1)[name = tensor("op_1263")]; + tensor var_1264_perm_0 = const()[name = tensor("op_1264_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_1267 = const()[name = tensor("op_1267"), val = tensor([5, 12, 256])]; + tensor var_1264 = transpose(perm = var_1264_perm_0, x = var_1263)[name = tensor("transpose_23")]; + tensor query_1 = reshape(shape = var_1267, x = var_1264)[name = tensor("query_1")]; tensor query_3_perm_0 = const()[name = tensor("query_3_perm_0"), val = tensor([1, 0, 2])]; tensor query_3 = transpose(perm = query_3_perm_0, x = query_1)[name = tensor("transpose_22")]; - tensor var_1247 = linear(bias = decoder_self_attn2_0_in_proj_bias, weight = decoder_self_attn2_0_in_proj_weight, x = query_3)[name = tensor("linear_43")]; + tensor var_1290 = linear(bias = inner_decoder_self_attn2_0_in_proj_bias, weight = inner_decoder_self_attn2_0_in_proj_weight, x = query_3)[name = tensor("linear_43")]; tensor concat_1 = const()[name = tensor("concat_1"), val = tensor([12, 5, 3, 256])]; - tensor var_1249 = reshape(shape = concat_1, x = var_1247)[name = tensor("op_1249")]; - tensor var_1250_axes_0 = const()[name = tensor("op_1250_axes_0"), val = tensor([0])]; - tensor var_1250 = expand_dims(axes = var_1250_axes_0, x = var_1249)[name = tensor("op_1250")]; - tensor var_1251_perm_0 = const()[name = tensor("op_1251_perm_0"), val = tensor([-2, 1, 2, 0, 4])]; - tensor var_1252_axes_0 = const()[name = tensor("op_1252_axes_0"), val = tensor([-2])]; - tensor var_1251 = transpose(perm = var_1251_perm_0, x = var_1250)[name = tensor("transpose_21")]; - tensor var_1252 = squeeze(axes = var_1252_axes_0, x = var_1251)[name = tensor("op_1252")]; + tensor var_1292 = reshape(shape = concat_1, x = var_1290)[name = tensor("op_1292")]; + tensor var_1293_axes_0 = const()[name = tensor("op_1293_axes_0"), val = tensor([0])]; + tensor var_1293 = expand_dims(axes = var_1293_axes_0, x = var_1292)[name = tensor("op_1293")]; + tensor var_1294_perm_0 = const()[name = tensor("op_1294_perm_0"), val = tensor([-2, 1, 2, 0, 4])]; + tensor var_1295_axes_0 = const()[name = tensor("op_1295_axes_0"), val = tensor([-2])]; + tensor var_1294 = transpose(perm = var_1294_perm_0, x = var_1293)[name = tensor("transpose_21")]; + tensor var_1295 = squeeze(axes = var_1295_axes_0, x = var_1294)[name = tensor("op_1295")]; tensor q_11_begin_0 = const()[name = tensor("q_11_begin_0"), val = tensor([0, 0, 0, 0])]; tensor q_11_end_0 = const()[name = tensor("q_11_end_0"), val = tensor([1, 12, 5, 256])]; tensor q_11_end_mask_0 = const()[name = tensor("q_11_end_mask_0"), val = tensor([false, true, true, true])]; tensor q_11_squeeze_mask_0 = const()[name = tensor("q_11_squeeze_mask_0"), val = tensor([true, false, false, false])]; - tensor q_11 = slice_by_index(begin = q_11_begin_0, end = q_11_end_0, end_mask = q_11_end_mask_0, squeeze_mask = q_11_squeeze_mask_0, x = var_1252)[name = tensor("q_11")]; + tensor q_11 = slice_by_index(begin = q_11_begin_0, end = q_11_end_0, end_mask = q_11_end_mask_0, squeeze_mask = q_11_squeeze_mask_0, x = var_1295)[name = tensor("q_11")]; tensor k_11_begin_0 = const()[name = tensor("k_11_begin_0"), val = tensor([1, 0, 0, 0])]; tensor k_11_end_0 = const()[name = tensor("k_11_end_0"), val = tensor([2, 12, 5, 256])]; tensor k_11_end_mask_0 = const()[name = tensor("k_11_end_mask_0"), val = tensor([false, true, true, true])]; tensor k_11_squeeze_mask_0 = const()[name = tensor("k_11_squeeze_mask_0"), val = tensor([true, false, false, false])]; - tensor k_11 = slice_by_index(begin = k_11_begin_0, end = k_11_end_0, end_mask = k_11_end_mask_0, squeeze_mask = k_11_squeeze_mask_0, x = var_1252)[name = tensor("k_11")]; + tensor k_11 = slice_by_index(begin = k_11_begin_0, end = k_11_end_0, end_mask = k_11_end_mask_0, squeeze_mask = k_11_squeeze_mask_0, x = var_1295)[name = tensor("k_11")]; tensor v_11_begin_0 = const()[name = tensor("v_11_begin_0"), val = tensor([2, 0, 0, 0])]; tensor v_11_end_0 = const()[name = tensor("v_11_end_0"), val = tensor([3, 12, 5, 256])]; tensor v_11_end_mask_0 = const()[name = tensor("v_11_end_mask_0"), val = tensor([false, true, true, true])]; tensor v_11_squeeze_mask_0 = const()[name = tensor("v_11_squeeze_mask_0"), val = tensor([true, false, false, false])]; - tensor v_11 = slice_by_index(begin = v_11_begin_0, end = v_11_end_0, end_mask = v_11_end_mask_0, squeeze_mask = v_11_squeeze_mask_0, x = var_1252)[name = tensor("v_11")]; - tensor var_1260 = const()[name = tensor("op_1260"), val = tensor([12, 20, 64])]; - tensor var_1261 = reshape(shape = var_1260, x = q_11)[name = tensor("op_1261")]; + tensor v_11 = slice_by_index(begin = v_11_begin_0, end = v_11_end_0, end_mask = v_11_end_mask_0, squeeze_mask = v_11_squeeze_mask_0, x = var_1295)[name = tensor("v_11")]; + tensor var_1303 = const()[name = tensor("op_1303"), val = tensor([12, 20, 64])]; + tensor var_1304 = reshape(shape = var_1303, x = q_11)[name = tensor("op_1304")]; tensor q_13_perm_0 = const()[name = tensor("q_13_perm_0"), val = tensor([1, 0, 2])]; - tensor var_1267 = const()[name = tensor("op_1267"), val = tensor([12, 20, 64])]; - tensor var_1268 = reshape(shape = var_1267, x = k_11)[name = tensor("op_1268")]; + tensor var_1310 = const()[name = tensor("op_1310"), val = tensor([12, 20, 64])]; + tensor var_1311 = reshape(shape = var_1310, x = k_11)[name = tensor("op_1311")]; tensor k_13_perm_0 = const()[name = tensor("k_13_perm_0"), val = tensor([1, 0, 2])]; - tensor var_1274 = const()[name = tensor("op_1274"), val = tensor([12, 20, 64])]; - tensor var_1275 = reshape(shape = var_1274, x = v_11)[name = tensor("op_1275")]; + tensor var_1317 = const()[name = tensor("op_1317"), val = tensor([12, 20, 64])]; + tensor var_1318 = reshape(shape = var_1317, x = v_11)[name = tensor("op_1318")]; tensor v_13_perm_0 = const()[name = tensor("v_13_perm_0"), val = tensor([1, 0, 2])]; - tensor var_1278 = const()[name = tensor("op_1278"), val = tensor([5, 4, 12, 64])]; - tensor q_13 = transpose(perm = q_13_perm_0, x = var_1261)[name = tensor("transpose_20")]; - tensor q_15 = reshape(shape = var_1278, x = q_13)[name = tensor("q_15")]; - tensor var_1280 = const()[name = tensor("op_1280"), val = tensor([5, 4, 12, 64])]; - tensor k_13 = transpose(perm = k_13_perm_0, x = var_1268)[name = tensor("transpose_19")]; - tensor k_15 = reshape(shape = var_1280, x = k_13)[name = tensor("k_15")]; - tensor var_1282 = const()[name = tensor("op_1282"), val = tensor([5, 4, 12, 64])]; - tensor v_13 = transpose(perm = v_13_perm_0, x = var_1275)[name = tensor("transpose_18")]; - tensor v_15 = reshape(shape = var_1282, x = v_13)[name = tensor("v_15")]; + tensor var_1321 = const()[name = tensor("op_1321"), val = tensor([5, 4, 12, 64])]; + tensor q_13 = transpose(perm = q_13_perm_0, x = var_1304)[name = tensor("transpose_20")]; + tensor q_15 = reshape(shape = var_1321, x = q_13)[name = tensor("q_15")]; + tensor var_1323 = const()[name = tensor("op_1323"), val = tensor([5, 4, 12, 64])]; + tensor k_13 = transpose(perm = k_13_perm_0, x = var_1311)[name = tensor("transpose_19")]; + tensor k_15 = reshape(shape = var_1323, x = k_13)[name = tensor("k_15")]; + tensor var_1325 = const()[name = tensor("op_1325"), val = tensor([5, 4, 12, 64])]; + tensor v_13 = transpose(perm = v_13_perm_0, x = var_1318)[name = tensor("transpose_18")]; + tensor v_15 = reshape(shape = var_1325, x = v_13)[name = tensor("v_15")]; tensor mul_1_y_0 = const()[name = tensor("mul_1_y_0"), val = tensor(0x1p-3)]; tensor mul_1 = mul(x = q_15, y = mul_1_y_0)[name = tensor("mul_1")]; tensor matmul_0_transpose_y_0 = const()[name = tensor("matmul_0_transpose_y_0"), val = tensor(true)]; @@ -1159,30 +1178,30 @@ program(1.0) tensor attn_output_1_transpose_x_0 = const()[name = tensor("attn_output_1_transpose_x_0"), val = tensor(false)]; tensor attn_output_1_transpose_y_0 = const()[name = tensor("attn_output_1_transpose_y_0"), val = tensor(false)]; tensor attn_output_1 = matmul(transpose_x = attn_output_1_transpose_x_0, transpose_y = attn_output_1_transpose_y_0, x = softmax_0, y = v_15)[name = tensor("attn_output_1")]; - tensor var_1285 = const()[name = tensor("op_1285"), val = tensor([2, 0, 1, 3])]; - tensor var_1290 = const()[name = tensor("op_1290"), val = tensor([60, 256])]; - tensor var_1286 = transpose(perm = var_1285, x = attn_output_1)[name = tensor("transpose_17")]; - tensor attn_output_3 = reshape(shape = var_1290, x = var_1286)[name = tensor("attn_output_3")]; - tensor attn_output_5 = linear(bias = decoder_self_attn2_0_out_proj_bias, weight = decoder_self_attn2_0_out_proj_weight, x = attn_output_3)[name = tensor("linear_44")]; - tensor var_1294 = const()[name = tensor("op_1294"), val = tensor([12, 5, 256])]; - tensor attn_output_7 = reshape(shape = var_1294, x = attn_output_5)[name = tensor("attn_output_7")]; + tensor var_1328 = const()[name = tensor("op_1328"), val = tensor([2, 0, 1, 3])]; + tensor var_1333 = const()[name = tensor("op_1333"), val = tensor([60, 256])]; + tensor var_1329 = transpose(perm = var_1328, x = attn_output_1)[name = tensor("transpose_17")]; + tensor attn_output_3 = reshape(shape = var_1333, x = var_1329)[name = tensor("attn_output_3")]; + tensor attn_output_5 = linear(bias = inner_decoder_self_attn2_0_out_proj_bias, weight = inner_decoder_self_attn2_0_out_proj_weight, x = attn_output_3)[name = tensor("linear_44")]; + tensor var_1337 = const()[name = tensor("op_1337"), val = tensor([12, 5, 256])]; + tensor attn_output_7 = reshape(shape = var_1337, x = attn_output_5)[name = tensor("attn_output_7")]; tensor sa2_out_1_perm_0 = const()[name = tensor("sa2_out_1_perm_0"), val = tensor([1, 0, 2])]; tensor sa2_out_1 = transpose(perm = sa2_out_1_perm_0, x = attn_output_7)[name = tensor("transpose_16")]; - tensor input_175 = add(x = query_1, y = sa2_out_1)[name = tensor("input_175")]; - tensor input_177_axes_0 = const()[name = tensor("input_177_axes_0"), val = tensor([-1])]; - tensor input_177 = layer_norm(axes = input_177_axes_0, beta = decoder_norm21_0_bias, epsilon = var_1050, gamma = decoder_norm21_0_weight, x = input_175)[name = tensor("input_177")]; - tensor input_179 = linear(bias = decoder_linear1_0_bias, weight = decoder_linear1_0_weight, x = input_177)[name = tensor("linear_45")]; - tensor input_181 = relu(x = input_179)[name = tensor("input_181")]; - tensor ffn_out_1 = linear(bias = decoder_linear2_0_bias, weight = decoder_linear2_0_weight, x = input_181)[name = tensor("linear_46")]; - tensor input_183 = add(x = input_177, y = ffn_out_1)[name = tensor("input_183")]; + tensor input_177 = add(x = query_1, y = sa2_out_1)[name = tensor("input_177")]; + tensor input_179_axes_0 = const()[name = tensor("input_179_axes_0"), val = tensor([-1])]; + tensor input_179 = layer_norm(axes = input_179_axes_0, beta = inner_decoder_norm21_0_bias, epsilon = var_76, gamma = inner_decoder_norm21_0_weight, x = input_177)[name = tensor("input_179")]; + tensor input_181 = linear(bias = inner_decoder_linear1_0_bias, weight = inner_decoder_linear1_0_weight, x = input_179)[name = tensor("linear_45")]; + tensor input_183 = relu(x = input_181)[name = tensor("input_183")]; + tensor ffn_out_1 = linear(bias = inner_decoder_linear2_0_bias, weight = inner_decoder_linear2_0_weight, x = input_183)[name = tensor("linear_46")]; + tensor input_185 = add(x = input_179, y = ffn_out_1)[name = tensor("input_185")]; tensor xt_3_axes_0 = const()[name = tensor("xt_3_axes_0"), val = tensor([-1])]; - tensor xt_3 = layer_norm(axes = xt_3_axes_0, beta = decoder_norm22_0_bias, epsilon = var_1050, gamma = decoder_norm22_0_weight, x = input_183)[name = tensor("xt_3")]; - tensor var_1314 = const()[name = tensor("op_1314"), val = tensor([1, 5, 12, 256])]; - tensor x_31 = reshape(shape = var_1314, x = xt_3)[name = tensor("x_31")]; - tensor var_1316_perm_0 = const()[name = tensor("op_1316_perm_0"), val = tensor([0, 2, 1, 3])]; - tensor var_1320 = const()[name = tensor("op_1320"), val = tensor([12, 5, 256])]; - tensor var_1316 = transpose(perm = var_1316_perm_0, x = x_31)[name = tensor("transpose_15")]; - tensor x = reshape(shape = var_1320, x = var_1316)[name = tensor("x")]; + tensor xt_3 = layer_norm(axes = xt_3_axes_0, beta = inner_decoder_norm22_0_bias, epsilon = var_76, gamma = inner_decoder_norm22_0_weight, x = input_185)[name = tensor("xt_3")]; + tensor var_1357 = const()[name = tensor("op_1357"), val = tensor([1, 5, 12, 256])]; + tensor x_31 = reshape(shape = var_1357, x = xt_3)[name = tensor("x_31")]; + tensor var_1359_perm_0 = const()[name = tensor("op_1359_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_1363 = const()[name = tensor("op_1363"), val = tensor([12, 5, 256])]; + tensor var_1359 = transpose(perm = var_1359_perm_0, x = x_31)[name = tensor("transpose_15")]; + tensor x = reshape(shape = var_1363, x = var_1359)[name = tensor("x")]; tensor prev_kv_begin_0 = const()[name = tensor("prev_kv_begin_0"), val = tensor([1, 0, 0, 0, 0])]; tensor prev_kv_end_0 = const()[name = tensor("prev_kv_end_0"), val = tensor([2, 12, 4, 64, 64])]; tensor prev_kv_end_mask_0 = const()[name = tensor("prev_kv_end_mask_0"), val = tensor([false, true, true, true, true])]; @@ -1193,120 +1212,120 @@ program(1.0) tensor prev_scale_end_mask_0 = const()[name = tensor("prev_scale_end_mask_0"), val = tensor([false, true])]; tensor prev_scale_squeeze_mask_0 = const()[name = tensor("prev_scale_squeeze_mask_0"), val = tensor([true, false])]; tensor prev_scale = slice_by_index(begin = prev_scale_begin_0, end = prev_scale_end_0, end_mask = prev_scale_end_mask_0, squeeze_mask = prev_scale_squeeze_mask_0, x = dec_scale)[name = tensor("prev_scale")]; - tensor var_1328 = linear(bias = decoder_q_proj_1_bias, weight = decoder_q_proj_1_weight, x = x)[name = tensor("linear_47")]; - tensor var_1329 = const()[name = tensor("op_1329"), val = tensor([12, 5, 4, 64])]; - tensor var_1330 = reshape(shape = var_1329, x = var_1328)[name = tensor("op_1330")]; + tensor var_1371 = linear(bias = inner_decoder_q_proj_1_bias, weight = inner_decoder_q_proj_1_weight, x = x)[name = tensor("linear_47")]; + tensor var_1372 = const()[name = tensor("op_1372"), val = tensor([12, 5, 4, 64])]; + tensor var_1373 = reshape(shape = var_1372, x = var_1371)[name = tensor("op_1373")]; tensor q_17_perm_0 = const()[name = tensor("q_17_perm_0"), val = tensor([0, 2, 1, 3])]; - tensor var_1334 = linear(bias = decoder_k_proj_1_bias, weight = decoder_k_proj_1_weight, x = x)[name = tensor("linear_48")]; - tensor var_1335 = const()[name = tensor("op_1335"), val = tensor(0x1p-3)]; - tensor var_1336 = mul(x = var_1334, y = var_1335)[name = tensor("op_1336")]; - tensor var_1337 = const()[name = tensor("op_1337"), val = tensor([12, 5, 4, 64])]; - tensor var_1338 = reshape(shape = var_1337, x = var_1336)[name = tensor("op_1338")]; + tensor var_1377 = linear(bias = inner_decoder_k_proj_1_bias, weight = inner_decoder_k_proj_1_weight, x = x)[name = tensor("linear_48")]; + tensor var_1378 = const()[name = tensor("op_1378"), val = tensor(0x1p-3)]; + tensor var_1379 = mul(x = var_1377, y = var_1378)[name = tensor("op_1379")]; + tensor var_1380 = const()[name = tensor("op_1380"), val = tensor([12, 5, 4, 64])]; + tensor var_1381 = reshape(shape = var_1380, x = var_1379)[name = tensor("op_1381")]; tensor k_17_perm_0 = const()[name = tensor("k_17_perm_0"), val = tensor([0, 2, 1, 3])]; - tensor var_1342 = linear(bias = decoder_v_proj_1_bias, weight = decoder_v_proj_1_weight, x = x)[name = tensor("linear_49")]; - tensor var_1343 = const()[name = tensor("op_1343"), val = tensor([12, 5, 4, 64])]; - tensor var_1344 = reshape(shape = var_1343, x = var_1342)[name = tensor("op_1344")]; + tensor var_1385 = linear(bias = inner_decoder_v_proj_1_bias, weight = inner_decoder_v_proj_1_weight, x = x)[name = tensor("linear_49")]; + tensor var_1386 = const()[name = tensor("op_1386"), val = tensor([12, 5, 4, 64])]; + tensor var_1387 = reshape(shape = var_1386, x = var_1385)[name = tensor("op_1387")]; tensor v_17_perm_0 = const()[name = tensor("v_17_perm_0"), val = tensor([0, 2, 1, 3])]; - tensor input_187 = linear(bias = decoder_g_proj_1_bias, weight = decoder_g_proj_1_weight, x = x)[name = tensor("linear_50")]; + tensor input_189 = linear(bias = inner_decoder_g_proj_1_bias, weight = inner_decoder_g_proj_1_weight, x = x)[name = tensor("linear_50")]; tensor sqrt_s0 = sqrt(x = prev_scale)[name = tensor("sqrt_s0")]; tensor s_t = add(x = prev_scale, y = cumsum_mask_1)[name = tensor("s_t")]; tensor const_32 = const()[name = tensor("const_32"), val = tensor(0x1.fffffep+127)]; - tensor clip_3 = clip(alpha = var_1055, beta = const_32, x = s_t)[name = tensor("clip_3")]; + tensor clip_3 = clip(alpha = var_69, beta = const_32, x = s_t)[name = tensor("clip_3")]; tensor sqrt_s_t = sqrt(x = clip_3)[name = tensor("sqrt_s_t")]; tensor qk_transpose_x_1 = const()[name = tensor("qk_transpose_x_1"), val = tensor(false)]; tensor qk_transpose_y_1 = const()[name = tensor("qk_transpose_y_1"), val = tensor(true)]; - tensor k_17 = transpose(perm = k_17_perm_0, x = var_1338)[name = tensor("transpose_13")]; - tensor q_17 = transpose(perm = q_17_perm_0, x = var_1330)[name = tensor("transpose_14")]; + tensor k_17 = transpose(perm = k_17_perm_0, x = var_1381)[name = tensor("transpose_13")]; + tensor q_17 = transpose(perm = q_17_perm_0, x = var_1373)[name = tensor("transpose_14")]; tensor qk = matmul(transpose_x = qk_transpose_x_1, transpose_y = qk_transpose_y_1, x = q_17, y = k_17)[name = tensor("qk")]; - tensor var_1359 = const()[name = tensor("op_1359"), val = tensor([5, 1])]; - tensor var_1360 = reshape(shape = var_1359, x = sqrt_s_t)[name = tensor("op_1360")]; - tensor M = real_div(x = causal_with_valid_1, y = var_1360)[name = tensor("M")]; - tensor var_1362 = mul(x = qk, y = M)[name = tensor("op_1362")]; - tensor inner_transpose_x_0 = const()[name = tensor("inner_transpose_x_0"), val = tensor(false)]; - tensor inner_transpose_y_0 = const()[name = tensor("inner_transpose_y_0"), val = tensor(false)]; - tensor v_17 = transpose(perm = v_17_perm_0, x = var_1344)[name = tensor("transpose_12")]; - tensor inner = matmul(transpose_x = inner_transpose_x_0, transpose_y = inner_transpose_y_0, x = var_1362, y = v_17)[name = tensor("inner")]; - tensor var_1364_transpose_x_0 = const()[name = tensor("op_1364_transpose_x_0"), val = tensor(false)]; - tensor var_1364_transpose_y_0 = const()[name = tensor("op_1364_transpose_y_0"), val = tensor(false)]; - tensor var_1364 = matmul(transpose_x = var_1364_transpose_x_0, transpose_y = var_1364_transpose_y_0, x = q_17, y = prev_kv)[name = tensor("op_1364")]; - tensor var_1365 = real_div(x = sqrt_s0, y = sqrt_s_t)[name = tensor("op_1365")]; - tensor var_1366 = const()[name = tensor("op_1366"), val = tensor([1, 1, 5, 1])]; - tensor var_1367 = reshape(shape = var_1366, x = var_1365)[name = tensor("op_1367")]; - tensor cross = mul(x = var_1364, y = var_1367)[name = tensor("cross")]; - tensor out_31 = add(x = inner, y = cross)[name = tensor("out_31")]; - tensor v_masked = mul(x = v_17, y = var_1190)[name = tensor("v_masked")]; - tensor var_1373 = mul(x = prev_kv, y = sqrt_s0)[name = tensor("op_1373")]; - tensor var_1375_transpose_x_1 = const()[name = tensor("op_1375_transpose_x_1"), val = tensor(true)]; - tensor var_1375_transpose_y_1 = const()[name = tensor("op_1375_transpose_y_1"), val = tensor(false)]; - tensor var_1375 = matmul(transpose_x = var_1375_transpose_x_1, transpose_y = var_1375_transpose_y_1, x = k_17, y = v_masked)[name = tensor("op_1375")]; - tensor new_kv_unnorm = add(x = var_1373, y = var_1375)[name = tensor("new_kv_unnorm")]; - tensor new_scale = add(x = prev_scale, y = var_1198)[name = tensor("new_scale")]; + tensor var_1402 = const()[name = tensor("op_1402"), val = tensor([5, 1])]; + tensor var_1403 = reshape(shape = var_1402, x = sqrt_s_t)[name = tensor("op_1403")]; + tensor M = real_div(x = causal_with_valid_1, y = var_1403)[name = tensor("M")]; + tensor var_1405 = mul(x = qk, y = M)[name = tensor("op_1405")]; + tensor inner_11_transpose_x_0 = const()[name = tensor("inner_11_transpose_x_0"), val = tensor(false)]; + tensor inner_11_transpose_y_0 = const()[name = tensor("inner_11_transpose_y_0"), val = tensor(false)]; + tensor v_17 = transpose(perm = v_17_perm_0, x = var_1387)[name = tensor("transpose_12")]; + tensor inner_11 = matmul(transpose_x = inner_11_transpose_x_0, transpose_y = inner_11_transpose_y_0, x = var_1405, y = v_17)[name = tensor("inner_11")]; + tensor var_1407_transpose_x_0 = const()[name = tensor("op_1407_transpose_x_0"), val = tensor(false)]; + tensor var_1407_transpose_y_0 = const()[name = tensor("op_1407_transpose_y_0"), val = tensor(false)]; + tensor var_1407 = matmul(transpose_x = var_1407_transpose_x_0, transpose_y = var_1407_transpose_y_0, x = q_17, y = prev_kv)[name = tensor("op_1407")]; + tensor var_1408 = real_div(x = sqrt_s0, y = sqrt_s_t)[name = tensor("op_1408")]; + tensor var_1409 = const()[name = tensor("op_1409"), val = tensor([1, 1, 5, 1])]; + tensor var_1410 = reshape(shape = var_1409, x = var_1408)[name = tensor("op_1410")]; + tensor cross = mul(x = var_1407, y = var_1410)[name = tensor("cross")]; + tensor out_31 = add(x = inner_11, y = cross)[name = tensor("out_31")]; + tensor v_masked = mul(x = v_17, y = var_1233)[name = tensor("v_masked")]; + tensor var_1416 = mul(x = prev_kv, y = sqrt_s0)[name = tensor("op_1416")]; + tensor var_1418_transpose_x_1 = const()[name = tensor("op_1418_transpose_x_1"), val = tensor(true)]; + tensor var_1418_transpose_y_1 = const()[name = tensor("op_1418_transpose_y_1"), val = tensor(false)]; + tensor var_1418 = matmul(transpose_x = var_1418_transpose_x_1, transpose_y = var_1418_transpose_y_1, x = k_17, y = v_masked)[name = tensor("op_1418")]; + tensor new_kv_unnorm = add(x = var_1416, y = var_1418)[name = tensor("new_kv_unnorm")]; + tensor new_scale = add(x = prev_scale, y = var_1241)[name = tensor("new_scale")]; tensor const_33 = const()[name = tensor("const_33"), val = tensor(0x1.fffffep+127)]; - tensor clip_4 = clip(alpha = var_1055, beta = const_33, x = new_scale)[name = tensor("clip_4")]; + tensor clip_4 = clip(alpha = var_69, beta = const_33, x = new_scale)[name = tensor("clip_4")]; tensor sqrt_new_scale = sqrt(x = clip_4)[name = tensor("sqrt_new_scale")]; tensor nkv = real_div(x = new_kv_unnorm, y = sqrt_new_scale)[name = tensor("nkv")]; - tensor var_1384_perm_0 = const()[name = tensor("op_1384_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_1427_perm_0 = const()[name = tensor("op_1427_perm_0"), val = tensor([0, 2, 1, 3])]; tensor out_33_axes_0 = const()[name = tensor("out_33_axes_0"), val = tensor([-1])]; - tensor var_1384 = transpose(perm = var_1384_perm_0, x = out_31)[name = tensor("transpose_11")]; - tensor out_33 = layer_norm(axes = out_33_axes_0, epsilon = var_1052, x = var_1384)[name = tensor("out_33")]; - tensor var_1388 = const()[name = tensor("op_1388"), val = tensor([12, 5, 256])]; - tensor out = reshape(shape = var_1388, x = out_33)[name = tensor("out")]; - tensor var_1390 = silu(x = input_187)[name = tensor("op_1390")]; - tensor input_189 = mul(x = var_1390, y = out)[name = tensor("input_189")]; - tensor ret_out = linear(bias = decoder_out_proj_1_bias, weight = decoder_out_proj_1_weight, x = input_189)[name = tensor("linear_51")]; - tensor input_191 = add(x = x, y = ret_out)[name = tensor("input_191")]; + tensor var_1427 = transpose(perm = var_1427_perm_0, x = out_31)[name = tensor("transpose_11")]; + tensor out_33 = layer_norm(axes = out_33_axes_0, epsilon = var_84, x = var_1427)[name = tensor("out_33")]; + tensor var_1431 = const()[name = tensor("op_1431"), val = tensor([12, 5, 256])]; + tensor out = reshape(shape = var_1431, x = out_33)[name = tensor("out")]; + tensor var_1433 = silu(x = input_189)[name = tensor("op_1433")]; + tensor input_191 = mul(x = var_1433, y = out)[name = tensor("input_191")]; + tensor ret_out = linear(bias = inner_decoder_out_proj_1_bias, weight = inner_decoder_out_proj_1_weight, x = input_191)[name = tensor("linear_51")]; + tensor input_193 = add(x = x, y = ret_out)[name = tensor("input_193")]; tensor xt_5_axes_0 = const()[name = tensor("xt_5_axes_0"), val = tensor([-1])]; - tensor xt_5 = layer_norm(axes = xt_5_axes_0, beta = decoder_norm11_1_bias, epsilon = var_1050, gamma = decoder_norm11_1_weight, x = input_191)[name = tensor("xt_5")]; - tensor var_1400 = const()[name = tensor("op_1400"), val = tensor([1, 12, 5, 256])]; - tensor var_1401 = reshape(shape = var_1400, x = xt_5)[name = tensor("op_1401")]; - tensor var_1402_perm_0 = const()[name = tensor("op_1402_perm_0"), val = tensor([0, 2, 1, 3])]; - tensor var_1405 = const()[name = tensor("op_1405"), val = tensor([5, 12, 256])]; - tensor var_1402 = transpose(perm = var_1402_perm_0, x = var_1401)[name = tensor("transpose_10")]; - tensor query_5 = reshape(shape = var_1405, x = var_1402)[name = tensor("query_5")]; + tensor xt_5 = layer_norm(axes = xt_5_axes_0, beta = inner_decoder_norm11_1_bias, epsilon = var_76, gamma = inner_decoder_norm11_1_weight, x = input_193)[name = tensor("xt_5")]; + tensor var_1443 = const()[name = tensor("op_1443"), val = tensor([1, 12, 5, 256])]; + tensor var_1444 = reshape(shape = var_1443, x = xt_5)[name = tensor("op_1444")]; + tensor var_1445_perm_0 = const()[name = tensor("op_1445_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_1448 = const()[name = tensor("op_1448"), val = tensor([5, 12, 256])]; + tensor var_1445 = transpose(perm = var_1445_perm_0, x = var_1444)[name = tensor("transpose_10")]; + tensor query_5 = reshape(shape = var_1448, x = var_1445)[name = tensor("query_5")]; tensor query_perm_0 = const()[name = tensor("query_perm_0"), val = tensor([1, 0, 2])]; tensor query = transpose(perm = query_perm_0, x = query_5)[name = tensor("transpose_9")]; - tensor var_1428 = linear(bias = decoder_self_attn2_1_in_proj_bias, weight = decoder_self_attn2_1_in_proj_weight, x = query)[name = tensor("linear_52")]; + tensor var_1471 = linear(bias = inner_decoder_self_attn2_1_in_proj_bias, weight = inner_decoder_self_attn2_1_in_proj_weight, x = query)[name = tensor("linear_52")]; tensor concat_2 = const()[name = tensor("concat_2"), val = tensor([12, 5, 3, 256])]; - tensor var_1430 = reshape(shape = concat_2, x = var_1428)[name = tensor("op_1430")]; - tensor var_1431_axes_0 = const()[name = tensor("op_1431_axes_0"), val = tensor([0])]; - tensor var_1431 = expand_dims(axes = var_1431_axes_0, x = var_1430)[name = tensor("op_1431")]; - tensor var_1432_perm_0 = const()[name = tensor("op_1432_perm_0"), val = tensor([-2, 1, 2, 0, 4])]; - tensor var_1433_axes_0 = const()[name = tensor("op_1433_axes_0"), val = tensor([-2])]; - tensor var_1432 = transpose(perm = var_1432_perm_0, x = var_1431)[name = tensor("transpose_8")]; - tensor var_1433 = squeeze(axes = var_1433_axes_0, x = var_1432)[name = tensor("op_1433")]; + tensor var_1473 = reshape(shape = concat_2, x = var_1471)[name = tensor("op_1473")]; + tensor var_1474_axes_0 = const()[name = tensor("op_1474_axes_0"), val = tensor([0])]; + tensor var_1474 = expand_dims(axes = var_1474_axes_0, x = var_1473)[name = tensor("op_1474")]; + tensor var_1475_perm_0 = const()[name = tensor("op_1475_perm_0"), val = tensor([-2, 1, 2, 0, 4])]; + tensor var_1476_axes_0 = const()[name = tensor("op_1476_axes_0"), val = tensor([-2])]; + tensor var_1475 = transpose(perm = var_1475_perm_0, x = var_1474)[name = tensor("transpose_8")]; + tensor var_1476 = squeeze(axes = var_1476_axes_0, x = var_1475)[name = tensor("op_1476")]; tensor q_19_begin_0 = const()[name = tensor("q_19_begin_0"), val = tensor([0, 0, 0, 0])]; tensor q_19_end_0 = const()[name = tensor("q_19_end_0"), val = tensor([1, 12, 5, 256])]; tensor q_19_end_mask_0 = const()[name = tensor("q_19_end_mask_0"), val = tensor([false, true, true, true])]; tensor q_19_squeeze_mask_0 = const()[name = tensor("q_19_squeeze_mask_0"), val = tensor([true, false, false, false])]; - tensor q_19 = slice_by_index(begin = q_19_begin_0, end = q_19_end_0, end_mask = q_19_end_mask_0, squeeze_mask = q_19_squeeze_mask_0, x = var_1433)[name = tensor("q_19")]; + tensor q_19 = slice_by_index(begin = q_19_begin_0, end = q_19_end_0, end_mask = q_19_end_mask_0, squeeze_mask = q_19_squeeze_mask_0, x = var_1476)[name = tensor("q_19")]; tensor k_19_begin_0 = const()[name = tensor("k_19_begin_0"), val = tensor([1, 0, 0, 0])]; tensor k_19_end_0 = const()[name = tensor("k_19_end_0"), val = tensor([2, 12, 5, 256])]; tensor k_19_end_mask_0 = const()[name = tensor("k_19_end_mask_0"), val = tensor([false, true, true, true])]; tensor k_19_squeeze_mask_0 = const()[name = tensor("k_19_squeeze_mask_0"), val = tensor([true, false, false, false])]; - tensor k_19 = slice_by_index(begin = k_19_begin_0, end = k_19_end_0, end_mask = k_19_end_mask_0, squeeze_mask = k_19_squeeze_mask_0, x = var_1433)[name = tensor("k_19")]; + tensor k_19 = slice_by_index(begin = k_19_begin_0, end = k_19_end_0, end_mask = k_19_end_mask_0, squeeze_mask = k_19_squeeze_mask_0, x = var_1476)[name = tensor("k_19")]; tensor v_19_begin_0 = const()[name = tensor("v_19_begin_0"), val = tensor([2, 0, 0, 0])]; tensor v_19_end_0 = const()[name = tensor("v_19_end_0"), val = tensor([3, 12, 5, 256])]; tensor v_19_end_mask_0 = const()[name = tensor("v_19_end_mask_0"), val = tensor([false, true, true, true])]; tensor v_19_squeeze_mask_0 = const()[name = tensor("v_19_squeeze_mask_0"), val = tensor([true, false, false, false])]; - tensor v_19 = slice_by_index(begin = v_19_begin_0, end = v_19_end_0, end_mask = v_19_end_mask_0, squeeze_mask = v_19_squeeze_mask_0, x = var_1433)[name = tensor("v_19")]; - tensor var_1441 = const()[name = tensor("op_1441"), val = tensor([12, 20, 64])]; - tensor var_1442 = reshape(shape = var_1441, x = q_19)[name = tensor("op_1442")]; + tensor v_19 = slice_by_index(begin = v_19_begin_0, end = v_19_end_0, end_mask = v_19_end_mask_0, squeeze_mask = v_19_squeeze_mask_0, x = var_1476)[name = tensor("v_19")]; + tensor var_1484 = const()[name = tensor("op_1484"), val = tensor([12, 20, 64])]; + tensor var_1485 = reshape(shape = var_1484, x = q_19)[name = tensor("op_1485")]; tensor q_21_perm_0 = const()[name = tensor("q_21_perm_0"), val = tensor([1, 0, 2])]; - tensor var_1448 = const()[name = tensor("op_1448"), val = tensor([12, 20, 64])]; - tensor var_1449 = reshape(shape = var_1448, x = k_19)[name = tensor("op_1449")]; + tensor var_1491 = const()[name = tensor("op_1491"), val = tensor([12, 20, 64])]; + tensor var_1492 = reshape(shape = var_1491, x = k_19)[name = tensor("op_1492")]; tensor k_21_perm_0 = const()[name = tensor("k_21_perm_0"), val = tensor([1, 0, 2])]; - tensor var_1455 = const()[name = tensor("op_1455"), val = tensor([12, 20, 64])]; - tensor var_1456 = reshape(shape = var_1455, x = v_19)[name = tensor("op_1456")]; + tensor var_1498 = const()[name = tensor("op_1498"), val = tensor([12, 20, 64])]; + tensor var_1499 = reshape(shape = var_1498, x = v_19)[name = tensor("op_1499")]; tensor v_21_perm_0 = const()[name = tensor("v_21_perm_0"), val = tensor([1, 0, 2])]; - tensor var_1459 = const()[name = tensor("op_1459"), val = tensor([5, 4, 12, 64])]; - tensor q_21 = transpose(perm = q_21_perm_0, x = var_1442)[name = tensor("transpose_7")]; - tensor q = reshape(shape = var_1459, x = q_21)[name = tensor("q")]; - tensor var_1461 = const()[name = tensor("op_1461"), val = tensor([5, 4, 12, 64])]; - tensor k_21 = transpose(perm = k_21_perm_0, x = var_1449)[name = tensor("transpose_6")]; - tensor k = reshape(shape = var_1461, x = k_21)[name = tensor("k")]; - tensor var_1463 = const()[name = tensor("op_1463"), val = tensor([5, 4, 12, 64])]; - tensor v_21 = transpose(perm = v_21_perm_0, x = var_1456)[name = tensor("transpose_5")]; - tensor v = reshape(shape = var_1463, x = v_21)[name = tensor("v")]; + tensor var_1502 = const()[name = tensor("op_1502"), val = tensor([5, 4, 12, 64])]; + tensor q_21 = transpose(perm = q_21_perm_0, x = var_1485)[name = tensor("transpose_7")]; + tensor q = reshape(shape = var_1502, x = q_21)[name = tensor("q")]; + tensor var_1504 = const()[name = tensor("op_1504"), val = tensor([5, 4, 12, 64])]; + tensor k_21 = transpose(perm = k_21_perm_0, x = var_1492)[name = tensor("transpose_6")]; + tensor k = reshape(shape = var_1504, x = k_21)[name = tensor("k")]; + tensor var_1506 = const()[name = tensor("op_1506"), val = tensor([5, 4, 12, 64])]; + tensor v_21 = transpose(perm = v_21_perm_0, x = var_1499)[name = tensor("transpose_5")]; + tensor v = reshape(shape = var_1506, x = v_21)[name = tensor("v")]; tensor mul_3_y_0 = const()[name = tensor("mul_3_y_0"), val = tensor(0x1p-3)]; tensor mul_3 = mul(x = q, y = mul_3_y_0)[name = tensor("mul_3")]; tensor matmul_1_transpose_y_0 = const()[name = tensor("matmul_1_transpose_y_0"), val = tensor(true)]; @@ -1317,36 +1336,36 @@ program(1.0) tensor attn_output_9_transpose_x_0 = const()[name = tensor("attn_output_9_transpose_x_0"), val = tensor(false)]; tensor attn_output_9_transpose_y_0 = const()[name = tensor("attn_output_9_transpose_y_0"), val = tensor(false)]; tensor attn_output_9 = matmul(transpose_x = attn_output_9_transpose_x_0, transpose_y = attn_output_9_transpose_y_0, x = softmax_1, y = v)[name = tensor("attn_output_9")]; - tensor var_1466 = const()[name = tensor("op_1466"), val = tensor([2, 0, 1, 3])]; - tensor var_1471 = const()[name = tensor("op_1471"), val = tensor([60, 256])]; - tensor var_1467 = transpose(perm = var_1466, x = attn_output_9)[name = tensor("transpose_4")]; - tensor attn_output_11 = reshape(shape = var_1471, x = var_1467)[name = tensor("attn_output_11")]; - tensor attn_output_13 = linear(bias = decoder_self_attn2_1_out_proj_bias, weight = decoder_self_attn2_1_out_proj_weight, x = attn_output_11)[name = tensor("linear_53")]; - tensor var_1475 = const()[name = tensor("op_1475"), val = tensor([12, 5, 256])]; - tensor attn_output = reshape(shape = var_1475, x = attn_output_13)[name = tensor("attn_output")]; + tensor var_1509 = const()[name = tensor("op_1509"), val = tensor([2, 0, 1, 3])]; + tensor var_1514 = const()[name = tensor("op_1514"), val = tensor([60, 256])]; + tensor var_1510 = transpose(perm = var_1509, x = attn_output_9)[name = tensor("transpose_4")]; + tensor attn_output_11 = reshape(shape = var_1514, x = var_1510)[name = tensor("attn_output_11")]; + tensor attn_output_13 = linear(bias = inner_decoder_self_attn2_1_out_proj_bias, weight = inner_decoder_self_attn2_1_out_proj_weight, x = attn_output_11)[name = tensor("linear_53")]; + tensor var_1518 = const()[name = tensor("op_1518"), val = tensor([12, 5, 256])]; + tensor attn_output = reshape(shape = var_1518, x = attn_output_13)[name = tensor("attn_output")]; tensor sa2_out_perm_0 = const()[name = tensor("sa2_out_perm_0"), val = tensor([1, 0, 2])]; tensor sa2_out = transpose(perm = sa2_out_perm_0, x = attn_output)[name = tensor("transpose_3")]; - tensor input_193 = add(x = query_5, y = sa2_out)[name = tensor("input_193")]; - tensor input_195_axes_0 = const()[name = tensor("input_195_axes_0"), val = tensor([-1])]; - tensor input_195 = layer_norm(axes = input_195_axes_0, beta = decoder_norm21_1_bias, epsilon = var_1050, gamma = decoder_norm21_1_weight, x = input_193)[name = tensor("input_195")]; - tensor input_197 = linear(bias = decoder_linear1_1_bias, weight = decoder_linear1_1_weight, x = input_195)[name = tensor("linear_54")]; - tensor input_199 = relu(x = input_197)[name = tensor("input_199")]; - tensor ffn_out = linear(bias = decoder_linear2_1_bias, weight = decoder_linear2_1_weight, x = input_199)[name = tensor("linear_55")]; - tensor input_201 = add(x = input_195, y = ffn_out)[name = tensor("input_201")]; + tensor input_195 = add(x = query_5, y = sa2_out)[name = tensor("input_195")]; + tensor input_197_axes_0 = const()[name = tensor("input_197_axes_0"), val = tensor([-1])]; + tensor input_197 = layer_norm(axes = input_197_axes_0, beta = inner_decoder_norm21_1_bias, epsilon = var_76, gamma = inner_decoder_norm21_1_weight, x = input_195)[name = tensor("input_197")]; + tensor input_199 = linear(bias = inner_decoder_linear1_1_bias, weight = inner_decoder_linear1_1_weight, x = input_197)[name = tensor("linear_54")]; + tensor input_201 = relu(x = input_199)[name = tensor("input_201")]; + tensor ffn_out = linear(bias = inner_decoder_linear2_1_bias, weight = inner_decoder_linear2_1_weight, x = input_201)[name = tensor("linear_55")]; + tensor input_203 = add(x = input_197, y = ffn_out)[name = tensor("input_203")]; tensor xt_axes_0 = const()[name = tensor("xt_axes_0"), val = tensor([-1])]; - tensor xt = layer_norm(axes = xt_axes_0, beta = decoder_norm22_1_bias, epsilon = var_1050, gamma = decoder_norm22_1_weight, x = input_201)[name = tensor("xt")]; - tensor var_1495 = const()[name = tensor("op_1495"), val = tensor([1, 5, 12, 256])]; - tensor input = reshape(shape = var_1495, x = xt)[name = tensor("input")]; - tensor var_1497 = const()[name = tensor("op_1497"), val = tensor([-1])]; - tensor var_1498 = reduce_l2_norm(axes = var_1497, keep_dims = var_1053, x = input)[name = tensor("op_1498")]; + tensor xt = layer_norm(axes = xt_axes_0, beta = inner_decoder_norm22_1_bias, epsilon = var_76, gamma = inner_decoder_norm22_1_weight, x = input_203)[name = tensor("xt")]; + tensor var_1538 = const()[name = tensor("op_1538"), val = tensor([1, 5, 12, 256])]; + tensor input = reshape(shape = var_1538, x = xt)[name = tensor("input")]; + tensor var_1540 = const()[name = tensor("op_1540"), val = tensor([-1])]; + tensor var_1541 = reduce_l2_norm(axes = var_1540, keep_dims = var_75, x = input)[name = tensor("op_1541")]; tensor const_42 = const()[name = tensor("const_42"), val = tensor(0x1.fffffep+127)]; - tensor clip_5 = clip(alpha = var_1045, beta = const_42, x = var_1498)[name = tensor("clip_5")]; - tensor var_1500 = real_div(x = input, y = clip_5)[name = tensor("op_1500")]; + tensor clip_5 = clip(alpha = var_90, beta = const_42, x = var_1541)[name = tensor("clip_5")]; + tensor var_1543 = real_div(x = input, y = clip_5)[name = tensor("op_1543")]; tensor concat_6 = const()[name = tensor("concat_6"), val = tensor([5, 1, 256])]; tensor reshape_0 = reshape(shape = concat_6, x = emb)[name = tensor("reshape_0")]; tensor transpose_1_perm_0 = const()[name = tensor("transpose_1_perm_0"), val = tensor([0, 1, 3, 2])]; tensor concat_7 = const()[name = tensor("concat_7"), val = tensor([5, 256, 12])]; - tensor transpose_1 = transpose(perm = transpose_1_perm_0, x = var_1500)[name = tensor("transpose_2")]; + tensor transpose_1 = transpose(perm = transpose_1_perm_0, x = var_1543)[name = tensor("transpose_2")]; tensor reshape_1 = reshape(shape = concat_7, x = transpose_1)[name = tensor("reshape_1")]; tensor matmul_2_transpose_x_0 = const()[name = tensor("matmul_2_transpose_x_0"), val = tensor(false)]; tensor matmul_2_transpose_y_0 = const()[name = tensor("matmul_2_transpose_y_0"), val = tensor(false)]; @@ -1357,10 +1376,10 @@ program(1.0) tensor output_end_0 = const()[name = tensor("output_end_0"), val = tensor([1, 5, 11])]; tensor output_end_mask_0 = const()[name = tensor("output_end_mask_0"), val = tensor([true, true, false])]; tensor output = slice_by_index(begin = output_begin_0, end = output_end_0, end_mask = output_end_mask_0, x = reshape_2)[name = tensor("output")]; - tensor probs = sigmoid(x = output)[name = tensor("op_1504")]; - tensor var_1506_axis_0 = const()[name = tensor("op_1506_axis_0"), val = tensor(0)]; - tensor dec_kv_new = stack(axis = var_1506_axis_0, values = (var_1202, nkv))[name = tensor("op_1506")]; - tensor var_1508_axis_0 = const()[name = tensor("op_1508_axis_0"), val = tensor(0)]; - tensor dec_scale_new = stack(axis = var_1508_axis_0, values = (new_scale_9, new_scale))[name = tensor("op_1508")]; + tensor probs = sigmoid(x = output)[name = tensor("op_1547")]; + tensor var_1549_axis_0 = const()[name = tensor("op_1549_axis_0"), val = tensor(0)]; + tensor dec_kv_new = stack(axis = var_1549_axis_0, values = (var_1245, nkv))[name = tensor("op_1549")]; + tensor var_1551_axis_0 = const()[name = tensor("op_1551_axis_0"), val = tensor(0)]; + tensor dec_scale_new = stack(axis = var_1551_axis_0, values = (new_scale_9, new_scale))[name = tensor("op_1551")]; } -> (probs, enc_kv_new, enc_scale_new, enc_conv_cache_new, cnn_window_new, dec_kv_new, dec_scale_new); } \ No newline at end of file diff --git a/optimized/dih3/500ms/ls_eend_dih3_500ms.mlpackage/Data/com.apple.CoreML/model.mlmodel b/optimized/dih3/500ms/ls_eend_dih3_500ms.mlpackage/Data/com.apple.CoreML/model.mlmodel index 55956bf9724e5d26a5683d705abec5e007070969..176f6bc60a62201091f417bee7b486615bb89b1c 100644 --- a/optimized/dih3/500ms/ls_eend_dih3_500ms.mlpackage/Data/com.apple.CoreML/model.mlmodel +++ b/optimized/dih3/500ms/ls_eend_dih3_500ms.mlpackage/Data/com.apple.CoreML/model.mlmodel @@ -1,3 +1,3 @@ version https://git-lfs.github.com/spec/v1 -oid sha256:caa9e47ded51d738809bca4aef08aede86aea03f444055505cb42b1a9539bb14 -size 196629 +oid sha256:c4b74a9b812800e63bff390dd09aee47559545f428d8bb10703a36598cefc708 +size 203229 diff --git a/optimized/dih3/500ms/ls_eend_dih3_500ms.mlpackage/Manifest.json b/optimized/dih3/500ms/ls_eend_dih3_500ms.mlpackage/Manifest.json index eb273aebdeb752ae1ebf3dffb84bfed3c48e15a9..f0929c8403f40931887ab605b60dabc479f935aa 100644 --- a/optimized/dih3/500ms/ls_eend_dih3_500ms.mlpackage/Manifest.json +++ b/optimized/dih3/500ms/ls_eend_dih3_500ms.mlpackage/Manifest.json @@ -1,18 +1,18 @@ { "fileFormatVersion": "1.0.0", "itemInfoEntries": { - "36D483E8-31EB-4224-A050-837C09E85C85": { - "author": "com.apple.CoreML", - "description": "CoreML Model Specification", - "name": "model.mlmodel", - "path": "com.apple.CoreML/model.mlmodel" - }, - "F67F7E02-4DC3-4166-A52A-2D9FE21F41D8": { + "70FFA018-D34B-4B71-9B7D-3D3754E252B7": { "author": "com.apple.CoreML", "description": "CoreML Model Weights", "name": "weights", "path": "com.apple.CoreML/weights" + }, + "CE8A955D-9842-41C5-A7BC-A91B484DBFCD": { + "author": "com.apple.CoreML", + "description": "CoreML Model Specification", + "name": "model.mlmodel", + "path": "com.apple.CoreML/model.mlmodel" } }, - "rootModelIdentifier": "36D483E8-31EB-4224-A050-837C09E85C85" + "rootModelIdentifier": "CE8A955D-9842-41C5-A7BC-A91B484DBFCD" }